{
 "S1sh4::signage": {
  "fp": "d69c7cf46dd44ae9",
  "inscriptions": [],
  "cues": [],
  "dropped": []
 },
 "era_assess::111ff5bcb749c1a3": {
  "subjects": [
   {
    "subject_native": "20세기 초 전라도 시골 오일장 좌판 거리",
    "search_terms_native": [
     "일제강점기 장터 좌판 사진",
     "근대 한국 시골 장터 풍경",
     "20세기 초 5일장 옹기전"
    ],
    "language_lock_native": "반드시 한국어로만 검색해야 하며 다른 언어로 번역하거나 외래어를 섞지 마십시오.",
    "reason_ko": "생성 모델이 민속촌 세트장이나 판타지 사극 풍으로 왜곡하기 쉬우며 거적을 깐 실제 좌판 배치와 장터 구조 고증이 필요함."
   }
  ],
  "subject_text": "복내장터 좌판 거리·장터 입구·골목\n흙바닥 길 양쪽에 옹기전, 짚신전, 포목전과 생선 좌판이 빽빽한 장터. 입구와 좁은 골목이 이어지고 국밥집 연기가 오른다.",
  "identity": "canonical",
  "scope_id": "L01",
  "scope_role": "location_exterior",
  "scope_sha": "002e59ee82cfcbeb"
 },
 "era_fail::335670d9fbadced3": {
  "stage": "research",
  "subject": "20세기 초 전라도 시골 오일장 좌판 거리",
  "terms": [
   "일제강점기 장터 좌판 사진",
   "근대 한국 시골 장터 풍경",
   "20세기 초 5일장 옹기전"
  ],
  "status": "no_usable",
  "queries": [
   [
    "일제강점기 전라도 장터 옹기전 사진"
   ],
   [
    "\"옹기 시장\" \"사진\" \"1900\"",
    "\"전주\" \"장터\" \"옛 사진\""
   ]
  ],
  "candidate_urls": [
   "https://img.koreatimes.co.kr/upload/thumbnailV2/K2019041300016.jpg",
   "https://blog.kakaocdn.net/dna/AFlTC/btsIUSZVbRb/AAAAAAAAAAAAAAAAAAAAAILfeQbfx8dSG0izVMcSGpc9s0tGfieBbi0SDCz0DQNr/img.jpg?allow_ip=&allow_referer=&credential=yqXZFxpELC7KVnFOS48ylbz2pIh7yKj8&expires=1777561199&signature=zUNYorwchUiQ6XuKCuruBfTzQHg%3D",
   "https://www.koreaimg.com/data/rental/goods/middle/original%5B1%5D%5B2%5D.jpg",
   "https://i1.media.daumcdn.net/uf/image/U01/agora/538FADE14D4C14002E"
  ],
  "coarse": {
   "eligible": [],
   "chosen_index": 0,
   "reason": "종류·보임·기준을 다 만족하는 후보가 없다 — 이 라운드에선 안 고른다",
   "single_judge": true,
   "rejected_judges": {}
  },
  "verdicts": [
   {
    "index": 1,
    "object_type_match": "yes",
    "visible": true,
    "criteria_match": "unsure",
    "similarity": 95
   },
   {
    "index": 2,
    "object_type_match": "no",
    "visible": true,
    "criteria_match": "unsure",
    "similarity": 30
   },
   {
    "index": 3,
    "object_type_match": "yes",
    "visible": true,
    "criteria_match": "unsure",
    "similarity": 90
   },
   {
    "index": 4,
    "object_type_match": "yes",
    "visible": true,
    "criteria_match": "unsure",
    "similarity": 75
   }
  ],
  "chosen_reason_ko": "",
  "attempts": 5
 },
 "groupbg::market_pottery_front": {
  "input_fingerprint": "f8a03c78e2eb26ec",
  "meta": {
   "model": "gpt-image-2",
   "size": "1536x864",
   "pack": "11.202607220237",
   "contract": "bgfirst_full_v3",
   "group_sig": {
    "key": "market_pottery_front",
    "tags": [
     "S1sh4"
    ]
   },
   "context_sig": "38f6b8eb84f57060"
  },
  "prompt": "Create ONE empty live-action location background photograph — NO PEOPLE, no figures, no body parts, no silhouettes, no shadows or reflections of people anywhere.\n\"Empty\" means no people only: KEEP the location's inherent occupants and stock that define the place — animals in an animal shelter, pen or farm, goods and produce in a market, moored boats in a harbour — unless the shot text explicitly removes them.\nThe FIRST attached image is a thin-line storyboard sketch of a shot that happens at this location: use it ONLY as spatial evidence — what this place contains, how its ground, structures and landmarks are arranged and proportioned. Ignore the sketched people and arrows entirely, and do NOT copy its line style: render a fully photographic, physically plausible real place that fits THE LOCATION text below.\nHUMAN-SCALE CALIBRATION: derive every structure's true size from human-scale elements — a door ≈ 2m, a window ≈ 1–1.5m wide, one storey ≈ 2.5–3m; never inflate a small structure or shrink a large one.\n\nTHE LOCATION — early 20th-century Korea, Jeolla region; all people are Korean unless explicitly stated otherwise: The outdoor frontage of an earthenware stall in a crowded market, under harsh midday sun.\n\nLOCATION DETAIL (descriptive evidence about this place from the production's location records — evidence, not a staging order): use it to understand what kind of place this is and what it permanently contains. Carry over ONLY its enduring physical features — layout, structures, ground and materials, fixed equipment and installations. Do NOT treat any lighting, weather or time-of-day wording, momentary object states, depicted events or subjective impressions in it as fixed facts of the place; the TIME OF DAY lock below is the sole authority for time and lighting.\n복내장터 좌판 거리·장터 입구·골목: 다양한 좌판과 짐꾼, 상인들이 밀집한 넓은 흙바닥 노천 장터. 포목전, 짚신전, 옹기전, 닭장, 생선 좌판이 빽빽하게 들어서 있으며, 벽에 관의 방이 붙어 있고 상황에 따라 찢어진 멍석과 부서진 좌판이 널브러져 있다. (특징: 넓은 흙바닥; 나무와 짚으로 얽어 만든 노점 좌판 (옹기, 짚신, 채소, 포목 등); 나무 닭장과 광주리; 방(벽보)이 붙어 있는 벽; 깨진 항아리와 뒤집힌 멍석, 부서진 나무판자들)\n\nSCENE EVIDENCE (verbatim quotes from the screenplay about this place — treat them as evidence of what the location physically contains and looks like; stage the PLACE those moments happen in, but do NOT depict the momentary actions, people or staged props themselves):\n- 옹기전 앞, 이관회. 헌병 장교의 총구가 천천히 그의 턱밑으로 올라온다.\n\nTIME OF DAY (lock): day, harsh sunlight.\n\nRender ONE photorealistic empty location photograph, 16:9, neutral enough that every shot of this place can be staged from it later. No readable writing anywhere: surfaces that would carry writing may be present, but stage any wording out of legibility — an oblique angle, distance, shallow focus. No captions, watermarks or overlay text.",
  "effective_prompt": "Create ONE empty live-action location background photograph — NO PEOPLE, no figures, no body parts, no silhouettes, no shadows or reflections of people anywhere.\n\"Empty\" means no people only: KEEP the location's inherent occupants and stock that define the place — animals in an animal shelter, pen or farm, goods and produce in a market, moored boats in a harbour — unless the shot text explicitly removes them.\nThe FIRST attached image is a thin-line storyboard sketch of a shot that happens at this location: use it ONLY as spatial evidence — what this place contains, how its ground, structures and landmarks are arranged and proportioned. Ignore the sketched people and arrows entirely, and do NOT copy its line style: render a fully photographic, physically plausible real place that fits THE LOCATION text below.\nHUMAN-SCALE CALIBRATION: derive every structure's true size from human-scale elements — a door ≈ 2m, a window ≈ 1–1.5m wide, one storey ≈ 2.5–3m; never inflate a small structure or shrink a large one.\n\nTHE LOCATION — early 20th-century Korea, Jeolla region; all people are Korean unless explicitly stated otherwise: The outdoor frontage of an earthenware stall in a crowded market, under harsh midday sun.\n\nLOCATION DETAIL (descriptive evidence about this place from the production's location records — evidence, not a staging order): use it to understand what kind of place this is and what it permanently contains. Carry over ONLY its enduring physical features — layout, structures, ground and materials, fixed equipment and installations. Do NOT treat any lighting, weather or time-of-day wording, momentary object states, depicted events or subjective impressions in it as fixed facts of the place; the TIME OF DAY lock below is the sole authority for time and lighting.\n복내장터 좌판 거리·장터 입구·골목: 다양한 좌판과 짐꾼, 상인들이 밀집한 넓은 흙바닥 노천 장터. 포목전, 짚신전, 옹기전, 닭장, 생선 좌판이 빽빽하게 들어서 있으며, 벽에 관의 방이 붙어 있고 상황에 따라 찢어진 멍석과 부서진 좌판이 널브러져 있다. (특징: 넓은 흙바닥; 나무와 짚으로 얽어 만든 노점 좌판 (옹기, 짚신, 채소, 포목 등); 나무 닭장과 광주리; 방(벽보)이 붙어 있는 벽; 깨진 항아리와 뒤집힌 멍석, 부서진 나무판자들)\n\nSCENE EVIDENCE (verbatim quotes from the screenplay about this place — treat them as evidence of what the location physically contains and looks like; stage the PLACE those moments happen in, but do NOT depict the momentary actions, people or staged props themselves):\n- 옹기전 앞, 이관회. 헌병 장교의 총구가 천천히 그의 턱밑으로 올라온다.\n\nTIME OF DAY (lock): day, harsh sunlight.\n\nRender ONE photorealistic empty location photograph, 16:9, neutral enough that every shot of this place can be staged from it later. No readable writing anywhere: surfaces that would carry writing may be present, but stage any wording out of legibility — an oblique angle, distance, shallow focus. No captions, watermarks or overlay text.",
  "bg_path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/db00517b-165b-4005-a287-f28936b59b95/images/272bd7af-9294-4490-98c6-88a40e9deacc/scene/recipe/groupbg_market_pottery_front_836383.png",
  "asset_id": "1df6d064-1b12-4716-9367-28fcecd83b8a",
  "input_asset_ids": [
   "c6ee6c62-d713-4a4c-bed5-b419c1abb827"
  ],
  "origin_tag": "S1sh4",
  "place_text": "The outdoor frontage of an earthenware stall in a crowded market, under harsh midday sun.",
  "origin_inputs": {
   "place_text": "The outdoor frontage of an earthenware stall in a crowded market, under harsh midday sun.",
   "time_of_day_en": "day, harsh sunlight",
   "conti_asset_id": "c6ee6c62-d713-4a4c-bed5-b419c1abb827"
  }
 },
 "S1sh4::bgfirst_bg": {
  "input_fingerprint": "a6913f2837a8b9aa",
  "prompt": "Create the EMPTY BACKGROUND PLATE for one film shot — NO PEOPLE, no figures, no body parts, no sketch lines, no arrows anywhere.\n\"Empty\" means no people only: KEEP the location's inherent occupants and stock that define the place — animals in an animal shelter, pen or farm, goods in a market, moored boats in a harbour — unless the shot text explicitly removes them.\nThe FIRST attached image is a thin-line storyboard sketch: use ONLY its camera angle, horizon, perspective and the placement/size of buildings and set masses — ignore the sketched people and arrows entirely. The SECOND attached image (LOCATION PHOTOGRAPH) is the real place: take its architecture, materials, signage and fixed features, and RE-PROJECT them into the sketch's camera. If the photograph's camera differs from the sketch's, the sketch's camera wins.\nHUMAN-SCALE CALIBRATION: derive every structure's true size from human-scale elements — a door ≈ 2m, a window ≈ 1–1.5m wide, one storey ≈ 2.5–3m; never inflate a small structure or shrink a large one.\n\nSHOT TEXT this background must serve (Korean): 헌병 장교의 총구가 이관회의 턱밑 살에 바짝 눌린 측면 클로즈업.\n\nLOCATION (lock): The outdoor frontage of an earthenware stall in a crowded market, under harsh midday sun.\n\nTIME OF DAY (lock): day, harsh sunlight.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: close-up\n- KEY BACKGROUND ELEMENTS: 장교의 총 (Muzzle pressed tightly beneath 이관회's chin) — Seen from the side, with its muzzle angled upward toward the underside of the chin; used as A narrow diagonal connects the officer's hand to the point of pressure; the weapon occupies substantially less space than the human features; 옹기 좌판 (Present behind the confrontation); used as Soft peripheral context preserves the market location without competing with the contact.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Harsh midday sunlight retains natural skin contrast and restrained saturation, making the threat stark without stylized lighting.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nRender ONE photorealistic empty location photograph, 16:9, that this shot can be staged inside later. No readable writing anywhere: surfaces that would carry writing may be present, but stage any wording out of legibility — an oblique angle, distance, shallow focus. No captions, watermarks or overlay text.",
  "effective_prompt": "Create the EMPTY BACKGROUND PLATE for one film shot — NO PEOPLE, no figures, no body parts, no sketch lines, no arrows anywhere.\n\"Empty\" means no people only: KEEP the location's inherent occupants and stock that define the place — animals in an animal shelter, pen or farm, goods in a market, moored boats in a harbour — unless the shot text explicitly removes them.\nThe FIRST attached image is a thin-line storyboard sketch: use ONLY its camera angle, horizon, perspective and the placement/size of buildings and set masses — ignore the sketched people and arrows entirely. The SECOND attached image (LOCATION PHOTOGRAPH) is the real place: take its architecture, materials, signage and fixed features, and RE-PROJECT them into the sketch's camera. If the photograph's camera differs from the sketch's, the sketch's camera wins.\nHUMAN-SCALE CALIBRATION: derive every structure's true size from human-scale elements — a door ≈ 2m, a window ≈ 1–1.5m wide, one storey ≈ 2.5–3m; never inflate a small structure or shrink a large one.\n\nSHOT TEXT this background must serve (Korean): 헌병 장교의 총구가 이관회의 턱밑 살에 바짝 눌린 측면 클로즈업.\n\nLOCATION (lock): The outdoor frontage of an earthenware stall in a crowded market, under harsh midday sun.\n\nTIME OF DAY (lock): day, harsh sunlight.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: close-up\n- KEY BACKGROUND ELEMENTS: 장교의 총 (Muzzle pressed tightly beneath 이관회's chin) — Seen from the side, with its muzzle angled upward toward the underside of the chin; used as A narrow diagonal connects the officer's hand to the point of pressure; the weapon occupies substantially less space than the human features; 옹기 좌판 (Present behind the confrontation); used as Soft peripheral context preserves the market location without competing with the contact.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Harsh midday sunlight retains natural skin contrast and restrained saturation, making the threat stark without stylized lighting.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nRender ONE photorealistic empty location photograph, 16:9, that this shot can be staged inside later. No readable writing anywhere: surfaces that would carry writing may be present, but stage any wording out of legibility — an oblique angle, distance, shallow focus. No captions, watermarks or overlay text.",
  "bg_path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/db00517b-165b-4005-a287-f28936b59b95/images/272bd7af-9294-4490-98c6-88a40e9deacc/scene/recipe/S1sh4__bgfirst_bg.png",
  "asset_id": "23232690-4c79-45b6-887a-266366c69dc4",
  "input_asset_ids": [
   "c6ee6c62-d713-4a4c-bed5-b419c1abb827",
   "1df6d064-1b12-4716-9367-28fcecd83b8a"
  ]
 },
 "S1sh4": {
  "input_fingerprint": "88c2fec8a8868ff0",
  "prompt": "Create ONE FINAL photorealistic live-action film still of the moment below — early 20th-century Korea, Jeolla region; all people are Korean unless explicitly stated otherwise. TIME OF DAY (lock): day, harsh sunlight.\n\nSHOT TEXT (authoritative, Korean): 헌병 장교의 총구가 이관회의 턱밑 살에 바짝 눌린 측면 클로즈업.\n\nLOCATION (lock): The outdoor frontage of an earthenware stall in a crowded market, under harsh midday sun. The shot takes place here — the attached LOCATION PHOTOGRAPH shows the exact spot.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: close-up\n- KEY BACKGROUND ELEMENTS: 장교의 총 (Muzzle pressed tightly beneath 이관회's chin) — Seen from the side, with its muzzle angled upward toward the underside of the chin; used as A narrow diagonal connects the officer's hand to the point of pressure; the weapon occupies substantially less space than the human features; 옹기 좌판 (Present behind the confrontation); used as Soft peripheral context preserves the market location without competing with the contact.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Harsh midday sunlight retains natural skin contrast and restrained saturation, making the threat stark without stylized lighting.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): Harsh midday sunlight covers the market; smoke rises beside the pottery stalls. 이관회: He stands in front of the pottery stall with a muzzle pressed beneath his chin.\n\nTHE HAND THAT IS DOING THIS: the object at the centre of this shot is being held, operated, read, aimed or handed over by 헌병 장교 right now, so 헌병 장교's hand — and as much of the wrist and forearm as the framing reaches — is in the frame, gripping or resting on that object exactly the way its real use requires. Match that hand to 헌병 장교: its size, build, skin, age, grooming, sleeve and anything worn on it belong to that person and to no one else.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 이관회 (한국인 남성, 청년 나이대, 자연스러운 검은 머리) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.",
  "roll_prompts": {
   "A": "Stage the shot. The FIRST attached image (SHOT BACKGROUND) is the finished empty background of this shot — keep it EXACTLY: its camera, perspective, architecture, lighting and every fixed feature stay untouched. The SECOND attached image (LAYOUT SKETCH) tells you ONLY where the people go: each sketched person's position, screen size, pose and the gaze/motion arrows. Ignore the sketch's background lines. The CHARACTER REFERENCE photographs show the real people.\nPlace the real people into the background at exactly the sketched positions, sizes and poses, following the arrow directions. No sketch lines or arrows may remain.\n\nCreate ONE FINAL photorealistic live-action film still of the moment below — early 20th-century Korea, Jeolla region; all people are Korean unless explicitly stated otherwise. TIME OF DAY (lock): day, harsh sunlight.\n\nSHOT TEXT (authoritative, Korean): 헌병 장교의 총구가 이관회의 턱밑 살에 바짝 눌린 측면 클로즈업.\n\nLOCATION (lock): The outdoor frontage of an earthenware stall in a crowded market, under harsh midday sun. The shot takes place here — the FIRST attached image (SHOT BACKGROUND) is this exact place, already built: its ground, structures, horizon, materials and lighting are the finished truth of this location and must not be redesigned or replaced. No location photograph is attached — read the place from that image alone, and add no scenery, structure, vehicle or fixture that it does not already show. This lock governs the place only; the figures in the shot follow the staging and pose instructions.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: close-up\n- KEY BACKGROUND ELEMENTS: 장교의 총 (Muzzle pressed tightly beneath 이관회's chin) — Seen from the side, with its muzzle angled upward toward the underside of the chin; used as A narrow diagonal connects the officer's hand to the point of pressure; the weapon occupies substantially less space than the human features; 옹기 좌판 (Present behind the confrontation); used as Soft peripheral context preserves the market location without competing with the contact.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Harsh midday sunlight retains natural skin contrast and restrained saturation, making the threat stark without stylized lighting.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): Harsh midday sunlight covers the market; smoke rises beside the pottery stalls. 이관회: He stands in front of the pottery stall with a muzzle pressed beneath his chin.\n\nTHE HAND THAT IS DOING THIS: the object at the centre of this shot is being held, operated, read, aimed or handed over by 헌병 장교 right now, so 헌병 장교's hand — and as much of the wrist and forearm as the framing reaches — is in the frame, gripping or resting on that object exactly the way its real use requires. Match that hand to 헌병 장교: its size, build, skin, age, grooming, sleeve and anything worn on it belong to that person and to no one else.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 이관회 (한국인 남성, 청년 나이대, 자연스러운 검은 머리) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.",
   "B": "Create ONE FINAL photorealistic live-action film still of the moment below — early 20th-century Korea, Jeolla region; all people are Korean unless explicitly stated otherwise. TIME OF DAY (lock): day, harsh sunlight.\n\nSHOT TEXT (authoritative, Korean): 헌병 장교의 총구가 이관회의 턱밑 살에 바짝 눌린 측면 클로즈업.\n\nLOCATION (lock): The outdoor frontage of an earthenware stall in a crowded market, under harsh midday sun. The shot takes place here — the attached LOCATION PHOTOGRAPH shows the exact spot.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: close-up\n- KEY BACKGROUND ELEMENTS: 장교의 총 (Muzzle pressed tightly beneath 이관회's chin) — Seen from the side, with its muzzle angled upward toward the underside of the chin; used as A narrow diagonal connects the officer's hand to the point of pressure; the weapon occupies substantially less space than the human features; 옹기 좌판 (Present behind the confrontation); used as Soft peripheral context preserves the market location without competing with the contact.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Harsh midday sunlight retains natural skin contrast and restrained saturation, making the threat stark without stylized lighting.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): Harsh midday sunlight covers the market; smoke rises beside the pottery stalls. 이관회: He stands in front of the pottery stall with a muzzle pressed beneath his chin.\n\nTHE HAND THAT IS DOING THIS: the object at the centre of this shot is being held, operated, read, aimed or handed over by 헌병 장교 right now, so 헌병 장교's hand — and as much of the wrist and forearm as the framing reaches — is in the frame, gripping or resting on that object exactly the way its real use requires. Match that hand to 헌병 장교: its size, build, skin, age, grooming, sleeve and anything worn on it belong to that person and to no one else.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 이관회 (한국인 남성, 청년 나이대, 자연스러운 검은 머리) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.\n\nCOMPOSITION VARIATION (this candidate only): keep the place, the people and the moment exactly as specified, and keep from the CAMERA & FRAME contract its subject scale, screen placement and key background elements exactly. For this candidate only, its camera angle and height are deliberately left open: choose a camera position distinctly different from the obvious one, as a film director picking a second setup on the same blocking. Stay true to any explicit framing the SHOT TEXT itself dictates."
  },
  "roll_refs": {
   "A": [
    {
     "label": "SHOT BACKGROUND",
     "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/db00517b-165b-4005-a287-f28936b59b95/images/272bd7af-9294-4490-98c6-88a40e9deacc/scene/recipe/S1sh4__bgfirst_bg.png",
     "asset_id": "23232690-4c79-45b6-887a-266366c69dc4",
     "role": "bgfirst_bg"
    },
    {
     "label": "LAYOUT SKETCH (people placement only)",
     "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/db00517b-165b-4005-a287-f28936b59b95/images/272bd7af-9294-4490-98c6-88a40e9deacc/conti/conti_S1sh4.png",
     "asset_id": "c6ee6c62-d713-4a4c-bed5-b419c1abb827",
     "role": "conti_light"
    },
    {
     "label": "CHARACTER REFERENCE — 이관회: the exact person appearing in this shot; match face, hair and build exactly.",
     "path": "<bytes:933540>",
     "asset_id": "851b450f-ac84-47fe-ac5a-03a56f63fcae",
     "role": "character_ref"
    },
    {
     "label": "PROP REFERENCE — 근대식 소총: the exact object appearing in this shot; match its look, material and wear exactly.",
     "path": "<bytes:828568>",
     "asset_id": "7543806c-8087-4caa-8d7b-56d10864fd3c",
     "role": "prop_ref"
    }
   ],
   "B": [
    {
     "label": "LOCATION PHOTOGRAPH — the exact place of this shot: its architecture, materials, fixed features and lighting mood are spatial truth; stage the moment inside this place. Never copy its camera framing.",
     "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/db00517b-165b-4005-a287-f28936b59b95/images/272bd7af-9294-4490-98c6-88a40e9deacc/scene/recipe/groupbg_market_pottery_front_836383.png",
     "asset_id": "1df6d064-1b12-4716-9367-28fcecd83b8a",
     "role": "bgfirst_group_bg"
    },
    {
     "label": "CHARACTER REFERENCE — 이관회: the exact person appearing in this shot; match face, hair and build exactly.",
     "path": "<bytes:933540>",
     "asset_id": "851b450f-ac84-47fe-ac5a-03a56f63fcae",
     "role": "character_ref"
    },
    {
     "label": "PROP REFERENCE — 근대식 소총: the exact object appearing in this shot; match its look, material and wear exactly.",
     "path": "<bytes:828568>",
     "asset_id": "7543806c-8087-4caa-8d7b-56d10864fd3c",
     "role": "prop_ref"
    }
   ]
  },
  "shot_run_spend_attempt_count": 1,
  "judge_flip_skipped": "judge_fn_owns_order",
  "readings": [
   {
    "label": "A",
    "direction": "장교의 시선은 이관회를 향하고 있으며, 총구는 위를 향해 이관회의 턱밑 살에 바짝 눌려 있다.",
    "built_space": "배경에 옹기 좌판이 있는 야외 시장이 보이며 위치 레퍼런스와 일치한다.",
    "entities": "이관회의 외모는 레퍼런스와 일치한다. 지시문에서 허용하지 않은 장교의 얼굴과 상체가 프레임에 나타난다. 총은 권총 부품과 소총 개머리판이 섞인 알 수 없는 형태다.",
    "hard_violations": [
     "[gemini-pro] 허용되지 않은 인물(장교의 얼굴과 상체) 등장",
     "[gemini-pro] 물리적으로 불가능한 구조의 발명된 사물(기괴하게 섞인 총)"
    ],
    "physics": "장교의 오른손이 권총 손잡이를 쥐고 있으나, 개머리판이 손목 아래로 뒤쪽을 향해 뻗어 있어 총을 제대로 지지하거나 사용할 수 없는 물리적으로 불가능한 구조다."
   },
   {
    "label": "B",
    "direction": "총열이 수평을 유지한 채 이관회의 턱밑이 아닌 목(목젖 부근)을 누르고 있다.",
    "built_space": "흐릿한 배경에 옹기 좌판과 야외 시장의 모습이 보이며 위치 레퍼런스와 일치한다.",
    "entities": "이관회의 외모는 레퍼런스와 일치한다. 총을 쥔 팔의 소매가 장교복이 아닌 이관회와 같은 거친 삼베 재질 옷감이다. 총은 레퍼런스에 비해 지나치게 두껍고 거대하며 총구 끝부분의 디테일이 생략되었다.",
    "hard_violations": [
     "[gpt-high] 화면 왼쪽 시장 통로에 샷 텍스트가 등장시키지 않은 배경 인물 두 명이 추가되어 있다."
    ],
    "physics": "왼손이 총열 위를 덮어쥐는 오버핸드 그립으로 허공에서 총을 지탱하고 있으며, 해당 손의 해부학적 구조와 자세는 물리적으로 가능하다."
   }
  ],
  "cross_model_order": {
   "policy": "cross_model_order_v2_slot0_fwd_slot1_rev_parallel2",
   "slots": [
    {
     "model": "gemini-pro",
     "order": "forward",
     "display_to_canonical": {
      "A": "A",
      "B": "B"
     },
     "raw": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 6,
        "verdict_ko": "비록 총구가 턱밑이 아닌 목을 향하고 총을 쥔 손의 소매가 평복이라는 단점이 있으나, 치명적인 물리적 오류나 규칙 위반 없이 클로즈업 구도를 안정적으로 구현했습니다."
       },
       {
        "label": "A",
        "score": 2,
        "verdict_ko": "총의 형태가 권총과 소총이 기괴하게 섞인 물리적으로 불가능한 구조이며, 지시문이 허용하지 않은 장교의 얼굴과 상체가 등장하여 심각한 규칙 위반입니다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "장교의 시선은 이관회를 향하고 있으며, 총구는 위를 향해 이관회의 턱밑 살에 바짝 눌려 있다.",
        "built_space": "배경에 옹기 좌판이 있는 야외 시장이 보이며 위치 레퍼런스와 일치한다.",
        "entities": "이관회의 외모는 레퍼런스와 일치한다. 지시문에서 허용하지 않은 장교의 얼굴과 상체가 프레임에 나타난다. 총은 권총 부품과 소총 개머리판이 섞인 알 수 없는 형태다.",
        "hard_violations": [
         "허용되지 않은 인물(장교의 얼굴과 상체) 등장",
         "물리적으로 불가능한 구조의 발명된 사물(기괴하게 섞인 총)"
        ],
        "physics": "장교의 오른손이 권총 손잡이를 쥐고 있으나, 개머리판이 손목 아래로 뒤쪽을 향해 뻗어 있어 총을 제대로 지지하거나 사용할 수 없는 물리적으로 불가능한 구조다."
       },
       {
        "label": "B",
        "direction": "총열이 수평을 유지한 채 이관회의 턱밑이 아닌 목(목젖 부근)을 누르고 있다.",
        "built_space": "흐릿한 배경에 옹기 좌판과 야외 시장의 모습이 보이며 위치 레퍼런스와 일치한다.",
        "entities": "이관회의 외모는 레퍼런스와 일치한다. 총을 쥔 팔의 소매가 장교복이 아닌 이관회와 같은 거친 삼베 재질 옷감이다. 총은 레퍼런스에 비해 지나치게 두껍고 거대하며 총구 끝부분의 디테일이 생략되었다.",
        "hard_violations": [],
        "physics": "왼손이 총열 위를 덮어쥐는 오버핸드 그립으로 허공에서 총을 지탱하고 있으며, 해당 손의 해부학적 구조와 자세는 물리적으로 가능하다."
       }
      ],
      "all_candidates_fail": false
     },
     "normalized": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 6,
        "verdict_ko": "비록 총구가 턱밑이 아닌 목을 향하고 총을 쥔 손의 소매가 평복이라는 단점이 있으나, 치명적인 물리적 오류나 규칙 위반 없이 클로즈업 구도를 안정적으로 구현했습니다."
       },
       {
        "label": "A",
        "score": 2,
        "verdict_ko": "총의 형태가 권총과 소총이 기괴하게 섞인 물리적으로 불가능한 구조이며, 지시문이 허용하지 않은 장교의 얼굴과 상체가 등장하여 심각한 규칙 위반입니다."
       }
      ],
      "all_candidates_fail": false,
      "readings": [
       {
        "label": "A",
        "direction": "장교의 시선은 이관회를 향하고 있으며, 총구는 위를 향해 이관회의 턱밑 살에 바짝 눌려 있다.",
        "built_space": "배경에 옹기 좌판이 있는 야외 시장이 보이며 위치 레퍼런스와 일치한다.",
        "entities": "이관회의 외모는 레퍼런스와 일치한다. 지시문에서 허용하지 않은 장교의 얼굴과 상체가 프레임에 나타난다. 총은 권총 부품과 소총 개머리판이 섞인 알 수 없는 형태다.",
        "hard_violations": [
         "허용되지 않은 인물(장교의 얼굴과 상체) 등장",
         "물리적으로 불가능한 구조의 발명된 사물(기괴하게 섞인 총)"
        ],
        "physics": "장교의 오른손이 권총 손잡이를 쥐고 있으나, 개머리판이 손목 아래로 뒤쪽을 향해 뻗어 있어 총을 제대로 지지하거나 사용할 수 없는 물리적으로 불가능한 구조다."
       },
       {
        "label": "B",
        "direction": "총열이 수평을 유지한 채 이관회의 턱밑이 아닌 목(목젖 부근)을 누르고 있다.",
        "built_space": "흐릿한 배경에 옹기 좌판과 야외 시장의 모습이 보이며 위치 레퍼런스와 일치한다.",
        "entities": "이관회의 외모는 레퍼런스와 일치한다. 총을 쥔 팔의 소매가 장교복이 아닌 이관회와 같은 거친 삼베 재질 옷감이다. 총은 레퍼런스에 비해 지나치게 두껍고 거대하며 총구 끝부분의 디테일이 생략되었다.",
        "hard_violations": [],
        "physics": "왼손이 총열 위를 덮어쥐는 오버핸드 그립으로 허공에서 총을 지탱하고 있으며, 해당 손의 해부학적 구조와 자세는 물리적으로 가능하다."
       }
      ]
     },
     "ok": true
    },
    {
     "model": "gpt-high",
     "order": "reverse",
     "display_to_canonical": {
      "A": "B",
      "B": "A"
     },
     "raw": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 3,
        "verdict_ko": "측면 클로즈업과 연기는 맞지만, 불필요한 배경 인물 두 명이 추가됐고 총구가 턱밑을 향해 위로 압박하는 대신 아래로 기울어져 결정적인 위협 동작이 어긋납니다."
       },
       {
        "label": "B",
        "score": 7,
        "verdict_ko": "장교가 잡은 총이 위쪽 대각선으로 이관회의 턱밑을 압박하는 핵심 관계는 맞지만, 배경이 지나치게 선명하고 소총의 형태·비례가 참조와 다릅니다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "이관회는 화면 왼쪽의 총을 든 사람 쪽을 바라본다. 총은 왼쪽 손에서 오른쪽 턱으로 이어지지만 총구 쪽으로 내려가는 기울기다. 총열 윗면이 턱 아래에 닿고 총구 끝은 접촉점보다 오른쪽으로 돌출되어 보여, 총구가 위를 향해 턱밑 살에 바짝 눌리는 요구와 다르다.",
        "built_space": "오른쪽 뒤에 옹기를 놓은 목재 선반 한 구역, 아래쪽에 낮은 목재 진열대 한 줄, 왼쪽에 흙길과 시장 가설물이 보인다. 선반의 전체 단수는 얼굴과 흐림 때문에 확인하기 어렵다. 옹기는 진열대에 놓여 있고 가게 앞 인물의 위치도 공간과 연결된다. 연기가 진열대 옆에서 올라오며 배경은 비교적 부드럽다. 중복된 고정 설비나 반사는 보이지 않는다.",
        "entities": "이관회는 검은 가르마 머리의 한국인 청년 남성으로 보이며, 참조와 유사한 얼굴 및 거친 베이지색 저고리를 갖췄다. 왼쪽에는 성인 손과 손목, 갈색 소매가 보이지만 군복 식별 요소가 없어 헌병 장교의 손이라는 확인은 약하다. 소총의 낡은 목재, 긴 금속 총열, 황동 띠는 참조와 가깝지만 참조의 긴 총검은 보이지 않고 총구 아래 부속 형태도 다르다. 왼쪽 먼 배경에는 지시되지 않은 사람 두 명이 보인다. 읽을 수 있는 문구는 없다.",
        "hard_violations": [
         "화면 왼쪽 시장 통로에 샷 텍스트가 등장시키지 않은 배경 인물 두 명이 추가되어 있다."
        ],
        "physics": "왼쪽 손이 소총의 목재 앞부분을 아래에서 받치고 손가락으로 감싸므로 총의 지지는 확인된다. 이관회의 머리는 목과 상체에 자연스럽게 연결되어 있고 부유하는 신체나 물체는 없다. 다만 턱을 받치는 총열 옆면의 접촉에 가까워, 총구 끝으로 턱밑을 위로 밀어붙이는 힘의 방향은 구현되지 않았다."
       },
       {
        "label": "B",
        "direction": "장교와 이관회는 서로의 얼굴 쪽을 바라본다. 장교의 손에서 시작하는 총은 오른쪽 위로 올라가며 끝이 이관회의 턱밑에 밀착한다. 총구 끝의 일부는 턱에 가려지지만 압박 대상과 위쪽 조준 방향은 분명하다.",
        "built_space": "오른쪽에는 네 단가량이 드러나는 목재 옹기 선반 한 구역과 앞쪽의 낮은 진열대 한 줄이 있고, 위에는 초가 처마와 말아 올린 발이 보인다. 왼쪽으로 흙길과 목재 시장 가설물이 이어져 참조 장소의 재료와 배치에 부합한다. 이관회는 좌판 바로 앞, 장교는 통로 쪽에 서 있다. 고정 설비의 부당한 중복이나 불가능한 반사는 없다. 다만 길과 건물까지 선명해 옹기 좌판을 부드러운 주변 맥락으로 제한하라는 요구에는 덜 맞는다.",
        "entities": "이관회는 참조와 유사한 검은 머리, 청년 남성의 얼굴, 거친 베이지색 저고리를 갖췄다. 왼쪽 장교는 동아시아계 성인 남성으로 보이며 군모와 군복을 착용했고 총을 쥔 손이 같은 군복 소매로 연결된다. 다른 사람은 보이지 않는다. 총에는 낡은 금속과 목재가 있지만, 방아쇠 부근부터 턱까지의 총열이 참조보다 현저히 짧고 발화 장치·황동 띠·총검 구성도 일치하지 않는다. 옹기와 강한 낮 햇빛은 있으나 좌판 옆 연기는 확인되지 않는다. 옷깃 표장은 보이지만 판독 가능한 문구는 없다.",
        "hard_violations": [],
        "physics": "장교의 손이 방아쇠울 뒤 손잡이 부위를 감싸 총을 지지하며 손목과 팔뚝도 이어져 보인다. 그 손에서 턱밑 접촉점으로 힘을 가하는 위쪽 대각선이 성립한다. 이관회가 턱을 조금 든 자세도 압박에 대한 반응으로 가능하다. 두 사람의 하체는 프레임 밖이지만 보이는 상체에 부유나 지지 불가능의 징후는 없고, 옹기들은 선반과 진열대에 놓여 있다."
       }
      ],
      "all_candidates_fail": false
     },
     "normalized": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 3,
        "verdict_ko": "측면 클로즈업과 연기는 맞지만, 불필요한 배경 인물 두 명이 추가됐고 총구가 턱밑을 향해 위로 압박하는 대신 아래로 기울어져 결정적인 위협 동작이 어긋납니다."
       },
       {
        "label": "A",
        "score": 7,
        "verdict_ko": "장교가 잡은 총이 위쪽 대각선으로 이관회의 턱밑을 압박하는 핵심 관계는 맞지만, 배경이 지나치게 선명하고 소총의 형태·비례가 참조와 다릅니다."
       }
      ],
      "all_candidates_fail": false,
      "readings": [
       {
        "label": "B",
        "direction": "이관회는 화면 왼쪽의 총을 든 사람 쪽을 바라본다. 총은 왼쪽 손에서 오른쪽 턱으로 이어지지만 총구 쪽으로 내려가는 기울기다. 총열 윗면이 턱 아래에 닿고 총구 끝은 접촉점보다 오른쪽으로 돌출되어 보여, 총구가 위를 향해 턱밑 살에 바짝 눌리는 요구와 다르다.",
        "built_space": "오른쪽 뒤에 옹기를 놓은 목재 선반 한 구역, 아래쪽에 낮은 목재 진열대 한 줄, 왼쪽에 흙길과 시장 가설물이 보인다. 선반의 전체 단수는 얼굴과 흐림 때문에 확인하기 어렵다. 옹기는 진열대에 놓여 있고 가게 앞 인물의 위치도 공간과 연결된다. 연기가 진열대 옆에서 올라오며 배경은 비교적 부드럽다. 중복된 고정 설비나 반사는 보이지 않는다.",
        "entities": "이관회는 검은 가르마 머리의 한국인 청년 남성으로 보이며, 참조와 유사한 얼굴 및 거친 베이지색 저고리를 갖췄다. 왼쪽에는 성인 손과 손목, 갈색 소매가 보이지만 군복 식별 요소가 없어 헌병 장교의 손이라는 확인은 약하다. 소총의 낡은 목재, 긴 금속 총열, 황동 띠는 참조와 가깝지만 참조의 긴 총검은 보이지 않고 총구 아래 부속 형태도 다르다. 왼쪽 먼 배경에는 지시되지 않은 사람 두 명이 보인다. 읽을 수 있는 문구는 없다.",
        "hard_violations": [
         "화면 왼쪽 시장 통로에 샷 텍스트가 등장시키지 않은 배경 인물 두 명이 추가되어 있다."
        ],
        "physics": "왼쪽 손이 소총의 목재 앞부분을 아래에서 받치고 손가락으로 감싸므로 총의 지지는 확인된다. 이관회의 머리는 목과 상체에 자연스럽게 연결되어 있고 부유하는 신체나 물체는 없다. 다만 턱을 받치는 총열 옆면의 접촉에 가까워, 총구 끝으로 턱밑을 위로 밀어붙이는 힘의 방향은 구현되지 않았다."
       },
       {
        "label": "A",
        "direction": "장교와 이관회는 서로의 얼굴 쪽을 바라본다. 장교의 손에서 시작하는 총은 오른쪽 위로 올라가며 끝이 이관회의 턱밑에 밀착한다. 총구 끝의 일부는 턱에 가려지지만 압박 대상과 위쪽 조준 방향은 분명하다.",
        "built_space": "오른쪽에는 네 단가량이 드러나는 목재 옹기 선반 한 구역과 앞쪽의 낮은 진열대 한 줄이 있고, 위에는 초가 처마와 말아 올린 발이 보인다. 왼쪽으로 흙길과 목재 시장 가설물이 이어져 참조 장소의 재료와 배치에 부합한다. 이관회는 좌판 바로 앞, 장교는 통로 쪽에 서 있다. 고정 설비의 부당한 중복이나 불가능한 반사는 없다. 다만 길과 건물까지 선명해 옹기 좌판을 부드러운 주변 맥락으로 제한하라는 요구에는 덜 맞는다.",
        "entities": "이관회는 참조와 유사한 검은 머리, 청년 남성의 얼굴, 거친 베이지색 저고리를 갖췄다. 왼쪽 장교는 동아시아계 성인 남성으로 보이며 군모와 군복을 착용했고 총을 쥔 손이 같은 군복 소매로 연결된다. 다른 사람은 보이지 않는다. 총에는 낡은 금속과 목재가 있지만, 방아쇠 부근부터 턱까지의 총열이 참조보다 현저히 짧고 발화 장치·황동 띠·총검 구성도 일치하지 않는다. 옹기와 강한 낮 햇빛은 있으나 좌판 옆 연기는 확인되지 않는다. 옷깃 표장은 보이지만 판독 가능한 문구는 없다.",
        "hard_violations": [],
        "physics": "장교의 손이 방아쇠울 뒤 손잡이 부위를 감싸 총을 지지하며 손목과 팔뚝도 이어져 보인다. 그 손에서 턱밑 접촉점으로 힘을 가하는 위쪽 대각선이 성립한다. 이관회가 턱을 조금 든 자세도 압박에 대한 반응으로 가능하다. 두 사람의 하체는 프레임 밖이지만 보이는 상체에 부유나 지지 불가능의 징후는 없고, 옹기들은 선반과 진열대에 놓여 있다."
       }
      ]
     },
     "ok": true
    }
   ],
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "slot_winner_match": false,
   "slot_winner": {
    "gemini-pro": "B",
    "gpt-high": "A"
   },
   "route": "cross_slot_combined"
  },
  "dual": {
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "normalized": {
    "A": 1.333,
    "B": 1.429
   },
   "adjusted": {
    "A": 1.083,
    "B": 1.179
   },
   "violations": {
    "A": [
     "[gemini-pro] 허용되지 않은 인물(장교의 얼굴과 상체) 등장",
     "[gemini-pro] 물리적으로 불가능한 구조의 발명된 사물(기괴하게 섞인 총)"
    ],
    "B": [
     "[gpt-high] 화면 왼쪽 시장 통로에 샷 텍스트가 등장시키지 않은 배경 인물 두 명이 추가되어 있다."
    ]
   },
   "per_model_winner": {
    "gemini-pro": "B",
    "gpt-high": "A"
   },
   "agreed": false
  },
  "totals": {
   "B": 1179,
   "A": 1083
  },
  "selected": "B",
  "ranking": [
   "B",
   "A"
  ],
  "verdicts": [
   {
    "label": "B",
    "score": 1179,
    "verdict_ko": "비록 총구가 턱밑이 아닌 목을 향하고 총을 쥔 손의 소매가 평복이라는 단점이 있으나, 치명적인 물리적 오류나 규칙 위반 없이 클로즈업 구도를 안정적으로 구현했습니다.  ★위반: [gpt-high] 화면 왼쪽 시장 통로에 샷 텍스트가 등장시키지 않은 배경 인물 두 명이 추가되어 있다."
   },
   {
    "label": "A",
    "score": 1083,
    "verdict_ko": "총의 형태가 권총과 소총이 기괴하게 섞인 물리적으로 불가능한 구조이며, 지시문이 허용하지 않은 장교의 얼굴과 상체가 등장하여 심각한 규칙 위반입니다.  ★위반: [gemini-pro] 허용되지 않은 인물(장교의 얼굴과 상체) 등장 / [gemini-pro] 물리적으로 불가능한 구조의 발명된 사물(기괴하게 섞인 총)"
   }
  ],
  "refs": [
   {
    "label": "LOCATION PHOTOGRAPH — the exact place of this shot: its architecture, materials, fixed features and lighting mood are spatial truth; stage the moment inside this place. Never copy its camera framing.",
    "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/db00517b-165b-4005-a287-f28936b59b95/images/272bd7af-9294-4490-98c6-88a40e9deacc/scene/recipe/groupbg_market_pottery_front_836383.png",
    "asset_id": "1df6d064-1b12-4716-9367-28fcecd83b8a",
    "role": "bgfirst_group_bg"
   },
   {
    "label": "CHARACTER REFERENCE — 이관회: the exact person appearing in this shot; match face, hair and build exactly.",
    "path": "<bytes:933540>",
    "asset_id": "851b450f-ac84-47fe-ac5a-03a56f63fcae",
    "role": "character_ref"
   },
   {
    "label": "PROP REFERENCE — 근대식 소총: the exact object appearing in this shot; match its look, material and wear exactly.",
    "path": "<bytes:828568>",
    "asset_id": "7543806c-8087-4caa-8d7b-56d10864fd3c",
    "role": "prop_ref"
   }
  ],
  "critique_skipped": true,
  "shot_run_produced": false,
  "shot_run_uid": "06aa13fe-7afe-7fbf-8efc-875c4c419a12",
  "bgfirst": {
   "bg_path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/db00517b-165b-4005-a287-f28936b59b95/images/272bd7af-9294-4490-98c6-88a40e9deacc/scene/recipe/S1sh4__bgfirst_bg.png",
   "bg_asset_id": "23232690-4c79-45b6-887a-266366c69dc4",
   "bg_record_key": "S1sh4::bgfirst_bg",
   "chain_winner": false,
   "authority": "groupbg",
   "group_key": "market_pottery_front",
   "groupbg_asset_id": "1df6d064-1b12-4716-9367-28fcecd83b8a"
  },
  "ref_mode": "그룹 배경+엔티티 (2택1: 무콘티 승)",
  "share_plan": {
   "ref_plan": "background"
  }
 },
 "S1sh4::cine": {
  "applied": true,
  "attempted_at": "2026-09-09T00:31:14.107564+00:00",
  "fingerprint": "b70a7d1d3ac021469218504caac55607450fc1c33bc8d6171634e330699173e3",
  "fingerprint_version": 2,
  "provider": "mai",
  "endpoint": "openrouter/chat-completions",
  "model": "microsoft/mai-image-2.6",
  "pack": "24.202608252115",
  "source_file": "S1sh4_sel.png",
  "source_sha256": "23f9a7796346caf29f7c3fbae2133764bd76f6186d72bf0fc07a86127f1a1fdd",
  "file": "S1sh4_cine.png",
  "staged_sha256": "bb355564fe43b4ad47ff6227572d8fa4a6da66cb8dbd22393ab19a5f360577bc",
  "latency_ms": 29484
 },
 "S1sh10::signage": {
  "fp": "06dbd1dc72ac67a9",
  "inscriptions": [],
  "cues": [],
  "dropped": []
 },
 "groupbg::market_center": {
  "input_fingerprint": "2b9f4a8b8c895dd3",
  "meta": {
   "model": "gpt-image-2",
   "size": "1536x864",
   "pack": "11.202607220237",
   "contract": "bgfirst_full_v3",
   "group_sig": {
    "key": "market_center",
    "tags": [
     "S1sh10",
     "S1sh14",
     "S30sh21",
     "S30sh38",
     "S30sh50"
    ]
   },
   "context_sig": "1e4dc90d39b30b70"
  },
  "prompt": "Create ONE empty live-action location background photograph — NO PEOPLE, no figures, no body parts, no silhouettes, no shadows or reflections of people anywhere.\n\"Empty\" means no people only: KEEP the location's inherent occupants and stock that define the place — animals in an animal shelter, pen or farm, goods and produce in a market, moored boats in a harbour — unless the shot text explicitly removes them.\nThe FIRST attached image is a thin-line storyboard sketch of a shot that happens at this location: use it ONLY as spatial evidence — what this place contains, how its ground, structures and landmarks are arranged and proportioned. Ignore the sketched people and arrows entirely, and do NOT copy its line style: render a fully photographic, physically plausible real place that fits THE LOCATION text below.\nHUMAN-SCALE CALIBRATION: derive every structure's true size from human-scale elements — a door ≈ 2m, a window ≈ 1–1.5m wide, one storey ≈ 2.5–3m; never inflate a small structure or shrink a large one.\n\nTHE LOCATION — early 20th-century Korea, Jeolla region; all people are Korean unless explicitly stated otherwise: The open dirt space at the center of a sunlit market, approached from the surrounding stalls.\n\nLOCATION DETAIL (descriptive evidence about this place from the production's location records — evidence, not a staging order): use it to understand what kind of place this is and what it permanently contains. Carry over ONLY its enduring physical features — layout, structures, ground and materials, fixed equipment and installations. Do NOT treat any lighting, weather or time-of-day wording, momentary object states, depicted events or subjective impressions in it as fixed facts of the place; the TIME OF DAY lock below is the sole authority for time and lighting.\n복내장터 좌판 거리·장터 입구·골목: 다양한 좌판과 짐꾼, 상인들이 밀집한 넓은 흙바닥 노천 장터. 포목전, 짚신전, 옹기전, 닭장, 생선 좌판이 빽빽하게 들어서 있으며, 벽에 관의 방이 붙어 있고 상황에 따라 찢어진 멍석과 부서진 좌판이 널브러져 있다. (특징: 넓은 흙바닥; 나무와 짚으로 얽어 만든 노점 좌판 (옹기, 짚신, 채소, 포목 등); 나무 닭장과 광주리; 방(벽보)이 붙어 있는 벽; 깨진 항아리와 뒤집힌 멍석, 부서진 나무판자들)\n\nSCENE EVIDENCE (verbatim quotes from the screenplay about this place — treat them as evidence of what the location physically contains and looks like; stage the PLACE those moments happen in, but do NOT depict the momentary actions, people or staged props themselves):\n- 무거운 지게가 흙바닥에 거칠게 내려앉는다. 헌병들의 시선이 장터 한복판으로 쏠린다.\n- 규홍이 장터 한복판으로 성큼성큼 걸어 나온다.\n- 규홍의 주먹이 헌병의 턱에 꽂힌다. 헌병이 옹기 좌판으로 나가떨어진다. 와장창!\n\nTIME OF DAY (lock): day, harsh sunlight.\n\nRender ONE photorealistic empty location photograph, 16:9, neutral enough that every shot of this place can be staged from it later. No readable writing anywhere: surfaces that would carry writing may be present, but stage any wording out of legibility — an oblique angle, distance, shallow focus. No captions, watermarks or overlay text.",
  "effective_prompt": "Create ONE empty live-action location background photograph — NO PEOPLE, no figures, no body parts, no silhouettes, no shadows or reflections of people anywhere.\n\"Empty\" means no people only: KEEP the location's inherent occupants and stock that define the place — animals in an animal shelter, pen or farm, goods and produce in a market, moored boats in a harbour — unless the shot text explicitly removes them.\nThe FIRST attached image is a thin-line storyboard sketch of a shot that happens at this location: use it ONLY as spatial evidence — what this place contains, how its ground, structures and landmarks are arranged and proportioned. Ignore the sketched people and arrows entirely, and do NOT copy its line style: render a fully photographic, physically plausible real place that fits THE LOCATION text below.\nHUMAN-SCALE CALIBRATION: derive every structure's true size from human-scale elements — a door ≈ 2m, a window ≈ 1–1.5m wide, one storey ≈ 2.5–3m; never inflate a small structure or shrink a large one.\n\nTHE LOCATION — early 20th-century Korea, Jeolla region; all people are Korean unless explicitly stated otherwise: The open dirt space at the center of a sunlit market, approached from the surrounding stalls.\n\nLOCATION DETAIL (descriptive evidence about this place from the production's location records — evidence, not a staging order): use it to understand what kind of place this is and what it permanently contains. Carry over ONLY its enduring physical features — layout, structures, ground and materials, fixed equipment and installations. Do NOT treat any lighting, weather or time-of-day wording, momentary object states, depicted events or subjective impressions in it as fixed facts of the place; the TIME OF DAY lock below is the sole authority for time and lighting.\n복내장터 좌판 거리·장터 입구·골목: 다양한 좌판과 짐꾼, 상인들이 밀집한 넓은 흙바닥 노천 장터. 포목전, 짚신전, 옹기전, 닭장, 생선 좌판이 빽빽하게 들어서 있으며, 벽에 관의 방이 붙어 있고 상황에 따라 찢어진 멍석과 부서진 좌판이 널브러져 있다. (특징: 넓은 흙바닥; 나무와 짚으로 얽어 만든 노점 좌판 (옹기, 짚신, 채소, 포목 등); 나무 닭장과 광주리; 방(벽보)이 붙어 있는 벽; 깨진 항아리와 뒤집힌 멍석, 부서진 나무판자들)\n\nSCENE EVIDENCE (verbatim quotes from the screenplay about this place — treat them as evidence of what the location physically contains and looks like; stage the PLACE those moments happen in, but do NOT depict the momentary actions, people or staged props themselves):\n- 무거운 지게가 흙바닥에 거칠게 내려앉는다. 헌병들의 시선이 장터 한복판으로 쏠린다.\n- 규홍이 장터 한복판으로 성큼성큼 걸어 나온다.\n- 규홍의 주먹이 헌병의 턱에 꽂힌다. 헌병이 옹기 좌판으로 나가떨어진다. 와장창!\n\nTIME OF DAY (lock): day, harsh sunlight.\n\nRender ONE photorealistic empty location photograph, 16:9, neutral enough that every shot of this place can be staged from it later. No readable writing anywhere: surfaces that would carry writing may be present, but stage any wording out of legibility — an oblique angle, distance, shallow focus. No captions, watermarks or overlay text.",
  "bg_path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/db00517b-165b-4005-a287-f28936b59b95/images/272bd7af-9294-4490-98c6-88a40e9deacc/scene/recipe/groupbg_market_center_fa2dce.png",
  "asset_id": "39a96dca-cf22-4c62-bacb-511754a80868",
  "input_asset_ids": [
   "5f12e311-afbc-4a66-84b1-e7a0a230bef2"
  ],
  "origin_tag": "S1sh10",
  "place_text": "The open dirt space at the center of a sunlit market, approached from the surrounding stalls.",
  "origin_inputs": {
   "place_text": "The open dirt space at the center of a sunlit market, approached from the surrounding stalls.",
   "time_of_day_en": "day, harsh sunlight",
   "conti_asset_id": "5f12e311-afbc-4a66-84b1-e7a0a230bef2"
  }
 },
 "S1sh10::bgfirst_bg": {
  "input_fingerprint": "0cbbc6914a434112",
  "prompt": "Create the EMPTY BACKGROUND PLATE for one film shot — NO PEOPLE, no figures, no body parts, no sketch lines, no arrows anywhere.\n\"Empty\" means no people only: KEEP the location's inherent occupants and stock that define the place — animals in an animal shelter, pen or farm, goods in a market, moored boats in a harbour — unless the shot text explicitly removes them.\nThe FIRST attached image is a thin-line storyboard sketch: use ONLY its camera angle, horizon, perspective and the placement/size of buildings and set masses — ignore the sketched people and arrows entirely. The SECOND attached image (LOCATION PHOTOGRAPH) is the real place: take its architecture, materials, signage and fixed features, and RE-PROJECT them into the sketch's camera. If the photograph's camera differs from the sketch's, the sketch's camera wins.\nHUMAN-SCALE CALIBRATION: derive every structure's true size from human-scale elements — a door ≈ 2m, a window ≈ 1–1.5m wide, one storey ≈ 2.5–3m; never inflate a small structure or shrink a large one.\n\nSHOT TEXT this background must serve (Korean): 양팔을 자연스럽게 내린 빈손의 안규홍이 헌병들을 향해 성큼 걸음을 내디뎌 한쪽 발이 공중에 떠 있는 전신.\n\nLOCATION (lock): The open dirt space at the center of a sunlit market, approached from the surrounding stalls.\n\nTIME OF DAY (lock): day, harsh sunlight.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: wide shot\n- FRAME LAYOUT: police-side end of the market aisle in the lower-left of the frame, foreground; clear ground beneath the approaching man's lifted foot in the lower-right of the frame, midground.\n- KEY BACKGROUND ELEMENTS: 장터 통로의 흙바닥 (Visible beneath 안규홍's approaching feet); used as An uninterrupted ground strip makes the lifted foot and advancing stride unmistakable; 옹기 좌판 (Bordering the confrontation area); used as Peripheral stall shapes maintain the aisle's depth and realistic human scale.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Strong daytime sunlight and muted earth tones preserve the market's hard-edged realism without turning his arrival into a heroic lighting effect.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nRender ONE photorealistic empty location photograph, 16:9, that this shot can be staged inside later. No readable writing anywhere: surfaces that would carry writing may be present, but stage any wording out of legibility — an oblique angle, distance, shallow focus. No captions, watermarks or overlay text.",
  "effective_prompt": "Create the EMPTY BACKGROUND PLATE for one film shot — NO PEOPLE, no figures, no body parts, no sketch lines, no arrows anywhere.\n\"Empty\" means no people only: KEEP the location's inherent occupants and stock that define the place — animals in an animal shelter, pen or farm, goods in a market, moored boats in a harbour — unless the shot text explicitly removes them.\nThe FIRST attached image is a thin-line storyboard sketch: use ONLY its camera angle, horizon, perspective and the placement/size of buildings and set masses — ignore the sketched people and arrows entirely. The SECOND attached image (LOCATION PHOTOGRAPH) is the real place: take its architecture, materials, signage and fixed features, and RE-PROJECT them into the sketch's camera. If the photograph's camera differs from the sketch's, the sketch's camera wins.\nHUMAN-SCALE CALIBRATION: derive every structure's true size from human-scale elements — a door ≈ 2m, a window ≈ 1–1.5m wide, one storey ≈ 2.5–3m; never inflate a small structure or shrink a large one.\n\nSHOT TEXT this background must serve (Korean): 양팔을 자연스럽게 내린 빈손의 안규홍이 헌병들을 향해 성큼 걸음을 내디뎌 한쪽 발이 공중에 떠 있는 전신.\n\nLOCATION (lock): The open dirt space at the center of a sunlit market, approached from the surrounding stalls.\n\nTIME OF DAY (lock): day, harsh sunlight.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: wide shot\n- FRAME LAYOUT: police-side end of the market aisle in the lower-left of the frame, foreground; clear ground beneath the approaching man's lifted foot in the lower-right of the frame, midground.\n- KEY BACKGROUND ELEMENTS: 장터 통로의 흙바닥 (Visible beneath 안규홍's approaching feet); used as An uninterrupted ground strip makes the lifted foot and advancing stride unmistakable; 옹기 좌판 (Bordering the confrontation area); used as Peripheral stall shapes maintain the aisle's depth and realistic human scale.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Strong daytime sunlight and muted earth tones preserve the market's hard-edged realism without turning his arrival into a heroic lighting effect.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nRender ONE photorealistic empty location photograph, 16:9, that this shot can be staged inside later. No readable writing anywhere: surfaces that would carry writing may be present, but stage any wording out of legibility — an oblique angle, distance, shallow focus. No captions, watermarks or overlay text.",
  "bg_path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/db00517b-165b-4005-a287-f28936b59b95/images/272bd7af-9294-4490-98c6-88a40e9deacc/scene/recipe/S1sh10__bgfirst_bg.png",
  "asset_id": "e9684387-9242-4865-88ee-81c863d1434f",
  "input_asset_ids": [
   "5f12e311-afbc-4a66-84b1-e7a0a230bef2",
   "39a96dca-cf22-4c62-bacb-511754a80868"
  ]
 },
 "S1sh10": {
  "input_fingerprint": "1c17df5335453d8a",
  "prompt": "Create ONE FINAL photorealistic live-action film still of the moment below — early 20th-century Korea, Jeolla region; all people are Korean unless explicitly stated otherwise. TIME OF DAY (lock): day, harsh sunlight.\n\nSHOT TEXT (authoritative, Korean): 양팔을 자연스럽게 내린 빈손의 안규홍이 헌병들을 향해 성큼 걸음을 내디뎌 한쪽 발이 공중에 떠 있는 전신.\n\nLOCATION (lock): The open dirt space at the center of a sunlit market, approached from the surrounding stalls. The shot takes place here — the attached LOCATION PHOTOGRAPH shows the exact spot.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: wide shot\n- FRAME LAYOUT: police-side end of the market aisle in the lower-left of the frame, foreground; clear ground beneath the approaching man's lifted foot in the lower-right of the frame, midground.\n- KEY BACKGROUND ELEMENTS: 장터 통로의 흙바닥 (Visible beneath 안규홍's approaching feet); used as An uninterrupted ground strip makes the lifted foot and advancing stride unmistakable; 옹기 좌판 (Bordering the confrontation area); used as Peripheral stall shapes maintain the aisle's depth and realistic human scale.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Strong daytime sunlight and muted earth tones preserve the market's hard-edged realism without turning his arrival into a heroic lighting effect.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The heavy carrying frame now rests on the dirt in the market center; harsh sunlight and stall smoke persist. 안규홍: He advances empty-handed, wearing a dirt-stained jeogori.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 안규홍 (한국인 남성, 청년 나이대, 검은 머리, 단단히 묶은 머리) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.",
  "roll_prompts": {
   "A": "Stage the shot. The FIRST attached image (SHOT BACKGROUND) is the finished empty background of this shot — keep it EXACTLY: its camera, perspective, architecture, lighting and every fixed feature stay untouched. The SECOND attached image (LAYOUT SKETCH) tells you ONLY where the people go: each sketched person's position, screen size, pose and the gaze/motion arrows. Ignore the sketch's background lines. The CHARACTER REFERENCE photographs show the real people.\nPlace the real people into the background at exactly the sketched positions, sizes and poses, following the arrow directions. No sketch lines or arrows may remain.\n\nCreate ONE FINAL photorealistic live-action film still of the moment below — early 20th-century Korea, Jeolla region; all people are Korean unless explicitly stated otherwise. TIME OF DAY (lock): day, harsh sunlight.\n\nSHOT TEXT (authoritative, Korean): 양팔을 자연스럽게 내린 빈손의 안규홍이 헌병들을 향해 성큼 걸음을 내디뎌 한쪽 발이 공중에 떠 있는 전신.\n\nLOCATION (lock): The open dirt space at the center of a sunlit market, approached from the surrounding stalls. The shot takes place here — the FIRST attached image (SHOT BACKGROUND) is this exact place, already built: its ground, structures, horizon, materials and lighting are the finished truth of this location and must not be redesigned or replaced. No location photograph is attached — read the place from that image alone, and add no scenery, structure, vehicle or fixture that it does not already show. This lock governs the place only; the figures in the shot follow the staging and pose instructions.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: wide shot\n- FRAME LAYOUT: police-side end of the market aisle in the lower-left of the frame, foreground; clear ground beneath the approaching man's lifted foot in the lower-right of the frame, midground.\n- KEY BACKGROUND ELEMENTS: 장터 통로의 흙바닥 (Visible beneath 안규홍's approaching feet); used as An uninterrupted ground strip makes the lifted foot and advancing stride unmistakable; 옹기 좌판 (Bordering the confrontation area); used as Peripheral stall shapes maintain the aisle's depth and realistic human scale.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Strong daytime sunlight and muted earth tones preserve the market's hard-edged realism without turning his arrival into a heroic lighting effect.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The heavy carrying frame now rests on the dirt in the market center; harsh sunlight and stall smoke persist. 안규홍: He advances empty-handed, wearing a dirt-stained jeogori.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 안규홍 (한국인 남성, 청년 나이대, 검은 머리, 단단히 묶은 머리) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.",
   "B": "Create ONE FINAL photorealistic live-action film still of the moment below — early 20th-century Korea, Jeolla region; all people are Korean unless explicitly stated otherwise. TIME OF DAY (lock): day, harsh sunlight.\n\nSHOT TEXT (authoritative, Korean): 양팔을 자연스럽게 내린 빈손의 안규홍이 헌병들을 향해 성큼 걸음을 내디뎌 한쪽 발이 공중에 떠 있는 전신.\n\nLOCATION (lock): The open dirt space at the center of a sunlit market, approached from the surrounding stalls. The shot takes place here — the attached LOCATION PHOTOGRAPH shows the exact spot.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: wide shot\n- FRAME LAYOUT: police-side end of the market aisle in the lower-left of the frame, foreground; clear ground beneath the approaching man's lifted foot in the lower-right of the frame, midground.\n- KEY BACKGROUND ELEMENTS: 장터 통로의 흙바닥 (Visible beneath 안규홍's approaching feet); used as An uninterrupted ground strip makes the lifted foot and advancing stride unmistakable; 옹기 좌판 (Bordering the confrontation area); used as Peripheral stall shapes maintain the aisle's depth and realistic human scale.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Strong daytime sunlight and muted earth tones preserve the market's hard-edged realism without turning his arrival into a heroic lighting effect.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The heavy carrying frame now rests on the dirt in the market center; harsh sunlight and stall smoke persist. 안규홍: He advances empty-handed, wearing a dirt-stained jeogori.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 안규홍 (한국인 남성, 청년 나이대, 검은 머리, 단단히 묶은 머리) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.\n\nCOMPOSITION VARIATION (this candidate only): keep the place, the people and the moment exactly as specified, and keep from the CAMERA & FRAME contract its subject scale, screen placement and key background elements exactly. For this candidate only, its camera angle and height are deliberately left open: choose a camera position distinctly different from the obvious one, as a film director picking a second setup on the same blocking. Stay true to any explicit framing the SHOT TEXT itself dictates."
  },
  "roll_refs": {
   "A": [
    {
     "label": "SHOT BACKGROUND",
     "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/db00517b-165b-4005-a287-f28936b59b95/images/272bd7af-9294-4490-98c6-88a40e9deacc/scene/recipe/S1sh10__bgfirst_bg.png",
     "asset_id": "e9684387-9242-4865-88ee-81c863d1434f",
     "role": "bgfirst_bg"
    },
    {
     "label": "LAYOUT SKETCH (people placement only)",
     "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/db00517b-165b-4005-a287-f28936b59b95/images/272bd7af-9294-4490-98c6-88a40e9deacc/conti/conti_S1sh10.png",
     "asset_id": "5f12e311-afbc-4a66-84b1-e7a0a230bef2",
     "role": "conti_light"
    },
    {
     "label": "CHARACTER REFERENCE — 안규홍: the exact person appearing in this shot; match face, hair and build exactly.",
     "path": "<bytes:1028384>",
     "asset_id": "3b5894db-ce55-4f34-b966-5e725c3aa952",
     "role": "character_ref"
    }
   ],
   "B": [
    {
     "label": "LOCATION PHOTOGRAPH — the exact place of this shot: its architecture, materials, fixed features and lighting mood are spatial truth; stage the moment inside this place. Never copy its camera framing.",
     "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/db00517b-165b-4005-a287-f28936b59b95/images/272bd7af-9294-4490-98c6-88a40e9deacc/scene/recipe/groupbg_market_center_fa2dce.png",
     "asset_id": "39a96dca-cf22-4c62-bacb-511754a80868",
     "role": "bgfirst_group_bg"
    },
    {
     "label": "CHARACTER REFERENCE — 안규홍: the exact person appearing in this shot; match face, hair and build exactly.",
     "path": "<bytes:1028384>",
     "asset_id": "3b5894db-ce55-4f34-b966-5e725c3aa952",
     "role": "character_ref"
    }
   ]
  },
  "shot_run_spend_attempt_count": 1,
  "judge_flip_skipped": "judge_fn_owns_order",
  "readings": [
   {
    "label": "A",
    "direction": "안규홍이 왼쪽 전경의 헌병들을 향해 시선을 고정하고 성큼 걸어가고 있음.",
    "built_space": "흙바닥으로 된 장터 중앙 통로이며, 양옆으로 옹기 좌판이 레퍼런스 구도에 맞게 배치되어 공간의 깊이감을 줌.",
    "entities": "안규홍과 헌병 2명의 외형 및 복장이 묘사된 대로 잘 나타나나, 배경 멀리에 텍스트에 명시되지 않은 작은 행인들이 존재함.",
    "hard_violations": [
     "[gemini-pro] 샷 텍스트에 명시되지 않은 배경 인물(행인) 추가됨",
     "[gpt-high] 숏에 지정되지 않은 행인 세 명을 뒤쪽 통로와 양쪽 좌판에 추가했다."
    ],
    "physics": "안규홍이 걷는 과정에서 오른발이 지면에서 명확히 떨어져 공중에 떠 있으며, 빈손인 양팔은 자연스럽게 내려진 상태로 무게 중심 이동이 자연스러움."
   },
   {
    "label": "B",
    "direction": "안규홍이 왼쪽 전경의 헌병들을 향해 걸어가고 있음.",
    "built_space": "장터 통로가 구현되었으며, 통로 중앙 바닥에 짐을 나르는 지게가 놓여 있어 '바닥에 놓인 짐틀' 설정을 반영함.",
    "entities": "안규홍과 헌병들이 등장하나, 헌병 바로 뒤와 우측에 텍스트에 없는 인물(마을 주민)들이 뚜렷하게 추가됨.",
    "hard_violations": [
     "[gemini-pro] 샷 텍스트에 명시되지 않은 주변 인물(마을 주민) 추가됨",
     "[gpt-high] 숏에 지정되지 않은 행인 네 명을 좌판 주변과 오른쪽 가장자리에 추가했다."
    ],
    "physics": "안규홍의 양발이 모두 지면에 닿아 있어 텍스트가 강제한 '한쪽 발이 공중에 떠 있는' 걷는 찰나의 동작이 구현되지 않음."
   }
  ],
  "cross_model_order": {
   "policy": "cross_model_order_v2_slot0_fwd_slot1_rev_parallel2",
   "slots": [
    {
     "model": "gemini-pro",
     "order": "forward",
     "display_to_canonical": {
      "A": "A",
      "B": "B"
     },
     "raw": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 8,
        "verdict_ko": "샷 텍스트와 프레이밍이 요구한 '공중에 떠 있는 한쪽 발'과 그 아래의 빈 지면을 완벽히 구현했으나, 배경에 명시되지 않은 인물이 작게 포함되어 감점되었습니다."
       },
       {
        "label": "B",
        "score": 4,
        "verdict_ko": "바닥에 내려놓은 짐틀의 묘사는 좋으나, 가장 중요한 '공중에 떠 있는 발' 동작이 전혀 구현되지 않았고 텍스트에 없는 인물이 뚜렷하게 등장해 지시를 위반했습니다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "안규홍이 왼쪽 전경의 헌병들을 향해 시선을 고정하고 성큼 걸어가고 있음.",
        "built_space": "흙바닥으로 된 장터 중앙 통로이며, 양옆으로 옹기 좌판이 레퍼런스 구도에 맞게 배치되어 공간의 깊이감을 줌.",
        "entities": "안규홍과 헌병 2명의 외형 및 복장이 묘사된 대로 잘 나타나나, 배경 멀리에 텍스트에 명시되지 않은 작은 행인들이 존재함.",
        "hard_violations": [
         "샷 텍스트에 명시되지 않은 배경 인물(행인) 추가됨"
        ],
        "physics": "안규홍이 걷는 과정에서 오른발이 지면에서 명확히 떨어져 공중에 떠 있으며, 빈손인 양팔은 자연스럽게 내려진 상태로 무게 중심 이동이 자연스러움."
       },
       {
        "label": "B",
        "direction": "안규홍이 왼쪽 전경의 헌병들을 향해 걸어가고 있음.",
        "built_space": "장터 통로가 구현되었으며, 통로 중앙 바닥에 짐을 나르는 지게가 놓여 있어 '바닥에 놓인 짐틀' 설정을 반영함.",
        "entities": "안규홍과 헌병들이 등장하나, 헌병 바로 뒤와 우측에 텍스트에 없는 인물(마을 주민)들이 뚜렷하게 추가됨.",
        "hard_violations": [
         "샷 텍스트에 명시되지 않은 주변 인물(마을 주민) 추가됨"
        ],
        "physics": "안규홍의 양발이 모두 지면에 닿아 있어 텍스트가 강제한 '한쪽 발이 공중에 떠 있는' 걷는 찰나의 동작이 구현되지 않음."
       }
      ],
      "all_candidates_fail": false
     },
     "normalized": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 8,
        "verdict_ko": "샷 텍스트와 프레이밍이 요구한 '공중에 떠 있는 한쪽 발'과 그 아래의 빈 지면을 완벽히 구현했으나, 배경에 명시되지 않은 인물이 작게 포함되어 감점되었습니다."
       },
       {
        "label": "B",
        "score": 4,
        "verdict_ko": "바닥에 내려놓은 짐틀의 묘사는 좋으나, 가장 중요한 '공중에 떠 있는 발' 동작이 전혀 구현되지 않았고 텍스트에 없는 인물이 뚜렷하게 등장해 지시를 위반했습니다."
       }
      ],
      "all_candidates_fail": false,
      "readings": [
       {
        "label": "A",
        "direction": "안규홍이 왼쪽 전경의 헌병들을 향해 시선을 고정하고 성큼 걸어가고 있음.",
        "built_space": "흙바닥으로 된 장터 중앙 통로이며, 양옆으로 옹기 좌판이 레퍼런스 구도에 맞게 배치되어 공간의 깊이감을 줌.",
        "entities": "안규홍과 헌병 2명의 외형 및 복장이 묘사된 대로 잘 나타나나, 배경 멀리에 텍스트에 명시되지 않은 작은 행인들이 존재함.",
        "hard_violations": [
         "샷 텍스트에 명시되지 않은 배경 인물(행인) 추가됨"
        ],
        "physics": "안규홍이 걷는 과정에서 오른발이 지면에서 명확히 떨어져 공중에 떠 있으며, 빈손인 양팔은 자연스럽게 내려진 상태로 무게 중심 이동이 자연스러움."
       },
       {
        "label": "B",
        "direction": "안규홍이 왼쪽 전경의 헌병들을 향해 걸어가고 있음.",
        "built_space": "장터 통로가 구현되었으며, 통로 중앙 바닥에 짐을 나르는 지게가 놓여 있어 '바닥에 놓인 짐틀' 설정을 반영함.",
        "entities": "안규홍과 헌병들이 등장하나, 헌병 바로 뒤와 우측에 텍스트에 없는 인물(마을 주민)들이 뚜렷하게 추가됨.",
        "hard_violations": [
         "샷 텍스트에 명시되지 않은 주변 인물(마을 주민) 추가됨"
        ],
        "physics": "안규홍의 양발이 모두 지면에 닿아 있어 텍스트가 강제한 '한쪽 발이 공중에 떠 있는' 걷는 찰나의 동작이 구현되지 않음."
       }
      ]
     },
     "ok": true
    },
    {
     "model": "gpt-high",
     "order": "reverse",
     "display_to_canonical": {
      "A": "B",
      "B": "A"
     },
     "raw": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 3,
        "verdict_ko": "헌병을 향한 빈손 전진과 지면에 놓인 운반틀은 보이지만, 불필요한 행인들이 추가되었고 전신 크기와 배경 압축이 요구한 와이드 구도에서 더 벗어난다."
       },
       {
        "label": "B",
        "score": 4,
        "verdict_ko": "왼쪽 전경의 헌병과 오른쪽 중경의 전신·들린 발, 장터 공간과 의상이 더 충실하지만, 추가 행인들과 보이지 않는 운반틀 때문에 완전한 합격은 아니다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "안규홍의 얼굴과 시선은 화면 왼쪽 전경의 헌병들을 향하며, 몸과 내딛는 발도 그들 쪽으로 접근한다. 헌병 두 명은 안규홍을 바라본다. 소총은 몸 옆에서 세워 휴대하고 있어 안규홍에게 겨누는 자세는 아니다.",
        "built_space": "왼쪽에 흙벽·기와와 차양 아래 옹기 좌판들이 있고, 오른쪽에도 차양과 좌판이 있다. 뒤쪽 중앙에는 문 하나와 그 앞으로 올라가는 계단이 보인다. 중앙 흙길에 운반틀 하나가 놓여 있고 안규홍의 발밑까지 길이 이어진다. 장소의 주요 재료와 배치는 닮았지만, 배경 문과 계단이 참조보다 크게 압축되어 보이고 안규홍도 화면 높이를 크게 차지한다. 헌병 두 명은 왼쪽 전경, 안규홍은 오른쪽 중경에 있다.",
        "entities": "안규홍은 참조와 유사한 얼굴·체격의 젊은 동아시아계 남성으로, 검은 머리를 뒤로 정리했고 때 묻은 밝은 저고리와 바지, 짚신을 착용했다. 양손은 비어 있고 팔은 내려와 있다. 다만 참조의 허리띠가 없고 옷매무새가 다르다. 헌병 두 명 외에 왼쪽 좌판의 여성, 오른쪽 좌판의 남성, 오른쪽 가장자리의 부분 인물 두 명이 추가되어 있다. 운반틀은 바닥에 있지만 상자와 긴 가로대가 강조되어 참조 없는 지게의 정확한 형태는 확인하기 어렵다. 강한 낮빛은 맞으나 좌판 연기는 뚜렷하지 않고 읽을 수 있는 글자는 없다.",
        "hard_violations": [
         "숏에 지정되지 않은 행인 네 명을 좌판 주변과 오른쪽 가장자리에 추가했다."
        ],
        "physics": "안규홍의 화면 오른쪽 발은 흙바닥을 딛고 체중을 받으며, 앞으로 나온 화면 왼쪽 발은 밑창 아래에 작은 틈과 분리된 그림자가 보여 들린 상태로 읽힌다. 보폭은 다소 작지만 실제 보행으로 가능한 자세다. 운반틀은 지면에, 옹기는 좌판과 지면에 놓여 있다. 소총은 멜빵과 손으로 지지되며, 근거 없이 떠 있는 몸이나 물체는 보이지 않는다."
       },
       {
        "label": "B",
        "direction": "안규홍의 몸은 왼쪽 전경의 헌병들이 있는 통로 입구 쪽으로 전진한다. 얼굴은 거의 정면이고 눈길은 약간 화면 왼쪽이어서 헌병 쪽과 대체로 맞지만, A만큼 대상을 분명하게 응시하지는 않는다. 헌병 두 명은 안규홍 쪽으로 고개를 돌렸다. 전경의 소총은 위를 향해 몸 옆에 세워져 있으며 사람에게 겨눠지지 않는다.",
        "built_space": "왼쪽 흙벽·기와·멍석 차양과 옹기 좌판, 오른쪽 천막·바구니·옹기 좌판, 오른쪽 전경의 닭장 하나, 먼 중앙의 문 하나와 계단이 보인다. 참조의 장터 깊이와 양옆 좌판의 상대적 크기가 A보다 가깝다. 헌병 두 명이 왼쪽 전경을 차지하고 안규홍의 전신은 오른쪽 중경에 있으며, 들린 발 아래와 앞쪽에 끊기지 않는 흙바닥이 확보되어 있다. 다만 넓게 드러난 중앙 지면에서 내려놓은 운반틀은 보이지 않는다.",
        "entities": "안규홍은 참조와 유사한 젊은 동아시아계 남성의 얼굴과 체격, 뒤로 정리한 검은 머리, 오염된 밝은 저고리·바지·허리띠·짚신을 갖췄다. 두 손은 비었고 양팔도 자연스럽게 내려와 있다. 헌병 두 명 외에 왼쪽 좌판, 뒤쪽 통로 중앙, 오른쪽 좌판에 행인 세 명이 추가되어 있다. 옹기와 닭장 등 장소 소품은 참조에 부합한다. 강한 햇빛과 흙빛 색조는 맞지만 운반틀과 뚜렷한 좌판 연기는 확인되지 않는다. 읽을 수 있는 글자는 없다.",
        "hard_violations": [
         "숏에 지정되지 않은 행인 세 명을 뒤쪽 통로와 양쪽 좌판에 추가했다."
        ],
        "physics": "안규홍의 화면 왼쪽 발은 지면을 딛고 몸을 지지하고, 화면 오른쪽 다리는 무릎을 굽힌 채 발 전체가 땅에서 떨어져 있다. 발 아래 빈 공간이 명확하며 보행 중 발을 앞으로 회수하는 순간으로 물리적으로 가능하다. 다만 길게 성큼 내딛는 순간보다는 보행 중간 단계에 가깝다. 닭장과 그릇은 목재 좌판에, 바구니와 항아리는 좌판 또는 지면에 놓여 있고 소총은 몸과 휴대 장비로 지지된다. 지지 없이 떠 있는 물체는 없다."
       }
      ],
      "all_candidates_fail": true
     },
     "normalized": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 3,
        "verdict_ko": "헌병을 향한 빈손 전진과 지면에 놓인 운반틀은 보이지만, 불필요한 행인들이 추가되었고 전신 크기와 배경 압축이 요구한 와이드 구도에서 더 벗어난다."
       },
       {
        "label": "A",
        "score": 4,
        "verdict_ko": "왼쪽 전경의 헌병과 오른쪽 중경의 전신·들린 발, 장터 공간과 의상이 더 충실하지만, 추가 행인들과 보이지 않는 운반틀 때문에 완전한 합격은 아니다."
       }
      ],
      "all_candidates_fail": true,
      "readings": [
       {
        "label": "B",
        "direction": "안규홍의 얼굴과 시선은 화면 왼쪽 전경의 헌병들을 향하며, 몸과 내딛는 발도 그들 쪽으로 접근한다. 헌병 두 명은 안규홍을 바라본다. 소총은 몸 옆에서 세워 휴대하고 있어 안규홍에게 겨누는 자세는 아니다.",
        "built_space": "왼쪽에 흙벽·기와와 차양 아래 옹기 좌판들이 있고, 오른쪽에도 차양과 좌판이 있다. 뒤쪽 중앙에는 문 하나와 그 앞으로 올라가는 계단이 보인다. 중앙 흙길에 운반틀 하나가 놓여 있고 안규홍의 발밑까지 길이 이어진다. 장소의 주요 재료와 배치는 닮았지만, 배경 문과 계단이 참조보다 크게 압축되어 보이고 안규홍도 화면 높이를 크게 차지한다. 헌병 두 명은 왼쪽 전경, 안규홍은 오른쪽 중경에 있다.",
        "entities": "안규홍은 참조와 유사한 얼굴·체격의 젊은 동아시아계 남성으로, 검은 머리를 뒤로 정리했고 때 묻은 밝은 저고리와 바지, 짚신을 착용했다. 양손은 비어 있고 팔은 내려와 있다. 다만 참조의 허리띠가 없고 옷매무새가 다르다. 헌병 두 명 외에 왼쪽 좌판의 여성, 오른쪽 좌판의 남성, 오른쪽 가장자리의 부분 인물 두 명이 추가되어 있다. 운반틀은 바닥에 있지만 상자와 긴 가로대가 강조되어 참조 없는 지게의 정확한 형태는 확인하기 어렵다. 강한 낮빛은 맞으나 좌판 연기는 뚜렷하지 않고 읽을 수 있는 글자는 없다.",
        "hard_violations": [
         "숏에 지정되지 않은 행인 네 명을 좌판 주변과 오른쪽 가장자리에 추가했다."
        ],
        "physics": "안규홍의 화면 오른쪽 발은 흙바닥을 딛고 체중을 받으며, 앞으로 나온 화면 왼쪽 발은 밑창 아래에 작은 틈과 분리된 그림자가 보여 들린 상태로 읽힌다. 보폭은 다소 작지만 실제 보행으로 가능한 자세다. 운반틀은 지면에, 옹기는 좌판과 지면에 놓여 있다. 소총은 멜빵과 손으로 지지되며, 근거 없이 떠 있는 몸이나 물체는 보이지 않는다."
       },
       {
        "label": "A",
        "direction": "안규홍의 몸은 왼쪽 전경의 헌병들이 있는 통로 입구 쪽으로 전진한다. 얼굴은 거의 정면이고 눈길은 약간 화면 왼쪽이어서 헌병 쪽과 대체로 맞지만, A만큼 대상을 분명하게 응시하지는 않는다. 헌병 두 명은 안규홍 쪽으로 고개를 돌렸다. 전경의 소총은 위를 향해 몸 옆에 세워져 있으며 사람에게 겨눠지지 않는다.",
        "built_space": "왼쪽 흙벽·기와·멍석 차양과 옹기 좌판, 오른쪽 천막·바구니·옹기 좌판, 오른쪽 전경의 닭장 하나, 먼 중앙의 문 하나와 계단이 보인다. 참조의 장터 깊이와 양옆 좌판의 상대적 크기가 A보다 가깝다. 헌병 두 명이 왼쪽 전경을 차지하고 안규홍의 전신은 오른쪽 중경에 있으며, 들린 발 아래와 앞쪽에 끊기지 않는 흙바닥이 확보되어 있다. 다만 넓게 드러난 중앙 지면에서 내려놓은 운반틀은 보이지 않는다.",
        "entities": "안규홍은 참조와 유사한 젊은 동아시아계 남성의 얼굴과 체격, 뒤로 정리한 검은 머리, 오염된 밝은 저고리·바지·허리띠·짚신을 갖췄다. 두 손은 비었고 양팔도 자연스럽게 내려와 있다. 헌병 두 명 외에 왼쪽 좌판, 뒤쪽 통로 중앙, 오른쪽 좌판에 행인 세 명이 추가되어 있다. 옹기와 닭장 등 장소 소품은 참조에 부합한다. 강한 햇빛과 흙빛 색조는 맞지만 운반틀과 뚜렷한 좌판 연기는 확인되지 않는다. 읽을 수 있는 글자는 없다.",
        "hard_violations": [
         "숏에 지정되지 않은 행인 세 명을 뒤쪽 통로와 양쪽 좌판에 추가했다."
        ],
        "physics": "안규홍의 화면 왼쪽 발은 지면을 딛고 몸을 지지하고, 화면 오른쪽 다리는 무릎을 굽힌 채 발 전체가 땅에서 떨어져 있다. 발 아래 빈 공간이 명확하며 보행 중 발을 앞으로 회수하는 순간으로 물리적으로 가능하다. 다만 길게 성큼 내딛는 순간보다는 보행 중간 단계에 가깝다. 닭장과 그릇은 목재 좌판에, 바구니와 항아리는 좌판 또는 지면에 놓여 있고 소총은 몸과 휴대 장비로 지지된다. 지지 없이 떠 있는 물체는 없다."
       }
      ]
     },
     "ok": true
    }
   ],
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "slot_winner_match": true,
   "slot_winner": {
    "gemini-pro": "A",
    "gpt-high": "A"
   },
   "route": "cross_slot_agree"
  },
  "dual": {
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "normalized": {
    "A": 2.0,
    "B": 1.25
   },
   "adjusted": {
    "A": 1.75,
    "B": 1.0
   },
   "violations": {
    "A": [
     "[gemini-pro] 샷 텍스트에 명시되지 않은 배경 인물(행인) 추가됨",
     "[gpt-high] 숏에 지정되지 않은 행인 세 명을 뒤쪽 통로와 양쪽 좌판에 추가했다."
    ],
    "B": [
     "[gemini-pro] 샷 텍스트에 명시되지 않은 주변 인물(마을 주민) 추가됨",
     "[gpt-high] 숏에 지정되지 않은 행인 네 명을 좌판 주변과 오른쪽 가장자리에 추가했다."
    ]
   },
   "per_model_winner": {
    "gemini-pro": "A",
    "gpt-high": "A"
   },
   "agreed": true
  },
  "totals": {
   "A": 1750,
   "B": 1000
  },
  "selected": "A",
  "ranking": [
   "A",
   "B"
  ],
  "verdicts": [
   {
    "label": "A",
    "score": 1750,
    "verdict_ko": "샷 텍스트와 프레이밍이 요구한 '공중에 떠 있는 한쪽 발'과 그 아래의 빈 지면을 완벽히 구현했으나, 배경에 명시되지 않은 인물이 작게 포함되어 감점되었습니다.  ★위반: [gemini-pro] 샷 텍스트에 명시되지 않은 배경 인물(행인) 추가됨 / [gpt-high] 숏에 지정되지 않은 행인 세 명을 뒤쪽 통로와 양쪽 좌판에 추가했다."
   },
   {
    "label": "B",
    "score": 1000,
    "verdict_ko": "바닥에 내려놓은 짐틀의 묘사는 좋으나, 가장 중요한 '공중에 떠 있는 발' 동작이 전혀 구현되지 않았고 텍스트에 없는 인물이 뚜렷하게 등장해 지시를 위반했습니다.  ★위반: [gemini-pro] 샷 텍스트에 명시되지 않은 주변 인물(마을 주민) 추가됨 / [gpt-high] 숏에 지정되지 않은 행인 네 명을 좌판 주변과 오른쪽 가장자리에 추가했다."
   }
  ],
  "refs": [
   {
    "label": "LOCATION PHOTOGRAPH — the exact place of this shot: its architecture, materials, fixed features and lighting mood are spatial truth; stage the moment inside this place. Never copy its camera framing.",
    "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/db00517b-165b-4005-a287-f28936b59b95/images/272bd7af-9294-4490-98c6-88a40e9deacc/scene/recipe/groupbg_market_center_fa2dce.png",
    "asset_id": "39a96dca-cf22-4c62-bacb-511754a80868",
    "role": "bgfirst_group_bg"
   },
   {
    "label": "CHARACTER REFERENCE — 안규홍: the exact person appearing in this shot; match face, hair and build exactly.",
    "path": "<bytes:1028384>",
    "asset_id": "3b5894db-ce55-4f34-b966-5e725c3aa952",
    "role": "character_ref"
   }
  ],
  "critique_skipped": true,
  "shot_run_produced": false,
  "shot_run_uid": "06aa1401-5375-7f1b-a7e9-1ea7bd3c65ef",
  "bgfirst": {
   "bg_path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/db00517b-165b-4005-a287-f28936b59b95/images/272bd7af-9294-4490-98c6-88a40e9deacc/scene/recipe/S1sh10__bgfirst_bg.png",
   "bg_asset_id": "e9684387-9242-4865-88ee-81c863d1434f",
   "bg_record_key": "S1sh10::bgfirst_bg",
   "chain_winner": true,
   "authority": "groupbg",
   "group_key": "market_center",
   "groupbg_asset_id": "39a96dca-cf22-4c62-bacb-511754a80868"
  },
  "ref_mode": "재투영 배경+콘티+엔티티 (2택1: 체인 승)",
  "share_plan": {
   "ref_plan": "background"
  }
 },
 "S1sh10::cine": {
  "applied": true,
  "attempted_at": "2026-09-09T00:36:08.623302+00:00",
  "fingerprint": "2818df7586ac3a825a1acf5d850f6e5fc500371c83f93adbe470d7a9e1dd59a3",
  "fingerprint_version": 2,
  "provider": "mai",
  "endpoint": "openrouter/chat-completions",
  "model": "microsoft/mai-image-2.6",
  "pack": "24.202608252115",
  "source_file": "S1sh10_sel.png",
  "source_sha256": "cb90102fb2e8c5c03b1fcdd77c8d233c62383d5a5fd2581899f12edae490064d",
  "file": "S1sh10_cine.png",
  "staged_sha256": "3b41bf6a4af4bb3305082fb2fdf3cfa1ee5a37d4ab1d365987f0deac6668e884",
  "latency_ms": 26939
 },
 "S1sh14::signage": {
  "fp": "a035c10539b0b5a2",
  "inscriptions": [],
  "cues": [],
  "dropped": []
 },
 "S1sh14": {
  "input_fingerprint": "768290518f5e7a1a",
  "prompt": "Create ONE FINAL photorealistic live-action film still of the moment below — early 20th-century Korea, Jeolla region; all people are Korean unless explicitly stated otherwise. TIME OF DAY (lock): day, harsh sunlight.\n\nSHOT TEXT (authoritative, Korean): 안규홍의 굵은 주먹이 헌병 장교의 턱에 정통으로 격돌하여 장교의 턱뼈가 옆으로 밀린 찰나.\n\nLOCATION (lock): The outdoor confrontation space beside the earthenware stalls in the market's central dirt thoroughfare. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: close-up\n- KEY BACKGROUND ELEMENTS: 장터 좌판 (Still present behind the fight); used as A narrow, unfocused background band anchors the collision in the established market space.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Maintain the same hard daytime sunlight and restrained saturation, allowing natural facial modeling to carry the violence without exaggerated impact lighting.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The clothing, hair and overall look of 안규홍 — who appear both in that photo and in this shot — are LOCKED to that photo. Anyone else visible in that photo is NOT in this shot: never carry their face, body or clothing onto anyone here. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nPREVIOUS STILL USAGE (follow exactly — what to take from the attached still and what to exclude; it governs place and objects only, never who is in this shot or how the camera sees them): Take the sunlit dirt ground and market-stall materials visible in the reference. Exclude the earlier walking action and any airborne dust unique to that instant.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The carrying frame remains on the dirt in the sunlit market center. 안규홍: He wears the dirt-stained jeogori and has no object in his striking hand. 헌병 장교: His head is knocked back at the start of his backward fall.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 안규홍 (한국인 남성, 청년 나이대, 검은 머리, 단단히 묶은 머리); 헌병 장교 (일본인 남성, 성인 나이대, 자연스러운 검은 머리) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.",
  "roll_prompts": {
   "A": "Create ONE FINAL photorealistic live-action film still of the moment below — early 20th-century Korea, Jeolla region; all people are Korean unless explicitly stated otherwise. TIME OF DAY (lock): day, harsh sunlight.\n\nSHOT TEXT (authoritative, Korean): 안규홍의 굵은 주먹이 헌병 장교의 턱에 정통으로 격돌하여 장교의 턱뼈가 옆으로 밀린 찰나.\n\nLOCATION (lock): The outdoor confrontation space beside the earthenware stalls in the market's central dirt thoroughfare. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: close-up\n- KEY BACKGROUND ELEMENTS: 장터 좌판 (Still present behind the fight); used as A narrow, unfocused background band anchors the collision in the established market space.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Maintain the same hard daytime sunlight and restrained saturation, allowing natural facial modeling to carry the violence without exaggerated impact lighting.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The clothing, hair and overall look of 안규홍 — who appear both in that photo and in this shot — are LOCKED to that photo. Anyone else visible in that photo is NOT in this shot: never carry their face, body or clothing onto anyone here. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nPREVIOUS STILL USAGE (follow exactly — what to take from the attached still and what to exclude; it governs place and objects only, never who is in this shot or how the camera sees them): Take the sunlit dirt ground and market-stall materials visible in the reference. Exclude the earlier walking action and any airborne dust unique to that instant.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The carrying frame remains on the dirt in the sunlit market center. 안규홍: He wears the dirt-stained jeogori and has no object in his striking hand. 헌병 장교: His head is knocked back at the start of his backward fall.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 안규홍 (한국인 남성, 청년 나이대, 검은 머리, 단단히 묶은 머리); 헌병 장교 (일본인 남성, 성인 나이대, 자연스러운 검은 머리) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.",
   "B": "Create ONE FINAL photorealistic live-action film still of the moment below — early 20th-century Korea, Jeolla region; all people are Korean unless explicitly stated otherwise. TIME OF DAY (lock): day, harsh sunlight.\n\nSHOT TEXT (authoritative, Korean): 안규홍의 굵은 주먹이 헌병 장교의 턱에 정통으로 격돌하여 장교의 턱뼈가 옆으로 밀린 찰나.\n\nLOCATION (lock): The outdoor confrontation space beside the earthenware stalls in the market's central dirt thoroughfare. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: close-up\n- KEY BACKGROUND ELEMENTS: 장터 좌판 (Still present behind the fight); used as A narrow, unfocused background band anchors the collision in the established market space.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Maintain the same hard daytime sunlight and restrained saturation, allowing natural facial modeling to carry the violence without exaggerated impact lighting.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The clothing, hair and overall look of 안규홍 — who appear both in that photo and in this shot — are LOCKED to that photo. Anyone else visible in that photo is NOT in this shot: never carry their face, body or clothing onto anyone here. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nPREVIOUS STILL USAGE (follow exactly — what to take from the attached still and what to exclude; it governs place and objects only, never who is in this shot or how the camera sees them): Take the sunlit dirt ground and market-stall materials visible in the reference. Exclude the earlier walking action and any airborne dust unique to that instant.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The carrying frame remains on the dirt in the sunlit market center. 안규홍: He wears the dirt-stained jeogori and has no object in his striking hand. 헌병 장교: His head is knocked back at the start of his backward fall.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 안규홍 (한국인 남성, 청년 나이대, 검은 머리, 단단히 묶은 머리); 헌병 장교 (일본인 남성, 성인 나이대, 자연스러운 검은 머리) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.\n\nCOMPOSITION VARIATION (this candidate only): keep the place, the people and the moment exactly as specified, and keep from the CAMERA & FRAME contract its subject scale, screen placement and key background elements exactly. For this candidate only, its camera angle and height are deliberately left open: choose a camera position distinctly different from the obvious one, as a film director picking a second setup on the same blocking. Stay true to any explicit framing the SHOT TEXT itself dictates."
  },
  "shot_run_spend_attempt_count": 1,
  "readings": [
   {
    "label": "A",
    "direction": "안규홍의 오른쪽 주먹이 장교의 턱을 향해 있으며, 장교의 시선은 위를 향함.",
    "built_space": "아웃포커싱된 배경에 장터 좌판과 항아리들이 위치함.",
    "entities": "안규홍의 외양은 일치하나, 장교는 캐릭터 레퍼런스의 복장(짙은 녹색, 훈장)이 아닌 이전 샷의 연한 카키색 복장을 입고 있음. 장교의 눈동자가 하얗게 뒤집힘.",
    "hard_violations": [
     "[gemini-pro] 장교의 눈동자가 하얗게 변형됨 (해부학적 변형 금지 위반)",
     "[gemini-pro] 장교의 복장이 캐릭터 레퍼런스와 불일치함"
    ],
    "physics": "주먹이 턱에 닿아 고개가 뒤로 젖혀지고 있으나 타격의 체중이 실린 느낌이 다소 부족함."
   },
   {
    "label": "B",
    "direction": "안규홍의 오른쪽 주먹이 장교의 턱에 정통으로 꽂히며 장교의 턱이 옆으로 밀림.",
    "built_space": "프롬프트에 지시된 대로 뒤쪽에 장터 좌판과 항아리들이 아웃포커스된 띠 형태로 자리 잡고 있음.",
    "entities": "안규홍의 더러워진 저고리와 묶은 머리, 장교의 얼굴 및 복장(훈장, 계급장 등)이 캐릭터 레퍼런스와 정확히 일치함.",
    "hard_violations": [
     "[gpt-high] 장교 소매에 ‘헌병’이라는 읽을 수 있는 글자가 노출되어, 화면 어디에도 판독 가능한 문자를 두지 말라는 금지 조건을 위반한다."
    ],
    "physics": "주먹의 타격에 의해 장교의 턱뼈가 밀리고 상체가 뒤로 넘어가는 물리적 동작이 자연스럽게 구현됨."
   }
  ],
  "cross_model_order": {
   "policy": "cross_model_order_v2_slot0_fwd_slot1_rev_parallel2",
   "slots": [
    {
     "model": "gemini-pro",
     "order": "forward",
     "display_to_canonical": {
      "A": "A",
      "B": "B"
     },
     "raw": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 9,
        "verdict_ko": "캐릭터 레퍼런스의 복장과 외양을 정확히 반영하였으며, 타격 순간의 물리적 묘사와 표정 연기가 프롬프트와 완벽히 일치합니다."
       },
       {
        "label": "A",
        "score": 3,
        "verdict_ko": "장교의 눈동자가 하얗게 변하는 해부학적 변형 금지 규정을 위반했으며, 지정된 캐릭터 레퍼런스가 아닌 이전 샷의 엑스트라 복장을 차용했습니다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "안규홍의 오른쪽 주먹이 장교의 턱을 향해 있으며, 장교의 시선은 위를 향함.",
        "built_space": "아웃포커싱된 배경에 장터 좌판과 항아리들이 위치함.",
        "entities": "안규홍의 외양은 일치하나, 장교는 캐릭터 레퍼런스의 복장(짙은 녹색, 훈장)이 아닌 이전 샷의 연한 카키색 복장을 입고 있음. 장교의 눈동자가 하얗게 뒤집힘.",
        "hard_violations": [
         "장교의 눈동자가 하얗게 변형됨 (해부학적 변형 금지 위반)",
         "장교의 복장이 캐릭터 레퍼런스와 불일치함"
        ],
        "physics": "주먹이 턱에 닿아 고개가 뒤로 젖혀지고 있으나 타격의 체중이 실린 느낌이 다소 부족함."
       },
       {
        "label": "B",
        "direction": "안규홍의 오른쪽 주먹이 장교의 턱에 정통으로 꽂히며 장교의 턱이 옆으로 밀림.",
        "built_space": "프롬프트에 지시된 대로 뒤쪽에 장터 좌판과 항아리들이 아웃포커스된 띠 형태로 자리 잡고 있음.",
        "entities": "안규홍의 더러워진 저고리와 묶은 머리, 장교의 얼굴 및 복장(훈장, 계급장 등)이 캐릭터 레퍼런스와 정확히 일치함.",
        "hard_violations": [],
        "physics": "주먹의 타격에 의해 장교의 턱뼈가 밀리고 상체가 뒤로 넘어가는 물리적 동작이 자연스럽게 구현됨."
       }
      ],
      "all_candidates_fail": false
     },
     "normalized": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 9,
        "verdict_ko": "캐릭터 레퍼런스의 복장과 외양을 정확히 반영하였으며, 타격 순간의 물리적 묘사와 표정 연기가 프롬프트와 완벽히 일치합니다."
       },
       {
        "label": "A",
        "score": 3,
        "verdict_ko": "장교의 눈동자가 하얗게 변하는 해부학적 변형 금지 규정을 위반했으며, 지정된 캐릭터 레퍼런스가 아닌 이전 샷의 엑스트라 복장을 차용했습니다."
       }
      ],
      "all_candidates_fail": false,
      "readings": [
       {
        "label": "A",
        "direction": "안규홍의 오른쪽 주먹이 장교의 턱을 향해 있으며, 장교의 시선은 위를 향함.",
        "built_space": "아웃포커싱된 배경에 장터 좌판과 항아리들이 위치함.",
        "entities": "안규홍의 외양은 일치하나, 장교는 캐릭터 레퍼런스의 복장(짙은 녹색, 훈장)이 아닌 이전 샷의 연한 카키색 복장을 입고 있음. 장교의 눈동자가 하얗게 뒤집힘.",
        "hard_violations": [
         "장교의 눈동자가 하얗게 변형됨 (해부학적 변형 금지 위반)",
         "장교의 복장이 캐릭터 레퍼런스와 불일치함"
        ],
        "physics": "주먹이 턱에 닿아 고개가 뒤로 젖혀지고 있으나 타격의 체중이 실린 느낌이 다소 부족함."
       },
       {
        "label": "B",
        "direction": "안규홍의 오른쪽 주먹이 장교의 턱에 정통으로 꽂히며 장교의 턱이 옆으로 밀림.",
        "built_space": "프롬프트에 지시된 대로 뒤쪽에 장터 좌판과 항아리들이 아웃포커스된 띠 형태로 자리 잡고 있음.",
        "entities": "안규홍의 더러워진 저고리와 묶은 머리, 장교의 얼굴 및 복장(훈장, 계급장 등)이 캐릭터 레퍼런스와 정확히 일치함.",
        "hard_violations": [],
        "physics": "주먹의 타격에 의해 장교의 턱뼈가 밀리고 상체가 뒤로 넘어가는 물리적 동작이 자연스럽게 구현됨."
       }
      ]
     },
     "ok": true
    },
    {
     "model": "gpt-high",
     "order": "reverse",
     "display_to_canonical": {
      "A": "B",
      "B": "A"
     },
     "raw": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 3,
        "verdict_ko": "장교의 뒤로 젖혀지는 반응과 제복은 비교적 충실하지만, 소매의 읽히는 ‘헌병’ 글자가 금지 조건을 위반하고 허리까지 담은 구도도 요구한 클로즈업보다 넓습니다."
       },
       {
        "label": "B",
        "score": 7,
        "verdict_ko": "주먹이 턱에 닿는 순간을 더 밀착해 포착하고 시선과 타격 방향도 맞지만, 장교의 밝은 제복과 붉은 깃은 지정 인물보다 이전 장면 병사들의 복장에 가까워 연속성이 떨어집니다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "왼쪽 안규홍은 오른쪽 장교의 얼굴을 보며 주먹을 오른쪽으로 뻗는다. 주먹은 장교의 턱 옆 아래쪽, 목과 맞닿는 경계에 접촉한다. 장교의 머리는 화면 오른쪽 뒤로 기울고 얼굴은 타격 쪽으로 돌아 있어 충격 반응은 보이지만, 턱 정면을 정확히 가격하는 접점은 다소 낮다. 겨누는 무기는 없다.",
        "built_space": "뒤쪽 중앙에는 항아리와 그릇을 얹은 목제 좌판이 있고, 오른쪽에는 바구니와 도기를 둔 별도 진열 구역이 보인다. 흙벽, 목재 기둥, 햇빛을 받은 흙길은 이전 시장의 재료와 부합한다. 두 사람은 좌판 앞 통로에 있으며 구조물과 충돌하지 않는다. 다만 배경과 몸통이 차지하는 면적이 커서 좁고 흐릿한 배경 띠를 둔 클로즈업보다 넓다. 반사면이나 명백하게 중복된 고정 설비는 없다.",
        "entities": "두 명의 성인 동아시아계 남성만 보인다. 안규홍의 검은 묶음머리와 때 묻은 밝은 저고리는 참조와 대체로 맞으며, 옆얼굴만 보여 정확한 얼굴 일치는 제한적으로 확인된다. 장교는 검은 머리, 올리브색 제복, 견장, 훈장, 가죽 띠와 군모를 갖춰 지정 참조에 비교적 가깝다. 국적은 외모만으로 확정할 수 없다. 타격하는 손에는 물건이 없다. 소매 표장의 ‘헌병’은 명확히 읽힌다. 지게의 지면 잔류 여부는 이 구도에서 확인되지 않는다.",
        "hard_violations": [
         "장교 소매에 ‘헌병’이라는 읽을 수 있는 글자가 노출되어, 화면 어디에도 판독 가능한 문자를 두지 말라는 금지 조건을 위반한다."
        ],
        "physics": "주먹은 손목과 소매 속 팔을 통해 안규홍의 어깨로 이어지고, 장교의 턱 아래에 실제로 접촉한다. 장교의 머리는 목과 몸통에 연결된 채 뒤로 기울어 넘어지기 시작하는 반응으로 가능하다. 발은 화면 밖이므로 지면 접촉은 확인할 수 없지만, 몸 전체가 공중에 떠 있는 묘사는 아니다. 군모는 머리에 얹혀 있고 훈장과 가죽 장비는 옷과 띠에 부착되어 있다. 입 주변의 작은 비말은 충격으로 튄 것으로 읽힌다."
       },
       {
        "label": "B",
        "direction": "오른쪽 안규홍의 시선은 왼쪽 장교의 얼굴에 고정되어 있고, 뻗은 주먹은 화면 왼쪽으로 향해 장교의 턱 옆에 직접 닿는다. 장교의 턱은 들리고 머리는 뒤로 젖혀지며 눈은 위쪽으로 돌아간다. 타격 목표와 주먹의 도달 지점이 일치한다. 무기를 겨누는 동작은 없다.",
        "built_space": "왼쪽에는 도기와 그릇을 놓은 목제 좌판, 중앙 뒤에는 낮은 진열대, 오른쪽 뒤에는 차양 아래 이어지는 좌판이 보인다. 흙벽, 거친 목재 지지대, 초가 차양과 흙길은 이전 시장의 재료 구성을 유지한다. 인물들은 좌판 앞 통로에 서 있으며 설비와 겹쳐 서 있는 문제는 없다. 얼굴과 충돌 지점이 A보다 크게 잡히지만 배경은 여전히 좁은 띠 이상으로 노출된다. 불가능한 반사나 식별 가능한 고정 설비 중복은 없다.",
        "entities": "두 명의 성인 동아시아계 남성만 등장한다. 안규홍은 뒤로 단단히 넘긴 검은 머리, 참조와 유사한 얼굴, 때 묻은 밝은 저고리를 갖췄다. 장교는 검은 머리와 군모를 착용했으나 얼굴이 참조보다 나이 들어 보이고, 밝은 카키색 제복과 붉은 깃은 지정 장교의 어두운 올리브색 제복보다 이전 장면 병사들의 복장에 가깝다. 한국인과 일본인이라는 국적 자체는 외모만으로 확인되지 않는다. 타격하는 손은 빈 주먹이며 반대 손도 몸 앞에서 주먹을 쥐고 있다. 읽히는 문자는 없다. 지게는 화면 밖이어서 상태를 판단할 수 없다.",
        "hard_violations": [],
        "physics": "타격 주먹은 손목과 전완에 자연스럽게 이어지고 턱에 접촉한다. 팔의 뻗음과 장교의 뒤로 들리는 턱이 같은 충격 작용으로 설명된다. 장교의 몸통은 아직 비교적 세워져 있어 후방 낙상의 시작은 머리 반응으로 주로 표현된다. 두 사람의 하체는 잘려 있지만 근거 없이 공중에 떠 있는 몸은 보이지 않는다. 군모는 머리에 걸쳐 있고 가죽 띠는 어깨에 지지된다. 작은 비말은 입 주변에서 타격으로 튄 것으로 해석 가능하다."
       }
      ],
      "all_candidates_fail": false
     },
     "normalized": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 3,
        "verdict_ko": "장교의 뒤로 젖혀지는 반응과 제복은 비교적 충실하지만, 소매의 읽히는 ‘헌병’ 글자가 금지 조건을 위반하고 허리까지 담은 구도도 요구한 클로즈업보다 넓습니다."
       },
       {
        "label": "A",
        "score": 7,
        "verdict_ko": "주먹이 턱에 닿는 순간을 더 밀착해 포착하고 시선과 타격 방향도 맞지만, 장교의 밝은 제복과 붉은 깃은 지정 인물보다 이전 장면 병사들의 복장에 가까워 연속성이 떨어집니다."
       }
      ],
      "all_candidates_fail": false,
      "readings": [
       {
        "label": "B",
        "direction": "왼쪽 안규홍은 오른쪽 장교의 얼굴을 보며 주먹을 오른쪽으로 뻗는다. 주먹은 장교의 턱 옆 아래쪽, 목과 맞닿는 경계에 접촉한다. 장교의 머리는 화면 오른쪽 뒤로 기울고 얼굴은 타격 쪽으로 돌아 있어 충격 반응은 보이지만, 턱 정면을 정확히 가격하는 접점은 다소 낮다. 겨누는 무기는 없다.",
        "built_space": "뒤쪽 중앙에는 항아리와 그릇을 얹은 목제 좌판이 있고, 오른쪽에는 바구니와 도기를 둔 별도 진열 구역이 보인다. 흙벽, 목재 기둥, 햇빛을 받은 흙길은 이전 시장의 재료와 부합한다. 두 사람은 좌판 앞 통로에 있으며 구조물과 충돌하지 않는다. 다만 배경과 몸통이 차지하는 면적이 커서 좁고 흐릿한 배경 띠를 둔 클로즈업보다 넓다. 반사면이나 명백하게 중복된 고정 설비는 없다.",
        "entities": "두 명의 성인 동아시아계 남성만 보인다. 안규홍의 검은 묶음머리와 때 묻은 밝은 저고리는 참조와 대체로 맞으며, 옆얼굴만 보여 정확한 얼굴 일치는 제한적으로 확인된다. 장교는 검은 머리, 올리브색 제복, 견장, 훈장, 가죽 띠와 군모를 갖춰 지정 참조에 비교적 가깝다. 국적은 외모만으로 확정할 수 없다. 타격하는 손에는 물건이 없다. 소매 표장의 ‘헌병’은 명확히 읽힌다. 지게의 지면 잔류 여부는 이 구도에서 확인되지 않는다.",
        "hard_violations": [
         "장교 소매에 ‘헌병’이라는 읽을 수 있는 글자가 노출되어, 화면 어디에도 판독 가능한 문자를 두지 말라는 금지 조건을 위반한다."
        ],
        "physics": "주먹은 손목과 소매 속 팔을 통해 안규홍의 어깨로 이어지고, 장교의 턱 아래에 실제로 접촉한다. 장교의 머리는 목과 몸통에 연결된 채 뒤로 기울어 넘어지기 시작하는 반응으로 가능하다. 발은 화면 밖이므로 지면 접촉은 확인할 수 없지만, 몸 전체가 공중에 떠 있는 묘사는 아니다. 군모는 머리에 얹혀 있고 훈장과 가죽 장비는 옷과 띠에 부착되어 있다. 입 주변의 작은 비말은 충격으로 튄 것으로 읽힌다."
       },
       {
        "label": "A",
        "direction": "오른쪽 안규홍의 시선은 왼쪽 장교의 얼굴에 고정되어 있고, 뻗은 주먹은 화면 왼쪽으로 향해 장교의 턱 옆에 직접 닿는다. 장교의 턱은 들리고 머리는 뒤로 젖혀지며 눈은 위쪽으로 돌아간다. 타격 목표와 주먹의 도달 지점이 일치한다. 무기를 겨누는 동작은 없다.",
        "built_space": "왼쪽에는 도기와 그릇을 놓은 목제 좌판, 중앙 뒤에는 낮은 진열대, 오른쪽 뒤에는 차양 아래 이어지는 좌판이 보인다. 흙벽, 거친 목재 지지대, 초가 차양과 흙길은 이전 시장의 재료 구성을 유지한다. 인물들은 좌판 앞 통로에 서 있으며 설비와 겹쳐 서 있는 문제는 없다. 얼굴과 충돌 지점이 A보다 크게 잡히지만 배경은 여전히 좁은 띠 이상으로 노출된다. 불가능한 반사나 식별 가능한 고정 설비 중복은 없다.",
        "entities": "두 명의 성인 동아시아계 남성만 등장한다. 안규홍은 뒤로 단단히 넘긴 검은 머리, 참조와 유사한 얼굴, 때 묻은 밝은 저고리를 갖췄다. 장교는 검은 머리와 군모를 착용했으나 얼굴이 참조보다 나이 들어 보이고, 밝은 카키색 제복과 붉은 깃은 지정 장교의 어두운 올리브색 제복보다 이전 장면 병사들의 복장에 가깝다. 한국인과 일본인이라는 국적 자체는 외모만으로 확인되지 않는다. 타격하는 손은 빈 주먹이며 반대 손도 몸 앞에서 주먹을 쥐고 있다. 읽히는 문자는 없다. 지게는 화면 밖이어서 상태를 판단할 수 없다.",
        "hard_violations": [],
        "physics": "타격 주먹은 손목과 전완에 자연스럽게 이어지고 턱에 접촉한다. 팔의 뻗음과 장교의 뒤로 들리는 턱이 같은 충격 작용으로 설명된다. 장교의 몸통은 아직 비교적 세워져 있어 후방 낙상의 시작은 머리 반응으로 주로 표현된다. 두 사람의 하체는 잘려 있지만 근거 없이 공중에 떠 있는 몸은 보이지 않는다. 군모는 머리에 걸쳐 있고 가죽 띠는 어깨에 지지된다. 작은 비말은 입 주변에서 타격으로 튄 것으로 해석 가능하다."
       }
      ]
     },
     "ok": true
    }
   ],
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "slot_winner_match": false,
   "slot_winner": {
    "gemini-pro": "B",
    "gpt-high": "A"
   },
   "route": "cross_slot_combined"
  },
  "dual": {
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "normalized": {
    "A": 1.333,
    "B": 1.429
   },
   "adjusted": {
    "A": 1.083,
    "B": 1.179
   },
   "violations": {
    "A": [
     "[gemini-pro] 장교의 눈동자가 하얗게 변형됨 (해부학적 변형 금지 위반)",
     "[gemini-pro] 장교의 복장이 캐릭터 레퍼런스와 불일치함"
    ],
    "B": [
     "[gpt-high] 장교 소매에 ‘헌병’이라는 읽을 수 있는 글자가 노출되어, 화면 어디에도 판독 가능한 문자를 두지 말라는 금지 조건을 위반한다."
    ]
   },
   "per_model_winner": {
    "gemini-pro": "B",
    "gpt-high": "A"
   },
   "agreed": false
  },
  "totals": {
   "B": 1179,
   "A": 1083
  },
  "selected": "B",
  "ranking": [
   "B",
   "A"
  ],
  "verdicts": [
   {
    "label": "B",
    "score": 1179,
    "verdict_ko": "캐릭터 레퍼런스의 복장과 외양을 정확히 반영하였으며, 타격 순간의 물리적 묘사와 표정 연기가 프롬프트와 완벽히 일치합니다.  ★위반: [gpt-high] 장교 소매에 ‘헌병’이라는 읽을 수 있는 글자가 노출되어, 화면 어디에도 판독 가능한 문자를 두지 말라는 금지 조건을 위반한다."
   },
   {
    "label": "A",
    "score": 1083,
    "verdict_ko": "장교의 눈동자가 하얗게 변하는 해부학적 변형 금지 규정을 위반했으며, 지정된 캐릭터 레퍼런스가 아닌 이전 샷의 엑스트라 복장을 차용했습니다.  ★위반: [gemini-pro] 장교의 눈동자가 하얗게 변형됨 (해부학적 변형 금지 위반) / [gemini-pro] 장교의 복장이 캐릭터 레퍼런스와 불일치함"
   }
  ],
  "refs": [
   {
    "label": "PREVIOUS SHOT STILL — a visually related earlier shot of this same place: the location's look, materials, fixed features and lighting mood are LOCKED to this photo; never copy its camera framing. The clothing, hair and overall look of 안규홍 — who appear both in that photo and in this shot — are LOCKED to that photo. Anyone else visible in that photo is NOT in this shot: never carry their face, body or clothing onto anyone here. If this photo shows a character who cannot move (dead or unconscious), that character's exact pose, position and orientation are ALSO LOCKED — treat their body as a fixed prop of the set that only the camera moves around.",
    "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/db00517b-165b-4005-a287-f28936b59b95/images/272bd7af-9294-4490-98c6-88a40e9deacc/scene/recipe/S1sh10_sel.png",
    "asset_id": "f1ae73c0-86b8-44fb-90bd-89aa539b1b61",
    "role": "prev_still"
   },
   {
    "label": "CHARACTER REFERENCE — 안규홍: the exact person appearing in this shot; match face, hair and build exactly.",
    "path": "<bytes:1028384>",
    "asset_id": "3b5894db-ce55-4f34-b966-5e725c3aa952",
    "role": "character_ref"
   },
   {
    "label": "CHARACTER REFERENCE — 헌병 장교: the exact person appearing in this shot; match face, hair and build exactly.",
    "path": "<bytes:862147>",
    "asset_id": "83cd18a8-56b7-4e8a-8937-e155fe16c041",
    "role": "character_ref"
   }
  ],
  "critique_skipped": true,
  "shot_run_produced": false,
  "shot_run_uid": "06aa1401-578c-76d6-94fd-ad763ba584d4",
  "ref_mode": "prev+엔티티",
  "share_plan": {
   "ref_plan": "prev",
   "prev_anchor_tag": "S1sh10"
  }
 },
 "S1sh14::cine": {
  "applied": true,
  "attempted_at": "2026-09-09T00:37:39.262339+00:00",
  "fingerprint": "abee98226ac34cfdab7643a67e2043d1392c212cbbcb23d2da9cfc5671e1ddaf",
  "fingerprint_version": 2,
  "provider": "mai",
  "endpoint": "openrouter/chat-completions",
  "model": "microsoft/mai-image-2.6",
  "pack": "24.202608252115",
  "source_file": "S1sh14_sel.png",
  "source_sha256": "1e7a92f40e5a0f11da6be1f23aab899242e8b70d7418792c5871e6f3c408dbb2",
  "file": "S1sh14_cine.png",
  "staged_sha256": "d84d2d025883707f1720118be8ff4a7631bfac69b52fa38914154401e932a781",
  "latency_ms": 25577
 },
 "S2sh1::signage": {
  "fp": "ea2e39332d012318",
  "inscriptions": [],
  "cues": [],
  "dropped": []
 },
 "era_assess::30a460a465c9f9f1": {
  "subjects": [
   {
    "subject_native": "20세기 초 한국 전라도 전통 민가의 부뚜막과 아궁이",
    "search_terms_native": [
     "전통 민가 부엌 아궁이",
     "초가집 부뚜막 가마솥",
     "조선시대 민가 부엌 실내",
     "옛날 시골 부엌 살림"
    ],
    "language_lock_native": "반드시 한국어로만 검색해야 하며 다른 언어로 번역하거나 외래어를 섞지 마십시오.",
    "reason_ko": "일반 AI 모델은 온돌 난방과 연결된 고유의 흙부뚜막 구조 대신 서구식 벽난로나 일본식 가마도 형태로 왜곡해 생성하기 쉽습니다."
   }
  ],
  "subject_text": "안규홍의 집 부엌\n장작을 넣는 아궁이와 약탕관이 있는 소박한 부엌. 구석에 지게를 둘 자리가 있고, 어두운 실내에 불빛이 낮게 번진다.",
  "identity": "canonical",
  "scope_id": "L03",
  "scope_role": "location_interior",
  "scope_sha": "637032a5af3eb4c7"
 },
 "era_fail::b529ef4c817c9125": {
  "stage": "research",
  "subject": "20세기 초 한국 전라도 전통 민가의 부뚜막과 아궁이",
  "terms": [
   "전통 민가 부엌 아궁이",
   "초가집 부뚜막 가마솥",
   "조선시대 민가 부엌 실내",
   "옛날 시골 부엌 살림"
  ],
  "status": "no_usable",
  "queries": [
   [
    "전라도 전통 민가 부엌 부뚜막 아궁이 가마솥",
    "국립민속박물관 부엌 부뚜막 아궁이 살림"
   ],
   [
    "낙안읍성 초가집 부엌 내부 부뚜막"
   ]
  ],
  "candidate_urls": [
   "https://www.hansik.or.kr/files/magazine/uploads/oldimg/202301/233a952af610fccd6270a53f344b8ccc_crop.jpg?t=18601504673",
   "https://gongu.copyright.or.kr/gongu/wrt/cmmn/wrtFileImageView.do?filePath=L2Rpc2sxL25ld2RhdGEvMjAxNC8yMS9DTFM2L2RpZ2lfMTExMjk1MDRfMDEyMDE0MTExMzA3&thumbAt=Y&thumbSe=b_tbumb&wrtSn=11129504&wrtTy=10006",
   "https://cdn.igj.co.kr/news/photo/201504/38258_78995_3158.jpg",
   "https://tong.visitkorea.or.kr/cms/resource/80/3379380_image2_1.JPG"
  ],
  "coarse": {
   "eligible": [],
   "chosen_index": 0,
   "reason": "종류·보임·기준을 다 만족하는 후보가 없다 — 이 라운드에선 안 고른다",
   "single_judge": true,
   "rejected_judges": {}
  },
  "verdicts": [
   {
    "index": 1,
    "object_type_match": "yes",
    "visible": true,
    "criteria_match": "unsure",
    "similarity": 95
   },
   {
    "index": 2,
    "object_type_match": "yes",
    "visible": true,
    "criteria_match": "unsure",
    "similarity": 80
   },
   {
    "index": 3,
    "object_type_match": "yes",
    "visible": true,
    "criteria_match": "unsure",
    "similarity": 100
   },
   {
    "index": 4,
    "object_type_match": "yes",
    "visible": true,
    "criteria_match": "unsure",
    "similarity": 90
   }
  ],
  "chosen_reason_ko": "",
  "attempts": 5
 },
 "S2sh1::bgfirst_bg": {
  "input_fingerprint": "b9938c2ec8270f81",
  "prompt": "Create the EMPTY BACKGROUND PLATE for one film shot — NO PEOPLE, no figures, no body parts, no sketch lines, no arrows anywhere.\n\"Empty\" means no people only: KEEP the location's inherent occupants and stock that define the place — animals in an animal shelter, pen or farm, goods in a market, moored boats in a harbour — unless the shot text explicitly removes them.\nThe FIRST attached image is a thin-line storyboard sketch: use ONLY its camera angle, horizon, perspective and the placement/size of buildings and set masses — ignore the sketched people and arrows entirely. The SECOND attached image (LOCATION PHOTOGRAPH) is the real place: take its architecture, materials, signage and fixed features, and RE-PROJECT them into the sketch's camera. If the photograph's camera differs from the sketch's, the sketch's camera wins.\nHUMAN-SCALE CALIBRATION: derive every structure's true size from human-scale elements — a door ≈ 2m, a window ≈ 1–1.5m wide, one storey ≈ 2.5–3m; never inflate a small structure or shrink a large one.\n\nSHOT TEXT this background must serve (Korean): 어스름한 새벽, 아궁이 불씨를 향해 입술을 모으고 입김을 뿜어내는 안규홍의 쪼그려 앉은 옆얼굴.\n\nLOCATION (lock): The hearth mouth inside a modest home's kitchen at dawn, with a crouching space directly before the rekindling red embers.\n\nTIME OF DAY (lock): dawn.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: close-up\n- KEY BACKGROUND ELEMENTS: 아궁이 입구 (Containing a fading ember that brightens as 안규홍 blows) — The opening is seen obliquely beside his profile, ahead of and below his mouth; used as Occupies less than a third of the frame and supplies the visible destination of his breath.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Dim dawn ambience and the ember's returning red glow support a subdued, intimate image with gently preserved facial detail.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nRender ONE photorealistic empty location photograph, 16:9, that this shot can be staged inside later. No readable writing anywhere: surfaces that would carry writing may be present, but stage any wording out of legibility — an oblique angle, distance, shallow focus. No captions, watermarks or overlay text.",
  "effective_prompt": "Create the EMPTY BACKGROUND PLATE for one film shot — NO PEOPLE, no figures, no body parts, no sketch lines, no arrows anywhere.\n\"Empty\" means no people only: KEEP the location's inherent occupants and stock that define the place — animals in an animal shelter, pen or farm, goods in a market, moored boats in a harbour — unless the shot text explicitly removes them.\nThe FIRST attached image is a thin-line storyboard sketch: use ONLY its camera angle, horizon, perspective and the placement/size of buildings and set masses — ignore the sketched people and arrows entirely. The SECOND attached image (LOCATION PHOTOGRAPH) is the real place: take its architecture, materials, signage and fixed features, and RE-PROJECT them into the sketch's camera. If the photograph's camera differs from the sketch's, the sketch's camera wins.\nHUMAN-SCALE CALIBRATION: derive every structure's true size from human-scale elements — a door ≈ 2m, a window ≈ 1–1.5m wide, one storey ≈ 2.5–3m; never inflate a small structure or shrink a large one.\n\nSHOT TEXT this background must serve (Korean): 어스름한 새벽, 아궁이 불씨를 향해 입술을 모으고 입김을 뿜어내는 안규홍의 쪼그려 앉은 옆얼굴.\n\nLOCATION (lock): The hearth mouth inside a modest home's kitchen at dawn, with a crouching space directly before the rekindling red embers.\n\nTIME OF DAY (lock): dawn.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: close-up\n- KEY BACKGROUND ELEMENTS: 아궁이 입구 (Containing a fading ember that brightens as 안규홍 blows) — The opening is seen obliquely beside his profile, ahead of and below his mouth; used as Occupies less than a third of the frame and supplies the visible destination of his breath.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Dim dawn ambience and the ember's returning red glow support a subdued, intimate image with gently preserved facial detail.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nRender ONE photorealistic empty location photograph, 16:9, that this shot can be staged inside later. No readable writing anywhere: surfaces that would carry writing may be present, but stage any wording out of legibility — an oblique angle, distance, shallow focus. No captions, watermarks or overlay text.",
  "bg_path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/db00517b-165b-4005-a287-f28936b59b95/images/272bd7af-9294-4490-98c6-88a40e9deacc/scene/recipe/S2sh1__bgfirst_bg.png",
  "asset_id": "d3f51f81-7354-42de-ab2f-a6e421bb5630",
  "input_asset_ids": [
   "d94541fe-9ddd-41d9-8586-47426b75a334",
   "07dcde76-24fe-4940-acad-6a24344f1cbb"
  ]
 },
 "S2sh1": {
  "input_fingerprint": "2d1b3a8df07491da",
  "prompt": "Create ONE FINAL photorealistic live-action film still of the moment below — early 20th-century Korea, Jeolla region; all people are Korean unless explicitly stated otherwise. TIME OF DAY (lock): dawn.\n\nSHOT TEXT (authoritative, Korean): 어스름한 새벽, 아궁이 불씨를 향해 입술을 모으고 입김을 뿜어내는 안규홍의 쪼그려 앉은 옆얼굴.\n\nLOCATION (lock): The hearth mouth inside a modest home's kitchen at dawn, with a crouching space directly before the rekindling red embers. The shot takes place here — the attached LOCATION PHOTOGRAPH shows the exact spot.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: close-up\n- KEY BACKGROUND ELEMENTS: 아궁이 입구 (Containing a fading ember that brightens as 안규홍 blows) — The opening is seen obliquely beside his profile, ahead of and below his mouth; used as Occupies less than a third of the frame and supplies the visible destination of his breath.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Dim dawn ambience and the ember's returning red glow support a subdued, intimate image with gently preserved facial detail.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): A dying ember glows inside the hearth in the dim dawn; the carrying frame is in a corner. 안규홍: He crouches sleepily before the hearth, with a cold-rice bundle tucked at his waist.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 안규홍 (한국인 남성, 청년 나이대, 검은 머리, 단단히 묶은 머리) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.",
  "roll_prompts": {
   "A": "Stage the shot. The FIRST attached image (SHOT BACKGROUND) is the finished empty background of this shot — keep it EXACTLY: its camera, perspective, architecture, lighting and every fixed feature stay untouched. The SECOND attached image (LAYOUT SKETCH) tells you ONLY where the people go: each sketched person's position, screen size, pose and the gaze/motion arrows. Ignore the sketch's background lines. The CHARACTER REFERENCE photographs show the real people.\nPlace the real people into the background at exactly the sketched positions, sizes and poses, following the arrow directions. No sketch lines or arrows may remain.\n\nCreate ONE FINAL photorealistic live-action film still of the moment below — early 20th-century Korea, Jeolla region; all people are Korean unless explicitly stated otherwise. TIME OF DAY (lock): dawn.\n\nSHOT TEXT (authoritative, Korean): 어스름한 새벽, 아궁이 불씨를 향해 입술을 모으고 입김을 뿜어내는 안규홍의 쪼그려 앉은 옆얼굴.\n\nLOCATION (lock): The hearth mouth inside a modest home's kitchen at dawn, with a crouching space directly before the rekindling red embers. The shot takes place here — the FIRST attached image (SHOT BACKGROUND) is this exact place, already built: its ground, structures, horizon, materials and lighting are the finished truth of this location and must not be redesigned or replaced. No location photograph is attached — read the place from that image alone, and add no scenery, structure, vehicle or fixture that it does not already show. This lock governs the place only; the figures in the shot follow the staging and pose instructions.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: close-up\n- KEY BACKGROUND ELEMENTS: 아궁이 입구 (Containing a fading ember that brightens as 안규홍 blows) — The opening is seen obliquely beside his profile, ahead of and below his mouth; used as Occupies less than a third of the frame and supplies the visible destination of his breath.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Dim dawn ambience and the ember's returning red glow support a subdued, intimate image with gently preserved facial detail.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): A dying ember glows inside the hearth in the dim dawn; the carrying frame is in a corner. 안규홍: He crouches sleepily before the hearth, with a cold-rice bundle tucked at his waist.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 안규홍 (한국인 남성, 청년 나이대, 검은 머리, 단단히 묶은 머리) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.",
   "B": "Create ONE FINAL photorealistic live-action film still of the moment below — early 20th-century Korea, Jeolla region; all people are Korean unless explicitly stated otherwise. TIME OF DAY (lock): dawn.\n\nSHOT TEXT (authoritative, Korean): 어스름한 새벽, 아궁이 불씨를 향해 입술을 모으고 입김을 뿜어내는 안규홍의 쪼그려 앉은 옆얼굴.\n\nLOCATION (lock): The hearth mouth inside a modest home's kitchen at dawn, with a crouching space directly before the rekindling red embers. The shot takes place here — the attached LOCATION PHOTOGRAPH shows the exact spot.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: close-up\n- KEY BACKGROUND ELEMENTS: 아궁이 입구 (Containing a fading ember that brightens as 안규홍 blows) — The opening is seen obliquely beside his profile, ahead of and below his mouth; used as Occupies less than a third of the frame and supplies the visible destination of his breath.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Dim dawn ambience and the ember's returning red glow support a subdued, intimate image with gently preserved facial detail.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): A dying ember glows inside the hearth in the dim dawn; the carrying frame is in a corner. 안규홍: He crouches sleepily before the hearth, with a cold-rice bundle tucked at his waist.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 안규홍 (한국인 남성, 청년 나이대, 검은 머리, 단단히 묶은 머리) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.\n\nCOMPOSITION VARIATION (this candidate only): keep the place, the people and the moment exactly as specified, and keep from the CAMERA & FRAME contract its subject scale, screen placement and key background elements exactly. For this candidate only, its camera angle and height are deliberately left open: choose a camera position distinctly different from the obvious one, as a film director picking a second setup on the same blocking. Stay true to any explicit framing the SHOT TEXT itself dictates."
  },
  "roll_refs": {
   "A": [
    {
     "label": "SHOT BACKGROUND",
     "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/db00517b-165b-4005-a287-f28936b59b95/images/272bd7af-9294-4490-98c6-88a40e9deacc/scene/recipe/S2sh1__bgfirst_bg.png",
     "asset_id": "d3f51f81-7354-42de-ab2f-a6e421bb5630",
     "role": "bgfirst_bg"
    },
    {
     "label": "LAYOUT SKETCH (people placement only)",
     "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/db00517b-165b-4005-a287-f28936b59b95/images/272bd7af-9294-4490-98c6-88a40e9deacc/conti/conti_S2sh1.png",
     "asset_id": "d94541fe-9ddd-41d9-8586-47426b75a334",
     "role": "conti_light"
    },
    {
     "label": "CHARACTER REFERENCE — 안규홍: the exact person appearing in this shot; match face, hair and build exactly.",
     "path": "<bytes:1028384>",
     "asset_id": "3b5894db-ce55-4f34-b966-5e725c3aa952",
     "role": "character_ref"
    }
   ],
   "B": [
    {
     "label": "LOCATION PHOTOGRAPH — the exact place of this shot: its architecture, materials, fixed features and lighting mood are spatial truth; stage the moment inside this place. Never copy its camera framing.",
     "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/db00517b-165b-4005-a287-f28936b59b95/episodes/272bd7af-9294-4490-98c6-88a40e9deacc/images/background_chain/L03B02.png",
     "asset_id": "07dcde76-24fe-4940-acad-6a24344f1cbb",
     "role": "location_plate"
    },
    {
     "label": "CHARACTER REFERENCE — 안규홍: the exact person appearing in this shot; match face, hair and build exactly.",
     "path": "<bytes:1028384>",
     "asset_id": "3b5894db-ce55-4f34-b966-5e725c3aa952",
     "role": "character_ref"
    }
   ]
  },
  "shot_run_spend_attempt_count": 1,
  "judge_flip_skipped": "judge_fn_owns_order",
  "readings": [
   {
    "label": "A",
    "direction": "우측 아궁이 쪽으로 시선과 얼굴 방향을 두고 입술을 모으고 있으나, 입김은 시각적으로 보이지 않음.",
    "built_space": "아궁이와 외부로 통하는 문의 위치 및 형태가 레퍼런스 공간을 잘 반영하여 구현됨.",
    "entities": "안규홍의 특징과 복장을 잘 반영하였으나, 요구된 소품인 식은 밥 덩어리가 묘사되지 않음.",
    "hard_violations": [
     "[gpt-high] 참조의 왼쪽 부뚜막으로 보이는 미장 구조를 유지하면서 출입구 앞 오른쪽에 별도의 부뚜막을 추가해, 하나여야 할 고정 설비를 중복 배치했다."
    ],
    "physics": "바닥에 몸을 낮추고 있으나 쪼그려 앉았다기보다는 엉덩이를 바닥에 댄 것에 가까운 자세로 지지됨."
   },
   {
    "label": "B",
    "direction": "시선과 모은 입술이 우측의 아궁이 불씨를 향하고 있으며, 뿜어져 나오는 하얀 입김이 아궁이 쪽으로 명확히 향함.",
    "built_space": "레퍼런스의 아궁이 질감과 형태가 우측에 배치되어 있고 인물은 그 앞의 바닥에 위치함.",
    "entities": "안규홍의 외모, 헤어스타일, 복장 레퍼런스와 일치하는 남성이 등장하나, 허리에 명시된 식은 밥 덩어리는 보이지 않음.",
    "hard_violations": [],
    "physics": "두 발로 바닥을 디디고 한 손을 무릎에 얹어 안정적으로 쪼그려 앉은 자세가 물리적으로 자연스러움."
   }
  ],
  "cross_model_order": {
   "policy": "cross_model_order_v2_slot0_fwd_slot1_rev_parallel2",
   "slots": [
    {
     "model": "gemini-pro",
     "order": "forward",
     "display_to_canonical": {
      "A": "A",
      "B": "B"
     },
     "raw": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 8,
        "verdict_ko": "인물의 쪼그려 앉은 자세와 아궁이를 향해 뿜어내는 입김을 매우 사실적으로 구현하였으나, 허리에 찬 식은 밥 덩어리가 누락되었습니다."
       },
       {
        "label": "A",
        "score": 6,
        "verdict_ko": "레퍼런스의 공간과 인물을 잘 반영했으나, 프롬프트의 핵심인 '입김'이 시각적으로 표현되지 않았고 요구된 소품(밥 덩어리)이 없습니다."
       }
      ],
      "readings": [
       {
        "label": "B",
        "direction": "시선과 모은 입술이 우측의 아궁이 불씨를 향하고 있으며, 뿜어져 나오는 하얀 입김이 아궁이 쪽으로 명확히 향함.",
        "built_space": "레퍼런스의 아궁이 질감과 형태가 우측에 배치되어 있고 인물은 그 앞의 바닥에 위치함.",
        "entities": "안규홍의 외모, 헤어스타일, 복장 레퍼런스와 일치하는 남성이 등장하나, 허리에 명시된 식은 밥 덩어리는 보이지 않음.",
        "hard_violations": [],
        "physics": "두 발로 바닥을 디디고 한 손을 무릎에 얹어 안정적으로 쪼그려 앉은 자세가 물리적으로 자연스러움."
       },
       {
        "label": "A",
        "direction": "우측 아궁이 쪽으로 시선과 얼굴 방향을 두고 입술을 모으고 있으나, 입김은 시각적으로 보이지 않음.",
        "built_space": "아궁이와 외부로 통하는 문의 위치 및 형태가 레퍼런스 공간을 잘 반영하여 구현됨.",
        "entities": "안규홍의 특징과 복장을 잘 반영하였으나, 요구된 소품인 식은 밥 덩어리가 묘사되지 않음.",
        "hard_violations": [],
        "physics": "바닥에 몸을 낮추고 있으나 쪼그려 앉았다기보다는 엉덩이를 바닥에 댄 것에 가까운 자세로 지지됨."
       }
      ],
      "all_candidates_fail": false
     },
     "normalized": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 8,
        "verdict_ko": "인물의 쪼그려 앉은 자세와 아궁이를 향해 뿜어내는 입김을 매우 사실적으로 구현하였으나, 허리에 찬 식은 밥 덩어리가 누락되었습니다."
       },
       {
        "label": "A",
        "score": 6,
        "verdict_ko": "레퍼런스의 공간과 인물을 잘 반영했으나, 프롬프트의 핵심인 '입김'이 시각적으로 표현되지 않았고 요구된 소품(밥 덩어리)이 없습니다."
       }
      ],
      "all_candidates_fail": false,
      "readings": [
       {
        "label": "B",
        "direction": "시선과 모은 입술이 우측의 아궁이 불씨를 향하고 있으며, 뿜어져 나오는 하얀 입김이 아궁이 쪽으로 명확히 향함.",
        "built_space": "레퍼런스의 아궁이 질감과 형태가 우측에 배치되어 있고 인물은 그 앞의 바닥에 위치함.",
        "entities": "안규홍의 외모, 헤어스타일, 복장 레퍼런스와 일치하는 남성이 등장하나, 허리에 명시된 식은 밥 덩어리는 보이지 않음.",
        "hard_violations": [],
        "physics": "두 발로 바닥을 디디고 한 손을 무릎에 얹어 안정적으로 쪼그려 앉은 자세가 물리적으로 자연스러움."
       },
       {
        "label": "A",
        "direction": "우측 아궁이 쪽으로 시선과 얼굴 방향을 두고 입술을 모으고 있으나, 입김은 시각적으로 보이지 않음.",
        "built_space": "아궁이와 외부로 통하는 문의 위치 및 형태가 레퍼런스 공간을 잘 반영하여 구현됨.",
        "entities": "안규홍의 특징과 복장을 잘 반영하였으나, 요구된 소품인 식은 밥 덩어리가 묘사되지 않음.",
        "hard_violations": [],
        "physics": "바닥에 몸을 낮추고 있으나 쪼그려 앉았다기보다는 엉덩이를 바닥에 댄 것에 가까운 자세로 지지됨."
       }
      ]
     },
     "ok": true
    },
    {
     "model": "gpt-high",
     "order": "reverse",
     "display_to_canonical": {
      "A": "B",
      "B": "A"
     },
     "raw": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 8,
        "verdict_ko": "불씨 쪽으로 내려가는 입김과 크게 잡힌 옆얼굴이 핵심 순간에 충실하며, 주변 그릇 배치와 다소 강한 불빛은 참조와 차이가 있다."
       },
       {
        "label": "B",
        "score": 3,
        "verdict_ko": "기존 부뚜막처럼 보이는 왼쪽 구조를 남긴 채 출입구 앞에 부뚜막을 하나 더 배치해 장소의 고정 구조를 위반하며, 입김의 목표 방향도 A보다 불명확하다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "남자는 오른쪽 아래의 아궁이를 내려다보며 입술을 모았다. 입에서 나온 가느다란 입김이 오른쪽 아래로 뻗어 아궁이 입구와 내부 불씨 방향으로 이어진다. 카메라를 보지 않으며, 입김의 목적지가 화면 안에 보인다.",
        "built_space": "오른쪽에 흙으로 만든 부뚜막 한 몸체와 사각 아궁이 입구 하나가 보인다. 입구는 얼굴 앞쪽 아래에 비스듬히 놓이고 화면의 3분의 1보다 작다. 남자는 입구 바로 앞 왼쪽에서 몸을 낮추고 있다. 왼쪽 뒤에는 외부로 열린 문이 있으며, 부뚜막 위에는 검은 그릇 두 개와 오른쪽 끝에 잘린 용기 일부, 바닥에는 검은 용기들과 장작이 보인다. 흙벽과 그을린 미장 표면은 참조와 유사하지만 그릇 배치까지 동일하지는 않다. 불가능한 반사나 중복 아궁이는 보이지 않는다.",
        "entities": "보이는 사람은 청년 한국인 남성으로 설정된 안규홍 한 명이며, 검은 머리를 뒤로 단단히 묶었다. 얼굴 윤곽과 코, 귀가 참조 인물과 대체로 부합하고, 때 묻은 회백색 전통 상의와 바지도 맞는다. 아궁이 안에는 붉은 숯불과 작은 불꽃이 있다. 허리는 일부 보이지만 찬밥 보따리라고 식별할 물건은 없고, 구석의 지게는 이 구도에서 확인되지 않는다. 읽을 수 있는 글자는 없다.",
        "hard_violations": [],
        "physics": "상체를 앞으로 기울이고 무릎을 깊게 접은 자세이며, 한 손은 올라온 무릎 위에 놓여 있다. 발의 접지는 화면 밖이지만 보이는 관절과 몸의 연결은 자연스럽고 부유를 시사하지 않는다. 그릇은 부뚜막이나 바닥에 놓이고 장작과 불씨는 아궁이 바닥에 받쳐져 있다. 입김은 입에서 시작하며, 불씨를 불어 살리는 동작으로 설명 가능한 자세다."
       },
       {
        "label": "B",
        "direction": "남자의 눈은 오른쪽 아래 아궁이 쪽을 향하고 입술은 오므려져 있다. 다만 입술의 축은 비교적 수평이며, 불씨는 훨씬 오른쪽 아래에 있다. 입김의 흐름도 보이지 않아 실제로 불씨까지 겨냥해 부는지는 A만큼 명확하지 않다.",
        "built_space": "왼쪽 뒤에는 참조의 부뚜막처럼 보이는 상판과 둥근 흙 미장 몸체가 남아 있고, 오른쪽 출입구 앞에는 아궁이 입구 하나와 솥 하나를 갖춘 별도의 부뚜막 몸체가 있다. 즉 입구는 하나만 보이지만 부뚜막 덩어리는 두 곳에 배치되어 있다. 참조에서는 부뚜막이 문 왼쪽에 하나뿐인데, 여기서는 추가 구조가 문 앞을 차지한다. 문틀, 바구니 하나, 작은 그릇 하나, 바닥 돌 하나와 바깥 돌담은 참조를 닮았다. 얼굴 옆의 입구 크기는 3분의 1 미만이지만, 얼굴보다 주변 공간과 무릎에 배분한 화면이 A보다 넓다.",
        "entities": "청년 한국인 남성으로 설정된 안규홍 한 명이 보이며, 묶은 검은 머리와 회백색의 낡은 전통 의상은 참조와 대체로 맞는다. 얼굴은 자연스러운 사람의 옆얼굴이고 눈에도 비정상적인 표현이 없다. 솥과 뚜껑, 붉은 불씨가 식별되지만 불꽃은 사그라진 불씨만 있는 상태보다 강하다. 찬밥 보따리는 식별되지 않고 지게가 놓일 구석은 충분히 드러나지 않는다. 읽을 수 있는 글자는 없다.",
        "hard_violations": [
         "참조의 왼쪽 부뚜막으로 보이는 미장 구조를 유지하면서 출입구 앞 오른쪽에 별도의 부뚜막을 추가해, 하나여야 할 고정 설비를 중복 배치했다."
        ],
        "physics": "남자는 무릎을 접고 상체를 앞으로 숙였으며, 하체 일부가 화면 아래로 잘린다. 발은 보이지 않지만 몸이 공중에 떠 있다는 증거는 없고, 보이는 자세는 웅크려 앉는 동작으로 가능하다. 솥은 오른쪽 부뚜막 상판에 받쳐져 있고, 바구니와 돌은 바닥에 놓여 있으며, 불씨와 장작은 아궁이 안쪽 바닥에 놓여 있다. 물체의 무지지 부유나 불가능한 관절은 보이지 않는다."
       }
      ],
      "all_candidates_fail": false
     },
     "normalized": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 8,
        "verdict_ko": "불씨 쪽으로 내려가는 입김과 크게 잡힌 옆얼굴이 핵심 순간에 충실하며, 주변 그릇 배치와 다소 강한 불빛은 참조와 차이가 있다."
       },
       {
        "label": "A",
        "score": 3,
        "verdict_ko": "기존 부뚜막처럼 보이는 왼쪽 구조를 남긴 채 출입구 앞에 부뚜막을 하나 더 배치해 장소의 고정 구조를 위반하며, 입김의 목표 방향도 A보다 불명확하다."
       }
      ],
      "all_candidates_fail": false,
      "readings": [
       {
        "label": "B",
        "direction": "남자는 오른쪽 아래의 아궁이를 내려다보며 입술을 모았다. 입에서 나온 가느다란 입김이 오른쪽 아래로 뻗어 아궁이 입구와 내부 불씨 방향으로 이어진다. 카메라를 보지 않으며, 입김의 목적지가 화면 안에 보인다.",
        "built_space": "오른쪽에 흙으로 만든 부뚜막 한 몸체와 사각 아궁이 입구 하나가 보인다. 입구는 얼굴 앞쪽 아래에 비스듬히 놓이고 화면의 3분의 1보다 작다. 남자는 입구 바로 앞 왼쪽에서 몸을 낮추고 있다. 왼쪽 뒤에는 외부로 열린 문이 있으며, 부뚜막 위에는 검은 그릇 두 개와 오른쪽 끝에 잘린 용기 일부, 바닥에는 검은 용기들과 장작이 보인다. 흙벽과 그을린 미장 표면은 참조와 유사하지만 그릇 배치까지 동일하지는 않다. 불가능한 반사나 중복 아궁이는 보이지 않는다.",
        "entities": "보이는 사람은 청년 한국인 남성으로 설정된 안규홍 한 명이며, 검은 머리를 뒤로 단단히 묶었다. 얼굴 윤곽과 코, 귀가 참조 인물과 대체로 부합하고, 때 묻은 회백색 전통 상의와 바지도 맞는다. 아궁이 안에는 붉은 숯불과 작은 불꽃이 있다. 허리는 일부 보이지만 찬밥 보따리라고 식별할 물건은 없고, 구석의 지게는 이 구도에서 확인되지 않는다. 읽을 수 있는 글자는 없다.",
        "hard_violations": [],
        "physics": "상체를 앞으로 기울이고 무릎을 깊게 접은 자세이며, 한 손은 올라온 무릎 위에 놓여 있다. 발의 접지는 화면 밖이지만 보이는 관절과 몸의 연결은 자연스럽고 부유를 시사하지 않는다. 그릇은 부뚜막이나 바닥에 놓이고 장작과 불씨는 아궁이 바닥에 받쳐져 있다. 입김은 입에서 시작하며, 불씨를 불어 살리는 동작으로 설명 가능한 자세다."
       },
       {
        "label": "A",
        "direction": "남자의 눈은 오른쪽 아래 아궁이 쪽을 향하고 입술은 오므려져 있다. 다만 입술의 축은 비교적 수평이며, 불씨는 훨씬 오른쪽 아래에 있다. 입김의 흐름도 보이지 않아 실제로 불씨까지 겨냥해 부는지는 A만큼 명확하지 않다.",
        "built_space": "왼쪽 뒤에는 참조의 부뚜막처럼 보이는 상판과 둥근 흙 미장 몸체가 남아 있고, 오른쪽 출입구 앞에는 아궁이 입구 하나와 솥 하나를 갖춘 별도의 부뚜막 몸체가 있다. 즉 입구는 하나만 보이지만 부뚜막 덩어리는 두 곳에 배치되어 있다. 참조에서는 부뚜막이 문 왼쪽에 하나뿐인데, 여기서는 추가 구조가 문 앞을 차지한다. 문틀, 바구니 하나, 작은 그릇 하나, 바닥 돌 하나와 바깥 돌담은 참조를 닮았다. 얼굴 옆의 입구 크기는 3분의 1 미만이지만, 얼굴보다 주변 공간과 무릎에 배분한 화면이 A보다 넓다.",
        "entities": "청년 한국인 남성으로 설정된 안규홍 한 명이 보이며, 묶은 검은 머리와 회백색의 낡은 전통 의상은 참조와 대체로 맞는다. 얼굴은 자연스러운 사람의 옆얼굴이고 눈에도 비정상적인 표현이 없다. 솥과 뚜껑, 붉은 불씨가 식별되지만 불꽃은 사그라진 불씨만 있는 상태보다 강하다. 찬밥 보따리는 식별되지 않고 지게가 놓일 구석은 충분히 드러나지 않는다. 읽을 수 있는 글자는 없다.",
        "hard_violations": [
         "참조의 왼쪽 부뚜막으로 보이는 미장 구조를 유지하면서 출입구 앞 오른쪽에 별도의 부뚜막을 추가해, 하나여야 할 고정 설비를 중복 배치했다."
        ],
        "physics": "남자는 무릎을 접고 상체를 앞으로 숙였으며, 하체 일부가 화면 아래로 잘린다. 발은 보이지 않지만 몸이 공중에 떠 있다는 증거는 없고, 보이는 자세는 웅크려 앉는 동작으로 가능하다. 솥은 오른쪽 부뚜막 상판에 받쳐져 있고, 바구니와 돌은 바닥에 놓여 있으며, 불씨와 장작은 아궁이 안쪽 바닥에 놓여 있다. 물체의 무지지 부유나 불가능한 관절은 보이지 않는다."
       }
      ]
     },
     "ok": true
    }
   ],
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "slot_winner_match": true,
   "slot_winner": {
    "gemini-pro": "B",
    "gpt-high": "B"
   },
   "route": "cross_slot_agree"
  },
  "dual": {
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "normalized": {
    "A": 1.125,
    "B": 2.0
   },
   "adjusted": {
    "A": 0.875,
    "B": 2.0
   },
   "violations": {
    "A": [
     "[gpt-high] 참조의 왼쪽 부뚜막으로 보이는 미장 구조를 유지하면서 출입구 앞 오른쪽에 별도의 부뚜막을 추가해, 하나여야 할 고정 설비를 중복 배치했다."
    ]
   },
   "per_model_winner": {
    "gemini-pro": "B",
    "gpt-high": "B"
   },
   "agreed": true
  },
  "totals": {
   "B": 2000,
   "A": 875
  },
  "selected": "B",
  "ranking": [
   "B",
   "A"
  ],
  "verdicts": [
   {
    "label": "B",
    "score": 2000,
    "verdict_ko": "인물의 쪼그려 앉은 자세와 아궁이를 향해 뿜어내는 입김을 매우 사실적으로 구현하였으나, 허리에 찬 식은 밥 덩어리가 누락되었습니다."
   },
   {
    "label": "A",
    "score": 875,
    "verdict_ko": "레퍼런스의 공간과 인물을 잘 반영했으나, 프롬프트의 핵심인 '입김'이 시각적으로 표현되지 않았고 요구된 소품(밥 덩어리)이 없습니다.  ★위반: [gpt-high] 참조의 왼쪽 부뚜막으로 보이는 미장 구조를 유지하면서 출입구 앞 오른쪽에 별도의 부뚜막을 추가해, 하나여야 할 고정 설비를 중복 배치했다."
   }
  ],
  "refs": [
   {
    "label": "LOCATION PHOTOGRAPH — the exact place of this shot: its architecture, materials, fixed features and lighting mood are spatial truth; stage the moment inside this place. Never copy its camera framing.",
    "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/db00517b-165b-4005-a287-f28936b59b95/episodes/272bd7af-9294-4490-98c6-88a40e9deacc/images/background_chain/L03B02.png",
    "asset_id": "07dcde76-24fe-4940-acad-6a24344f1cbb",
    "role": "location_plate"
   },
   {
    "label": "CHARACTER REFERENCE — 안규홍: the exact person appearing in this shot; match face, hair and build exactly.",
    "path": "<bytes:1028384>",
    "asset_id": "3b5894db-ce55-4f34-b966-5e725c3aa952",
    "role": "character_ref"
   }
  ],
  "critique_skipped": true,
  "shot_run_produced": false,
  "shot_run_uid": "06aa1401-5906-779c-b469-73fa7771807f",
  "bgfirst": {
   "bg_path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/db00517b-165b-4005-a287-f28936b59b95/images/272bd7af-9294-4490-98c6-88a40e9deacc/scene/recipe/S2sh1__bgfirst_bg.png",
   "bg_asset_id": "d3f51f81-7354-42de-ab2f-a6e421bb5630",
   "bg_record_key": "S2sh1::bgfirst_bg",
   "chain_winner": false,
   "authority": "plate"
  },
  "ref_mode": "플레이트+엔티티 (2택1: 무콘티 승)",
  "share_plan": {
   "ref_plan": "background"
  }
 },
 "S2sh1::cine": {
  "applied": true,
  "attempted_at": "2026-09-09T00:41:03.394064+00:00",
  "fingerprint": "50e4f932368c4306c546303c189945cf40432fd6795ea6ddf36fae64f4d83e99",
  "fingerprint_version": 2,
  "provider": "mai",
  "endpoint": "openrouter/chat-completions",
  "model": "microsoft/mai-image-2.6",
  "pack": "24.202608252115",
  "source_file": "S2sh1_sel.png",
  "source_sha256": "37bf42f09ae4b5cce57e56b79f6a8af35a54b0a4a0fe6ae297a7638a52ba4a7e",
  "file": "S2sh1_cine.png",
  "staged_sha256": "5867d9c804f87ac24eb80ab87a602989d19a9589788f67b12c8a22afc87ba61f",
  "latency_ms": 28812
 },
 "S2sh13::signage": {
  "fp": "2818baadc5471cc8",
  "inscriptions": [],
  "cues": [],
  "dropped": []
 },
 "S2sh13::bgfirst_bg": {
  "input_fingerprint": "a95bbb4a83ee0fd3",
  "prompt": "Create the EMPTY BACKGROUND PLATE for one film shot — NO PEOPLE, no figures, no body parts, no sketch lines, no arrows anywhere.\n\"Empty\" means no people only: KEEP the location's inherent occupants and stock that define the place — animals in an animal shelter, pen or farm, goods in a market, moored boats in a harbour — unless the shot text explicitly removes them.\nThe FIRST attached image is a thin-line storyboard sketch: use ONLY its camera angle, horizon, perspective and the placement/size of buildings and set masses — ignore the sketched people and arrows entirely. The SECOND attached image (LOCATION PHOTOGRAPH) is the real place: take its architecture, materials, signage and fixed features, and RE-PROJECT them into the sketch's camera. If the photograph's camera differs from the sketch's, the sketch's camera wins.\nHUMAN-SCALE CALIBRATION: derive every structure's true size from human-scale elements — a door ≈ 2m, a window ≈ 1–1.5m wide, one storey ≈ 2.5–3m; never inflate a small structure or shrink a large one.\n\nSHOT TEXT this background must serve (Korean): 문틈에 얼굴만 내민 채 능청스럽게 눈웃음을 치는 안규홍의 얼굴 클로즈업.\n\nLOCATION (lock): The gap in a reopened rustic wicket gate at a modest home's outdoor entrance in the early dawn.\n\nTIME OF DAY (lock): dawn.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: close-up\n- FRAME LAYOUT: gate edge beside the protruding face in the middle-right of the frame, foreground.\n- KEY BACKGROUND ELEMENTS: 사립문 (Reopened, with 안규홍's face protruding through the opening) — An oblique inner-facing edge is visible beside his face; the gate conceals his body; used as A narrow lateral boundary makes the playful reappearance legible without letting the gate outweigh the face.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Keep the gate-side dawn ambience subdued and naturally modeled, with muted saturation and no added hearth-colored illumination.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nRender ONE photorealistic empty location photograph, 16:9, that this shot can be staged inside later. No readable writing anywhere: surfaces that would carry writing may be present, but stage any wording out of legibility — an oblique angle, distance, shallow focus. No captions, watermarks or overlay text.",
  "effective_prompt": "Create the EMPTY BACKGROUND PLATE for one film shot — NO PEOPLE, no figures, no body parts, no sketch lines, no arrows anywhere.\n\"Empty\" means no people only: KEEP the location's inherent occupants and stock that define the place — animals in an animal shelter, pen or farm, goods in a market, moored boats in a harbour — unless the shot text explicitly removes them.\nThe FIRST attached image is a thin-line storyboard sketch: use ONLY its camera angle, horizon, perspective and the placement/size of buildings and set masses — ignore the sketched people and arrows entirely. The SECOND attached image (LOCATION PHOTOGRAPH) is the real place: take its architecture, materials, signage and fixed features, and RE-PROJECT them into the sketch's camera. If the photograph's camera differs from the sketch's, the sketch's camera wins.\nHUMAN-SCALE CALIBRATION: derive every structure's true size from human-scale elements — a door ≈ 2m, a window ≈ 1–1.5m wide, one storey ≈ 2.5–3m; never inflate a small structure or shrink a large one.\n\nSHOT TEXT this background must serve (Korean): 문틈에 얼굴만 내민 채 능청스럽게 눈웃음을 치는 안규홍의 얼굴 클로즈업.\n\nLOCATION (lock): The gap in a reopened rustic wicket gate at a modest home's outdoor entrance in the early dawn.\n\nTIME OF DAY (lock): dawn.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: close-up\n- FRAME LAYOUT: gate edge beside the protruding face in the middle-right of the frame, foreground.\n- KEY BACKGROUND ELEMENTS: 사립문 (Reopened, with 안규홍's face protruding through the opening) — An oblique inner-facing edge is visible beside his face; the gate conceals his body; used as A narrow lateral boundary makes the playful reappearance legible without letting the gate outweigh the face.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Keep the gate-side dawn ambience subdued and naturally modeled, with muted saturation and no added hearth-colored illumination.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nRender ONE photorealistic empty location photograph, 16:9, that this shot can be staged inside later. No readable writing anywhere: surfaces that would carry writing may be present, but stage any wording out of legibility — an oblique angle, distance, shallow focus. No captions, watermarks or overlay text.",
  "bg_path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/db00517b-165b-4005-a287-f28936b59b95/images/272bd7af-9294-4490-98c6-88a40e9deacc/scene/recipe/S2sh13__bgfirst_bg.png",
  "asset_id": "b1d60c72-e8d0-4095-b880-092f1128a9f9",
  "input_asset_ids": [
   "23016f46-9848-4b3b-ae34-c239ed6abd95",
   "8263e5ec-d509-48ee-bca3-977ed63622f2"
  ]
 },
 "S2sh13": {
  "input_fingerprint": "48602c65999068c2",
  "prompt": "Create ONE FINAL photorealistic live-action film still of the moment below — early 20th-century Korea, Jeolla region; all people are Korean unless explicitly stated otherwise. TIME OF DAY (lock): dawn.\n\nSHOT TEXT (authoritative, Korean): 문틈에 얼굴만 내민 채 능청스럽게 눈웃음을 치는 안규홍의 얼굴 클로즈업.\n\nLOCATION (lock): The gap in a reopened rustic wicket gate at a modest home's outdoor entrance in the early dawn. The shot takes place here — the attached LOCATION PHOTOGRAPH shows the exact spot.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: close-up\n- FRAME LAYOUT: gate edge beside the protruding face in the middle-right of the frame, foreground.\n- KEY BACKGROUND ELEMENTS: 사립문 (Reopened, with 안규홍's face protruding through the opening) — An oblique inner-facing edge is visible beside his face; the gate conceals his body; used as A narrow lateral boundary makes the playful reappearance legible without letting the gate outweigh the face.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Keep the gate-side dawn ambience subdued and naturally modeled, with muted saturation and no added hearth-colored illumination.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The gate has reopened; the lidded medicine pot contains all the herbs, and an added log burns in the hearth. 안규홍: He is outside with his carrying frame and the rice bundle tucked back at his waist, leaning his face through the gate.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 안규홍 (한국인 남성, 청년 나이대, 검은 머리, 단단히 묶은 머리) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.",
  "roll_prompts": {
   "A": "Stage the shot. The FIRST attached image (SHOT BACKGROUND) is the finished empty background of this shot — keep it EXACTLY: its camera, perspective, architecture, lighting and every fixed feature stay untouched. The SECOND attached image (LAYOUT SKETCH) tells you ONLY where the people go: each sketched person's position, screen size, pose and the gaze/motion arrows. Ignore the sketch's background lines. The CHARACTER REFERENCE photographs show the real people.\nPlace the real people into the background at exactly the sketched positions, sizes and poses, following the arrow directions. No sketch lines or arrows may remain.\n\nCreate ONE FINAL photorealistic live-action film still of the moment below — early 20th-century Korea, Jeolla region; all people are Korean unless explicitly stated otherwise. TIME OF DAY (lock): dawn.\n\nSHOT TEXT (authoritative, Korean): 문틈에 얼굴만 내민 채 능청스럽게 눈웃음을 치는 안규홍의 얼굴 클로즈업.\n\nLOCATION (lock): The gap in a reopened rustic wicket gate at a modest home's outdoor entrance in the early dawn. The shot takes place here — the FIRST attached image (SHOT BACKGROUND) is this exact place, already built: its ground, structures, horizon, materials and lighting are the finished truth of this location and must not be redesigned or replaced. No location photograph is attached — read the place from that image alone, and add no scenery, structure, vehicle or fixture that it does not already show. This lock governs the place only; the figures in the shot follow the staging and pose instructions.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: close-up\n- FRAME LAYOUT: gate edge beside the protruding face in the middle-right of the frame, foreground.\n- KEY BACKGROUND ELEMENTS: 사립문 (Reopened, with 안규홍's face protruding through the opening) — An oblique inner-facing edge is visible beside his face; the gate conceals his body; used as A narrow lateral boundary makes the playful reappearance legible without letting the gate outweigh the face.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Keep the gate-side dawn ambience subdued and naturally modeled, with muted saturation and no added hearth-colored illumination.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The gate has reopened; the lidded medicine pot contains all the herbs, and an added log burns in the hearth. 안규홍: He is outside with his carrying frame and the rice bundle tucked back at his waist, leaning his face through the gate.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 안규홍 (한국인 남성, 청년 나이대, 검은 머리, 단단히 묶은 머리) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.",
   "B": "Create ONE FINAL photorealistic live-action film still of the moment below — early 20th-century Korea, Jeolla region; all people are Korean unless explicitly stated otherwise. TIME OF DAY (lock): dawn.\n\nSHOT TEXT (authoritative, Korean): 문틈에 얼굴만 내민 채 능청스럽게 눈웃음을 치는 안규홍의 얼굴 클로즈업.\n\nLOCATION (lock): The gap in a reopened rustic wicket gate at a modest home's outdoor entrance in the early dawn. The shot takes place here — the attached LOCATION PHOTOGRAPH shows the exact spot.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: close-up\n- FRAME LAYOUT: gate edge beside the protruding face in the middle-right of the frame, foreground.\n- KEY BACKGROUND ELEMENTS: 사립문 (Reopened, with 안규홍's face protruding through the opening) — An oblique inner-facing edge is visible beside his face; the gate conceals his body; used as A narrow lateral boundary makes the playful reappearance legible without letting the gate outweigh the face.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Keep the gate-side dawn ambience subdued and naturally modeled, with muted saturation and no added hearth-colored illumination.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The gate has reopened; the lidded medicine pot contains all the herbs, and an added log burns in the hearth. 안규홍: He is outside with his carrying frame and the rice bundle tucked back at his waist, leaning his face through the gate.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 안규홍 (한국인 남성, 청년 나이대, 검은 머리, 단단히 묶은 머리) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.\n\nCOMPOSITION VARIATION (this candidate only): keep the place, the people and the moment exactly as specified, and keep from the CAMERA & FRAME contract its subject scale, screen placement and key background elements exactly. For this candidate only, its camera angle and height are deliberately left open: choose a camera position distinctly different from the obvious one, as a film director picking a second setup on the same blocking. Stay true to any explicit framing the SHOT TEXT itself dictates."
  },
  "roll_refs": {
   "A": [
    {
     "label": "SHOT BACKGROUND",
     "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/db00517b-165b-4005-a287-f28936b59b95/images/272bd7af-9294-4490-98c6-88a40e9deacc/scene/recipe/S2sh13__bgfirst_bg.png",
     "asset_id": "b1d60c72-e8d0-4095-b880-092f1128a9f9",
     "role": "bgfirst_bg"
    },
    {
     "label": "LAYOUT SKETCH (people placement only)",
     "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/db00517b-165b-4005-a287-f28936b59b95/images/272bd7af-9294-4490-98c6-88a40e9deacc/conti/conti_S2sh13.png",
     "asset_id": "23016f46-9848-4b3b-ae34-c239ed6abd95",
     "role": "conti_light"
    },
    {
     "label": "CHARACTER REFERENCE — 안규홍: the exact person appearing in this shot; match face, hair and build exactly.",
     "path": "<bytes:1028384>",
     "asset_id": "3b5894db-ce55-4f34-b966-5e725c3aa952",
     "role": "character_ref"
    }
   ],
   "B": [
    {
     "label": "LOCATION PHOTOGRAPH — the exact place of this shot: its architecture, materials, fixed features and lighting mood are spatial truth; stage the moment inside this place. Never copy its camera framing.",
     "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/db00517b-165b-4005-a287-f28936b59b95/episodes/272bd7af-9294-4490-98c6-88a40e9deacc/images/background_chain/L03B01.png",
     "asset_id": "8263e5ec-d509-48ee-bca3-977ed63622f2",
     "role": "location_plate"
    },
    {
     "label": "CHARACTER REFERENCE — 안규홍: the exact person appearing in this shot; match face, hair and build exactly.",
     "path": "<bytes:1028384>",
     "asset_id": "3b5894db-ce55-4f34-b966-5e725c3aa952",
     "role": "character_ref"
    }
   ]
  },
  "shot_run_spend_attempt_count": 1,
  "judge_flip_skipped": "judge_fn_owns_order",
  "readings": [
   {
    "label": "A",
    "direction": "안규홍의 시선은 카메라를 향해 똑바로 장난스럽게 던져집니다.",
    "built_space": "레퍼런스와 동일하게 왼쪽에는 섶나무 문짝, 오른쪽에는 두꺼운 나무 기둥이 배치되어 있으며, 그 사이의 틈을 통해 인물이 내다보고 있습니다.",
    "entities": "안규홍의 얼굴과 헤어스타일이 레퍼런스와 일치하며, 지시된 능청스러운 눈웃음을 잘 표현하고 있습니다.",
    "hard_violations": [],
    "physics": "몸은 오른쪽 문기둥 뒤에 가려져 보이지 않으나, 얼굴이 내밀어진 각도와 위치가 자연스럽게 지탱되고 있습니다."
   },
   {
    "label": "B",
    "direction": "안규홍의 시선은 카메라를 향하고 있습니다.",
    "built_space": "레퍼런스와 동일한 섶나무 문짝과 나무 기둥이 열려 있는 상태를 정확히 재현했습니다.",
    "entities": "안규홍의 외모는 레퍼런스와 일치하나, 지시와 달리 상반신(한복) 전체가 문틈으로 크게 노출되어 있습니다.",
    "hard_violations": [],
    "physics": "오른쪽 기둥을 잡고 있는 손과 몸의 자세는 물리적으로 자연스럽게 지탱되고 있습니다."
   }
  ],
  "cross_model_order": {
   "policy": "cross_model_order_v2_slot0_fwd_slot1_rev_parallel2",
   "slots": [
    {
     "model": "gemini-pro",
     "order": "forward",
     "display_to_canonical": {
      "A": "A",
      "B": "B"
     },
     "raw": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 9,
        "verdict_ko": "프롬프트가 요구한 '얼굴만 내민 채'와 '문이 몸을 가린다'는 지시사항을 완벽하게 따르며, 인물의 장난스러운 표정과 클로즈업 구도를 매우 충실하게 구현했습니다."
       },
       {
        "label": "B",
        "score": 5,
        "verdict_ko": "인물의 표정과 배경은 훌륭하나, 몸통과 팔이 문틈으로 훤히 드러나 '문이 몸을 가리고 얼굴만 내민다'는 핵심 지시사항을 위반했습니다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "안규홍의 시선은 카메라를 향해 똑바로 장난스럽게 던져집니다.",
        "built_space": "레퍼런스와 동일하게 왼쪽에는 섶나무 문짝, 오른쪽에는 두꺼운 나무 기둥이 배치되어 있으며, 그 사이의 틈을 통해 인물이 내다보고 있습니다.",
        "entities": "안규홍의 얼굴과 헤어스타일이 레퍼런스와 일치하며, 지시된 능청스러운 눈웃음을 잘 표현하고 있습니다.",
        "hard_violations": [],
        "physics": "몸은 오른쪽 문기둥 뒤에 가려져 보이지 않으나, 얼굴이 내밀어진 각도와 위치가 자연스럽게 지탱되고 있습니다."
       },
       {
        "label": "B",
        "direction": "안규홍의 시선은 카메라를 향하고 있습니다.",
        "built_space": "레퍼런스와 동일한 섶나무 문짝과 나무 기둥이 열려 있는 상태를 정확히 재현했습니다.",
        "entities": "안규홍의 외모는 레퍼런스와 일치하나, 지시와 달리 상반신(한복) 전체가 문틈으로 크게 노출되어 있습니다.",
        "hard_violations": [],
        "physics": "오른쪽 기둥을 잡고 있는 손과 몸의 자세는 물리적으로 자연스럽게 지탱되고 있습니다."
       }
      ],
      "all_candidates_fail": false
     },
     "normalized": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 9,
        "verdict_ko": "프롬프트가 요구한 '얼굴만 내민 채'와 '문이 몸을 가린다'는 지시사항을 완벽하게 따르며, 인물의 장난스러운 표정과 클로즈업 구도를 매우 충실하게 구현했습니다."
       },
       {
        "label": "B",
        "score": 5,
        "verdict_ko": "인물의 표정과 배경은 훌륭하나, 몸통과 팔이 문틈으로 훤히 드러나 '문이 몸을 가리고 얼굴만 내민다'는 핵심 지시사항을 위반했습니다."
       }
      ],
      "all_candidates_fail": false,
      "readings": [
       {
        "label": "A",
        "direction": "안규홍의 시선은 카메라를 향해 똑바로 장난스럽게 던져집니다.",
        "built_space": "레퍼런스와 동일하게 왼쪽에는 섶나무 문짝, 오른쪽에는 두꺼운 나무 기둥이 배치되어 있으며, 그 사이의 틈을 통해 인물이 내다보고 있습니다.",
        "entities": "안규홍의 얼굴과 헤어스타일이 레퍼런스와 일치하며, 지시된 능청스러운 눈웃음을 잘 표현하고 있습니다.",
        "hard_violations": [],
        "physics": "몸은 오른쪽 문기둥 뒤에 가려져 보이지 않으나, 얼굴이 내밀어진 각도와 위치가 자연스럽게 지탱되고 있습니다."
       },
       {
        "label": "B",
        "direction": "안규홍의 시선은 카메라를 향하고 있습니다.",
        "built_space": "레퍼런스와 동일한 섶나무 문짝과 나무 기둥이 열려 있는 상태를 정확히 재현했습니다.",
        "entities": "안규홍의 외모는 레퍼런스와 일치하나, 지시와 달리 상반신(한복) 전체가 문틈으로 크게 노출되어 있습니다.",
        "hard_violations": [],
        "physics": "오른쪽 기둥을 잡고 있는 손과 몸의 자세는 물리적으로 자연스럽게 지탱되고 있습니다."
       }
      ]
     },
     "ok": true
    },
    {
     "model": "gpt-high",
     "order": "reverse",
     "display_to_canonical": {
      "A": "B",
      "B": "A"
     },
     "raw": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 6,
        "verdict_ko": "능청스러운 눈웃음과 새벽 분위기는 맞지만, 얼굴이 작고 어깨·가슴·팔까지 드러나 ‘얼굴만 내민 얼굴 클로즈업’이라는 핵심 구도에서 벗어난다."
       },
       {
        "label": "B",
        "score": 9,
        "verdict_ko": "중앙 오른쪽에 얼굴을 크게 내밀고 몸은 문 뒤에 감춰, 문틈에서 능청스럽게 눈웃음 짓는 얼굴 클로즈업을 가장 충실히 구현한다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "고개를 화면 왼쪽으로 기울여 문틈 앞으로 내밀고, 시선은 카메라 가까운 문 안쪽을 향한다. 눈을 가늘게 접고 입꼬리를 올려 능청스러운 눈웃음을 표현한다. 별도로 겨누거나 이동하는 물체는 없다.",
        "built_space": "왼쪽 문짝에는 세로 나뭇살, 밧줄로 묶은 가로대 두 개와 아래쪽 넓은 판재 하나가 보이고, 오른쪽에는 홈에 박힌 나무 걸쇠 하나가 있다. 참고 장소의 기본 구성이 유지되지만 나뭇살 사이가 더 성기다. 열린 틈의 중앙 오른쪽에 인물이 있으나 얼굴뿐 아니라 상체와 굽힌 팔도 보인다. 양쪽 문이 화면 대부분을 차지해 얼굴 클로즈업보다 넓은 구도다. 반사면은 없다.",
        "entities": "한국인 청년 남성 설정에 부합하는 인물 한 명이며 검은 머리를 뒤로 단단히 넘겼다. 묶은 부분 자체는 가려져 있다. 참고 인물과 대체로 유사한 얼굴과 때 탄 밝은색 전통 상의가 보인다. 사립문의 목재와 밧줄은 실제 재료로 읽히며, 조명은 차갑고 절제된 새벽빛이다. 지게·허리의 쌀 꾸러미·약탕기·아궁이는 이 구도에서 확인되지 않는다. 읽을 수 있는 글자는 없다.",
        "hard_violations": [],
        "physics": "머리는 보이는 목과 상체에 자연스럽게 연결되고 상체가 문틈으로 기울어져 있다. 굽힌 팔과 문 가장자리 가까운 손 일부가 보인다. 발과 지면 접촉은 화면 밖이지만, 서서 허리를 기울이는 동작으로 가능한 자세이며 공중에 뜬 몸이나 지지 없는 물체는 보이지 않는다."
       },
       {
        "label": "B",
        "direction": "얼굴을 문 가장자리 너머 화면 왼쪽으로 내밀고, 눈은 카메라 가까운 문 안쪽을 향한다. 고개를 살짝 기울인 채 눈가를 접고 비스듬히 미소 지어 장난스럽고 능청스러운 재등장으로 읽힌다. 겨누는 도구나 이동 물체는 없다.",
        "built_space": "왼쪽에는 촘촘한 세로 나뭇살과 밧줄로 묶은 가로대 두 개, 아래쪽 넓은 판재 하나가 있고, 오른쪽 목재 면에는 나무 걸쇠 하나가 보인다. 참고 장소의 재료와 고정물 구성이 잘 유지된다. 얼굴은 중앙 오른쪽 전경에 크게 놓이고, 바로 옆 문 가장자리의 두께와 비스듬한 안쪽 면이 보인다. 몸은 오른쪽 문 뒤에 가려져 있으며 틈 너머에는 흙길과 흐릿한 담장·수목이 보인다. 반사면은 없다.",
        "entities": "인물은 한 명으로, 한국인 청년 남성 설정에 맞는 외모와 뒤로 단단히 넘긴 검은 머리를 지녔다. 참고 인물의 얼굴형과 주요 이목구비에 대체로 부합하며 눈은 정상적인 해부학적 형태다. 턱 아래 옷깃만 조금 보이므로 나머지 의상과 지게·쌀 꾸러미는 판단 대상이 아니다. 약탕기와 아궁이도 화면 밖이다. 목재의 마모와 피부 질감은 실물처럼 표현되고, 새벽빛에는 별도의 주황색 화염 조명이 없다. 읽을 수 있는 글자는 없다.",
        "hard_violations": [],
        "physics": "턱 아래 목과 옷깃 일부가 문 뒤로 이어져, 몸을 숨긴 채 목과 상체를 기울여 얼굴만 내미는 자세로 성립한다. 하체의 지지는 가려져 확인되지 않지만 머리가 분리되거나 떠 있는 것으로 보이지 않는다. 문짝과 걸쇠도 구조에 붙어 있으며 지지 없이 떠 있는 물체는 없다."
       }
      ],
      "all_candidates_fail": false
     },
     "normalized": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 6,
        "verdict_ko": "능청스러운 눈웃음과 새벽 분위기는 맞지만, 얼굴이 작고 어깨·가슴·팔까지 드러나 ‘얼굴만 내민 얼굴 클로즈업’이라는 핵심 구도에서 벗어난다."
       },
       {
        "label": "A",
        "score": 9,
        "verdict_ko": "중앙 오른쪽에 얼굴을 크게 내밀고 몸은 문 뒤에 감춰, 문틈에서 능청스럽게 눈웃음 짓는 얼굴 클로즈업을 가장 충실히 구현한다."
       }
      ],
      "all_candidates_fail": false,
      "readings": [
       {
        "label": "B",
        "direction": "고개를 화면 왼쪽으로 기울여 문틈 앞으로 내밀고, 시선은 카메라 가까운 문 안쪽을 향한다. 눈을 가늘게 접고 입꼬리를 올려 능청스러운 눈웃음을 표현한다. 별도로 겨누거나 이동하는 물체는 없다.",
        "built_space": "왼쪽 문짝에는 세로 나뭇살, 밧줄로 묶은 가로대 두 개와 아래쪽 넓은 판재 하나가 보이고, 오른쪽에는 홈에 박힌 나무 걸쇠 하나가 있다. 참고 장소의 기본 구성이 유지되지만 나뭇살 사이가 더 성기다. 열린 틈의 중앙 오른쪽에 인물이 있으나 얼굴뿐 아니라 상체와 굽힌 팔도 보인다. 양쪽 문이 화면 대부분을 차지해 얼굴 클로즈업보다 넓은 구도다. 반사면은 없다.",
        "entities": "한국인 청년 남성 설정에 부합하는 인물 한 명이며 검은 머리를 뒤로 단단히 넘겼다. 묶은 부분 자체는 가려져 있다. 참고 인물과 대체로 유사한 얼굴과 때 탄 밝은색 전통 상의가 보인다. 사립문의 목재와 밧줄은 실제 재료로 읽히며, 조명은 차갑고 절제된 새벽빛이다. 지게·허리의 쌀 꾸러미·약탕기·아궁이는 이 구도에서 확인되지 않는다. 읽을 수 있는 글자는 없다.",
        "hard_violations": [],
        "physics": "머리는 보이는 목과 상체에 자연스럽게 연결되고 상체가 문틈으로 기울어져 있다. 굽힌 팔과 문 가장자리 가까운 손 일부가 보인다. 발과 지면 접촉은 화면 밖이지만, 서서 허리를 기울이는 동작으로 가능한 자세이며 공중에 뜬 몸이나 지지 없는 물체는 보이지 않는다."
       },
       {
        "label": "A",
        "direction": "얼굴을 문 가장자리 너머 화면 왼쪽으로 내밀고, 눈은 카메라 가까운 문 안쪽을 향한다. 고개를 살짝 기울인 채 눈가를 접고 비스듬히 미소 지어 장난스럽고 능청스러운 재등장으로 읽힌다. 겨누는 도구나 이동 물체는 없다.",
        "built_space": "왼쪽에는 촘촘한 세로 나뭇살과 밧줄로 묶은 가로대 두 개, 아래쪽 넓은 판재 하나가 있고, 오른쪽 목재 면에는 나무 걸쇠 하나가 보인다. 참고 장소의 재료와 고정물 구성이 잘 유지된다. 얼굴은 중앙 오른쪽 전경에 크게 놓이고, 바로 옆 문 가장자리의 두께와 비스듬한 안쪽 면이 보인다. 몸은 오른쪽 문 뒤에 가려져 있으며 틈 너머에는 흙길과 흐릿한 담장·수목이 보인다. 반사면은 없다.",
        "entities": "인물은 한 명으로, 한국인 청년 남성 설정에 맞는 외모와 뒤로 단단히 넘긴 검은 머리를 지녔다. 참고 인물의 얼굴형과 주요 이목구비에 대체로 부합하며 눈은 정상적인 해부학적 형태다. 턱 아래 옷깃만 조금 보이므로 나머지 의상과 지게·쌀 꾸러미는 판단 대상이 아니다. 약탕기와 아궁이도 화면 밖이다. 목재의 마모와 피부 질감은 실물처럼 표현되고, 새벽빛에는 별도의 주황색 화염 조명이 없다. 읽을 수 있는 글자는 없다.",
        "hard_violations": [],
        "physics": "턱 아래 목과 옷깃 일부가 문 뒤로 이어져, 몸을 숨긴 채 목과 상체를 기울여 얼굴만 내미는 자세로 성립한다. 하체의 지지는 가려져 확인되지 않지만 머리가 분리되거나 떠 있는 것으로 보이지 않는다. 문짝과 걸쇠도 구조에 붙어 있으며 지지 없이 떠 있는 물체는 없다."
       }
      ]
     },
     "ok": true
    }
   ],
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "slot_winner_match": true,
   "slot_winner": {
    "gemini-pro": "A",
    "gpt-high": "A"
   },
   "route": "cross_slot_agree"
  },
  "dual": {
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "normalized": {
    "A": 2.0,
    "B": 1.222
   },
   "adjusted": {
    "A": 2.0,
    "B": 1.222
   },
   "violations": {},
   "per_model_winner": {
    "gemini-pro": "A",
    "gpt-high": "A"
   },
   "agreed": true
  },
  "totals": {
   "A": 2000,
   "B": 1222
  },
  "selected": "A",
  "ranking": [
   "A",
   "B"
  ],
  "verdicts": [
   {
    "label": "A",
    "score": 2000,
    "verdict_ko": "프롬프트가 요구한 '얼굴만 내민 채'와 '문이 몸을 가린다'는 지시사항을 완벽하게 따르며, 인물의 장난스러운 표정과 클로즈업 구도를 매우 충실하게 구현했습니다."
   },
   {
    "label": "B",
    "score": 1222,
    "verdict_ko": "인물의 표정과 배경은 훌륭하나, 몸통과 팔이 문틈으로 훤히 드러나 '문이 몸을 가리고 얼굴만 내민다'는 핵심 지시사항을 위반했습니다."
   }
  ],
  "refs": [
   {
    "label": "LOCATION PHOTOGRAPH — the exact place of this shot: its architecture, materials, fixed features and lighting mood are spatial truth; stage the moment inside this place. Never copy its camera framing.",
    "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/db00517b-165b-4005-a287-f28936b59b95/episodes/272bd7af-9294-4490-98c6-88a40e9deacc/images/background_chain/L03B01.png",
    "asset_id": "8263e5ec-d509-48ee-bca3-977ed63622f2",
    "role": "location_plate"
   },
   {
    "label": "CHARACTER REFERENCE — 안규홍: the exact person appearing in this shot; match face, hair and build exactly.",
    "path": "<bytes:1028384>",
    "asset_id": "3b5894db-ce55-4f34-b966-5e725c3aa952",
    "role": "character_ref"
   }
  ],
  "critique_skipped": true,
  "shot_run_produced": false,
  "shot_run_uid": "06aa1403-a478-7c00-b727-87360d00fa47",
  "bgfirst": {
   "bg_path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/db00517b-165b-4005-a287-f28936b59b95/images/272bd7af-9294-4490-98c6-88a40e9deacc/scene/recipe/S2sh13__bgfirst_bg.png",
   "bg_asset_id": "b1d60c72-e8d0-4095-b880-092f1128a9f9",
   "bg_record_key": "S2sh13::bgfirst_bg",
   "chain_winner": true,
   "authority": "plate"
  },
  "ref_mode": "재투영 배경+콘티+엔티티 (2택1: 체인 승)",
  "share_plan": {
   "ref_plan": "background"
  }
 },
 "S2sh13::cine": {
  "applied": true,
  "attempted_at": "2026-09-09T00:43:40.898328+00:00",
  "fingerprint": "e52de610d460c5db3bebf3fd1a3e3d002aa2d28215c89fdff549ae449ed6cf41",
  "fingerprint_version": 2,
  "provider": "mai",
  "endpoint": "openrouter/chat-completions",
  "model": "microsoft/mai-image-2.6",
  "pack": "24.202608252115",
  "source_file": "S2sh13_sel.png",
  "source_sha256": "ebc4fa77817ad2823abd19dace29e81d6d734fe94869db8e64dac7bf6bf6e121",
  "file": "S2sh13_cine.png",
  "staged_sha256": "5389a20eee3e65ddb6b8d88edf009fe53e4c65fda80c51f1da73ab471a7397fa",
  "latency_ms": 28162
 },
 "S2sh14::signage": {
  "fp": "ebe426ef16797a07",
  "inscriptions": [],
  "cues": [],
  "dropped": []
 },
 "S2sh14": {
  "input_fingerprint": "6835e2b7910bb3c9",
  "prompt": "Create ONE FINAL photorealistic live-action film still of the moment below — early 20th-century Korea, Jeolla region; all people are Korean unless explicitly stated otherwise. TIME OF DAY (lock): dawn.\n\nSHOT TEXT (authoritative, Korean): 다시 닫힌 사립문을 바라보며 입가에 옅은 미소를 띠는 안규홍의 어머니의 얼굴.\n\nLOCATION (lock): The home-side vantage facing the closed rustic wicket gate at dawn; the mother's exact position relative to the kitchen and yard is unspecified. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: close-up\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Softly restrained dawn ambience preserves the mother's facial detail and quiet warmth without introducing a new light source or color cue.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The people visible in that photo are NOT in this shot: never carry their faces, bodies or clothing onto anyone here. Each person in THIS shot is defined solely by the PEOPLE list, their own CHARACTER REFERENCE images and their wardrobe notes. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The gate is closed again; white steam rises from the spout of the lidded medicine pot. 안규홍의 어머니: She remains standing, looking at the closed gate with a faint smile.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 안규홍의 어머니 (한국인 여성, 중노년 나이대, 자연스러운 연령대의 얼굴) — wearing: 오랜 병환으로 핏기가 없는 얼굴에 어울리는 색 바랜 흰색 무명 치마와 저고리. — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.",
  "roll_prompts": {
   "A": "Create ONE FINAL photorealistic live-action film still of the moment below — early 20th-century Korea, Jeolla region; all people are Korean unless explicitly stated otherwise. TIME OF DAY (lock): dawn.\n\nSHOT TEXT (authoritative, Korean): 다시 닫힌 사립문을 바라보며 입가에 옅은 미소를 띠는 안규홍의 어머니의 얼굴.\n\nLOCATION (lock): The home-side vantage facing the closed rustic wicket gate at dawn; the mother's exact position relative to the kitchen and yard is unspecified. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: close-up\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Softly restrained dawn ambience preserves the mother's facial detail and quiet warmth without introducing a new light source or color cue.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The people visible in that photo are NOT in this shot: never carry their faces, bodies or clothing onto anyone here. Each person in THIS shot is defined solely by the PEOPLE list, their own CHARACTER REFERENCE images and their wardrobe notes. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The gate is closed again; white steam rises from the spout of the lidded medicine pot. 안규홍의 어머니: She remains standing, looking at the closed gate with a faint smile.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 안규홍의 어머니 (한국인 여성, 중노년 나이대, 자연스러운 연령대의 얼굴) — wearing: 오랜 병환으로 핏기가 없는 얼굴에 어울리는 색 바랜 흰색 무명 치마와 저고리. — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.",
   "B": "Create ONE FINAL photorealistic live-action film still of the moment below — early 20th-century Korea, Jeolla region; all people are Korean unless explicitly stated otherwise. TIME OF DAY (lock): dawn.\n\nSHOT TEXT (authoritative, Korean): 다시 닫힌 사립문을 바라보며 입가에 옅은 미소를 띠는 안규홍의 어머니의 얼굴.\n\nLOCATION (lock): The home-side vantage facing the closed rustic wicket gate at dawn; the mother's exact position relative to the kitchen and yard is unspecified. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: close-up\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Softly restrained dawn ambience preserves the mother's facial detail and quiet warmth without introducing a new light source or color cue.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The people visible in that photo are NOT in this shot: never carry their faces, bodies or clothing onto anyone here. Each person in THIS shot is defined solely by the PEOPLE list, their own CHARACTER REFERENCE images and their wardrobe notes. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The gate is closed again; white steam rises from the spout of the lidded medicine pot. 안규홍의 어머니: She remains standing, looking at the closed gate with a faint smile.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 안규홍의 어머니 (한국인 여성, 중노년 나이대, 자연스러운 연령대의 얼굴) — wearing: 오랜 병환으로 핏기가 없는 얼굴에 어울리는 색 바랜 흰색 무명 치마와 저고리. — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.\n\nCOMPOSITION VARIATION (this candidate only): keep the place, the people and the moment exactly as specified, and keep from the CAMERA & FRAME contract its subject scale, screen placement and key background elements exactly. For this candidate only, its camera angle and height are deliberately left open: choose a camera position distinctly different from the obvious one, as a film director picking a second setup on the same blocking. Stay true to any explicit framing the SHOT TEXT itself dictates."
  },
  "shot_run_spend_attempt_count": 1,
  "initial_roll_all_fail": true,
  "readings": [
   {
    "label": "A",
    "direction": "여자의 시선이 화면의 왼쪽 평면을 향하고 있어, Z축 깊이감이 있는 배경의 사립문에 닿지 않고 화면 밖을 향합니다.",
    "built_space": "배경 왼쪽에 사립문이 있으나 문틈이 벌어져 빛이 들어옵니다. 화면 하단 중앙의 부뚜막 형태가 레퍼런스의 아궁이 구조와 일치하지 않습니다.",
    "entities": "중노년의 한국인 여성, 옅은 미소, 빛바랜 흰색 옷은 지시문과 일치합니다. 그러나 약탕기 대신 현대식 손잡이가 달린 양은/스텐 냄비가 놓여 있어 시대상에 맞지 않습니다.",
    "hard_violations": [],
    "physics": "여성은 바닥을 딛고 서 있으며, 냄비는 부뚜막 위에 안정적으로 놓여 있고 김이 자연스럽게 위로 피어오릅니다."
   },
   {
    "label": "B",
    "direction": "여자가 화면 왼쪽을 바라보고 있으나, 사립문은 여자의 어깨 너머 뒤쪽 배경에 위치하여 시선이 목표물(사립문)을 향하지 않습니다.",
    "built_space": "배경에 완전히 닫힌 사립문이 자리 잡고 있습니다. 왼쪽의 아궁이는 레퍼런스 이미지에서 보여준 흙벽돌 구조 및 형태적 특징을 잘 유지하고 있습니다.",
    "entities": "중노년의 한국 여성, 옅은 미소, 흰색 무명 저고리 모두 잘 표현되었습니다. 아궁이 위에는 뚜껑 덮인 솥이 있고 김이 피어오릅니다.",
    "hard_violations": [],
    "physics": "여성의 서 있는 자세, 아궁이 위 솥의 거치 상태, 위로 피어오르는 김의 방향 모두 물리적 지탱과 작용이 자연스럽게 표현되었습니다."
   }
  ],
  "cross_model_order": {
   "policy": "cross_model_order_v2_slot0_fwd_slot1_rev_parallel2",
   "slots": [
    {
     "model": "gemini-pro",
     "order": "forward",
     "display_to_canonical": {
      "A": "A",
      "B": "B"
     },
     "raw": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 4,
        "verdict_ko": "배경의 사립문과 아궁이 구조를 레퍼런스에 맞게 잘 구현했으나, 여자의 시선이 사립문을 향하지 않아 핵심 행동 지시를 실패했습니다."
       },
       {
        "label": "A",
        "score": 2,
        "verdict_ko": "여자의 시선이 사립문을 벗어났을 뿐만 아니라, 문이 덜 닫혀 있고 부뚜막에 시대에 맞지 않는 현대식 냄비가 배치되어 지시문과 어긋납니다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "여자의 시선이 화면의 왼쪽 평면을 향하고 있어, Z축 깊이감이 있는 배경의 사립문에 닿지 않고 화면 밖을 향합니다.",
        "built_space": "배경 왼쪽에 사립문이 있으나 문틈이 벌어져 빛이 들어옵니다. 화면 하단 중앙의 부뚜막 형태가 레퍼런스의 아궁이 구조와 일치하지 않습니다.",
        "entities": "중노년의 한국인 여성, 옅은 미소, 빛바랜 흰색 옷은 지시문과 일치합니다. 그러나 약탕기 대신 현대식 손잡이가 달린 양은/스텐 냄비가 놓여 있어 시대상에 맞지 않습니다.",
        "hard_violations": [],
        "physics": "여성은 바닥을 딛고 서 있으며, 냄비는 부뚜막 위에 안정적으로 놓여 있고 김이 자연스럽게 위로 피어오릅니다."
       },
       {
        "label": "B",
        "direction": "여자가 화면 왼쪽을 바라보고 있으나, 사립문은 여자의 어깨 너머 뒤쪽 배경에 위치하여 시선이 목표물(사립문)을 향하지 않습니다.",
        "built_space": "배경에 완전히 닫힌 사립문이 자리 잡고 있습니다. 왼쪽의 아궁이는 레퍼런스 이미지에서 보여준 흙벽돌 구조 및 형태적 특징을 잘 유지하고 있습니다.",
        "entities": "중노년의 한국 여성, 옅은 미소, 흰색 무명 저고리 모두 잘 표현되었습니다. 아궁이 위에는 뚜껑 덮인 솥이 있고 김이 피어오릅니다.",
        "hard_violations": [],
        "physics": "여성의 서 있는 자세, 아궁이 위 솥의 거치 상태, 위로 피어오르는 김의 방향 모두 물리적 지탱과 작용이 자연스럽게 표현되었습니다."
       }
      ],
      "all_candidates_fail": true
     },
     "normalized": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 4,
        "verdict_ko": "배경의 사립문과 아궁이 구조를 레퍼런스에 맞게 잘 구현했으나, 여자의 시선이 사립문을 향하지 않아 핵심 행동 지시를 실패했습니다."
       },
       {
        "label": "A",
        "score": 2,
        "verdict_ko": "여자의 시선이 사립문을 벗어났을 뿐만 아니라, 문이 덜 닫혀 있고 부뚜막에 시대에 맞지 않는 현대식 냄비가 배치되어 지시문과 어긋납니다."
       }
      ],
      "all_candidates_fail": true,
      "readings": [
       {
        "label": "A",
        "direction": "여자의 시선이 화면의 왼쪽 평면을 향하고 있어, Z축 깊이감이 있는 배경의 사립문에 닿지 않고 화면 밖을 향합니다.",
        "built_space": "배경 왼쪽에 사립문이 있으나 문틈이 벌어져 빛이 들어옵니다. 화면 하단 중앙의 부뚜막 형태가 레퍼런스의 아궁이 구조와 일치하지 않습니다.",
        "entities": "중노년의 한국인 여성, 옅은 미소, 빛바랜 흰색 옷은 지시문과 일치합니다. 그러나 약탕기 대신 현대식 손잡이가 달린 양은/스텐 냄비가 놓여 있어 시대상에 맞지 않습니다.",
        "hard_violations": [],
        "physics": "여성은 바닥을 딛고 서 있으며, 냄비는 부뚜막 위에 안정적으로 놓여 있고 김이 자연스럽게 위로 피어오릅니다."
       },
       {
        "label": "B",
        "direction": "여자가 화면 왼쪽을 바라보고 있으나, 사립문은 여자의 어깨 너머 뒤쪽 배경에 위치하여 시선이 목표물(사립문)을 향하지 않습니다.",
        "built_space": "배경에 완전히 닫힌 사립문이 자리 잡고 있습니다. 왼쪽의 아궁이는 레퍼런스 이미지에서 보여준 흙벽돌 구조 및 형태적 특징을 잘 유지하고 있습니다.",
        "entities": "중노년의 한국 여성, 옅은 미소, 흰색 무명 저고리 모두 잘 표현되었습니다. 아궁이 위에는 뚜껑 덮인 솥이 있고 김이 피어오릅니다.",
        "hard_violations": [],
        "physics": "여성의 서 있는 자세, 아궁이 위 솥의 거치 상태, 위로 피어오르는 김의 방향 모두 물리적 지탱과 작용이 자연스럽게 표현되었습니다."
       }
      ]
     },
     "ok": true
    },
    {
     "model": "gpt-high",
     "order": "reverse",
     "display_to_canonical": {
      "A": "B",
      "B": "A"
     },
     "raw": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 3,
        "verdict_ko": "새벽빛과 옅은 미소는 맞지만, 얼굴 클로즈업보다 넓은 상반신 구도이며 시선이 뒤쪽의 닫힌 문을 향하지 않는다."
       },
       {
        "label": "B",
        "score": 5,
        "verdict_ko": "얼굴 중심 클로즈업과 잔잔한 미소는 더 충실하지만, 닫힌 문을 뒤에 두고 카메라 왼편을 바라보아 핵심 시선 지시는 여전히 실패한다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "어머니의 얼굴과 눈은 카메라 가까운 화면 왼쪽을 향한다. 닫힌 문은 어머니의 왼쪽 뒤 배경에 있으므로, 보이는 시선의 목표는 문이 아니라 카메라 왼편의 화면 밖 공간이다. 무기나 손에 든 지시 대상은 없다.",
        "built_space": "중앙 왼쪽에 두 짝으로 구성된 닫힌 나무문 한 개, 왼쪽 전경에 그을린 흙 아궁이 한 개가 보인다. 그 위와 주변에는 여러 검은 용기와 금속 뚜껑 냄비가 있고, 왼쪽 벽에는 국자 한 개, 오른쪽 벽에는 흐릿한 걸이 도구들이 있다. 어머니는 문과 아궁이보다 카메라 가까운 오른쪽에 있다. 흙벽·목재·그을음은 참조와 유사하지만, 참조에서 문과 전체 배치가 드러나지 않아 정확한 공간 일치는 확인할 수 없다. 얼굴뿐 아니라 허리 부근까지 보여 요구된 클로즈업보다 넓다.",
        "entities": "중노년 한국인 여성 설정에 부합하는 외모의 인물 한 명만 보이며, 회색 머리와 자연스러운 주름, 색 바랜 흰색 계열 무명 저고리가 보인다. 입가에는 옅은 미소가 있지만 오랜 병환의 창백함은 뚜렷하지 않다. 닫힌 나무문은 있으나 성긴 사립문보다는 판재 대문처럼 보인다. 뚜껑 있는 냄비 위로 흰 김이 올라오지만 약탕기의 주둥이는 식별되지 않는다. 읽을 수 있는 글자나 이전 장면의 남자는 없다.",
        "hard_violations": [],
        "physics": "머리·목·상체의 연결과 옷의 처짐은 자연스럽고 서 있는 자세와 양립한다. 발은 프레임 밖이므로 바닥 접촉은 확인할 수 없지만 부유의 증거는 없다. 냄비들은 아궁이 상판에 놓여 지지되고, 문은 문틀 안에 서 있다. 김의 상승은 가열된 용기에서 발생하는 현상으로 타당하지만 주둥이에서 나오는지는 확인되지 않는다."
       },
       {
        "label": "B",
        "direction": "어머니는 카메라 왼편 앞쪽의 화면 밖을 응시한다. 닫힌 문은 왼쪽 후방 배경에 있어 눈과 얼굴의 방향이 그 문에 닿지 않는다. 옅은 미소는 보이지만 '다시 닫힌 사립문을 바라보는' 관계는 성립하지 않는다.",
        "built_space": "왼쪽 배경에 두 짝으로 된 닫힌 나무문 한 개, 하단에 일부 잘린 흙 아궁이 상판 한 개와 뚜껑 냄비 한 개가 보인다. 뒤쪽 흙벽에는 둥근 도구 하나가 걸려 있다. 어머니는 오른쪽 전경에서 얼굴과 어깨·가슴 윗부분을 차지하여 A보다 요구한 클로즈업에 가깝다. 흙과 낡은 나무의 재질은 참조와 어울리지만 문 배치의 동일성은 참조만으로 확정할 수 없다. 조명은 참조의 푸른 새벽빛보다 다소 따뜻하고 밝다.",
        "entities": "중노년 한국인 여성 설정에 부합하는 외모의 인물 한 명이며, 자연스러운 노화와 흰머리 섞인 쪽머리, 낡은 흰색 계열 무명 저고리, 잔잔한 입가의 미소가 보인다. 병색과 창백함은 약하다. 치마는 클로즈업 밖이므로 판정 대상이 아니다. 닫힌 판재 나무문과 뚜껑 있는 검은 냄비, 흰 김은 있으나 사립문의 성긴 구조와 약탕기 주둥이는 확인되지 않는다. 추가 인물이나 읽을 수 있는 문자는 없다.",
        "hard_violations": [],
        "physics": "머리는 목과 상체에 정상적으로 지지되고 상체는 곧게 서 있어 기립 상태와 모순되지 않는다. 하체는 화면 밖이며 지지 불가능한 자세는 보이지 않는다. 냄비는 흙 상판 위에 놓여 있고 김은 그 위에서 상승한다. 다만 김의 발생 지점이 주둥이인지 뚜껑 주변인지는 흐림 때문에 구분되지 않는다."
       }
      ],
      "all_candidates_fail": true
     },
     "normalized": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 3,
        "verdict_ko": "새벽빛과 옅은 미소는 맞지만, 얼굴 클로즈업보다 넓은 상반신 구도이며 시선이 뒤쪽의 닫힌 문을 향하지 않는다."
       },
       {
        "label": "A",
        "score": 5,
        "verdict_ko": "얼굴 중심 클로즈업과 잔잔한 미소는 더 충실하지만, 닫힌 문을 뒤에 두고 카메라 왼편을 바라보아 핵심 시선 지시는 여전히 실패한다."
       }
      ],
      "all_candidates_fail": true,
      "readings": [
       {
        "label": "B",
        "direction": "어머니의 얼굴과 눈은 카메라 가까운 화면 왼쪽을 향한다. 닫힌 문은 어머니의 왼쪽 뒤 배경에 있으므로, 보이는 시선의 목표는 문이 아니라 카메라 왼편의 화면 밖 공간이다. 무기나 손에 든 지시 대상은 없다.",
        "built_space": "중앙 왼쪽에 두 짝으로 구성된 닫힌 나무문 한 개, 왼쪽 전경에 그을린 흙 아궁이 한 개가 보인다. 그 위와 주변에는 여러 검은 용기와 금속 뚜껑 냄비가 있고, 왼쪽 벽에는 국자 한 개, 오른쪽 벽에는 흐릿한 걸이 도구들이 있다. 어머니는 문과 아궁이보다 카메라 가까운 오른쪽에 있다. 흙벽·목재·그을음은 참조와 유사하지만, 참조에서 문과 전체 배치가 드러나지 않아 정확한 공간 일치는 확인할 수 없다. 얼굴뿐 아니라 허리 부근까지 보여 요구된 클로즈업보다 넓다.",
        "entities": "중노년 한국인 여성 설정에 부합하는 외모의 인물 한 명만 보이며, 회색 머리와 자연스러운 주름, 색 바랜 흰색 계열 무명 저고리가 보인다. 입가에는 옅은 미소가 있지만 오랜 병환의 창백함은 뚜렷하지 않다. 닫힌 나무문은 있으나 성긴 사립문보다는 판재 대문처럼 보인다. 뚜껑 있는 냄비 위로 흰 김이 올라오지만 약탕기의 주둥이는 식별되지 않는다. 읽을 수 있는 글자나 이전 장면의 남자는 없다.",
        "hard_violations": [],
        "physics": "머리·목·상체의 연결과 옷의 처짐은 자연스럽고 서 있는 자세와 양립한다. 발은 프레임 밖이므로 바닥 접촉은 확인할 수 없지만 부유의 증거는 없다. 냄비들은 아궁이 상판에 놓여 지지되고, 문은 문틀 안에 서 있다. 김의 상승은 가열된 용기에서 발생하는 현상으로 타당하지만 주둥이에서 나오는지는 확인되지 않는다."
       },
       {
        "label": "A",
        "direction": "어머니는 카메라 왼편 앞쪽의 화면 밖을 응시한다. 닫힌 문은 왼쪽 후방 배경에 있어 눈과 얼굴의 방향이 그 문에 닿지 않는다. 옅은 미소는 보이지만 '다시 닫힌 사립문을 바라보는' 관계는 성립하지 않는다.",
        "built_space": "왼쪽 배경에 두 짝으로 된 닫힌 나무문 한 개, 하단에 일부 잘린 흙 아궁이 상판 한 개와 뚜껑 냄비 한 개가 보인다. 뒤쪽 흙벽에는 둥근 도구 하나가 걸려 있다. 어머니는 오른쪽 전경에서 얼굴과 어깨·가슴 윗부분을 차지하여 A보다 요구한 클로즈업에 가깝다. 흙과 낡은 나무의 재질은 참조와 어울리지만 문 배치의 동일성은 참조만으로 확정할 수 없다. 조명은 참조의 푸른 새벽빛보다 다소 따뜻하고 밝다.",
        "entities": "중노년 한국인 여성 설정에 부합하는 외모의 인물 한 명이며, 자연스러운 노화와 흰머리 섞인 쪽머리, 낡은 흰색 계열 무명 저고리, 잔잔한 입가의 미소가 보인다. 병색과 창백함은 약하다. 치마는 클로즈업 밖이므로 판정 대상이 아니다. 닫힌 판재 나무문과 뚜껑 있는 검은 냄비, 흰 김은 있으나 사립문의 성긴 구조와 약탕기 주둥이는 확인되지 않는다. 추가 인물이나 읽을 수 있는 문자는 없다.",
        "hard_violations": [],
        "physics": "머리는 목과 상체에 정상적으로 지지되고 상체는 곧게 서 있어 기립 상태와 모순되지 않는다. 하체는 화면 밖이며 지지 불가능한 자세는 보이지 않는다. 냄비는 흙 상판 위에 놓여 있고 김은 그 위에서 상승한다. 다만 김의 발생 지점이 주둥이인지 뚜껑 주변인지는 흐림 때문에 구분되지 않는다."
       }
      ]
     },
     "ok": true
    }
   ],
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "slot_winner_match": false,
   "slot_winner": {
    "gemini-pro": "B",
    "gpt-high": "A"
   },
   "route": "cross_slot_combined"
  },
  "dual": {
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "normalized": {
    "A": 1.5,
    "B": 1.6
   },
   "adjusted": {
    "A": 1.5,
    "B": 1.6
   },
   "violations": {},
   "per_model_winner": {
    "gemini-pro": "B",
    "gpt-high": "A"
   },
   "agreed": false
  },
  "totals": {
   "B": 1600,
   "A": 1500
  },
  "selected": "B",
  "ranking": [
   "B",
   "A"
  ],
  "verdicts": [
   {
    "label": "B",
    "score": 1600,
    "verdict_ko": "배경의 사립문과 아궁이 구조를 레퍼런스에 맞게 잘 구현했으나, 여자의 시선이 사립문을 향하지 않아 핵심 행동 지시를 실패했습니다."
   },
   {
    "label": "A",
    "score": 1500,
    "verdict_ko": "여자의 시선이 사립문을 벗어났을 뿐만 아니라, 문이 덜 닫혀 있고 부뚜막에 시대에 맞지 않는 현대식 냄비가 배치되어 지시문과 어긋납니다."
   }
  ],
  "refs": [
   {
    "label": "PREVIOUS SHOT STILL — a visually related earlier shot of this same place: the location's look, materials, fixed features and lighting mood are LOCKED to this photo; never copy its camera framing. The people visible in that photo are NOT in this shot: never carry their faces, bodies or clothing onto anyone here. Each person in THIS shot is defined solely by the PEOPLE list, their own CHARACTER REFERENCE images and their wardrobe notes. If this photo shows a character who cannot move (dead or unconscious), that character's exact pose, position and orientation are ALSO LOCKED — treat their body as a fixed prop of the set that only the camera moves around.",
    "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/db00517b-165b-4005-a287-f28936b59b95/images/272bd7af-9294-4490-98c6-88a40e9deacc/scene/recipe/S2sh1_sel.png",
    "asset_id": "74c63b98-8685-4cc9-b967-a3f1b3bed8da",
    "role": "prev_still"
   }
  ],
  "critique_skipped": true,
  "needs_reshoot": true,
  "shot_run_produced": false,
  "shot_run_uid": "06aa1403-a749-7396-8211-16424e17796c",
  "ref_mode": "prev+엔티티",
  "share_plan": {
   "ref_plan": "prev",
   "prev_anchor_tag": "S2sh1"
  }
 },
 "S2sh14::cine": {
  "applied": true,
  "attempted_at": "2026-09-09T00:45:32.287966+00:00",
  "fingerprint": "7cfd7ae36f27c910ae15d3420c5c096a15c5547d8ca351a37e9dfca654571fd1",
  "fingerprint_version": 2,
  "provider": "mai",
  "endpoint": "openrouter/chat-completions",
  "model": "microsoft/mai-image-2.6",
  "pack": "24.202608252115",
  "source_file": "S2sh14_sel.png",
  "source_sha256": "b94b484848f2dba0fbc6ac2c7e0672199419730421618b6f571103e946d6ed8f",
  "file": "S2sh14_cine.png",
  "staged_sha256": "4acccbea10af6aabe7872b1ae65befcb860eddcd6468e945fb8bb78c1ed1c638",
  "latency_ms": 26642
 },
 "S3sh5::signage": {
  "fp": "90ef94894944969e",
  "inscriptions": [],
  "cues": [],
  "dropped": []
 },
 "era_assess::f9c2b0f4a2c30a08": {
  "subjects": [
   {
    "subject_native": "20세기 초 전라도 전통 가옥 마당과 부속채(곳간)",
    "search_terms_native": [
     "전라도 전통 한옥 마당",
     "남도 고택 곳간채",
     "전라도 전통 가옥 안마당",
     "조선 후기 부농 가옥 마당"
    ],
    "language_lock_native": "반드시 한국어로만 검색해야 하며 다른 언어로 번역하거나 외래어를 섞지 마십시오.",
    "reason_ko": "AI가 한옥의 마당 및 곳간 등의 부속채 구조를 중국이나 일본의 전통 가옥 양식과 혼동하기 쉽기 때문입니다."
   }
  ],
  "subject_text": "박제현 집 마당·일터·대문·곳간 앞\n사랑채와 안채, 곳간이 면한 넓은 흙마당. 작업용 평상과 볏짐이 놓이고, 큰 대문 곁에 툇돌이 있다.",
  "identity": "canonical",
  "scope_id": "L08",
  "scope_role": "location_exterior",
  "scope_sha": "5a65b38515620901"
 },
 "era_fail::fe018e58a475315f": {
  "stage": "research",
  "subject": "20세기 초 전라도 전통 가옥 마당과 부속채(곳간)",
  "terms": [
   "전라도 전통 한옥 마당",
   "남도 고택 곳간채",
   "전라도 전통 가옥 안마당",
   "조선 후기 부농 가옥 마당"
  ],
  "status": "no_usable",
  "queries": [
   [
    "전라도 고택 곳간채 안마당 나주",
    "보성 이금재 고택 곳간채 마당",
    "구례 운조루 안마당 곳간"
   ],
   [
    "보성 이정래 고택 이금재 고택 곳간채 국가유산",
    "나주 남파 고택 곳간채 안채 1918 국가유산"
   ],
   [
    "\"보성 이금재 고택\" \"곳간\"",
    "\"보성 이정래 고택\" \"1900\""
   ]
  ],
  "candidate_urls": [
   "https://minio.nculture.org/amsweb-opt/multimedia_assets/154/85600/90944/c/%EB%82%98%EC%A3%BC-%ED%99%8D%EA%B8%B0%EC%B0%BD%EA%B0%80%EC%98%A5-%283%29-medium-size.jpg",
   "https://blog.kakaocdn.net/dna/DUFto/btsNRQo9rKY/AAAAAAAAAAAAAAAAAAAAALZQbyTxgRGo3ZzzaOB9E35klnSE2pXTTdyTQ6odKCTu/img.jpg?allow_ip=&allow_referer=&credential=yqXZFxpELC7KVnFOS48ylbz2pIh7yKj8&expires=1777561199&signature=NzT0P0Ponb1aiX6b4RlNkXFBoF4%3D",
   "https://img.hankyung.com/photo/202011/AKR20201015124200805_01_i_P4.jpg",
   "https://blog.kakaocdn.net/dna/UQo86/btsNS5S3yyE/AAAAAAAAAAAAAAAAAAAAAOle6Qo9tWvpO4ay8-34nFZxnnW45sx_te5OrWTB9WvA/img.jpg?allow_ip=&allow_referer=&credential=yqXZFxpELC7KVnFOS48ylbz2pIh7yKj8&expires=1777561199&signature=dbAskhybcjsKuvBIRAh6UomQ998%3D"
  ],
  "coarse": {
   "eligible": [],
   "chosen_index": 0,
   "reason": "종류·보임·기준을 다 만족하는 후보가 없다 — 이 라운드에선 안 고른다",
   "single_judge": true,
   "rejected_judges": {}
  },
  "verdicts": [
   {
    "index": 1,
    "object_type_match": "yes",
    "visible": true,
    "criteria_match": "unsure",
    "similarity": 90
   },
   {
    "index": 2,
    "object_type_match": "yes",
    "visible": true,
    "criteria_match": "unsure",
    "similarity": 85
   },
   {
    "index": 3,
    "object_type_match": "unsure",
    "visible": true,
    "criteria_match": "unsure",
    "similarity": 30
   },
   {
    "index": 4,
    "object_type_match": "yes",
    "visible": true,
    "criteria_match": "unsure",
    "similarity": 85
   }
  ],
  "chosen_reason_ko": "",
  "attempts": 5
 },
 "S3sh5::bgfirst_bg": {
  "input_fingerprint": "471e81d2f8f6c1a0",
  "prompt": "Create the EMPTY BACKGROUND PLATE for one film shot — NO PEOPLE, no figures, no body parts, no sketch lines, no arrows anywhere.\n\"Empty\" means no people only: KEEP the location's inherent occupants and stock that define the place — animals in an animal shelter, pen or farm, goods in a market, moored boats in a harbour — unless the shot text explicitly removes them.\nThe FIRST attached image is a thin-line storyboard sketch: use ONLY its camera angle, horizon, perspective and the placement/size of buildings and set masses — ignore the sketched people and arrows entirely. The SECOND attached image (LOCATION PHOTOGRAPH) is the real place: take its architecture, materials, signage and fixed features, and RE-PROJECT them into the sketch's camera. If the photograph's camera differs from the sketch's, the sketch's camera wins.\nHUMAN-SCALE CALIBRATION: derive every structure's true size from human-scale elements — a door ≈ 2m, a window ≈ 1–1.5m wide, one storey ≈ 2.5–3m; never inflate a small structure or shrink a large one.\n\nSHOT TEXT this background must serve (Korean): 양쪽에서 볏짐을 번쩍 들어 올려 허공에 띄운 채 멈춘, 팔 근육이 팽팽하게 부풀어 오른 안규홍과 장신 조선인 일꾼의 상체.\n\nLOCATION (lock): The sun-beaten dirt work yard of a traditional estate, at the spot where a large straw load is being lifted onto a carrying frame.\n\nTIME OF DAY (lock): morning, harsh sunlight.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: medium shot\n- KEY BACKGROUND ELEMENTS: 볏짐 (Lifted clear of the ground and held between the two workers) — Opposite gripped edges are visible beneath the men's arms; used as Spans the lower third as a shared load, leaving both sets of hands, forearms, and upper bodies unobstructed; 마당의 흙바닥 (Below the suspended bundle); used as Small lateral glimpses establish clearance beneath the load without adding airborne dust.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Strong morning sunlight renders the physical strain with natural contrast, muted earth tones, and unpolished texture.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nRender ONE photorealistic empty location photograph, 16:9, that this shot can be staged inside later. No readable writing anywhere: surfaces that would carry writing may be present, but stage any wording out of legibility — an oblique angle, distance, shallow focus. No captions, watermarks or overlay text.",
  "effective_prompt": "Create the EMPTY BACKGROUND PLATE for one film shot — NO PEOPLE, no figures, no body parts, no sketch lines, no arrows anywhere.\n\"Empty\" means no people only: KEEP the location's inherent occupants and stock that define the place — animals in an animal shelter, pen or farm, goods in a market, moored boats in a harbour — unless the shot text explicitly removes them.\nThe FIRST attached image is a thin-line storyboard sketch: use ONLY its camera angle, horizon, perspective and the placement/size of buildings and set masses — ignore the sketched people and arrows entirely. The SECOND attached image (LOCATION PHOTOGRAPH) is the real place: take its architecture, materials, signage and fixed features, and RE-PROJECT them into the sketch's camera. If the photograph's camera differs from the sketch's, the sketch's camera wins.\nHUMAN-SCALE CALIBRATION: derive every structure's true size from human-scale elements — a door ≈ 2m, a window ≈ 1–1.5m wide, one storey ≈ 2.5–3m; never inflate a small structure or shrink a large one.\n\nSHOT TEXT this background must serve (Korean): 양쪽에서 볏짐을 번쩍 들어 올려 허공에 띄운 채 멈춘, 팔 근육이 팽팽하게 부풀어 오른 안규홍과 장신 조선인 일꾼의 상체.\n\nLOCATION (lock): The sun-beaten dirt work yard of a traditional estate, at the spot where a large straw load is being lifted onto a carrying frame.\n\nTIME OF DAY (lock): morning, harsh sunlight.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: medium shot\n- KEY BACKGROUND ELEMENTS: 볏짐 (Lifted clear of the ground and held between the two workers) — Opposite gripped edges are visible beneath the men's arms; used as Spans the lower third as a shared load, leaving both sets of hands, forearms, and upper bodies unobstructed; 마당의 흙바닥 (Below the suspended bundle); used as Small lateral glimpses establish clearance beneath the load without adding airborne dust.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Strong morning sunlight renders the physical strain with natural contrast, muted earth tones, and unpolished texture.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nRender ONE photorealistic empty location photograph, 16:9, that this shot can be staged inside later. No readable writing anywhere: surfaces that would carry writing may be present, but stage any wording out of legibility — an oblique angle, distance, shallow focus. No captions, watermarks or overlay text.",
  "bg_path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/db00517b-165b-4005-a287-f28936b59b95/images/272bd7af-9294-4490-98c6-88a40e9deacc/scene/recipe/S3sh5__bgfirst_bg.png",
  "asset_id": "e853c50a-c64b-4fda-8cf9-b920ac435aa3",
  "input_asset_ids": [
   "3e173d0f-3019-472e-bff1-f56a6ba53c08",
   "71e0793a-2c9b-4326-bdfc-5abf8fffaac8"
  ]
 },
 "S3sh5": {
  "input_fingerprint": "95713b2e3241e953",
  "prompt": "Create ONE FINAL photorealistic live-action film still of the moment below — early 20th-century Korea, Jeolla region; all people are Korean unless explicitly stated otherwise. TIME OF DAY (lock): morning, harsh sunlight.\n\nSHOT TEXT (authoritative, Korean): 양쪽에서 볏짐을 번쩍 들어 올려 허공에 띄운 채 멈춘, 팔 근육이 팽팽하게 부풀어 오른 안규홍과 장신 조선인 일꾼의 상체.\n\nLOCATION (lock): The sun-beaten dirt work yard of a traditional estate, at the spot where a large straw load is being lifted onto a carrying frame. The shot takes place here — the attached LOCATION PHOTOGRAPH shows the exact spot.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: medium shot\n- KEY BACKGROUND ELEMENTS: 볏짐 (Lifted clear of the ground and held between the two workers) — Opposite gripped edges are visible beneath the men's arms; used as Spans the lower third as a shared load, leaving both sets of hands, forearms, and upper bodies unobstructed; 마당의 흙바닥 (Below the suspended bundle); used as Small lateral glimpses establish clearance beneath the load without adding airborne dust.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Strong morning sunlight renders the physical strain with natural contrast, muted earth tones, and unpolished texture.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): Blazing sunshine covers the dirt yard and wooden platform; the fallen rice-straw load is now raised. 안규홍: He stands gripping a corner of the load, wearing worn straw shoes.\n\nTHE HAND THAT IS DOING THIS: the object at the centre of this shot is being held, operated, read, aimed or handed over by 안규홍과 장신 조선인 일꾼 right now, so 안규홍과 장신 조선인 일꾼's hand — and as much of the wrist and forearm as the framing reaches — is in the frame, gripping or resting on that object exactly the way its real use requires. Match that hand to 안규홍과 장신 조선인 일꾼: its size, build, skin, age, grooming, sleeve and anything worn on it belong to that person and to no one else.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 안규홍 (한국인 남성, 청년 나이대, 검은 머리, 단단히 묶은 머리) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.",
  "roll_prompts": {
   "A": "Stage the shot. The FIRST attached image (SHOT BACKGROUND) is the finished empty background of this shot — keep it EXACTLY: its camera, perspective, architecture, lighting and every fixed feature stay untouched. The SECOND attached image (LAYOUT SKETCH) tells you ONLY where the people go: each sketched person's position, screen size, pose and the gaze/motion arrows. Ignore the sketch's background lines. The CHARACTER REFERENCE photographs show the real people.\nPlace the real people into the background at exactly the sketched positions, sizes and poses, following the arrow directions. No sketch lines or arrows may remain.\n\nCreate ONE FINAL photorealistic live-action film still of the moment below — early 20th-century Korea, Jeolla region; all people are Korean unless explicitly stated otherwise. TIME OF DAY (lock): morning, harsh sunlight.\n\nSHOT TEXT (authoritative, Korean): 양쪽에서 볏짐을 번쩍 들어 올려 허공에 띄운 채 멈춘, 팔 근육이 팽팽하게 부풀어 오른 안규홍과 장신 조선인 일꾼의 상체.\n\nLOCATION (lock): The sun-beaten dirt work yard of a traditional estate, at the spot where a large straw load is being lifted onto a carrying frame. The shot takes place here — the FIRST attached image (SHOT BACKGROUND) is this exact place, already built: its ground, structures, horizon, materials and lighting are the finished truth of this location and must not be redesigned or replaced. No location photograph is attached — read the place from that image alone, and add no scenery, structure, vehicle or fixture that it does not already show. This lock governs the place only; the figures in the shot follow the staging and pose instructions.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: medium shot\n- KEY BACKGROUND ELEMENTS: 볏짐 (Lifted clear of the ground and held between the two workers) — Opposite gripped edges are visible beneath the men's arms; used as Spans the lower third as a shared load, leaving both sets of hands, forearms, and upper bodies unobstructed; 마당의 흙바닥 (Below the suspended bundle); used as Small lateral glimpses establish clearance beneath the load without adding airborne dust.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Strong morning sunlight renders the physical strain with natural contrast, muted earth tones, and unpolished texture.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): Blazing sunshine covers the dirt yard and wooden platform; the fallen rice-straw load is now raised. 안규홍: He stands gripping a corner of the load, wearing worn straw shoes.\n\nTHE HAND THAT IS DOING THIS: the object at the centre of this shot is being held, operated, read, aimed or handed over by 안규홍과 장신 조선인 일꾼 right now, so 안규홍과 장신 조선인 일꾼's hand — and as much of the wrist and forearm as the framing reaches — is in the frame, gripping or resting on that object exactly the way its real use requires. Match that hand to 안규홍과 장신 조선인 일꾼: its size, build, skin, age, grooming, sleeve and anything worn on it belong to that person and to no one else.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 안규홍 (한국인 남성, 청년 나이대, 검은 머리, 단단히 묶은 머리) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.",
   "B": "Create ONE FINAL photorealistic live-action film still of the moment below — early 20th-century Korea, Jeolla region; all people are Korean unless explicitly stated otherwise. TIME OF DAY (lock): morning, harsh sunlight.\n\nSHOT TEXT (authoritative, Korean): 양쪽에서 볏짐을 번쩍 들어 올려 허공에 띄운 채 멈춘, 팔 근육이 팽팽하게 부풀어 오른 안규홍과 장신 조선인 일꾼의 상체.\n\nLOCATION (lock): The sun-beaten dirt work yard of a traditional estate, at the spot where a large straw load is being lifted onto a carrying frame. The shot takes place here — the attached LOCATION PHOTOGRAPH shows the exact spot.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: medium shot\n- KEY BACKGROUND ELEMENTS: 볏짐 (Lifted clear of the ground and held between the two workers) — Opposite gripped edges are visible beneath the men's arms; used as Spans the lower third as a shared load, leaving both sets of hands, forearms, and upper bodies unobstructed; 마당의 흙바닥 (Below the suspended bundle); used as Small lateral glimpses establish clearance beneath the load without adding airborne dust.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Strong morning sunlight renders the physical strain with natural contrast, muted earth tones, and unpolished texture.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): Blazing sunshine covers the dirt yard and wooden platform; the fallen rice-straw load is now raised. 안규홍: He stands gripping a corner of the load, wearing worn straw shoes.\n\nTHE HAND THAT IS DOING THIS: the object at the centre of this shot is being held, operated, read, aimed or handed over by 안규홍과 장신 조선인 일꾼 right now, so 안규홍과 장신 조선인 일꾼's hand — and as much of the wrist and forearm as the framing reaches — is in the frame, gripping or resting on that object exactly the way its real use requires. Match that hand to 안규홍과 장신 조선인 일꾼: its size, build, skin, age, grooming, sleeve and anything worn on it belong to that person and to no one else.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 안규홍 (한국인 남성, 청년 나이대, 검은 머리, 단단히 묶은 머리) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.\n\nCOMPOSITION VARIATION (this candidate only): keep the place, the people and the moment exactly as specified, and keep from the CAMERA & FRAME contract its subject scale, screen placement and key background elements exactly. For this candidate only, its camera angle and height are deliberately left open: choose a camera position distinctly different from the obvious one, as a film director picking a second setup on the same blocking. Stay true to any explicit framing the SHOT TEXT itself dictates."
  },
  "roll_refs": {
   "A": [
    {
     "label": "SHOT BACKGROUND",
     "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/db00517b-165b-4005-a287-f28936b59b95/images/272bd7af-9294-4490-98c6-88a40e9deacc/scene/recipe/S3sh5__bgfirst_bg.png",
     "asset_id": "e853c50a-c64b-4fda-8cf9-b920ac435aa3",
     "role": "bgfirst_bg"
    },
    {
     "label": "LAYOUT SKETCH (people placement only)",
     "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/db00517b-165b-4005-a287-f28936b59b95/images/272bd7af-9294-4490-98c6-88a40e9deacc/conti/conti_S3sh5.png",
     "asset_id": "3e173d0f-3019-472e-bff1-f56a6ba53c08",
     "role": "conti_light"
    },
    {
     "label": "CHARACTER REFERENCE — 안규홍: the exact person appearing in this shot; match face, hair and build exactly.",
     "path": "<bytes:1028384>",
     "asset_id": "3b5894db-ce55-4f34-b966-5e725c3aa952",
     "role": "character_ref"
    }
   ],
   "B": [
    {
     "label": "LOCATION PHOTOGRAPH — the exact place of this shot: its architecture, materials, fixed features and lighting mood are spatial truth; stage the moment inside this place. Never copy its camera framing.",
     "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/db00517b-165b-4005-a287-f28936b59b95/episodes/272bd7af-9294-4490-98c6-88a40e9deacc/images/background_chain/L08B04.png",
     "asset_id": "71e0793a-2c9b-4326-bdfc-5abf8fffaac8",
     "role": "location_plate"
    },
    {
     "label": "CHARACTER REFERENCE — 안규홍: the exact person appearing in this shot; match face, hair and build exactly.",
     "path": "<bytes:1028384>",
     "asset_id": "3b5894db-ce55-4f34-b966-5e725c3aa952",
     "role": "character_ref"
    }
   ]
  },
  "shot_run_spend_attempt_count": 1,
  "judge_flip_skipped": "judge_fn_owns_order",
  "readings": [
   {
    "label": "A",
    "direction": "두 남자의 시선은 자신들이 들고 있는 볏짐과 나무 들것을 향해 있습니다.",
    "built_space": "참조 이미지와 동일한 흙바닥 마당이며, 기와를 얹은 작은 문과 흙담, 목조 마루가 올바른 위치에 있습니다. 아침 햇살이 적절한 방향에서 비춥니다.",
    "entities": "왼쪽의 안규홍은 참조 이미지의 인물과 얼굴 및 헤어스타일이 일치하나, 옷이 지나치게 깨끗하게 묘사되었습니다. 오른쪽의 장신 일꾼이 함께 서 있으며, 볏짐은 나무 들것 위에 올려진 채 묘사되었습니다.",
    "hard_violations": [],
    "physics": "두 남자가 나무 들것의 양쪽 손잡이를 각각 두 손으로 쥐고 들어 올리고 있으며, 볏짐은 들것에 의해 지탱되어 허공에 안정적으로 떠 있습니다."
   },
   {
    "label": "B",
    "direction": "두 사람의 시선이 양손으로 직접 쥐고 있는 볏짐의 중심부를 향해 있습니다.",
    "built_space": "참조 이미지의 마당을 정확히 반영하여 배경에 기와문과 담장, 툇마루가 보이며, 강한 대비를 이루는 아침 햇빛이 공간을 채우고 있습니다.",
    "entities": "왼쪽의 안규홍은 참조 이미지의 얼굴, 묶은 머리뿐만 아니라 흙 묻은 낡은 복장까지 정확하게 일치합니다. 오른쪽 일꾼은 프롬프트대로 근육질의 팔과 상체가 강조되어 있으며 프레임 의도에 맞게 얼굴이 부분적으로 잘려 있습니다. 볏짐은 밧줄로 묶인 형태로 두 사람 사이에 위치합니다.",
    "hard_violations": [],
    "physics": "두 사람의 양손이 볏짐의 양쪽 끝을 직접 꽉 쥐고 있으며, 팽팽한 팔 근육의 긴장감과 함께 볏짐이 허공에 들려 있습니다. 바닥과 볏짐 사이의 빈 공간이 잘 보입니다."
   }
  ],
  "cross_model_order": {
   "policy": "cross_model_order_v2_slot0_fwd_slot1_rev_parallel2",
   "slots": [
    {
     "model": "gemini-pro",
     "order": "forward",
     "display_to_canonical": {
      "A": "A",
      "B": "B"
     },
     "raw": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 10,
        "verdict_ko": "지시된 미디엄 샷 프레이밍과 상체 중심의 구도를 정확히 구현했으며, 안규홍의 복장 디테일과 팽팽한 팔 근육의 묘사가 프롬프트와 완벽히 일치합니다."
       },
       {
        "label": "A",
        "score": 7,
        "verdict_ko": "배경과 상황은 잘 묘사되었으나, 요구된 미디엄 샷(상체)보다 프레이밍이 넓게 잡혔고 안규홍의 의상에 있는 흙먼지 등 오염된 디테일이 누락되었습니다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "두 남자의 시선은 자신들이 들고 있는 볏짐과 나무 들것을 향해 있습니다.",
        "built_space": "참조 이미지와 동일한 흙바닥 마당이며, 기와를 얹은 작은 문과 흙담, 목조 마루가 올바른 위치에 있습니다. 아침 햇살이 적절한 방향에서 비춥니다.",
        "entities": "왼쪽의 안규홍은 참조 이미지의 인물과 얼굴 및 헤어스타일이 일치하나, 옷이 지나치게 깨끗하게 묘사되었습니다. 오른쪽의 장신 일꾼이 함께 서 있으며, 볏짐은 나무 들것 위에 올려진 채 묘사되었습니다.",
        "hard_violations": [],
        "physics": "두 남자가 나무 들것의 양쪽 손잡이를 각각 두 손으로 쥐고 들어 올리고 있으며, 볏짐은 들것에 의해 지탱되어 허공에 안정적으로 떠 있습니다."
       },
       {
        "label": "B",
        "direction": "두 사람의 시선이 양손으로 직접 쥐고 있는 볏짐의 중심부를 향해 있습니다.",
        "built_space": "참조 이미지의 마당을 정확히 반영하여 배경에 기와문과 담장, 툇마루가 보이며, 강한 대비를 이루는 아침 햇빛이 공간을 채우고 있습니다.",
        "entities": "왼쪽의 안규홍은 참조 이미지의 얼굴, 묶은 머리뿐만 아니라 흙 묻은 낡은 복장까지 정확하게 일치합니다. 오른쪽 일꾼은 프롬프트대로 근육질의 팔과 상체가 강조되어 있으며 프레임 의도에 맞게 얼굴이 부분적으로 잘려 있습니다. 볏짐은 밧줄로 묶인 형태로 두 사람 사이에 위치합니다.",
        "hard_violations": [],
        "physics": "두 사람의 양손이 볏짐의 양쪽 끝을 직접 꽉 쥐고 있으며, 팽팽한 팔 근육의 긴장감과 함께 볏짐이 허공에 들려 있습니다. 바닥과 볏짐 사이의 빈 공간이 잘 보입니다."
       }
      ],
      "all_candidates_fail": false
     },
     "normalized": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 10,
        "verdict_ko": "지시된 미디엄 샷 프레이밍과 상체 중심의 구도를 정확히 구현했으며, 안규홍의 복장 디테일과 팽팽한 팔 근육의 묘사가 프롬프트와 완벽히 일치합니다."
       },
       {
        "label": "A",
        "score": 7,
        "verdict_ko": "배경과 상황은 잘 묘사되었으나, 요구된 미디엄 샷(상체)보다 프레이밍이 넓게 잡혔고 안규홍의 의상에 있는 흙먼지 등 오염된 디테일이 누락되었습니다."
       }
      ],
      "all_candidates_fail": false,
      "readings": [
       {
        "label": "A",
        "direction": "두 남자의 시선은 자신들이 들고 있는 볏짐과 나무 들것을 향해 있습니다.",
        "built_space": "참조 이미지와 동일한 흙바닥 마당이며, 기와를 얹은 작은 문과 흙담, 목조 마루가 올바른 위치에 있습니다. 아침 햇살이 적절한 방향에서 비춥니다.",
        "entities": "왼쪽의 안규홍은 참조 이미지의 인물과 얼굴 및 헤어스타일이 일치하나, 옷이 지나치게 깨끗하게 묘사되었습니다. 오른쪽의 장신 일꾼이 함께 서 있으며, 볏짐은 나무 들것 위에 올려진 채 묘사되었습니다.",
        "hard_violations": [],
        "physics": "두 남자가 나무 들것의 양쪽 손잡이를 각각 두 손으로 쥐고 들어 올리고 있으며, 볏짐은 들것에 의해 지탱되어 허공에 안정적으로 떠 있습니다."
       },
       {
        "label": "B",
        "direction": "두 사람의 시선이 양손으로 직접 쥐고 있는 볏짐의 중심부를 향해 있습니다.",
        "built_space": "참조 이미지의 마당을 정확히 반영하여 배경에 기와문과 담장, 툇마루가 보이며, 강한 대비를 이루는 아침 햇빛이 공간을 채우고 있습니다.",
        "entities": "왼쪽의 안규홍은 참조 이미지의 얼굴, 묶은 머리뿐만 아니라 흙 묻은 낡은 복장까지 정확하게 일치합니다. 오른쪽 일꾼은 프롬프트대로 근육질의 팔과 상체가 강조되어 있으며 프레임 의도에 맞게 얼굴이 부분적으로 잘려 있습니다. 볏짐은 밧줄로 묶인 형태로 두 사람 사이에 위치합니다.",
        "hard_violations": [],
        "physics": "두 사람의 양손이 볏짐의 양쪽 끝을 직접 꽉 쥐고 있으며, 팽팽한 팔 근육의 긴장감과 함께 볏짐이 허공에 들려 있습니다. 바닥과 볏짐 사이의 빈 공간이 잘 보입니다."
       }
      ]
     },
     "ok": true
    },
    {
     "model": "gpt-high",
     "order": "reverse",
     "display_to_canonical": {
      "A": "B",
      "B": "A"
     },
     "raw": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 8,
        "verdict_ko": "볏짐 양끝을 직접 움켜쥔 손과 팽팽한 팔, 상체 중심의 구도가 핵심 순간에 더 충실하지만, 짐이 하단 3분의 1보다 크게 차지하고 오른쪽 인물의 머리가 잘린다."
       },
       {
        "label": "B",
        "score": 6,
        "verdict_ko": "장소와 공중에 든 짐의 지지는 자연스럽지만, 볏짐 모서리가 아닌 운반틀을 들어 올리며 허벅지까지 넓힌 구도와 비교적 편 팔이 요구된 상체 중심의 긴장감을 약화한다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "왼쪽 안규홍은 아래쪽 볏짐과 손 부위를 바라본다. 오른쪽 일꾼은 얼굴 대부분이 잘려 정확한 시선은 확인되지 않지만, 몸과 드러난 얼굴은 짐 쪽으로 숙여져 있다. 두 사람의 손은 가운데 공유한 볏짐의 서로 반대쪽 가장자리를 향해 들어가 있다.",
        "built_space": "뒤 중앙에 기와 대문 한 곳과 이를 받치는 두 기둥, 연결된 돌담이 있고 오른쪽에는 기와 처마 아래 흰 벽과 목재 기둥, 창 한 곳, 이어진 툇마루가 보인다. 참조의 대문·담·오른쪽 건물 배치와 재료가 부합한다. 두 사람은 마당에서 짐을 사이에 두고 서 있으며, 짐 아래 중앙과 양옆으로 흙바닥이 드러난다. 복제된 설비나 반사는 없다.",
        "entities": "남성은 두 명뿐이다. 왼쪽은 청년 한국인 남성 설정에 부합하는 외모이며, 검게 묶은 머리와 때 탄 옅은 저고리·허리띠가 안규홍 참조와 대체로 맞는다. 오른쪽은 굵은 팔과 민소매 작업복을 가진 성인 남성으로, 큰 체격은 보이지만 얼굴과 전신이 잘려 민족적 외모와 정확한 키 차이는 확인하기 어렵다. 실제 벼 이삭과 줄기가 묶인 큰 볏짐이 보인다. 운반틀과 짚신은 화면에서 확인되지 않는다. 읽을 수 있는 글씨나 추가 인물은 없다.",
        "hard_violations": [],
        "physics": "각 인물의 가까운 손이 볏짐의 반대편 가장자리를 아래에서 감싸 쥐어 지지한다. 손가락 접촉과 전완의 긴장이 보이고, 몸을 숙인 자세는 무거운 짐을 들어 버티는 동작으로 가능하다. 짐은 아래 흙바닥과 분리되어 있으며 근거 없이 떠 있지 않다. 발은 화면 밖이고 다리는 아래로 이어진다. 다만 먼쪽 손들은 볏짐에 가려져 양손 모두를 드러내라는 요구는 충족하지 못한다."
       },
       {
        "label": "B",
        "direction": "두 남성 모두 고개와 시선을 가운데 볏짐 및 자신의 손 부위로 내린다. 서로 마주 선 상태에서 손을 운반틀 양끝으로 뻗는다. 작업 대상에 대한 방향은 맞지만, 실제로 움켜쥔 대상은 볏짐 모서리가 아니라 나무 가로대다.",
        "built_space": "뒤 왼쪽에 기와 대문 한 곳과 두 문기둥, 연결된 돌담이 있고, 왼쪽 가장자리에 낮은 평상 한 개가 보인다. 오른쪽에는 창 한 곳과 어두운 문 개구부, 여러 목재 기둥이 놓인 연속된 툇마루가 있다. 참조 장소의 주요 배치와 재료를 유지한다. 두 사람은 흙마당에서 운반틀 양옆에 서 있다. 짐 아래와 주변의 바닥이 넓게 노출되어, 요구된 작은 측면 틈보다 마당의 비중이 크다.",
        "entities": "한국인 남성 설정에 부합하는 외모의 성인 남성 두 명이 보인다. 왼쪽 안규홍은 청년의 얼굴, 검은 묶음머리, 낡은 옅은 저고리와 허리띠로 참조에 대체로 부합한다. 오른쪽 일꾼은 조금 더 높게 서 있고 머리띠와 옅은 작업복을 착용한다. 볏짚 한 짐과 그 아래 밧줄로 결속한 목제 운반틀 한 개가 보인다. 짚신은 화면 밖이다. 추가 인물이나 읽을 수 있는 글씨는 없다.",
        "hard_violations": [],
        "physics": "가까운 두 손이 나무 가로대 양끝을 감싸 쥐고, 볏짐은 그 운반틀 위에 놓여 지지된다. 따라서 짐과 틀이 함께 들린 상태는 물리적으로 가능하며 아래 바닥과의 간격도 보인다. 두 사람의 다리는 아래로 이어지고 몸이 공중에 떠 있지 않다. 다만 팔을 비교적 길게 편 낮은 들기 자세라 번쩍 들어 멈춘 순간의 근육 팽창은 약하다. 먼쪽 손은 짐에 가려 완전한 파지 상태를 확인하기 어렵다."
       }
      ],
      "all_candidates_fail": false
     },
     "normalized": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 8,
        "verdict_ko": "볏짐 양끝을 직접 움켜쥔 손과 팽팽한 팔, 상체 중심의 구도가 핵심 순간에 더 충실하지만, 짐이 하단 3분의 1보다 크게 차지하고 오른쪽 인물의 머리가 잘린다."
       },
       {
        "label": "A",
        "score": 6,
        "verdict_ko": "장소와 공중에 든 짐의 지지는 자연스럽지만, 볏짐 모서리가 아닌 운반틀을 들어 올리며 허벅지까지 넓힌 구도와 비교적 편 팔이 요구된 상체 중심의 긴장감을 약화한다."
       }
      ],
      "all_candidates_fail": false,
      "readings": [
       {
        "label": "B",
        "direction": "왼쪽 안규홍은 아래쪽 볏짐과 손 부위를 바라본다. 오른쪽 일꾼은 얼굴 대부분이 잘려 정확한 시선은 확인되지 않지만, 몸과 드러난 얼굴은 짐 쪽으로 숙여져 있다. 두 사람의 손은 가운데 공유한 볏짐의 서로 반대쪽 가장자리를 향해 들어가 있다.",
        "built_space": "뒤 중앙에 기와 대문 한 곳과 이를 받치는 두 기둥, 연결된 돌담이 있고 오른쪽에는 기와 처마 아래 흰 벽과 목재 기둥, 창 한 곳, 이어진 툇마루가 보인다. 참조의 대문·담·오른쪽 건물 배치와 재료가 부합한다. 두 사람은 마당에서 짐을 사이에 두고 서 있으며, 짐 아래 중앙과 양옆으로 흙바닥이 드러난다. 복제된 설비나 반사는 없다.",
        "entities": "남성은 두 명뿐이다. 왼쪽은 청년 한국인 남성 설정에 부합하는 외모이며, 검게 묶은 머리와 때 탄 옅은 저고리·허리띠가 안규홍 참조와 대체로 맞는다. 오른쪽은 굵은 팔과 민소매 작업복을 가진 성인 남성으로, 큰 체격은 보이지만 얼굴과 전신이 잘려 민족적 외모와 정확한 키 차이는 확인하기 어렵다. 실제 벼 이삭과 줄기가 묶인 큰 볏짐이 보인다. 운반틀과 짚신은 화면에서 확인되지 않는다. 읽을 수 있는 글씨나 추가 인물은 없다.",
        "hard_violations": [],
        "physics": "각 인물의 가까운 손이 볏짐의 반대편 가장자리를 아래에서 감싸 쥐어 지지한다. 손가락 접촉과 전완의 긴장이 보이고, 몸을 숙인 자세는 무거운 짐을 들어 버티는 동작으로 가능하다. 짐은 아래 흙바닥과 분리되어 있으며 근거 없이 떠 있지 않다. 발은 화면 밖이고 다리는 아래로 이어진다. 다만 먼쪽 손들은 볏짐에 가려져 양손 모두를 드러내라는 요구는 충족하지 못한다."
       },
       {
        "label": "A",
        "direction": "두 남성 모두 고개와 시선을 가운데 볏짐 및 자신의 손 부위로 내린다. 서로 마주 선 상태에서 손을 운반틀 양끝으로 뻗는다. 작업 대상에 대한 방향은 맞지만, 실제로 움켜쥔 대상은 볏짐 모서리가 아니라 나무 가로대다.",
        "built_space": "뒤 왼쪽에 기와 대문 한 곳과 두 문기둥, 연결된 돌담이 있고, 왼쪽 가장자리에 낮은 평상 한 개가 보인다. 오른쪽에는 창 한 곳과 어두운 문 개구부, 여러 목재 기둥이 놓인 연속된 툇마루가 있다. 참조 장소의 주요 배치와 재료를 유지한다. 두 사람은 흙마당에서 운반틀 양옆에 서 있다. 짐 아래와 주변의 바닥이 넓게 노출되어, 요구된 작은 측면 틈보다 마당의 비중이 크다.",
        "entities": "한국인 남성 설정에 부합하는 외모의 성인 남성 두 명이 보인다. 왼쪽 안규홍은 청년의 얼굴, 검은 묶음머리, 낡은 옅은 저고리와 허리띠로 참조에 대체로 부합한다. 오른쪽 일꾼은 조금 더 높게 서 있고 머리띠와 옅은 작업복을 착용한다. 볏짚 한 짐과 그 아래 밧줄로 결속한 목제 운반틀 한 개가 보인다. 짚신은 화면 밖이다. 추가 인물이나 읽을 수 있는 글씨는 없다.",
        "hard_violations": [],
        "physics": "가까운 두 손이 나무 가로대 양끝을 감싸 쥐고, 볏짐은 그 운반틀 위에 놓여 지지된다. 따라서 짐과 틀이 함께 들린 상태는 물리적으로 가능하며 아래 바닥과의 간격도 보인다. 두 사람의 다리는 아래로 이어지고 몸이 공중에 떠 있지 않다. 다만 팔을 비교적 길게 편 낮은 들기 자세라 번쩍 들어 멈춘 순간의 근육 팽창은 약하다. 먼쪽 손은 짐에 가려 완전한 파지 상태를 확인하기 어렵다."
       }
      ]
     },
     "ok": true
    }
   ],
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "slot_winner_match": true,
   "slot_winner": {
    "gemini-pro": "B",
    "gpt-high": "B"
   },
   "route": "cross_slot_agree"
  },
  "dual": {
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "normalized": {
    "A": 1.45,
    "B": 2.0
   },
   "adjusted": {
    "A": 1.45,
    "B": 2.0
   },
   "violations": {},
   "per_model_winner": {
    "gemini-pro": "B",
    "gpt-high": "B"
   },
   "agreed": true
  },
  "totals": {
   "B": 2000,
   "A": 1450
  },
  "selected": "B",
  "ranking": [
   "B",
   "A"
  ],
  "verdicts": [
   {
    "label": "B",
    "score": 2000,
    "verdict_ko": "지시된 미디엄 샷 프레이밍과 상체 중심의 구도를 정확히 구현했으며, 안규홍의 복장 디테일과 팽팽한 팔 근육의 묘사가 프롬프트와 완벽히 일치합니다."
   },
   {
    "label": "A",
    "score": 1450,
    "verdict_ko": "배경과 상황은 잘 묘사되었으나, 요구된 미디엄 샷(상체)보다 프레이밍이 넓게 잡혔고 안규홍의 의상에 있는 흙먼지 등 오염된 디테일이 누락되었습니다."
   }
  ],
  "refs": [
   {
    "label": "LOCATION PHOTOGRAPH — the exact place of this shot: its architecture, materials, fixed features and lighting mood are spatial truth; stage the moment inside this place. Never copy its camera framing.",
    "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/db00517b-165b-4005-a287-f28936b59b95/episodes/272bd7af-9294-4490-98c6-88a40e9deacc/images/background_chain/L08B04.png",
    "asset_id": "71e0793a-2c9b-4326-bdfc-5abf8fffaac8",
    "role": "location_plate"
   },
   {
    "label": "CHARACTER REFERENCE — 안규홍: the exact person appearing in this shot; match face, hair and build exactly.",
    "path": "<bytes:1028384>",
    "asset_id": "3b5894db-ce55-4f34-b966-5e725c3aa952",
    "role": "character_ref"
   }
  ],
  "critique_skipped": true,
  "shot_run_produced": false,
  "shot_run_uid": "06aa1403-a8bb-7278-bf74-c45932ba1189",
  "bgfirst": {
   "bg_path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/db00517b-165b-4005-a287-f28936b59b95/images/272bd7af-9294-4490-98c6-88a40e9deacc/scene/recipe/S3sh5__bgfirst_bg.png",
   "bg_asset_id": "e853c50a-c64b-4fda-8cf9-b920ac435aa3",
   "bg_record_key": "S3sh5::bgfirst_bg",
   "chain_winner": false,
   "authority": "plate"
  },
  "ref_mode": "플레이트+엔티티 (2택1: 무콘티 승)",
  "share_plan": {
   "ref_plan": "background"
  }
 },
 "S3sh5::cine": {
  "applied": true,
  "attempted_at": "2026-09-09T00:49:18.766448+00:00",
  "fingerprint": "3c68dd58a2aa9cb2b88e4f92842412b34510cb2fcd55c653cf178a25c765e322",
  "fingerprint_version": 2,
  "provider": "mai",
  "endpoint": "openrouter/chat-completions",
  "model": "microsoft/mai-image-2.6",
  "pack": "24.202608252115",
  "source_file": "S3sh5_sel.png",
  "source_sha256": "d15716a4004f070c6c9904523c26052e39234a40bc06e2efc57fa38999588911",
  "file": "S3sh5_cine.png",
  "staged_sha256": "150f6a965da49ff15091b241a2d624966f02f95ee10a44076f27fc8f45e09904",
  "latency_ms": 27976
 },
 "S3sh18::signage": {
  "fp": "f99392fa71b65903",
  "inscriptions": [],
  "cues": [],
  "dropped": []
 },
 "S3sh18::bgfirst_bg": {
  "input_fingerprint": "002d52e50e34b94c",
  "prompt": "Create the EMPTY BACKGROUND PLATE for one film shot — NO PEOPLE, no figures, no body parts, no sketch lines, no arrows anywhere.\n\"Empty\" means no people only: KEEP the location's inherent occupants and stock that define the place — animals in an animal shelter, pen or farm, goods in a market, moored boats in a harbour — unless the shot text explicitly removes them.\nThe FIRST attached image is a thin-line storyboard sketch: use ONLY its camera angle, horizon, perspective and the placement/size of buildings and set masses — ignore the sketched people and arrows entirely. The SECOND attached image (LOCATION PHOTOGRAPH) is the real place: take its architecture, materials, signage and fixed features, and RE-PROJECT them into the sketch's camera. If the photograph's camera differs from the sketch's, the sketch's camera wins.\nHUMAN-SCALE CALIBRATION: derive every structure's true size from human-scale elements — a door ≈ 2m, a window ≈ 1–1.5m wide, one storey ≈ 2.5–3m; never inflate a small structure or shrink a large one.\n\nSHOT TEXT this background must serve (Korean): 대문 쪽을 향해 걷는 중인 mid-stride 자세의 안규홍 뒷모습을 노려보며 꼿꼿이 선 박진성의 전신.\n\nLOCATION (lock): Inside the estate's guest-wing room beside its open doorway, with morning daylight entering from the yard and a view toward the main gate.\n\nTIME OF DAY (lock): morning, harsh sunlight.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: wide shot\n- FRAME LAYOUT: interior area beside the doorway occupied by 박진성 in the middle-left of the frame, midground; open doorway framing the departing 안규홍 in the middle-right of the frame, background.\n- KEY BACKGROUND ELEMENTS: 사랑채 출입문 (Open onto the yard) — Seen from inside at an oblique angle, with the departing 안규홍 visible through the opening; used as Separates the rigid interior figure from the receding exterior figure without breaking their shared spatial relationship; 안규홍의 지게 (Carried on his back as he leaves) — Its rear is visible above and beside his back; used as Maintains his working silhouette at a natural scale within the distant doorway.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Retain the strong morning daylight outside and naturally subdued interior exposure, preserving readable figures without artificial dramatic color separation.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nRender ONE photorealistic empty location photograph, 16:9, that this shot can be staged inside later. No readable writing anywhere: surfaces that would carry writing may be present, but stage any wording out of legibility — an oblique angle, distance, shallow focus. No captions, watermarks or overlay text.\n\nThe THIRD attached image (STRUCTURE LOOK) is the identity source of the fixed structure at this location: its faces, openings, levels, materials and signage are truth. Where it conflicts with the LOCATION PHOTOGRAPH about the structure itself, the STRUCTURE LOOK wins; the photograph still governs the surroundings, time of day and lighting.",
  "effective_prompt": "Create the EMPTY BACKGROUND PLATE for one film shot — NO PEOPLE, no figures, no body parts, no sketch lines, no arrows anywhere.\n\"Empty\" means no people only: KEEP the location's inherent occupants and stock that define the place — animals in an animal shelter, pen or farm, goods in a market, moored boats in a harbour — unless the shot text explicitly removes them.\nThe FIRST attached image is a thin-line storyboard sketch: use ONLY its camera angle, horizon, perspective and the placement/size of buildings and set masses — ignore the sketched people and arrows entirely. The SECOND attached image (LOCATION PHOTOGRAPH) is the real place: take its architecture, materials, signage and fixed features, and RE-PROJECT them into the sketch's camera. If the photograph's camera differs from the sketch's, the sketch's camera wins.\nHUMAN-SCALE CALIBRATION: derive every structure's true size from human-scale elements — a door ≈ 2m, a window ≈ 1–1.5m wide, one storey ≈ 2.5–3m; never inflate a small structure or shrink a large one.\n\nSHOT TEXT this background must serve (Korean): 대문 쪽을 향해 걷는 중인 mid-stride 자세의 안규홍 뒷모습을 노려보며 꼿꼿이 선 박진성의 전신.\n\nLOCATION (lock): Inside the estate's guest-wing room beside its open doorway, with morning daylight entering from the yard and a view toward the main gate.\n\nTIME OF DAY (lock): morning, harsh sunlight.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: wide shot\n- FRAME LAYOUT: interior area beside the doorway occupied by 박진성 in the middle-left of the frame, midground; open doorway framing the departing 안규홍 in the middle-right of the frame, background.\n- KEY BACKGROUND ELEMENTS: 사랑채 출입문 (Open onto the yard) — Seen from inside at an oblique angle, with the departing 안규홍 visible through the opening; used as Separates the rigid interior figure from the receding exterior figure without breaking their shared spatial relationship; 안규홍의 지게 (Carried on his back as he leaves) — Its rear is visible above and beside his back; used as Maintains his working silhouette at a natural scale within the distant doorway.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Retain the strong morning daylight outside and naturally subdued interior exposure, preserving readable figures without artificial dramatic color separation.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nRender ONE photorealistic empty location photograph, 16:9, that this shot can be staged inside later. No readable writing anywhere: surfaces that would carry writing may be present, but stage any wording out of legibility — an oblique angle, distance, shallow focus. No captions, watermarks or overlay text.\n\nThe THIRD attached image (STRUCTURE LOOK) is the identity source of the fixed structure at this location: its faces, openings, levels, materials and signage are truth. Where it conflicts with the LOCATION PHOTOGRAPH about the structure itself, the STRUCTURE LOOK wins; the photograph still governs the surroundings, time of day and lighting.",
  "bg_path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/db00517b-165b-4005-a287-f28936b59b95/images/272bd7af-9294-4490-98c6-88a40e9deacc/scene/recipe/S3sh18__bgfirst_bg.png",
  "asset_id": "cb877610-afef-4846-9ca9-416de5702a10",
  "input_asset_ids": [
   "f536a455-5400-48c3-b56e-7a44b413639e",
   "71e0793a-2c9b-4326-bdfc-5abf8fffaac8",
   "0d12e89a-d0a7-4e88-a172-83d7a9b982f6"
  ]
 },
 "S3sh18": {
  "input_fingerprint": "b602d1e35cc86233",
  "prompt": "Create ONE FINAL photorealistic live-action film still of the moment below — early 20th-century Korea, Jeolla region; all people are Korean unless explicitly stated otherwise. TIME OF DAY (lock): morning, harsh sunlight.\n\nSHOT TEXT (authoritative, Korean): 대문 쪽을 향해 걷는 중인 mid-stride 자세의 안규홍 뒷모습을 노려보며 꼿꼿이 선 박진성의 전신.\n\nLOCATION (lock): Inside the estate's guest-wing room beside its open doorway, with morning daylight entering from the yard and a view toward the main gate. The shot takes place here — the attached LOCATION PHOTOGRAPH shows the exact spot.\n\nSTRUCTURE LOOK AUTHORITY: the attached STRUCTURE LOOK photograph is the identity of the fixed structure at this location — wherever that structure appears in the frame, its shape, proportions, openings, materials and colors are LOCKED to it. The LOCATION PHOTOGRAPH remains the authority for this shot's sub-space, surroundings, time of day and lighting. If the two conflict on the structure itself, the STRUCTURE LOOK photo wins; for everything else, the LOCATION PHOTOGRAPH wins.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: wide shot\n- FRAME LAYOUT: interior area beside the doorway occupied by 박진성 in the middle-left of the frame, midground; open doorway framing the departing 안규홍 in the middle-right of the frame, background.\n- KEY BACKGROUND ELEMENTS: 사랑채 출입문 (Open onto the yard) — Seen from inside at an oblique angle, with the departing 안규홍 visible through the opening; used as Separates the rigid interior figure from the receding exterior figure without breaking their shared spatial relationship; 안규홍의 지게 (Carried on his back as he leaves) — Its rear is visible above and beside his back; used as Maintains his working silhouette at a natural scale within the distant doorway.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Retain the strong morning daylight outside and naturally subdued interior exposure, preserving readable figures without artificial dramatic color separation.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The sarangchae door remains open; a drinking table stands inside with a cup set down upon it. 안규홍: He walks toward the gate carrying his frame, with the small jade-colored cloth bundle tucked into it and worn straw shoes on his feet. 박진성: He remains standing after rising abruptly.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 안규홍 (한국인 남성, 청년 나이대, 검은 머리, 단단히 묶은 머리); 박진성 (한국인 남성, 성인 나이대, 자연스러운 검은 머리) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.",
  "roll_prompts": {
   "A": "Stage the shot. The FIRST attached image (SHOT BACKGROUND) is the finished empty background of this shot — keep it EXACTLY: its camera, perspective, architecture, lighting and every fixed feature stay untouched. The SECOND attached image (LAYOUT SKETCH) tells you ONLY where the people go: each sketched person's position, screen size, pose and the gaze/motion arrows. Ignore the sketch's background lines. The CHARACTER REFERENCE photographs show the real people.\nPlace the real people into the background at exactly the sketched positions, sizes and poses, following the arrow directions. No sketch lines or arrows may remain.\n\nCreate ONE FINAL photorealistic live-action film still of the moment below — early 20th-century Korea, Jeolla region; all people are Korean unless explicitly stated otherwise. TIME OF DAY (lock): morning, harsh sunlight.\n\nSHOT TEXT (authoritative, Korean): 대문 쪽을 향해 걷는 중인 mid-stride 자세의 안규홍 뒷모습을 노려보며 꼿꼿이 선 박진성의 전신.\n\nLOCATION (lock): Inside the estate's guest-wing room beside its open doorway, with morning daylight entering from the yard and a view toward the main gate. The shot takes place here — the FIRST attached image (SHOT BACKGROUND) is this exact place, already built: its ground, structures, horizon, materials and lighting are the finished truth of this location and must not be redesigned or replaced. No location photograph is attached — read the place from that image alone, and add no scenery, structure, vehicle or fixture that it does not already show. This lock governs the place only; the figures in the shot follow the staging and pose instructions.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: wide shot\n- FRAME LAYOUT: interior area beside the doorway occupied by 박진성 in the middle-left of the frame, midground; open doorway framing the departing 안규홍 in the middle-right of the frame, background.\n- KEY BACKGROUND ELEMENTS: 사랑채 출입문 (Open onto the yard) — Seen from inside at an oblique angle, with the departing 안규홍 visible through the opening; used as Separates the rigid interior figure from the receding exterior figure without breaking their shared spatial relationship; 안규홍의 지게 (Carried on his back as he leaves) — Its rear is visible above and beside his back; used as Maintains his working silhouette at a natural scale within the distant doorway.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Retain the strong morning daylight outside and naturally subdued interior exposure, preserving readable figures without artificial dramatic color separation.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The sarangchae door remains open; a drinking table stands inside with a cup set down upon it. 안규홍: He walks toward the gate carrying his frame, with the small jade-colored cloth bundle tucked into it and worn straw shoes on his feet. 박진성: He remains standing after rising abruptly.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 안규홍 (한국인 남성, 청년 나이대, 검은 머리, 단단히 묶은 머리); 박진성 (한국인 남성, 성인 나이대, 자연스러운 검은 머리) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.",
   "B": "Create ONE FINAL photorealistic live-action film still of the moment below — early 20th-century Korea, Jeolla region; all people are Korean unless explicitly stated otherwise. TIME OF DAY (lock): morning, harsh sunlight.\n\nSHOT TEXT (authoritative, Korean): 대문 쪽을 향해 걷는 중인 mid-stride 자세의 안규홍 뒷모습을 노려보며 꼿꼿이 선 박진성의 전신.\n\nLOCATION (lock): Inside the estate's guest-wing room beside its open doorway, with morning daylight entering from the yard and a view toward the main gate. The shot takes place here — the attached LOCATION PHOTOGRAPH shows the exact spot.\n\nSTRUCTURE LOOK AUTHORITY: the attached STRUCTURE LOOK photograph is the identity of the fixed structure at this location — wherever that structure appears in the frame, its shape, proportions, openings, materials and colors are LOCKED to it. The LOCATION PHOTOGRAPH remains the authority for this shot's sub-space, surroundings, time of day and lighting. If the two conflict on the structure itself, the STRUCTURE LOOK photo wins; for everything else, the LOCATION PHOTOGRAPH wins.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: wide shot\n- FRAME LAYOUT: interior area beside the doorway occupied by 박진성 in the middle-left of the frame, midground; open doorway framing the departing 안규홍 in the middle-right of the frame, background.\n- KEY BACKGROUND ELEMENTS: 사랑채 출입문 (Open onto the yard) — Seen from inside at an oblique angle, with the departing 안규홍 visible through the opening; used as Separates the rigid interior figure from the receding exterior figure without breaking their shared spatial relationship; 안규홍의 지게 (Carried on his back as he leaves) — Its rear is visible above and beside his back; used as Maintains his working silhouette at a natural scale within the distant doorway.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Retain the strong morning daylight outside and naturally subdued interior exposure, preserving readable figures without artificial dramatic color separation.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The sarangchae door remains open; a drinking table stands inside with a cup set down upon it. 안규홍: He walks toward the gate carrying his frame, with the small jade-colored cloth bundle tucked into it and worn straw shoes on his feet. 박진성: He remains standing after rising abruptly.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 안규홍 (한국인 남성, 청년 나이대, 검은 머리, 단단히 묶은 머리); 박진성 (한국인 남성, 성인 나이대, 자연스러운 검은 머리) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.\n\nCOMPOSITION VARIATION (this candidate only): keep the place, the people and the moment exactly as specified, and keep from the CAMERA & FRAME contract its subject scale, screen placement and key background elements exactly. For this candidate only, its camera angle and height are deliberately left open: choose a camera position distinctly different from the obvious one, as a film director picking a second setup on the same blocking. Stay true to any explicit framing the SHOT TEXT itself dictates."
  },
  "roll_refs": {
   "A": [
    {
     "label": "SHOT BACKGROUND",
     "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/db00517b-165b-4005-a287-f28936b59b95/images/272bd7af-9294-4490-98c6-88a40e9deacc/scene/recipe/S3sh18__bgfirst_bg.png",
     "asset_id": "cb877610-afef-4846-9ca9-416de5702a10",
     "role": "bgfirst_bg"
    },
    {
     "label": "LAYOUT SKETCH (people placement only)",
     "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/db00517b-165b-4005-a287-f28936b59b95/images/272bd7af-9294-4490-98c6-88a40e9deacc/conti/conti_S3sh18.png",
     "asset_id": "f536a455-5400-48c3-b56e-7a44b413639e",
     "role": "conti_light"
    },
    {
     "label": "CHARACTER REFERENCE — 안규홍: the exact person appearing in this shot; match face, hair and build exactly.",
     "path": "<bytes:1028384>",
     "asset_id": "3b5894db-ce55-4f34-b966-5e725c3aa952",
     "role": "character_ref"
    },
    {
     "label": "CHARACTER REFERENCE — 박진성: the exact person appearing in this shot; match face, hair and build exactly.",
     "path": "<bytes:912297>",
     "asset_id": "3690b9ae-055b-43ee-9b74-794f83fca3d5",
     "role": "character_ref"
    }
   ],
   "B": [
    {
     "label": "LOCATION PHOTOGRAPH — the exact place of this shot: its spatial layout, surroundings, fixed features, time of day and lighting mood are spatial truth; stage the moment inside this place. If a STRUCTURE LOOK photograph is also attached, that photo wins for the fixed structure itself — this photograph wins for everything around it. Never copy its camera framing.",
     "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/db00517b-165b-4005-a287-f28936b59b95/episodes/272bd7af-9294-4490-98c6-88a40e9deacc/images/background_chain/L08B04.png",
     "asset_id": "71e0793a-2c9b-4326-bdfc-5abf8fffaac8",
     "role": "location_plate"
    },
    {
     "label": "STRUCTURE LOOK — the confirmed photograph of the fixed structure at this location: wherever the structure appears in the frame, its shape, proportions, materials, colors and openings are LOCKED to this photo. Never copy its camera framing, time of day or lighting — the shot text and the LOCATION PHOTOGRAPH are the authorities for those.",
     "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/db00517b-165b-4005-a287-f28936b59b95/images/background_chain/seed_bg_tiled_house_sel.png",
     "asset_id": "0d12e89a-d0a7-4e88-a172-83d7a9b982f6",
     "role": "structure_seed_look"
    },
    {
     "label": "CHARACTER REFERENCE — 안규홍: the exact person appearing in this shot; match face, hair and build exactly.",
     "path": "<bytes:1028384>",
     "asset_id": "3b5894db-ce55-4f34-b966-5e725c3aa952",
     "role": "character_ref"
    },
    {
     "label": "CHARACTER REFERENCE — 박진성: the exact person appearing in this shot; match face, hair and build exactly.",
     "path": "<bytes:912297>",
     "asset_id": "3690b9ae-055b-43ee-9b74-794f83fca3d5",
     "role": "character_ref"
    }
   ]
  },
  "shot_run_spend_attempt_count": 1,
  "judge_flip_skipped": "judge_fn_owns_order",
  "readings": [
   {
    "label": "A",
    "direction": "박진성의 시선과 몸의 방향이 열린 문을 통해 멀어지는 안규홍을 향하고 있습니다. 안규홍은 마당 너머의 대문을 타겟으로 삼아 똑바로 걷고 있습니다.",
    "built_space": "나무 바닥으로 된 실내의 좌측에 나무 탁자와 술잔이 하나 놓여 있습니다. 바깥으로 열린 2짝의 나무 문이 안규홍의 뒷모습을 프레이밍하고 있으며, 마당 너머 기와를 얹은 대문과 흙담이 기준 사진의 구조와 유사한 거리감으로 배치되어 있습니다.",
    "entities": "박진성은 참조 이미지와 일치하는 푸른색 도포와 갓을 착용한 성인 남성으로 묘사되었습니다. 안규홍은 흰색 옷과 짚신을 착용하고 있으며, 묶은 머리가 보이고 등에 지게와 둥근 옥색 보따리를 매달고 있습니다. 두 인물 모두 프롬프트가 요구한 외형과 일치합니다.",
    "hard_violations": [],
    "physics": "안규홍은 왼발을 땅에 단단히 딛고 오른발을 들어 올린 걷는 자세(mid-stride)를 물리적으로 자연스럽게 취하고 있으며, 지게는 그의 어깨와 등에 정상적으로 지지되어 있습니다. 박진성 역시 나무 바닥 위에 두 발로 안정적으로 서 있으며, 탁자 위의 술잔도 바닥면에 정상적으로 안착해 있습니다."
   },
   {
    "label": "B",
    "direction": "박진성이 문밖의 안규홍을 응시하고 있으며, 안규홍은 마당을 가로질러 대문을 향해 걷고 있습니다.",
    "built_space": "실내 바닥이 흙바닥으로 묘사되었으며 일반적인 한옥 사랑채에 맞지 않는 두껍고 투박한 성곽용 나무 문이 달려 있어 기준 공간과 이질적입니다. 좌측에 소반과 술잔이 놓여 있고, 바깥 마당의 대문은 공간의 깊이감에 비해 너무 가깝고 거대하게 렌더링되었습니다.",
    "entities": "박진성의 푸른 한복과 갓, 안규홍의 흰 옷과 짚신, 지게, 옥색 보따리 모두 존재하며 참조된 인물의 특성을 가지고 있습니다.",
    "hard_violations": [],
    "physics": "안규홍은 왼발을 땅에 딛고 걷는 포즈를 취하고 있으며 지게를 등에 메고 있습니다. 박진성은 바닥에 서 있으며 특별히 지지되지 않고 공중에 떠 있는 물체나 인물은 없습니다."
   }
  ],
  "cross_model_order": {
   "policy": "cross_model_order_v2_slot0_fwd_slot1_rev_parallel2",
   "slots": [
    {
     "model": "gemini-pro",
     "order": "forward",
     "display_to_canonical": {
      "A": "A",
      "B": "B"
     },
     "raw": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 9,
        "verdict_ko": "요구된 화면 구도(좌측 박진성, 우측 문밖 안규홍)를 완벽하게 구현하였으며, 안규홍의 걷는 자세와 지게, 옥색 보따리, 술상 등의 세부 요소를 지시문과 참조에 맞게 충실히 렌더링한 우수한 결과물입니다."
       },
       {
        "label": "B",
        "score": 4,
        "verdict_ko": "실내 바닥을 흙바닥으로 묘사하고 육중한 문을 달아 참조된 장소의 건축 양식(Location lock)을 크게 훼손하였으며, 박진성을 화면의 좌측 끝으로 밀어내어 요구된 프레이밍 구도를 지키지 못했습니다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "박진성의 시선과 몸의 방향이 열린 문을 통해 멀어지는 안규홍을 향하고 있습니다. 안규홍은 마당 너머의 대문을 타겟으로 삼아 똑바로 걷고 있습니다.",
        "built_space": "나무 바닥으로 된 실내의 좌측에 나무 탁자와 술잔이 하나 놓여 있습니다. 바깥으로 열린 2짝의 나무 문이 안규홍의 뒷모습을 프레이밍하고 있으며, 마당 너머 기와를 얹은 대문과 흙담이 기준 사진의 구조와 유사한 거리감으로 배치되어 있습니다.",
        "entities": "박진성은 참조 이미지와 일치하는 푸른색 도포와 갓을 착용한 성인 남성으로 묘사되었습니다. 안규홍은 흰색 옷과 짚신을 착용하고 있으며, 묶은 머리가 보이고 등에 지게와 둥근 옥색 보따리를 매달고 있습니다. 두 인물 모두 프롬프트가 요구한 외형과 일치합니다.",
        "hard_violations": [],
        "physics": "안규홍은 왼발을 땅에 단단히 딛고 오른발을 들어 올린 걷는 자세(mid-stride)를 물리적으로 자연스럽게 취하고 있으며, 지게는 그의 어깨와 등에 정상적으로 지지되어 있습니다. 박진성 역시 나무 바닥 위에 두 발로 안정적으로 서 있으며, 탁자 위의 술잔도 바닥면에 정상적으로 안착해 있습니다."
       },
       {
        "label": "B",
        "direction": "박진성이 문밖의 안규홍을 응시하고 있으며, 안규홍은 마당을 가로질러 대문을 향해 걷고 있습니다.",
        "built_space": "실내 바닥이 흙바닥으로 묘사되었으며 일반적인 한옥 사랑채에 맞지 않는 두껍고 투박한 성곽용 나무 문이 달려 있어 기준 공간과 이질적입니다. 좌측에 소반과 술잔이 놓여 있고, 바깥 마당의 대문은 공간의 깊이감에 비해 너무 가깝고 거대하게 렌더링되었습니다.",
        "entities": "박진성의 푸른 한복과 갓, 안규홍의 흰 옷과 짚신, 지게, 옥색 보따리 모두 존재하며 참조된 인물의 특성을 가지고 있습니다.",
        "hard_violations": [],
        "physics": "안규홍은 왼발을 땅에 딛고 걷는 포즈를 취하고 있으며 지게를 등에 메고 있습니다. 박진성은 바닥에 서 있으며 특별히 지지되지 않고 공중에 떠 있는 물체나 인물은 없습니다."
       }
      ],
      "all_candidates_fail": false
     },
     "normalized": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 9,
        "verdict_ko": "요구된 화면 구도(좌측 박진성, 우측 문밖 안규홍)를 완벽하게 구현하였으며, 안규홍의 걷는 자세와 지게, 옥색 보따리, 술상 등의 세부 요소를 지시문과 참조에 맞게 충실히 렌더링한 우수한 결과물입니다."
       },
       {
        "label": "B",
        "score": 4,
        "verdict_ko": "실내 바닥을 흙바닥으로 묘사하고 육중한 문을 달아 참조된 장소의 건축 양식(Location lock)을 크게 훼손하였으며, 박진성을 화면의 좌측 끝으로 밀어내어 요구된 프레이밍 구도를 지키지 못했습니다."
       }
      ],
      "all_candidates_fail": false,
      "readings": [
       {
        "label": "A",
        "direction": "박진성의 시선과 몸의 방향이 열린 문을 통해 멀어지는 안규홍을 향하고 있습니다. 안규홍은 마당 너머의 대문을 타겟으로 삼아 똑바로 걷고 있습니다.",
        "built_space": "나무 바닥으로 된 실내의 좌측에 나무 탁자와 술잔이 하나 놓여 있습니다. 바깥으로 열린 2짝의 나무 문이 안규홍의 뒷모습을 프레이밍하고 있으며, 마당 너머 기와를 얹은 대문과 흙담이 기준 사진의 구조와 유사한 거리감으로 배치되어 있습니다.",
        "entities": "박진성은 참조 이미지와 일치하는 푸른색 도포와 갓을 착용한 성인 남성으로 묘사되었습니다. 안규홍은 흰색 옷과 짚신을 착용하고 있으며, 묶은 머리가 보이고 등에 지게와 둥근 옥색 보따리를 매달고 있습니다. 두 인물 모두 프롬프트가 요구한 외형과 일치합니다.",
        "hard_violations": [],
        "physics": "안규홍은 왼발을 땅에 단단히 딛고 오른발을 들어 올린 걷는 자세(mid-stride)를 물리적으로 자연스럽게 취하고 있으며, 지게는 그의 어깨와 등에 정상적으로 지지되어 있습니다. 박진성 역시 나무 바닥 위에 두 발로 안정적으로 서 있으며, 탁자 위의 술잔도 바닥면에 정상적으로 안착해 있습니다."
       },
       {
        "label": "B",
        "direction": "박진성이 문밖의 안규홍을 응시하고 있으며, 안규홍은 마당을 가로질러 대문을 향해 걷고 있습니다.",
        "built_space": "실내 바닥이 흙바닥으로 묘사되었으며 일반적인 한옥 사랑채에 맞지 않는 두껍고 투박한 성곽용 나무 문이 달려 있어 기준 공간과 이질적입니다. 좌측에 소반과 술잔이 놓여 있고, 바깥 마당의 대문은 공간의 깊이감에 비해 너무 가깝고 거대하게 렌더링되었습니다.",
        "entities": "박진성의 푸른 한복과 갓, 안규홍의 흰 옷과 짚신, 지게, 옥색 보따리 모두 존재하며 참조된 인물의 특성을 가지고 있습니다.",
        "hard_violations": [],
        "physics": "안규홍은 왼발을 땅에 딛고 걷는 포즈를 취하고 있으며 지게를 등에 메고 있습니다. 박진성은 바닥에 서 있으며 특별히 지지되지 않고 공중에 떠 있는 물체나 인물은 없습니다."
       }
      ]
     },
     "ok": true
    },
    {
     "model": "gpt-high",
     "order": "reverse",
     "display_to_canonical": {
      "A": "B",
      "B": "A"
     },
     "raw": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 5,
        "verdict_ko": "전신과 퇴장 동작은 맞지만, 안규홍이 배경 인물치고 크게 잡히고 흙바닥·문턱 공간과 기와 대문이 장소 및 구조 기준에서 벗어난다."
       },
       {
        "label": "B",
        "score": 7,
        "verdict_ko": "중앙 왼쪽에 선 박진성과 오른쪽 먼 문밖으로 걷는 안규홍의 크기·시선·공간 관계가 더 정확하지만, 기와 대문은 구조 기준의 초가 대문과 다르다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "박진성은 오른쪽으로 얼굴을 돌려 문밖 안규홍 쪽을 바라본다. 안규홍은 카메라에 등을 보이며 앞쪽 대문을 향해 걷는다. 지게의 뒤쪽과 옥색 보따리가 카메라를 향한다. 박진성의 시선 방향은 맞지만 강하게 노려보는 표정은 뚜렷하지 않다.",
        "built_space": "중앙 오른쪽에 열린 출입구 하나와 나무 문짝 하나, 바깥에 대문 하나가 보인다. 실내에는 잔 하나를 올린 작은 상 하나가 있고 왼쪽에는 다른 낮은 목제 가구 일부가 보인다. 박진성은 문 왼쪽 실내에, 안규홍은 문밖 마당에 있다. 다만 실내가 흙바닥이고 낮은 문턱이 곧바로 마당에 이어져, 참고 사진의 높인 목조 마루와 툇마루 관계를 충분히 재현하지 않는다. 바깥 대문과 담장은 장소 사진의 기와·돌 형태를 따르지만, 우선권이 있는 구조 사진의 초가 대문·황토 담장과 다르다.",
        "entities": "인물은 두 명뿐이다. 박진성은 성인 한국인 남성 설정과 부합하는 외형이며 검은 갓, 갓끈, 하늘색 도포, 흰 바지와 검정·흰 신발을 착용한다. 옆얼굴은 보이지만 정면 기준과의 정확한 얼굴 일치는 제한적으로만 확인된다. 안규홍은 검은 머리를 묶은 청년 남성의 뒷모습으로, 때 탄 밝은 작업복과 짚신을 착용한다. 나무 지게와 옥색 보따리가 보이며 지게 하단에는 추가 짚더미도 실려 있다. 상 위 잔은 확인되고 읽을 수 있는 글자는 없다.",
        "hard_violations": [],
        "physics": "박진성은 두 발을 바닥에 붙이고 곧게 서 있다. 안규홍은 한쪽 발로 땅을 딛고 반대쪽 발을 뒤로 들어 보행 중간 동작을 만든다. 지게는 등에 밀착되어 어깨 쪽으로 걸쳐져 있고, 보따리와 짚은 지게 가로대에 얹혀 있다. 상은 다리로 바닥에 서며 잔은 상판에 놓여 있다. 명백하게 지지 없이 떠 있는 몸이나 물체는 없다."
       },
       {
        "label": "B",
        "direction": "박진성의 몸과 고개는 오른쪽 문밖 안규홍을 향한다. 안규홍은 등을 보이고 대문 쪽으로 멀어지는 방향으로 걷는다. 지게의 뒷면과 옥색 보따리가 관객에게 보인다. 시선 대상은 맞지만 박진성의 얼굴이 뒤쪽 각도로 보여 노려보는 눈빛 자체는 확인하기 어렵다.",
        "built_space": "오른쪽에 큰 출입구 하나가 있고, 왼쪽으로 열린 문짝 하나와 오른쪽의 좁은 고정 구획이 보인다. 실내 왼쪽에는 낮은 상 하나와 잔 하나, 마당 오른쪽에는 평상 일부, 먼 배경에는 대문 하나와 담장이 있다. 박진성은 목조 실내 바닥의 중앙 왼쪽에 전신으로 서 있고, 안규홍은 중앙 오른쪽 문틀 안의 먼 마당에 작게 배치된다. 높인 나무 바닥과 마당 사이 단차가 장소 사진의 목조 공간에 더 가깝다. 다만 대문은 기와지붕이며 담장 상단도 구조 사진의 초가 형태와 다르다.",
        "entities": "추가 인물 없이 두 남성만 보인다. 박진성은 성인 한국인 남성 설정에 부합하는 외형이고 검은 갓, 하늘색 도포, 흰 바지와 검정·흰 신발이 참고 복장과 맞는다. 얼굴은 대부분 가려져 정확한 동일 인물 여부는 판별하기 어렵다. 안규홍은 묶은 검은 머리, 밝고 낡은 작업복, 짚신을 착용한 청년 남성의 뒷모습이다. 등에는 자연스러운 크기의 나무 지게가 있고 아래쪽 가로대 부근에 작은 옥색 천 보따리가 들어 있다. 실내 상과 내려놓은 잔도 확인되며 읽을 수 있는 글자는 없다.",
        "hard_violations": [],
        "physics": "박진성의 두 발은 나무 바닥에 안정적으로 닿아 있으며 상체는 꼿꼿하다. 안규홍은 오른쪽 발로 마당을 딛고 왼쪽 발을 뒤로 들어 실제 걸음의 중간 순간을 보인다. 지게는 어깨와 등에 걸려 있고 보따리는 지게 하단에 받쳐져 있다. 상과 평상은 다리로 지면에 지지되며 잔은 상판 위에 놓인다. 지지 없는 부유나 불가능한 관절 자세는 보이지 않는다."
       }
      ],
      "all_candidates_fail": true
     },
     "normalized": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 5,
        "verdict_ko": "전신과 퇴장 동작은 맞지만, 안규홍이 배경 인물치고 크게 잡히고 흙바닥·문턱 공간과 기와 대문이 장소 및 구조 기준에서 벗어난다."
       },
       {
        "label": "A",
        "score": 7,
        "verdict_ko": "중앙 왼쪽에 선 박진성과 오른쪽 먼 문밖으로 걷는 안규홍의 크기·시선·공간 관계가 더 정확하지만, 기와 대문은 구조 기준의 초가 대문과 다르다."
       }
      ],
      "all_candidates_fail": true,
      "readings": [
       {
        "label": "B",
        "direction": "박진성은 오른쪽으로 얼굴을 돌려 문밖 안규홍 쪽을 바라본다. 안규홍은 카메라에 등을 보이며 앞쪽 대문을 향해 걷는다. 지게의 뒤쪽과 옥색 보따리가 카메라를 향한다. 박진성의 시선 방향은 맞지만 강하게 노려보는 표정은 뚜렷하지 않다.",
        "built_space": "중앙 오른쪽에 열린 출입구 하나와 나무 문짝 하나, 바깥에 대문 하나가 보인다. 실내에는 잔 하나를 올린 작은 상 하나가 있고 왼쪽에는 다른 낮은 목제 가구 일부가 보인다. 박진성은 문 왼쪽 실내에, 안규홍은 문밖 마당에 있다. 다만 실내가 흙바닥이고 낮은 문턱이 곧바로 마당에 이어져, 참고 사진의 높인 목조 마루와 툇마루 관계를 충분히 재현하지 않는다. 바깥 대문과 담장은 장소 사진의 기와·돌 형태를 따르지만, 우선권이 있는 구조 사진의 초가 대문·황토 담장과 다르다.",
        "entities": "인물은 두 명뿐이다. 박진성은 성인 한국인 남성 설정과 부합하는 외형이며 검은 갓, 갓끈, 하늘색 도포, 흰 바지와 검정·흰 신발을 착용한다. 옆얼굴은 보이지만 정면 기준과의 정확한 얼굴 일치는 제한적으로만 확인된다. 안규홍은 검은 머리를 묶은 청년 남성의 뒷모습으로, 때 탄 밝은 작업복과 짚신을 착용한다. 나무 지게와 옥색 보따리가 보이며 지게 하단에는 추가 짚더미도 실려 있다. 상 위 잔은 확인되고 읽을 수 있는 글자는 없다.",
        "hard_violations": [],
        "physics": "박진성은 두 발을 바닥에 붙이고 곧게 서 있다. 안규홍은 한쪽 발로 땅을 딛고 반대쪽 발을 뒤로 들어 보행 중간 동작을 만든다. 지게는 등에 밀착되어 어깨 쪽으로 걸쳐져 있고, 보따리와 짚은 지게 가로대에 얹혀 있다. 상은 다리로 바닥에 서며 잔은 상판에 놓여 있다. 명백하게 지지 없이 떠 있는 몸이나 물체는 없다."
       },
       {
        "label": "A",
        "direction": "박진성의 몸과 고개는 오른쪽 문밖 안규홍을 향한다. 안규홍은 등을 보이고 대문 쪽으로 멀어지는 방향으로 걷는다. 지게의 뒷면과 옥색 보따리가 관객에게 보인다. 시선 대상은 맞지만 박진성의 얼굴이 뒤쪽 각도로 보여 노려보는 눈빛 자체는 확인하기 어렵다.",
        "built_space": "오른쪽에 큰 출입구 하나가 있고, 왼쪽으로 열린 문짝 하나와 오른쪽의 좁은 고정 구획이 보인다. 실내 왼쪽에는 낮은 상 하나와 잔 하나, 마당 오른쪽에는 평상 일부, 먼 배경에는 대문 하나와 담장이 있다. 박진성은 목조 실내 바닥의 중앙 왼쪽에 전신으로 서 있고, 안규홍은 중앙 오른쪽 문틀 안의 먼 마당에 작게 배치된다. 높인 나무 바닥과 마당 사이 단차가 장소 사진의 목조 공간에 더 가깝다. 다만 대문은 기와지붕이며 담장 상단도 구조 사진의 초가 형태와 다르다.",
        "entities": "추가 인물 없이 두 남성만 보인다. 박진성은 성인 한국인 남성 설정에 부합하는 외형이고 검은 갓, 하늘색 도포, 흰 바지와 검정·흰 신발이 참고 복장과 맞는다. 얼굴은 대부분 가려져 정확한 동일 인물 여부는 판별하기 어렵다. 안규홍은 묶은 검은 머리, 밝고 낡은 작업복, 짚신을 착용한 청년 남성의 뒷모습이다. 등에는 자연스러운 크기의 나무 지게가 있고 아래쪽 가로대 부근에 작은 옥색 천 보따리가 들어 있다. 실내 상과 내려놓은 잔도 확인되며 읽을 수 있는 글자는 없다.",
        "hard_violations": [],
        "physics": "박진성의 두 발은 나무 바닥에 안정적으로 닿아 있으며 상체는 꼿꼿하다. 안규홍은 오른쪽 발로 마당을 딛고 왼쪽 발을 뒤로 들어 실제 걸음의 중간 순간을 보인다. 지게는 어깨와 등에 걸려 있고 보따리는 지게 하단에 받쳐져 있다. 상과 평상은 다리로 지면에 지지되며 잔은 상판 위에 놓인다. 지지 없는 부유나 불가능한 관절 자세는 보이지 않는다."
       }
      ]
     },
     "ok": true
    }
   ],
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "slot_winner_match": true,
   "slot_winner": {
    "gemini-pro": "A",
    "gpt-high": "A"
   },
   "route": "cross_slot_agree"
  },
  "dual": {
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "normalized": {
    "A": 2.0,
    "B": 1.159
   },
   "adjusted": {
    "A": 2.0,
    "B": 1.159
   },
   "violations": {},
   "per_model_winner": {
    "gemini-pro": "A",
    "gpt-high": "A"
   },
   "agreed": true
  },
  "totals": {
   "A": 2000,
   "B": 1159
  },
  "selected": "A",
  "ranking": [
   "A",
   "B"
  ],
  "verdicts": [
   {
    "label": "A",
    "score": 2000,
    "verdict_ko": "요구된 화면 구도(좌측 박진성, 우측 문밖 안규홍)를 완벽하게 구현하였으며, 안규홍의 걷는 자세와 지게, 옥색 보따리, 술상 등의 세부 요소를 지시문과 참조에 맞게 충실히 렌더링한 우수한 결과물입니다."
   },
   {
    "label": "B",
    "score": 1159,
    "verdict_ko": "실내 바닥을 흙바닥으로 묘사하고 육중한 문을 달아 참조된 장소의 건축 양식(Location lock)을 크게 훼손하였으며, 박진성을 화면의 좌측 끝으로 밀어내어 요구된 프레이밍 구도를 지키지 못했습니다."
   }
  ],
  "refs": [
   {
    "label": "LOCATION PHOTOGRAPH — the exact place of this shot: its spatial layout, surroundings, fixed features, time of day and lighting mood are spatial truth; stage the moment inside this place. If a STRUCTURE LOOK photograph is also attached, that photo wins for the fixed structure itself — this photograph wins for everything around it. Never copy its camera framing.",
    "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/db00517b-165b-4005-a287-f28936b59b95/episodes/272bd7af-9294-4490-98c6-88a40e9deacc/images/background_chain/L08B04.png",
    "asset_id": "71e0793a-2c9b-4326-bdfc-5abf8fffaac8",
    "role": "location_plate"
   },
   {
    "label": "STRUCTURE LOOK — the confirmed photograph of the fixed structure at this location: wherever the structure appears in the frame, its shape, proportions, materials, colors and openings are LOCKED to this photo. Never copy its camera framing, time of day or lighting — the shot text and the LOCATION PHOTOGRAPH are the authorities for those.",
    "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/db00517b-165b-4005-a287-f28936b59b95/images/background_chain/seed_bg_tiled_house_sel.png",
    "asset_id": "0d12e89a-d0a7-4e88-a172-83d7a9b982f6",
    "role": "structure_seed_look"
   },
   {
    "label": "CHARACTER REFERENCE — 안규홍: the exact person appearing in this shot; match face, hair and build exactly.",
    "path": "<bytes:1028384>",
    "asset_id": "3b5894db-ce55-4f34-b966-5e725c3aa952",
    "role": "character_ref"
   },
   {
    "label": "CHARACTER REFERENCE — 박진성: the exact person appearing in this shot; match face, hair and build exactly.",
    "path": "<bytes:912297>",
    "asset_id": "3690b9ae-055b-43ee-9b74-794f83fca3d5",
    "role": "character_ref"
   }
  ],
  "critique_skipped": true,
  "shot_run_produced": false,
  "shot_run_uid": "06aa1406-e87d-77e4-8164-da0cd90fe4c1",
  "bgfirst": {
   "bg_path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/db00517b-165b-4005-a287-f28936b59b95/images/272bd7af-9294-4490-98c6-88a40e9deacc/scene/recipe/S3sh18__bgfirst_bg.png",
   "bg_asset_id": "cb877610-afef-4846-9ca9-416de5702a10",
   "bg_record_key": "S3sh18::bgfirst_bg",
   "chain_winner": true,
   "authority": "plate",
   "seed_attached": true
  },
  "ref_mode": "재투영 배경+콘티+엔티티 (2택1: 체인 승)",
  "share_plan": {
   "ref_plan": "background"
  },
  "lane_policy": "ab_select_ready"
 },
 "S3sh18::cine": {
  "applied": true,
  "attempted_at": "2026-09-09T00:52:42.868157+00:00",
  "fingerprint": "38966ba426567979b2f5ee3c30aa9e46f54feb160c9061715469e654926181ae",
  "fingerprint_version": 2,
  "provider": "mai",
  "endpoint": "openrouter/chat-completions",
  "model": "microsoft/mai-image-2.6",
  "pack": "24.202608252115",
  "source_file": "S3sh18_sel.png",
  "source_sha256": "82c74d360fd4757565938d1af7d70551e406ef705274b996337fe789b7c8f50e",
  "file": "S3sh18_cine.png",
  "staged_sha256": "bd200e37376cc10f81dd280c9dd67a1653910e9f4d64119f9ed6f6e2a9d2afd3",
  "latency_ms": 24700
 },
 "S3sh26::signage": {
  "fp": "9db538302db5aa9e",
  "inscriptions": [],
  "cues": [],
  "dropped": []
 },
 "S3sh26::bgfirst_bg": {
  "input_fingerprint": "18d32d6dd381b02d",
  "prompt": "Create the EMPTY BACKGROUND PLATE for one film shot — NO PEOPLE, no figures, no body parts, no sketch lines, no arrows anywhere.\n\"Empty\" means no people only: KEEP the location's inherent occupants and stock that define the place — animals in an animal shelter, pen or farm, goods in a market, moored boats in a harbour — unless the shot text explicitly removes them.\nThe FIRST attached image is a thin-line storyboard sketch: use ONLY its camera angle, horizon, perspective and the placement/size of buildings and set masses — ignore the sketched people and arrows entirely. The SECOND attached image (LOCATION PHOTOGRAPH) is the real place: take its architecture, materials, signage and fixed features, and RE-PROJECT them into the sketch's camera. If the photograph's camera differs from the sketch's, the sketch's camera wins.\nHUMAN-SCALE CALIBRATION: derive every structure's true size from human-scale elements — a door ≈ 2m, a window ≈ 1–1.5m wide, one storey ≈ 2.5–3m; never inflate a small structure or shrink a large one.\n\nSHOT TEXT this background must serve (Korean): 달래를 마주보며 참지 못하고 활짝 웃는 안규홍의 흙먼지 묻은 정면.\n\nLOCATION (lock): Beside the stone step at the main gate inside a traditional estate's outdoor yard.\n\nTIME OF DAY (lock): morning, harsh sunlight.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: close-up\n- KEY BACKGROUND ELEMENTS: 안규홍의 지게 (Still carried on his back) — A small oblique portion is visible behind his shoulder; used as Keeps the intimate smile attached to his ongoing departure and physical workload.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Maintain the established morning sunlight with restrained saturation and natural skin texture, allowing the shared laughter rather than a lighting change to provide warmth.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nRender ONE photorealistic empty location photograph, 16:9, that this shot can be staged inside later. No readable writing anywhere: surfaces that would carry writing may be present, but stage any wording out of legibility — an oblique angle, distance, shallow focus. No captions, watermarks or overlay text.",
  "effective_prompt": "Create the EMPTY BACKGROUND PLATE for one film shot — NO PEOPLE, no figures, no body parts, no sketch lines, no arrows anywhere.\n\"Empty\" means no people only: KEEP the location's inherent occupants and stock that define the place — animals in an animal shelter, pen or farm, goods in a market, moored boats in a harbour — unless the shot text explicitly removes them.\nThe FIRST attached image is a thin-line storyboard sketch: use ONLY its camera angle, horizon, perspective and the placement/size of buildings and set masses — ignore the sketched people and arrows entirely. The SECOND attached image (LOCATION PHOTOGRAPH) is the real place: take its architecture, materials, signage and fixed features, and RE-PROJECT them into the sketch's camera. If the photograph's camera differs from the sketch's, the sketch's camera wins.\nHUMAN-SCALE CALIBRATION: derive every structure's true size from human-scale elements — a door ≈ 2m, a window ≈ 1–1.5m wide, one storey ≈ 2.5–3m; never inflate a small structure or shrink a large one.\n\nSHOT TEXT this background must serve (Korean): 달래를 마주보며 참지 못하고 활짝 웃는 안규홍의 흙먼지 묻은 정면.\n\nLOCATION (lock): Beside the stone step at the main gate inside a traditional estate's outdoor yard.\n\nTIME OF DAY (lock): morning, harsh sunlight.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: close-up\n- KEY BACKGROUND ELEMENTS: 안규홍의 지게 (Still carried on his back) — A small oblique portion is visible behind his shoulder; used as Keeps the intimate smile attached to his ongoing departure and physical workload.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Maintain the established morning sunlight with restrained saturation and natural skin texture, allowing the shared laughter rather than a lighting change to provide warmth.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nRender ONE photorealistic empty location photograph, 16:9, that this shot can be staged inside later. No readable writing anywhere: surfaces that would carry writing may be present, but stage any wording out of legibility — an oblique angle, distance, shallow focus. No captions, watermarks or overlay text.",
  "bg_path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/db00517b-165b-4005-a287-f28936b59b95/images/272bd7af-9294-4490-98c6-88a40e9deacc/scene/recipe/S3sh26__bgfirst_bg.png",
  "asset_id": "965ad17e-3be6-4b17-9809-1b80afd43baf",
  "input_asset_ids": [
   "8e95141e-5e59-4329-a4a2-daac49ee8836",
   "71e0793a-2c9b-4326-bdfc-5abf8fffaac8"
  ]
 },
 "S3sh26": {
  "input_fingerprint": "122f5332bbbdbf1b",
  "prompt": "Create ONE FINAL photorealistic live-action film still of the moment below — early 20th-century Korea, Jeolla region; all people are Korean unless explicitly stated otherwise. TIME OF DAY (lock): morning, harsh sunlight.\n\nSHOT TEXT (authoritative, Korean): 달래를 마주보며 참지 못하고 활짝 웃는 안규홍의 흙먼지 묻은 정면.\n\nLOCATION (lock): Beside the stone step at the main gate inside a traditional estate's outdoor yard. The shot takes place here — the attached LOCATION PHOTOGRAPH shows the exact spot.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: close-up\n- KEY BACKGROUND ELEMENTS: 안규홍의 지게 (Still carried on his back) — A small oblique portion is visible behind his shoulder; used as Keeps the intimate smile attached to his ongoing departure and physical workload.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Maintain the established morning sunlight with restrained saturation and natural skin texture, allowing the shared laughter rather than a lighting change to provide warmth.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The sarangchae door remains open with the drinking table inside; strong sunlight continues over the yard. 안규홍: He still wears his loaded carrying frame and worn straw shoes, holding the partly eaten hot potato.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 안규홍 (한국인 남성, 청년 나이대, 검은 머리, 단단히 묶은 머리) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.",
  "roll_prompts": {
   "A": "Stage the shot. The FIRST attached image (SHOT BACKGROUND) is the finished empty background of this shot — keep it EXACTLY: its camera, perspective, architecture, lighting and every fixed feature stay untouched. The SECOND attached image (LAYOUT SKETCH) tells you ONLY where the people go: each sketched person's position, screen size, pose and the gaze/motion arrows. Ignore the sketch's background lines. The CHARACTER REFERENCE photographs show the real people.\nPlace the real people into the background at exactly the sketched positions, sizes and poses, following the arrow directions. No sketch lines or arrows may remain.\n\nCreate ONE FINAL photorealistic live-action film still of the moment below — early 20th-century Korea, Jeolla region; all people are Korean unless explicitly stated otherwise. TIME OF DAY (lock): morning, harsh sunlight.\n\nSHOT TEXT (authoritative, Korean): 달래를 마주보며 참지 못하고 활짝 웃는 안규홍의 흙먼지 묻은 정면.\n\nLOCATION (lock): Beside the stone step at the main gate inside a traditional estate's outdoor yard. The shot takes place here — the FIRST attached image (SHOT BACKGROUND) is this exact place, already built: its ground, structures, horizon, materials and lighting are the finished truth of this location and must not be redesigned or replaced. No location photograph is attached — read the place from that image alone, and add no scenery, structure, vehicle or fixture that it does not already show. This lock governs the place only; the figures in the shot follow the staging and pose instructions.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: close-up\n- KEY BACKGROUND ELEMENTS: 안규홍의 지게 (Still carried on his back) — A small oblique portion is visible behind his shoulder; used as Keeps the intimate smile attached to his ongoing departure and physical workload.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Maintain the established morning sunlight with restrained saturation and natural skin texture, allowing the shared laughter rather than a lighting change to provide warmth.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The sarangchae door remains open with the drinking table inside; strong sunlight continues over the yard. 안규홍: He still wears his loaded carrying frame and worn straw shoes, holding the partly eaten hot potato.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 안규홍 (한국인 남성, 청년 나이대, 검은 머리, 단단히 묶은 머리) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.",
   "B": "Create ONE FINAL photorealistic live-action film still of the moment below — early 20th-century Korea, Jeolla region; all people are Korean unless explicitly stated otherwise. TIME OF DAY (lock): morning, harsh sunlight.\n\nSHOT TEXT (authoritative, Korean): 달래를 마주보며 참지 못하고 활짝 웃는 안규홍의 흙먼지 묻은 정면.\n\nLOCATION (lock): Beside the stone step at the main gate inside a traditional estate's outdoor yard. The shot takes place here — the attached LOCATION PHOTOGRAPH shows the exact spot.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: close-up\n- KEY BACKGROUND ELEMENTS: 안규홍의 지게 (Still carried on his back) — A small oblique portion is visible behind his shoulder; used as Keeps the intimate smile attached to his ongoing departure and physical workload.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Maintain the established morning sunlight with restrained saturation and natural skin texture, allowing the shared laughter rather than a lighting change to provide warmth.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The sarangchae door remains open with the drinking table inside; strong sunlight continues over the yard. 안규홍: He still wears his loaded carrying frame and worn straw shoes, holding the partly eaten hot potato.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 안규홍 (한국인 남성, 청년 나이대, 검은 머리, 단단히 묶은 머리) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.\n\nCOMPOSITION VARIATION (this candidate only): keep the place, the people and the moment exactly as specified, and keep from the CAMERA & FRAME contract its subject scale, screen placement and key background elements exactly. For this candidate only, its camera angle and height are deliberately left open: choose a camera position distinctly different from the obvious one, as a film director picking a second setup on the same blocking. Stay true to any explicit framing the SHOT TEXT itself dictates."
  },
  "roll_refs": {
   "A": [
    {
     "label": "SHOT BACKGROUND",
     "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/db00517b-165b-4005-a287-f28936b59b95/images/272bd7af-9294-4490-98c6-88a40e9deacc/scene/recipe/S3sh26__bgfirst_bg.png",
     "asset_id": "965ad17e-3be6-4b17-9809-1b80afd43baf",
     "role": "bgfirst_bg"
    },
    {
     "label": "LAYOUT SKETCH (people placement only)",
     "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/db00517b-165b-4005-a287-f28936b59b95/images/272bd7af-9294-4490-98c6-88a40e9deacc/conti/conti_S3sh26.png",
     "asset_id": "8e95141e-5e59-4329-a4a2-daac49ee8836",
     "role": "conti_light"
    },
    {
     "label": "CHARACTER REFERENCE — 안규홍: the exact person appearing in this shot; match face, hair and build exactly.",
     "path": "<bytes:1028384>",
     "asset_id": "3b5894db-ce55-4f34-b966-5e725c3aa952",
     "role": "character_ref"
    }
   ],
   "B": [
    {
     "label": "LOCATION PHOTOGRAPH — the exact place of this shot: its architecture, materials, fixed features and lighting mood are spatial truth; stage the moment inside this place. Never copy its camera framing.",
     "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/db00517b-165b-4005-a287-f28936b59b95/episodes/272bd7af-9294-4490-98c6-88a40e9deacc/images/background_chain/L08B04.png",
     "asset_id": "71e0793a-2c9b-4326-bdfc-5abf8fffaac8",
     "role": "location_plate"
    },
    {
     "label": "CHARACTER REFERENCE — 안규홍: the exact person appearing in this shot; match face, hair and build exactly.",
     "path": "<bytes:1028384>",
     "asset_id": "3b5894db-ce55-4f34-b966-5e725c3aa952",
     "role": "character_ref"
    }
   ]
  },
  "shot_run_spend_attempt_count": 1,
  "judge_flip_skipped": "judge_fn_owns_order",
  "readings": [
   {
    "label": "A",
    "direction": "안규홍이 왼쪽 전경에 위치한 달래를 향해 고개를 돌려 바라보고 있음.",
    "built_space": "마당, 대문, 우측 건물의 기단이 보이나 카메라 앵글로 인해 사랑채 내부와 술상은 프레임 밖으로 제외됨.",
    "entities": "안규홍의 얼굴에 흙먼지가 묻어 있고 먹다 남은 감자를 들고 있으나, 레퍼런스에 없는 머리띠를 착용하여 인물 외형이 변형됨.",
    "hard_violations": [
     "[gemini-pro] 레퍼런스에 없는 머리띠 추가 (인물 외형 위반)",
     "[gemini-pro] 등에 맨 지게 나무 기둥의 구조가 물리적으로 불가능하고 공중에 떠 있는 듯한 형태 오류",
     "[gpt-high] 화면 왼쪽 전경에 여성을 등장시켜 안규홍만 허용한 명시적 인물 목록을 위반한다.",
     "[gpt-high] 안규홍을 지정된 대문 돌계단 옆이 아니라, 대문과 떨어진 사랑채 툇마루 바로 옆에 배치했다."
    ],
    "physics": "손으로 감자를 쥐고 있는 모습은 유지되나, 등 뒤의 지게 구조물들이 서로 제대로 연결되지 않은 채 얹혀 있음."
   },
   {
    "label": "B",
    "direction": "안규홍이 정면에 가깝게 위치한 상태에서 우측 전경의 달래를 바라보며 활짝 웃고 있음.",
    "built_space": "배경으로 마당, 솟을대문, 기단부가 클로즈업 샷에 맞게 배치되어 있으며 자연스러운 깊이감을 줌.",
    "entities": "안규홍의 단단히 묶은 머리(머리띠 없음), 흙먼지 묻은 얼굴, 먹다 남은 뜨거운 감자, 어깨 너머로 보이는 짚이 실린 지게 모두 지시 및 레퍼런스와 정확히 일치함.",
    "hard_violations": [
     "[gpt-high] 화면 오른쪽 전경에 여성을 등장시켰다. 가시 인물을 안규홍 한 명으로 제한한 명시적 인물 목록을 위반한다."
    ],
    "physics": "손으로 감자를 안정적으로 쥐고 있으며, 어깨에 걸쳐진 지게 끈과 짊어진 자세가 물리적으로 자연스러움."
   }
  ],
  "cross_model_order": {
   "policy": "cross_model_order_v2_slot0_fwd_slot1_rev_parallel2",
   "slots": [
    {
     "model": "gemini-pro",
     "order": "forward",
     "display_to_canonical": {
      "A": "A",
      "B": "B"
     },
     "raw": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 9,
        "verdict_ko": "샷 텍스트가 요구한 안규홍의 '정면' 구도를 훌륭하게 구현했으며, 레퍼런스와 일치하는 헤어스타일 및 지게의 디테일을 자연스럽게 살려냈습니다."
       },
       {
        "label": "A",
        "score": 4,
        "verdict_ko": "요구된 정면 샷 대신 측면 구도를 렌더링하였고, 레퍼런스에 없는 머리띠를 추가했으며 지게의 형태가 물리적으로 심하게 왜곡되었습니다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "안규홍이 왼쪽 전경에 위치한 달래를 향해 고개를 돌려 바라보고 있음.",
        "built_space": "마당, 대문, 우측 건물의 기단이 보이나 카메라 앵글로 인해 사랑채 내부와 술상은 프레임 밖으로 제외됨.",
        "entities": "안규홍의 얼굴에 흙먼지가 묻어 있고 먹다 남은 감자를 들고 있으나, 레퍼런스에 없는 머리띠를 착용하여 인물 외형이 변형됨.",
        "hard_violations": [
         "레퍼런스에 없는 머리띠 추가 (인물 외형 위반)",
         "등에 맨 지게 나무 기둥의 구조가 물리적으로 불가능하고 공중에 떠 있는 듯한 형태 오류"
        ],
        "physics": "손으로 감자를 쥐고 있는 모습은 유지되나, 등 뒤의 지게 구조물들이 서로 제대로 연결되지 않은 채 얹혀 있음."
       },
       {
        "label": "B",
        "direction": "안규홍이 정면에 가깝게 위치한 상태에서 우측 전경의 달래를 바라보며 활짝 웃고 있음.",
        "built_space": "배경으로 마당, 솟을대문, 기단부가 클로즈업 샷에 맞게 배치되어 있으며 자연스러운 깊이감을 줌.",
        "entities": "안규홍의 단단히 묶은 머리(머리띠 없음), 흙먼지 묻은 얼굴, 먹다 남은 뜨거운 감자, 어깨 너머로 보이는 짚이 실린 지게 모두 지시 및 레퍼런스와 정확히 일치함.",
        "hard_violations": [],
        "physics": "손으로 감자를 안정적으로 쥐고 있으며, 어깨에 걸쳐진 지게 끈과 짊어진 자세가 물리적으로 자연스러움."
       }
      ],
      "all_candidates_fail": false
     },
     "normalized": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 9,
        "verdict_ko": "샷 텍스트가 요구한 안규홍의 '정면' 구도를 훌륭하게 구현했으며, 레퍼런스와 일치하는 헤어스타일 및 지게의 디테일을 자연스럽게 살려냈습니다."
       },
       {
        "label": "A",
        "score": 4,
        "verdict_ko": "요구된 정면 샷 대신 측면 구도를 렌더링하였고, 레퍼런스에 없는 머리띠를 추가했으며 지게의 형태가 물리적으로 심하게 왜곡되었습니다."
       }
      ],
      "all_candidates_fail": false,
      "readings": [
       {
        "label": "A",
        "direction": "안규홍이 왼쪽 전경에 위치한 달래를 향해 고개를 돌려 바라보고 있음.",
        "built_space": "마당, 대문, 우측 건물의 기단이 보이나 카메라 앵글로 인해 사랑채 내부와 술상은 프레임 밖으로 제외됨.",
        "entities": "안규홍의 얼굴에 흙먼지가 묻어 있고 먹다 남은 감자를 들고 있으나, 레퍼런스에 없는 머리띠를 착용하여 인물 외형이 변형됨.",
        "hard_violations": [
         "레퍼런스에 없는 머리띠 추가 (인물 외형 위반)",
         "등에 맨 지게 나무 기둥의 구조가 물리적으로 불가능하고 공중에 떠 있는 듯한 형태 오류"
        ],
        "physics": "손으로 감자를 쥐고 있는 모습은 유지되나, 등 뒤의 지게 구조물들이 서로 제대로 연결되지 않은 채 얹혀 있음."
       },
       {
        "label": "B",
        "direction": "안규홍이 정면에 가깝게 위치한 상태에서 우측 전경의 달래를 바라보며 활짝 웃고 있음.",
        "built_space": "배경으로 마당, 솟을대문, 기단부가 클로즈업 샷에 맞게 배치되어 있으며 자연스러운 깊이감을 줌.",
        "entities": "안규홍의 단단히 묶은 머리(머리띠 없음), 흙먼지 묻은 얼굴, 먹다 남은 뜨거운 감자, 어깨 너머로 보이는 짚이 실린 지게 모두 지시 및 레퍼런스와 정확히 일치함.",
        "hard_violations": [],
        "physics": "손으로 감자를 안정적으로 쥐고 있으며, 어깨에 걸쳐진 지게 끈과 짊어진 자세가 물리적으로 자연스러움."
       }
      ]
     },
     "ok": true
    },
    {
     "model": "gpt-high",
     "order": "reverse",
     "display_to_canonical": {
      "A": "B",
      "B": "A"
     },
     "raw": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 3,
        "verdict_ko": "정면의 흙먼지 묻은 활짝 웃음과 어깨 뒤 지게는 더 충실하지만, 허용되지 않은 여성 전경과 넓은 상반신 구도로 인물 제한 및 클로즈업 지시를 위반한다."
       },
       {
        "label": "B",
        "score": 2,
        "verdict_ko": "허용되지 않은 여성을 추가했고, 대문 돌계단 옆 정면 클로즈업 대신 사랑채 앞에서 돌아보는 넓은 구도로 지게까지 크게 드러낸다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "안규홍은 얼굴을 거의 정면으로 보이며 화면 오른쪽 전경 여성의 얼굴을 향해 눈을 두고 활짝 웃는다. 여성도 안규홍 쪽으로 고개를 향한다. 달래를 마주보는 시선 관계 자체는 읽히지만, 등장 허용 목록에는 이 여성이 없다.",
        "built_space": "뒤쪽에 기와 대문 한 곳과 그 아래 낮은 돌 문턱, 연결된 돌담이 보인다. 오른쪽에는 사랑채 툇마루 일부와 돌 디딤돌 한 개가 보인다. 참고의 재료와 주요 배치는 대체로 유지되며 안규홍은 대문 앞마당에 있다. 다만 배경과 몸통을 많이 포함해 지정된 얼굴 클로즈업보다 넓다. 중복 시설이나 불가능한 반사는 보이지 않는다.",
        "entities": "안규홍으로 제시된 청년 한국인 남성 한 명이 검은 머리를 단단히 묶고 흙먼지 묻은 옅은색 저고리를 입었다. 참고보다 얼굴이 조금 더 나이 들어 보이나 주요 인상과 복장은 가깝다. 얼굴에도 흙먼지가 있고 입을 크게 벌려 웃는다. 어깨 뒤에는 나무 지게의 세로대와 가로대 일부, 짚짐이 보인다. 손에는 일부 먹은 감자가 있다. 오른쪽에 검은 머리를 묶은 여성의 머리와 어깨가 추가되어 있다. 짚신과 실내 술상은 화면 밖이므로 판정하지 않는다. 읽을 수 있는 글자는 보이지 않는다.",
        "hard_violations": [
         "화면 오른쪽 전경에 여성을 등장시켰다. 가시 인물을 안규홍 한 명으로 제한한 명시적 인물 목록을 위반한다."
        ],
        "physics": "감자는 손가락과 손바닥으로 받쳐 쥐고 있다. 지게는 양쪽 어깨의 멜끈으로 몸에 연결되고 짐은 등 뒤 지게에 얹혀 있어 지지가 읽힌다. 웃으며 조금 앞으로 기울어진 상체는 가능한 자세다. 발은 화면 밖이며 공중에 떠 있다는 징후는 없다. 여성의 보이는 머리와 어깨도 자연스럽게 이어진다."
       },
       {
        "label": "B",
        "direction": "안규홍은 몸통을 비스듬히 둔 채 고개를 화면 왼쪽 전경 여성에게 돌려 웃는다. 눈도 그 여성의 얼굴 쪽을 향하고 여성은 그를 바라본다. 상대를 향한 웃음은 맞지만, 정면을 보여주는 순간보다 출발하다 뒤돌아보는 순간으로 읽힌다.",
        "built_space": "왼쪽 뒤에 기와 대문 한 곳, 열린 문틈과 낮은 돌계단, 그 오른쪽 독립된 돌 한 개가 보인다. 오른쪽에는 여러 목재 기둥이 이어지는 사랑채 툇마루와 열린 어두운 출입구가 있다. 참고 장소의 대문·담·사랑채 관계는 가깝지만 안규홍은 멀리 보이는 대문 돌계단이 아니라 사랑채 툇마루 바로 옆에 서 있다. 넓은 마당까지 포함되어 클로즈업 지시와도 다르다. 불가능한 반사는 보이지 않는다.",
        "entities": "청년 한국인 남성의 검은 묶은 머리, 흙 묻은 얼굴, 낡은 옅은색 저고리와 활짝 웃는 표정이 보인다. 참고에는 없는 넓은 천 머리띠를 착용했다. 손에는 일부 먹은 감자가 있으며, 오른쪽에는 나무 지게의 긴 세로대와 가로대, 결박끈이 크게 드러난다. 적재물은 잘려 있어 충분히 확인되지 않는다. 왼쪽 전경에는 여성의 뒤통수와 어깨가 있다. 짚신과 실내 술상은 화면 밖이므로 감점하지 않는다. 읽을 수 있는 글자는 없다.",
        "hard_violations": [
         "화면 왼쪽 전경에 여성을 등장시켜 안규홍만 허용한 명시적 인물 목록을 위반한다.",
         "안규홍을 지정된 대문 돌계단 옆이 아니라, 대문과 떨어진 사랑채 툇마루 바로 옆에 배치했다."
        ],
        "physics": "감자는 구부린 손가락과 손바닥이 지지한다. 지게는 어깨 위 굵은 멜끈과 등 뒤 접촉으로 지지되며 나무 부재들은 밧줄로 결박되어 있다. 몸통을 비스듬히 두고 목을 돌려 웃는 자세는 물리적으로 가능하다. 하체는 잘려 있지만 떠 있는 신체나 지지 없는 물체는 보이지 않는다."
       }
      ],
      "all_candidates_fail": true
     },
     "normalized": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 3,
        "verdict_ko": "정면의 흙먼지 묻은 활짝 웃음과 어깨 뒤 지게는 더 충실하지만, 허용되지 않은 여성 전경과 넓은 상반신 구도로 인물 제한 및 클로즈업 지시를 위반한다."
       },
       {
        "label": "A",
        "score": 2,
        "verdict_ko": "허용되지 않은 여성을 추가했고, 대문 돌계단 옆 정면 클로즈업 대신 사랑채 앞에서 돌아보는 넓은 구도로 지게까지 크게 드러낸다."
       }
      ],
      "all_candidates_fail": true,
      "readings": [
       {
        "label": "B",
        "direction": "안규홍은 얼굴을 거의 정면으로 보이며 화면 오른쪽 전경 여성의 얼굴을 향해 눈을 두고 활짝 웃는다. 여성도 안규홍 쪽으로 고개를 향한다. 달래를 마주보는 시선 관계 자체는 읽히지만, 등장 허용 목록에는 이 여성이 없다.",
        "built_space": "뒤쪽에 기와 대문 한 곳과 그 아래 낮은 돌 문턱, 연결된 돌담이 보인다. 오른쪽에는 사랑채 툇마루 일부와 돌 디딤돌 한 개가 보인다. 참고의 재료와 주요 배치는 대체로 유지되며 안규홍은 대문 앞마당에 있다. 다만 배경과 몸통을 많이 포함해 지정된 얼굴 클로즈업보다 넓다. 중복 시설이나 불가능한 반사는 보이지 않는다.",
        "entities": "안규홍으로 제시된 청년 한국인 남성 한 명이 검은 머리를 단단히 묶고 흙먼지 묻은 옅은색 저고리를 입었다. 참고보다 얼굴이 조금 더 나이 들어 보이나 주요 인상과 복장은 가깝다. 얼굴에도 흙먼지가 있고 입을 크게 벌려 웃는다. 어깨 뒤에는 나무 지게의 세로대와 가로대 일부, 짚짐이 보인다. 손에는 일부 먹은 감자가 있다. 오른쪽에 검은 머리를 묶은 여성의 머리와 어깨가 추가되어 있다. 짚신과 실내 술상은 화면 밖이므로 판정하지 않는다. 읽을 수 있는 글자는 보이지 않는다.",
        "hard_violations": [
         "화면 오른쪽 전경에 여성을 등장시켰다. 가시 인물을 안규홍 한 명으로 제한한 명시적 인물 목록을 위반한다."
        ],
        "physics": "감자는 손가락과 손바닥으로 받쳐 쥐고 있다. 지게는 양쪽 어깨의 멜끈으로 몸에 연결되고 짐은 등 뒤 지게에 얹혀 있어 지지가 읽힌다. 웃으며 조금 앞으로 기울어진 상체는 가능한 자세다. 발은 화면 밖이며 공중에 떠 있다는 징후는 없다. 여성의 보이는 머리와 어깨도 자연스럽게 이어진다."
       },
       {
        "label": "A",
        "direction": "안규홍은 몸통을 비스듬히 둔 채 고개를 화면 왼쪽 전경 여성에게 돌려 웃는다. 눈도 그 여성의 얼굴 쪽을 향하고 여성은 그를 바라본다. 상대를 향한 웃음은 맞지만, 정면을 보여주는 순간보다 출발하다 뒤돌아보는 순간으로 읽힌다.",
        "built_space": "왼쪽 뒤에 기와 대문 한 곳, 열린 문틈과 낮은 돌계단, 그 오른쪽 독립된 돌 한 개가 보인다. 오른쪽에는 여러 목재 기둥이 이어지는 사랑채 툇마루와 열린 어두운 출입구가 있다. 참고 장소의 대문·담·사랑채 관계는 가깝지만 안규홍은 멀리 보이는 대문 돌계단이 아니라 사랑채 툇마루 바로 옆에 서 있다. 넓은 마당까지 포함되어 클로즈업 지시와도 다르다. 불가능한 반사는 보이지 않는다.",
        "entities": "청년 한국인 남성의 검은 묶은 머리, 흙 묻은 얼굴, 낡은 옅은색 저고리와 활짝 웃는 표정이 보인다. 참고에는 없는 넓은 천 머리띠를 착용했다. 손에는 일부 먹은 감자가 있으며, 오른쪽에는 나무 지게의 긴 세로대와 가로대, 결박끈이 크게 드러난다. 적재물은 잘려 있어 충분히 확인되지 않는다. 왼쪽 전경에는 여성의 뒤통수와 어깨가 있다. 짚신과 실내 술상은 화면 밖이므로 감점하지 않는다. 읽을 수 있는 글자는 없다.",
        "hard_violations": [
         "화면 왼쪽 전경에 여성을 등장시켜 안규홍만 허용한 명시적 인물 목록을 위반한다.",
         "안규홍을 지정된 대문 돌계단 옆이 아니라, 대문과 떨어진 사랑채 툇마루 바로 옆에 배치했다."
        ],
        "physics": "감자는 구부린 손가락과 손바닥이 지지한다. 지게는 어깨 위 굵은 멜끈과 등 뒤 접촉으로 지지되며 나무 부재들은 밧줄로 결박되어 있다. 몸통을 비스듬히 두고 목을 돌려 웃는 자세는 물리적으로 가능하다. 하체는 잘려 있지만 떠 있는 신체나 지지 없는 물체는 보이지 않는다."
       }
      ]
     },
     "ok": true
    }
   ],
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "slot_winner_match": true,
   "slot_winner": {
    "gemini-pro": "B",
    "gpt-high": "B"
   },
   "route": "cross_slot_agree"
  },
  "dual": {
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "normalized": {
    "A": 1.111,
    "B": 2.0
   },
   "adjusted": {
    "A": 0.861,
    "B": 1.75
   },
   "violations": {
    "A": [
     "[gemini-pro] 레퍼런스에 없는 머리띠 추가 (인물 외형 위반)",
     "[gemini-pro] 등에 맨 지게 나무 기둥의 구조가 물리적으로 불가능하고 공중에 떠 있는 듯한 형태 오류",
     "[gpt-high] 화면 왼쪽 전경에 여성을 등장시켜 안규홍만 허용한 명시적 인물 목록을 위반한다.",
     "[gpt-high] 안규홍을 지정된 대문 돌계단 옆이 아니라, 대문과 떨어진 사랑채 툇마루 바로 옆에 배치했다."
    ],
    "B": [
     "[gpt-high] 화면 오른쪽 전경에 여성을 등장시켰다. 가시 인물을 안규홍 한 명으로 제한한 명시적 인물 목록을 위반한다."
    ]
   },
   "per_model_winner": {
    "gemini-pro": "B",
    "gpt-high": "B"
   },
   "agreed": true
  },
  "totals": {
   "B": 1750,
   "A": 861
  },
  "selected": "B",
  "ranking": [
   "B",
   "A"
  ],
  "verdicts": [
   {
    "label": "B",
    "score": 1750,
    "verdict_ko": "샷 텍스트가 요구한 안규홍의 '정면' 구도를 훌륭하게 구현했으며, 레퍼런스와 일치하는 헤어스타일 및 지게의 디테일을 자연스럽게 살려냈습니다.  ★위반: [gpt-high] 화면 오른쪽 전경에 여성을 등장시켰다. 가시 인물을 안규홍 한 명으로 제한한 명시적 인물 목록을 위반한다."
   },
   {
    "label": "A",
    "score": 861,
    "verdict_ko": "요구된 정면 샷 대신 측면 구도를 렌더링하였고, 레퍼런스에 없는 머리띠를 추가했으며 지게의 형태가 물리적으로 심하게 왜곡되었습니다.  ★위반: [gemini-pro] 레퍼런스에 없는 머리띠 추가 (인물 외형 위반) / [gemini-pro] 등에 맨 지게 나무 기둥의 구조가 물리적으로 불가능하고 공중에 떠 있는 듯한 형태 오류 / [gpt-high] 화면 왼쪽 전경에 여성을 등장시켜 안규홍만 허용한 명시적 인물 목록을 위반한다. / [gpt-high] 안규홍을 지정된 대문 돌계단 옆이 아니라, 대문과 떨어진 사랑채 툇마루 바로 옆에 배치했다."
   }
  ],
  "refs": [
   {
    "label": "LOCATION PHOTOGRAPH — the exact place of this shot: its architecture, materials, fixed features and lighting mood are spatial truth; stage the moment inside this place. Never copy its camera framing.",
    "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/db00517b-165b-4005-a287-f28936b59b95/episodes/272bd7af-9294-4490-98c6-88a40e9deacc/images/background_chain/L08B04.png",
    "asset_id": "71e0793a-2c9b-4326-bdfc-5abf8fffaac8",
    "role": "location_plate"
   },
   {
    "label": "CHARACTER REFERENCE — 안규홍: the exact person appearing in this shot; match face, hair and build exactly.",
    "path": "<bytes:1028384>",
    "asset_id": "3b5894db-ce55-4f34-b966-5e725c3aa952",
    "role": "character_ref"
   }
  ],
  "critique_skipped": true,
  "shot_run_produced": false,
  "shot_run_uid": "06aa1406-eb89-7624-96d1-c99807c18ee6",
  "bgfirst": {
   "bg_path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/db00517b-165b-4005-a287-f28936b59b95/images/272bd7af-9294-4490-98c6-88a40e9deacc/scene/recipe/S3sh26__bgfirst_bg.png",
   "bg_asset_id": "965ad17e-3be6-4b17-9809-1b80afd43baf",
   "bg_record_key": "S3sh26::bgfirst_bg",
   "chain_winner": false,
   "authority": "plate"
  },
  "ref_mode": "플레이트+엔티티 (2택1: 무콘티 승)",
  "share_plan": {
   "ref_plan": "background"
  }
 },
 "S3sh26::cine": {
  "applied": true,
  "attempted_at": "2026-09-09T00:55:33.797278+00:00",
  "fingerprint": "d1acdd06611a6f013a08ea9f4e7b8345dc1d7e64154c47d43bba0a539b6f30ef",
  "fingerprint_version": 2,
  "provider": "mai",
  "endpoint": "openrouter/chat-completions",
  "model": "microsoft/mai-image-2.6",
  "pack": "24.202608252115",
  "source_file": "S3sh26_sel.png",
  "source_sha256": "4202a492340f2a2ddf4ae869e1df63f49997127f52d9d03788eb73266c640e6b",
  "file": "S3sh26_cine.png",
  "staged_sha256": "4981cc594fc5e29db4277027d973e418b986ee965dd8917bc133883409f8437b",
  "latency_ms": 25539
 },
 "S4sh17::signage": {
  "fp": "e6e6ca75278aecf5",
  "inscriptions": [],
  "cues": [],
  "dropped": []
 },
 "groupbg::market_shoe_front": {
  "input_fingerprint": "a91398405a10c23d",
  "meta": {
   "model": "gpt-image-2",
   "size": "1536x864",
   "pack": "11.202607220237",
   "contract": "bgfirst_full_v3",
   "group_sig": {
    "key": "market_shoe_front",
    "tags": [
     "S10sh13",
     "S10sh15",
     "S10sh6",
     "S48sh10",
     "S48sh12",
     "S48sh23",
     "S4sh17",
     "S4sh24",
     "S4sh30",
     "S51sh15",
     "S51sh2",
     "S51sh9",
     "S53sh10",
     "S53sh2",
     "S53sh6",
     "S55sh11",
     "S55sh34",
     "S55sh40",
     "S57sh12",
     "S5sh21",
     "S5sh28",
     "S5sh39",
     "S6sh1"
    ]
   },
   "context_sig": "5ac107de8e91ae53"
  },
  "prompt": "Create ONE empty live-action location background photograph — NO PEOPLE, no figures, no body parts, no silhouettes, no shadows or reflections of people anywhere.\n\"Empty\" means no people only: KEEP the location's inherent occupants and stock that define the place — animals in an animal shelter, pen or farm, goods and produce in a market, moored boats in a harbour — unless the shot text explicitly removes them.\nThe FIRST attached image is a thin-line storyboard sketch of a shot that happens at this location: use it ONLY as spatial evidence — what this place contains, how its ground, structures and landmarks are arranged and proportioned. Ignore the sketched people and arrows entirely, and do NOT copy its line style: render a fully photographic, physically plausible real place that fits THE LOCATION text below.\nHUMAN-SCALE CALIBRATION: derive every structure's true size from human-scale elements — a door ≈ 2m, a window ≈ 1–1.5m wide, one storey ≈ 2.5–3m; never inflate a small structure or shrink a large one.\n\nTHE LOCATION — early 20th-century Korea, Jeolla region; all people are Korean unless explicitly stated otherwise: The outdoor frontage of a straw-sandal stall along a densely packed daytime market lane.\n\nLOCATION DETAIL (descriptive evidence about this place from the production's location records — evidence, not a staging order): use it to understand what kind of place this is and what it permanently contains. Carry over ONLY its enduring physical features — layout, structures, ground and materials, fixed equipment and installations. Do NOT treat any lighting, weather or time-of-day wording, momentary object states, depicted events or subjective impressions in it as fixed facts of the place; the TIME OF DAY lock below is the sole authority for time and lighting.\n복내장터 좌판 거리·장터 입구·골목: 다양한 좌판과 짐꾼, 상인들이 밀집한 넓은 흙바닥 노천 장터. 포목전, 짚신전, 옹기전, 닭장, 생선 좌판이 빽빽하게 들어서 있으며, 벽에 관의 방이 붙어 있고 상황에 따라 찢어진 멍석과 부서진 좌판이 널브러져 있다. (특징: 넓은 흙바닥; 나무와 짚으로 얽어 만든 노점 좌판 (옹기, 짚신, 채소, 포목 등); 나무 닭장과 광주리; 방(벽보)이 붙어 있는 벽; 깨진 항아리와 뒤집힌 멍석, 부서진 나무판자들)\n\nSCENE EVIDENCE (verbatim quotes from the screenplay about this place — treat them as evidence of what the location physically contains and looks like; stage the PLACE those moments happen in, but do NOT depict the momentary actions, people or staged props themselves):\n- 배상도 일행이 짚신전 앞으로 온다. 상도와 규홍의 눈이 마주친다.\n- 탁! 규홍이 지게를 털썩 내려놓는다. 최 서방 앞에 쪼그려 앉아 흙 묻은 짚신을 집어 옷소매로 툭, 툭 턴다.\n- 규홍이 몸을 돌려 달린다. 장터 사람들이 슬쩍슬쩍 몸을 비킨다.\n- 조금 전 싸움이 끝난 장터.\n사람들이 쓰러진 일본인들을 멀찍이 둘러싸고 웅성거린다.\n- 짚신전 앞.\n최 서방의 아들이 엉망이 된 멍석을 걷어내고 있다.\n- 짚신전 앞. 최 서방과 눈이 마주친다.\n- 관회가 뒤를 돌아본다. 뒤쪽에서도 나온다. 길이 막혔다.\n- 둘은 다시 등을 맞댄다.\n- 규홍이 가려는 순간 일진회 대장이 길을 막는다.\n- 둘이 그대로 흙바닥에 주저앉는다.\n- 저 멀리 저잣거리 한복판에 일진회 시체들과 함께 관회가 쓰러져 있다.\n\nTIME OF DAY (lock): day.\n\nRender ONE photorealistic empty location photograph, 16:9, neutral enough that every shot of this place can be staged from it later. No readable writing anywhere: surfaces that would carry writing may be present, but stage any wording out of legibility — an oblique angle, distance, shallow focus. No captions, watermarks or overlay text.",
  "effective_prompt": "Create ONE empty live-action location background photograph — NO PEOPLE, no figures, no body parts, no silhouettes, no shadows or reflections of people anywhere.\n\"Empty\" means no people only: KEEP the location's inherent occupants and stock that define the place — animals in an animal shelter, pen or farm, goods and produce in a market, moored boats in a harbour — unless the shot text explicitly removes them.\nThe FIRST attached image is a thin-line storyboard sketch of a shot that happens at this location: use it ONLY as spatial evidence — what this place contains, how its ground, structures and landmarks are arranged and proportioned. Ignore the sketched people and arrows entirely, and do NOT copy its line style: render a fully photographic, physically plausible real place that fits THE LOCATION text below.\nHUMAN-SCALE CALIBRATION: derive every structure's true size from human-scale elements — a door ≈ 2m, a window ≈ 1–1.5m wide, one storey ≈ 2.5–3m; never inflate a small structure or shrink a large one.\n\nTHE LOCATION — early 20th-century Korea, Jeolla region; all people are Korean unless explicitly stated otherwise: The outdoor frontage of a straw-sandal stall along a densely packed daytime market lane.\n\nLOCATION DETAIL (descriptive evidence about this place from the production's location records — evidence, not a staging order): use it to understand what kind of place this is and what it permanently contains. Carry over ONLY its enduring physical features — layout, structures, ground and materials, fixed equipment and installations. Do NOT treat any lighting, weather or time-of-day wording, momentary object states, depicted events or subjective impressions in it as fixed facts of the place; the TIME OF DAY lock below is the sole authority for time and lighting.\n복내장터 좌판 거리·장터 입구·골목: 다양한 좌판과 짐꾼, 상인들이 밀집한 넓은 흙바닥 노천 장터. 포목전, 짚신전, 옹기전, 닭장, 생선 좌판이 빽빽하게 들어서 있으며, 벽에 관의 방이 붙어 있고 상황에 따라 찢어진 멍석과 부서진 좌판이 널브러져 있다. (특징: 넓은 흙바닥; 나무와 짚으로 얽어 만든 노점 좌판 (옹기, 짚신, 채소, 포목 등); 나무 닭장과 광주리; 방(벽보)이 붙어 있는 벽; 깨진 항아리와 뒤집힌 멍석, 부서진 나무판자들)\n\nSCENE EVIDENCE (verbatim quotes from the screenplay about this place — treat them as evidence of what the location physically contains and looks like; stage the PLACE those moments happen in, but do NOT depict the momentary actions, people or staged props themselves):\n- 배상도 일행이 짚신전 앞으로 온다. 상도와 규홍의 눈이 마주친다.\n- 탁! 규홍이 지게를 털썩 내려놓는다. 최 서방 앞에 쪼그려 앉아 흙 묻은 짚신을 집어 옷소매로 툭, 툭 턴다.\n- 규홍이 몸을 돌려 달린다. 장터 사람들이 슬쩍슬쩍 몸을 비킨다.\n- 조금 전 싸움이 끝난 장터.\n사람들이 쓰러진 일본인들을 멀찍이 둘러싸고 웅성거린다.\n- 짚신전 앞.\n최 서방의 아들이 엉망이 된 멍석을 걷어내고 있다.\n- 짚신전 앞. 최 서방과 눈이 마주친다.\n- 관회가 뒤를 돌아본다. 뒤쪽에서도 나온다. 길이 막혔다.\n- 둘은 다시 등을 맞댄다.\n- 규홍이 가려는 순간 일진회 대장이 길을 막는다.\n- 둘이 그대로 흙바닥에 주저앉는다.\n- 저 멀리 저잣거리 한복판에 일진회 시체들과 함께 관회가 쓰러져 있다.\n\nTIME OF DAY (lock): day.\n\nRender ONE photorealistic empty location photograph, 16:9, neutral enough that every shot of this place can be staged from it later. No readable writing anywhere: surfaces that would carry writing may be present, but stage any wording out of legibility — an oblique angle, distance, shallow focus. No captions, watermarks or overlay text.",
  "bg_path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/db00517b-165b-4005-a287-f28936b59b95/images/272bd7af-9294-4490-98c6-88a40e9deacc/scene/recipe/groupbg_market_shoe_front_6546f9.png",
  "asset_id": "a82710e0-4099-4422-b0e7-8cd69a1606bb",
  "input_asset_ids": [
   "47ad740d-fb4b-46a1-8596-baf4ba0db6b0"
  ],
  "origin_tag": "S4sh17",
  "place_text": "The outdoor frontage of a straw-sandal stall along a densely packed daytime market lane.",
  "origin_inputs": {
   "place_text": "The outdoor frontage of a straw-sandal stall along a densely packed daytime market lane.",
   "time_of_day_en": "day",
   "conti_asset_id": "47ad740d-fb4b-46a1-8596-baf4ba0db6b0"
  }
 },
 "S4sh17::bgfirst_bg": {
  "input_fingerprint": "9f59f486777d9359",
  "prompt": "Create the EMPTY BACKGROUND PLATE for one film shot — NO PEOPLE, no figures, no body parts, no sketch lines, no arrows anywhere.\n\"Empty\" means no people only: KEEP the location's inherent occupants and stock that define the place — animals in an animal shelter, pen or farm, goods in a market, moored boats in a harbour — unless the shot text explicitly removes them.\nThe FIRST attached image is a thin-line storyboard sketch: use ONLY its camera angle, horizon, perspective and the placement/size of buildings and set masses — ignore the sketched people and arrows entirely. The SECOND attached image (LOCATION PHOTOGRAPH) is the real place: take its architecture, materials, signage and fixed features, and RE-PROJECT them into the sketch's camera. If the photograph's camera differs from the sketch's, the sketch's camera wins.\nHUMAN-SCALE CALIBRATION: derive every structure's true size from human-scale elements — a door ≈ 2m, a window ≈ 1–1.5m wide, one storey ≈ 2.5–3m; never inflate a small structure or shrink a large one.\n\nSHOT TEXT this background must serve (Korean): 짚신전 앞으로 다가온 배상도와 안규홍의 시선이 공중에서 부딪히는 순간.\n\nLOCATION (lock): The outdoor frontage of a straw-sandal stall along a densely packed daytime market lane.\n\nTIME OF DAY (lock): day.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: medium shot\n- KEY BACKGROUND ELEMENTS: 짚신전 (Stocked with straw shoes beside the encounter) — The display is seen obliquely from its aisle side; used as A peripheral background strip establishes the location without introducing additional visible people; 배상도의 장부 (In his possession after checking the market accounts) — Held low and obliquely, with no readable page content presented to the camera; used as A small lower-frame element identifies the administrative threat without displacing the faces; 안규홍의 지게 (Carried on his back) — Seen at an angle behind his shoulder; used as Retains his working silhouette and continuity with his arrival at the stall.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Natural daytime ambience, muted saturation, and sober facial contrast allow the exchange to turn tense without an unsupported change of light source.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nRender ONE photorealistic empty location photograph, 16:9, that this shot can be staged inside later. No readable writing anywhere: surfaces that would carry writing may be present, but stage any wording out of legibility — an oblique angle, distance, shallow focus. No captions, watermarks or overlay text.",
  "effective_prompt": "Create the EMPTY BACKGROUND PLATE for one film shot — NO PEOPLE, no figures, no body parts, no sketch lines, no arrows anywhere.\n\"Empty\" means no people only: KEEP the location's inherent occupants and stock that define the place — animals in an animal shelter, pen or farm, goods in a market, moored boats in a harbour — unless the shot text explicitly removes them.\nThe FIRST attached image is a thin-line storyboard sketch: use ONLY its camera angle, horizon, perspective and the placement/size of buildings and set masses — ignore the sketched people and arrows entirely. The SECOND attached image (LOCATION PHOTOGRAPH) is the real place: take its architecture, materials, signage and fixed features, and RE-PROJECT them into the sketch's camera. If the photograph's camera differs from the sketch's, the sketch's camera wins.\nHUMAN-SCALE CALIBRATION: derive every structure's true size from human-scale elements — a door ≈ 2m, a window ≈ 1–1.5m wide, one storey ≈ 2.5–3m; never inflate a small structure or shrink a large one.\n\nSHOT TEXT this background must serve (Korean): 짚신전 앞으로 다가온 배상도와 안규홍의 시선이 공중에서 부딪히는 순간.\n\nLOCATION (lock): The outdoor frontage of a straw-sandal stall along a densely packed daytime market lane.\n\nTIME OF DAY (lock): day.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: medium shot\n- KEY BACKGROUND ELEMENTS: 짚신전 (Stocked with straw shoes beside the encounter) — The display is seen obliquely from its aisle side; used as A peripheral background strip establishes the location without introducing additional visible people; 배상도의 장부 (In his possession after checking the market accounts) — Held low and obliquely, with no readable page content presented to the camera; used as A small lower-frame element identifies the administrative threat without displacing the faces; 안규홍의 지게 (Carried on his back) — Seen at an angle behind his shoulder; used as Retains his working silhouette and continuity with his arrival at the stall.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Natural daytime ambience, muted saturation, and sober facial contrast allow the exchange to turn tense without an unsupported change of light source.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nRender ONE photorealistic empty location photograph, 16:9, that this shot can be staged inside later. No readable writing anywhere: surfaces that would carry writing may be present, but stage any wording out of legibility — an oblique angle, distance, shallow focus. No captions, watermarks or overlay text.",
  "bg_path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/db00517b-165b-4005-a287-f28936b59b95/images/272bd7af-9294-4490-98c6-88a40e9deacc/scene/recipe/S4sh17__bgfirst_bg.png",
  "asset_id": "31405fa2-f469-42ec-b507-99fd65efe61c",
  "input_asset_ids": [
   "47ad740d-fb4b-46a1-8596-baf4ba0db6b0",
   "a82710e0-4099-4422-b0e7-8cd69a1606bb"
  ]
 },
 "S4sh17": {
  "input_fingerprint": "d95457f3d72ff799",
  "prompt": "Create ONE FINAL photorealistic live-action film still of the moment below — early 20th-century Korea, Jeolla region; all people are Korean unless explicitly stated otherwise. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 짚신전 앞으로 다가온 배상도와 안규홍의 시선이 공중에서 부딪히는 순간.\n\nLOCATION (lock): The outdoor frontage of a straw-sandal stall along a densely packed daytime market lane. The shot takes place here — the attached LOCATION PHOTOGRAPH shows the exact spot.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: medium shot\n- KEY BACKGROUND ELEMENTS: 짚신전 (Stocked with straw shoes beside the encounter) — The display is seen obliquely from its aisle side; used as A peripheral background strip establishes the location without introducing additional visible people; 배상도의 장부 (In his possession after checking the market accounts) — Held low and obliquely, with no readable page content presented to the camera; used as A small lower-frame element identifies the administrative threat without displacing the faces; 안규홍의 지게 (Carried on his back) — Seen at an angle behind his shoulder; used as Retains his working silhouette and continuity with his arrival at the stall.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Natural daytime ambience, muted saturation, and sober facial contrast allow the exchange to turn tense without an unsupported change of light source.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The jade-colored delivery bundle rests beside the straw-shoe stall, and the heavy vegetable basket remains on the platform; steam rises from the soup cauldron. 안규홍: He still wears his carrying frame and old straw shoes with a big toe protruding; the potato is fully eaten. 배상도: He carries the tax ledger at the straw-shoe stall.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 안규홍 (한국인 남성, 청년 나이대, 검은 머리, 단단히 묶은 머리); 배상도 (한국인 남성, 성인 나이대, 자연스러운 검은 머리) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.",
  "roll_prompts": {
   "A": "Stage the shot. The FIRST attached image (SHOT BACKGROUND) is the finished empty background of this shot — keep it EXACTLY: its camera, perspective, architecture, lighting and every fixed feature stay untouched. The SECOND attached image (LAYOUT SKETCH) tells you ONLY where the people go: each sketched person's position, screen size, pose and the gaze/motion arrows. Ignore the sketch's background lines. The CHARACTER REFERENCE photographs show the real people.\nPlace the real people into the background at exactly the sketched positions, sizes and poses, following the arrow directions. No sketch lines or arrows may remain.\n\nCreate ONE FINAL photorealistic live-action film still of the moment below — early 20th-century Korea, Jeolla region; all people are Korean unless explicitly stated otherwise. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 짚신전 앞으로 다가온 배상도와 안규홍의 시선이 공중에서 부딪히는 순간.\n\nLOCATION (lock): The outdoor frontage of a straw-sandal stall along a densely packed daytime market lane. The shot takes place here — the FIRST attached image (SHOT BACKGROUND) is this exact place, already built: its ground, structures, horizon, materials and lighting are the finished truth of this location and must not be redesigned or replaced. No location photograph is attached — read the place from that image alone, and add no scenery, structure, vehicle or fixture that it does not already show. This lock governs the place only; the figures in the shot follow the staging and pose instructions.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: medium shot\n- KEY BACKGROUND ELEMENTS: 짚신전 (Stocked with straw shoes beside the encounter) — The display is seen obliquely from its aisle side; used as A peripheral background strip establishes the location without introducing additional visible people; 배상도의 장부 (In his possession after checking the market accounts) — Held low and obliquely, with no readable page content presented to the camera; used as A small lower-frame element identifies the administrative threat without displacing the faces; 안규홍의 지게 (Carried on his back) — Seen at an angle behind his shoulder; used as Retains his working silhouette and continuity with his arrival at the stall.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Natural daytime ambience, muted saturation, and sober facial contrast allow the exchange to turn tense without an unsupported change of light source.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The jade-colored delivery bundle rests beside the straw-shoe stall, and the heavy vegetable basket remains on the platform; steam rises from the soup cauldron. 안규홍: He still wears his carrying frame and old straw shoes with a big toe protruding; the potato is fully eaten. 배상도: He carries the tax ledger at the straw-shoe stall.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 안규홍 (한국인 남성, 청년 나이대, 검은 머리, 단단히 묶은 머리); 배상도 (한국인 남성, 성인 나이대, 자연스러운 검은 머리) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.",
   "B": "Create ONE FINAL photorealistic live-action film still of the moment below — early 20th-century Korea, Jeolla region; all people are Korean unless explicitly stated otherwise. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 짚신전 앞으로 다가온 배상도와 안규홍의 시선이 공중에서 부딪히는 순간.\n\nLOCATION (lock): The outdoor frontage of a straw-sandal stall along a densely packed daytime market lane. The shot takes place here — the attached LOCATION PHOTOGRAPH shows the exact spot.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: medium shot\n- KEY BACKGROUND ELEMENTS: 짚신전 (Stocked with straw shoes beside the encounter) — The display is seen obliquely from its aisle side; used as A peripheral background strip establishes the location without introducing additional visible people; 배상도의 장부 (In his possession after checking the market accounts) — Held low and obliquely, with no readable page content presented to the camera; used as A small lower-frame element identifies the administrative threat without displacing the faces; 안규홍의 지게 (Carried on his back) — Seen at an angle behind his shoulder; used as Retains his working silhouette and continuity with his arrival at the stall.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Natural daytime ambience, muted saturation, and sober facial contrast allow the exchange to turn tense without an unsupported change of light source.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The jade-colored delivery bundle rests beside the straw-shoe stall, and the heavy vegetable basket remains on the platform; steam rises from the soup cauldron. 안규홍: He still wears his carrying frame and old straw shoes with a big toe protruding; the potato is fully eaten. 배상도: He carries the tax ledger at the straw-shoe stall.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 안규홍 (한국인 남성, 청년 나이대, 검은 머리, 단단히 묶은 머리); 배상도 (한국인 남성, 성인 나이대, 자연스러운 검은 머리) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.\n\nCOMPOSITION VARIATION (this candidate only): keep the place, the people and the moment exactly as specified, and keep from the CAMERA & FRAME contract its subject scale, screen placement and key background elements exactly. For this candidate only, its camera angle and height are deliberately left open: choose a camera position distinctly different from the obvious one, as a film director picking a second setup on the same blocking. Stay true to any explicit framing the SHOT TEXT itself dictates."
  },
  "roll_refs": {
   "A": [
    {
     "label": "SHOT BACKGROUND",
     "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/db00517b-165b-4005-a287-f28936b59b95/images/272bd7af-9294-4490-98c6-88a40e9deacc/scene/recipe/S4sh17__bgfirst_bg.png",
     "asset_id": "31405fa2-f469-42ec-b507-99fd65efe61c",
     "role": "bgfirst_bg"
    },
    {
     "label": "LAYOUT SKETCH (people placement only)",
     "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/db00517b-165b-4005-a287-f28936b59b95/images/272bd7af-9294-4490-98c6-88a40e9deacc/conti/conti_S4sh17.png",
     "asset_id": "47ad740d-fb4b-46a1-8596-baf4ba0db6b0",
     "role": "conti_light"
    },
    {
     "label": "CHARACTER REFERENCE — 안규홍: the exact person appearing in this shot; match face, hair and build exactly.",
     "path": "<bytes:1028384>",
     "asset_id": "3b5894db-ce55-4f34-b966-5e725c3aa952",
     "role": "character_ref"
    },
    {
     "label": "CHARACTER REFERENCE — 배상도: the exact person appearing in this shot; match face, hair and build exactly.",
     "path": "<bytes:851718>",
     "asset_id": "034f5325-975e-4474-854b-e48359710bf8",
     "role": "character_ref"
    }
   ],
   "B": [
    {
     "label": "LOCATION PHOTOGRAPH — the exact place of this shot: its architecture, materials, fixed features and lighting mood are spatial truth; stage the moment inside this place. Never copy its camera framing.",
     "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/db00517b-165b-4005-a287-f28936b59b95/images/272bd7af-9294-4490-98c6-88a40e9deacc/scene/recipe/groupbg_market_shoe_front_6546f9.png",
     "asset_id": "a82710e0-4099-4422-b0e7-8cd69a1606bb",
     "role": "bgfirst_group_bg"
    },
    {
     "label": "CHARACTER REFERENCE — 안규홍: the exact person appearing in this shot; match face, hair and build exactly.",
     "path": "<bytes:1028384>",
     "asset_id": "3b5894db-ce55-4f34-b966-5e725c3aa952",
     "role": "character_ref"
    },
    {
     "label": "CHARACTER REFERENCE — 배상도: the exact person appearing in this shot; match face, hair and build exactly.",
     "path": "<bytes:851718>",
     "asset_id": "034f5325-975e-4474-854b-e48359710bf8",
     "role": "character_ref"
    }
   ]
  },
  "shot_run_spend_attempt_count": 1,
  "judge_flip_skipped": "judge_fn_owns_order",
  "readings": [
   {
    "label": "A",
    "direction": "안규홍과 배상도가 서로 마주 서서 시선을 공중에서 정확하게 부딪히고 있다.",
    "built_space": "왼쪽에 짚신전 매대가 사선 구도로 배치되어 있으며, 배경으로는 가마솥과 시장 길이 원본 사진의 위치에 맞게 자리 잡고 있다.",
    "entities": "지시된 대로 안규홍과 배상도 단 두 사람만 등장한다. 안규홍은 짚신을 신고 지게를 메고 있으며, 평상 위에는 옥색 보따리가 놓여 있다. 무거운 채소 바구니와 배상도의 모자는 구현되지 않았다.",
    "hard_violations": [],
    "physics": "두 사람 모두 흙바닥에 안정적으로 발을 디디고 서 있으며, 안규홍은 어깨에 지게를 메고 배상도는 손과 팔을 이용해 장부를 자연스럽게 지탱하고 있다."
   },
   {
    "label": "B",
    "direction": "두 인물이 비교적 가까운 거리에서 서로의 눈을 똑바로 쳐다보며 대치하고 있다.",
    "built_space": "좌측에 짚신전이 위치하고 배경에 가마솥에서 김이 나는 모습이 있으나, 인물들 주변으로 지시되지 않은 복잡한 시장 상황이 채워져 있다.",
    "entities": "안규홍과 배상도 외에 지시되지 않은 수많은 배경 인물(여성, 남성 행인 및 상인들)이 화면을 채우고 있다. 배상도의 장부에는 식별 가능한 한자와 표 형태의 글씨가 쓰여 있다.",
    "hard_violations": [
     "[gemini-pro] 프롬프트가 엄격히 금지한 다수의 배경 인물(군중) 추가",
     "[gemini-pro] 장부 겉면에 식별 가능한 텍스트 및 표 노출",
     "[gpt-high] 허용된 두 인물 외에 다수의 행인과 부분적으로 가려진 인체가 추가되었다.",
     "[gpt-high] 장부 페이지에 읽을 수 있는 글자가 노출되어 문자 금지 지시를 위반한다."
    ],
    "physics": "안규홍의 지게와 배상도의 장부 모두 신체에 닿아 무리 없이 지탱되고 있으며, 배경의 인물들 역시 땅을 딛고 자연스럽게 서 있거나 걷고 있다."
   }
  ],
  "cross_model_order": {
   "policy": "cross_model_order_v2_slot0_fwd_slot1_rev_parallel2",
   "slots": [
    {
     "model": "gemini-pro",
     "order": "forward",
     "display_to_canonical": {
      "A": "A",
      "B": "B"
     },
     "raw": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 7,
        "verdict_ko": "지정된 미디엄 샷 대신 전신이 보이는 풀 샷으로 프레이밍된 점은 아쉬우나, 금지된 추가 인물을 철저히 배제하고 옥색 보따리와 장부 등 명시된 소품의 상태를 무난히 구현하여 우위를 점합니다."
       },
       {
        "label": "B",
        "score": 2,
        "verdict_ko": "미디엄 샷의 화면 크기는 지시사항에 부합하나, 프롬프트에서 엄격히 금지한 다수의 추가 군중이 배경에 등장하고 장부 겉면에 텍스트가 노출되는 등 치명적인 하드 위반이 발생했습니다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "안규홍과 배상도가 서로 마주 서서 시선을 공중에서 정확하게 부딪히고 있다.",
        "built_space": "왼쪽에 짚신전 매대가 사선 구도로 배치되어 있으며, 배경으로는 가마솥과 시장 길이 원본 사진의 위치에 맞게 자리 잡고 있다.",
        "entities": "지시된 대로 안규홍과 배상도 단 두 사람만 등장한다. 안규홍은 짚신을 신고 지게를 메고 있으며, 평상 위에는 옥색 보따리가 놓여 있다. 무거운 채소 바구니와 배상도의 모자는 구현되지 않았다.",
        "hard_violations": [],
        "physics": "두 사람 모두 흙바닥에 안정적으로 발을 디디고 서 있으며, 안규홍은 어깨에 지게를 메고 배상도는 손과 팔을 이용해 장부를 자연스럽게 지탱하고 있다."
       },
       {
        "label": "B",
        "direction": "두 인물이 비교적 가까운 거리에서 서로의 눈을 똑바로 쳐다보며 대치하고 있다.",
        "built_space": "좌측에 짚신전이 위치하고 배경에 가마솥에서 김이 나는 모습이 있으나, 인물들 주변으로 지시되지 않은 복잡한 시장 상황이 채워져 있다.",
        "entities": "안규홍과 배상도 외에 지시되지 않은 수많은 배경 인물(여성, 남성 행인 및 상인들)이 화면을 채우고 있다. 배상도의 장부에는 식별 가능한 한자와 표 형태의 글씨가 쓰여 있다.",
        "hard_violations": [
         "프롬프트가 엄격히 금지한 다수의 배경 인물(군중) 추가",
         "장부 겉면에 식별 가능한 텍스트 및 표 노출"
        ],
        "physics": "안규홍의 지게와 배상도의 장부 모두 신체에 닿아 무리 없이 지탱되고 있으며, 배경의 인물들 역시 땅을 딛고 자연스럽게 서 있거나 걷고 있다."
       }
      ],
      "all_candidates_fail": false
     },
     "normalized": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 7,
        "verdict_ko": "지정된 미디엄 샷 대신 전신이 보이는 풀 샷으로 프레이밍된 점은 아쉬우나, 금지된 추가 인물을 철저히 배제하고 옥색 보따리와 장부 등 명시된 소품의 상태를 무난히 구현하여 우위를 점합니다."
       },
       {
        "label": "B",
        "score": 2,
        "verdict_ko": "미디엄 샷의 화면 크기는 지시사항에 부합하나, 프롬프트에서 엄격히 금지한 다수의 추가 군중이 배경에 등장하고 장부 겉면에 텍스트가 노출되는 등 치명적인 하드 위반이 발생했습니다."
       }
      ],
      "all_candidates_fail": false,
      "readings": [
       {
        "label": "A",
        "direction": "안규홍과 배상도가 서로 마주 서서 시선을 공중에서 정확하게 부딪히고 있다.",
        "built_space": "왼쪽에 짚신전 매대가 사선 구도로 배치되어 있으며, 배경으로는 가마솥과 시장 길이 원본 사진의 위치에 맞게 자리 잡고 있다.",
        "entities": "지시된 대로 안규홍과 배상도 단 두 사람만 등장한다. 안규홍은 짚신을 신고 지게를 메고 있으며, 평상 위에는 옥색 보따리가 놓여 있다. 무거운 채소 바구니와 배상도의 모자는 구현되지 않았다.",
        "hard_violations": [],
        "physics": "두 사람 모두 흙바닥에 안정적으로 발을 디디고 서 있으며, 안규홍은 어깨에 지게를 메고 배상도는 손과 팔을 이용해 장부를 자연스럽게 지탱하고 있다."
       },
       {
        "label": "B",
        "direction": "두 인물이 비교적 가까운 거리에서 서로의 눈을 똑바로 쳐다보며 대치하고 있다.",
        "built_space": "좌측에 짚신전이 위치하고 배경에 가마솥에서 김이 나는 모습이 있으나, 인물들 주변으로 지시되지 않은 복잡한 시장 상황이 채워져 있다.",
        "entities": "안규홍과 배상도 외에 지시되지 않은 수많은 배경 인물(여성, 남성 행인 및 상인들)이 화면을 채우고 있다. 배상도의 장부에는 식별 가능한 한자와 표 형태의 글씨가 쓰여 있다.",
        "hard_violations": [
         "프롬프트가 엄격히 금지한 다수의 배경 인물(군중) 추가",
         "장부 겉면에 식별 가능한 텍스트 및 표 노출"
        ],
        "physics": "안규홍의 지게와 배상도의 장부 모두 신체에 닿아 무리 없이 지탱되고 있으며, 배경의 인물들 역시 땅을 딛고 자연스럽게 서 있거나 걷고 있다."
       }
      ]
     },
     "ok": true
    },
    {
     "model": "gpt-high",
     "order": "reverse",
     "display_to_canonical": {
      "A": "B",
      "B": "A"
     },
     "raw": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 2,
        "verdict_ko": "두 사람의 눈맞춤은 맞지만, 금지된 행인 다수와 읽을 수 있는 장부 글자가 등장하여 실격이며 장부 면도 카메라를 향한다."
       },
       {
        "label": "B",
        "score": 6,
        "verdict_ko": "추가 인물 없이 눈맞춤과 장부·지게의 물리적 배치는 지켰으나, 미디엄 숏 대신 전신 와이드로 넓혀 얼굴 중심의 대치를 약화했다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "왼쪽 안규홍은 오른쪽 배상도의 눈을, 배상도는 안규홍의 눈을 바라봐 시선의 충돌은 성립한다. 배상도의 장부는 아래에 들려 있지만 페이지 면이 관객 쪽으로 거의 정면 노출되어, 비스듬히 숨기라는 지시와 어긋난다.",
        "built_space": "왼쪽에 짚신을 세 단으로 건 진열대 하나, 그 옆 발 하나, 아래 목재 평상 하나가 보이고 뒤쪽에는 화덕 위 가마솥 하나가 있다. 목조 처마와 흙길은 장소 참고와 유사하다. 두 주인공은 평상 옆 통로에 서 있으나 그 사이와 뒤 통로를 여러 행인이 채운다. 진열대가 화면 왼쪽의 큰 면적을 차지하여 주변부의 좁은 배경 띠라는 요구보다 강하게 드러난다.",
        "entities": "안규홍은 검은 머리를 뒤로 묶은 젊은 한국인 남성으로 보이며 낡은 담색 옷과 지게를 갖췄다. 배상도는 성인 한국인 남성으로 보이고 흰옷과 검은 조끼는 맞지만 참고의 검은 모자는 없다. 두 얼굴의 정확한 동일성은 참고와 차이가 있다. 지정되지 않은 남녀 행인이 다수 보인다. 장부에는 판독 가능한 글자와 표가 노출된다. 평상 위 큰 바구니의 내용물은 채소보다 짚 제품으로 보이고, 지게에는 별도의 바구니가 실려 있다. 가마솥에서는 김이 난다. 옥색 보따리는 보이지 않으며 주인공의 발은 화면 밖이라 짚신과 엄지발가락 상태는 판단할 수 없다.",
        "hard_violations": [
         "허용된 두 인물 외에 다수의 행인과 부분적으로 가려진 인체가 추가되었다.",
         "장부 페이지에 읽을 수 있는 글자가 노출되어 문자 금지 지시를 위반한다."
        ],
        "physics": "안규홍의 지게는 어깨끈과 등에 의해 지지되고, 실린 바구니는 지게의 가로대에 받쳐져 있다. 배상도는 손으로 장부 아래쪽을 잡고 있다. 평상은 다리로, 바구니는 평상으로, 가마솥은 화덕으로 지지된다. 주인공의 발은 잘렸지만 상체와 다리 자세에 부유나 불가능한 동작은 보이지 않는다."
       },
       {
        "label": "B",
        "direction": "왼쪽 안규홍과 오른쪽 배상도가 서로의 눈과 얼굴을 정면으로 바라본다. 배상도는 장부를 허벅지 옆에 낮게 들고 있으며 닫힌 표지만 카메라에 보여 페이지 내용은 드러나지 않는다. 안규홍의 손은 어깨의 지게끈을 잡고 있다.",
        "built_space": "왼쪽 짚신전에는 세 단 진열대 하나, 옆의 발 하나, 앞쪽 평상 하나가 보인다. 평상에는 보따리 하나와 큰 바구니 하나가 놓였고, 뒤쪽 낮은 받침에는 작은 나무통 하나, 그 옆에는 화덕과 가마솥 하나가 있다. 목조 상점과 흙길의 배열은 장소 참고에 가깝고 두 사람 외에는 아무도 없다. 다만 두 사람의 발부터 머리까지와 넓은 시장길을 담은 전신 와이드여서 지정된 미디엄 숏이 아니며, 짚신전도 주변부의 좁은 띠보다 크게 보인다.",
        "entities": "안규홍은 묶은 검은 머리의 젊은 한국인 남성으로 보이며 낡은 담색 옷, 등 뒤 지게, 발가락이 드러나는 짚신을 갖췄다. 배상도는 한국인 성인 남성으로 보이고 흰옷과 검은 조끼를 입었지만, 참고보다 나이 들어 보이는 얼굴이며 검은 모자가 없다. 장부는 닫혀 있고 읽을 수 있는 글자는 없다. 옥색 보따리는 평상 위에 놓여 있다. 큰 바구니는 존재하지만 내용물이 채소가 아니라 짚신 또는 짚 제품으로 보인다. 가마솥의 김이 보이고 먹다 남은 감자는 없다.",
        "hard_violations": [],
        "physics": "두 사람 모두 흙바닥에 발을 딛고 서 있다. 안규홍의 지게는 어깨끈과 등에 지지되며 한 손이 끈을 잡아 하중을 받는다. 배상도의 장부는 손에 확실히 잡혀 있다. 보따리와 큰 바구니는 평상에, 작은 나무통은 낮은 받침에, 가마솥은 화덕에 놓여 있다. 지지 없이 뜬 물체나 불가능한 신체 자세는 보이지 않는다."
       }
      ],
      "all_candidates_fail": true
     },
     "normalized": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 2,
        "verdict_ko": "두 사람의 눈맞춤은 맞지만, 금지된 행인 다수와 읽을 수 있는 장부 글자가 등장하여 실격이며 장부 면도 카메라를 향한다."
       },
       {
        "label": "A",
        "score": 6,
        "verdict_ko": "추가 인물 없이 눈맞춤과 장부·지게의 물리적 배치는 지켰으나, 미디엄 숏 대신 전신 와이드로 넓혀 얼굴 중심의 대치를 약화했다."
       }
      ],
      "all_candidates_fail": true,
      "readings": [
       {
        "label": "B",
        "direction": "왼쪽 안규홍은 오른쪽 배상도의 눈을, 배상도는 안규홍의 눈을 바라봐 시선의 충돌은 성립한다. 배상도의 장부는 아래에 들려 있지만 페이지 면이 관객 쪽으로 거의 정면 노출되어, 비스듬히 숨기라는 지시와 어긋난다.",
        "built_space": "왼쪽에 짚신을 세 단으로 건 진열대 하나, 그 옆 발 하나, 아래 목재 평상 하나가 보이고 뒤쪽에는 화덕 위 가마솥 하나가 있다. 목조 처마와 흙길은 장소 참고와 유사하다. 두 주인공은 평상 옆 통로에 서 있으나 그 사이와 뒤 통로를 여러 행인이 채운다. 진열대가 화면 왼쪽의 큰 면적을 차지하여 주변부의 좁은 배경 띠라는 요구보다 강하게 드러난다.",
        "entities": "안규홍은 검은 머리를 뒤로 묶은 젊은 한국인 남성으로 보이며 낡은 담색 옷과 지게를 갖췄다. 배상도는 성인 한국인 남성으로 보이고 흰옷과 검은 조끼는 맞지만 참고의 검은 모자는 없다. 두 얼굴의 정확한 동일성은 참고와 차이가 있다. 지정되지 않은 남녀 행인이 다수 보인다. 장부에는 판독 가능한 글자와 표가 노출된다. 평상 위 큰 바구니의 내용물은 채소보다 짚 제품으로 보이고, 지게에는 별도의 바구니가 실려 있다. 가마솥에서는 김이 난다. 옥색 보따리는 보이지 않으며 주인공의 발은 화면 밖이라 짚신과 엄지발가락 상태는 판단할 수 없다.",
        "hard_violations": [
         "허용된 두 인물 외에 다수의 행인과 부분적으로 가려진 인체가 추가되었다.",
         "장부 페이지에 읽을 수 있는 글자가 노출되어 문자 금지 지시를 위반한다."
        ],
        "physics": "안규홍의 지게는 어깨끈과 등에 의해 지지되고, 실린 바구니는 지게의 가로대에 받쳐져 있다. 배상도는 손으로 장부 아래쪽을 잡고 있다. 평상은 다리로, 바구니는 평상으로, 가마솥은 화덕으로 지지된다. 주인공의 발은 잘렸지만 상체와 다리 자세에 부유나 불가능한 동작은 보이지 않는다."
       },
       {
        "label": "A",
        "direction": "왼쪽 안규홍과 오른쪽 배상도가 서로의 눈과 얼굴을 정면으로 바라본다. 배상도는 장부를 허벅지 옆에 낮게 들고 있으며 닫힌 표지만 카메라에 보여 페이지 내용은 드러나지 않는다. 안규홍의 손은 어깨의 지게끈을 잡고 있다.",
        "built_space": "왼쪽 짚신전에는 세 단 진열대 하나, 옆의 발 하나, 앞쪽 평상 하나가 보인다. 평상에는 보따리 하나와 큰 바구니 하나가 놓였고, 뒤쪽 낮은 받침에는 작은 나무통 하나, 그 옆에는 화덕과 가마솥 하나가 있다. 목조 상점과 흙길의 배열은 장소 참고에 가깝고 두 사람 외에는 아무도 없다. 다만 두 사람의 발부터 머리까지와 넓은 시장길을 담은 전신 와이드여서 지정된 미디엄 숏이 아니며, 짚신전도 주변부의 좁은 띠보다 크게 보인다.",
        "entities": "안규홍은 묶은 검은 머리의 젊은 한국인 남성으로 보이며 낡은 담색 옷, 등 뒤 지게, 발가락이 드러나는 짚신을 갖췄다. 배상도는 한국인 성인 남성으로 보이고 흰옷과 검은 조끼를 입었지만, 참고보다 나이 들어 보이는 얼굴이며 검은 모자가 없다. 장부는 닫혀 있고 읽을 수 있는 글자는 없다. 옥색 보따리는 평상 위에 놓여 있다. 큰 바구니는 존재하지만 내용물이 채소가 아니라 짚신 또는 짚 제품으로 보인다. 가마솥의 김이 보이고 먹다 남은 감자는 없다.",
        "hard_violations": [],
        "physics": "두 사람 모두 흙바닥에 발을 딛고 서 있다. 안규홍의 지게는 어깨끈과 등에 지지되며 한 손이 끈을 잡아 하중을 받는다. 배상도의 장부는 손에 확실히 잡혀 있다. 보따리와 큰 바구니는 평상에, 작은 나무통은 낮은 받침에, 가마솥은 화덕에 놓여 있다. 지지 없이 뜬 물체나 불가능한 신체 자세는 보이지 않는다."
       }
      ]
     },
     "ok": true
    }
   ],
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "slot_winner_match": true,
   "slot_winner": {
    "gemini-pro": "A",
    "gpt-high": "A"
   },
   "route": "cross_slot_agree"
  },
  "dual": {
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "normalized": {
    "A": 2.0,
    "B": 0.619
   },
   "adjusted": {
    "A": 2.0,
    "B": 0.369
   },
   "violations": {
    "B": [
     "[gemini-pro] 프롬프트가 엄격히 금지한 다수의 배경 인물(군중) 추가",
     "[gemini-pro] 장부 겉면에 식별 가능한 텍스트 및 표 노출",
     "[gpt-high] 허용된 두 인물 외에 다수의 행인과 부분적으로 가려진 인체가 추가되었다.",
     "[gpt-high] 장부 페이지에 읽을 수 있는 글자가 노출되어 문자 금지 지시를 위반한다."
    ]
   },
   "per_model_winner": {
    "gemini-pro": "A",
    "gpt-high": "A"
   },
   "agreed": true
  },
  "totals": {
   "A": 2000,
   "B": 369
  },
  "selected": "A",
  "ranking": [
   "A",
   "B"
  ],
  "verdicts": [
   {
    "label": "A",
    "score": 2000,
    "verdict_ko": "지정된 미디엄 샷 대신 전신이 보이는 풀 샷으로 프레이밍된 점은 아쉬우나, 금지된 추가 인물을 철저히 배제하고 옥색 보따리와 장부 등 명시된 소품의 상태를 무난히 구현하여 우위를 점합니다."
   },
   {
    "label": "B",
    "score": 369,
    "verdict_ko": "미디엄 샷의 화면 크기는 지시사항에 부합하나, 프롬프트에서 엄격히 금지한 다수의 추가 군중이 배경에 등장하고 장부 겉면에 텍스트가 노출되는 등 치명적인 하드 위반이 발생했습니다.  ★위반: [gemini-pro] 프롬프트가 엄격히 금지한 다수의 배경 인물(군중) 추가 / [gemini-pro] 장부 겉면에 식별 가능한 텍스트 및 표 노출 / [gpt-high] 허용된 두 인물 외에 다수의 행인과 부분적으로 가려진 인체가 추가되었다. / [gpt-high] 장부 페이지에 읽을 수 있는 글자가 노출되어 문자 금지 지시를 위반한다."
   }
  ],
  "refs": [
   {
    "label": "LOCATION PHOTOGRAPH — the exact place of this shot: its architecture, materials, fixed features and lighting mood are spatial truth; stage the moment inside this place. Never copy its camera framing.",
    "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/db00517b-165b-4005-a287-f28936b59b95/images/272bd7af-9294-4490-98c6-88a40e9deacc/scene/recipe/groupbg_market_shoe_front_6546f9.png",
    "asset_id": "a82710e0-4099-4422-b0e7-8cd69a1606bb",
    "role": "bgfirst_group_bg"
   },
   {
    "label": "CHARACTER REFERENCE — 안규홍: the exact person appearing in this shot; match face, hair and build exactly.",
    "path": "<bytes:1028384>",
    "asset_id": "3b5894db-ce55-4f34-b966-5e725c3aa952",
    "role": "character_ref"
   },
   {
    "label": "CHARACTER REFERENCE — 배상도: the exact person appearing in this shot; match face, hair and build exactly.",
    "path": "<bytes:851718>",
    "asset_id": "034f5325-975e-4474-854b-e48359710bf8",
    "role": "character_ref"
   }
  ],
  "critique_skipped": true,
  "shot_run_produced": false,
  "shot_run_uid": "06aa1406-ee64-7e26-9c76-67387fdf8a38",
  "bgfirst": {
   "bg_path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/db00517b-165b-4005-a287-f28936b59b95/images/272bd7af-9294-4490-98c6-88a40e9deacc/scene/recipe/S4sh17__bgfirst_bg.png",
   "bg_asset_id": "31405fa2-f469-42ec-b507-99fd65efe61c",
   "bg_record_key": "S4sh17::bgfirst_bg",
   "chain_winner": true,
   "authority": "groupbg",
   "group_key": "market_shoe_front",
   "groupbg_asset_id": "a82710e0-4099-4422-b0e7-8cd69a1606bb"
  },
  "ref_mode": "재투영 배경+콘티+엔티티 (2택1: 체인 승)",
  "share_plan": {
   "ref_plan": "background"
  }
 },
 "S4sh17::cine": {
  "applied": true,
  "attempted_at": "2026-09-09T01:00:27.437179+00:00",
  "fingerprint": "1b01afe2a2b0f38689adac81f54b20e0d353cddb004a00651b0377cceba09a87",
  "fingerprint_version": 2,
  "provider": "mai",
  "endpoint": "openrouter/chat-completions",
  "model": "microsoft/mai-image-2.6",
  "pack": "24.202608252115",
  "source_file": "S4sh17_sel.png",
  "source_sha256": "f74495463aecfdfdfd45ca08e5381e23b8153035b48bd71df73fc8485953ef84",
  "file": "S4sh17_cine.png",
  "staged_sha256": "6f57d2138295237bd9d8fde80c2ce894e38357ee3f3b7905d46d7f1ef54ad4ba",
  "latency_ms": 26232
 },
 "S4sh24::signage": {
  "fp": "b5ca0bcff15a2463",
  "inscriptions": [],
  "cues": [],
  "dropped": []
 },
 "S4sh24": {
  "input_fingerprint": "14b4042c697eec77",
  "prompt": "Create ONE FINAL photorealistic live-action film still of the moment below — early 20th-century Korea, Jeolla region; all people are Korean unless explicitly stated otherwise. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 나서는 안규홍의 굵은 팔뚝을 이관회의 두 손이 꽉 움켜쥔 찰나.\n\nLOCATION (lock): The dirt standing space directly in front of the market's straw-sandal stall, beside the vendor's mat. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: close-up\n- KEY BACKGROUND ELEMENTS: 짚신전 (The confrontation remains immediately beside the stall) — Only a soft, partial view of the stall is visible beyond the cropped torsos; used as Maintains the location without distracting from the joined hands and arm.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Natural daylight with restrained saturation and consistent, naturalistic contrast keeps the interrupted movement tense without theatrical emphasis.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The clothing, hair and overall look of 안규홍 — who appear both in that photo and in this shot — are LOCKED to that photo. Anyone else visible in that photo is NOT in this shot: never carry their face, body or clothing onto anyone here. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nPREVIOUS STILL USAGE (follow exactly — what to take from the attached still and what to exclude; it governs place and objects only, never who is in this shot or how the camera sees them): Take the straw-sandal stall, its woven display materials, and the sunlit dirt ground visible in the reference. Exclude the ledger-checking action and do not retain the earlier undisturbed arrangement of loose sandals.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The delivery bundle remains beside the stall, and the vegetable basket remains on the platform. 안규홍: He has halted with his carrying frame still on and his big toe protruding through an old straw shoe. 이관회: He stands with his arms extended in a firm restraining grip.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 안규홍 (한국인 남성, 청년 나이대, 검은 머리, 단단히 묶은 머리); 이관회 (한국인 남성, 청년 나이대, 자연스러운 검은 머리) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.",
  "roll_prompts": {
   "A": "Create ONE FINAL photorealistic live-action film still of the moment below — early 20th-century Korea, Jeolla region; all people are Korean unless explicitly stated otherwise. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 나서는 안규홍의 굵은 팔뚝을 이관회의 두 손이 꽉 움켜쥔 찰나.\n\nLOCATION (lock): The dirt standing space directly in front of the market's straw-sandal stall, beside the vendor's mat. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: close-up\n- KEY BACKGROUND ELEMENTS: 짚신전 (The confrontation remains immediately beside the stall) — Only a soft, partial view of the stall is visible beyond the cropped torsos; used as Maintains the location without distracting from the joined hands and arm.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Natural daylight with restrained saturation and consistent, naturalistic contrast keeps the interrupted movement tense without theatrical emphasis.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The clothing, hair and overall look of 안규홍 — who appear both in that photo and in this shot — are LOCKED to that photo. Anyone else visible in that photo is NOT in this shot: never carry their face, body or clothing onto anyone here. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nPREVIOUS STILL USAGE (follow exactly — what to take from the attached still and what to exclude; it governs place and objects only, never who is in this shot or how the camera sees them): Take the straw-sandal stall, its woven display materials, and the sunlit dirt ground visible in the reference. Exclude the ledger-checking action and do not retain the earlier undisturbed arrangement of loose sandals.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The delivery bundle remains beside the stall, and the vegetable basket remains on the platform. 안규홍: He has halted with his carrying frame still on and his big toe protruding through an old straw shoe. 이관회: He stands with his arms extended in a firm restraining grip.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 안규홍 (한국인 남성, 청년 나이대, 검은 머리, 단단히 묶은 머리); 이관회 (한국인 남성, 청년 나이대, 자연스러운 검은 머리) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.",
   "B": "Create ONE FINAL photorealistic live-action film still of the moment below — early 20th-century Korea, Jeolla region; all people are Korean unless explicitly stated otherwise. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 나서는 안규홍의 굵은 팔뚝을 이관회의 두 손이 꽉 움켜쥔 찰나.\n\nLOCATION (lock): The dirt standing space directly in front of the market's straw-sandal stall, beside the vendor's mat. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: close-up\n- KEY BACKGROUND ELEMENTS: 짚신전 (The confrontation remains immediately beside the stall) — Only a soft, partial view of the stall is visible beyond the cropped torsos; used as Maintains the location without distracting from the joined hands and arm.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Natural daylight with restrained saturation and consistent, naturalistic contrast keeps the interrupted movement tense without theatrical emphasis.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The clothing, hair and overall look of 안규홍 — who appear both in that photo and in this shot — are LOCKED to that photo. Anyone else visible in that photo is NOT in this shot: never carry their face, body or clothing onto anyone here. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nPREVIOUS STILL USAGE (follow exactly — what to take from the attached still and what to exclude; it governs place and objects only, never who is in this shot or how the camera sees them): Take the straw-sandal stall, its woven display materials, and the sunlit dirt ground visible in the reference. Exclude the ledger-checking action and do not retain the earlier undisturbed arrangement of loose sandals.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The delivery bundle remains beside the stall, and the vegetable basket remains on the platform. 안규홍: He has halted with his carrying frame still on and his big toe protruding through an old straw shoe. 이관회: He stands with his arms extended in a firm restraining grip.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 안규홍 (한국인 남성, 청년 나이대, 검은 머리, 단단히 묶은 머리); 이관회 (한국인 남성, 청년 나이대, 자연스러운 검은 머리) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.\n\nCOMPOSITION VARIATION (this candidate only): keep the place, the people and the moment exactly as specified, and keep from the CAMERA & FRAME contract its subject scale, screen placement and key background elements exactly. For this candidate only, its camera angle and height are deliberately left open: choose a camera position distinctly different from the obvious one, as a film director picking a second setup on the same blocking. Stay true to any explicit framing the SHOT TEXT itself dictates."
  },
  "shot_run_spend_attempt_count": 1,
  "readings": [
   {
    "label": "A",
    "direction": "카메라와 시선이 두 인물의 팔이 교차하는 중심부에 집중되어 있음. 오른쪽 인물의 두 손이 왼쪽 인물의 굵은 오른팔을 강하게 붙잡고 있음.",
    "built_space": "짚신전 앞 흙바닥 뷰. 배경에 매달린 짚신들이 아웃포커싱되어 부드럽게 보이며, 프레이밍 지시에 따라 좌판의 일부만 제한적으로 노출됨.",
    "entities": "안규홍(왼쪽): 흰색의 낡은 옷, 지게 착용. 이관회(오른쪽): 갈색 거친 질감의 저고리 착용. 두 인물의 복장 모두 레퍼런스와 일치함.",
    "hard_violations": [],
    "physics": "두 인물이 지면에 서서 힘을 겨루고 있으며, 팔을 뻗어 상대방의 팔뚝과 손목을 쥔 손의 형태와 그립이 물리적으로 자연스럽게 지탱됨."
   },
   {
    "label": "B",
    "direction": "두 인물이 서로의 얼굴을 마주보며 대치하고 있음. 오른손을 서로 맞잡고, 오른쪽 인물의 왼손이 상대의 손목을 잡고 있음.",
    "built_space": "짚신전 앞 흙바닥 뷰. 배경의 짚신 좌판이 넓게 보이며, 툇마루 위 이전 샷의 초록색 보따리가 그대로 놓여 있음.",
    "entities": "안규홍(왼쪽)과 이관회(오른쪽)의 얼굴, 헤어스타일, 복장 모두 캐릭터 레퍼런스와 일치함. 안규홍은 지게를 매고 있음.",
    "hard_violations": [],
    "physics": "두 인물이 지면에 안정적으로 서서 손을 맞잡고 힘을 주고 있으며, 맞잡은 팔의 근육과 자세가 동작을 잘 뒷받침함."
   }
  ],
  "cross_model_order": {
   "policy": "cross_model_order_v2_slot0_fwd_slot1_rev_parallel2",
   "slots": [
    {
     "model": "gemini-pro",
     "order": "forward",
     "display_to_canonical": {
      "A": "A",
      "B": "B"
     },
     "raw": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 9,
        "verdict_ko": "지시문이 엄격하게 요구한 클로즈업 프레이밍(얼굴이 배제된 토르소 크롭)과 '두 손으로 팔뚝을 꽉 움켜쥐는' 핵심 동작을 정확하게 구현했습니다."
       },
       {
        "label": "B",
        "score": 4,
        "verdict_ko": "프레이밍 지시(클로즈업, 크롭된 토르소)를 완전히 무시하고 인물들의 얼굴과 상반신을 모두 노출시켰으며, 동작 또한 두 손으로 팔뚝을 잡는 것이 아닌 손을 맞잡는 형태로 임의 변형되었습니다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "카메라와 시선이 두 인물의 팔이 교차하는 중심부에 집중되어 있음. 오른쪽 인물의 두 손이 왼쪽 인물의 굵은 오른팔을 강하게 붙잡고 있음.",
        "built_space": "짚신전 앞 흙바닥 뷰. 배경에 매달린 짚신들이 아웃포커싱되어 부드럽게 보이며, 프레이밍 지시에 따라 좌판의 일부만 제한적으로 노출됨.",
        "entities": "안규홍(왼쪽): 흰색의 낡은 옷, 지게 착용. 이관회(오른쪽): 갈색 거친 질감의 저고리 착용. 두 인물의 복장 모두 레퍼런스와 일치함.",
        "hard_violations": [],
        "physics": "두 인물이 지면에 서서 힘을 겨루고 있으며, 팔을 뻗어 상대방의 팔뚝과 손목을 쥔 손의 형태와 그립이 물리적으로 자연스럽게 지탱됨."
       },
       {
        "label": "B",
        "direction": "두 인물이 서로의 얼굴을 마주보며 대치하고 있음. 오른손을 서로 맞잡고, 오른쪽 인물의 왼손이 상대의 손목을 잡고 있음.",
        "built_space": "짚신전 앞 흙바닥 뷰. 배경의 짚신 좌판이 넓게 보이며, 툇마루 위 이전 샷의 초록색 보따리가 그대로 놓여 있음.",
        "entities": "안규홍(왼쪽)과 이관회(오른쪽)의 얼굴, 헤어스타일, 복장 모두 캐릭터 레퍼런스와 일치함. 안규홍은 지게를 매고 있음.",
        "hard_violations": [],
        "physics": "두 인물이 지면에 안정적으로 서서 손을 맞잡고 힘을 주고 있으며, 맞잡은 팔의 근육과 자세가 동작을 잘 뒷받침함."
       }
      ],
      "all_candidates_fail": false
     },
     "normalized": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 9,
        "verdict_ko": "지시문이 엄격하게 요구한 클로즈업 프레이밍(얼굴이 배제된 토르소 크롭)과 '두 손으로 팔뚝을 꽉 움켜쥐는' 핵심 동작을 정확하게 구현했습니다."
       },
       {
        "label": "B",
        "score": 4,
        "verdict_ko": "프레이밍 지시(클로즈업, 크롭된 토르소)를 완전히 무시하고 인물들의 얼굴과 상반신을 모두 노출시켰으며, 동작 또한 두 손으로 팔뚝을 잡는 것이 아닌 손을 맞잡는 형태로 임의 변형되었습니다."
       }
      ],
      "all_candidates_fail": false,
      "readings": [
       {
        "label": "A",
        "direction": "카메라와 시선이 두 인물의 팔이 교차하는 중심부에 집중되어 있음. 오른쪽 인물의 두 손이 왼쪽 인물의 굵은 오른팔을 강하게 붙잡고 있음.",
        "built_space": "짚신전 앞 흙바닥 뷰. 배경에 매달린 짚신들이 아웃포커싱되어 부드럽게 보이며, 프레이밍 지시에 따라 좌판의 일부만 제한적으로 노출됨.",
        "entities": "안규홍(왼쪽): 흰색의 낡은 옷, 지게 착용. 이관회(오른쪽): 갈색 거친 질감의 저고리 착용. 두 인물의 복장 모두 레퍼런스와 일치함.",
        "hard_violations": [],
        "physics": "두 인물이 지면에 서서 힘을 겨루고 있으며, 팔을 뻗어 상대방의 팔뚝과 손목을 쥔 손의 형태와 그립이 물리적으로 자연스럽게 지탱됨."
       },
       {
        "label": "B",
        "direction": "두 인물이 서로의 얼굴을 마주보며 대치하고 있음. 오른손을 서로 맞잡고, 오른쪽 인물의 왼손이 상대의 손목을 잡고 있음.",
        "built_space": "짚신전 앞 흙바닥 뷰. 배경의 짚신 좌판이 넓게 보이며, 툇마루 위 이전 샷의 초록색 보따리가 그대로 놓여 있음.",
        "entities": "안규홍(왼쪽)과 이관회(오른쪽)의 얼굴, 헤어스타일, 복장 모두 캐릭터 레퍼런스와 일치함. 안규홍은 지게를 매고 있음.",
        "hard_violations": [],
        "physics": "두 인물이 지면에 안정적으로 서서 손을 맞잡고 힘을 주고 있으며, 맞잡은 팔의 근육과 자세가 동작을 잘 뒷받침함."
       }
      ]
     },
     "ok": true
    },
    {
     "model": "gpt-high",
     "order": "reverse",
     "display_to_canonical": {
      "A": "B",
      "B": "A"
     },
     "raw": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 5,
        "verdict_ko": "얼굴과 지게까지 드러낸 넓은 구도이며, 한 손은 손목을 잡고 다른 손은 맞잡는 형태라 ‘두 손으로 굵은 팔뚝을 움켜쥔 찰나’의 클로즈업과 다릅니다."
       },
       {
        "label": "B",
        "score": 9,
        "verdict_ko": "잘린 몸통 사이에서 이관회의 두 손이 안규홍의 팔뚝을 단단히 붙잡는 클로즈업으로, 핵심 동작과 흐릿한 짚신전 배경을 충실히 구현했습니다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "왼쪽 안규홍과 오른쪽 이관회가 서로의 얼굴을 바라봅니다. 안규홍의 팔은 오른쪽 위로 굽혀져 있고, 이관회의 한 손은 그 손목 부근을 감싸지만 다른 손은 안규홍의 손과 맞잡혀 있습니다. 두 손 모두 팔뚝을 붙잡는 방향과 접촉은 아닙니다.",
        "built_space": "두 사람 뒤에 목재 짚신 진열대 하나와 낮은 판매 평상 하나가 보입니다. 짚신은 세 높이의 가로 진열부에 걸려 있으며, 평상 위에는 청록색 보따리 하나와 큰 바구니 하나가 놓여 있습니다. 두 사람은 평상 앞에 있지만, 배경과 얼굴을 넓고 비교적 선명하게 보여 주어 요청한 부분적이고 부드러운 배경보다 비중이 큽니다.",
        "entities": "한국인 청년 남성으로 설정된 두 인물만 보입니다. 안규홍의 검은 묶음머리, 밝은색 낡은 옷과 목제 지게는 참조와 대체로 맞지만, 소매를 어깨 가까이까지 올려 이전 장면보다 상완이 크게 노출됩니다. 이관회는 자연스러운 검은 머리와 갈색의 해진 저고리를 입어 인물 참조에 부합합니다. 바구니 내용물은 채소로 확인되지 않으며, 신발과 발가락은 화면 밖입니다. 읽을 수 있는 글자는 없습니다.",
        "hard_violations": [],
        "physics": "지게는 어깨끈으로 안규홍의 몸에 지지되고, 보따리와 바구니는 평상에 놓여 있습니다. 두 사람의 손과 팔은 신체에 자연스럽게 연결되어 있고 접촉도 가능합니다. 다만 팔을 높여 서로 손을 맞잡은 자세는 나서려는 팔뚝을 두 손으로 제지하는 순간보다 힘겨루기에 가깝습니다. 발은 잘려 있어 지면 접촉은 확인할 수 없지만 공중에 떠 있는 신체나 물체는 보이지 않습니다."
       },
       {
        "label": "B",
        "direction": "이관회의 두 팔이 오른쪽에서 왼쪽 안규홍의 팔뚝으로 뻗습니다. 한 손은 팔뚝 중간을 아래에서 감싸고, 다른 손은 손목에 가까운 팔뚝을 붙잡습니다. 두 손의 목표가 동일한 팔뚝으로 명확하며, 안규홍의 팔이 오른쪽으로 나가는 것을 막는 관계입니다. 얼굴과 시선은 화면 밖입니다.",
        "built_space": "잘린 두 몸통 뒤로 목재 짚신 진열대 하나, 낮은 평상 하나와 햇빛을 받은 흙바닥이 부분적으로 보입니다. 짚신이 걸린 가로 진열부는 적어도 두 높이에서 보이며, 아래쪽에는 둥근 짚 직물이 기대어 있습니다. 인물들은 평상 바로 앞의 흙 공간에 있고, 배경은 흐려져 손과 팔을 방해하지 않습니다. 고정 시설의 중복이나 불가능한 공간 배치는 보이지 않습니다.",
        "entities": "안규홍의 밝은색 낡은 저고리, 걷어 올린 소매, 굵은 팔뚝과 지게 일부가 보입니다. 이관회는 참조의 갈색 해진 저고리와 밝은색 바지를 입고 있습니다. 보이는 팔과 손은 청년 남성의 외형에 부합하지만, 얼굴과 머리가 잘려 정확한 얼굴 정체성과 머리 모양은 검증할 수 없습니다. 두 인물 외의 사람이나 읽을 수 있는 글자는 없습니다. 보따리, 채소 바구니와 신발은 이 클로즈업 범위에서 확인되지 않습니다.",
        "hard_violations": [],
        "physics": "이관회의 두 손이 팔뚝의 서로 다른 지점을 실제로 감싸며 손가락과 엄지의 맞물림이 제지력을 전달합니다. 안규홍의 잡힌 팔과 아래로 내려온 반대 손은 각각 몸에 연결되어 있습니다. 지게는 몸을 가로지르는 끈과 등에 의해 지지됩니다. 발은 화면 밖이지만, 보이는 상체와 팔의 자세는 서서 상대의 전진을 막는 동작으로 가능하며 지지 없이 떠 있는 물체는 없습니다."
       }
      ],
      "all_candidates_fail": false
     },
     "normalized": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 5,
        "verdict_ko": "얼굴과 지게까지 드러낸 넓은 구도이며, 한 손은 손목을 잡고 다른 손은 맞잡는 형태라 ‘두 손으로 굵은 팔뚝을 움켜쥔 찰나’의 클로즈업과 다릅니다."
       },
       {
        "label": "A",
        "score": 9,
        "verdict_ko": "잘린 몸통 사이에서 이관회의 두 손이 안규홍의 팔뚝을 단단히 붙잡는 클로즈업으로, 핵심 동작과 흐릿한 짚신전 배경을 충실히 구현했습니다."
       }
      ],
      "all_candidates_fail": false,
      "readings": [
       {
        "label": "B",
        "direction": "왼쪽 안규홍과 오른쪽 이관회가 서로의 얼굴을 바라봅니다. 안규홍의 팔은 오른쪽 위로 굽혀져 있고, 이관회의 한 손은 그 손목 부근을 감싸지만 다른 손은 안규홍의 손과 맞잡혀 있습니다. 두 손 모두 팔뚝을 붙잡는 방향과 접촉은 아닙니다.",
        "built_space": "두 사람 뒤에 목재 짚신 진열대 하나와 낮은 판매 평상 하나가 보입니다. 짚신은 세 높이의 가로 진열부에 걸려 있으며, 평상 위에는 청록색 보따리 하나와 큰 바구니 하나가 놓여 있습니다. 두 사람은 평상 앞에 있지만, 배경과 얼굴을 넓고 비교적 선명하게 보여 주어 요청한 부분적이고 부드러운 배경보다 비중이 큽니다.",
        "entities": "한국인 청년 남성으로 설정된 두 인물만 보입니다. 안규홍의 검은 묶음머리, 밝은색 낡은 옷과 목제 지게는 참조와 대체로 맞지만, 소매를 어깨 가까이까지 올려 이전 장면보다 상완이 크게 노출됩니다. 이관회는 자연스러운 검은 머리와 갈색의 해진 저고리를 입어 인물 참조에 부합합니다. 바구니 내용물은 채소로 확인되지 않으며, 신발과 발가락은 화면 밖입니다. 읽을 수 있는 글자는 없습니다.",
        "hard_violations": [],
        "physics": "지게는 어깨끈으로 안규홍의 몸에 지지되고, 보따리와 바구니는 평상에 놓여 있습니다. 두 사람의 손과 팔은 신체에 자연스럽게 연결되어 있고 접촉도 가능합니다. 다만 팔을 높여 서로 손을 맞잡은 자세는 나서려는 팔뚝을 두 손으로 제지하는 순간보다 힘겨루기에 가깝습니다. 발은 잘려 있어 지면 접촉은 확인할 수 없지만 공중에 떠 있는 신체나 물체는 보이지 않습니다."
       },
       {
        "label": "A",
        "direction": "이관회의 두 팔이 오른쪽에서 왼쪽 안규홍의 팔뚝으로 뻗습니다. 한 손은 팔뚝 중간을 아래에서 감싸고, 다른 손은 손목에 가까운 팔뚝을 붙잡습니다. 두 손의 목표가 동일한 팔뚝으로 명확하며, 안규홍의 팔이 오른쪽으로 나가는 것을 막는 관계입니다. 얼굴과 시선은 화면 밖입니다.",
        "built_space": "잘린 두 몸통 뒤로 목재 짚신 진열대 하나, 낮은 평상 하나와 햇빛을 받은 흙바닥이 부분적으로 보입니다. 짚신이 걸린 가로 진열부는 적어도 두 높이에서 보이며, 아래쪽에는 둥근 짚 직물이 기대어 있습니다. 인물들은 평상 바로 앞의 흙 공간에 있고, 배경은 흐려져 손과 팔을 방해하지 않습니다. 고정 시설의 중복이나 불가능한 공간 배치는 보이지 않습니다.",
        "entities": "안규홍의 밝은색 낡은 저고리, 걷어 올린 소매, 굵은 팔뚝과 지게 일부가 보입니다. 이관회는 참조의 갈색 해진 저고리와 밝은색 바지를 입고 있습니다. 보이는 팔과 손은 청년 남성의 외형에 부합하지만, 얼굴과 머리가 잘려 정확한 얼굴 정체성과 머리 모양은 검증할 수 없습니다. 두 인물 외의 사람이나 읽을 수 있는 글자는 없습니다. 보따리, 채소 바구니와 신발은 이 클로즈업 범위에서 확인되지 않습니다.",
        "hard_violations": [],
        "physics": "이관회의 두 손이 팔뚝의 서로 다른 지점을 실제로 감싸며 손가락과 엄지의 맞물림이 제지력을 전달합니다. 안규홍의 잡힌 팔과 아래로 내려온 반대 손은 각각 몸에 연결되어 있습니다. 지게는 몸을 가로지르는 끈과 등에 의해 지지됩니다. 발은 화면 밖이지만, 보이는 상체와 팔의 자세는 서서 상대의 전진을 막는 동작으로 가능하며 지지 없이 떠 있는 물체는 없습니다."
       }
      ]
     },
     "ok": true
    }
   ],
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "slot_winner_match": true,
   "slot_winner": {
    "gemini-pro": "A",
    "gpt-high": "A"
   },
   "route": "cross_slot_agree"
  },
  "dual": {
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "normalized": {
    "A": 2.0,
    "B": 1.0
   },
   "adjusted": {
    "A": 2.0,
    "B": 1.0
   },
   "violations": {},
   "per_model_winner": {
    "gemini-pro": "A",
    "gpt-high": "A"
   },
   "agreed": true
  },
  "totals": {
   "A": 2000,
   "B": 1000
  },
  "selected": "A",
  "ranking": [
   "A",
   "B"
  ],
  "verdicts": [
   {
    "label": "A",
    "score": 2000,
    "verdict_ko": "지시문이 엄격하게 요구한 클로즈업 프레이밍(얼굴이 배제된 토르소 크롭)과 '두 손으로 팔뚝을 꽉 움켜쥐는' 핵심 동작을 정확하게 구현했습니다."
   },
   {
    "label": "B",
    "score": 1000,
    "verdict_ko": "프레이밍 지시(클로즈업, 크롭된 토르소)를 완전히 무시하고 인물들의 얼굴과 상반신을 모두 노출시켰으며, 동작 또한 두 손으로 팔뚝을 잡는 것이 아닌 손을 맞잡는 형태로 임의 변형되었습니다."
   }
  ],
  "refs": [
   {
    "label": "PREVIOUS SHOT STILL — a visually related earlier shot of this same place: the location's look, materials, fixed features and lighting mood are LOCKED to this photo; never copy its camera framing. The clothing, hair and overall look of 안규홍 — who appear both in that photo and in this shot — are LOCKED to that photo. Anyone else visible in that photo is NOT in this shot: never carry their face, body or clothing onto anyone here. If this photo shows a character who cannot move (dead or unconscious), that character's exact pose, position and orientation are ALSO LOCKED — treat their body as a fixed prop of the set that only the camera moves around.",
    "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/db00517b-165b-4005-a287-f28936b59b95/images/272bd7af-9294-4490-98c6-88a40e9deacc/scene/recipe/S4sh17_sel.png",
    "asset_id": "e97a0186-83a8-4435-bc84-d2f038b63284",
    "role": "prev_still"
   },
   {
    "label": "CHARACTER REFERENCE — 안규홍: the exact person appearing in this shot; match face, hair and build exactly.",
    "path": "<bytes:1028384>",
    "asset_id": "3b5894db-ce55-4f34-b966-5e725c3aa952",
    "role": "character_ref"
   },
   {
    "label": "CHARACTER REFERENCE — 이관회: the exact person appearing in this shot; match face, hair and build exactly.",
    "path": "<bytes:933540>",
    "asset_id": "851b450f-ac84-47fe-ac5a-03a56f63fcae",
    "role": "character_ref"
   }
  ],
  "critique_skipped": true,
  "shot_run_produced": false,
  "shot_run_uid": "06aa1406-f282-7d8c-b9ca-d1ee6819f4c4",
  "ref_mode": "prev+엔티티",
  "share_plan": {
   "ref_plan": "prev",
   "prev_anchor_tag": "S4sh17"
  }
 },
 "S4sh24::cine": {
  "applied": true,
  "attempted_at": "2026-09-09T01:01:58.078278+00:00",
  "fingerprint": "1aa1162c5c6974c9e8de333110bc4d3eed9dd625c0a3b421e0aa26e7aafead30",
  "fingerprint_version": 2,
  "provider": "mai",
  "endpoint": "openrouter/chat-completions",
  "model": "microsoft/mai-image-2.6",
  "pack": "24.202608252115",
  "source_file": "S4sh24_sel.png",
  "source_sha256": "dd91668bbbda4edf16fdd0f8bafb197e7ce1cc30aeea6bf83baf5c9a0cd86c04",
  "file": "S4sh24_cine.png",
  "staged_sha256": "35d483a5d5c510718db5e5c8efb16de7c12a8d4a4237f25b13d1b25a5fe2c729",
  "latency_ms": 24004
 },
 "S4sh30::signage": {
  "fp": "22df9897053e4e42",
  "inscriptions": [],
  "cues": [],
  "dropped": []
 },
 "S4sh30": {
  "input_fingerprint": "10fe62ded6ee1126",
  "prompt": "Create ONE FINAL photorealistic live-action film still of the moment below — early 20th-century Korea, Jeolla region; all people are Korean unless explicitly stated otherwise. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 고개를 든 채 눈빛이 서늘하게 굳은 안규홍의 정면 클로즈업.\n\nLOCATION (lock): The sunlit dirt frontage of the market's straw-sandal stall, where a loose sandal has stopped at a customer's feet. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: close-up\n- KEY BACKGROUND ELEMENTS: 짚신전 (Still behind the confrontation, now largely outside the narrow field of focus) — A partial oblique view remains along the frame edge; used as Provides a subdued continuity cue behind the isolated face.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Maintain the market's natural daylight and muted saturation, preserving readable eyes and sober facial contrast without a new lighting cue.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The clothing, hair and overall look of 안규홍 — who appear both in that photo and in this shot — are LOCKED to that photo. Anyone else visible in that photo is NOT in this shot: never carry their face, body or clothing onto anyone here. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nPREVIOUS STILL USAGE (follow exactly — what to take from the attached still and what to exclude; it governs place and objects only, never who is in this shot or how the camera sees them): Take the stall materials, dirt ground, and harsh daylight from the reference. Exclude the earlier inspection action and any intact display arrangement contradicted by the vendor's fall.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): One new straw shoe lies on the dirt at the stall edge; the delivery bundle remains beside the stall. 안규홍: He still wears his carrying frame and damaged straw shoes, standing with the loose new shoe at his toes.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 안규홍 (한국인 남성, 청년 나이대, 검은 머리, 단단히 묶은 머리) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.",
  "roll_prompts": {
   "A": "Create ONE FINAL photorealistic live-action film still of the moment below — early 20th-century Korea, Jeolla region; all people are Korean unless explicitly stated otherwise. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 고개를 든 채 눈빛이 서늘하게 굳은 안규홍의 정면 클로즈업.\n\nLOCATION (lock): The sunlit dirt frontage of the market's straw-sandal stall, where a loose sandal has stopped at a customer's feet. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: close-up\n- KEY BACKGROUND ELEMENTS: 짚신전 (Still behind the confrontation, now largely outside the narrow field of focus) — A partial oblique view remains along the frame edge; used as Provides a subdued continuity cue behind the isolated face.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Maintain the market's natural daylight and muted saturation, preserving readable eyes and sober facial contrast without a new lighting cue.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The clothing, hair and overall look of 안규홍 — who appear both in that photo and in this shot — are LOCKED to that photo. Anyone else visible in that photo is NOT in this shot: never carry their face, body or clothing onto anyone here. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nPREVIOUS STILL USAGE (follow exactly — what to take from the attached still and what to exclude; it governs place and objects only, never who is in this shot or how the camera sees them): Take the stall materials, dirt ground, and harsh daylight from the reference. Exclude the earlier inspection action and any intact display arrangement contradicted by the vendor's fall.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): One new straw shoe lies on the dirt at the stall edge; the delivery bundle remains beside the stall. 안규홍: He still wears his carrying frame and damaged straw shoes, standing with the loose new shoe at his toes.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 안규홍 (한국인 남성, 청년 나이대, 검은 머리, 단단히 묶은 머리) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.",
   "B": "Create ONE FINAL photorealistic live-action film still of the moment below — early 20th-century Korea, Jeolla region; all people are Korean unless explicitly stated otherwise. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 고개를 든 채 눈빛이 서늘하게 굳은 안규홍의 정면 클로즈업.\n\nLOCATION (lock): The sunlit dirt frontage of the market's straw-sandal stall, where a loose sandal has stopped at a customer's feet. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: close-up\n- KEY BACKGROUND ELEMENTS: 짚신전 (Still behind the confrontation, now largely outside the narrow field of focus) — A partial oblique view remains along the frame edge; used as Provides a subdued continuity cue behind the isolated face.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Maintain the market's natural daylight and muted saturation, preserving readable eyes and sober facial contrast without a new lighting cue.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The clothing, hair and overall look of 안규홍 — who appear both in that photo and in this shot — are LOCKED to that photo. Anyone else visible in that photo is NOT in this shot: never carry their face, body or clothing onto anyone here. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nPREVIOUS STILL USAGE (follow exactly — what to take from the attached still and what to exclude; it governs place and objects only, never who is in this shot or how the camera sees them): Take the stall materials, dirt ground, and harsh daylight from the reference. Exclude the earlier inspection action and any intact display arrangement contradicted by the vendor's fall.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): One new straw shoe lies on the dirt at the stall edge; the delivery bundle remains beside the stall. 안규홍: He still wears his carrying frame and damaged straw shoes, standing with the loose new shoe at his toes.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 안규홍 (한국인 남성, 청년 나이대, 검은 머리, 단단히 묶은 머리) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.\n\nCOMPOSITION VARIATION (this candidate only): keep the place, the people and the moment exactly as specified, and keep from the CAMERA & FRAME contract its subject scale, screen placement and key background elements exactly. For this candidate only, its camera angle and height are deliberately left open: choose a camera position distinctly different from the obvious one, as a film director picking a second setup on the same blocking. Stay true to any explicit framing the SHOT TEXT itself dictates."
  },
  "shot_run_spend_attempt_count": 1,
  "readings": [
   {
    "label": "A",
    "direction": "카메라 렌즈 정면을 향해 서늘하고 단호한 눈빛으로 시선을 고정하고 있음.",
    "built_space": "화면 우측 가장자리에 짚신이 걸린 가판대가 아웃포커싱 처리되어 지시대로 배치되었으며, 인물 주변과 뒤편으로 자연광이 비추는 흙바닥 공간이 형성됨.",
    "entities": "레퍼런스와 일치하는 얼굴, 헤어스타일, 의상, 지게를 착용한 안규홍 단 한 명만 프레임 내에 올바르게 등장함.",
    "hard_violations": [],
    "physics": "어깨에 지게를 멘 채 두 발로 흙바닥에 곧게 서 있는 자세가 중력과 지지면에 맞게 자연스럽게 유지됨."
   },
   {
    "label": "B",
    "direction": "화면 우측 전경에 위치한 타인의 다리와 발 쪽으로 고개를 들어 위를 비스듬히 올려다보고 있음.",
    "built_space": "전경의 흙바닥에 새 짚신이 놓여 있고 배경에 짚신 가판대가 보이나, 카메라 뷰포인트가 지면 수준으로 지나치게 낮게 설정됨.",
    "entities": "안규홍 본인의 외형은 레퍼런스와 일치하나, 샷 텍스트에 명시되지 않은 제3자의 다리와 발이 우측 화면 전경에 크게 침범하여 등장함.",
    "hard_violations": [
     "[gemini-pro] invented people or objects (프롬프트 텍스트에 존재하지 않는 제3자의 다리와 발이 전경에 추가됨)",
     "[gemini-pro] a person placed where the staging does not put them (이전 샷 설정에 따라 '서 있는' 상태를 유지해야 하나 바닥에 엎드린 자세로 잘못 배치됨)",
     "[gpt-high] 등장해서는 안 되는 다른 사람의 다리와 신은 발을 오른쪽 전경에 추가했다.",
     "[gpt-high] 서 있어야 하는 안규홍을 다른 사람의 발과 얼굴이 비슷한 높이에 놓이도록 지면 가까이 낮춘 위치에 배치했다."
    ],
    "physics": "바닥에 상체를 과도하게 숙이고 엎드려 고개만 든 상태이며, 이는 두 발로 서서 지지하고 있어야 할 기존 물리적 설정과 완전히 모순됨."
   }
  ],
  "cross_model_order": {
   "policy": "cross_model_order_v2_slot0_fwd_slot1_rev_parallel2",
   "slots": [
    {
     "model": "gemini-pro",
     "order": "forward",
     "display_to_canonical": {
      "A": "A",
      "B": "B"
     },
     "raw": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 9,
        "verdict_ko": "정면 클로즈업이라는 지정된 프레이밍을 정확히 지켰으며, 텍스트에 없는 인물 추가 없이 안규홍의 서 있는 자세와 배경의 짚신전 요소를 완벽하게 구현했습니다."
       },
       {
        "label": "B",
        "score": 2,
        "verdict_ko": "지정되지 않은 타인의 다리가 전경에 등장하여 인물 제한 규칙을 심각하게 위반했고, 서 있어야 한다는 설정과 달리 바닥에 엎드린 자세를 취해 실격입니다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "카메라 렌즈 정면을 향해 서늘하고 단호한 눈빛으로 시선을 고정하고 있음.",
        "built_space": "화면 우측 가장자리에 짚신이 걸린 가판대가 아웃포커싱 처리되어 지시대로 배치되었으며, 인물 주변과 뒤편으로 자연광이 비추는 흙바닥 공간이 형성됨.",
        "entities": "레퍼런스와 일치하는 얼굴, 헤어스타일, 의상, 지게를 착용한 안규홍 단 한 명만 프레임 내에 올바르게 등장함.",
        "hard_violations": [],
        "physics": "어깨에 지게를 멘 채 두 발로 흙바닥에 곧게 서 있는 자세가 중력과 지지면에 맞게 자연스럽게 유지됨."
       },
       {
        "label": "B",
        "direction": "화면 우측 전경에 위치한 타인의 다리와 발 쪽으로 고개를 들어 위를 비스듬히 올려다보고 있음.",
        "built_space": "전경의 흙바닥에 새 짚신이 놓여 있고 배경에 짚신 가판대가 보이나, 카메라 뷰포인트가 지면 수준으로 지나치게 낮게 설정됨.",
        "entities": "안규홍 본인의 외형은 레퍼런스와 일치하나, 샷 텍스트에 명시되지 않은 제3자의 다리와 발이 우측 화면 전경에 크게 침범하여 등장함.",
        "hard_violations": [
         "invented people or objects (프롬프트 텍스트에 존재하지 않는 제3자의 다리와 발이 전경에 추가됨)",
         "a person placed where the staging does not put them (이전 샷 설정에 따라 '서 있는' 상태를 유지해야 하나 바닥에 엎드린 자세로 잘못 배치됨)"
        ],
        "physics": "바닥에 상체를 과도하게 숙이고 엎드려 고개만 든 상태이며, 이는 두 발로 서서 지지하고 있어야 할 기존 물리적 설정과 완전히 모순됨."
       }
      ],
      "all_candidates_fail": false
     },
     "normalized": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 9,
        "verdict_ko": "정면 클로즈업이라는 지정된 프레이밍을 정확히 지켰으며, 텍스트에 없는 인물 추가 없이 안규홍의 서 있는 자세와 배경의 짚신전 요소를 완벽하게 구현했습니다."
       },
       {
        "label": "B",
        "score": 2,
        "verdict_ko": "지정되지 않은 타인의 다리가 전경에 등장하여 인물 제한 규칙을 심각하게 위반했고, 서 있어야 한다는 설정과 달리 바닥에 엎드린 자세를 취해 실격입니다."
       }
      ],
      "all_candidates_fail": false,
      "readings": [
       {
        "label": "A",
        "direction": "카메라 렌즈 정면을 향해 서늘하고 단호한 눈빛으로 시선을 고정하고 있음.",
        "built_space": "화면 우측 가장자리에 짚신이 걸린 가판대가 아웃포커싱 처리되어 지시대로 배치되었으며, 인물 주변과 뒤편으로 자연광이 비추는 흙바닥 공간이 형성됨.",
        "entities": "레퍼런스와 일치하는 얼굴, 헤어스타일, 의상, 지게를 착용한 안규홍 단 한 명만 프레임 내에 올바르게 등장함.",
        "hard_violations": [],
        "physics": "어깨에 지게를 멘 채 두 발로 흙바닥에 곧게 서 있는 자세가 중력과 지지면에 맞게 자연스럽게 유지됨."
       },
       {
        "label": "B",
        "direction": "화면 우측 전경에 위치한 타인의 다리와 발 쪽으로 고개를 들어 위를 비스듬히 올려다보고 있음.",
        "built_space": "전경의 흙바닥에 새 짚신이 놓여 있고 배경에 짚신 가판대가 보이나, 카메라 뷰포인트가 지면 수준으로 지나치게 낮게 설정됨.",
        "entities": "안규홍 본인의 외형은 레퍼런스와 일치하나, 샷 텍스트에 명시되지 않은 제3자의 다리와 발이 우측 화면 전경에 크게 침범하여 등장함.",
        "hard_violations": [
         "invented people or objects (프롬프트 텍스트에 존재하지 않는 제3자의 다리와 발이 전경에 추가됨)",
         "a person placed where the staging does not put them (이전 샷 설정에 따라 '서 있는' 상태를 유지해야 하나 바닥에 엎드린 자세로 잘못 배치됨)"
        ],
        "physics": "바닥에 상체를 과도하게 숙이고 엎드려 고개만 든 상태이며, 이는 두 발로 서서 지지하고 있어야 할 기존 물리적 설정과 완전히 모순됨."
       }
      ]
     },
     "ok": true
    },
    {
     "model": "gpt-high",
     "order": "reverse",
     "display_to_canonical": {
      "A": "B",
      "B": "A"
     },
     "raw": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 2,
        "verdict_ko": "서늘한 눈빛은 보이지만, 서 있는 정면 클로즈업을 지면 가까이 몸을 낮춘 장면으로 바꾸고 금지된 다른 사람의 다리까지 넣었다."
       },
       {
        "label": "B",
        "score": 8,
        "verdict_ko": "안규홍 혼자 고개를 바로 들고 굳은 눈빛을 보이는 정면 연출과 지게·낮빛의 연속성이 맞지만, 가슴까지 포함해 요구한 클로즈업보다 다소 넓다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "얼굴은 거의 정면이지만 상체와 머리가 지면 가까이 낮아져 있고, 눈은 위쪽의 화면 밖 상대를 올려다본다. 오른쪽에는 그 상대처럼 읽히는 다리가 있다. 서서 고개를 든 정면 응시와는 다른 대치 구도다.",
        "built_space": "전경의 흙바닥 위에 느슨한 짚신 한 짝과 오른쪽 인물의 신은 발이 보인다. 뒤에는 목재 짚신 진열대 하나와 가로 선반들이 있으며, 짚신전이 가장자리의 작은 연속성 단서가 아니라 얼굴 뒤 넓은 영역을 차지한다. 안규홍의 얼굴이 다른 사람의 발목 높이 가까이에 배치되어 지정된 서 있는 위치와 맞지 않는다.",
        "entities": "안규홍으로 보이는 검은 머리를 뒤로 묶은 젊은 동아시아계 남성 한 명은 참조와 대체로 유사한 얼굴, 때 탄 밝은 저고리, 목재 지게와 어깨끈을 갖췄다. 오른쪽에는 별도 인물의 바짓단·맨발 일부·짚신이 추가되어 있다. 새 짚신으로 보이는 한 짝은 보이지만 안규홍 자신의 발끝에 놓인 것으로 읽히지 않는다. 안규홍의 낡은 신과 배달 꾸러미는 프레임에서 확인할 수 없고, 읽을 수 있는 글자는 없다.",
        "hard_violations": [
         "등장해서는 안 되는 다른 사람의 다리와 신은 발을 오른쪽 전경에 추가했다.",
         "서 있어야 하는 안규홍을 다른 사람의 발과 얼굴이 비슷한 높이에 놓이도록 지면 가까이 낮춘 위치에 배치했다."
        ],
        "physics": "전경의 짚신은 흙바닥에 놓여 있고 오른쪽 발도 바닥에 닿아 있다. 지게는 어깨끈과 등에 의해 지지되는 것으로 보인다. 안규홍의 손·무릎·발은 가려져 정확한 하체 지지는 확인할 수 없지만, 몸을 숙리거나 웅크려 만들 수 있는 자세이므로 공중에 떠 있다고 단정할 근거는 없다."
       },
       {
        "label": "B",
        "direction": "안규홍의 얼굴과 상체가 정면을 향하고 눈은 렌즈 가까운 정면의 화면 밖 대치 상대를 응시한다. 고개를 숙이지 않았으며, 좁아진 눈과 다문 입이 서늘하게 굳은 표정을 만든다.",
        "built_space": "오른쪽 가장자리에 목재 짚신 진열대 하나가 비스듬히 보이며, 상단·중단·하단의 세 진열 구역을 구분할 수 있다. 왼쪽에는 햇빛과 기둥 그림자가 드리운 흙길, 뒤에는 목재 기둥들이 있다. 안규홍은 진열대 앞에 곧게 선 상체로 배치되어 공간 관계가 자연스럽다. 다만 가슴 아래까지 담아 배경과 의복의 면적이 좁은 얼굴 클로즈업 요구보다 크다.",
        "entities": "한 명의 젊은 동아시아계 남성만 보이며, 참조의 안규홍과 유사한 얼굴 윤곽과 이목구비, 뒤로 단단히 넘겨 묶은 검은 머리, 때 탄 밝은 저고리를 갖췄다. 등 뒤의 목재 지게 하나와 양어깨의 끈이 확인된다. 발밑의 새 짚신, 본인의 손상된 짚신, 배달 꾸러미는 이 상반신 프레임 밖이어서 평가할 수 없다. 다른 사람이나 읽을 수 있는 글자는 없다.",
        "hard_violations": [],
        "physics": "머리는 곧게 세운 목과 상체에 자연스럽게 연결되고, 지게는 양쪽 어깨끈을 통해 몸에 걸려 있다. 발은 프레임 밖이지만 보이는 상체에 부유나 불가능한 관절 배치는 없다. 배경 짚신들은 선반 위에 놓이거나 진열대에 기대어 있다."
       }
      ],
      "all_candidates_fail": false
     },
     "normalized": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 2,
        "verdict_ko": "서늘한 눈빛은 보이지만, 서 있는 정면 클로즈업을 지면 가까이 몸을 낮춘 장면으로 바꾸고 금지된 다른 사람의 다리까지 넣었다."
       },
       {
        "label": "A",
        "score": 8,
        "verdict_ko": "안규홍 혼자 고개를 바로 들고 굳은 눈빛을 보이는 정면 연출과 지게·낮빛의 연속성이 맞지만, 가슴까지 포함해 요구한 클로즈업보다 다소 넓다."
       }
      ],
      "all_candidates_fail": false,
      "readings": [
       {
        "label": "B",
        "direction": "얼굴은 거의 정면이지만 상체와 머리가 지면 가까이 낮아져 있고, 눈은 위쪽의 화면 밖 상대를 올려다본다. 오른쪽에는 그 상대처럼 읽히는 다리가 있다. 서서 고개를 든 정면 응시와는 다른 대치 구도다.",
        "built_space": "전경의 흙바닥 위에 느슨한 짚신 한 짝과 오른쪽 인물의 신은 발이 보인다. 뒤에는 목재 짚신 진열대 하나와 가로 선반들이 있으며, 짚신전이 가장자리의 작은 연속성 단서가 아니라 얼굴 뒤 넓은 영역을 차지한다. 안규홍의 얼굴이 다른 사람의 발목 높이 가까이에 배치되어 지정된 서 있는 위치와 맞지 않는다.",
        "entities": "안규홍으로 보이는 검은 머리를 뒤로 묶은 젊은 동아시아계 남성 한 명은 참조와 대체로 유사한 얼굴, 때 탄 밝은 저고리, 목재 지게와 어깨끈을 갖췄다. 오른쪽에는 별도 인물의 바짓단·맨발 일부·짚신이 추가되어 있다. 새 짚신으로 보이는 한 짝은 보이지만 안규홍 자신의 발끝에 놓인 것으로 읽히지 않는다. 안규홍의 낡은 신과 배달 꾸러미는 프레임에서 확인할 수 없고, 읽을 수 있는 글자는 없다.",
        "hard_violations": [
         "등장해서는 안 되는 다른 사람의 다리와 신은 발을 오른쪽 전경에 추가했다.",
         "서 있어야 하는 안규홍을 다른 사람의 발과 얼굴이 비슷한 높이에 놓이도록 지면 가까이 낮춘 위치에 배치했다."
        ],
        "physics": "전경의 짚신은 흙바닥에 놓여 있고 오른쪽 발도 바닥에 닿아 있다. 지게는 어깨끈과 등에 의해 지지되는 것으로 보인다. 안규홍의 손·무릎·발은 가려져 정확한 하체 지지는 확인할 수 없지만, 몸을 숙리거나 웅크려 만들 수 있는 자세이므로 공중에 떠 있다고 단정할 근거는 없다."
       },
       {
        "label": "A",
        "direction": "안규홍의 얼굴과 상체가 정면을 향하고 눈은 렌즈 가까운 정면의 화면 밖 대치 상대를 응시한다. 고개를 숙이지 않았으며, 좁아진 눈과 다문 입이 서늘하게 굳은 표정을 만든다.",
        "built_space": "오른쪽 가장자리에 목재 짚신 진열대 하나가 비스듬히 보이며, 상단·중단·하단의 세 진열 구역을 구분할 수 있다. 왼쪽에는 햇빛과 기둥 그림자가 드리운 흙길, 뒤에는 목재 기둥들이 있다. 안규홍은 진열대 앞에 곧게 선 상체로 배치되어 공간 관계가 자연스럽다. 다만 가슴 아래까지 담아 배경과 의복의 면적이 좁은 얼굴 클로즈업 요구보다 크다.",
        "entities": "한 명의 젊은 동아시아계 남성만 보이며, 참조의 안규홍과 유사한 얼굴 윤곽과 이목구비, 뒤로 단단히 넘겨 묶은 검은 머리, 때 탄 밝은 저고리를 갖췄다. 등 뒤의 목재 지게 하나와 양어깨의 끈이 확인된다. 발밑의 새 짚신, 본인의 손상된 짚신, 배달 꾸러미는 이 상반신 프레임 밖이어서 평가할 수 없다. 다른 사람이나 읽을 수 있는 글자는 없다.",
        "hard_violations": [],
        "physics": "머리는 곧게 세운 목과 상체에 자연스럽게 연결되고, 지게는 양쪽 어깨끈을 통해 몸에 걸려 있다. 발은 프레임 밖이지만 보이는 상체에 부유나 불가능한 관절 배치는 없다. 배경 짚신들은 선반 위에 놓이거나 진열대에 기대어 있다."
       }
      ]
     },
     "ok": true
    }
   ],
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "slot_winner_match": true,
   "slot_winner": {
    "gemini-pro": "A",
    "gpt-high": "A"
   },
   "route": "cross_slot_agree"
  },
  "dual": {
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "normalized": {
    "A": 2.0,
    "B": 0.472
   },
   "adjusted": {
    "A": 2.0,
    "B": 0.222
   },
   "violations": {
    "B": [
     "[gemini-pro] invented people or objects (프롬프트 텍스트에 존재하지 않는 제3자의 다리와 발이 전경에 추가됨)",
     "[gemini-pro] a person placed where the staging does not put them (이전 샷 설정에 따라 '서 있는' 상태를 유지해야 하나 바닥에 엎드린 자세로 잘못 배치됨)",
     "[gpt-high] 등장해서는 안 되는 다른 사람의 다리와 신은 발을 오른쪽 전경에 추가했다.",
     "[gpt-high] 서 있어야 하는 안규홍을 다른 사람의 발과 얼굴이 비슷한 높이에 놓이도록 지면 가까이 낮춘 위치에 배치했다."
    ]
   },
   "per_model_winner": {
    "gemini-pro": "A",
    "gpt-high": "A"
   },
   "agreed": true
  },
  "totals": {
   "A": 2000,
   "B": 222
  },
  "selected": "A",
  "ranking": [
   "A",
   "B"
  ],
  "verdicts": [
   {
    "label": "A",
    "score": 2000,
    "verdict_ko": "정면 클로즈업이라는 지정된 프레이밍을 정확히 지켰으며, 텍스트에 없는 인물 추가 없이 안규홍의 서 있는 자세와 배경의 짚신전 요소를 완벽하게 구현했습니다."
   },
   {
    "label": "B",
    "score": 222,
    "verdict_ko": "지정되지 않은 타인의 다리가 전경에 등장하여 인물 제한 규칙을 심각하게 위반했고, 서 있어야 한다는 설정과 달리 바닥에 엎드린 자세를 취해 실격입니다.  ★위반: [gemini-pro] invented people or objects (프롬프트 텍스트에 존재하지 않는 제3자의 다리와 발이 전경에 추가됨) / [gemini-pro] a person placed where the staging does not put them (이전 샷 설정에 따라 '서 있는' 상태를 유지해야 하나 바닥에 엎드린 자세로 잘못 배치됨) / [gpt-high] 등장해서는 안 되는 다른 사람의 다리와 신은 발을 오른쪽 전경에 추가했다. / [gpt-high] 서 있어야 하는 안규홍을 다른 사람의 발과 얼굴이 비슷한 높이에 놓이도록 지면 가까이 낮춘 위치에 배치했다."
   }
  ],
  "refs": [
   {
    "label": "PREVIOUS SHOT STILL — a visually related earlier shot of this same place: the location's look, materials, fixed features and lighting mood are LOCKED to this photo; never copy its camera framing. The clothing, hair and overall look of 안규홍 — who appear both in that photo and in this shot — are LOCKED to that photo. Anyone else visible in that photo is NOT in this shot: never carry their face, body or clothing onto anyone here. If this photo shows a character who cannot move (dead or unconscious), that character's exact pose, position and orientation are ALSO LOCKED — treat their body as a fixed prop of the set that only the camera moves around.",
    "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/db00517b-165b-4005-a287-f28936b59b95/images/272bd7af-9294-4490-98c6-88a40e9deacc/scene/recipe/S4sh24_sel.png",
    "asset_id": "c9cd1553-8d4b-453f-a8d4-5f3848271e57",
    "role": "prev_still"
   },
   {
    "label": "CHARACTER REFERENCE — 안규홍: the exact person appearing in this shot; match face, hair and build exactly.",
    "path": "<bytes:1028384>",
    "asset_id": "3b5894db-ce55-4f34-b966-5e725c3aa952",
    "role": "character_ref"
   }
  ],
  "critique_skipped": true,
  "shot_run_produced": false,
  "shot_run_uid": "06aa1406-f409-7129-b99c-3962074c2b4b",
  "ref_mode": "prev+엔티티",
  "share_plan": {
   "ref_plan": "prev",
   "prev_anchor_tag": "S4sh24"
  }
 },
 "S4sh30::cine": {
  "applied": true,
  "attempted_at": "2026-09-09T01:03:17.909841+00:00",
  "fingerprint": "8bd500cf05f606f16fa54a9cfdb6753cf2bbd7154651e43cbe8df3afd7ec7bdc",
  "fingerprint_version": 2,
  "provider": "mai",
  "endpoint": "openrouter/chat-completions",
  "model": "microsoft/mai-image-2.6",
  "pack": "24.202608252115",
  "source_file": "S4sh30_sel.png",
  "source_sha256": "ae1b4e450f96ecc6e506186abee156e8813914bdabf71df07bd0b5934ed877e5",
  "file": "S4sh30_cine.png",
  "staged_sha256": "45df1d407b99c9e649f50605b965131fe4464c9d41e8ef51823a87f5019579c3",
  "latency_ms": 22650
 },
 "S5sh21::signage": {
  "fp": "59697e790a4756a3",
  "inscriptions": [],
  "cues": [],
  "dropped": []
 },
 "S5sh21": {
  "input_fingerprint": "398e082d0f4453d2",
  "prompt": "Create ONE FINAL photorealistic live-action film still of the moment below — early 20th-century Korea, Jeolla region; all people are Korean unless explicitly stated otherwise. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 치켜든 주먹보다 먼저 안규홍의 거대한 주먹이 일본인 징세원1의 안면에 정통으로 꽂힌 찰나.\n\nLOCATION (lock): The outdoor market lane immediately outside the straw-sandal stall, near a set-down wooden carrying frame. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: medium shot\n- KEY BACKGROUND ELEMENTS: 지게 (Carried on 안규홍's back after he has shouldered it again) — Its side is partly visible behind his shoulder, not between the fist and the collector; used as Connects the counterpunch to his interrupted attempt to leave and supplies bodily scale.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Continuous natural daylight, muted color, and unembellished contrast render the impact abruptly and physically rather than heroically.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The clothing, hair and overall look of 안규홍 — who appear both in that photo and in this shot — are LOCKED to that photo. Anyone else visible in that photo is NOT in this shot: never carry their face, body or clothing onto anyone here. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nPREVIOUS STILL USAGE (follow exactly — what to take from the attached still and what to exclude; it governs place and objects only, never who is in this shot or how the camera sees them): Take the straw-sandal stall's fixed features, woven materials, and daytime light. Exclude the earlier tax-inspection action and the pristine arrangement of loose merchandise.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The recovered straw shoe is back inside the stall's mesh bag; the delivery bundle remains beside the stall. 안규홍: He has shouldered his carrying frame again, still wearing torn straw shoes and carrying the coin bundle at his waist; his mouth is bleeding from a punch. 일본인 징세원1: His face takes the punch at the beginning of a backward fall.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 안규홍 (한국인 남성, 청년 나이대, 검은 머리, 단단히 묶은 머리); 일본인 징세원1 (일본인 남성, 성인 나이대, 자연스러운 검은 머리) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.",
  "roll_prompts": {
   "A": "Create ONE FINAL photorealistic live-action film still of the moment below — early 20th-century Korea, Jeolla region; all people are Korean unless explicitly stated otherwise. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 치켜든 주먹보다 먼저 안규홍의 거대한 주먹이 일본인 징세원1의 안면에 정통으로 꽂힌 찰나.\n\nLOCATION (lock): The outdoor market lane immediately outside the straw-sandal stall, near a set-down wooden carrying frame. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: medium shot\n- KEY BACKGROUND ELEMENTS: 지게 (Carried on 안규홍's back after he has shouldered it again) — Its side is partly visible behind his shoulder, not between the fist and the collector; used as Connects the counterpunch to his interrupted attempt to leave and supplies bodily scale.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Continuous natural daylight, muted color, and unembellished contrast render the impact abruptly and physically rather than heroically.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The clothing, hair and overall look of 안규홍 — who appear both in that photo and in this shot — are LOCKED to that photo. Anyone else visible in that photo is NOT in this shot: never carry their face, body or clothing onto anyone here. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nPREVIOUS STILL USAGE (follow exactly — what to take from the attached still and what to exclude; it governs place and objects only, never who is in this shot or how the camera sees them): Take the straw-sandal stall's fixed features, woven materials, and daytime light. Exclude the earlier tax-inspection action and the pristine arrangement of loose merchandise.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The recovered straw shoe is back inside the stall's mesh bag; the delivery bundle remains beside the stall. 안규홍: He has shouldered his carrying frame again, still wearing torn straw shoes and carrying the coin bundle at his waist; his mouth is bleeding from a punch. 일본인 징세원1: His face takes the punch at the beginning of a backward fall.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 안규홍 (한국인 남성, 청년 나이대, 검은 머리, 단단히 묶은 머리); 일본인 징세원1 (일본인 남성, 성인 나이대, 자연스러운 검은 머리) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.",
   "B": "Create ONE FINAL photorealistic live-action film still of the moment below — early 20th-century Korea, Jeolla region; all people are Korean unless explicitly stated otherwise. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 치켜든 주먹보다 먼저 안규홍의 거대한 주먹이 일본인 징세원1의 안면에 정통으로 꽂힌 찰나.\n\nLOCATION (lock): The outdoor market lane immediately outside the straw-sandal stall, near a set-down wooden carrying frame. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: medium shot\n- KEY BACKGROUND ELEMENTS: 지게 (Carried on 안규홍's back after he has shouldered it again) — Its side is partly visible behind his shoulder, not between the fist and the collector; used as Connects the counterpunch to his interrupted attempt to leave and supplies bodily scale.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Continuous natural daylight, muted color, and unembellished contrast render the impact abruptly and physically rather than heroically.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The clothing, hair and overall look of 안규홍 — who appear both in that photo and in this shot — are LOCKED to that photo. Anyone else visible in that photo is NOT in this shot: never carry their face, body or clothing onto anyone here. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nPREVIOUS STILL USAGE (follow exactly — what to take from the attached still and what to exclude; it governs place and objects only, never who is in this shot or how the camera sees them): Take the straw-sandal stall's fixed features, woven materials, and daytime light. Exclude the earlier tax-inspection action and the pristine arrangement of loose merchandise.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The recovered straw shoe is back inside the stall's mesh bag; the delivery bundle remains beside the stall. 안규홍: He has shouldered his carrying frame again, still wearing torn straw shoes and carrying the coin bundle at his waist; his mouth is bleeding from a punch. 일본인 징세원1: His face takes the punch at the beginning of a backward fall.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 안규홍 (한국인 남성, 청년 나이대, 검은 머리, 단단히 묶은 머리); 일본인 징세원1 (일본인 남성, 성인 나이대, 자연스러운 검은 머리) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.\n\nCOMPOSITION VARIATION (this candidate only): keep the place, the people and the moment exactly as specified, and keep from the CAMERA & FRAME contract its subject scale, screen placement and key background elements exactly. For this candidate only, its camera angle and height are deliberately left open: choose a camera position distinctly different from the obvious one, as a film director picking a second setup on the same blocking. Stay true to any explicit framing the SHOT TEXT itself dictates."
  },
  "shot_run_spend_attempt_count": 1,
  "readings": [
   {
    "label": "A",
    "direction": "안규홍의 뻗은 주먹이 일본인 징세원1의 안면(턱과 볼)을 정확히 타격하고 있으며, 타격의 힘으로 징세원의 고개와 시선이 뒤로 꺾이고 있음.",
    "built_space": "짚신 가판대와 야외 시장 골목이 이전 샷의 구조를 유지하고 있으며, 가판대 평상 위에 초록색 망사 자루와 네모난 배달용 짐 꾸러미가 지시된 대로 배치되어 있음.",
    "entities": "안규홍은 입가에 피를 흘리며 흰색 한복을 입고 등에 지게를 메고 있음. 일본인 징세원1은 레퍼런스에 맞춰 검은 양복과 중절모를 착용하고 있으며 인물의 정체성이 정확히 일치함.",
    "hard_violations": [],
    "physics": "안규홍은 두 발로 안정적으로 땅을 딛고 체중을 실어 타격하고 있으며, 징세원1은 펀치의 충격으로 인해 몸이 뒤로 젖혀지며 넘어지기 시작하는 자연스러운 물리적 반응을 보여줌."
   },
   {
    "label": "B",
    "direction": "안규홍의 주먹이 일본인 징세원1의 안면을 향해 뻗어 명중하고 있음.",
    "built_space": "짚신 가판대의 형태는 유지되고 있으나, 프롬프트에서 요구된 망사 자루와 특정 짐 꾸러미의 명확한 배치가 보이지 않음.",
    "entities": "안규홍과 징세원1의 기본적인 외형은 레퍼런스와 부합하나, 지게가 불필요하게 징세원의 몸에 하나 더 생성되어 있음.",
    "hard_violations": [
     "[gemini-pro] 징세원1의 등과 팔을 물리적으로 불가능하게 관통하며 공중에 떠 있는 형태의 지게 (physically impossible anatomy or staging / duplicated and unsupported object)",
     "[gpt-high] 등에 멘 지게와 내려놓은 지게 외에 징세원 등 뒤에도 별도의 큰 지게를 추가하여 운반 소품을 부당하게 증식시켰다."
    ],
    "physics": "안규홍의 타격 자세는 성립되나, 징세원1의 신체를 관통하여 붙어 있는 지게가 어떠한 물리적 지지대도 없이 허공에 융합되어 있어 심각한 물리 법칙 위반을 보임."
   }
  ],
  "cross_model_order": {
   "policy": "cross_model_order_v2_slot0_fwd_slot1_rev_parallel2",
   "slots": [
    {
     "model": "gemini-pro",
     "order": "forward",
     "display_to_canonical": {
      "A": "A",
      "B": "B"
     },
     "raw": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 10,
        "verdict_ko": "레퍼런스와 일치하는 인물 외형, 안규홍의 주먹이 적중하는 찰나와 뒤로 밀려나는 징세원의 물리적 동선을 완벽하게 구현하였으며, 가판대 주변의 망사 자루와 짐 꾸러미 배치까지 지시사항을 충실히 따름."
       },
       {
        "label": "B",
        "score": 2,
        "verdict_ko": "타격 순간의 연출은 존재하나, 징세원의 등과 팔을 관통하는 복제된 지게가 등장하는 치명적인 물리적 오류와 구조적 위반이 발생함."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "안규홍의 뻗은 주먹이 일본인 징세원1의 안면(턱과 볼)을 정확히 타격하고 있으며, 타격의 힘으로 징세원의 고개와 시선이 뒤로 꺾이고 있음.",
        "built_space": "짚신 가판대와 야외 시장 골목이 이전 샷의 구조를 유지하고 있으며, 가판대 평상 위에 초록색 망사 자루와 네모난 배달용 짐 꾸러미가 지시된 대로 배치되어 있음.",
        "entities": "안규홍은 입가에 피를 흘리며 흰색 한복을 입고 등에 지게를 메고 있음. 일본인 징세원1은 레퍼런스에 맞춰 검은 양복과 중절모를 착용하고 있으며 인물의 정체성이 정확히 일치함.",
        "hard_violations": [],
        "physics": "안규홍은 두 발로 안정적으로 땅을 딛고 체중을 실어 타격하고 있으며, 징세원1은 펀치의 충격으로 인해 몸이 뒤로 젖혀지며 넘어지기 시작하는 자연스러운 물리적 반응을 보여줌."
       },
       {
        "label": "B",
        "direction": "안규홍의 주먹이 일본인 징세원1의 안면을 향해 뻗어 명중하고 있음.",
        "built_space": "짚신 가판대의 형태는 유지되고 있으나, 프롬프트에서 요구된 망사 자루와 특정 짐 꾸러미의 명확한 배치가 보이지 않음.",
        "entities": "안규홍과 징세원1의 기본적인 외형은 레퍼런스와 부합하나, 지게가 불필요하게 징세원의 몸에 하나 더 생성되어 있음.",
        "hard_violations": [
         "징세원1의 등과 팔을 물리적으로 불가능하게 관통하며 공중에 떠 있는 형태의 지게 (physically impossible anatomy or staging / duplicated and unsupported object)"
        ],
        "physics": "안규홍의 타격 자세는 성립되나, 징세원1의 신체를 관통하여 붙어 있는 지게가 어떠한 물리적 지지대도 없이 허공에 융합되어 있어 심각한 물리 법칙 위반을 보임."
       }
      ],
      "all_candidates_fail": false
     },
     "normalized": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 10,
        "verdict_ko": "레퍼런스와 일치하는 인물 외형, 안규홍의 주먹이 적중하는 찰나와 뒤로 밀려나는 징세원의 물리적 동선을 완벽하게 구현하였으며, 가판대 주변의 망사 자루와 짐 꾸러미 배치까지 지시사항을 충실히 따름."
       },
       {
        "label": "B",
        "score": 2,
        "verdict_ko": "타격 순간의 연출은 존재하나, 징세원의 등과 팔을 관통하는 복제된 지게가 등장하는 치명적인 물리적 오류와 구조적 위반이 발생함."
       }
      ],
      "all_candidates_fail": false,
      "readings": [
       {
        "label": "A",
        "direction": "안규홍의 뻗은 주먹이 일본인 징세원1의 안면(턱과 볼)을 정확히 타격하고 있으며, 타격의 힘으로 징세원의 고개와 시선이 뒤로 꺾이고 있음.",
        "built_space": "짚신 가판대와 야외 시장 골목이 이전 샷의 구조를 유지하고 있으며, 가판대 평상 위에 초록색 망사 자루와 네모난 배달용 짐 꾸러미가 지시된 대로 배치되어 있음.",
        "entities": "안규홍은 입가에 피를 흘리며 흰색 한복을 입고 등에 지게를 메고 있음. 일본인 징세원1은 레퍼런스에 맞춰 검은 양복과 중절모를 착용하고 있으며 인물의 정체성이 정확히 일치함.",
        "hard_violations": [],
        "physics": "안규홍은 두 발로 안정적으로 땅을 딛고 체중을 실어 타격하고 있으며, 징세원1은 펀치의 충격으로 인해 몸이 뒤로 젖혀지며 넘어지기 시작하는 자연스러운 물리적 반응을 보여줌."
       },
       {
        "label": "B",
        "direction": "안규홍의 주먹이 일본인 징세원1의 안면을 향해 뻗어 명중하고 있음.",
        "built_space": "짚신 가판대의 형태는 유지되고 있으나, 프롬프트에서 요구된 망사 자루와 특정 짐 꾸러미의 명확한 배치가 보이지 않음.",
        "entities": "안규홍과 징세원1의 기본적인 외형은 레퍼런스와 부합하나, 지게가 불필요하게 징세원의 몸에 하나 더 생성되어 있음.",
        "hard_violations": [
         "징세원1의 등과 팔을 물리적으로 불가능하게 관통하며 공중에 떠 있는 형태의 지게 (physically impossible anatomy or staging / duplicated and unsupported object)"
        ],
        "physics": "안규홍의 타격 자세는 성립되나, 징세원1의 신체를 관통하여 붙어 있는 지게가 어떠한 물리적 지지대도 없이 허공에 융합되어 있어 심각한 물리 법칙 위반을 보임."
       }
      ]
     },
     "ok": true
    },
    {
     "model": "gpt-high",
     "order": "reverse",
     "display_to_canonical": {
      "A": "B",
      "B": "A"
     },
     "raw": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 3,
        "verdict_ko": "주먹은 징세원의 안면에 닿지만 지게를 세 개로 증식시켰고, 징세원의 두 손도 펴져 있어 치켜든 주먹을 앞지르는 반격이라는 순간이 약하다."
       },
       {
        "label": "B",
        "score": 8,
        "verdict_ko": "미디엄 숏 안에서 안면에 꽂힌 주먹과 뒤로 기우는 징세원, 안규홍의 등에 멘 지게를 명확히 연결하지만 징세원의 선행 주먹은 충분히 치켜들려 있지 않다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "오른쪽 안규홍은 왼쪽 징세원의 얼굴을 바라보며 주먹을 뻗고, 주먹은 입과 볼 부위에 실제로 닿는다. 징세원의 머리와 상체는 타격 방향을 따라 왼쪽 뒤로 기운다. 다만 징세원의 두 손은 앞쪽으로 펴져 있어 먼저 치켜든 공격 주먹은 보이지 않는다.",
        "built_space": "목조 짚신 가게에 세 단의 짚신 진열대, 대나무 발, 낮은 판매 마루와 바구니가 보이며 두 사람은 그 바로 앞에 있다. 재료와 낮의 빛은 이전 장소와 대체로 맞는다. 그러나 안규홍 등 뒤에 하나, 징세원 등 뒤에 큰 것 하나, 중앙 아래에 내려놓은 작은 것 하나로 지게가 세 개 보인다.",
        "entities": "등장인물은 두 명뿐이다. 안규홍은 검은 머리를 뒤로 단단히 묶은 젊은 동아시아계 남성으로, 참조의 얼굴 윤곽과 낡은 베이지색 옷, 입가 출혈을 대체로 따른다. 징세원은 성인 동아시아계 남성이며 검은 모자·양복·넥타이는 해당 인물 참조와 맞고 이전 장면의 다른 남성 복장을 가져오지 않았다. 국적 자체는 외모로 확정할 수 없다. 짚신이 든 망태와 배달 꾸러미, 허리의 돈 꾸러미는 확인되지 않으며 발은 화면 밖이다. 읽을 수 있는 글자는 없다.",
        "hard_violations": [
         "등에 멘 지게와 내려놓은 지게 외에 징세원 등 뒤에도 별도의 큰 지게를 추가하여 운반 소품을 부당하게 증식시켰다."
        ],
        "physics": "안규홍의 타격 팔은 어깨에서 주먹까지 연결되고 손목은 징세원의 얼굴에 접촉한다. 징세원은 골반 아래가 화면 밖으로 이어지는 상태에서 상체가 뒤로 기울어, 공중 부양보다는 넘어지기 시작하는 자세로 읽힌다. 발의 접지는 프레임 밖이다. 안규홍의 지게는 어깨끈으로 지지되고 작은 지게는 아래 끝이 지면에 닿는다. 징세원 뒤의 추가 지게는 끈이나 잡은 손이 보이지 않으며 하단이 바구니 부근과 겹쳐 정확한 지지 관계가 불명확하다."
       },
       {
        "label": "B",
        "direction": "왼쪽 안규홍의 시선과 뻗은 주먹은 오른쪽 징세원의 얼굴을 향하고, 주먹이 볼과 턱 옆에 밀착한다. 징세원은 눈을 감으며 머리와 상체가 오른쪽 뒤로 젖혀진다. 한 손은 주먹을 쥐고 있으나 낮게 남아 있고 다른 손은 가슴 높이에 있어, 치켜든 주먹을 앞질렀다는 관계는 불완전하다.",
        "built_space": "두 사람의 허벅지 위를 담는 미디엄 숏이며 짚신 가게의 목조 기둥, 세 단 진열대, 대나무 발과 낮은 마루가 배경을 이룬다. 지게는 안규홍이 멘 한 개와 왼쪽 마루에 내려놓아 벽에 기댄 빈 틀 한 개가 보인다. 멘 지게는 안규홍의 어깨 뒤에 있으며 주먹과 징세원 사이를 가리지 않는다. 청록색 보따리와 짚신 망태는 인물 뒤 마루에, 묶은 배달 꾸러미는 오른쪽 마루에 놓여 있다.",
        "entities": "추가 인물 없이 젊은 안규홍과 성인 징세원 두 명만 보인다. 안규홍의 얼굴, 검은 묶음머리, 체격과 낡은 베이지색 옷은 참조에 가깝고 입가의 피도 명확하다. 징세원의 얼굴은 충격 표정으로 변형되어 있으나 검은 머리, 검은 모자와 양복, 흰 셔츠와 넥타이가 인물 참조에 부합한다. 두 사람 모두 동아시아계 외관이며 국적은 시각적으로 확정할 수 없다. 망태 안에 짚신이 보이고 배달 꾸러미도 존재한다. 돈 꾸러미는 가려져 확인되지 않고 신발은 프레임 밖이다. 읽을 수 있는 글자는 없다.",
        "hard_violations": [],
        "physics": "안규홍은 몸을 앞으로 실어 팔을 뻗고 있으며 주먹과 얼굴의 접촉, 징세원의 뒤로 젖혀진 상체가 같은 충격 동작으로 연결된다. 징세원은 하체가 화면 아래로 이어져 있어 근거 없는 공중 부양은 아니다. 두 사람의 발 접지는 촬영 범위 밖이다. 멘 지게는 어깨끈과 등으로 지지되고, 빈 나무틀은 마루와 벽에 기대며, 망태와 꾸러미는 마루에 놓여 있다. 얼굴 근처의 작은 비말은 타격으로 튄 것으로 읽힌다."
       }
      ],
      "all_candidates_fail": false
     },
     "normalized": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 3,
        "verdict_ko": "주먹은 징세원의 안면에 닿지만 지게를 세 개로 증식시켰고, 징세원의 두 손도 펴져 있어 치켜든 주먹을 앞지르는 반격이라는 순간이 약하다."
       },
       {
        "label": "A",
        "score": 8,
        "verdict_ko": "미디엄 숏 안에서 안면에 꽂힌 주먹과 뒤로 기우는 징세원, 안규홍의 등에 멘 지게를 명확히 연결하지만 징세원의 선행 주먹은 충분히 치켜들려 있지 않다."
       }
      ],
      "all_candidates_fail": false,
      "readings": [
       {
        "label": "B",
        "direction": "오른쪽 안규홍은 왼쪽 징세원의 얼굴을 바라보며 주먹을 뻗고, 주먹은 입과 볼 부위에 실제로 닿는다. 징세원의 머리와 상체는 타격 방향을 따라 왼쪽 뒤로 기운다. 다만 징세원의 두 손은 앞쪽으로 펴져 있어 먼저 치켜든 공격 주먹은 보이지 않는다.",
        "built_space": "목조 짚신 가게에 세 단의 짚신 진열대, 대나무 발, 낮은 판매 마루와 바구니가 보이며 두 사람은 그 바로 앞에 있다. 재료와 낮의 빛은 이전 장소와 대체로 맞는다. 그러나 안규홍 등 뒤에 하나, 징세원 등 뒤에 큰 것 하나, 중앙 아래에 내려놓은 작은 것 하나로 지게가 세 개 보인다.",
        "entities": "등장인물은 두 명뿐이다. 안규홍은 검은 머리를 뒤로 단단히 묶은 젊은 동아시아계 남성으로, 참조의 얼굴 윤곽과 낡은 베이지색 옷, 입가 출혈을 대체로 따른다. 징세원은 성인 동아시아계 남성이며 검은 모자·양복·넥타이는 해당 인물 참조와 맞고 이전 장면의 다른 남성 복장을 가져오지 않았다. 국적 자체는 외모로 확정할 수 없다. 짚신이 든 망태와 배달 꾸러미, 허리의 돈 꾸러미는 확인되지 않으며 발은 화면 밖이다. 읽을 수 있는 글자는 없다.",
        "hard_violations": [
         "등에 멘 지게와 내려놓은 지게 외에 징세원 등 뒤에도 별도의 큰 지게를 추가하여 운반 소품을 부당하게 증식시켰다."
        ],
        "physics": "안규홍의 타격 팔은 어깨에서 주먹까지 연결되고 손목은 징세원의 얼굴에 접촉한다. 징세원은 골반 아래가 화면 밖으로 이어지는 상태에서 상체가 뒤로 기울어, 공중 부양보다는 넘어지기 시작하는 자세로 읽힌다. 발의 접지는 프레임 밖이다. 안규홍의 지게는 어깨끈으로 지지되고 작은 지게는 아래 끝이 지면에 닿는다. 징세원 뒤의 추가 지게는 끈이나 잡은 손이 보이지 않으며 하단이 바구니 부근과 겹쳐 정확한 지지 관계가 불명확하다."
       },
       {
        "label": "A",
        "direction": "왼쪽 안규홍의 시선과 뻗은 주먹은 오른쪽 징세원의 얼굴을 향하고, 주먹이 볼과 턱 옆에 밀착한다. 징세원은 눈을 감으며 머리와 상체가 오른쪽 뒤로 젖혀진다. 한 손은 주먹을 쥐고 있으나 낮게 남아 있고 다른 손은 가슴 높이에 있어, 치켜든 주먹을 앞질렀다는 관계는 불완전하다.",
        "built_space": "두 사람의 허벅지 위를 담는 미디엄 숏이며 짚신 가게의 목조 기둥, 세 단 진열대, 대나무 발과 낮은 마루가 배경을 이룬다. 지게는 안규홍이 멘 한 개와 왼쪽 마루에 내려놓아 벽에 기댄 빈 틀 한 개가 보인다. 멘 지게는 안규홍의 어깨 뒤에 있으며 주먹과 징세원 사이를 가리지 않는다. 청록색 보따리와 짚신 망태는 인물 뒤 마루에, 묶은 배달 꾸러미는 오른쪽 마루에 놓여 있다.",
        "entities": "추가 인물 없이 젊은 안규홍과 성인 징세원 두 명만 보인다. 안규홍의 얼굴, 검은 묶음머리, 체격과 낡은 베이지색 옷은 참조에 가깝고 입가의 피도 명확하다. 징세원의 얼굴은 충격 표정으로 변형되어 있으나 검은 머리, 검은 모자와 양복, 흰 셔츠와 넥타이가 인물 참조에 부합한다. 두 사람 모두 동아시아계 외관이며 국적은 시각적으로 확정할 수 없다. 망태 안에 짚신이 보이고 배달 꾸러미도 존재한다. 돈 꾸러미는 가려져 확인되지 않고 신발은 프레임 밖이다. 읽을 수 있는 글자는 없다.",
        "hard_violations": [],
        "physics": "안규홍은 몸을 앞으로 실어 팔을 뻗고 있으며 주먹과 얼굴의 접촉, 징세원의 뒤로 젖혀진 상체가 같은 충격 동작으로 연결된다. 징세원은 하체가 화면 아래로 이어져 있어 근거 없는 공중 부양은 아니다. 두 사람의 발 접지는 촬영 범위 밖이다. 멘 지게는 어깨끈과 등으로 지지되고, 빈 나무틀은 마루와 벽에 기대며, 망태와 꾸러미는 마루에 놓여 있다. 얼굴 근처의 작은 비말은 타격으로 튄 것으로 읽힌다."
       }
      ]
     },
     "ok": true
    }
   ],
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "slot_winner_match": true,
   "slot_winner": {
    "gemini-pro": "A",
    "gpt-high": "A"
   },
   "route": "cross_slot_agree"
  },
  "dual": {
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "normalized": {
    "A": 2.0,
    "B": 0.575
   },
   "adjusted": {
    "A": 2.0,
    "B": 0.325
   },
   "violations": {
    "B": [
     "[gemini-pro] 징세원1의 등과 팔을 물리적으로 불가능하게 관통하며 공중에 떠 있는 형태의 지게 (physically impossible anatomy or staging / duplicated and unsupported object)",
     "[gpt-high] 등에 멘 지게와 내려놓은 지게 외에 징세원 등 뒤에도 별도의 큰 지게를 추가하여 운반 소품을 부당하게 증식시켰다."
    ]
   },
   "per_model_winner": {
    "gemini-pro": "A",
    "gpt-high": "A"
   },
   "agreed": true
  },
  "totals": {
   "A": 2000,
   "B": 325
  },
  "selected": "A",
  "ranking": [
   "A",
   "B"
  ],
  "verdicts": [
   {
    "label": "A",
    "score": 2000,
    "verdict_ko": "레퍼런스와 일치하는 인물 외형, 안규홍의 주먹이 적중하는 찰나와 뒤로 밀려나는 징세원의 물리적 동선을 완벽하게 구현하였으며, 가판대 주변의 망사 자루와 짐 꾸러미 배치까지 지시사항을 충실히 따름."
   },
   {
    "label": "B",
    "score": 325,
    "verdict_ko": "타격 순간의 연출은 존재하나, 징세원의 등과 팔을 관통하는 복제된 지게가 등장하는 치명적인 물리적 오류와 구조적 위반이 발생함.  ★위반: [gemini-pro] 징세원1의 등과 팔을 물리적으로 불가능하게 관통하며 공중에 떠 있는 형태의 지게 (physically impossible anatomy or staging / duplicated and unsupported object) / [gpt-high] 등에 멘 지게와 내려놓은 지게 외에 징세원 등 뒤에도 별도의 큰 지게를 추가하여 운반 소품을 부당하게 증식시켰다."
   }
  ],
  "refs": [
   {
    "label": "PREVIOUS SHOT STILL — a visually related earlier shot of this same place: the location's look, materials, fixed features and lighting mood are LOCKED to this photo; never copy its camera framing. The clothing, hair and overall look of 안규홍 — who appear both in that photo and in this shot — are LOCKED to that photo. Anyone else visible in that photo is NOT in this shot: never carry their face, body or clothing onto anyone here. If this photo shows a character who cannot move (dead or unconscious), that character's exact pose, position and orientation are ALSO LOCKED — treat their body as a fixed prop of the set that only the camera moves around.",
    "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/db00517b-165b-4005-a287-f28936b59b95/images/272bd7af-9294-4490-98c6-88a40e9deacc/scene/recipe/S4sh17_sel.png",
    "asset_id": "e97a0186-83a8-4435-bc84-d2f038b63284",
    "role": "prev_still"
   },
   {
    "label": "CHARACTER REFERENCE — 안규홍: the exact person appearing in this shot; match face, hair and build exactly.",
    "path": "<bytes:1028384>",
    "asset_id": "3b5894db-ce55-4f34-b966-5e725c3aa952",
    "role": "character_ref"
   },
   {
    "label": "CHARACTER REFERENCE — 일본인 징세원1: the exact person appearing in this shot; match face, hair and build exactly.",
    "path": "<bytes:713416>",
    "asset_id": "47009e21-b41e-49ba-91a0-8e4392776be7",
    "role": "character_ref"
   }
  ],
  "critique_skipped": true,
  "shot_run_produced": false,
  "shot_run_uid": "06aa1406-f589-79f1-951b-e15a9ffaabff",
  "ref_mode": "prev+엔티티",
  "share_plan": {
   "ref_plan": "prev",
   "prev_anchor_tag": "S4sh17"
  }
 },
 "S5sh21::cine": {
  "applied": true,
  "attempted_at": "2026-09-09T01:04:38.981909+00:00",
  "fingerprint": "a32e178e097ce5f1814bd1b5e54a83366c7608f17694b523884982a694430541",
  "fingerprint_version": 2,
  "provider": "mai",
  "endpoint": "openrouter/chat-completions",
  "model": "microsoft/mai-image-2.6",
  "pack": "24.202608252115",
  "source_file": "S5sh21_sel.png",
  "source_sha256": "ae0863e855887b308e11653a41895dda4cb0a03c2b299de7c52716f28bf17e90",
  "file": "S5sh21_cine.png",
  "staged_sha256": "d9765682cf317d9d29e8e7121bdcc56c21b10e7397983ed24d31311c4ba0a584",
  "latency_ms": 26487
 },
 "S5sh28::signage": {
  "fp": "753ce9d1e09c24bf",
  "inscriptions": [],
  "cues": [],
  "dropped": []
 },
 "S5sh28": {
  "input_fingerprint": "c847c94924da7bcd",
  "prompt": "Create ONE FINAL photorealistic live-action film still of the moment below — early 20th-century Korea, Jeolla region; all people are Korean unless explicitly stated otherwise. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 안규홍이 다급히 쪼그려 앉아 바닥에 뻗은 일본인 징세원1의 코밑에 검지손가락을 갖다 댄 손 클로즈업.\n\nLOCATION (lock): Ground level on the market's dirt lane outside the straw-sandal stall, where the first tax collector has fallen. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: close-up\n- KEY BACKGROUND ELEMENTS: 장터 바닥 (Visible beneath the fallen collector's head and beside 안규홍's crouched knee); used as Provides a narrow spatial reference that prevents the hand and face from appearing disconnected.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Keep the same natural daylight and restrained contrast, allowing the hand and breathing reference to read clearly without a diagnostic or stylized glow.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The clothing, hair and overall look of 안규홍, 일본인 징세원1 — who appear both in that photo and in this shot — are LOCKED to that photo. Anyone else visible in that photo is NOT in this shot: never carry their face, body or clothing onto anyone here. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nPREVIOUS STILL USAGE (follow exactly — what to take from the attached still and what to exclude; it governs place and objects only, never who is in this shot or how the camera sees them): Take the dirt surface and straw-sandal stall materials visible in the reference. Exclude the earlier inspection action and do not restore merchandise disturbed during the struggle.\n\nIMMOBILE CHARACTER POSE — CANONICAL (identical wherever this character appears in ANY panel; on any conflict THIS POSE WINS): He is stretched out motionless on the market ground, his body supported by the ground rather than seated or propped against anything. The screenplay does not specify whether he is face-up or face-down, his head angle, or the exact arrangement of his arms and legs.\n\nIMMOBILE BODIES OBEY GRAVITY: a person who is dead or unconscious\nexerts NO muscular effort. Every part of their body — head, torso,\narms, hands, fingers, legs — rests fully on whatever supports it\n(floor, wall, furniture, their own lap) and hangs or slumps with\ngravity. NEVER show any part of an immobile person's body lifted,\nraised, held up in the air, or posed as if presenting something:\nan object in their grip stays clenched in a hand that itself lies\nfallen on a support — the hand does not hold the object up. If the\ncanonical pose leaves a body part unspecified, resolve it as the\nmost gravity-compliant, fully-supported position.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The returned straw shoe remains inside the mesh bag beside the stall. 안규홍: He crouches with his carrying frame still shouldered, blood in his mouth and the coin bundle still at his waist. 일본인 징세원1: He lies motionless on the dirt but is breathing.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 안규홍 (한국인 남성, 청년 나이대, 검은 머리, 단단히 묶은 머리); 일본인 징세원1 (일본인 남성, 성인 나이대, 자연스러운 검은 머리) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.",
  "roll_prompts": {
   "A": "Create ONE FINAL photorealistic live-action film still of the moment below — early 20th-century Korea, Jeolla region; all people are Korean unless explicitly stated otherwise. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 안규홍이 다급히 쪼그려 앉아 바닥에 뻗은 일본인 징세원1의 코밑에 검지손가락을 갖다 댄 손 클로즈업.\n\nLOCATION (lock): Ground level on the market's dirt lane outside the straw-sandal stall, where the first tax collector has fallen. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: close-up\n- KEY BACKGROUND ELEMENTS: 장터 바닥 (Visible beneath the fallen collector's head and beside 안규홍's crouched knee); used as Provides a narrow spatial reference that prevents the hand and face from appearing disconnected.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Keep the same natural daylight and restrained contrast, allowing the hand and breathing reference to read clearly without a diagnostic or stylized glow.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The clothing, hair and overall look of 안규홍, 일본인 징세원1 — who appear both in that photo and in this shot — are LOCKED to that photo. Anyone else visible in that photo is NOT in this shot: never carry their face, body or clothing onto anyone here. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nPREVIOUS STILL USAGE (follow exactly — what to take from the attached still and what to exclude; it governs place and objects only, never who is in this shot or how the camera sees them): Take the dirt surface and straw-sandal stall materials visible in the reference. Exclude the earlier inspection action and do not restore merchandise disturbed during the struggle.\n\nIMMOBILE CHARACTER POSE — CANONICAL (identical wherever this character appears in ANY panel; on any conflict THIS POSE WINS): He is stretched out motionless on the market ground, his body supported by the ground rather than seated or propped against anything. The screenplay does not specify whether he is face-up or face-down, his head angle, or the exact arrangement of his arms and legs.\n\nIMMOBILE BODIES OBEY GRAVITY: a person who is dead or unconscious\nexerts NO muscular effort. Every part of their body — head, torso,\narms, hands, fingers, legs — rests fully on whatever supports it\n(floor, wall, furniture, their own lap) and hangs or slumps with\ngravity. NEVER show any part of an immobile person's body lifted,\nraised, held up in the air, or posed as if presenting something:\nan object in their grip stays clenched in a hand that itself lies\nfallen on a support — the hand does not hold the object up. If the\ncanonical pose leaves a body part unspecified, resolve it as the\nmost gravity-compliant, fully-supported position.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The returned straw shoe remains inside the mesh bag beside the stall. 안규홍: He crouches with his carrying frame still shouldered, blood in his mouth and the coin bundle still at his waist. 일본인 징세원1: He lies motionless on the dirt but is breathing.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 안규홍 (한국인 남성, 청년 나이대, 검은 머리, 단단히 묶은 머리); 일본인 징세원1 (일본인 남성, 성인 나이대, 자연스러운 검은 머리) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.",
   "B": "Create ONE FINAL photorealistic live-action film still of the moment below — early 20th-century Korea, Jeolla region; all people are Korean unless explicitly stated otherwise. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 안규홍이 다급히 쪼그려 앉아 바닥에 뻗은 일본인 징세원1의 코밑에 검지손가락을 갖다 댄 손 클로즈업.\n\nLOCATION (lock): Ground level on the market's dirt lane outside the straw-sandal stall, where the first tax collector has fallen. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: close-up\n- KEY BACKGROUND ELEMENTS: 장터 바닥 (Visible beneath the fallen collector's head and beside 안규홍's crouched knee); used as Provides a narrow spatial reference that prevents the hand and face from appearing disconnected.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Keep the same natural daylight and restrained contrast, allowing the hand and breathing reference to read clearly without a diagnostic or stylized glow.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The clothing, hair and overall look of 안규홍, 일본인 징세원1 — who appear both in that photo and in this shot — are LOCKED to that photo. Anyone else visible in that photo is NOT in this shot: never carry their face, body or clothing onto anyone here. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nPREVIOUS STILL USAGE (follow exactly — what to take from the attached still and what to exclude; it governs place and objects only, never who is in this shot or how the camera sees them): Take the dirt surface and straw-sandal stall materials visible in the reference. Exclude the earlier inspection action and do not restore merchandise disturbed during the struggle.\n\nIMMOBILE CHARACTER POSE — CANONICAL (identical wherever this character appears in ANY panel; on any conflict THIS POSE WINS): He is stretched out motionless on the market ground, his body supported by the ground rather than seated or propped against anything. The screenplay does not specify whether he is face-up or face-down, his head angle, or the exact arrangement of his arms and legs.\n\nIMMOBILE BODIES OBEY GRAVITY: a person who is dead or unconscious\nexerts NO muscular effort. Every part of their body — head, torso,\narms, hands, fingers, legs — rests fully on whatever supports it\n(floor, wall, furniture, their own lap) and hangs or slumps with\ngravity. NEVER show any part of an immobile person's body lifted,\nraised, held up in the air, or posed as if presenting something:\nan object in their grip stays clenched in a hand that itself lies\nfallen on a support — the hand does not hold the object up. If the\ncanonical pose leaves a body part unspecified, resolve it as the\nmost gravity-compliant, fully-supported position.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The returned straw shoe remains inside the mesh bag beside the stall. 안규홍: He crouches with his carrying frame still shouldered, blood in his mouth and the coin bundle still at his waist. 일본인 징세원1: He lies motionless on the dirt but is breathing.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 안규홍 (한국인 남성, 청년 나이대, 검은 머리, 단단히 묶은 머리); 일본인 징세원1 (일본인 남성, 성인 나이대, 자연스러운 검은 머리) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.\n\nCOMPOSITION VARIATION (this candidate only): keep the place, the people and the moment exactly as specified, and keep from the CAMERA & FRAME contract its subject scale, screen placement and key background elements exactly. For this candidate only, its camera angle and height are deliberately left open: choose a camera position distinctly different from the obvious one, as a film director picking a second setup on the same blocking. Stay true to any explicit framing the SHOT TEXT itself dictates."
  },
  "shot_run_spend_attempt_count": 1,
  "readings": [
   {
    "label": "A",
    "direction": "안규홍의 뻗은 오른손 검지손가락이 쓰러진 징세원의 코밑을 정확하게 가리키며 닿아 있음.",
    "built_space": "짚신이 진열된 목조 가판대와 장터의 흙바닥이 지시된 위치에 맞게 배경으로 올바르게 구성되어 있음.",
    "entities": "안규홍(얼굴은 화면 밖에 있으나 낡은 전통 의상, 지게, 허리의 엽전 꾸러미 일치함), 일본인 징세원1(정장과 검은 모자를 착용하고 있으며, 이전 컷에 따라 흰 장갑을 끼지 않은 상태가 유지됨).",
    "hard_violations": [],
    "physics": "징세원은 흙바닥에 평평하게 쓰러져 신체가 지면에 완전히 지지되어 있고, 안규홍은 다리로 체중을 받치며 쪼그려 앉은 자세를 물리적으로 어색함 없이 유지하고 있음."
   },
   {
    "label": "B",
    "direction": "안규홍의 오른손 검지손가락이 쓰러진 징세원의 코밑으로 정확히 향하고 있음.",
    "built_space": "짚신 가판대와 흙바닥이 자연스럽게 배치되어 있으며 인물들의 공간적 위치와 잘 맞음.",
    "entities": "안규홍(입가의 피, 지게, 전통 의상은 일치하나 허리의 엽전 꾸러미가 없음), 일본인 징세원1(정장을 입고 있으나 이전 컷에 없던 흰 장갑을 끼고 있어 의상 고정 조건을 위반함).",
    "hard_violations": [],
    "physics": "징세원은 바닥에 쓰러져 중력에 맞게 몸이 늘어져 있으며, 안규홍은 바닥에 발을 단단히 딛고 안정적으로 쪼그려 앉아 있음."
   }
  ],
  "cross_model_order": {
   "policy": "cross_model_order_v2_slot0_fwd_slot1_rev_parallel2",
   "slots": [
    {
     "model": "gemini-pro",
     "order": "forward",
     "display_to_canonical": {
      "A": "A",
      "B": "B"
     },
     "raw": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 5,
        "verdict_ko": "두 후보 모두 '손 클로즈업'이라는 프레이밍 지시를 어기고 샷을 너무 넓게 잡았으나, A는 인물의 얼굴을 잘라내어 의도에 더 가깝고 이전 샷의 상태(흰 장갑 없음)와 소품(엽전 꾸러미) 조건을 정확하게 지켰습니다."
       },
       {
        "label": "B",
        "score": 3,
        "verdict_ko": "클로즈업 지시를 완전히 무시하고 인물들의 상반신 전체를 보여주었으며, 이전 샷에 없던 징세원의 흰 장갑을 임의로 추가하고 안규홍의 동전 꾸러미를 누락하여 우선순위 조건들을 다수 위반했습니다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "안규홍의 뻗은 오른손 검지손가락이 쓰러진 징세원의 코밑을 정확하게 가리키며 닿아 있음.",
        "built_space": "짚신이 진열된 목조 가판대와 장터의 흙바닥이 지시된 위치에 맞게 배경으로 올바르게 구성되어 있음.",
        "entities": "안규홍(얼굴은 화면 밖에 있으나 낡은 전통 의상, 지게, 허리의 엽전 꾸러미 일치함), 일본인 징세원1(정장과 검은 모자를 착용하고 있으며, 이전 컷에 따라 흰 장갑을 끼지 않은 상태가 유지됨).",
        "hard_violations": [],
        "physics": "징세원은 흙바닥에 평평하게 쓰러져 신체가 지면에 완전히 지지되어 있고, 안규홍은 다리로 체중을 받치며 쪼그려 앉은 자세를 물리적으로 어색함 없이 유지하고 있음."
       },
       {
        "label": "B",
        "direction": "안규홍의 오른손 검지손가락이 쓰러진 징세원의 코밑으로 정확히 향하고 있음.",
        "built_space": "짚신 가판대와 흙바닥이 자연스럽게 배치되어 있으며 인물들의 공간적 위치와 잘 맞음.",
        "entities": "안규홍(입가의 피, 지게, 전통 의상은 일치하나 허리의 엽전 꾸러미가 없음), 일본인 징세원1(정장을 입고 있으나 이전 컷에 없던 흰 장갑을 끼고 있어 의상 고정 조건을 위반함).",
        "hard_violations": [],
        "physics": "징세원은 바닥에 쓰러져 중력에 맞게 몸이 늘어져 있으며, 안규홍은 바닥에 발을 단단히 딛고 안정적으로 쪼그려 앉아 있음."
       }
      ],
      "all_candidates_fail": false
     },
     "normalized": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 5,
        "verdict_ko": "두 후보 모두 '손 클로즈업'이라는 프레이밍 지시를 어기고 샷을 너무 넓게 잡았으나, A는 인물의 얼굴을 잘라내어 의도에 더 가깝고 이전 샷의 상태(흰 장갑 없음)와 소품(엽전 꾸러미) 조건을 정확하게 지켰습니다."
       },
       {
        "label": "B",
        "score": 3,
        "verdict_ko": "클로즈업 지시를 완전히 무시하고 인물들의 상반신 전체를 보여주었으며, 이전 샷에 없던 징세원의 흰 장갑을 임의로 추가하고 안규홍의 동전 꾸러미를 누락하여 우선순위 조건들을 다수 위반했습니다."
       }
      ],
      "all_candidates_fail": false,
      "readings": [
       {
        "label": "A",
        "direction": "안규홍의 뻗은 오른손 검지손가락이 쓰러진 징세원의 코밑을 정확하게 가리키며 닿아 있음.",
        "built_space": "짚신이 진열된 목조 가판대와 장터의 흙바닥이 지시된 위치에 맞게 배경으로 올바르게 구성되어 있음.",
        "entities": "안규홍(얼굴은 화면 밖에 있으나 낡은 전통 의상, 지게, 허리의 엽전 꾸러미 일치함), 일본인 징세원1(정장과 검은 모자를 착용하고 있으며, 이전 컷에 따라 흰 장갑을 끼지 않은 상태가 유지됨).",
        "hard_violations": [],
        "physics": "징세원은 흙바닥에 평평하게 쓰러져 신체가 지면에 완전히 지지되어 있고, 안규홍은 다리로 체중을 받치며 쪼그려 앉은 자세를 물리적으로 어색함 없이 유지하고 있음."
       },
       {
        "label": "B",
        "direction": "안규홍의 오른손 검지손가락이 쓰러진 징세원의 코밑으로 정확히 향하고 있음.",
        "built_space": "짚신 가판대와 흙바닥이 자연스럽게 배치되어 있으며 인물들의 공간적 위치와 잘 맞음.",
        "entities": "안규홍(입가의 피, 지게, 전통 의상은 일치하나 허리의 엽전 꾸러미가 없음), 일본인 징세원1(정장을 입고 있으나 이전 컷에 없던 흰 장갑을 끼고 있어 의상 고정 조건을 위반함).",
        "hard_violations": [],
        "physics": "징세원은 바닥에 쓰러져 중력에 맞게 몸이 늘어져 있으며, 안규홍은 바닥에 발을 단단히 딛고 안정적으로 쪼그려 앉아 있음."
       }
      ]
     },
     "ok": true
    },
    {
     "model": "gpt-high",
     "order": "reverse",
     "display_to_canonical": {
      "A": "B",
      "B": "A"
     },
     "raw": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 4,
        "verdict_ko": "쪼그린 자세와 지면에 누운 징세원은 자연스럽지만, 손 클로즈업 대신 두 사람과 가게를 넓게 담았고 검지가 코밑보다는 코끝에 닿아 있다."
       },
       {
        "label": "B",
        "score": 6,
        "verdict_ko": "검지가 콧구멍 바로 아래에 놓이고 머리와 무릎 사이 흙바닥도 이어져 핵심 행동은 더 정확하지만, 여전히 손 클로즈업보다 구도가 넓다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "안규홍은 징세원의 코와 자신의 손끝을 내려다본다. 뻗은 검지는 코끝 아래쪽과 콧방울 경계에 닿아 있어, 코밑에서 숨을 확인하기보다 코끝을 건드리는 인상이 강하다. 징세원은 눈을 감고 얼굴을 위로 향한다.",
        "built_space": "흙길 뒤로 목조 판매대 한 줄과 수직 기둥들, 왼쪽에 기대 놓은 나무틀 하나, 위아래로 진열된 짚신 약 열 개와 일부 잘린 짚신이 보인다. 판매대에는 바구니 하나와 녹색 자루가 보인다. 목재와 짚 재질은 장소 참조와 유사하지만 바구니는 참조에서 확인되지 않는다. 징세원은 판매대 앞 흙바닥에 누워 있고 안규홍은 그 옆에 쪼그려 있다. 공간 관계는 성립하나 가게와 두 사람을 너무 많이 포함한다.",
        "entities": "등장인물은 두 명뿐이다. 안규홍은 검은 머리를 묶은 젊은 동아시아계 남성으로, 참조와 유사한 얼굴과 해진 베이지색 옷, 짚신, 입 주변의 피, 어깨끈이 보인다. 징세원은 참조와 유사한 검은 머리의 성인 동아시아계 남성이며 검은 양복과 흰 셔츠, 검은 넥타이를 착용했다. 국적 자체는 외형만으로 검증할 수 없다. 검은 모자는 머리 옆에 떨어져 있다. 흰 장갑은 인물 참조에는 있으나 복장이 우선 고정된 이전 장면의 맨손과 다르다. 지게 본체와 허리 동전 꾸러미는 이 구도에서 확인하기 어렵고, 자루 속 짚신도 식별되지 않는다. 읽을 수 있는 글자는 없다.",
        "hard_violations": [],
        "physics": "징세원의 뒤통수와 몸통, 팔은 흙바닥에 놓여 있다. 장갑 낀 손은 손등과 손목 쪽이 바닥에 받쳐져 있고 손가락만 느슨하게 굽어 있어 공중 부양으로 볼 근거는 없다. 안규홍은 접은 다리와 바닥에 닿은 발로 쪼그린 몸을 지탱하며, 검지를 뻗은 손은 팔에 연결되어 있다. 떨어진 모자도 지면에 놓여 있다."
       },
       {
        "label": "B",
        "direction": "안규홍의 검지 끝이 징세원의 콧구멍 바로 아래, 윗입술 위쪽에 놓여 있어 호흡 확인의 목표 지점이 명확하다. 안규홍의 얼굴은 프레임 밖이라 시선은 확인할 수 없다. 징세원은 눈을 감고 얼굴을 위쪽으로 향한다.",
        "built_space": "왼쪽 배경에는 목조 판매대 한 줄과 이를 받치는 기둥, 겹쳐 진열된 짚신 약 열 개, 묶인 꾸러미 하나가 보인다. 참조의 낡은 목재와 짚 재질을 유지한다. 징세원 머리 아래에서 안규홍의 굽힌 무릎 옆까지 흙바닥이 연속해서 보여 두 신체의 위치가 분명하다. 다만 징세원의 허리와 안규홍의 몸통까지 포함해 요구된 손 클로즈업보다 넓다.",
        "entities": "두 사람 외의 인물은 없다. 안규홍은 얼굴과 머리가 잘렸지만 보이는 손과 체격, 해진 베이지색 옷은 참조에 부합한다. 어깨끈과 오른쪽 가장자리의 나무 지게, 허리에 묶인 동전 묶음이 보인다. 입의 피와 머리 모양은 프레임 밖이므로 판단하지 않는다. 징세원은 참조와 유사한 검은 머리의 성인 동아시아계 남성으로 검은 모자와 양복, 흰 셔츠, 검은 넥타이를 착용했다. 외형만으로 일본 국적을 확인할 수는 없다. 반환한 짚신과 망 자루는 프레임 밖이며 읽을 수 있는 글자는 없다.",
        "hard_violations": [],
        "physics": "징세원의 뒤통수는 지면에 닿은 모자와 함께 낮게 놓여 있고 어깨와 몸통도 바닥에 누워 있다. 머리를 스스로 들어 올리는 자세는 보이지 않는다. 안규홍은 양다리를 깊게 굽혀 쪼그려 있으며 하단에 일부 보이는 발이 지면을 지지한다. 검지를 뻗은 손과 반대 손은 각각 팔에 정상적으로 연결되어 있다. 지게는 어깨끈으로, 동전 묶음은 허리의 끈으로 지탱된다."
       }
      ],
      "all_candidates_fail": true
     },
     "normalized": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 4,
        "verdict_ko": "쪼그린 자세와 지면에 누운 징세원은 자연스럽지만, 손 클로즈업 대신 두 사람과 가게를 넓게 담았고 검지가 코밑보다는 코끝에 닿아 있다."
       },
       {
        "label": "A",
        "score": 6,
        "verdict_ko": "검지가 콧구멍 바로 아래에 놓이고 머리와 무릎 사이 흙바닥도 이어져 핵심 행동은 더 정확하지만, 여전히 손 클로즈업보다 구도가 넓다."
       }
      ],
      "all_candidates_fail": true,
      "readings": [
       {
        "label": "B",
        "direction": "안규홍은 징세원의 코와 자신의 손끝을 내려다본다. 뻗은 검지는 코끝 아래쪽과 콧방울 경계에 닿아 있어, 코밑에서 숨을 확인하기보다 코끝을 건드리는 인상이 강하다. 징세원은 눈을 감고 얼굴을 위로 향한다.",
        "built_space": "흙길 뒤로 목조 판매대 한 줄과 수직 기둥들, 왼쪽에 기대 놓은 나무틀 하나, 위아래로 진열된 짚신 약 열 개와 일부 잘린 짚신이 보인다. 판매대에는 바구니 하나와 녹색 자루가 보인다. 목재와 짚 재질은 장소 참조와 유사하지만 바구니는 참조에서 확인되지 않는다. 징세원은 판매대 앞 흙바닥에 누워 있고 안규홍은 그 옆에 쪼그려 있다. 공간 관계는 성립하나 가게와 두 사람을 너무 많이 포함한다.",
        "entities": "등장인물은 두 명뿐이다. 안규홍은 검은 머리를 묶은 젊은 동아시아계 남성으로, 참조와 유사한 얼굴과 해진 베이지색 옷, 짚신, 입 주변의 피, 어깨끈이 보인다. 징세원은 참조와 유사한 검은 머리의 성인 동아시아계 남성이며 검은 양복과 흰 셔츠, 검은 넥타이를 착용했다. 국적 자체는 외형만으로 검증할 수 없다. 검은 모자는 머리 옆에 떨어져 있다. 흰 장갑은 인물 참조에는 있으나 복장이 우선 고정된 이전 장면의 맨손과 다르다. 지게 본체와 허리 동전 꾸러미는 이 구도에서 확인하기 어렵고, 자루 속 짚신도 식별되지 않는다. 읽을 수 있는 글자는 없다.",
        "hard_violations": [],
        "physics": "징세원의 뒤통수와 몸통, 팔은 흙바닥에 놓여 있다. 장갑 낀 손은 손등과 손목 쪽이 바닥에 받쳐져 있고 손가락만 느슨하게 굽어 있어 공중 부양으로 볼 근거는 없다. 안규홍은 접은 다리와 바닥에 닿은 발로 쪼그린 몸을 지탱하며, 검지를 뻗은 손은 팔에 연결되어 있다. 떨어진 모자도 지면에 놓여 있다."
       },
       {
        "label": "A",
        "direction": "안규홍의 검지 끝이 징세원의 콧구멍 바로 아래, 윗입술 위쪽에 놓여 있어 호흡 확인의 목표 지점이 명확하다. 안규홍의 얼굴은 프레임 밖이라 시선은 확인할 수 없다. 징세원은 눈을 감고 얼굴을 위쪽으로 향한다.",
        "built_space": "왼쪽 배경에는 목조 판매대 한 줄과 이를 받치는 기둥, 겹쳐 진열된 짚신 약 열 개, 묶인 꾸러미 하나가 보인다. 참조의 낡은 목재와 짚 재질을 유지한다. 징세원 머리 아래에서 안규홍의 굽힌 무릎 옆까지 흙바닥이 연속해서 보여 두 신체의 위치가 분명하다. 다만 징세원의 허리와 안규홍의 몸통까지 포함해 요구된 손 클로즈업보다 넓다.",
        "entities": "두 사람 외의 인물은 없다. 안규홍은 얼굴과 머리가 잘렸지만 보이는 손과 체격, 해진 베이지색 옷은 참조에 부합한다. 어깨끈과 오른쪽 가장자리의 나무 지게, 허리에 묶인 동전 묶음이 보인다. 입의 피와 머리 모양은 프레임 밖이므로 판단하지 않는다. 징세원은 참조와 유사한 검은 머리의 성인 동아시아계 남성으로 검은 모자와 양복, 흰 셔츠, 검은 넥타이를 착용했다. 외형만으로 일본 국적을 확인할 수는 없다. 반환한 짚신과 망 자루는 프레임 밖이며 읽을 수 있는 글자는 없다.",
        "hard_violations": [],
        "physics": "징세원의 뒤통수는 지면에 닿은 모자와 함께 낮게 놓여 있고 어깨와 몸통도 바닥에 누워 있다. 머리를 스스로 들어 올리는 자세는 보이지 않는다. 안규홍은 양다리를 깊게 굽혀 쪼그려 있으며 하단에 일부 보이는 발이 지면을 지지한다. 검지를 뻗은 손과 반대 손은 각각 팔에 정상적으로 연결되어 있다. 지게는 어깨끈으로, 동전 묶음은 허리의 끈으로 지탱된다."
       }
      ]
     },
     "ok": true
    }
   ],
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "slot_winner_match": true,
   "slot_winner": {
    "gemini-pro": "A",
    "gpt-high": "A"
   },
   "route": "cross_slot_agree"
  },
  "dual": {
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "normalized": {
    "A": 2.0,
    "B": 1.267
   },
   "adjusted": {
    "A": 2.0,
    "B": 1.267
   },
   "violations": {},
   "per_model_winner": {
    "gemini-pro": "A",
    "gpt-high": "A"
   },
   "agreed": true
  },
  "totals": {
   "A": 2000,
   "B": 1267
  },
  "selected": "A",
  "ranking": [
   "A",
   "B"
  ],
  "verdicts": [
   {
    "label": "A",
    "score": 2000,
    "verdict_ko": "두 후보 모두 '손 클로즈업'이라는 프레이밍 지시를 어기고 샷을 너무 넓게 잡았으나, A는 인물의 얼굴을 잘라내어 의도에 더 가깝고 이전 샷의 상태(흰 장갑 없음)와 소품(엽전 꾸러미) 조건을 정확하게 지켰습니다."
   },
   {
    "label": "B",
    "score": 1267,
    "verdict_ko": "클로즈업 지시를 완전히 무시하고 인물들의 상반신 전체를 보여주었으며, 이전 샷에 없던 징세원의 흰 장갑을 임의로 추가하고 안규홍의 동전 꾸러미를 누락하여 우선순위 조건들을 다수 위반했습니다."
   }
  ],
  "refs": [
   {
    "label": "PREVIOUS SHOT STILL — a visually related earlier shot of this same place: the location's look, materials, fixed features and lighting mood are LOCKED to this photo; never copy its camera framing. The clothing, hair and overall look of 안규홍, 일본인 징세원1 — who appear both in that photo and in this shot — are LOCKED to that photo. Anyone else visible in that photo is NOT in this shot: never carry their face, body or clothing onto anyone here. If this photo shows a character who cannot move (dead or unconscious), that character's exact pose, position and orientation are ALSO LOCKED — treat their body as a fixed prop of the set that only the camera moves around.",
    "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/db00517b-165b-4005-a287-f28936b59b95/images/272bd7af-9294-4490-98c6-88a40e9deacc/scene/recipe/S5sh21_sel.png",
    "asset_id": "c9e50fbc-151e-422d-b0f9-0b8022d0bb0c",
    "role": "prev_still"
   },
   {
    "label": "CHARACTER REFERENCE — 안규홍: the exact person appearing in this shot; match face, hair and build exactly.",
    "path": "<bytes:1028384>",
    "asset_id": "3b5894db-ce55-4f34-b966-5e725c3aa952",
    "role": "character_ref"
   },
   {
    "label": "CHARACTER REFERENCE — 일본인 징세원1: the exact person appearing in this shot; match face, hair and build exactly.",
    "path": "<bytes:1055496>",
    "asset_id": "cdacfcb2-5da3-4e19-b2f7-22a2df555c28",
    "role": "character_ref"
   }
  ],
  "critique_skipped": true,
  "shot_run_produced": false,
  "shot_run_uid": "06aa1406-f70f-7462-832d-733d3de79d81",
  "ref_mode": "prev+엔티티",
  "share_plan": {
   "ref_plan": "prev",
   "prev_anchor_tag": "S5sh21"
  }
 },
 "S5sh28::cine": {
  "applied": true,
  "attempted_at": "2026-09-09T01:06:14.786289+00:00",
  "fingerprint": "8069bf96e43457f5bb0ebf05eec15a59fc3e492265c77f35183a6e55670440c1",
  "fingerprint_version": 2,
  "provider": "mai",
  "endpoint": "openrouter/chat-completions",
  "model": "microsoft/mai-image-2.6",
  "pack": "24.202608252115",
  "source_file": "S5sh28_sel.png",
  "source_sha256": "9531329923a4c35a94a3105b98aa5bd92c52dc434587fc31c2b9ba6b19d7111d",
  "file": "S5sh28_cine.png",
  "staged_sha256": "6b8e8e85b663a291729dc1ddb912695a96e8c37acbd039df9b7171cb6c1a4624",
  "latency_ms": 27736
 },
 "S5sh39::signage": {
  "fp": "4aca90d2f35aee8a",
  "inscriptions": [],
  "cues": [],
  "dropped": []
 },
 "S5sh39": {
  "input_fingerprint": "a2d3f910f63d19a2",
  "prompt": "Create ONE FINAL photorealistic live-action film still of the moment below — early 20th-century Korea, Jeolla region; all people are Korean unless explicitly stated otherwise. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 사람들 사이로 벌어진 좁은 틈을 향해 한쪽 발을 강하게 내디디며 쏜살같이 달리는 안규홍의 mid-action 뒷모습.\n\nLOCATION (lock): A narrow opening through the crowd along the outdoor market lane leading away from the fight. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: wide shot\n- FRAME LAYOUT: narrow opening between the market people in the upper-center of the frame, background.\n- KEY BACKGROUND ELEMENTS: 사람들 사이의 좁은 틈 (Opening as people move aside for 안규홍) — Runs away from the camera between unevenly withdrawn bodies toward the upper center; used as Defines the escape vector while cropped edge figures glance down at their footing and preserve natural, unequal spacing.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Maintain the established natural daylight and muted palette, with clear bodily separation and no triumphant lighting change.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The clothing, hair and overall look of 안규홍 — who appear both in that photo and in this shot — are LOCKED to that photo. Anyone else visible in that photo is NOT in this shot: never carry their face, body or clothing onto anyone here. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The recovered shoe remains in the mesh bag; a passage has opened through the market. 안규홍: He runs with his carrying frame and worn straw shoes; his mouth remains injured and his coin bundle has been given away.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 안규홍 (한국인 남성, 청년 나이대, 검은 머리, 단단히 묶은 머리) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.",
  "roll_prompts": {
   "A": "Create ONE FINAL photorealistic live-action film still of the moment below — early 20th-century Korea, Jeolla region; all people are Korean unless explicitly stated otherwise. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 사람들 사이로 벌어진 좁은 틈을 향해 한쪽 발을 강하게 내디디며 쏜살같이 달리는 안규홍의 mid-action 뒷모습.\n\nLOCATION (lock): A narrow opening through the crowd along the outdoor market lane leading away from the fight. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: wide shot\n- FRAME LAYOUT: narrow opening between the market people in the upper-center of the frame, background.\n- KEY BACKGROUND ELEMENTS: 사람들 사이의 좁은 틈 (Opening as people move aside for 안규홍) — Runs away from the camera between unevenly withdrawn bodies toward the upper center; used as Defines the escape vector while cropped edge figures glance down at their footing and preserve natural, unequal spacing.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Maintain the established natural daylight and muted palette, with clear bodily separation and no triumphant lighting change.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The clothing, hair and overall look of 안규홍 — who appear both in that photo and in this shot — are LOCKED to that photo. Anyone else visible in that photo is NOT in this shot: never carry their face, body or clothing onto anyone here. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The recovered shoe remains in the mesh bag; a passage has opened through the market. 안규홍: He runs with his carrying frame and worn straw shoes; his mouth remains injured and his coin bundle has been given away.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 안규홍 (한국인 남성, 청년 나이대, 검은 머리, 단단히 묶은 머리) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.",
   "B": "Create ONE FINAL photorealistic live-action film still of the moment below — early 20th-century Korea, Jeolla region; all people are Korean unless explicitly stated otherwise. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 사람들 사이로 벌어진 좁은 틈을 향해 한쪽 발을 강하게 내디디며 쏜살같이 달리는 안규홍의 mid-action 뒷모습.\n\nLOCATION (lock): A narrow opening through the crowd along the outdoor market lane leading away from the fight. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: wide shot\n- FRAME LAYOUT: narrow opening between the market people in the upper-center of the frame, background.\n- KEY BACKGROUND ELEMENTS: 사람들 사이의 좁은 틈 (Opening as people move aside for 안규홍) — Runs away from the camera between unevenly withdrawn bodies toward the upper center; used as Defines the escape vector while cropped edge figures glance down at their footing and preserve natural, unequal spacing.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Maintain the established natural daylight and muted palette, with clear bodily separation and no triumphant lighting change.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The clothing, hair and overall look of 안규홍 — who appear both in that photo and in this shot — are LOCKED to that photo. Anyone else visible in that photo is NOT in this shot: never carry their face, body or clothing onto anyone here. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The recovered shoe remains in the mesh bag; a passage has opened through the market. 안규홍: He runs with his carrying frame and worn straw shoes; his mouth remains injured and his coin bundle has been given away.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 안규홍 (한국인 남성, 청년 나이대, 검은 머리, 단단히 묶은 머리) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.\n\nCOMPOSITION VARIATION (this candidate only): keep the place, the people and the moment exactly as specified, and keep from the CAMERA & FRAME contract its subject scale, screen placement and key background elements exactly. For this candidate only, its camera angle and height are deliberately left open: choose a camera position distinctly different from the obvious one, as a film director picking a second setup on the same blocking. Stay true to any explicit framing the SHOT TEXT itself dictates."
  },
  "shot_run_spend_attempt_count": 1,
  "readings": [
   {
    "label": "A",
    "direction": "안규홍이 카메라를 등지고 시장 골목의 정중앙을 향해 달려가고 있으며, 주변 군중은 바닥을 보지 않고 대부분 그를 쳐다보고 있습니다.",
    "built_space": "나무로 지어진 시장 건물과 짚신이 걸려 있는 모습이 보이나, 이전 샷에 있던 나무 계단 구조물 등 특정 디테일이 누락되었고 군중의 배치가 인위적으로 대칭을 이룹니다.",
    "entities": "안규홍의 복장과 지게는 레퍼런스와 일치하나 짚신이 든 망사 가방이 완전히 누락되었습니다. 군중은 자연스러운 불균형 배치라는 요구사항을 충족하지 못했습니다.",
    "hard_violations": [],
    "physics": "달리는 포즈에서 왼발이 앞으로 나갔음에도 왼팔이 함께 앞으로 뻗어 있어 인체공학적으로 매우 어색하고 부자연스러운 주행 자세를 보입니다."
   },
   {
    "label": "B",
    "direction": "안규홍이 화면 상단 중앙의 좁은 틈을 향해 달려가며, 프레임 가장자리에 잘린 군중들은 지시문대로 자신의 발밑을 내려다보고 있습니다.",
    "built_space": "이전 샷에서 확립된 좌측 마루 위의 작은 나무 계단 구조물과 네모난 짐 꾸러미 등의 배경 디테일이 훌륭하게 유지되었으며, 좁은 통로가 불규칙하게 형성되어 있습니다.",
    "entities": "안규홍은 레퍼런스에 맞는 외형을 갖추고 있으며, 특히 프롬프트의 지시대로 짚신이 든 망사 가방을 지게에 매달고 있습니다. 주변 군중의 모습과 시선 처리도 요구사항과 일치합니다.",
    "hard_violations": [],
    "physics": "왼발로 땅을 강하게 딛고 오른발을 뒤로 찬 달리기 자세가 역동적이고 물리적으로 자연스러우며, 체중 이동과 팔의 위치도 적절합니다."
   }
  ],
  "cross_model_order": {
   "policy": "cross_model_order_v2_slot0_fwd_slot1_rev_parallel2",
   "slots": [
    {
     "model": "gemini-pro",
     "order": "forward",
     "display_to_canonical": {
      "A": "A",
      "B": "B"
     },
     "raw": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 4,
        "verdict_ko": "군중이 너무 대칭적으로 도열해 있고 프롬프트에서 요구한 망사 가방(mesh bag)이 누락되었으며, 달리는 자세도 부자연스럽습니다."
       },
       {
        "label": "B",
        "score": 9,
        "verdict_ko": "이전 샷의 소품(나무 계단 구조물 등)을 잘 유지했고, 망사 가방을 지게에 매단 디테일과 군중이 바닥을 내려다보며 불규칙하게 길을 터주는 구도를 훌륭하게 구현했습니다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "안규홍이 카메라를 등지고 시장 골목의 정중앙을 향해 달려가고 있으며, 주변 군중은 바닥을 보지 않고 대부분 그를 쳐다보고 있습니다.",
        "built_space": "나무로 지어진 시장 건물과 짚신이 걸려 있는 모습이 보이나, 이전 샷에 있던 나무 계단 구조물 등 특정 디테일이 누락되었고 군중의 배치가 인위적으로 대칭을 이룹니다.",
        "entities": "안규홍의 복장과 지게는 레퍼런스와 일치하나 짚신이 든 망사 가방이 완전히 누락되었습니다. 군중은 자연스러운 불균형 배치라는 요구사항을 충족하지 못했습니다.",
        "hard_violations": [],
        "physics": "달리는 포즈에서 왼발이 앞으로 나갔음에도 왼팔이 함께 앞으로 뻗어 있어 인체공학적으로 매우 어색하고 부자연스러운 주행 자세를 보입니다."
       },
       {
        "label": "B",
        "direction": "안규홍이 화면 상단 중앙의 좁은 틈을 향해 달려가며, 프레임 가장자리에 잘린 군중들은 지시문대로 자신의 발밑을 내려다보고 있습니다.",
        "built_space": "이전 샷에서 확립된 좌측 마루 위의 작은 나무 계단 구조물과 네모난 짐 꾸러미 등의 배경 디테일이 훌륭하게 유지되었으며, 좁은 통로가 불규칙하게 형성되어 있습니다.",
        "entities": "안규홍은 레퍼런스에 맞는 외형을 갖추고 있으며, 특히 프롬프트의 지시대로 짚신이 든 망사 가방을 지게에 매달고 있습니다. 주변 군중의 모습과 시선 처리도 요구사항과 일치합니다.",
        "hard_violations": [],
        "physics": "왼발로 땅을 강하게 딛고 오른발을 뒤로 찬 달리기 자세가 역동적이고 물리적으로 자연스러우며, 체중 이동과 팔의 위치도 적절합니다."
       }
      ],
      "all_candidates_fail": false
     },
     "normalized": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 4,
        "verdict_ko": "군중이 너무 대칭적으로 도열해 있고 프롬프트에서 요구한 망사 가방(mesh bag)이 누락되었으며, 달리는 자세도 부자연스럽습니다."
       },
       {
        "label": "B",
        "score": 9,
        "verdict_ko": "이전 샷의 소품(나무 계단 구조물 등)을 잘 유지했고, 망사 가방을 지게에 매단 디테일과 군중이 바닥을 내려다보며 불규칙하게 길을 터주는 구도를 훌륭하게 구현했습니다."
       }
      ],
      "all_candidates_fail": false,
      "readings": [
       {
        "label": "A",
        "direction": "안규홍이 카메라를 등지고 시장 골목의 정중앙을 향해 달려가고 있으며, 주변 군중은 바닥을 보지 않고 대부분 그를 쳐다보고 있습니다.",
        "built_space": "나무로 지어진 시장 건물과 짚신이 걸려 있는 모습이 보이나, 이전 샷에 있던 나무 계단 구조물 등 특정 디테일이 누락되었고 군중의 배치가 인위적으로 대칭을 이룹니다.",
        "entities": "안규홍의 복장과 지게는 레퍼런스와 일치하나 짚신이 든 망사 가방이 완전히 누락되었습니다. 군중은 자연스러운 불균형 배치라는 요구사항을 충족하지 못했습니다.",
        "hard_violations": [],
        "physics": "달리는 포즈에서 왼발이 앞으로 나갔음에도 왼팔이 함께 앞으로 뻗어 있어 인체공학적으로 매우 어색하고 부자연스러운 주행 자세를 보입니다."
       },
       {
        "label": "B",
        "direction": "안규홍이 화면 상단 중앙의 좁은 틈을 향해 달려가며, 프레임 가장자리에 잘린 군중들은 지시문대로 자신의 발밑을 내려다보고 있습니다.",
        "built_space": "이전 샷에서 확립된 좌측 마루 위의 작은 나무 계단 구조물과 네모난 짐 꾸러미 등의 배경 디테일이 훌륭하게 유지되었으며, 좁은 통로가 불규칙하게 형성되어 있습니다.",
        "entities": "안규홍은 레퍼런스에 맞는 외형을 갖추고 있으며, 특히 프롬프트의 지시대로 짚신이 든 망사 가방을 지게에 매달고 있습니다. 주변 군중의 모습과 시선 처리도 요구사항과 일치합니다.",
        "hard_violations": [],
        "physics": "왼발로 땅을 강하게 딛고 오른발을 뒤로 찬 달리기 자세가 역동적이고 물리적으로 자연스러우며, 체중 이동과 팔의 위치도 적절합니다."
       }
      ]
     },
     "ok": true
    },
    {
     "model": "gpt-high",
     "order": "reverse",
     "display_to_canonical": {
      "A": "B",
      "B": "A"
     },
     "raw": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 8,
        "verdict_ko": "비대칭으로 비켜선 군중 사이를 향하는 와이드 뒷모습, 기존 짚신 가게와 망주머니를 잘 유지하지만 전력 질주의 추진력은 약하다."
       },
       {
        "label": "B",
        "score": 6,
        "verdict_ko": "달리는 뒷모습과 중앙 통로는 맞지만 군중이 정렬된 관람객처럼 보이고, 망주머니가 보이지 않으며 초가 중심 배경의 장소 연속성도 떨어진다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "안규홍의 머리와 몸통, 진행 방향이 모두 카메라 반대편 상단 중앙의 시장 통로를 향한다. 양쪽 가장자리 남성들은 고개를 숙여 발밑을 보고, 왼쪽 중경 남성은 안규홍 쪽으로 돌아본다. 오른쪽 여성은 통로 쪽을 바라본다. 군중이 서로 다른 거리만큼 물러나 있어 도주 방향이 자연스럽게 이어진다. 겨누는 무기나 지시 물체는 없다.",
        "built_space": "왼쪽에 목재 기둥과 대나무 발을 갖춘 짚신 가게가 있고, 여러 줄의 걸린 짚신, 낮은 목재 판매대 하나, 기대 세운 사각 목틀 하나, 묶음 꾸러미 하나가 보인다. 이는 이전 장면의 주요 재료와 집기를 상당 부분 이어간다. 오른쪽에는 멍석 차양과 작은 판매대가 있으며, 뒤쪽 기와지붕 사이로 흙길이 이어진다. 인물들은 판매대 위가 아니라 길 양쪽에 서 있다. 상단 중앙의 통로와 인물 전신을 담은 와이드 구도이며, 반사는 없다.",
        "entities": "중앙 인물은 검은 머리를 단단한 쪽머리로 묶은 청년 남성의 뒷모습이다. 얼굴과 입의 상처는 이 시점에서 확인할 수 없으며, 보이는 체격·때 탄 미색 옷·접어 올린 소매·짚신은 이전 장면과 대체로 맞는다. 목재 지게와 그 위에 매단 망주머니가 보이고, 주머니 속에는 짚신 밑창 두 개처럼 보이는 내용물이 있다. 회수한 신발의 수량까지 정확히 일치하는지는 불분명하다. 동전 꾸러미는 보이지 않는다. 주변에는 시대복을 입은 성인 남녀 군중이 있으며, 이전 장면의 검은 양복 남성은 없다. 읽을 수 있는 글자나 그래픽 표식은 없다.",
        "hard_violations": [],
        "physics": "안규홍은 왼발 앞부분으로 흙바닥을 디디고 오른쪽 무릎을 굽혀 발을 뒤로 들어 올렸다. 지면 지지와 다음 보폭이 가능한 달리기 자세이며, 공중에 정지한 몸이 아니다. 다만 상체가 거의 수직이고 보폭이 짧아 강하게 내달리는 순간보다는 가벼운 달리기에 가깝다. 지게는 어깨끈으로 몸에 고정되고 망주머니는 지게 상단에 묶여 있다. 오른쪽 군중의 바구니도 등에 멘 끈으로 지지된다. 주변 인물과 판매 물품은 각각 지면·판매대·벽걸이에 지지되어 있다."
       },
       {
        "label": "B",
        "direction": "안규홍은 카메라에서 멀어지며 상단 중앙의 군중 사이 통로를 향한다. 양쪽 군중은 대체로 안규홍이나 통로 안쪽을 바라보고 일부는 아래를 본다. 다만 가장자리 인물들의 발밑 확인보다는 주인공을 지켜보는 시선이 두드러진다. 진행 목표는 맞지만 양쪽 줄이 비교적 가지런해 자연스럽게 급히 비켜선 인상은 약하다. 겨누는 물체는 없다.",
        "built_space": "왼쪽에는 여러 줄의 짚신이 걸린 목재 상점과 대나무 발, 낮은 판매대 일부가 보인다. 중앙과 오른쪽은 큰 초가지붕 상점들이 차지하며, 오른쪽 처마에는 매단 바구니 하나가 보인다. 이전 장면의 묶음 꾸러미와 기대 세운 목틀은 확인되지 않으며, 목재 시장이라는 재료감은 이어져도 동일 장소의 구체적인 연결은 A보다 약하다. 양쪽 군중은 흙길 위에 길게 늘어서 중앙 통로를 만든다. 전신이 보이는 와이드 뒷모습 구도이며, 불가능한 반사는 없다.",
        "entities": "주인공은 검은 쪽머리, 미색의 낡은 옷, 목재 지게, 짚신을 갖춘 청년 남성으로 보인다. 등만 보여 얼굴의 동일성과 입의 상처는 판정할 수 없다. 지게 뒤가 넓게 드러나지만 망주머니나 그 안의 회수한 신발은 보이지 않아 휴대 상태를 확인할 수 없다. 동전 꾸러미는 없다. 양쪽에는 시대복을 입은 성인 남녀 군중이 있고 이전 장면의 양복 남성은 보이지 않는다. 읽을 수 있는 글자나 표식은 없다.",
        "hard_violations": [],
        "physics": "왼발 앞부분이 지면에 닿고 오른발은 무릎을 굽혀 뒤로 들려 있어 실제 달리기에서 가능한 지지 자세다. 몸이 근거 없이 떠 있지는 않다. 다만 몸통이 곧고 팔과 보폭의 추진 동작이 작아 쏜살같이 달리는 속도감은 부족하다. 지게는 어깨 쪽 끈으로 지지되며, 군중의 보이는 발은 지면에 놓여 있다. 처마의 바구니는 줄에 매달려 있고 판매대의 물건도 받침 위에 있다."
       }
      ],
      "all_candidates_fail": false
     },
     "normalized": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 8,
        "verdict_ko": "비대칭으로 비켜선 군중 사이를 향하는 와이드 뒷모습, 기존 짚신 가게와 망주머니를 잘 유지하지만 전력 질주의 추진력은 약하다."
       },
       {
        "label": "A",
        "score": 6,
        "verdict_ko": "달리는 뒷모습과 중앙 통로는 맞지만 군중이 정렬된 관람객처럼 보이고, 망주머니가 보이지 않으며 초가 중심 배경의 장소 연속성도 떨어진다."
       }
      ],
      "all_candidates_fail": false,
      "readings": [
       {
        "label": "B",
        "direction": "안규홍의 머리와 몸통, 진행 방향이 모두 카메라 반대편 상단 중앙의 시장 통로를 향한다. 양쪽 가장자리 남성들은 고개를 숙여 발밑을 보고, 왼쪽 중경 남성은 안규홍 쪽으로 돌아본다. 오른쪽 여성은 통로 쪽을 바라본다. 군중이 서로 다른 거리만큼 물러나 있어 도주 방향이 자연스럽게 이어진다. 겨누는 무기나 지시 물체는 없다.",
        "built_space": "왼쪽에 목재 기둥과 대나무 발을 갖춘 짚신 가게가 있고, 여러 줄의 걸린 짚신, 낮은 목재 판매대 하나, 기대 세운 사각 목틀 하나, 묶음 꾸러미 하나가 보인다. 이는 이전 장면의 주요 재료와 집기를 상당 부분 이어간다. 오른쪽에는 멍석 차양과 작은 판매대가 있으며, 뒤쪽 기와지붕 사이로 흙길이 이어진다. 인물들은 판매대 위가 아니라 길 양쪽에 서 있다. 상단 중앙의 통로와 인물 전신을 담은 와이드 구도이며, 반사는 없다.",
        "entities": "중앙 인물은 검은 머리를 단단한 쪽머리로 묶은 청년 남성의 뒷모습이다. 얼굴과 입의 상처는 이 시점에서 확인할 수 없으며, 보이는 체격·때 탄 미색 옷·접어 올린 소매·짚신은 이전 장면과 대체로 맞는다. 목재 지게와 그 위에 매단 망주머니가 보이고, 주머니 속에는 짚신 밑창 두 개처럼 보이는 내용물이 있다. 회수한 신발의 수량까지 정확히 일치하는지는 불분명하다. 동전 꾸러미는 보이지 않는다. 주변에는 시대복을 입은 성인 남녀 군중이 있으며, 이전 장면의 검은 양복 남성은 없다. 읽을 수 있는 글자나 그래픽 표식은 없다.",
        "hard_violations": [],
        "physics": "안규홍은 왼발 앞부분으로 흙바닥을 디디고 오른쪽 무릎을 굽혀 발을 뒤로 들어 올렸다. 지면 지지와 다음 보폭이 가능한 달리기 자세이며, 공중에 정지한 몸이 아니다. 다만 상체가 거의 수직이고 보폭이 짧아 강하게 내달리는 순간보다는 가벼운 달리기에 가깝다. 지게는 어깨끈으로 몸에 고정되고 망주머니는 지게 상단에 묶여 있다. 오른쪽 군중의 바구니도 등에 멘 끈으로 지지된다. 주변 인물과 판매 물품은 각각 지면·판매대·벽걸이에 지지되어 있다."
       },
       {
        "label": "A",
        "direction": "안규홍은 카메라에서 멀어지며 상단 중앙의 군중 사이 통로를 향한다. 양쪽 군중은 대체로 안규홍이나 통로 안쪽을 바라보고 일부는 아래를 본다. 다만 가장자리 인물들의 발밑 확인보다는 주인공을 지켜보는 시선이 두드러진다. 진행 목표는 맞지만 양쪽 줄이 비교적 가지런해 자연스럽게 급히 비켜선 인상은 약하다. 겨누는 물체는 없다.",
        "built_space": "왼쪽에는 여러 줄의 짚신이 걸린 목재 상점과 대나무 발, 낮은 판매대 일부가 보인다. 중앙과 오른쪽은 큰 초가지붕 상점들이 차지하며, 오른쪽 처마에는 매단 바구니 하나가 보인다. 이전 장면의 묶음 꾸러미와 기대 세운 목틀은 확인되지 않으며, 목재 시장이라는 재료감은 이어져도 동일 장소의 구체적인 연결은 A보다 약하다. 양쪽 군중은 흙길 위에 길게 늘어서 중앙 통로를 만든다. 전신이 보이는 와이드 뒷모습 구도이며, 불가능한 반사는 없다.",
        "entities": "주인공은 검은 쪽머리, 미색의 낡은 옷, 목재 지게, 짚신을 갖춘 청년 남성으로 보인다. 등만 보여 얼굴의 동일성과 입의 상처는 판정할 수 없다. 지게 뒤가 넓게 드러나지만 망주머니나 그 안의 회수한 신발은 보이지 않아 휴대 상태를 확인할 수 없다. 동전 꾸러미는 없다. 양쪽에는 시대복을 입은 성인 남녀 군중이 있고 이전 장면의 양복 남성은 보이지 않는다. 읽을 수 있는 글자나 표식은 없다.",
        "hard_violations": [],
        "physics": "왼발 앞부분이 지면에 닿고 오른발은 무릎을 굽혀 뒤로 들려 있어 실제 달리기에서 가능한 지지 자세다. 몸이 근거 없이 떠 있지는 않다. 다만 몸통이 곧고 팔과 보폭의 추진 동작이 작아 쏜살같이 달리는 속도감은 부족하다. 지게는 어깨 쪽 끈으로 지지되며, 군중의 보이는 발은 지면에 놓여 있다. 처마의 바구니는 줄에 매달려 있고 판매대의 물건도 받침 위에 있다."
       }
      ]
     },
     "ok": true
    }
   ],
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "slot_winner_match": true,
   "slot_winner": {
    "gemini-pro": "B",
    "gpt-high": "B"
   },
   "route": "cross_slot_agree"
  },
  "dual": {
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "normalized": {
    "A": 1.194,
    "B": 2.0
   },
   "adjusted": {
    "A": 1.194,
    "B": 2.0
   },
   "violations": {},
   "per_model_winner": {
    "gemini-pro": "B",
    "gpt-high": "B"
   },
   "agreed": true
  },
  "totals": {
   "A": 1194,
   "B": 2000
  },
  "selected": "B",
  "ranking": [
   "B",
   "A"
  ],
  "verdicts": [
   {
    "label": "A",
    "score": 1194,
    "verdict_ko": "군중이 너무 대칭적으로 도열해 있고 프롬프트에서 요구한 망사 가방(mesh bag)이 누락되었으며, 달리는 자세도 부자연스럽습니다."
   },
   {
    "label": "B",
    "score": 2000,
    "verdict_ko": "이전 샷의 소품(나무 계단 구조물 등)을 잘 유지했고, 망사 가방을 지게에 매단 디테일과 군중이 바닥을 내려다보며 불규칙하게 길을 터주는 구도를 훌륭하게 구현했습니다."
   }
  ],
  "refs": [
   {
    "label": "PREVIOUS SHOT STILL — a visually related earlier shot of this same place: the location's look, materials, fixed features and lighting mood are LOCKED to this photo; never copy its camera framing. The clothing, hair and overall look of 안규홍 — who appear both in that photo and in this shot — are LOCKED to that photo. Anyone else visible in that photo is NOT in this shot: never carry their face, body or clothing onto anyone here. If this photo shows a character who cannot move (dead or unconscious), that character's exact pose, position and orientation are ALSO LOCKED — treat their body as a fixed prop of the set that only the camera moves around.",
    "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/db00517b-165b-4005-a287-f28936b59b95/images/272bd7af-9294-4490-98c6-88a40e9deacc/scene/recipe/S5sh21_sel.png",
    "asset_id": "c9e50fbc-151e-422d-b0f9-0b8022d0bb0c",
    "role": "prev_still"
   },
   {
    "label": "CHARACTER REFERENCE — 안규홍: the exact person appearing in this shot; match face, hair and build exactly.",
    "path": "<bytes:1028384>",
    "asset_id": "3b5894db-ce55-4f34-b966-5e725c3aa952",
    "role": "character_ref"
   }
  ],
  "critique_skipped": true,
  "shot_run_produced": false,
  "shot_run_uid": "06aa1406-f891-7926-bc14-fd08117b222d",
  "ref_mode": "prev+엔티티",
  "share_plan": {
   "ref_plan": "prev",
   "prev_anchor_tag": "S5sh21"
  }
 },
 "S5sh39::cine": {
  "applied": true,
  "attempted_at": "2026-09-09T01:07:56.082929+00:00",
  "fingerprint": "0456d85c57dfe190692198314fa45becf33a529dbefe76856b8a278b6c35683a",
  "fingerprint_version": 2,
  "provider": "mai",
  "endpoint": "openrouter/chat-completions",
  "model": "microsoft/mai-image-2.6",
  "pack": "24.202608252115",
  "source_file": "S5sh39_sel.png",
  "source_sha256": "abe61eb89b9b8017b8945d6e4ebf09f7d5bcce721079a964a4085f7f91b62508",
  "file": "S5sh39_cine.png",
  "staged_sha256": "7312668ffaf93dd6fd76d1f082ebc4d41a06678441bbdc1d6c3aebc2e071b301",
  "latency_ms": 26468
 },
 "S6sh1::signage": {
  "fp": "f44f1d91d8ee9a95",
  "inscriptions": [],
  "cues": [],
  "dropped": []
 },
 "S6sh1": {
  "input_fingerprint": "ef7248f03e8c8423",
  "prompt": "Create ONE FINAL photorealistic live-action film still of the moment below — early 20th-century Korea, Jeolla region; all people are Korean unless explicitly stated otherwise. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 흙바닥에 대자로 뻗어 기절한 일본인 징세원1과 일본인 징세원2를 멀찍이 둥글게 둘러싸고 내려다보는 낡은 옷차림의 조선인 상인들.\n\nLOCATION (lock): The dirt fighting area outside the market's straw-sandal stall, now encircled at a distance by onlookers. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: wide shot\n- KEY BACKGROUND ELEMENTS: 흙바닥 (Exposed around the two fallen collectors); used as The visible clearance separates the bodies from the cautious onlookers and makes the encircling distance legible; 상인들의 낡은 옷차림 (Worn clothing visible around the perimeter); used as Provides an uneven human border through varied shoulder angles, bent elbows, and weight distribution rather than a uniform formation.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Natural daytime illumination and restrained saturation keep the aftermath plainly visible, with sober contrast rather than comic or sensational emphasis.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The people visible in that photo are NOT in this shot: never carry their faces, bodies or clothing onto anyone here. Each person in THIS shot is defined solely by the PEOPLE list, their own CHARACTER REFERENCE images and their wardrobe notes. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nPREVIOUS STILL USAGE (follow exactly — what to take from the attached still and what to exclude; it governs place and objects only, never who is in this shot or how the camera sees them): Take the straw-sandal stall, woven display materials, and sunlit dirt ground. Exclude the fleeing laborer and the earlier inspection action; do not restore the stall's undisturbed merchandise arrangement.\n\nIMMOBILE CHARACTER POSE — CANONICAL (identical wherever this character appears in ANY panel; on any conflict THIS POSE WINS): He is stretched out motionless on the market ground, his body supported by the ground rather than seated or propped against anything. The screenplay does not specify whether he is face-up or face-down, his head angle, or the exact arrangement of his arms and legs.\nIMMOBILE CHARACTER POSE — CANONICAL (identical wherever this character appears in ANY panel; on any conflict THIS POSE WINS): He is stretched out motionless on the market ground, his body supported by the ground rather than seated or propped against anything. The screenplay does not specify whether he is face-up or face-down, his head angle, or the exact arrangement of his arms and legs.\n\nIMMOBILE BODIES OBEY GRAVITY: a person who is dead or unconscious\nexerts NO muscular effort. Every part of their body — head, torso,\narms, hands, fingers, legs — rests fully on whatever supports it\n(floor, wall, furniture, their own lap) and hangs or slumps with\ngravity. NEVER show any part of an immobile person's body lifted,\nraised, held up in the air, or posed as if presenting something:\nan object in their grip stays clenched in a hand that itself lies\nfallen on a support — the hand does not hold the object up. If the\ncanonical pose leaves a body part unspecified, resolve it as the\nmost gravity-compliant, fully-supported position.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The same sunlit market remains immediately after the fight; the returned straw shoe is still in the stall's mesh bag. 일본인 징세원1: He lies unconscious but breathing on the dirt. 일본인 징세원2: He lies unconscious but breathing on the dirt.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 일본인 징세원1 (일본인 남성, 성인 나이대, 자연스러운 검은 머리); 일본인 징세원2 (일본인 남성, 성인 나이대, 자연스러운 검은 머리) — wearing: 서양식 검은색 양복 정장에 흰 와이셔츠. 팔에는 징세원을 나타내는 완장을 찼고, 가죽 구두를 신었다. — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.",
  "roll_prompts": {
   "A": "Create ONE FINAL photorealistic live-action film still of the moment below — early 20th-century Korea, Jeolla region; all people are Korean unless explicitly stated otherwise. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 흙바닥에 대자로 뻗어 기절한 일본인 징세원1과 일본인 징세원2를 멀찍이 둥글게 둘러싸고 내려다보는 낡은 옷차림의 조선인 상인들.\n\nLOCATION (lock): The dirt fighting area outside the market's straw-sandal stall, now encircled at a distance by onlookers. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: wide shot\n- KEY BACKGROUND ELEMENTS: 흙바닥 (Exposed around the two fallen collectors); used as The visible clearance separates the bodies from the cautious onlookers and makes the encircling distance legible; 상인들의 낡은 옷차림 (Worn clothing visible around the perimeter); used as Provides an uneven human border through varied shoulder angles, bent elbows, and weight distribution rather than a uniform formation.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Natural daytime illumination and restrained saturation keep the aftermath plainly visible, with sober contrast rather than comic or sensational emphasis.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The people visible in that photo are NOT in this shot: never carry their faces, bodies or clothing onto anyone here. Each person in THIS shot is defined solely by the PEOPLE list, their own CHARACTER REFERENCE images and their wardrobe notes. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nPREVIOUS STILL USAGE (follow exactly — what to take from the attached still and what to exclude; it governs place and objects only, never who is in this shot or how the camera sees them): Take the straw-sandal stall, woven display materials, and sunlit dirt ground. Exclude the fleeing laborer and the earlier inspection action; do not restore the stall's undisturbed merchandise arrangement.\n\nIMMOBILE CHARACTER POSE — CANONICAL (identical wherever this character appears in ANY panel; on any conflict THIS POSE WINS): He is stretched out motionless on the market ground, his body supported by the ground rather than seated or propped against anything. The screenplay does not specify whether he is face-up or face-down, his head angle, or the exact arrangement of his arms and legs.\nIMMOBILE CHARACTER POSE — CANONICAL (identical wherever this character appears in ANY panel; on any conflict THIS POSE WINS): He is stretched out motionless on the market ground, his body supported by the ground rather than seated or propped against anything. The screenplay does not specify whether he is face-up or face-down, his head angle, or the exact arrangement of his arms and legs.\n\nIMMOBILE BODIES OBEY GRAVITY: a person who is dead or unconscious\nexerts NO muscular effort. Every part of their body — head, torso,\narms, hands, fingers, legs — rests fully on whatever supports it\n(floor, wall, furniture, their own lap) and hangs or slumps with\ngravity. NEVER show any part of an immobile person's body lifted,\nraised, held up in the air, or posed as if presenting something:\nan object in their grip stays clenched in a hand that itself lies\nfallen on a support — the hand does not hold the object up. If the\ncanonical pose leaves a body part unspecified, resolve it as the\nmost gravity-compliant, fully-supported position.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The same sunlit market remains immediately after the fight; the returned straw shoe is still in the stall's mesh bag. 일본인 징세원1: He lies unconscious but breathing on the dirt. 일본인 징세원2: He lies unconscious but breathing on the dirt.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 일본인 징세원1 (일본인 남성, 성인 나이대, 자연스러운 검은 머리); 일본인 징세원2 (일본인 남성, 성인 나이대, 자연스러운 검은 머리) — wearing: 서양식 검은색 양복 정장에 흰 와이셔츠. 팔에는 징세원을 나타내는 완장을 찼고, 가죽 구두를 신었다. — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.",
   "B": "Create ONE FINAL photorealistic live-action film still of the moment below — early 20th-century Korea, Jeolla region; all people are Korean unless explicitly stated otherwise. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 흙바닥에 대자로 뻗어 기절한 일본인 징세원1과 일본인 징세원2를 멀찍이 둥글게 둘러싸고 내려다보는 낡은 옷차림의 조선인 상인들.\n\nLOCATION (lock): The dirt fighting area outside the market's straw-sandal stall, now encircled at a distance by onlookers. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: wide shot\n- KEY BACKGROUND ELEMENTS: 흙바닥 (Exposed around the two fallen collectors); used as The visible clearance separates the bodies from the cautious onlookers and makes the encircling distance legible; 상인들의 낡은 옷차림 (Worn clothing visible around the perimeter); used as Provides an uneven human border through varied shoulder angles, bent elbows, and weight distribution rather than a uniform formation.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Natural daytime illumination and restrained saturation keep the aftermath plainly visible, with sober contrast rather than comic or sensational emphasis.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The people visible in that photo are NOT in this shot: never carry their faces, bodies or clothing onto anyone here. Each person in THIS shot is defined solely by the PEOPLE list, their own CHARACTER REFERENCE images and their wardrobe notes. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nPREVIOUS STILL USAGE (follow exactly — what to take from the attached still and what to exclude; it governs place and objects only, never who is in this shot or how the camera sees them): Take the straw-sandal stall, woven display materials, and sunlit dirt ground. Exclude the fleeing laborer and the earlier inspection action; do not restore the stall's undisturbed merchandise arrangement.\n\nIMMOBILE CHARACTER POSE — CANONICAL (identical wherever this character appears in ANY panel; on any conflict THIS POSE WINS): He is stretched out motionless on the market ground, his body supported by the ground rather than seated or propped against anything. The screenplay does not specify whether he is face-up or face-down, his head angle, or the exact arrangement of his arms and legs.\nIMMOBILE CHARACTER POSE — CANONICAL (identical wherever this character appears in ANY panel; on any conflict THIS POSE WINS): He is stretched out motionless on the market ground, his body supported by the ground rather than seated or propped against anything. The screenplay does not specify whether he is face-up or face-down, his head angle, or the exact arrangement of his arms and legs.\n\nIMMOBILE BODIES OBEY GRAVITY: a person who is dead or unconscious\nexerts NO muscular effort. Every part of their body — head, torso,\narms, hands, fingers, legs — rests fully on whatever supports it\n(floor, wall, furniture, their own lap) and hangs or slumps with\ngravity. NEVER show any part of an immobile person's body lifted,\nraised, held up in the air, or posed as if presenting something:\nan object in their grip stays clenched in a hand that itself lies\nfallen on a support — the hand does not hold the object up. If the\ncanonical pose leaves a body part unspecified, resolve it as the\nmost gravity-compliant, fully-supported position.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The same sunlit market remains immediately after the fight; the returned straw shoe is still in the stall's mesh bag. 일본인 징세원1: He lies unconscious but breathing on the dirt. 일본인 징세원2: He lies unconscious but breathing on the dirt.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 일본인 징세원1 (일본인 남성, 성인 나이대, 자연스러운 검은 머리); 일본인 징세원2 (일본인 남성, 성인 나이대, 자연스러운 검은 머리) — wearing: 서양식 검은색 양복 정장에 흰 와이셔츠. 팔에는 징세원을 나타내는 완장을 찼고, 가죽 구두를 신었다. — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.\n\nCOMPOSITION VARIATION (this candidate only): keep the place, the people and the moment exactly as specified, and keep from the CAMERA & FRAME contract its subject scale, screen placement and key background elements exactly. For this candidate only, its camera angle and height are deliberately left open: choose a camera position distinctly different from the obvious one, as a film director picking a second setup on the same blocking. Stay true to any explicit framing the SHOT TEXT itself dictates."
  },
  "shot_run_spend_attempt_count": 1,
  "readings": [
   {
    "label": "A",
    "direction": "상인들의 시선이 바닥에 쓰러진 두 남자를 향해 있음.",
    "built_space": "짚신이 진열된 상점 앞의 흙바닥 공터로, 레퍼런스와 일치하는 배경 구조임.",
    "entities": "검은 양복에 흰 장갑을 낀 징세원 두 명과 낡은 한복 차림의 조선인 상인들. 그러나 배경의 상인 중 한 명에게 징세원 완장이 적용됨.",
    "hard_violations": [
     "[gemini-pro] 조선인 상인에게 징세원 완장이 잘못 적용됨 (leaked marker/text)",
     "[gpt-high] 중앙에 서 있는 상인의 남색 완장에 판독 가능한 한글이 노출되어, 읽을 수 있는 글자를 전혀 두지 말라는 지시를 위반한다."
    ],
    "physics": "바닥에 쓰러진 두 남자의 몸이 흙바닥에 자연스럽게 닿아 있음."
   },
   {
    "label": "B",
    "direction": "둥글게 모인 상인들이 중앙에 쓰러진 두 사람을 내려다보고 있음.",
    "built_space": "짚신 상점이 있는 시장 공터로, 인물들 사이의 흙바닥 공간이 넓게 확보됨.",
    "entities": "검은 양복, 흰 장갑, 완장, 구두를 착용한 두 명의 징세원과 이들을 둘러싼 다양한 자세의 조선인 상인들이 정확히 묘사됨.",
    "hard_violations": [],
    "physics": "쓰러진 인물들이 완전히 힘이 빠진 채로 바닥에 밀착되어 누워 있음."
   }
  ],
  "cross_model_order": {
   "policy": "cross_model_order_v2_slot0_fwd_slot1_rev_parallel2",
   "slots": [
    {
     "model": "gemini-pro",
     "order": "forward",
     "display_to_canonical": {
      "A": "A",
      "B": "B"
     },
     "raw": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 9,
        "verdict_ko": "프롬프트가 요구한 와이드 샷 구도와 쓰러진 인물들을 멀찍이 둘러싼 상인들의 모습을 정확하게 구현했으며, 복장과 쓰러진 자세 모두 훌륭합니다."
       },
       {
        "label": "A",
        "score": 4,
        "verdict_ko": "조선인 상인의 팔에 일본인 징세원의 완장이 잘못 렌더링되는 치명적인 오류가 발생했습니다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "상인들의 시선이 바닥에 쓰러진 두 남자를 향해 있음.",
        "built_space": "짚신이 진열된 상점 앞의 흙바닥 공터로, 레퍼런스와 일치하는 배경 구조임.",
        "entities": "검은 양복에 흰 장갑을 낀 징세원 두 명과 낡은 한복 차림의 조선인 상인들. 그러나 배경의 상인 중 한 명에게 징세원 완장이 적용됨.",
        "hard_violations": [
         "조선인 상인에게 징세원 완장이 잘못 적용됨 (leaked marker/text)"
        ],
        "physics": "바닥에 쓰러진 두 남자의 몸이 흙바닥에 자연스럽게 닿아 있음."
       },
       {
        "label": "B",
        "direction": "둥글게 모인 상인들이 중앙에 쓰러진 두 사람을 내려다보고 있음.",
        "built_space": "짚신 상점이 있는 시장 공터로, 인물들 사이의 흙바닥 공간이 넓게 확보됨.",
        "entities": "검은 양복, 흰 장갑, 완장, 구두를 착용한 두 명의 징세원과 이들을 둘러싼 다양한 자세의 조선인 상인들이 정확히 묘사됨.",
        "hard_violations": [],
        "physics": "쓰러진 인물들이 완전히 힘이 빠진 채로 바닥에 밀착되어 누워 있음."
       }
      ],
      "all_candidates_fail": false
     },
     "normalized": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 9,
        "verdict_ko": "프롬프트가 요구한 와이드 샷 구도와 쓰러진 인물들을 멀찍이 둘러싼 상인들의 모습을 정확하게 구현했으며, 복장과 쓰러진 자세 모두 훌륭합니다."
       },
       {
        "label": "A",
        "score": 4,
        "verdict_ko": "조선인 상인의 팔에 일본인 징세원의 완장이 잘못 렌더링되는 치명적인 오류가 발생했습니다."
       }
      ],
      "all_candidates_fail": false,
      "readings": [
       {
        "label": "A",
        "direction": "상인들의 시선이 바닥에 쓰러진 두 남자를 향해 있음.",
        "built_space": "짚신이 진열된 상점 앞의 흙바닥 공터로, 레퍼런스와 일치하는 배경 구조임.",
        "entities": "검은 양복에 흰 장갑을 낀 징세원 두 명과 낡은 한복 차림의 조선인 상인들. 그러나 배경의 상인 중 한 명에게 징세원 완장이 적용됨.",
        "hard_violations": [
         "조선인 상인에게 징세원 완장이 잘못 적용됨 (leaked marker/text)"
        ],
        "physics": "바닥에 쓰러진 두 남자의 몸이 흙바닥에 자연스럽게 닿아 있음."
       },
       {
        "label": "B",
        "direction": "둥글게 모인 상인들이 중앙에 쓰러진 두 사람을 내려다보고 있음.",
        "built_space": "짚신 상점이 있는 시장 공터로, 인물들 사이의 흙바닥 공간이 넓게 확보됨.",
        "entities": "검은 양복, 흰 장갑, 완장, 구두를 착용한 두 명의 징세원과 이들을 둘러싼 다양한 자세의 조선인 상인들이 정확히 묘사됨.",
        "hard_violations": [],
        "physics": "쓰러진 인물들이 완전히 힘이 빠진 채로 바닥에 밀착되어 누워 있음."
       }
      ]
     },
     "ok": true
    },
    {
     "model": "gpt-high",
     "order": "reverse",
     "display_to_canonical": {
      "A": "B",
      "B": "A"
     },
     "raw": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 8,
        "verdict_ko": "두 징세원과 상인들 사이의 넓은 흙바닥, 멀찍이 에워싼 원형 배치를 와이드 숏으로 명확히 구현하지만, 상인들의 자세가 지나치게 정렬되어 있다."
       },
       {
        "label": "B",
        "score": 3,
        "verdict_ko": "징세원들의 뻗은 자세는 분명하지만 상인들이 너무 가까우며, 중앙 상인에게 글자가 읽히는 완장까지 부여해 무문자 지시와 인물별 소품 배정을 위반한다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "둘레의 상인들은 대체로 몸과 얼굴을 중앙의 쓰러진 두 남자 쪽으로 돌리고 있다. 좌우와 전경 인물들의 내려다보는 방향은 명확하지만, 후경 중앙 일부는 고개가 비교적 꼿꼿해 정확한 시선 도착점이 덜 분명하다. 두 징세원은 얼굴이 위를 향한 채 눈을 감고 있으며, 겨누는 물건이나 이동 중인 몸은 없다.",
        "built_space": "왼쪽에 짚신을 여러 줄로 매단 목조 판매대 한 구역, 낮은 진열 평상, 기대 놓은 사다리 하나와 바닥의 멍석이 보인다. 오른쪽에는 목조 점포와 엮은 차양이 있고 중앙 뒤로 시장 통로가 이어진다. 참고의 목재·짚 진열·흙바닥 재료는 이어지지만 공간은 더 넓은 마당처럼 보인다. 약 스무 명 이상의 상인이 전경까지 이어지는 원형 둘레를 이루고, 두 몸과 둘레 사이에는 수 걸음 너비의 빈 흙바닥이 있다. 불가능한 반사나 명백한 고정 설비 중복은 보이지 않는다.",
        "entities": "중앙에는 검은 머리의 동아시아계 성인 남성 두 명만 쓰러져 있다. 검은 양복, 흰 셔츠, 검은 넥타이, 흰 장갑, 가죽 구두와 어두운 완장이 보인다. 얼굴이 작아 징세원1의 참고 얼굴과 정확히 일치하는지는 확인하기 어렵고, 일본인이라는 국적도 외모만으로 확정할 수 없다. 둘레에는 해진 회백색 저고리·바지와 치마를 입은 성인 남녀 상인들이 있다. 짚신과 엮은 생활용품은 보이지만 반환된 짚신이 든 특정 망 가방은 식별되지 않는다. 완장에는 작은 흰 흔적이 있으나 이 크기에서 확실히 읽히는 문구는 없다.",
        "hard_violations": [],
        "physics": "두 징세원은 등과 골반, 뻗은 다리 및 벌어진 팔을 흙바닥에 내려놓고 있다. 장갑 낀 손도 바닥 가까이에 떨어져 있으며, 들어 올린 물건이나 지지 없이 떠 있는 신체 부위는 보이지 않는다. 상인들은 두 발 또는 한쪽에 실린 체중으로 서 있고, 왼쪽 인물의 등짐 바구니는 등과 어깨끈에 지지된다. 다만 다수 상인의 팔과 어깨가 비슷하게 정렬되어 요청한 불균일한 인간 테두리보다 정적인 대형에 가깝다."
       },
       {
        "label": "B",
        "direction": "앞쪽 반원을 이루는 상인 대부분이 고개를 숙여 발앞의 두 징세원을 보고 있다. 중앙 뒤쪽 일부 인물은 정면을 향해 시선 대상이 덜 명확하다. 쓰러진 두 사람은 눈을 감고 얼굴을 위로 향하며, 무기나 겨누는 소품은 없다.",
        "built_space": "왼쪽에는 짚신을 매단 목조 판매대 한 구역과 낮은 평상, 펼쳐진 멍석이 있고, 오른쪽에는 엮은 차양과 큰 망이 걸려 있다. 뒤로 이어지는 좁은 점포 통로는 참고 장소의 재료와 밀도에 비교적 가깝다. 그러나 가까운 상인들이 두 몸의 머리와 손 주변까지 접근해 있어, 요청된 넓은 흙바닥 완충 공간 대신 좁은 반원형 관람 배치로 읽힌다. 두 몸은 화면 하단을 크게 차지하며 전경 쪽 원형 둘레는 보이지 않는다. 불가능한 반사는 없다.",
        "entities": "검은 머리의 동아시아계 성인 남성 두 명이 검은 양복, 흰 셔츠, 넥타이, 흰 장갑, 가죽 구두와 완장을 착용하고 누워 있다. 징세원1 참고의 전반적인 얼굴형과 복장에는 가까우나 동일 얼굴이라고 단정하기는 어렵다. 주변에는 낡은 조선식 옷차림의 성인 남녀가 있다. 그런데 중앙의 서 있는 남성 상인에게도 흰 한글이 보이는 징세원식 남색 완장이 붙어 있다. 짚신 진열과 망은 있으나 반환된 특정 짚신의 수납 상태는 확인되지 않는다.",
        "hard_violations": [
         "중앙에 서 있는 상인의 남색 완장에 판독 가능한 한글이 노출되어, 읽을 수 있는 글자를 전혀 두지 말라는 지시를 위반한다."
        ],
        "physics": "두 징세원은 등과 머리, 골반을 땅에 대고 팔다리를 벌리고 있다. 발은 뒤꿈치와 구두 가장자리로 지지되고, 장갑 낀 손과 팔도 바닥에 놓여 있어 공중에 떠 있는 물체나 몸은 보이지 않는다. 상인들의 발도 지면과 접촉한다. 자세 자체는 기절해 뻗은 몸으로 가능하지만 상인들의 가까운 발 위치 때문에 조심스럽게 거리를 두었다는 행동은 약해진다."
       }
      ],
      "all_candidates_fail": false
     },
     "normalized": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 8,
        "verdict_ko": "두 징세원과 상인들 사이의 넓은 흙바닥, 멀찍이 에워싼 원형 배치를 와이드 숏으로 명확히 구현하지만, 상인들의 자세가 지나치게 정렬되어 있다."
       },
       {
        "label": "A",
        "score": 3,
        "verdict_ko": "징세원들의 뻗은 자세는 분명하지만 상인들이 너무 가까우며, 중앙 상인에게 글자가 읽히는 완장까지 부여해 무문자 지시와 인물별 소품 배정을 위반한다."
       }
      ],
      "all_candidates_fail": false,
      "readings": [
       {
        "label": "B",
        "direction": "둘레의 상인들은 대체로 몸과 얼굴을 중앙의 쓰러진 두 남자 쪽으로 돌리고 있다. 좌우와 전경 인물들의 내려다보는 방향은 명확하지만, 후경 중앙 일부는 고개가 비교적 꼿꼿해 정확한 시선 도착점이 덜 분명하다. 두 징세원은 얼굴이 위를 향한 채 눈을 감고 있으며, 겨누는 물건이나 이동 중인 몸은 없다.",
        "built_space": "왼쪽에 짚신을 여러 줄로 매단 목조 판매대 한 구역, 낮은 진열 평상, 기대 놓은 사다리 하나와 바닥의 멍석이 보인다. 오른쪽에는 목조 점포와 엮은 차양이 있고 중앙 뒤로 시장 통로가 이어진다. 참고의 목재·짚 진열·흙바닥 재료는 이어지지만 공간은 더 넓은 마당처럼 보인다. 약 스무 명 이상의 상인이 전경까지 이어지는 원형 둘레를 이루고, 두 몸과 둘레 사이에는 수 걸음 너비의 빈 흙바닥이 있다. 불가능한 반사나 명백한 고정 설비 중복은 보이지 않는다.",
        "entities": "중앙에는 검은 머리의 동아시아계 성인 남성 두 명만 쓰러져 있다. 검은 양복, 흰 셔츠, 검은 넥타이, 흰 장갑, 가죽 구두와 어두운 완장이 보인다. 얼굴이 작아 징세원1의 참고 얼굴과 정확히 일치하는지는 확인하기 어렵고, 일본인이라는 국적도 외모만으로 확정할 수 없다. 둘레에는 해진 회백색 저고리·바지와 치마를 입은 성인 남녀 상인들이 있다. 짚신과 엮은 생활용품은 보이지만 반환된 짚신이 든 특정 망 가방은 식별되지 않는다. 완장에는 작은 흰 흔적이 있으나 이 크기에서 확실히 읽히는 문구는 없다.",
        "hard_violations": [],
        "physics": "두 징세원은 등과 골반, 뻗은 다리 및 벌어진 팔을 흙바닥에 내려놓고 있다. 장갑 낀 손도 바닥 가까이에 떨어져 있으며, 들어 올린 물건이나 지지 없이 떠 있는 신체 부위는 보이지 않는다. 상인들은 두 발 또는 한쪽에 실린 체중으로 서 있고, 왼쪽 인물의 등짐 바구니는 등과 어깨끈에 지지된다. 다만 다수 상인의 팔과 어깨가 비슷하게 정렬되어 요청한 불균일한 인간 테두리보다 정적인 대형에 가깝다."
       },
       {
        "label": "A",
        "direction": "앞쪽 반원을 이루는 상인 대부분이 고개를 숙여 발앞의 두 징세원을 보고 있다. 중앙 뒤쪽 일부 인물은 정면을 향해 시선 대상이 덜 명확하다. 쓰러진 두 사람은 눈을 감고 얼굴을 위로 향하며, 무기나 겨누는 소품은 없다.",
        "built_space": "왼쪽에는 짚신을 매단 목조 판매대 한 구역과 낮은 평상, 펼쳐진 멍석이 있고, 오른쪽에는 엮은 차양과 큰 망이 걸려 있다. 뒤로 이어지는 좁은 점포 통로는 참고 장소의 재료와 밀도에 비교적 가깝다. 그러나 가까운 상인들이 두 몸의 머리와 손 주변까지 접근해 있어, 요청된 넓은 흙바닥 완충 공간 대신 좁은 반원형 관람 배치로 읽힌다. 두 몸은 화면 하단을 크게 차지하며 전경 쪽 원형 둘레는 보이지 않는다. 불가능한 반사는 없다.",
        "entities": "검은 머리의 동아시아계 성인 남성 두 명이 검은 양복, 흰 셔츠, 넥타이, 흰 장갑, 가죽 구두와 완장을 착용하고 누워 있다. 징세원1 참고의 전반적인 얼굴형과 복장에는 가까우나 동일 얼굴이라고 단정하기는 어렵다. 주변에는 낡은 조선식 옷차림의 성인 남녀가 있다. 그런데 중앙의 서 있는 남성 상인에게도 흰 한글이 보이는 징세원식 남색 완장이 붙어 있다. 짚신 진열과 망은 있으나 반환된 특정 짚신의 수납 상태는 확인되지 않는다.",
        "hard_violations": [
         "중앙에 서 있는 상인의 남색 완장에 판독 가능한 한글이 노출되어, 읽을 수 있는 글자를 전혀 두지 말라는 지시를 위반한다."
        ],
        "physics": "두 징세원은 등과 머리, 골반을 땅에 대고 팔다리를 벌리고 있다. 발은 뒤꿈치와 구두 가장자리로 지지되고, 장갑 낀 손과 팔도 바닥에 놓여 있어 공중에 떠 있는 물체나 몸은 보이지 않는다. 상인들의 발도 지면과 접촉한다. 자세 자체는 기절해 뻗은 몸으로 가능하지만 상인들의 가까운 발 위치 때문에 조심스럽게 거리를 두었다는 행동은 약해진다."
       }
      ]
     },
     "ok": true
    }
   ],
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "slot_winner_match": true,
   "slot_winner": {
    "gemini-pro": "B",
    "gpt-high": "B"
   },
   "route": "cross_slot_agree"
  },
  "dual": {
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "normalized": {
    "A": 0.819,
    "B": 2.0
   },
   "adjusted": {
    "A": 0.569,
    "B": 2.0
   },
   "violations": {
    "A": [
     "[gemini-pro] 조선인 상인에게 징세원 완장이 잘못 적용됨 (leaked marker/text)",
     "[gpt-high] 중앙에 서 있는 상인의 남색 완장에 판독 가능한 한글이 노출되어, 읽을 수 있는 글자를 전혀 두지 말라는 지시를 위반한다."
    ]
   },
   "per_model_winner": {
    "gemini-pro": "B",
    "gpt-high": "B"
   },
   "agreed": true
  },
  "totals": {
   "B": 2000,
   "A": 569
  },
  "selected": "B",
  "ranking": [
   "B",
   "A"
  ],
  "verdicts": [
   {
    "label": "B",
    "score": 2000,
    "verdict_ko": "프롬프트가 요구한 와이드 샷 구도와 쓰러진 인물들을 멀찍이 둘러싼 상인들의 모습을 정확하게 구현했으며, 복장과 쓰러진 자세 모두 훌륭합니다."
   },
   {
    "label": "A",
    "score": 569,
    "verdict_ko": "조선인 상인의 팔에 일본인 징세원의 완장이 잘못 렌더링되는 치명적인 오류가 발생했습니다.  ★위반: [gemini-pro] 조선인 상인에게 징세원 완장이 잘못 적용됨 (leaked marker/text) / [gpt-high] 중앙에 서 있는 상인의 남색 완장에 판독 가능한 한글이 노출되어, 읽을 수 있는 글자를 전혀 두지 말라는 지시를 위반한다."
   }
  ],
  "refs": [
   {
    "label": "PREVIOUS SHOT STILL — a visually related earlier shot of this same place: the location's look, materials, fixed features and lighting mood are LOCKED to this photo; never copy its camera framing. The people visible in that photo are NOT in this shot: never carry their faces, bodies or clothing onto anyone here. Each person in THIS shot is defined solely by the PEOPLE list, their own CHARACTER REFERENCE images and their wardrobe notes. If this photo shows a character who cannot move (dead or unconscious), that character's exact pose, position and orientation are ALSO LOCKED — treat their body as a fixed prop of the set that only the camera moves around.",
    "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/db00517b-165b-4005-a287-f28936b59b95/images/272bd7af-9294-4490-98c6-88a40e9deacc/scene/recipe/S5sh39_sel.png",
    "asset_id": "95625762-ab80-442c-8b8d-465a12e4bf26",
    "role": "prev_still"
   },
   {
    "label": "CHARACTER REFERENCE — 일본인 징세원1: the exact person appearing in this shot; match face, hair and build exactly.",
    "path": "<bytes:1055496>",
    "asset_id": "cdacfcb2-5da3-4e19-b2f7-22a2df555c28",
    "role": "character_ref"
   }
  ],
  "critique_skipped": true,
  "shot_run_produced": false,
  "shot_run_uid": "06aa1406-fa1a-709c-b71b-2f0e4a5be4e4",
  "ref_mode": "prev+엔티티",
  "share_plan": {
   "ref_plan": "prev",
   "prev_anchor_tag": "S5sh39"
  }
 },
 "S6sh1::cine": {
  "applied": true,
  "attempted_at": "2026-09-09T01:09:23.809468+00:00",
  "fingerprint": "97ff0edfca5e09512f6a70f2159dbc1dc4edc2c08a08c8480405c5f0d0cf25db",
  "fingerprint_version": 2,
  "provider": "mai",
  "endpoint": "openrouter/chat-completions",
  "model": "microsoft/mai-image-2.6",
  "pack": "24.202608252115",
  "source_file": "S6sh1_sel.png",
  "source_sha256": "b3f91fdd4a0cd7972dde1816b0f14f3de1725e08444b46127c17f33de2e30805",
  "file": "S6sh1_cine.png",
  "staged_sha256": "b1c8a59533fee50b7b40b4640a199e83d3da3596435b669cbd3e130977fd3bba",
  "latency_ms": 28182
 },
 "S6sh4::signage": {
  "fp": "63c45b1272786f34",
  "inscriptions": [],
  "cues": [],
  "dropped": []
 },
 "groupbg::market_escape_alley": {
  "input_fingerprint": "89d384ae13953fbb",
  "meta": {
   "model": "gpt-image-2",
   "size": "1536x864",
   "pack": "11.202607220237",
   "contract": "bgfirst_full_v3",
   "group_sig": {
    "key": "market_escape_alley",
    "tags": [
     "S6sh4"
    ]
   },
   "context_sig": "3978a2c181a74ac7"
  },
  "prompt": "Create ONE empty live-action location background photograph — NO PEOPLE, no figures, no body parts, no silhouettes, no shadows or reflections of people anywhere.\n\"Empty\" means no people only: KEEP the location's inherent occupants and stock that define the place — animals in an animal shelter, pen or farm, goods and produce in a market, moored boats in a harbour — unless the shot text explicitly removes them.\nThe FIRST attached image is a thin-line storyboard sketch of a shot that happens at this location: use it ONLY as spatial evidence — what this place contains, how its ground, structures and landmarks are arranged and proportioned. Ignore the sketched people and arrows entirely, and do NOT copy its line style: render a fully photographic, physically plausible real place that fits THE LOCATION text below.\nHUMAN-SCALE CALIBRATION: derive every structure's true size from human-scale elements — a door ≈ 2m, a window ≈ 1–1.5m wide, one storey ≈ 2.5–3m; never inflate a small structure or shrink a large one.\n\nTHE LOCATION — early 20th-century Korea, Jeolla region; all people are Korean unless explicitly stated otherwise: The far end of a crowded outdoor market alley, along the route away from the fight.\n\nLOCATION DETAIL (descriptive evidence about this place from the production's location records — evidence, not a staging order): use it to understand what kind of place this is and what it permanently contains. Carry over ONLY its enduring physical features — layout, structures, ground and materials, fixed equipment and installations. Do NOT treat any lighting, weather or time-of-day wording, momentary object states, depicted events or subjective impressions in it as fixed facts of the place; the TIME OF DAY lock below is the sole authority for time and lighting.\n복내장터 좌판 거리·장터 입구·골목: 다양한 좌판과 짐꾼, 상인들이 밀집한 넓은 흙바닥 노천 장터. 포목전, 짚신전, 옹기전, 닭장, 생선 좌판이 빽빽하게 들어서 있으며, 벽에 관의 방이 붙어 있고 상황에 따라 찢어진 멍석과 부서진 좌판이 널브러져 있다. (특징: 넓은 흙바닥; 나무와 짚으로 얽어 만든 노점 좌판 (옹기, 짚신, 채소, 포목 등); 나무 닭장과 광주리; 방(벽보)이 붙어 있는 벽; 깨진 항아리와 뒤집힌 멍석, 부서진 나무판자들)\n\nSCENE EVIDENCE (verbatim quotes from the screenplay about this place — treat them as evidence of what the location physically contains and looks like; stage the PLACE those moments happen in, but do NOT depict the momentary actions, people or staged props themselves):\n- 사람들이 고개를 든다. 저 멀리 달아나는 규홍\n장터 골목, 사람들 사이로 소문이 옮겨 붙는다.\n\nTIME OF DAY (lock): day.\n\nRender ONE photorealistic empty location photograph, 16:9, neutral enough that every shot of this place can be staged from it later. No readable writing anywhere: surfaces that would carry writing may be present, but stage any wording out of legibility — an oblique angle, distance, shallow focus. No captions, watermarks or overlay text.",
  "effective_prompt": "Create ONE empty live-action location background photograph — NO PEOPLE, no figures, no body parts, no silhouettes, no shadows or reflections of people anywhere.\n\"Empty\" means no people only: KEEP the location's inherent occupants and stock that define the place — animals in an animal shelter, pen or farm, goods and produce in a market, moored boats in a harbour — unless the shot text explicitly removes them.\nThe FIRST attached image is a thin-line storyboard sketch of a shot that happens at this location: use it ONLY as spatial evidence — what this place contains, how its ground, structures and landmarks are arranged and proportioned. Ignore the sketched people and arrows entirely, and do NOT copy its line style: render a fully photographic, physically plausible real place that fits THE LOCATION text below.\nHUMAN-SCALE CALIBRATION: derive every structure's true size from human-scale elements — a door ≈ 2m, a window ≈ 1–1.5m wide, one storey ≈ 2.5–3m; never inflate a small structure or shrink a large one.\n\nTHE LOCATION — early 20th-century Korea, Jeolla region; all people are Korean unless explicitly stated otherwise: The far end of a crowded outdoor market alley, along the route away from the fight.\n\nLOCATION DETAIL (descriptive evidence about this place from the production's location records — evidence, not a staging order): use it to understand what kind of place this is and what it permanently contains. Carry over ONLY its enduring physical features — layout, structures, ground and materials, fixed equipment and installations. Do NOT treat any lighting, weather or time-of-day wording, momentary object states, depicted events or subjective impressions in it as fixed facts of the place; the TIME OF DAY lock below is the sole authority for time and lighting.\n복내장터 좌판 거리·장터 입구·골목: 다양한 좌판과 짐꾼, 상인들이 밀집한 넓은 흙바닥 노천 장터. 포목전, 짚신전, 옹기전, 닭장, 생선 좌판이 빽빽하게 들어서 있으며, 벽에 관의 방이 붙어 있고 상황에 따라 찢어진 멍석과 부서진 좌판이 널브러져 있다. (특징: 넓은 흙바닥; 나무와 짚으로 얽어 만든 노점 좌판 (옹기, 짚신, 채소, 포목 등); 나무 닭장과 광주리; 방(벽보)이 붙어 있는 벽; 깨진 항아리와 뒤집힌 멍석, 부서진 나무판자들)\n\nSCENE EVIDENCE (verbatim quotes from the screenplay about this place — treat them as evidence of what the location physically contains and looks like; stage the PLACE those moments happen in, but do NOT depict the momentary actions, people or staged props themselves):\n- 사람들이 고개를 든다. 저 멀리 달아나는 규홍\n장터 골목, 사람들 사이로 소문이 옮겨 붙는다.\n\nTIME OF DAY (lock): day.\n\nRender ONE photorealistic empty location photograph, 16:9, neutral enough that every shot of this place can be staged from it later. No readable writing anywhere: surfaces that would carry writing may be present, but stage any wording out of legibility — an oblique angle, distance, shallow focus. No captions, watermarks or overlay text.",
  "bg_path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/db00517b-165b-4005-a287-f28936b59b95/images/272bd7af-9294-4490-98c6-88a40e9deacc/scene/recipe/groupbg_market_escape_alley_2bc28c.png",
  "asset_id": "8916c734-358b-4cd6-857e-d37a2057a57a",
  "input_asset_ids": [
   "87e8fdb7-2c4f-4534-b30e-95287a82fa63"
  ],
  "origin_tag": "S6sh4",
  "place_text": "The far end of a crowded outdoor market alley, along the route away from the fight.",
  "origin_inputs": {
   "place_text": "The far end of a crowded outdoor market alley, along the route away from the fight.",
   "time_of_day_en": "day",
   "conti_asset_id": "87e8fdb7-2c4f-4534-b30e-95287a82fa63"
  }
 },
 "S6sh4::bgfirst_bg": {
  "input_fingerprint": "55946a2055f5b0d5",
  "prompt": "Create the EMPTY BACKGROUND PLATE for one film shot — NO PEOPLE, no figures, no body parts, no sketch lines, no arrows anywhere.\n\"Empty\" means no people only: KEEP the location's inherent occupants and stock that define the place — animals in an animal shelter, pen or farm, goods in a market, moored boats in a harbour — unless the shot text explicitly removes them.\nThe FIRST attached image is a thin-line storyboard sketch: use ONLY its camera angle, horizon, perspective and the placement/size of buildings and set masses — ignore the sketched people and arrows entirely. The SECOND attached image (LOCATION PHOTOGRAPH) is the real place: take its architecture, materials, signage and fixed features, and RE-PROJECT them into the sketch's camera. If the photograph's camera differs from the sketch's, the sketch's camera wins.\nHUMAN-SCALE CALIBRATION: derive every structure's true size from human-scale elements — a door ≈ 2m, a window ≈ 1–1.5m wide, one storey ≈ 2.5–3m; never inflate a small structure or shrink a large one.\n\nSHOT TEXT this background must serve (Korean): 북적이는 사람들 사이로 장터 골목 끝을 향해 뛰어가는 도중, 한쪽 다리를 길게 뻗고 다른 쪽 발은 지면에서 막 떨어진 자세의 안규홍의 작아진 뒷모습.\n\nLOCATION (lock): The far end of a crowded outdoor market alley, along the route away from the fight.\n\nTIME OF DAY (lock): day.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: wide shot\n- FRAME LAYOUT: far end of the market alley in the upper-center of the frame, background.\n- KEY BACKGROUND ELEMENTS: 장터 골목 (Occupied by people, with 안규홍 running toward its far end) — Its length recedes obliquely from the foreground edges toward the upper-center distance; used as Maintains a narrow sightline to the runner through staggered shoulders and uneven gaps.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Keep natural daylight and muted saturation continuous through the montage, with sufficient tonal separation to read the small runner without spotlighting him.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nRender ONE photorealistic empty location photograph, 16:9, that this shot can be staged inside later. No readable writing anywhere: surfaces that would carry writing may be present, but stage any wording out of legibility — an oblique angle, distance, shallow focus. No captions, watermarks or overlay text.",
  "effective_prompt": "Create the EMPTY BACKGROUND PLATE for one film shot — NO PEOPLE, no figures, no body parts, no sketch lines, no arrows anywhere.\n\"Empty\" means no people only: KEEP the location's inherent occupants and stock that define the place — animals in an animal shelter, pen or farm, goods in a market, moored boats in a harbour — unless the shot text explicitly removes them.\nThe FIRST attached image is a thin-line storyboard sketch: use ONLY its camera angle, horizon, perspective and the placement/size of buildings and set masses — ignore the sketched people and arrows entirely. The SECOND attached image (LOCATION PHOTOGRAPH) is the real place: take its architecture, materials, signage and fixed features, and RE-PROJECT them into the sketch's camera. If the photograph's camera differs from the sketch's, the sketch's camera wins.\nHUMAN-SCALE CALIBRATION: derive every structure's true size from human-scale elements — a door ≈ 2m, a window ≈ 1–1.5m wide, one storey ≈ 2.5–3m; never inflate a small structure or shrink a large one.\n\nSHOT TEXT this background must serve (Korean): 북적이는 사람들 사이로 장터 골목 끝을 향해 뛰어가는 도중, 한쪽 다리를 길게 뻗고 다른 쪽 발은 지면에서 막 떨어진 자세의 안규홍의 작아진 뒷모습.\n\nLOCATION (lock): The far end of a crowded outdoor market alley, along the route away from the fight.\n\nTIME OF DAY (lock): day.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: wide shot\n- FRAME LAYOUT: far end of the market alley in the upper-center of the frame, background.\n- KEY BACKGROUND ELEMENTS: 장터 골목 (Occupied by people, with 안규홍 running toward its far end) — Its length recedes obliquely from the foreground edges toward the upper-center distance; used as Maintains a narrow sightline to the runner through staggered shoulders and uneven gaps.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Keep natural daylight and muted saturation continuous through the montage, with sufficient tonal separation to read the small runner without spotlighting him.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nRender ONE photorealistic empty location photograph, 16:9, that this shot can be staged inside later. No readable writing anywhere: surfaces that would carry writing may be present, but stage any wording out of legibility — an oblique angle, distance, shallow focus. No captions, watermarks or overlay text.",
  "bg_path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/db00517b-165b-4005-a287-f28936b59b95/images/272bd7af-9294-4490-98c6-88a40e9deacc/scene/recipe/S6sh4__bgfirst_bg.png",
  "asset_id": "8698e3a6-8bb7-42c0-a471-c2dca29f0fb4",
  "input_asset_ids": [
   "87e8fdb7-2c4f-4534-b30e-95287a82fa63",
   "8916c734-358b-4cd6-857e-d37a2057a57a"
  ]
 },
 "S6sh4": {
  "input_fingerprint": "a5c9bc30654787cc",
  "prompt": "Create ONE FINAL photorealistic live-action film still of the moment below — early 20th-century Korea, Jeolla region; all people are Korean unless explicitly stated otherwise. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 북적이는 사람들 사이로 장터 골목 끝을 향해 뛰어가는 도중, 한쪽 다리를 길게 뻗고 다른 쪽 발은 지면에서 막 떨어진 자세의 안규홍의 작아진 뒷모습.\n\nLOCATION (lock): The far end of a crowded outdoor market alley, along the route away from the fight. The shot takes place here — the attached LOCATION PHOTOGRAPH shows the exact spot.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: wide shot\n- FRAME LAYOUT: far end of the market alley in the upper-center of the frame, background.\n- KEY BACKGROUND ELEMENTS: 장터 골목 (Occupied by people, with 안규홍 running toward its far end) — Its length recedes obliquely from the foreground edges toward the upper-center distance; used as Maintains a narrow sightline to the runner through staggered shoulders and uneven gaps.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Keep natural daylight and muted saturation continuous through the montage, with sufficient tonal separation to read the small runner without spotlighting him.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The sunlit market remains unchanged immediately after the fight. 안규홍: He continues fleeing with his carrying frame, worn straw shoes and injured mouth.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 안규홍 (한국인 남성, 청년 나이대, 검은 머리, 단단히 묶은 머리) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.",
  "roll_prompts": {
   "A": "Stage the shot. The FIRST attached image (SHOT BACKGROUND) is the finished empty background of this shot — keep it EXACTLY: its camera, perspective, architecture, lighting and every fixed feature stay untouched. The SECOND attached image (LAYOUT SKETCH) tells you ONLY where the people go: each sketched person's position, screen size, pose and the gaze/motion arrows. Ignore the sketch's background lines. The CHARACTER REFERENCE photographs show the real people.\nPlace the real people into the background at exactly the sketched positions, sizes and poses, following the arrow directions. No sketch lines or arrows may remain.\n\nCreate ONE FINAL photorealistic live-action film still of the moment below — early 20th-century Korea, Jeolla region; all people are Korean unless explicitly stated otherwise. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 북적이는 사람들 사이로 장터 골목 끝을 향해 뛰어가는 도중, 한쪽 다리를 길게 뻗고 다른 쪽 발은 지면에서 막 떨어진 자세의 안규홍의 작아진 뒷모습.\n\nLOCATION (lock): The far end of a crowded outdoor market alley, along the route away from the fight. The shot takes place here — the FIRST attached image (SHOT BACKGROUND) is this exact place, already built: its ground, structures, horizon, materials and lighting are the finished truth of this location and must not be redesigned or replaced. No location photograph is attached — read the place from that image alone, and add no scenery, structure, vehicle or fixture that it does not already show. This lock governs the place only; the figures in the shot follow the staging and pose instructions.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: wide shot\n- FRAME LAYOUT: far end of the market alley in the upper-center of the frame, background.\n- KEY BACKGROUND ELEMENTS: 장터 골목 (Occupied by people, with 안규홍 running toward its far end) — Its length recedes obliquely from the foreground edges toward the upper-center distance; used as Maintains a narrow sightline to the runner through staggered shoulders and uneven gaps.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Keep natural daylight and muted saturation continuous through the montage, with sufficient tonal separation to read the small runner without spotlighting him.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The sunlit market remains unchanged immediately after the fight. 안규홍: He continues fleeing with his carrying frame, worn straw shoes and injured mouth.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 안규홍 (한국인 남성, 청년 나이대, 검은 머리, 단단히 묶은 머리) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.",
   "B": "Create ONE FINAL photorealistic live-action film still of the moment below — early 20th-century Korea, Jeolla region; all people are Korean unless explicitly stated otherwise. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 북적이는 사람들 사이로 장터 골목 끝을 향해 뛰어가는 도중, 한쪽 다리를 길게 뻗고 다른 쪽 발은 지면에서 막 떨어진 자세의 안규홍의 작아진 뒷모습.\n\nLOCATION (lock): The far end of a crowded outdoor market alley, along the route away from the fight. The shot takes place here — the attached LOCATION PHOTOGRAPH shows the exact spot.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: wide shot\n- FRAME LAYOUT: far end of the market alley in the upper-center of the frame, background.\n- KEY BACKGROUND ELEMENTS: 장터 골목 (Occupied by people, with 안규홍 running toward its far end) — Its length recedes obliquely from the foreground edges toward the upper-center distance; used as Maintains a narrow sightline to the runner through staggered shoulders and uneven gaps.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Keep natural daylight and muted saturation continuous through the montage, with sufficient tonal separation to read the small runner without spotlighting him.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The sunlit market remains unchanged immediately after the fight. 안규홍: He continues fleeing with his carrying frame, worn straw shoes and injured mouth.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 안규홍 (한국인 남성, 청년 나이대, 검은 머리, 단단히 묶은 머리) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.\n\nCOMPOSITION VARIATION (this candidate only): keep the place, the people and the moment exactly as specified, and keep from the CAMERA & FRAME contract its subject scale, screen placement and key background elements exactly. For this candidate only, its camera angle and height are deliberately left open: choose a camera position distinctly different from the obvious one, as a film director picking a second setup on the same blocking. Stay true to any explicit framing the SHOT TEXT itself dictates."
  },
  "roll_refs": {
   "A": [
    {
     "label": "SHOT BACKGROUND",
     "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/db00517b-165b-4005-a287-f28936b59b95/images/272bd7af-9294-4490-98c6-88a40e9deacc/scene/recipe/S6sh4__bgfirst_bg.png",
     "asset_id": "8698e3a6-8bb7-42c0-a471-c2dca29f0fb4",
     "role": "bgfirst_bg"
    },
    {
     "label": "LAYOUT SKETCH (people placement only)",
     "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/db00517b-165b-4005-a287-f28936b59b95/images/272bd7af-9294-4490-98c6-88a40e9deacc/conti/conti_S6sh4.png",
     "asset_id": "87e8fdb7-2c4f-4534-b30e-95287a82fa63",
     "role": "conti_light"
    },
    {
     "label": "CHARACTER REFERENCE — 안규홍: the exact person appearing in this shot; match face, hair and build exactly.",
     "path": "<bytes:1028384>",
     "asset_id": "3b5894db-ce55-4f34-b966-5e725c3aa952",
     "role": "character_ref"
    }
   ],
   "B": [
    {
     "label": "LOCATION PHOTOGRAPH — the exact place of this shot: its architecture, materials, fixed features and lighting mood are spatial truth; stage the moment inside this place. Never copy its camera framing.",
     "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/db00517b-165b-4005-a287-f28936b59b95/images/272bd7af-9294-4490-98c6-88a40e9deacc/scene/recipe/groupbg_market_escape_alley_2bc28c.png",
     "asset_id": "8916c734-358b-4cd6-857e-d37a2057a57a",
     "role": "bgfirst_group_bg"
    },
    {
     "label": "CHARACTER REFERENCE — 안규홍: the exact person appearing in this shot; match face, hair and build exactly.",
     "path": "<bytes:1028384>",
     "asset_id": "3b5894db-ce55-4f34-b966-5e725c3aa952",
     "role": "character_ref"
    }
   ]
  },
  "shot_run_spend_attempt_count": 1,
  "judge_flip_skipped": "judge_fn_owns_order",
  "readings": [
   {
    "label": "A",
    "direction": "안규홍은 골목 끝을 향해 달려가고 있으며, 전경에 배치된 사람들은 골목 안쪽을 바라보고 있습니다.",
    "built_space": "장터 골목의 가판대, 바구니 등 로케이션 레퍼런스의 요소들이 배치되어 있으나, 공간의 입체감이 부족하고 평면적으로 보입니다.",
    "entities": "안규홍(흰 옷, 짚신, 작은 지게, 뒷모습)이 존재하지만, 주변 군중들이 모두 실제 사람이 아닌 2D 애니메이션 스타일의 그림으로 표현되어 규정을 위반했습니다.",
    "hard_violations": [
     "[gemini-pro] 군중이 실제 사람이 아닌 2D 평면 그래픽/일러스트로 렌더링됨 (invented people / graphic overlay)",
     "[gemini-pro] 안규홍의 두 발이 모두 지면에서 상당히 떨어져 있으나 도약의 역학이 없어 허공에 부유하고 있음 (unsupported body)",
     "[gpt-high] 사진처럼 렌더링된 장터 위에 윤곽선이 뚜렷한 삽화 인물들을 붙인 혼합매체 콜라주로 보여, 하나의 실사 촬영 프레임이라는 필수 조건을 위반한다."
    ],
    "physics": "안규홍의 발이 땅에서 완전히 떨어져 있고 뛰는 추진력이 느껴지지 않아 지탱하는 요소 없이 허공에 떠 있습니다. 전경의 인물들 또한 질량이나 무게감이 전혀 없는 평면입니다."
   },
   {
    "label": "B",
    "direction": "안규홍은 화면 중앙의 골목 끝을 향해 달려가고 있으며, 주변의 사람들은 각자 장터에서의 활동을 하며 다양한 방향을 향하고 있습니다.",
    "built_space": "흙바닥, 초가 지붕 가판대, 나무 기둥과 항아리 등 레퍼런스의 장터 구조물이 올바른 질감과 원근감으로 매우 사실적으로 배치되어 있습니다.",
    "entities": "안규홍(흰 옷, 짚신, 상투를 튼 뒷모습, 나무로 된 지게 형태의 구조물을 짊어짐)과 한국 전통 의상을 입은 사실적인 군중들, 그리고 말 등이 정확하고 생생하게 렌더링되었습니다.",
    "hard_violations": [
     "[gpt-high] 지시와 장소 참조에 없는 말 두 마리를 골목에 추가하여 인물들 사이의 도주 장면에 별도 동물과 장애물을 도입했다."
    ],
    "physics": "안규홍이 달리는 자세에서 왼발 앞꿈치가 지면을 딛고 체중을 지탱하고 있으며, 오른발은 들려 있어 제시된 역동적인 질주 자세가 물리적으로 자연스럽게 구현되었습니다."
   }
  ],
  "cross_model_order": {
   "policy": "cross_model_order_v2_slot0_fwd_slot1_rev_parallel2",
   "slots": [
    {
     "model": "gemini-pro",
     "order": "forward",
     "display_to_canonical": {
      "A": "A",
      "B": "B"
     },
     "raw": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 9,
        "verdict_ko": "포토리얼리즘 조건과 장터 로케이션을 사실적으로 완벽하게 구현했으며, 지게를 매고 달려가는 안규홍의 뒷모습과 북적이는 군중의 구도를 프롬프트에 맞게 정확히 연출했습니다."
       },
       {
        "label": "A",
        "score": 1,
        "verdict_ko": "전경 및 배경의 군중이 모두 2D 일러스트로 렌더링되어 실사 영화 스틸컷이라는 핵심 조건을 심각하게 위배했으며, 달리는 인물이 공중에 부자연스럽게 떠 있습니다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "안규홍은 골목 끝을 향해 달려가고 있으며, 전경에 배치된 사람들은 골목 안쪽을 바라보고 있습니다.",
        "built_space": "장터 골목의 가판대, 바구니 등 로케이션 레퍼런스의 요소들이 배치되어 있으나, 공간의 입체감이 부족하고 평면적으로 보입니다.",
        "entities": "안규홍(흰 옷, 짚신, 작은 지게, 뒷모습)이 존재하지만, 주변 군중들이 모두 실제 사람이 아닌 2D 애니메이션 스타일의 그림으로 표현되어 규정을 위반했습니다.",
        "hard_violations": [
         "군중이 실제 사람이 아닌 2D 평면 그래픽/일러스트로 렌더링됨 (invented people / graphic overlay)",
         "안규홍의 두 발이 모두 지면에서 상당히 떨어져 있으나 도약의 역학이 없어 허공에 부유하고 있음 (unsupported body)"
        ],
        "physics": "안규홍의 발이 땅에서 완전히 떨어져 있고 뛰는 추진력이 느껴지지 않아 지탱하는 요소 없이 허공에 떠 있습니다. 전경의 인물들 또한 질량이나 무게감이 전혀 없는 평면입니다."
       },
       {
        "label": "B",
        "direction": "안규홍은 화면 중앙의 골목 끝을 향해 달려가고 있으며, 주변의 사람들은 각자 장터에서의 활동을 하며 다양한 방향을 향하고 있습니다.",
        "built_space": "흙바닥, 초가 지붕 가판대, 나무 기둥과 항아리 등 레퍼런스의 장터 구조물이 올바른 질감과 원근감으로 매우 사실적으로 배치되어 있습니다.",
        "entities": "안규홍(흰 옷, 짚신, 상투를 튼 뒷모습, 나무로 된 지게 형태의 구조물을 짊어짐)과 한국 전통 의상을 입은 사실적인 군중들, 그리고 말 등이 정확하고 생생하게 렌더링되었습니다.",
        "hard_violations": [],
        "physics": "안규홍이 달리는 자세에서 왼발 앞꿈치가 지면을 딛고 체중을 지탱하고 있으며, 오른발은 들려 있어 제시된 역동적인 질주 자세가 물리적으로 자연스럽게 구현되었습니다."
       }
      ],
      "all_candidates_fail": false
     },
     "normalized": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 9,
        "verdict_ko": "포토리얼리즘 조건과 장터 로케이션을 사실적으로 완벽하게 구현했으며, 지게를 매고 달려가는 안규홍의 뒷모습과 북적이는 군중의 구도를 프롬프트에 맞게 정확히 연출했습니다."
       },
       {
        "label": "A",
        "score": 1,
        "verdict_ko": "전경 및 배경의 군중이 모두 2D 일러스트로 렌더링되어 실사 영화 스틸컷이라는 핵심 조건을 심각하게 위배했으며, 달리는 인물이 공중에 부자연스럽게 떠 있습니다."
       }
      ],
      "all_candidates_fail": false,
      "readings": [
       {
        "label": "A",
        "direction": "안규홍은 골목 끝을 향해 달려가고 있으며, 전경에 배치된 사람들은 골목 안쪽을 바라보고 있습니다.",
        "built_space": "장터 골목의 가판대, 바구니 등 로케이션 레퍼런스의 요소들이 배치되어 있으나, 공간의 입체감이 부족하고 평면적으로 보입니다.",
        "entities": "안규홍(흰 옷, 짚신, 작은 지게, 뒷모습)이 존재하지만, 주변 군중들이 모두 실제 사람이 아닌 2D 애니메이션 스타일의 그림으로 표현되어 규정을 위반했습니다.",
        "hard_violations": [
         "군중이 실제 사람이 아닌 2D 평면 그래픽/일러스트로 렌더링됨 (invented people / graphic overlay)",
         "안규홍의 두 발이 모두 지면에서 상당히 떨어져 있으나 도약의 역학이 없어 허공에 부유하고 있음 (unsupported body)"
        ],
        "physics": "안규홍의 발이 땅에서 완전히 떨어져 있고 뛰는 추진력이 느껴지지 않아 지탱하는 요소 없이 허공에 떠 있습니다. 전경의 인물들 또한 질량이나 무게감이 전혀 없는 평면입니다."
       },
       {
        "label": "B",
        "direction": "안규홍은 화면 중앙의 골목 끝을 향해 달려가고 있으며, 주변의 사람들은 각자 장터에서의 활동을 하며 다양한 방향을 향하고 있습니다.",
        "built_space": "흙바닥, 초가 지붕 가판대, 나무 기둥과 항아리 등 레퍼런스의 장터 구조물이 올바른 질감과 원근감으로 매우 사실적으로 배치되어 있습니다.",
        "entities": "안규홍(흰 옷, 짚신, 상투를 튼 뒷모습, 나무로 된 지게 형태의 구조물을 짊어짐)과 한국 전통 의상을 입은 사실적인 군중들, 그리고 말 등이 정확하고 생생하게 렌더링되었습니다.",
        "hard_violations": [],
        "physics": "안규홍이 달리는 자세에서 왼발 앞꿈치가 지면을 딛고 체중을 지탱하고 있으며, 오른발은 들려 있어 제시된 역동적인 질주 자세가 물리적으로 자연스럽게 구현되었습니다."
       }
      ]
     },
     "ok": true
    },
    {
     "model": "gpt-high",
     "order": "reverse",
     "display_to_canonical": {
      "A": "B",
      "B": "A"
     },
     "raw": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 4,
        "verdict_ko": "실사 인물과 장터 재질, 도주 방향은 맞지만, 안규홍이 너무 크고 가까우며 지시되지 않은 말 두 마리가 추가됐다."
       },
       {
        "label": "B",
        "score": 2,
        "verdict_ko": "상단 중앙의 작은 도주자 배치는 더 정확하지만, 사진 배경에 삽화 인물을 붙인 표현과 현대적인 군중 복장이 실사 시대극 조건을 크게 위반한다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "중앙의 안규홍은 등을 보이고 화면 상단 중앙의 골목 끝을 향한다. 얼굴과 눈은 보이지 않아 시선 자체는 확인할 수 없다. 주변 사람들은 좌판의 물건이나 서로를 바라보며, 달리는 사람을 일제히 쳐다보지는 않는다. 말 두 마리 중 가까운 말은 왼쪽 좌판 쪽으로 머리를 향하고, 먼 말은 골목 안쪽을 향한다.",
        "built_space": "오른쪽에는 기와를 얹은 돌담 한 줄과 목재 기둥으로 받친 짚 차양, 항아리 진열대가 있고, 왼쪽에는 큰 발 차양과 흰 천막 아래 바구니 좌판들이 이어진다. 좌우 좌판은 상단 중앙으로 수렴하여 참조 장소의 기본 구조를 유지한다. 오른쪽 앞에는 세로로 천을 늘어뜨린 진열대가 두드러진다. 군중은 양쪽 좌판과 골목 안에 분산되어 있지만 중앙 통로가 비교적 넓게 열려 있다. 안규홍은 화면 높이의 약 40%를 차지하여, 먼 배경의 작아진 뒷모습보다는 중경의 주인공으로 읽힌다. 반사면이나 불가능한 반사는 보이지 않는다.",
        "entities": "안규홍으로 읽히는 남성 한 명은 검은 묶음머리, 때 묻은 밝은 저고리와 바지, 짚신, 목제 지게를 갖췄다. 뒷모습이므로 참조 얼굴과 다친 입은 확인할 수 없으며 이는 구도상 자연스럽다. 군중은 한국의 옛 장터 복장을 한 남녀로 표현되어 있다. 바구니, 항아리, 목재 좌판과 천막은 실물 재질로 보인다. 다만 장면과 참조에 없는 말 두 마리가 추가되어 있다. 읽을 수 있는 글자는 없다.",
        "hard_violations": [
         "지시와 장소 참조에 없는 말 두 마리를 골목에 추가하여 인물들 사이의 도주 장면에 별도 동물과 장애물을 도입했다."
        ],
        "physics": "안규홍의 두 발은 지면에서 떨어져 보이지만, 몸을 앞으로 보내고 한쪽 발바닥을 뒤로 드러낸 짧은 달리기 체공으로 해석할 수 있다. 아래에 착지할 흙길이 있으므로 근거 없는 공중 부양은 아니다. 다만 두 다리의 벌어짐이 작아 한 다리를 길게 뻗는 지정 순간은 약하다. 지게는 등과 어깨에 밀착되어 몸에 지지되고, 바구니를 든 여성은 손으로 손잡이를 잡는다. 나머지 사람과 말은 지면에 서 있으며 진열품은 좌판에 놓여 있다."
       },
       {
        "label": "B",
        "direction": "상단 중앙의 지게 진 남성은 등을 보이며 골목 끝을 향해 달린다. 중앙 중경의 밝은 옷 남성도 같은 방향으로 걷고, 좌우 군중 중 여러 명은 중앙 통로나 도주자 쪽으로 고개를 돌린다. 오른쪽 좌판의 두 사람은 상품과 서로의 손 쪽을 향한다. 주인공의 눈은 보이지 않는다.",
        "built_space": "오른쪽의 기와 돌담 한 줄, 짚 차양과 목재 지주, 왼쪽의 큰 발과 흰 천막, 양쪽 좌판이 참조 장소에 가깝게 배치됐다. 전경 바닥의 목재 조각과 오른쪽 도기 파편도 보인다. 골목은 상단 중앙으로 좁아지고 도주자는 화면 높이의 약 20%로 작게 배치되어 A보다 지정 규모에 가깝다. 다만 군중이 양옆에 비교적 정렬되어, 불규칙한 어깨 사이로 겨우 보이는 북적임보다는 도주자를 위해 열린 통로처럼 보인다. 불가능한 반사는 없다.",
        "entities": "먼 도주자는 검은 묶음머리, 밝은 전통 상하의, 짚신과 지게를 갖춘 남성으로 표현됐다. 얼굴과 입의 부상은 뒷모습과 거리 때문에 확인할 수 없다. 중경에도 비슷한 밝은 전통복과 짚신의 남성이 있지만 별개 군중인지 주인공 중복인지 단정하기 어렵다. 군중에는 야구모자처럼 보이는 모자 두 개, 반소매 티셔츠와 현대적인 치마·신발이 보인다. 특히 사람들은 피부와 옷이 검은 외곽선 및 평면 명암으로 그려져 실제 배우로 보이지 않는다. 배경의 돌담과 좌판은 사진처럼 보여 인물 표현과 분리된다. 읽을 수 있는 글자는 없다.",
        "hard_violations": [
         "사진처럼 렌더링된 장터 위에 윤곽선이 뚜렷한 삽화 인물들을 붙인 혼합매체 콜라주로 보여, 하나의 실사 촬영 프레임이라는 필수 조건을 위반한다."
        ],
        "physics": "먼 도주자는 한 다리를 아래로 길게 내리고 다른 무릎을 굽힌 달리기 체공 자세이며, 바로 아래 흙길이 착지면이다. 따라서 두 발이 떠 있다는 이유만으로 지지 없는 부양으로 볼 수는 없다. 지게는 어깨끈과 등으로 지지된다. 중앙 중경 남성은 한 발을 땅에 두고 다른 발의 뒤꿈치와 발바닥을 들어 보행한다. 나머지 군중은 지면에 서고 좌판의 상품은 상판에 놓여 있어, 명백히 지지 없이 떠 있는 물체는 보이지 않는다."
       }
      ],
      "all_candidates_fail": true
     },
     "normalized": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 4,
        "verdict_ko": "실사 인물과 장터 재질, 도주 방향은 맞지만, 안규홍이 너무 크고 가까우며 지시되지 않은 말 두 마리가 추가됐다."
       },
       {
        "label": "A",
        "score": 2,
        "verdict_ko": "상단 중앙의 작은 도주자 배치는 더 정확하지만, 사진 배경에 삽화 인물을 붙인 표현과 현대적인 군중 복장이 실사 시대극 조건을 크게 위반한다."
       }
      ],
      "all_candidates_fail": true,
      "readings": [
       {
        "label": "B",
        "direction": "중앙의 안규홍은 등을 보이고 화면 상단 중앙의 골목 끝을 향한다. 얼굴과 눈은 보이지 않아 시선 자체는 확인할 수 없다. 주변 사람들은 좌판의 물건이나 서로를 바라보며, 달리는 사람을 일제히 쳐다보지는 않는다. 말 두 마리 중 가까운 말은 왼쪽 좌판 쪽으로 머리를 향하고, 먼 말은 골목 안쪽을 향한다.",
        "built_space": "오른쪽에는 기와를 얹은 돌담 한 줄과 목재 기둥으로 받친 짚 차양, 항아리 진열대가 있고, 왼쪽에는 큰 발 차양과 흰 천막 아래 바구니 좌판들이 이어진다. 좌우 좌판은 상단 중앙으로 수렴하여 참조 장소의 기본 구조를 유지한다. 오른쪽 앞에는 세로로 천을 늘어뜨린 진열대가 두드러진다. 군중은 양쪽 좌판과 골목 안에 분산되어 있지만 중앙 통로가 비교적 넓게 열려 있다. 안규홍은 화면 높이의 약 40%를 차지하여, 먼 배경의 작아진 뒷모습보다는 중경의 주인공으로 읽힌다. 반사면이나 불가능한 반사는 보이지 않는다.",
        "entities": "안규홍으로 읽히는 남성 한 명은 검은 묶음머리, 때 묻은 밝은 저고리와 바지, 짚신, 목제 지게를 갖췄다. 뒷모습이므로 참조 얼굴과 다친 입은 확인할 수 없으며 이는 구도상 자연스럽다. 군중은 한국의 옛 장터 복장을 한 남녀로 표현되어 있다. 바구니, 항아리, 목재 좌판과 천막은 실물 재질로 보인다. 다만 장면과 참조에 없는 말 두 마리가 추가되어 있다. 읽을 수 있는 글자는 없다.",
        "hard_violations": [
         "지시와 장소 참조에 없는 말 두 마리를 골목에 추가하여 인물들 사이의 도주 장면에 별도 동물과 장애물을 도입했다."
        ],
        "physics": "안규홍의 두 발은 지면에서 떨어져 보이지만, 몸을 앞으로 보내고 한쪽 발바닥을 뒤로 드러낸 짧은 달리기 체공으로 해석할 수 있다. 아래에 착지할 흙길이 있으므로 근거 없는 공중 부양은 아니다. 다만 두 다리의 벌어짐이 작아 한 다리를 길게 뻗는 지정 순간은 약하다. 지게는 등과 어깨에 밀착되어 몸에 지지되고, 바구니를 든 여성은 손으로 손잡이를 잡는다. 나머지 사람과 말은 지면에 서 있으며 진열품은 좌판에 놓여 있다."
       },
       {
        "label": "A",
        "direction": "상단 중앙의 지게 진 남성은 등을 보이며 골목 끝을 향해 달린다. 중앙 중경의 밝은 옷 남성도 같은 방향으로 걷고, 좌우 군중 중 여러 명은 중앙 통로나 도주자 쪽으로 고개를 돌린다. 오른쪽 좌판의 두 사람은 상품과 서로의 손 쪽을 향한다. 주인공의 눈은 보이지 않는다.",
        "built_space": "오른쪽의 기와 돌담 한 줄, 짚 차양과 목재 지주, 왼쪽의 큰 발과 흰 천막, 양쪽 좌판이 참조 장소에 가깝게 배치됐다. 전경 바닥의 목재 조각과 오른쪽 도기 파편도 보인다. 골목은 상단 중앙으로 좁아지고 도주자는 화면 높이의 약 20%로 작게 배치되어 A보다 지정 규모에 가깝다. 다만 군중이 양옆에 비교적 정렬되어, 불규칙한 어깨 사이로 겨우 보이는 북적임보다는 도주자를 위해 열린 통로처럼 보인다. 불가능한 반사는 없다.",
        "entities": "먼 도주자는 검은 묶음머리, 밝은 전통 상하의, 짚신과 지게를 갖춘 남성으로 표현됐다. 얼굴과 입의 부상은 뒷모습과 거리 때문에 확인할 수 없다. 중경에도 비슷한 밝은 전통복과 짚신의 남성이 있지만 별개 군중인지 주인공 중복인지 단정하기 어렵다. 군중에는 야구모자처럼 보이는 모자 두 개, 반소매 티셔츠와 현대적인 치마·신발이 보인다. 특히 사람들은 피부와 옷이 검은 외곽선 및 평면 명암으로 그려져 실제 배우로 보이지 않는다. 배경의 돌담과 좌판은 사진처럼 보여 인물 표현과 분리된다. 읽을 수 있는 글자는 없다.",
        "hard_violations": [
         "사진처럼 렌더링된 장터 위에 윤곽선이 뚜렷한 삽화 인물들을 붙인 혼합매체 콜라주로 보여, 하나의 실사 촬영 프레임이라는 필수 조건을 위반한다."
        ],
        "physics": "먼 도주자는 한 다리를 아래로 길게 내리고 다른 무릎을 굽힌 달리기 체공 자세이며, 바로 아래 흙길이 착지면이다. 따라서 두 발이 떠 있다는 이유만으로 지지 없는 부양으로 볼 수는 없다. 지게는 어깨끈과 등으로 지지된다. 중앙 중경 남성은 한 발을 땅에 두고 다른 발의 뒤꿈치와 발바닥을 들어 보행한다. 나머지 군중은 지면에 서고 좌판의 상품은 상판에 놓여 있어, 명백히 지지 없이 떠 있는 물체는 보이지 않는다."
       }
      ]
     },
     "ok": true
    }
   ],
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "slot_winner_match": true,
   "slot_winner": {
    "gemini-pro": "B",
    "gpt-high": "B"
   },
   "route": "cross_slot_agree"
  },
  "dual": {
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "normalized": {
    "A": 0.611,
    "B": 2.0
   },
   "adjusted": {
    "A": 0.361,
    "B": 1.75
   },
   "violations": {
    "A": [
     "[gemini-pro] 군중이 실제 사람이 아닌 2D 평면 그래픽/일러스트로 렌더링됨 (invented people / graphic overlay)",
     "[gemini-pro] 안규홍의 두 발이 모두 지면에서 상당히 떨어져 있으나 도약의 역학이 없어 허공에 부유하고 있음 (unsupported body)",
     "[gpt-high] 사진처럼 렌더링된 장터 위에 윤곽선이 뚜렷한 삽화 인물들을 붙인 혼합매체 콜라주로 보여, 하나의 실사 촬영 프레임이라는 필수 조건을 위반한다."
    ],
    "B": [
     "[gpt-high] 지시와 장소 참조에 없는 말 두 마리를 골목에 추가하여 인물들 사이의 도주 장면에 별도 동물과 장애물을 도입했다."
    ]
   },
   "per_model_winner": {
    "gemini-pro": "B",
    "gpt-high": "B"
   },
   "agreed": true
  },
  "totals": {
   "B": 1750,
   "A": 361
  },
  "selected": "B",
  "ranking": [
   "B",
   "A"
  ],
  "verdicts": [
   {
    "label": "B",
    "score": 1750,
    "verdict_ko": "포토리얼리즘 조건과 장터 로케이션을 사실적으로 완벽하게 구현했으며, 지게를 매고 달려가는 안규홍의 뒷모습과 북적이는 군중의 구도를 프롬프트에 맞게 정확히 연출했습니다.  ★위반: [gpt-high] 지시와 장소 참조에 없는 말 두 마리를 골목에 추가하여 인물들 사이의 도주 장면에 별도 동물과 장애물을 도입했다."
   },
   {
    "label": "A",
    "score": 361,
    "verdict_ko": "전경 및 배경의 군중이 모두 2D 일러스트로 렌더링되어 실사 영화 스틸컷이라는 핵심 조건을 심각하게 위배했으며, 달리는 인물이 공중에 부자연스럽게 떠 있습니다.  ★위반: [gemini-pro] 군중이 실제 사람이 아닌 2D 평면 그래픽/일러스트로 렌더링됨 (invented people / graphic overlay) / [gemini-pro] 안규홍의 두 발이 모두 지면에서 상당히 떨어져 있으나 도약의 역학이 없어 허공에 부유하고 있음 (unsupported body) / [gpt-high] 사진처럼 렌더링된 장터 위에 윤곽선이 뚜렷한 삽화 인물들을 붙인 혼합매체 콜라주로 보여, 하나의 실사 촬영 프레임이라는 필수 조건을 위반한다."
   }
  ],
  "refs": [
   {
    "label": "LOCATION PHOTOGRAPH — the exact place of this shot: its architecture, materials, fixed features and lighting mood are spatial truth; stage the moment inside this place. Never copy its camera framing.",
    "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/db00517b-165b-4005-a287-f28936b59b95/images/272bd7af-9294-4490-98c6-88a40e9deacc/scene/recipe/groupbg_market_escape_alley_2bc28c.png",
    "asset_id": "8916c734-358b-4cd6-857e-d37a2057a57a",
    "role": "bgfirst_group_bg"
   },
   {
    "label": "CHARACTER REFERENCE — 안규홍: the exact person appearing in this shot; match face, hair and build exactly.",
    "path": "<bytes:1028384>",
    "asset_id": "3b5894db-ce55-4f34-b966-5e725c3aa952",
    "role": "character_ref"
   }
  ],
  "critique_skipped": true,
  "shot_run_produced": false,
  "shot_run_uid": "06aa1406-fba2-79c7-8d99-fc792c80eb93",
  "bgfirst": {
   "bg_path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/db00517b-165b-4005-a287-f28936b59b95/images/272bd7af-9294-4490-98c6-88a40e9deacc/scene/recipe/S6sh4__bgfirst_bg.png",
   "bg_asset_id": "8698e3a6-8bb7-42c0-a471-c2dca29f0fb4",
   "bg_record_key": "S6sh4::bgfirst_bg",
   "chain_winner": false,
   "authority": "groupbg",
   "group_key": "market_escape_alley",
   "groupbg_asset_id": "8916c734-358b-4cd6-857e-d37a2057a57a"
  },
  "ref_mode": "그룹 배경+엔티티 (2택1: 무콘티 승)",
  "share_plan": {
   "ref_plan": "background"
  }
 },
 "S6sh4::cine": {
  "applied": true,
  "attempted_at": "2026-09-09T01:14:28.579441+00:00",
  "fingerprint": "d346c1708d52ed7f43a6a8bf6db9c3d4423ac061601d97e9d1468e56b529fca2",
  "fingerprint_version": 2,
  "provider": "mai",
  "endpoint": "openrouter/chat-completions",
  "model": "microsoft/mai-image-2.6",
  "pack": "24.202608252115",
  "source_file": "S6sh4_sel.png",
  "source_sha256": "9cf4c436e33343d2460f360f910e6d75c34d334986041d68428b594e46a67e56",
  "file": "S6sh4_cine.png",
  "staged_sha256": "ff19fbd9f31ad3102bc97df0c51961527ee1481262c052d5ce964f621ee4c852",
  "latency_ms": 27124
 },
 "S7sh4::signage": {
  "fp": "511968c9a4b07c7d",
  "inscriptions": [],
  "cues": [],
  "dropped": []
 },
 "S7sh4::bgfirst_bg": {
  "input_fingerprint": "75be6d7468ec688d",
  "prompt": "Create the EMPTY BACKGROUND PLATE for one film shot — NO PEOPLE, no figures, no body parts, no sketch lines, no arrows anywhere.\n\"Empty\" means no people only: KEEP the location's inherent occupants and stock that define the place — animals in an animal shelter, pen or farm, goods in a market, moored boats in a harbour — unless the shot text explicitly removes them.\nThe FIRST attached image is a thin-line storyboard sketch: use ONLY its camera angle, horizon, perspective and the placement/size of buildings and set masses — ignore the sketched people and arrows entirely. The SECOND attached image (LOCATION PHOTOGRAPH) is the real place: take its architecture, materials, signage and fixed features, and RE-PROJECT them into the sketch's camera. If the photograph's camera differs from the sketch's, the sketch's camera wins.\nHUMAN-SCALE CALIBRATION: derive every structure's true size from human-scale elements — a door ≈ 2m, a window ≈ 1–1.5m wide, one storey ≈ 2.5–3m; never inflate a small structure or shrink a large one.\n\nSHOT TEXT this background must serve (Korean): 조선인 순검들 뒤에 서서 턱을 치켜들고 안규홍을 손가락으로 가리키는 배상도의 의기양양한 상체.\n\nLOCATION (lock): Inside the main gate of a traditional estate's outdoor courtyard, behind the newly arrived constables in late-afternoon light.\n\nTIME OF DAY (lock): late afternoon.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: medium shot\n- KEY BACKGROUND ELEMENTS: 대문 (The entry passage is clear as the officers come into the courtyard) — Seen obliquely from its courtyard side behind the entering group; used as Anchors 배상도 behind the officers while their differing shoulder positions leave his face and pointing hand visible.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Natural late-afternoon light with restrained saturation and moderate contrast preserves the accusation's self-satisfaction without imposing a theatrical color cast.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nRender ONE photorealistic empty location photograph, 16:9, that this shot can be staged inside later. No readable writing anywhere: surfaces that would carry writing may be present, but stage any wording out of legibility — an oblique angle, distance, shallow focus. No captions, watermarks or overlay text.",
  "effective_prompt": "Create the EMPTY BACKGROUND PLATE for one film shot — NO PEOPLE, no figures, no body parts, no sketch lines, no arrows anywhere.\n\"Empty\" means no people only: KEEP the location's inherent occupants and stock that define the place — animals in an animal shelter, pen or farm, goods in a market, moored boats in a harbour — unless the shot text explicitly removes them.\nThe FIRST attached image is a thin-line storyboard sketch: use ONLY its camera angle, horizon, perspective and the placement/size of buildings and set masses — ignore the sketched people and arrows entirely. The SECOND attached image (LOCATION PHOTOGRAPH) is the real place: take its architecture, materials, signage and fixed features, and RE-PROJECT them into the sketch's camera. If the photograph's camera differs from the sketch's, the sketch's camera wins.\nHUMAN-SCALE CALIBRATION: derive every structure's true size from human-scale elements — a door ≈ 2m, a window ≈ 1–1.5m wide, one storey ≈ 2.5–3m; never inflate a small structure or shrink a large one.\n\nSHOT TEXT this background must serve (Korean): 조선인 순검들 뒤에 서서 턱을 치켜들고 안규홍을 손가락으로 가리키는 배상도의 의기양양한 상체.\n\nLOCATION (lock): Inside the main gate of a traditional estate's outdoor courtyard, behind the newly arrived constables in late-afternoon light.\n\nTIME OF DAY (lock): late afternoon.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: medium shot\n- KEY BACKGROUND ELEMENTS: 대문 (The entry passage is clear as the officers come into the courtyard) — Seen obliquely from its courtyard side behind the entering group; used as Anchors 배상도 behind the officers while their differing shoulder positions leave his face and pointing hand visible.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Natural late-afternoon light with restrained saturation and moderate contrast preserves the accusation's self-satisfaction without imposing a theatrical color cast.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nRender ONE photorealistic empty location photograph, 16:9, that this shot can be staged inside later. No readable writing anywhere: surfaces that would carry writing may be present, but stage any wording out of legibility — an oblique angle, distance, shallow focus. No captions, watermarks or overlay text.",
  "bg_path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/db00517b-165b-4005-a287-f28936b59b95/images/272bd7af-9294-4490-98c6-88a40e9deacc/scene/recipe/S7sh4__bgfirst_bg.png",
  "asset_id": "bb4dbf43-e434-4ef1-82d6-a9e5f3e104ab",
  "input_asset_ids": [
   "84d0c93b-13f3-4025-8095-d54b858c4cd0",
   "64556ed8-9e6e-4367-ade4-e6cfd6cae678"
  ]
 },
 "S7sh4": {
  "input_fingerprint": "5f07938fb53924bb",
  "prompt": "Create ONE FINAL photorealistic live-action film still of the moment below — early 20th-century Korea, Jeolla region; all people are Korean unless explicitly stated otherwise. TIME OF DAY (lock): late afternoon.\n\nSHOT TEXT (authoritative, Korean): 조선인 순검들 뒤에 서서 턱을 치켜들고 안규홍을 손가락으로 가리키는 배상도의 의기양양한 상체.\n\nLOCATION (lock): Inside the main gate of a traditional estate's outdoor courtyard, behind the newly arrived constables in late-afternoon light. The shot takes place here — the attached LOCATION PHOTOGRAPH shows the exact spot.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: medium shot\n- KEY BACKGROUND ELEMENTS: 대문 (The entry passage is clear as the officers come into the courtyard) — Seen obliquely from its courtyard side behind the entering group; used as Anchors 배상도 behind the officers while their differing shoulder positions leave his face and pointing hand visible.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Natural late-afternoon light with restrained saturation and moderate contrast preserves the accusation's self-satisfaction without imposing a theatrical color cast.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The carrying frame stands in the middle of the late-afternoon yard. 배상도: He stands behind the arriving police.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 배상도 (한국인 남성, 성인 나이대, 자연스러운 검은 머리); 조선인 순검(안규홍 체포·옥사 담당) (한국인 남성, 성인 나이대, 자연스러운 검은 머리) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.",
  "roll_prompts": {
   "A": "Stage the shot. The FIRST attached image (SHOT BACKGROUND) is the finished empty background of this shot — keep it EXACTLY: its camera, perspective, architecture, lighting and every fixed feature stay untouched. The SECOND attached image (LAYOUT SKETCH) tells you ONLY where the people go: each sketched person's position, screen size, pose and the gaze/motion arrows. Ignore the sketch's background lines. The CHARACTER REFERENCE photographs show the real people.\nPlace the real people into the background at exactly the sketched positions, sizes and poses, following the arrow directions. No sketch lines or arrows may remain.\n\nCreate ONE FINAL photorealistic live-action film still of the moment below — early 20th-century Korea, Jeolla region; all people are Korean unless explicitly stated otherwise. TIME OF DAY (lock): late afternoon.\n\nSHOT TEXT (authoritative, Korean): 조선인 순검들 뒤에 서서 턱을 치켜들고 안규홍을 손가락으로 가리키는 배상도의 의기양양한 상체.\n\nLOCATION (lock): Inside the main gate of a traditional estate's outdoor courtyard, behind the newly arrived constables in late-afternoon light. The shot takes place here — the FIRST attached image (SHOT BACKGROUND) is this exact place, already built: its ground, structures, horizon, materials and lighting are the finished truth of this location and must not be redesigned or replaced. No location photograph is attached — read the place from that image alone, and add no scenery, structure, vehicle or fixture that it does not already show. This lock governs the place only; the figures in the shot follow the staging and pose instructions.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: medium shot\n- KEY BACKGROUND ELEMENTS: 대문 (The entry passage is clear as the officers come into the courtyard) — Seen obliquely from its courtyard side behind the entering group; used as Anchors 배상도 behind the officers while their differing shoulder positions leave his face and pointing hand visible.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Natural late-afternoon light with restrained saturation and moderate contrast preserves the accusation's self-satisfaction without imposing a theatrical color cast.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The carrying frame stands in the middle of the late-afternoon yard. 배상도: He stands behind the arriving police.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 배상도 (한국인 남성, 성인 나이대, 자연스러운 검은 머리); 조선인 순검(안규홍 체포·옥사 담당) (한국인 남성, 성인 나이대, 자연스러운 검은 머리) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.",
   "B": "Create ONE FINAL photorealistic live-action film still of the moment below — early 20th-century Korea, Jeolla region; all people are Korean unless explicitly stated otherwise. TIME OF DAY (lock): late afternoon.\n\nSHOT TEXT (authoritative, Korean): 조선인 순검들 뒤에 서서 턱을 치켜들고 안규홍을 손가락으로 가리키는 배상도의 의기양양한 상체.\n\nLOCATION (lock): Inside the main gate of a traditional estate's outdoor courtyard, behind the newly arrived constables in late-afternoon light. The shot takes place here — the attached LOCATION PHOTOGRAPH shows the exact spot.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: medium shot\n- KEY BACKGROUND ELEMENTS: 대문 (The entry passage is clear as the officers come into the courtyard) — Seen obliquely from its courtyard side behind the entering group; used as Anchors 배상도 behind the officers while their differing shoulder positions leave his face and pointing hand visible.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Natural late-afternoon light with restrained saturation and moderate contrast preserves the accusation's self-satisfaction without imposing a theatrical color cast.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The carrying frame stands in the middle of the late-afternoon yard. 배상도: He stands behind the arriving police.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 배상도 (한국인 남성, 성인 나이대, 자연스러운 검은 머리); 조선인 순검(안규홍 체포·옥사 담당) (한국인 남성, 성인 나이대, 자연스러운 검은 머리) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.\n\nCOMPOSITION VARIATION (this candidate only): keep the place, the people and the moment exactly as specified, and keep from the CAMERA & FRAME contract its subject scale, screen placement and key background elements exactly. For this candidate only, its camera angle and height are deliberately left open: choose a camera position distinctly different from the obvious one, as a film director picking a second setup on the same blocking. Stay true to any explicit framing the SHOT TEXT itself dictates."
  },
  "roll_refs": {
   "A": [
    {
     "label": "SHOT BACKGROUND",
     "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/db00517b-165b-4005-a287-f28936b59b95/images/272bd7af-9294-4490-98c6-88a40e9deacc/scene/recipe/S7sh4__bgfirst_bg.png",
     "asset_id": "bb4dbf43-e434-4ef1-82d6-a9e5f3e104ab",
     "role": "bgfirst_bg"
    },
    {
     "label": "LAYOUT SKETCH (people placement only)",
     "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/db00517b-165b-4005-a287-f28936b59b95/images/272bd7af-9294-4490-98c6-88a40e9deacc/conti/conti_S7sh4.png",
     "asset_id": "84d0c93b-13f3-4025-8095-d54b858c4cd0",
     "role": "conti_light"
    },
    {
     "label": "CHARACTER REFERENCE — 배상도: the exact person appearing in this shot; match face, hair and build exactly.",
     "path": "<bytes:851718>",
     "asset_id": "034f5325-975e-4474-854b-e48359710bf8",
     "role": "character_ref"
    },
    {
     "label": "CHARACTER REFERENCE — 조선인 순검(안규홍 체포·옥사 담당): the exact person appearing in this shot; match face, hair and build exactly.",
     "path": "<bytes:1264194>",
     "asset_id": "0b736389-3dee-4967-9bc7-ff4ea80d9b44",
     "role": "character_ref"
    }
   ],
   "B": [
    {
     "label": "LOCATION PHOTOGRAPH — the exact place of this shot: its architecture, materials, fixed features and lighting mood are spatial truth; stage the moment inside this place. Never copy its camera framing.",
     "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/db00517b-165b-4005-a287-f28936b59b95/episodes/272bd7af-9294-4490-98c6-88a40e9deacc/images/background_chain/L08B01.png",
     "asset_id": "64556ed8-9e6e-4367-ade4-e6cfd6cae678",
     "role": "location_plate"
    },
    {
     "label": "CHARACTER REFERENCE — 배상도: the exact person appearing in this shot; match face, hair and build exactly.",
     "path": "<bytes:851718>",
     "asset_id": "034f5325-975e-4474-854b-e48359710bf8",
     "role": "character_ref"
    },
    {
     "label": "CHARACTER REFERENCE — 조선인 순검(안규홍 체포·옥사 담당): the exact person appearing in this shot; match face, hair and build exactly.",
     "path": "<bytes:1264194>",
     "asset_id": "0b736389-3dee-4967-9bc7-ff4ea80d9b44",
     "role": "character_ref"
    }
   ]
  },
  "shot_run_spend_attempt_count": 1,
  "judge_flip_skipped": "judge_fn_owns_order",
  "readings": [
   {
    "label": "A",
    "direction": "배상도는 프레임 왼쪽 바깥에 있을 안규홍을 향해 오른손 검지를 뻗어 정확히 가리키고 있으며, 턱을 치켜들고 거만한 시선으로 그 방향을 응시한다. 앞쪽의 순검 두 명도 가리키는 방향인 왼쪽 밖을 바라보고 있다.",
    "built_space": "기준 사진과 동일한 마당 내부로, 왼쪽에는 마루가 있는 기와집이 위치하고 우측 배경에는 기와지붕이 있는 대문이 비스듬히 보인다. 인물들은 마당 한가운데에 올바르게 서 있으며 카메라가 뒤에서 앞을 비추는 올바른 앵글을 유지하고 있다.",
    "entities": "배상도(한국인 남성, 성인)는 얼굴과 체형이 기준과 일치하지만, 기준 복장 중 검은 모자와 조끼가 누락된 흰색 한복 차림이다. 안규홍 체포를 담당한 조선인 순검(한국인 남성, 성인) 두 명은 검은색 제복과 모자를 쓴 뒷모습으로 등장하며, 지시되지 않은 다른 인물은 나타나지 않았다.",
    "hard_violations": [
     "[gpt-high] 순검들이 배상도를 향해 마주 서도록 배치되어, 배상도가 도착하는 순검들 뒤에 서 있다는 필수 동선 관계를 위반한다."
    ],
    "physics": "배상도와 순검들 모두 두 발로 마당의 흙바닥을 단단히 디디고 서 있다. 배상도가 들어 올린 오른팔은 어깨에 자연스럽게 연결되어 지탱되며 물리적으로 불가능한 자세이나 공중에 뜬 사물은 없다."
   },
   {
    "label": "B",
    "direction": "배상도는 화면 왼쪽 전경에 서 있는 망건 쓴 남성을 향해 오른손 검지를 뻗어 가리키고 있으며, 턱을 치켜든 채 그 인물을 내려다보고 있다. 뒤편의 순검들은 정면과 측면을 각각 응시하고 있다.",
    "built_space": "대문 바로 안쪽에서 바깥을 향하거나 그 반대를 바라보는 구도로, 양옆으로 나무 문짝 2개가 열려 있다. 배상도는 대문 한가운데 서 있어, 마당 한가운데로 들어서는 인물들의 거리를 묘사해야 하는 원본 스테이징과 구조적으로 맞지 않는다.",
    "entities": "배상도(한국인 남성, 성인)는 기준 사진과 일치하는 검은 모자와 조끼, 흰 한복을 착용했다. 조선인 순검(한국인 남성, 성인) 두 명이 제복 차림으로 뒤편에 등장한다. 그러나 프롬프트에 명시된 허용 인물 외에 망건을 쓴 신원 불명의 한국인 남성이 왼쪽 전경에 등장하여 지시를 위반했다.",
    "hard_violations": [
     "[gemini-pro] invented people (프롬프트의 PEOPLE 목록에 없는 망건 쓴 남성을 화면 왼쪽 전경에 임의로 추가함)",
     "[gemini-pro] person placed where the staging does not put them (순검들 뒤에 서야 하는 배상도를 순검들보다 앞쪽인 전경에 배치함)",
     "[gpt-high] 화면 왼쪽 전경에 허용된 배상도·순검 이외의 지목 대상 인물을 추가했다.",
     "[gpt-high] 배상도를 순검들 뒤가 아니라 두 순검보다 앞에 배치했다."
    ],
    "physics": "배상도와 순검들, 전경의 남성 모두 두 발로 지면을 딛고 서 있다. 배상도의 오른팔은 어깨와 몸통에 의해 지탱되나 손가락 형태가 약간 왜곡되어 있다. 지지대 없이 공중에 떠 있는 물체는 없다."
   }
  ],
  "cross_model_order": {
   "policy": "cross_model_order_v2_slot0_fwd_slot1_rev_parallel2",
   "slots": [
    {
     "model": "gemini-pro",
     "order": "forward",
     "display_to_canonical": {
      "A": "A",
      "B": "B"
     },
     "raw": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 8,
        "verdict_ko": "순검들 뒤에서 안규홍을 가리키는 프레임 구성과 지형 배치를 정확히 구현했으나, 배상도의 기준 복장인 모자와 조끼가 누락되었습니다."
       },
       {
        "label": "B",
        "score": 2,
        "verdict_ko": "배상도의 복장은 완벽하게 재현했으나, 순검들 앞에 서는 스테이징 오류와 지시되지 않은 인물을 추가하는 치명적인 위반을 범했습니다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "배상도는 프레임 왼쪽 바깥에 있을 안규홍을 향해 오른손 검지를 뻗어 정확히 가리키고 있으며, 턱을 치켜들고 거만한 시선으로 그 방향을 응시한다. 앞쪽의 순검 두 명도 가리키는 방향인 왼쪽 밖을 바라보고 있다.",
        "built_space": "기준 사진과 동일한 마당 내부로, 왼쪽에는 마루가 있는 기와집이 위치하고 우측 배경에는 기와지붕이 있는 대문이 비스듬히 보인다. 인물들은 마당 한가운데에 올바르게 서 있으며 카메라가 뒤에서 앞을 비추는 올바른 앵글을 유지하고 있다.",
        "entities": "배상도(한국인 남성, 성인)는 얼굴과 체형이 기준과 일치하지만, 기준 복장 중 검은 모자와 조끼가 누락된 흰색 한복 차림이다. 안규홍 체포를 담당한 조선인 순검(한국인 남성, 성인) 두 명은 검은색 제복과 모자를 쓴 뒷모습으로 등장하며, 지시되지 않은 다른 인물은 나타나지 않았다.",
        "hard_violations": [],
        "physics": "배상도와 순검들 모두 두 발로 마당의 흙바닥을 단단히 디디고 서 있다. 배상도가 들어 올린 오른팔은 어깨에 자연스럽게 연결되어 지탱되며 물리적으로 불가능한 자세이나 공중에 뜬 사물은 없다."
       },
       {
        "label": "B",
        "direction": "배상도는 화면 왼쪽 전경에 서 있는 망건 쓴 남성을 향해 오른손 검지를 뻗어 가리키고 있으며, 턱을 치켜든 채 그 인물을 내려다보고 있다. 뒤편의 순검들은 정면과 측면을 각각 응시하고 있다.",
        "built_space": "대문 바로 안쪽에서 바깥을 향하거나 그 반대를 바라보는 구도로, 양옆으로 나무 문짝 2개가 열려 있다. 배상도는 대문 한가운데 서 있어, 마당 한가운데로 들어서는 인물들의 거리를 묘사해야 하는 원본 스테이징과 구조적으로 맞지 않는다.",
        "entities": "배상도(한국인 남성, 성인)는 기준 사진과 일치하는 검은 모자와 조끼, 흰 한복을 착용했다. 조선인 순검(한국인 남성, 성인) 두 명이 제복 차림으로 뒤편에 등장한다. 그러나 프롬프트에 명시된 허용 인물 외에 망건을 쓴 신원 불명의 한국인 남성이 왼쪽 전경에 등장하여 지시를 위반했다.",
        "hard_violations": [
         "invented people (프롬프트의 PEOPLE 목록에 없는 망건 쓴 남성을 화면 왼쪽 전경에 임의로 추가함)",
         "person placed where the staging does not put them (순검들 뒤에 서야 하는 배상도를 순검들보다 앞쪽인 전경에 배치함)"
        ],
        "physics": "배상도와 순검들, 전경의 남성 모두 두 발로 지면을 딛고 서 있다. 배상도의 오른팔은 어깨와 몸통에 의해 지탱되나 손가락 형태가 약간 왜곡되어 있다. 지지대 없이 공중에 떠 있는 물체는 없다."
       }
      ],
      "all_candidates_fail": false
     },
     "normalized": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 8,
        "verdict_ko": "순검들 뒤에서 안규홍을 가리키는 프레임 구성과 지형 배치를 정확히 구현했으나, 배상도의 기준 복장인 모자와 조끼가 누락되었습니다."
       },
       {
        "label": "B",
        "score": 2,
        "verdict_ko": "배상도의 복장은 완벽하게 재현했으나, 순검들 앞에 서는 스테이징 오류와 지시되지 않은 인물을 추가하는 치명적인 위반을 범했습니다."
       }
      ],
      "all_candidates_fail": false,
      "readings": [
       {
        "label": "A",
        "direction": "배상도는 프레임 왼쪽 바깥에 있을 안규홍을 향해 오른손 검지를 뻗어 정확히 가리키고 있으며, 턱을 치켜들고 거만한 시선으로 그 방향을 응시한다. 앞쪽의 순검 두 명도 가리키는 방향인 왼쪽 밖을 바라보고 있다.",
        "built_space": "기준 사진과 동일한 마당 내부로, 왼쪽에는 마루가 있는 기와집이 위치하고 우측 배경에는 기와지붕이 있는 대문이 비스듬히 보인다. 인물들은 마당 한가운데에 올바르게 서 있으며 카메라가 뒤에서 앞을 비추는 올바른 앵글을 유지하고 있다.",
        "entities": "배상도(한국인 남성, 성인)는 얼굴과 체형이 기준과 일치하지만, 기준 복장 중 검은 모자와 조끼가 누락된 흰색 한복 차림이다. 안규홍 체포를 담당한 조선인 순검(한국인 남성, 성인) 두 명은 검은색 제복과 모자를 쓴 뒷모습으로 등장하며, 지시되지 않은 다른 인물은 나타나지 않았다.",
        "hard_violations": [],
        "physics": "배상도와 순검들 모두 두 발로 마당의 흙바닥을 단단히 디디고 서 있다. 배상도가 들어 올린 오른팔은 어깨에 자연스럽게 연결되어 지탱되며 물리적으로 불가능한 자세이나 공중에 뜬 사물은 없다."
       },
       {
        "label": "B",
        "direction": "배상도는 화면 왼쪽 전경에 서 있는 망건 쓴 남성을 향해 오른손 검지를 뻗어 가리키고 있으며, 턱을 치켜든 채 그 인물을 내려다보고 있다. 뒤편의 순검들은 정면과 측면을 각각 응시하고 있다.",
        "built_space": "대문 바로 안쪽에서 바깥을 향하거나 그 반대를 바라보는 구도로, 양옆으로 나무 문짝 2개가 열려 있다. 배상도는 대문 한가운데 서 있어, 마당 한가운데로 들어서는 인물들의 거리를 묘사해야 하는 원본 스테이징과 구조적으로 맞지 않는다.",
        "entities": "배상도(한국인 남성, 성인)는 기준 사진과 일치하는 검은 모자와 조끼, 흰 한복을 착용했다. 조선인 순검(한국인 남성, 성인) 두 명이 제복 차림으로 뒤편에 등장한다. 그러나 프롬프트에 명시된 허용 인물 외에 망건을 쓴 신원 불명의 한국인 남성이 왼쪽 전경에 등장하여 지시를 위반했다.",
        "hard_violations": [
         "invented people (프롬프트의 PEOPLE 목록에 없는 망건 쓴 남성을 화면 왼쪽 전경에 임의로 추가함)",
         "person placed where the staging does not put them (순검들 뒤에 서야 하는 배상도를 순검들보다 앞쪽인 전경에 배치함)"
        ],
        "physics": "배상도와 순검들, 전경의 남성 모두 두 발로 지면을 딛고 서 있다. 배상도의 오른팔은 어깨와 몸통에 의해 지탱되나 손가락 형태가 약간 왜곡되어 있다. 지지대 없이 공중에 떠 있는 물체는 없다."
       }
      ]
     },
     "ok": true
    },
    {
     "model": "gpt-high",
     "order": "reverse",
     "display_to_canonical": {
      "A": "B",
      "B": "A"
     },
     "raw": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 1,
        "verdict_ko": "턱을 든 표정과 배상도의 복장은 가깝지만, 순검 뒤가 아닌 앞에 서 있으며 금지된 전경 인물까지 추가되어 핵심 배치를 위반한다."
       },
       {
        "label": "B",
        "score": 3,
        "verdict_ko": "장소와 허용된 인물 수는 더 충실하지만, 순검들과 마주 선 배치이고 배상도의 복장도 달라 요구된 상체 중심 장면으로는 부적합하다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "배상도의 검지와 시선은 화면 왼쪽 전경의 흰옷 입은 남자를 향한다. 지목 대상은 분명하지만, 이 남자를 안규홍으로 표현했다면 화면에 허용되지 않은 인물을 보여 준 것이다. 순검 두 명도 대체로 같은 전경 방향을 보며 배상도의 뒤에 있다.",
        "built_space": "가까운 대문 개구부 하나와 양쪽 기둥, 왼쪽으로 열린 문짝 하나, 양옆 담장 및 안쪽 한옥이 보인다. 문 통로는 열려 있으나 카메라는 대문 너머 마당의 건물을 바라보므로 요구된 마당 쪽에서 대문을 비스듬히 보는 구도와 다르다. 배상도가 문 앞쪽을 차지하고 순검 두 명이 그보다 안쪽에 서 있어 '순검들 뒤의 배상도'라는 순서가 뒤집혔다.",
        "entities": "배상도와 제복 차림 순검 두 명 외에 왼쪽 전경 남자 한 명, 총 네 명이 보인다. 모두 성인 한국인 남성 설정에 어긋나는 뚜렷한 외형은 없다. 배상도의 검은 모자·조끼, 흰옷, 허리띠와 늘어진 장식은 참조와 가깝지만 얼굴은 정확히 일치하지 않는다. 순검들은 참조의 현대식 재킷 대신 시대극 제복을 입었으며 얼굴 동일성도 뚜렷하지 않다. 지게는 보이지 않지만 이 인물 중심 크롭에서는 누락만으로 위반이라 할 수 없다. 읽을 수 있는 글자는 없다.",
        "hard_violations": [
         "화면 왼쪽 전경에 허용된 배상도·순검 이외의 지목 대상 인물을 추가했다.",
         "배상도를 순검들 뒤가 아니라 두 순검보다 앞에 배치했다."
        ],
        "physics": "배상도의 지목하는 손은 뻗은 팔과 자연스럽게 연결되고, 반대팔은 몸 옆으로 내려와 있다. 장식과 순검의 휴대품은 허리띠에 매달려 있다. 발은 크롭되어 직접 접지를 확인할 수 없지만 몸통은 정상적으로 선 자세이며, 공중 부양이나 지지 없는 물체는 보이지 않는다."
       },
       {
        "label": "B",
        "direction": "배상도는 턱을 들고 화면 왼쪽으로 검지를 뻗으며 같은 방향을 본다. 손가락의 연장선은 중앙 전경 순검의 머리·어깨 쪽으로 이어져, 화면 밖 안규홍을 지목한다고 명확히 읽히지는 않는다. 순검 두 명은 등과 옆얼굴을 보이며 배상도 쪽을 향해 있어 그를 앞서 마당으로 들어오는 관계보다는 서로 마주 선 관계로 보인다.",
        "built_space": "왼쪽 한옥과 높은 마루, 뒤쪽 기와 담장, 오른쪽 뒤 작은 대문 한 개의 위치와 재료가 장소 참조에 가깝다. 마당 안에서 대문을 보는 시점도 맞는다. 다만 대문은 닫혀 보이며, 인물들 뒤로 확보되어야 할 열린 진입 통로가 드러나지 않는다. 순검 두 명이 전경, 배상도가 중경에 있지만 순검들의 몸 방향 때문에 '들어오는 순검을 뒤따르는 배상도'라는 공간 관계는 성립하지 않는다. 배상도는 허벅지까지 보이고 큰 전경 등들이 화면을 차지해 상체 중심 미디엄 샷의 집중도도 떨어진다.",
        "entities": "배상도 한 명과 순검 두 명으로 총 세 명이며 허용되지 않은 추가 인물은 없다. 보이는 인물들은 성인 한국인 남성 설정과 부합하고 검은 머리도 자연스럽다. 배상도는 참조의 검은 모자와 조끼 없이 베이지색 저고리와 바지를 입어 명확히 다른 복장이며 얼굴의 정확한 일치도 부족하다. 순검들은 모자와 밝은색 옷을 입고 있어 참조 복장과 다르고, 등 위주라 얼굴 동일성은 확인하기 어렵다. 지게는 보이지 않으나 전체 마당을 보여 주는 프레임은 아니므로 누락을 단정할 수 없다. 읽을 수 있는 글자는 없다.",
        "hard_violations": [
         "순검들이 배상도를 향해 마주 서도록 배치되어, 배상도가 도착하는 순검들 뒤에 서 있다는 필수 동선 관계를 위반한다."
        ],
        "physics": "배상도의 팔과 검지는 어깨에서 이어지는 가능한 지목 자세이며, 반대팔은 몸 뒤쪽으로 내려가 일부 가려진다. 세 사람 모두 하체가 프레임 밖으로 이어지는 직립 자세이고 발의 접지는 직접 보이지 않는다. 몸이나 물체가 지지 없이 떠 있는 징후는 없으며, 의복과 모자는 각각 몸과 머리에 정상적으로 지지된다."
       }
      ],
      "all_candidates_fail": true
     },
     "normalized": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 1,
        "verdict_ko": "턱을 든 표정과 배상도의 복장은 가깝지만, 순검 뒤가 아닌 앞에 서 있으며 금지된 전경 인물까지 추가되어 핵심 배치를 위반한다."
       },
       {
        "label": "A",
        "score": 3,
        "verdict_ko": "장소와 허용된 인물 수는 더 충실하지만, 순검들과 마주 선 배치이고 배상도의 복장도 달라 요구된 상체 중심 장면으로는 부적합하다."
       }
      ],
      "all_candidates_fail": true,
      "readings": [
       {
        "label": "B",
        "direction": "배상도의 검지와 시선은 화면 왼쪽 전경의 흰옷 입은 남자를 향한다. 지목 대상은 분명하지만, 이 남자를 안규홍으로 표현했다면 화면에 허용되지 않은 인물을 보여 준 것이다. 순검 두 명도 대체로 같은 전경 방향을 보며 배상도의 뒤에 있다.",
        "built_space": "가까운 대문 개구부 하나와 양쪽 기둥, 왼쪽으로 열린 문짝 하나, 양옆 담장 및 안쪽 한옥이 보인다. 문 통로는 열려 있으나 카메라는 대문 너머 마당의 건물을 바라보므로 요구된 마당 쪽에서 대문을 비스듬히 보는 구도와 다르다. 배상도가 문 앞쪽을 차지하고 순검 두 명이 그보다 안쪽에 서 있어 '순검들 뒤의 배상도'라는 순서가 뒤집혔다.",
        "entities": "배상도와 제복 차림 순검 두 명 외에 왼쪽 전경 남자 한 명, 총 네 명이 보인다. 모두 성인 한국인 남성 설정에 어긋나는 뚜렷한 외형은 없다. 배상도의 검은 모자·조끼, 흰옷, 허리띠와 늘어진 장식은 참조와 가깝지만 얼굴은 정확히 일치하지 않는다. 순검들은 참조의 현대식 재킷 대신 시대극 제복을 입었으며 얼굴 동일성도 뚜렷하지 않다. 지게는 보이지 않지만 이 인물 중심 크롭에서는 누락만으로 위반이라 할 수 없다. 읽을 수 있는 글자는 없다.",
        "hard_violations": [
         "화면 왼쪽 전경에 허용된 배상도·순검 이외의 지목 대상 인물을 추가했다.",
         "배상도를 순검들 뒤가 아니라 두 순검보다 앞에 배치했다."
        ],
        "physics": "배상도의 지목하는 손은 뻗은 팔과 자연스럽게 연결되고, 반대팔은 몸 옆으로 내려와 있다. 장식과 순검의 휴대품은 허리띠에 매달려 있다. 발은 크롭되어 직접 접지를 확인할 수 없지만 몸통은 정상적으로 선 자세이며, 공중 부양이나 지지 없는 물체는 보이지 않는다."
       },
       {
        "label": "A",
        "direction": "배상도는 턱을 들고 화면 왼쪽으로 검지를 뻗으며 같은 방향을 본다. 손가락의 연장선은 중앙 전경 순검의 머리·어깨 쪽으로 이어져, 화면 밖 안규홍을 지목한다고 명확히 읽히지는 않는다. 순검 두 명은 등과 옆얼굴을 보이며 배상도 쪽을 향해 있어 그를 앞서 마당으로 들어오는 관계보다는 서로 마주 선 관계로 보인다.",
        "built_space": "왼쪽 한옥과 높은 마루, 뒤쪽 기와 담장, 오른쪽 뒤 작은 대문 한 개의 위치와 재료가 장소 참조에 가깝다. 마당 안에서 대문을 보는 시점도 맞는다. 다만 대문은 닫혀 보이며, 인물들 뒤로 확보되어야 할 열린 진입 통로가 드러나지 않는다. 순검 두 명이 전경, 배상도가 중경에 있지만 순검들의 몸 방향 때문에 '들어오는 순검을 뒤따르는 배상도'라는 공간 관계는 성립하지 않는다. 배상도는 허벅지까지 보이고 큰 전경 등들이 화면을 차지해 상체 중심 미디엄 샷의 집중도도 떨어진다.",
        "entities": "배상도 한 명과 순검 두 명으로 총 세 명이며 허용되지 않은 추가 인물은 없다. 보이는 인물들은 성인 한국인 남성 설정과 부합하고 검은 머리도 자연스럽다. 배상도는 참조의 검은 모자와 조끼 없이 베이지색 저고리와 바지를 입어 명확히 다른 복장이며 얼굴의 정확한 일치도 부족하다. 순검들은 모자와 밝은색 옷을 입고 있어 참조 복장과 다르고, 등 위주라 얼굴 동일성은 확인하기 어렵다. 지게는 보이지 않으나 전체 마당을 보여 주는 프레임은 아니므로 누락을 단정할 수 없다. 읽을 수 있는 글자는 없다.",
        "hard_violations": [
         "순검들이 배상도를 향해 마주 서도록 배치되어, 배상도가 도착하는 순검들 뒤에 서 있다는 필수 동선 관계를 위반한다."
        ],
        "physics": "배상도의 팔과 검지는 어깨에서 이어지는 가능한 지목 자세이며, 반대팔은 몸 뒤쪽으로 내려가 일부 가려진다. 세 사람 모두 하체가 프레임 밖으로 이어지는 직립 자세이고 발의 접지는 직접 보이지 않는다. 몸이나 물체가 지지 없이 떠 있는 징후는 없으며, 의복과 모자는 각각 몸과 머리에 정상적으로 지지된다."
       }
      ]
     },
     "ok": true
    }
   ],
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "slot_winner_match": true,
   "slot_winner": {
    "gemini-pro": "A",
    "gpt-high": "A"
   },
   "route": "cross_slot_agree"
  },
  "dual": {
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "normalized": {
    "A": 2.0,
    "B": 0.583
   },
   "adjusted": {
    "A": 1.75,
    "B": 0.333
   },
   "violations": {
    "B": [
     "[gemini-pro] invented people (프롬프트의 PEOPLE 목록에 없는 망건 쓴 남성을 화면 왼쪽 전경에 임의로 추가함)",
     "[gemini-pro] person placed where the staging does not put them (순검들 뒤에 서야 하는 배상도를 순검들보다 앞쪽인 전경에 배치함)",
     "[gpt-high] 화면 왼쪽 전경에 허용된 배상도·순검 이외의 지목 대상 인물을 추가했다.",
     "[gpt-high] 배상도를 순검들 뒤가 아니라 두 순검보다 앞에 배치했다."
    ],
    "A": [
     "[gpt-high] 순검들이 배상도를 향해 마주 서도록 배치되어, 배상도가 도착하는 순검들 뒤에 서 있다는 필수 동선 관계를 위반한다."
    ]
   },
   "per_model_winner": {
    "gemini-pro": "A",
    "gpt-high": "A"
   },
   "agreed": true
  },
  "totals": {
   "A": 1750,
   "B": 333
  },
  "selected": "A",
  "ranking": [
   "A",
   "B"
  ],
  "verdicts": [
   {
    "label": "A",
    "score": 1750,
    "verdict_ko": "순검들 뒤에서 안규홍을 가리키는 프레임 구성과 지형 배치를 정확히 구현했으나, 배상도의 기준 복장인 모자와 조끼가 누락되었습니다.  ★위반: [gpt-high] 순검들이 배상도를 향해 마주 서도록 배치되어, 배상도가 도착하는 순검들 뒤에 서 있다는 필수 동선 관계를 위반한다."
   },
   {
    "label": "B",
    "score": 333,
    "verdict_ko": "배상도의 복장은 완벽하게 재현했으나, 순검들 앞에 서는 스테이징 오류와 지시되지 않은 인물을 추가하는 치명적인 위반을 범했습니다.  ★위반: [gemini-pro] invented people (프롬프트의 PEOPLE 목록에 없는 망건 쓴 남성을 화면 왼쪽 전경에 임의로 추가함) / [gemini-pro] person placed where the staging does not put them (순검들 뒤에 서야 하는 배상도를 순검들보다 앞쪽인 전경에 배치함) / [gpt-high] 화면 왼쪽 전경에 허용된 배상도·순검 이외의 지목 대상 인물을 추가했다. / [gpt-high] 배상도를 순검들 뒤가 아니라 두 순검보다 앞에 배치했다."
   }
  ],
  "refs": [
   {
    "label": "LOCATION PHOTOGRAPH — the exact place of this shot: its architecture, materials, fixed features and lighting mood are spatial truth; stage the moment inside this place. Never copy its camera framing.",
    "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/db00517b-165b-4005-a287-f28936b59b95/episodes/272bd7af-9294-4490-98c6-88a40e9deacc/images/background_chain/L08B01.png",
    "asset_id": "64556ed8-9e6e-4367-ade4-e6cfd6cae678",
    "role": "location_plate"
   },
   {
    "label": "CHARACTER REFERENCE — 배상도: the exact person appearing in this shot; match face, hair and build exactly.",
    "path": "<bytes:851718>",
    "asset_id": "034f5325-975e-4474-854b-e48359710bf8",
    "role": "character_ref"
   },
   {
    "label": "CHARACTER REFERENCE — 조선인 순검(안규홍 체포·옥사 담당): the exact person appearing in this shot; match face, hair and build exactly.",
    "path": "<bytes:1264194>",
    "asset_id": "0b736389-3dee-4967-9bc7-ff4ea80d9b44",
    "role": "character_ref"
   }
  ],
  "critique_skipped": true,
  "shot_run_produced": false,
  "shot_run_uid": "06aa1406-ffcc-748a-8eba-81ac46d78924",
  "bgfirst": {
   "bg_path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/db00517b-165b-4005-a287-f28936b59b95/images/272bd7af-9294-4490-98c6-88a40e9deacc/scene/recipe/S7sh4__bgfirst_bg.png",
   "bg_asset_id": "bb4dbf43-e434-4ef1-82d6-a9e5f3e104ab",
   "bg_record_key": "S7sh4::bgfirst_bg",
   "chain_winner": true,
   "authority": "plate"
  },
  "ref_mode": "재투영 배경+콘티+엔티티 (2택1: 체인 승)",
  "share_plan": {
   "ref_plan": "background"
  }
 },
 "S7sh4::cine": {
  "applied": true,
  "attempted_at": "2026-09-09T01:18:23.061075+00:00",
  "fingerprint": "f24991dda33d25f581e10b3adf3be655a67174ac7ad858b8d73a5cd2055e0513",
  "fingerprint_version": 2,
  "provider": "mai",
  "endpoint": "openrouter/chat-completions",
  "model": "microsoft/mai-image-2.6",
  "pack": "24.202608252115",
  "source_file": "S7sh4_sel.png",
  "source_sha256": "32b386fea8ebb3102e8b74de055d05e31936979a854a13e2376f9365e8756be2",
  "file": "S7sh4_cine.png",
  "staged_sha256": "8991107d057d65647a652fd9b8ab72a2637a6f0040e65a795fa57c71adac5c75",
  "latency_ms": 30388
 },
 "S7sh10::signage": {
  "fp": "02c1e35878a796ff",
  "inscriptions": [],
  "cues": [],
  "dropped": []
 },
 "S7sh10": {
  "input_fingerprint": "058959cab7938c2b",
  "prompt": "Create ONE FINAL photorealistic live-action film still of the moment below — early 20th-century Korea, Jeolla region; all people are Korean unless explicitly stated otherwise. TIME OF DAY (lock): late afternoon.\n\nSHOT TEXT (authoritative, Korean): 양팔이 조선인 순검들에게 단단히 붙들린 채 대문 밖을 향해 몸이 기울어 두 발이 흙바닥에 끌리고 있는 안규홍의 뒷모습.\n\nLOCATION (lock): The dirt exit route across a traditional estate's courtyard toward its open main gate. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: wide shot\n- KEY BACKGROUND ELEMENTS: 대문 (The passage remains clear for the departing escort) — Its courtyard-facing opening lies ahead of the group; used as Provides the destination for the officers' attention and the group's movement; 마당 흙바닥 (안규홍's feet drag across it); used as Retains visible contact beneath both feet without adding invented dust or drag marks.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Maintain the courtyard's natural late-afternoon illumination and subdued color, keeping the restrained body and dragging feet clearly readable.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The clothing, hair and overall look of 조선인 순검(안규홍 체포·옥사 담당) — who appear both in that photo and in this shot — are LOCKED to that photo. Anyone else visible in that photo is NOT in this shot: never carry their face, body or clothing onto anyone here. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The carrying frame remains abandoned in the center of the yard. 안규홍: His arms are twisted back in restraint, and his body has relaxed rather than resisted; he carries no frame. 조선인 순검(안규홍 체포·옥사 담당): He maintains an arm-restraint grip while escorting out.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 안규홍 (한국인 남성, 청년 나이대, 검은 머리, 단단히 묶은 머리); 조선인 순검(안규홍 체포·옥사 담당) (한국인 남성, 성인 나이대, 자연스러운 검은 머리) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.",
  "roll_prompts": {
   "A": "Create ONE FINAL photorealistic live-action film still of the moment below — early 20th-century Korea, Jeolla region; all people are Korean unless explicitly stated otherwise. TIME OF DAY (lock): late afternoon.\n\nSHOT TEXT (authoritative, Korean): 양팔이 조선인 순검들에게 단단히 붙들린 채 대문 밖을 향해 몸이 기울어 두 발이 흙바닥에 끌리고 있는 안규홍의 뒷모습.\n\nLOCATION (lock): The dirt exit route across a traditional estate's courtyard toward its open main gate. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: wide shot\n- KEY BACKGROUND ELEMENTS: 대문 (The passage remains clear for the departing escort) — Its courtyard-facing opening lies ahead of the group; used as Provides the destination for the officers' attention and the group's movement; 마당 흙바닥 (안규홍's feet drag across it); used as Retains visible contact beneath both feet without adding invented dust or drag marks.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Maintain the courtyard's natural late-afternoon illumination and subdued color, keeping the restrained body and dragging feet clearly readable.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The clothing, hair and overall look of 조선인 순검(안규홍 체포·옥사 담당) — who appear both in that photo and in this shot — are LOCKED to that photo. Anyone else visible in that photo is NOT in this shot: never carry their face, body or clothing onto anyone here. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The carrying frame remains abandoned in the center of the yard. 안규홍: His arms are twisted back in restraint, and his body has relaxed rather than resisted; he carries no frame. 조선인 순검(안규홍 체포·옥사 담당): He maintains an arm-restraint grip while escorting out.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 안규홍 (한국인 남성, 청년 나이대, 검은 머리, 단단히 묶은 머리); 조선인 순검(안규홍 체포·옥사 담당) (한국인 남성, 성인 나이대, 자연스러운 검은 머리) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.",
   "B": "Create ONE FINAL photorealistic live-action film still of the moment below — early 20th-century Korea, Jeolla region; all people are Korean unless explicitly stated otherwise. TIME OF DAY (lock): late afternoon.\n\nSHOT TEXT (authoritative, Korean): 양팔이 조선인 순검들에게 단단히 붙들린 채 대문 밖을 향해 몸이 기울어 두 발이 흙바닥에 끌리고 있는 안규홍의 뒷모습.\n\nLOCATION (lock): The dirt exit route across a traditional estate's courtyard toward its open main gate. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: wide shot\n- KEY BACKGROUND ELEMENTS: 대문 (The passage remains clear for the departing escort) — Its courtyard-facing opening lies ahead of the group; used as Provides the destination for the officers' attention and the group's movement; 마당 흙바닥 (안규홍's feet drag across it); used as Retains visible contact beneath both feet without adding invented dust or drag marks.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Maintain the courtyard's natural late-afternoon illumination and subdued color, keeping the restrained body and dragging feet clearly readable.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The clothing, hair and overall look of 조선인 순검(안규홍 체포·옥사 담당) — who appear both in that photo and in this shot — are LOCKED to that photo. Anyone else visible in that photo is NOT in this shot: never carry their face, body or clothing onto anyone here. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The carrying frame remains abandoned in the center of the yard. 안규홍: His arms are twisted back in restraint, and his body has relaxed rather than resisted; he carries no frame. 조선인 순검(안규홍 체포·옥사 담당): He maintains an arm-restraint grip while escorting out.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 안규홍 (한국인 남성, 청년 나이대, 검은 머리, 단단히 묶은 머리); 조선인 순검(안규홍 체포·옥사 담당) (한국인 남성, 성인 나이대, 자연스러운 검은 머리) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.\n\nCOMPOSITION VARIATION (this candidate only): keep the place, the people and the moment exactly as specified, and keep from the CAMERA & FRAME contract its subject scale, screen placement and key background elements exactly. For this candidate only, its camera angle and height are deliberately left open: choose a camera position distinctly different from the obvious one, as a film director picking a second setup on the same blocking. Stay true to any explicit framing the SHOT TEXT itself dictates."
  },
  "shot_run_spend_attempt_count": 1,
  "readings": [
   {
    "label": "A",
    "direction": "안규홍과 두 명의 순검이 배경에 있는 열린 대문을 향해 나아가고 있다. 안규홍의 몸은 뒤로 기울어진 채 대문 쪽을 향한다.",
    "built_space": "전통 건축물과 흙바닥으로 이루어진 마당, 그리고 배경의 열린 대문 등 공간 구조는 기준 이미지와 일치한다. 그러나 마당 중앙에 있어야 할 지게는 존재하지 않는다.",
    "entities": "안규홍(뒷모습, 흰옷, 묶은 머리)은 지문과 일치한다. 순검은 두 명이 등장해 양팔을 붙잡는 조건은 충족했으나, 기준 이미지의 검은색 제복이 아닌 파란색 위장 무늬 패턴의 잘못된 옷을 입고 있다.",
    "hard_violations": [
     "[gpt-high] 추가하지 말라고 명시한 길고 평행한 발 끌림 자국을 흙바닥에 만들어 넣었다."
    ],
    "physics": "두 명의 순검이 안규홍의 양팔을 단단히 붙잡아 뒤로 기울어진 체중을 지탱하고 있어 물리적 균형이 가능하다. 다만 프롬프트에서 명시적으로 금지한 두 줄의 끌린 자국(drag marks)이 바닥에 깊게 파여 있다."
   },
   {
    "label": "B",
    "direction": "안규홍과 한 명의 순검이 배경의 열린 대문을 향하고 있다. 안규홍의 몸이 대문을 향해 비정상적으로 심하게 누워 있다.",
    "built_space": "흙바닥 마당과 배경의 대문은 기준 이미지와 일치하며, 마당 중앙에 지문이 요구한 지게가 버려져 있다.",
    "entities": "안규홍(뒷모습, 흰옷)은 존재하나, 양팔을 붙들어야 할 순검이 단 한 명뿐이다. 순검의 제복(검은색, 모자, 견장)은 기준 이미지와 정확히 일치한다.",
    "hard_violations": [
     "[gemini-pro] Physically impossible anatomy or staging: 안규홍이 중력을 거슬러 뒤로 45도 이상 심하게 기울어져 있으나, 단 한 명의 순검이 왼팔만 가볍게 잡고 꼿꼿이 걷고 있어 체중을 지탱할 수 없는 물리적으로 불가능한 자세임.",
     "[gpt-high] 추가하지 말라고 명시한 긴 발 끌림 자국과 발밑의 먼지 구름을 만들어 넣었다."
    ],
    "physics": "안규홍의 체중이 뒤로 심하게 쏠려 넘어지기 직전의 각도임에도 불구하고, 옆의 순검 한 명이 전혀 힘을 주지 않고 한쪽 팔만 가볍게 잡고 있어 지탱될 수 없는 불가능한 자세다. 또한 프롬프트에서 금지한 흙먼지와 바닥의 끌린 자국이 강하게 묘사되어 있다."
   }
  ],
  "cross_model_order": {
   "policy": "cross_model_order_v2_slot0_fwd_slot1_rev_parallel2",
   "slots": [
    {
     "model": "gemini-pro",
     "order": "forward",
     "display_to_canonical": {
      "A": "A",
      "B": "B"
     },
     "raw": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 4,
        "verdict_ko": "두 명의 순검이 양팔을 단단히 붙들고 끌고 가는 핵심 동작과 물리적 균형은 제대로 연출되었으나, 순검의 제복이 기준 이미지와 전혀 다르고 마당의 지게가 누락되었으며 명시적으로 금지된 끌린 자국이 추가되었습니다."
       },
       {
        "label": "B",
        "score": 1,
        "verdict_ko": "순검의 제복과 마당의 지게는 기준과 일치하지만, 안규홍이 체중 지탱 없이 물리적으로 불가능한 각도로 뒤로 누워 있고 순검이 한 명뿐이며 금지된 흙먼지와 자국이 추가되어 실격입니다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "안규홍과 두 명의 순검이 배경에 있는 열린 대문을 향해 나아가고 있다. 안규홍의 몸은 뒤로 기울어진 채 대문 쪽을 향한다.",
        "built_space": "전통 건축물과 흙바닥으로 이루어진 마당, 그리고 배경의 열린 대문 등 공간 구조는 기준 이미지와 일치한다. 그러나 마당 중앙에 있어야 할 지게는 존재하지 않는다.",
        "entities": "안규홍(뒷모습, 흰옷, 묶은 머리)은 지문과 일치한다. 순검은 두 명이 등장해 양팔을 붙잡는 조건은 충족했으나, 기준 이미지의 검은색 제복이 아닌 파란색 위장 무늬 패턴의 잘못된 옷을 입고 있다.",
        "hard_violations": [],
        "physics": "두 명의 순검이 안규홍의 양팔을 단단히 붙잡아 뒤로 기울어진 체중을 지탱하고 있어 물리적 균형이 가능하다. 다만 프롬프트에서 명시적으로 금지한 두 줄의 끌린 자국(drag marks)이 바닥에 깊게 파여 있다."
       },
       {
        "label": "B",
        "direction": "안규홍과 한 명의 순검이 배경의 열린 대문을 향하고 있다. 안규홍의 몸이 대문을 향해 비정상적으로 심하게 누워 있다.",
        "built_space": "흙바닥 마당과 배경의 대문은 기준 이미지와 일치하며, 마당 중앙에 지문이 요구한 지게가 버려져 있다.",
        "entities": "안규홍(뒷모습, 흰옷)은 존재하나, 양팔을 붙들어야 할 순검이 단 한 명뿐이다. 순검의 제복(검은색, 모자, 견장)은 기준 이미지와 정확히 일치한다.",
        "hard_violations": [
         "Physically impossible anatomy or staging: 안규홍이 중력을 거슬러 뒤로 45도 이상 심하게 기울어져 있으나, 단 한 명의 순검이 왼팔만 가볍게 잡고 꼿꼿이 걷고 있어 체중을 지탱할 수 없는 물리적으로 불가능한 자세임."
        ],
        "physics": "안규홍의 체중이 뒤로 심하게 쏠려 넘어지기 직전의 각도임에도 불구하고, 옆의 순검 한 명이 전혀 힘을 주지 않고 한쪽 팔만 가볍게 잡고 있어 지탱될 수 없는 불가능한 자세다. 또한 프롬프트에서 금지한 흙먼지와 바닥의 끌린 자국이 강하게 묘사되어 있다."
       }
      ],
      "all_candidates_fail": false
     },
     "normalized": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 4,
        "verdict_ko": "두 명의 순검이 양팔을 단단히 붙들고 끌고 가는 핵심 동작과 물리적 균형은 제대로 연출되었으나, 순검의 제복이 기준 이미지와 전혀 다르고 마당의 지게가 누락되었으며 명시적으로 금지된 끌린 자국이 추가되었습니다."
       },
       {
        "label": "B",
        "score": 1,
        "verdict_ko": "순검의 제복과 마당의 지게는 기준과 일치하지만, 안규홍이 체중 지탱 없이 물리적으로 불가능한 각도로 뒤로 누워 있고 순검이 한 명뿐이며 금지된 흙먼지와 자국이 추가되어 실격입니다."
       }
      ],
      "all_candidates_fail": false,
      "readings": [
       {
        "label": "A",
        "direction": "안규홍과 두 명의 순검이 배경에 있는 열린 대문을 향해 나아가고 있다. 안규홍의 몸은 뒤로 기울어진 채 대문 쪽을 향한다.",
        "built_space": "전통 건축물과 흙바닥으로 이루어진 마당, 그리고 배경의 열린 대문 등 공간 구조는 기준 이미지와 일치한다. 그러나 마당 중앙에 있어야 할 지게는 존재하지 않는다.",
        "entities": "안규홍(뒷모습, 흰옷, 묶은 머리)은 지문과 일치한다. 순검은 두 명이 등장해 양팔을 붙잡는 조건은 충족했으나, 기준 이미지의 검은색 제복이 아닌 파란색 위장 무늬 패턴의 잘못된 옷을 입고 있다.",
        "hard_violations": [],
        "physics": "두 명의 순검이 안규홍의 양팔을 단단히 붙잡아 뒤로 기울어진 체중을 지탱하고 있어 물리적 균형이 가능하다. 다만 프롬프트에서 명시적으로 금지한 두 줄의 끌린 자국(drag marks)이 바닥에 깊게 파여 있다."
       },
       {
        "label": "B",
        "direction": "안규홍과 한 명의 순검이 배경의 열린 대문을 향하고 있다. 안규홍의 몸이 대문을 향해 비정상적으로 심하게 누워 있다.",
        "built_space": "흙바닥 마당과 배경의 대문은 기준 이미지와 일치하며, 마당 중앙에 지문이 요구한 지게가 버려져 있다.",
        "entities": "안규홍(뒷모습, 흰옷)은 존재하나, 양팔을 붙들어야 할 순검이 단 한 명뿐이다. 순검의 제복(검은색, 모자, 견장)은 기준 이미지와 정확히 일치한다.",
        "hard_violations": [
         "Physically impossible anatomy or staging: 안규홍이 중력을 거슬러 뒤로 45도 이상 심하게 기울어져 있으나, 단 한 명의 순검이 왼팔만 가볍게 잡고 꼿꼿이 걷고 있어 체중을 지탱할 수 없는 물리적으로 불가능한 자세임."
        ],
        "physics": "안규홍의 체중이 뒤로 심하게 쏠려 넘어지기 직전의 각도임에도 불구하고, 옆의 순검 한 명이 전혀 힘을 주지 않고 한쪽 팔만 가볍게 잡고 있어 지탱될 수 없는 불가능한 자세다. 또한 프롬프트에서 금지한 흙먼지와 바닥의 끌린 자국이 강하게 묘사되어 있다."
       }
      ]
     },
     "ok": true
    },
    {
     "model": "gpt-high",
     "order": "reverse",
     "display_to_canonical": {
      "A": "B",
      "B": "A"
     },
     "raw": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 2,
        "verdict_ko": "와이드 뒷모습과 버려진 지게는 보이지만 순검이 한 명뿐이고, 몸의 기울기가 대문 방향과 어긋나며 금지된 먼지와 끌린 자국까지 추가됐다."
       },
       {
        "label": "B",
        "score": 4,
        "verdict_ko": "두 순검이 양팔을 붙들고 열린 대문으로 이동하는 뒷모습은 더 충실하지만, 끌린 자국 추가와 순검 복장·대문 구조 변경으로 그대로 사용할 수는 없다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "열린 대문은 일행의 오른쪽 앞에 있다. 안규홍의 머리는 오른쪽 아래를 향하지만, 상체는 발보다 화면 왼쪽으로 기울어 대문 쪽으로 끌려가는 방향과 맞지 않는다. 순검은 안규홍 쪽으로 고개를 돌렸으며 대문을 주시하는 모습은 아니다. 허리에 매달린 긴 물건은 아래를 향하고 특정 대상을 겨누지 않는다.",
        "built_space": "왼쪽에 마루가 있는 한옥 한 동, 뒤쪽에 낮은 돌담, 중앙 오른쪽에 열린 대문 한 곳이 보인다. 지게 한 개가 마당 바닥에 놓여 있고 대문 개구부 자체는 비어 있다. 다만 기준 사진의 작은 독립 대문에 비해 문과 지붕이 훨씬 크고, 담 너머는 수목 중심 풍경 대신 여러 기와지붕으로 채워졌다. 일행은 대문에서 떨어진 마당 왼쪽에 서 있다.",
        "entities": "안규홍으로 보이는 청년 남성 한 명과 성인 남성 순검 한 명만 보여, 복수의 순검이 양팔을 붙드는 요구를 충족하지 못한다. 안규홍의 검게 묶은 머리, 밝은색 저고리와 바지, 짚신은 대체로 인물 참고와 맞지만 얼굴은 뒷모습이라 확인할 수 없다. 순검의 검정 제복은 인물 참고에 가깝지만 우선 고정된 이전 장면의 밝은색 옷과 다르다. 안규홍은 지게를 메지 않으며 지게는 바닥에 있다. 읽을 수 있는 글자는 없다.",
        "hard_violations": [
         "추가하지 말라고 명시한 긴 발 끌림 자국과 발밑의 먼지 구름을 만들어 넣었다."
        ],
        "physics": "순검은 두 부츠로 땅을 딛고 한 손으로 안규홍의 위팔을 잡고 있다. 안규홍은 왼발과 순검의 손에 지지되므로 몸 전체가 근거 없이 떠 있지는 않다. 그러나 오른쪽 짚신은 흙바닥에서 들려 보여 두 발이 동시에 지면에 닿아 끌리는 조건에 맞지 않는다. 두 손은 등 뒤에 모였지만 반대쪽 팔을 붙드는 순검의 손은 없다. 지게는 나무틀로 땅에 지지되고 순검의 긴 휴대물은 허리띠에 매달려 있다."
       },
       {
        "label": "B",
        "direction": "두 순검의 머리와 몸은 화면 오른쪽 앞의 열린 대문을 향하고, 안규홍도 그 사이에서 같은 방향으로 이동한다. 안규홍은 고개를 숙이고 상체를 앞으로 기울여 A보다 목적지를 향한 호송 동작이 명확하다. 얼굴과 눈은 뒤에서 보이지 않아 정확한 시선은 확인할 수 없다. 대문 밖에는 이동을 막는 사람이 없다.",
        "built_space": "왼쪽에 마루가 있는 한옥 한 동과 낮은 돌담이 있고, 오른쪽 앞에는 통과 가능한 대문 한 곳이 있다. 대문 좌우에는 각각 닫힌 목제 문이 있는 벽체가 붙어 있다. 이는 기준 사진의 낮은 담 사이에 놓인 작은 독립 대문보다 큰 문간채 구조여서 장소 연속성이 떨어진다. 세 사람은 마당 중앙에서 대문으로 향하며 통로는 비어 있다. 보이는 마당에는 버려진 지게가 확인되지 않지만 일행 뒤쪽 바닥 일부는 몸에 가려져 있다.",
        "entities": "안규홍 한 명과 양옆의 순검 두 명으로 필요한 인원 관계가 맞는다. 세 사람 모두 한국인 남성 설정과 모순되는 외형은 없지만 뒷모습이므로 얼굴 동일성은 판단할 수 없다. 안규홍의 묶은 검은 머리, 밝은색 한복과 짚신은 참고와 대체로 맞는다. 순검들은 검은 모자와 허리띠를 착용했으나 청색 무늬 상의는 이전 장면의 밝은색 복장에도, 별도 인물 참고의 검정 제복에도 맞지 않는다. 안규홍이 지게를 지지 않은 점은 맞고 읽을 수 있는 글자는 없다.",
        "hard_violations": [
         "추가하지 말라고 명시한 길고 평행한 발 끌림 자국을 흙바닥에 만들어 넣었다."
        ],
        "physics": "각 순검의 안쪽 손이 안규홍의 양쪽 위팔에 닿아 붙들고 있어 호송과 체중 지지의 근거가 있다. 순검들은 한쪽 발을 딛고 반대쪽 발을 들어 걷는 자세다. 안규홍의 팔은 등 뒤로 돌아가 있고 몸은 두 사람 사이에 처져 있어 힘을 뺀 상태로 읽힌다. 왼쪽 짚신은 바닥과 접촉하지만 오른쪽 짚신은 밑창을 드러낸 채 들린 듯해 두 발 모두의 지속적인 지면 접촉은 명확하지 않다. 몸 전체가 지지 없이 떠 있는 상황은 아니다."
       }
      ],
      "all_candidates_fail": true
     },
     "normalized": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 2,
        "verdict_ko": "와이드 뒷모습과 버려진 지게는 보이지만 순검이 한 명뿐이고, 몸의 기울기가 대문 방향과 어긋나며 금지된 먼지와 끌린 자국까지 추가됐다."
       },
       {
        "label": "A",
        "score": 4,
        "verdict_ko": "두 순검이 양팔을 붙들고 열린 대문으로 이동하는 뒷모습은 더 충실하지만, 끌린 자국 추가와 순검 복장·대문 구조 변경으로 그대로 사용할 수는 없다."
       }
      ],
      "all_candidates_fail": true,
      "readings": [
       {
        "label": "B",
        "direction": "열린 대문은 일행의 오른쪽 앞에 있다. 안규홍의 머리는 오른쪽 아래를 향하지만, 상체는 발보다 화면 왼쪽으로 기울어 대문 쪽으로 끌려가는 방향과 맞지 않는다. 순검은 안규홍 쪽으로 고개를 돌렸으며 대문을 주시하는 모습은 아니다. 허리에 매달린 긴 물건은 아래를 향하고 특정 대상을 겨누지 않는다.",
        "built_space": "왼쪽에 마루가 있는 한옥 한 동, 뒤쪽에 낮은 돌담, 중앙 오른쪽에 열린 대문 한 곳이 보인다. 지게 한 개가 마당 바닥에 놓여 있고 대문 개구부 자체는 비어 있다. 다만 기준 사진의 작은 독립 대문에 비해 문과 지붕이 훨씬 크고, 담 너머는 수목 중심 풍경 대신 여러 기와지붕으로 채워졌다. 일행은 대문에서 떨어진 마당 왼쪽에 서 있다.",
        "entities": "안규홍으로 보이는 청년 남성 한 명과 성인 남성 순검 한 명만 보여, 복수의 순검이 양팔을 붙드는 요구를 충족하지 못한다. 안규홍의 검게 묶은 머리, 밝은색 저고리와 바지, 짚신은 대체로 인물 참고와 맞지만 얼굴은 뒷모습이라 확인할 수 없다. 순검의 검정 제복은 인물 참고에 가깝지만 우선 고정된 이전 장면의 밝은색 옷과 다르다. 안규홍은 지게를 메지 않으며 지게는 바닥에 있다. 읽을 수 있는 글자는 없다.",
        "hard_violations": [
         "추가하지 말라고 명시한 긴 발 끌림 자국과 발밑의 먼지 구름을 만들어 넣었다."
        ],
        "physics": "순검은 두 부츠로 땅을 딛고 한 손으로 안규홍의 위팔을 잡고 있다. 안규홍은 왼발과 순검의 손에 지지되므로 몸 전체가 근거 없이 떠 있지는 않다. 그러나 오른쪽 짚신은 흙바닥에서 들려 보여 두 발이 동시에 지면에 닿아 끌리는 조건에 맞지 않는다. 두 손은 등 뒤에 모였지만 반대쪽 팔을 붙드는 순검의 손은 없다. 지게는 나무틀로 땅에 지지되고 순검의 긴 휴대물은 허리띠에 매달려 있다."
       },
       {
        "label": "A",
        "direction": "두 순검의 머리와 몸은 화면 오른쪽 앞의 열린 대문을 향하고, 안규홍도 그 사이에서 같은 방향으로 이동한다. 안규홍은 고개를 숙이고 상체를 앞으로 기울여 A보다 목적지를 향한 호송 동작이 명확하다. 얼굴과 눈은 뒤에서 보이지 않아 정확한 시선은 확인할 수 없다. 대문 밖에는 이동을 막는 사람이 없다.",
        "built_space": "왼쪽에 마루가 있는 한옥 한 동과 낮은 돌담이 있고, 오른쪽 앞에는 통과 가능한 대문 한 곳이 있다. 대문 좌우에는 각각 닫힌 목제 문이 있는 벽체가 붙어 있다. 이는 기준 사진의 낮은 담 사이에 놓인 작은 독립 대문보다 큰 문간채 구조여서 장소 연속성이 떨어진다. 세 사람은 마당 중앙에서 대문으로 향하며 통로는 비어 있다. 보이는 마당에는 버려진 지게가 확인되지 않지만 일행 뒤쪽 바닥 일부는 몸에 가려져 있다.",
        "entities": "안규홍 한 명과 양옆의 순검 두 명으로 필요한 인원 관계가 맞는다. 세 사람 모두 한국인 남성 설정과 모순되는 외형은 없지만 뒷모습이므로 얼굴 동일성은 판단할 수 없다. 안규홍의 묶은 검은 머리, 밝은색 한복과 짚신은 참고와 대체로 맞는다. 순검들은 검은 모자와 허리띠를 착용했으나 청색 무늬 상의는 이전 장면의 밝은색 복장에도, 별도 인물 참고의 검정 제복에도 맞지 않는다. 안규홍이 지게를 지지 않은 점은 맞고 읽을 수 있는 글자는 없다.",
        "hard_violations": [
         "추가하지 말라고 명시한 길고 평행한 발 끌림 자국을 흙바닥에 만들어 넣었다."
        ],
        "physics": "각 순검의 안쪽 손이 안규홍의 양쪽 위팔에 닿아 붙들고 있어 호송과 체중 지지의 근거가 있다. 순검들은 한쪽 발을 딛고 반대쪽 발을 들어 걷는 자세다. 안규홍의 팔은 등 뒤로 돌아가 있고 몸은 두 사람 사이에 처져 있어 힘을 뺀 상태로 읽힌다. 왼쪽 짚신은 바닥과 접촉하지만 오른쪽 짚신은 밑창을 드러낸 채 들린 듯해 두 발 모두의 지속적인 지면 접촉은 명확하지 않다. 몸 전체가 지지 없이 떠 있는 상황은 아니다."
       }
      ]
     },
     "ok": true
    }
   ],
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "slot_winner_match": true,
   "slot_winner": {
    "gemini-pro": "A",
    "gpt-high": "A"
   },
   "route": "cross_slot_agree"
  },
  "dual": {
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "normalized": {
    "A": 2.0,
    "B": 0.75
   },
   "adjusted": {
    "A": 1.75,
    "B": 0.5
   },
   "violations": {
    "B": [
     "[gemini-pro] Physically impossible anatomy or staging: 안규홍이 중력을 거슬러 뒤로 45도 이상 심하게 기울어져 있으나, 단 한 명의 순검이 왼팔만 가볍게 잡고 꼿꼿이 걷고 있어 체중을 지탱할 수 없는 물리적으로 불가능한 자세임.",
     "[gpt-high] 추가하지 말라고 명시한 긴 발 끌림 자국과 발밑의 먼지 구름을 만들어 넣었다."
    ],
    "A": [
     "[gpt-high] 추가하지 말라고 명시한 길고 평행한 발 끌림 자국을 흙바닥에 만들어 넣었다."
    ]
   },
   "per_model_winner": {
    "gemini-pro": "A",
    "gpt-high": "A"
   },
   "agreed": true
  },
  "totals": {
   "A": 1750,
   "B": 500
  },
  "selected": "A",
  "ranking": [
   "A",
   "B"
  ],
  "verdicts": [
   {
    "label": "A",
    "score": 1750,
    "verdict_ko": "두 명의 순검이 양팔을 단단히 붙들고 끌고 가는 핵심 동작과 물리적 균형은 제대로 연출되었으나, 순검의 제복이 기준 이미지와 전혀 다르고 마당의 지게가 누락되었으며 명시적으로 금지된 끌린 자국이 추가되었습니다.  ★위반: [gpt-high] 추가하지 말라고 명시한 길고 평행한 발 끌림 자국을 흙바닥에 만들어 넣었다."
   },
   {
    "label": "B",
    "score": 500,
    "verdict_ko": "순검의 제복과 마당의 지게는 기준과 일치하지만, 안규홍이 체중 지탱 없이 물리적으로 불가능한 각도로 뒤로 누워 있고 순검이 한 명뿐이며 금지된 흙먼지와 자국이 추가되어 실격입니다.  ★위반: [gemini-pro] Physically impossible anatomy or staging: 안규홍이 중력을 거슬러 뒤로 45도 이상 심하게 기울어져 있으나, 단 한 명의 순검이 왼팔만 가볍게 잡고 꼿꼿이 걷고 있어 체중을 지탱할 수 없는 물리적으로 불가능한 자세임. / [gpt-high] 추가하지 말라고 명시한 긴 발 끌림 자국과 발밑의 먼지 구름을 만들어 넣었다."
   }
  ],
  "refs": [
   {
    "label": "PREVIOUS SHOT STILL — a visually related earlier shot of this same place: the location's look, materials, fixed features and lighting mood are LOCKED to this photo; never copy its camera framing. The clothing, hair and overall look of 조선인 순검(안규홍 체포·옥사 담당) — who appear both in that photo and in this shot — are LOCKED to that photo. Anyone else visible in that photo is NOT in this shot: never carry their face, body or clothing onto anyone here. If this photo shows a character who cannot move (dead or unconscious), that character's exact pose, position and orientation are ALSO LOCKED — treat their body as a fixed prop of the set that only the camera moves around.",
    "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/db00517b-165b-4005-a287-f28936b59b95/images/272bd7af-9294-4490-98c6-88a40e9deacc/scene/recipe/S7sh4_sel.png",
    "asset_id": "13536a8b-761b-41a2-8755-42dbccce2212",
    "role": "prev_still"
   },
   {
    "label": "CHARACTER REFERENCE — 안규홍: the exact person appearing in this shot; match face, hair and build exactly.",
    "path": "<bytes:1028384>",
    "asset_id": "3b5894db-ce55-4f34-b966-5e725c3aa952",
    "role": "character_ref"
   },
   {
    "label": "CHARACTER REFERENCE — 조선인 순검(안규홍 체포·옥사 담당): the exact person appearing in this shot; match face, hair and build exactly.",
    "path": "<bytes:769097>",
    "asset_id": "f2c44e25-281b-405f-a293-5aa733a69f0b",
    "role": "character_ref"
   }
  ],
  "critique_skipped": true,
  "shot_run_produced": false,
  "shot_run_uid": "06aa1407-02c0-750b-a6af-16e2a5ef9008",
  "ref_mode": "prev+엔티티",
  "share_plan": {
   "ref_plan": "prev",
   "prev_anchor_tag": "S7sh4"
  },
  "lane_policy": "ab_select_bypass:prev"
 },
 "S7sh10::cine": {
  "applied": true,
  "attempted_at": "2026-09-09T01:20:09.769315+00:00",
  "fingerprint": "7bbb88eb5021cc6abfdb0cb120b899bc8592bd637d58920a59df8c437395462d",
  "fingerprint_version": 2,
  "provider": "mai",
  "endpoint": "openrouter/chat-completions",
  "model": "microsoft/mai-image-2.6",
  "pack": "24.202608252115",
  "source_file": "S7sh10_sel.png",
  "source_sha256": "80f191044cfe3c75ab7cdfe2bafebb7d5c30d7d5fce4d817b9e3083e75732f63",
  "file": "S7sh10_cine.png",
  "staged_sha256": "dc935402892b43228c6d52b1d1d5a09f32d857177115d38a83a096a790c25aee",
  "latency_ms": 26045
 },
 "S7sh11::signage": {
  "fp": "067442e862b06026",
  "inscriptions": [],
  "cues": [],
  "dropped": []
 },
 "S7sh11": {
  "input_fingerprint": "6312484c935aed3f",
  "prompt": "Create ONE FINAL photorealistic live-action film still of the moment below — early 20th-century Korea, Jeolla region; all people are Korean unless explicitly stated otherwise. TIME OF DAY (lock): late afternoon.\n\nSHOT TEXT (authoritative, Korean): 일행이 모두 빠져나간 텅 빈 마당 한가운데 덩그러니 남겨진 나무 지게.\n\nLOCATION (lock): The empty center of a traditional estate's dirt courtyard, with a lone wooden carrying frame left behind. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: wide shot\n- FRAME LAYOUT: empty courtyard surrounding the solitary carrying frame in the middle-center of the frame, midground.\n- KEY BACKGROUND ELEMENTS: 나무 지게 (Left alone in the middle of the empty courtyard) — Seen obliquely from above, with its carrying structure readable at a natural scale; used as Serves as the small central narrative anchor amid negative space; 마당 (Empty after the group has departed); used as Provides uninterrupted surrounding space that makes the absence of its owner visible; 마루 (Visible beyond the courtyard with no person in frame) — Its courtyard-facing edge lies across the distant upper portion of the composition; used as Supplies a restrained architectural scale reference.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Unchanged natural late-afternoon light and restrained saturation leave the empty courtyard sober and unadorned rather than artificially darkened.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The people visible in that photo are NOT in this shot: never carry their faces, bodies or clothing onto anyone here. Each person in THIS shot is defined solely by the PEOPLE list, their own CHARACTER REFERENCE images and their wardrobe notes. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The wooden carrying frame is left alone in the middle of the yard.\n\nNO PEOPLE IN THIS SHOT: the shot text shows only the place and its state — no living person appears in frame, and no face, and no figure standing, sitting or passing through. Do not invent someone to occupy the space.\n\nOne thing this rule does NOT forbid: the hand or arm of whoever is doing what the shot text describes. If the moment is an object being held, carried, operated, read, aimed, worn or handed over, the hand doing it belongs in frame, resting on or gripping that object exactly as its real use requires. An object photographed floating unsupported, with the hand that must be holding it erased, is a worse failure than the person you were avoiding — render the hand.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.",
  "roll_prompts": {
   "A": "Create ONE FINAL photorealistic live-action film still of the moment below — early 20th-century Korea, Jeolla region; all people are Korean unless explicitly stated otherwise. TIME OF DAY (lock): late afternoon.\n\nSHOT TEXT (authoritative, Korean): 일행이 모두 빠져나간 텅 빈 마당 한가운데 덩그러니 남겨진 나무 지게.\n\nLOCATION (lock): The empty center of a traditional estate's dirt courtyard, with a lone wooden carrying frame left behind. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: wide shot\n- FRAME LAYOUT: empty courtyard surrounding the solitary carrying frame in the middle-center of the frame, midground.\n- KEY BACKGROUND ELEMENTS: 나무 지게 (Left alone in the middle of the empty courtyard) — Seen obliquely from above, with its carrying structure readable at a natural scale; used as Serves as the small central narrative anchor amid negative space; 마당 (Empty after the group has departed); used as Provides uninterrupted surrounding space that makes the absence of its owner visible; 마루 (Visible beyond the courtyard with no person in frame) — Its courtyard-facing edge lies across the distant upper portion of the composition; used as Supplies a restrained architectural scale reference.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Unchanged natural late-afternoon light and restrained saturation leave the empty courtyard sober and unadorned rather than artificially darkened.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The people visible in that photo are NOT in this shot: never carry their faces, bodies or clothing onto anyone here. Each person in THIS shot is defined solely by the PEOPLE list, their own CHARACTER REFERENCE images and their wardrobe notes. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The wooden carrying frame is left alone in the middle of the yard.\n\nNO PEOPLE IN THIS SHOT: the shot text shows only the place and its state — no living person appears in frame, and no face, and no figure standing, sitting or passing through. Do not invent someone to occupy the space.\n\nOne thing this rule does NOT forbid: the hand or arm of whoever is doing what the shot text describes. If the moment is an object being held, carried, operated, read, aimed, worn or handed over, the hand doing it belongs in frame, resting on or gripping that object exactly as its real use requires. An object photographed floating unsupported, with the hand that must be holding it erased, is a worse failure than the person you were avoiding — render the hand.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.",
   "B": "Create ONE FINAL photorealistic live-action film still of the moment below — early 20th-century Korea, Jeolla region; all people are Korean unless explicitly stated otherwise. TIME OF DAY (lock): late afternoon.\n\nSHOT TEXT (authoritative, Korean): 일행이 모두 빠져나간 텅 빈 마당 한가운데 덩그러니 남겨진 나무 지게.\n\nLOCATION (lock): The empty center of a traditional estate's dirt courtyard, with a lone wooden carrying frame left behind. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: wide shot\n- FRAME LAYOUT: empty courtyard surrounding the solitary carrying frame in the middle-center of the frame, midground.\n- KEY BACKGROUND ELEMENTS: 나무 지게 (Left alone in the middle of the empty courtyard) — Seen obliquely from above, with its carrying structure readable at a natural scale; used as Serves as the small central narrative anchor amid negative space; 마당 (Empty after the group has departed); used as Provides uninterrupted surrounding space that makes the absence of its owner visible; 마루 (Visible beyond the courtyard with no person in frame) — Its courtyard-facing edge lies across the distant upper portion of the composition; used as Supplies a restrained architectural scale reference.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Unchanged natural late-afternoon light and restrained saturation leave the empty courtyard sober and unadorned rather than artificially darkened.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The people visible in that photo are NOT in this shot: never carry their faces, bodies or clothing onto anyone here. Each person in THIS shot is defined solely by the PEOPLE list, their own CHARACTER REFERENCE images and their wardrobe notes. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The wooden carrying frame is left alone in the middle of the yard.\n\nNO PEOPLE IN THIS SHOT: the shot text shows only the place and its state — no living person appears in frame, and no face, and no figure standing, sitting or passing through. Do not invent someone to occupy the space.\n\nOne thing this rule does NOT forbid: the hand or arm of whoever is doing what the shot text describes. If the moment is an object being held, carried, operated, read, aimed, worn or handed over, the hand doing it belongs in frame, resting on or gripping that object exactly as its real use requires. An object photographed floating unsupported, with the hand that must be holding it erased, is a worse failure than the person you were avoiding — render the hand.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.\n\nCOMPOSITION VARIATION (this candidate only): keep the place, the people and the moment exactly as specified, and keep from the CAMERA & FRAME contract its subject scale, screen placement and key background elements exactly. For this candidate only, its camera angle and height are deliberately left open: choose a camera position distinctly different from the obvious one, as a film director picking a second setup on the same blocking. Stay true to any explicit framing the SHOT TEXT itself dictates."
  },
  "shot_run_spend_attempt_count": 1,
  "readings": [
   {
    "label": "A",
    "direction": "카메라는 마당 한가운데 놓인 나무 지게를 비스듬히 내려다보고 있습니다.",
    "built_space": "흙마당을 배경으로 하고 있으나, 레퍼런스와 달리 왼쪽에 돌담이 있고 오른쪽에 마루가 배치되어 공간의 구조가 변형되었습니다.",
    "entities": "평평한 바닥 구조를 가진 나무 지게 1개가 있습니다. 인물은 존재하지 않습니다.",
    "hard_violations": [],
    "physics": "지게가 평평한 썰매 형태의 하단 받침을 통해 바닥에 놓여 지탱되고 있습니다."
   },
   {
    "label": "B",
    "direction": "카메라는 마당 중앙에 세워진 나무 지게를 약간 위에서 비스듬히 내려다보고 있습니다.",
    "built_space": "레퍼런스와 동일한 구조의 흙마당입니다. 왼쪽에 마루가 있는 전통 가옥이 있고, 오른쪽 배경에 열린 대문이 배치되어 있습니다. 공간의 고정된 특징과 방향이 레퍼런스와 정확히 일치합니다.",
    "entities": "나무 지게 1개가 중앙에 있습니다. 지시대로 사람은 아무도 없습니다.",
    "hard_violations": [],
    "physics": "지게는 두 개의 앞다리와 뒤쪽의 지지대를 이용해 흙바닥 위에 안정적으로 서 있습니다."
   }
  ],
  "cross_model_order": {
   "policy": "cross_model_order_v2_slot0_fwd_slot1_rev_parallel2",
   "slots": [
    {
     "model": "gemini-pro",
     "order": "forward",
     "display_to_canonical": {
      "A": "A",
      "B": "B"
     },
     "raw": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 10,
        "verdict_ko": "레퍼런스 이미지의 공간 구조(왼쪽의 마루, 오른쪽의 대문)와 바닥의 끌린 자국까지 완벽하게 유지하면서, 지시된 대로 마당 한가운데 덩그러니 놓인 지게를 인물 없이 훌륭하게 묘사했습니다."
       },
       {
        "label": "A",
        "score": 6,
        "verdict_ko": "지게와 빈 마당을 표현했으나, 레퍼런스와 비교해 건축물의 위치(마루가 오른쪽, 돌담이 왼쪽으로 변경됨)가 달라져 장소의 연속성이 떨어지며 지게의 형태도 다소 이질적입니다."
       }
      ],
      "readings": [
       {
        "label": "B",
        "direction": "카메라는 마당 중앙에 세워진 나무 지게를 약간 위에서 비스듬히 내려다보고 있습니다.",
        "built_space": "레퍼런스와 동일한 구조의 흙마당입니다. 왼쪽에 마루가 있는 전통 가옥이 있고, 오른쪽 배경에 열린 대문이 배치되어 있습니다. 공간의 고정된 특징과 방향이 레퍼런스와 정확히 일치합니다.",
        "entities": "나무 지게 1개가 중앙에 있습니다. 지시대로 사람은 아무도 없습니다.",
        "hard_violations": [],
        "physics": "지게는 두 개의 앞다리와 뒤쪽의 지지대를 이용해 흙바닥 위에 안정적으로 서 있습니다."
       },
       {
        "label": "A",
        "direction": "카메라는 마당 한가운데 놓인 나무 지게를 비스듬히 내려다보고 있습니다.",
        "built_space": "흙마당을 배경으로 하고 있으나, 레퍼런스와 달리 왼쪽에 돌담이 있고 오른쪽에 마루가 배치되어 공간의 구조가 변형되었습니다.",
        "entities": "평평한 바닥 구조를 가진 나무 지게 1개가 있습니다. 인물은 존재하지 않습니다.",
        "hard_violations": [],
        "physics": "지게가 평평한 썰매 형태의 하단 받침을 통해 바닥에 놓여 지탱되고 있습니다."
       }
      ],
      "all_candidates_fail": false
     },
     "normalized": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 10,
        "verdict_ko": "레퍼런스 이미지의 공간 구조(왼쪽의 마루, 오른쪽의 대문)와 바닥의 끌린 자국까지 완벽하게 유지하면서, 지시된 대로 마당 한가운데 덩그러니 놓인 지게를 인물 없이 훌륭하게 묘사했습니다."
       },
       {
        "label": "A",
        "score": 6,
        "verdict_ko": "지게와 빈 마당을 표현했으나, 레퍼런스와 비교해 건축물의 위치(마루가 오른쪽, 돌담이 왼쪽으로 변경됨)가 달라져 장소의 연속성이 떨어지며 지게의 형태도 다소 이질적입니다."
       }
      ],
      "all_candidates_fail": false,
      "readings": [
       {
        "label": "B",
        "direction": "카메라는 마당 중앙에 세워진 나무 지게를 약간 위에서 비스듬히 내려다보고 있습니다.",
        "built_space": "레퍼런스와 동일한 구조의 흙마당입니다. 왼쪽에 마루가 있는 전통 가옥이 있고, 오른쪽 배경에 열린 대문이 배치되어 있습니다. 공간의 고정된 특징과 방향이 레퍼런스와 정확히 일치합니다.",
        "entities": "나무 지게 1개가 중앙에 있습니다. 지시대로 사람은 아무도 없습니다.",
        "hard_violations": [],
        "physics": "지게는 두 개의 앞다리와 뒤쪽의 지지대를 이용해 흙바닥 위에 안정적으로 서 있습니다."
       },
       {
        "label": "A",
        "direction": "카메라는 마당 한가운데 놓인 나무 지게를 비스듬히 내려다보고 있습니다.",
        "built_space": "흙마당을 배경으로 하고 있으나, 레퍼런스와 달리 왼쪽에 돌담이 있고 오른쪽에 마루가 배치되어 공간의 구조가 변형되었습니다.",
        "entities": "평평한 바닥 구조를 가진 나무 지게 1개가 있습니다. 인물은 존재하지 않습니다.",
        "hard_violations": [],
        "physics": "지게가 평평한 썰매 형태의 하단 받침을 통해 바닥에 놓여 지탱되고 있습니다."
       }
      ]
     },
     "ok": true
    },
    {
     "model": "gpt-high",
     "order": "reverse",
     "display_to_canonical": {
      "A": "B",
      "B": "A"
     },
     "raw": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 8,
        "verdict_ko": "비스듬한 부감과 상단의 빈 마루, 열린 대문 주변의 장소 연속성이 지시에 더 충실하지만, 지게는 작은 중경의 서사적 중심치고 다소 크다."
       },
       {
        "label": "B",
        "score": 6,
        "verdict_ko": "빈 마당과 홀로 놓인 지게는 구현했으나, 낮은 시점과 크게 드러난 지붕들이 부감·여백 중심 구도를 약화하고 이전 장소와의 연결도 덜 명확하다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "사람이나 시선, 무기, 이동체는 없다. 지게의 멜빵이 달린 면이 카메라 쪽을 비스듬히 향하고, 뒤쪽 버팀대는 화면 오른쪽으로 벌어져 있다. 카메라는 지게와 흙바닥을 위에서 내려다본다.",
        "built_space": "상단 왼쪽에 긴 목조 마루 한 줄과 격자문들이 있고, 마루 앞에는 직사각형 디딤돌 두 개가 보인다. 상단 중앙에는 기와를 얹은 돌담, 오른쪽에는 열린 대문 한 곳과 그 양옆의 닫힌 판문 두 곳이 있다. 참고 이미지의 마루·짧은 돌담·열린 대문 관계와 재료가 잘 이어진다. 추가로 보이는 디딤돌은 더 넓게 드러난 마루 구간에 있어 중복 시설로 단정할 수 없다. 마당에는 사람이 없고 반사면도 없다.",
        "entities": "나무 지게 한 개가 화면 중앙 부근에 있으며, 나무 세로대와 가로대, 등판, 굵은 섬유 멜빵이 보인다. 빈 흙마당과 먼 마루가 모두 포함된다. 사람·얼굴·읽을 수 있는 글자는 없다. 목재와 흙의 마모, 낮은 채도의 따뜻한 오후 빛은 요구와 부합한다.",
        "hard_violations": [],
        "physics": "지게의 앞쪽 두 다리와 오른쪽 뒤 버팀대 끝이 흙바닥에 닿아 삼각형 지지 구조를 이룬다. 멜빵은 위쪽 나무대에 걸려 아래로 처지고, 지게의 긴 그림자도 지면에 이어진다. 공중에 떠 있거나 지지 없이 매달린 물체는 없다."
       },
       {
        "label": "B",
        "direction": "사람이나 시선, 무기, 이동체는 없다. 지게의 판재 면과 끈이 카메라를 향해 비스듬히 드러나고, 긴 세로대는 화면 왼쪽으로 기울어 올라간다. 바닥은 내려다보이지만 A보다 시선이 낮아 지게를 위에서 관찰하는 효과가 약하다.",
        "built_space": "오른쪽에는 목조 마루 한 줄과 여러 기둥, 디딤돌 한 개가 보인다. 중앙 배경에는 긴 기와 돌담이 이어지고 왼쪽 끝에는 닫힌 양짝 문 한 곳이 있다. 담 뒤로 여러 지붕이 크게 드러난다. 전통 목재·회벽·돌담 재료는 참고와 유사하지만, 참고의 열린 대문과 인접 건물 관계는 이 화면에서 확인되지 않는다. 마루 가장자리는 상단 오른쪽에만 놓이고 지붕과 하늘이 상당한 면적을 차지한다. 사람과 반사면은 없다.",
        "entities": "나무 지게 한 개에 긴 세로대 두 개, 가로대, 판재 등판과 매달린 끈들이 보인다. 주변은 빈 흙마당이고 오른쪽 마루에도 사람이 없다. 얼굴이나 읽을 수 있는 글자는 없다. 지게와 건축 재료는 실물처럼 보이며 따뜻한 오후 광선도 유지된다.",
        "hard_violations": [],
        "physics": "지게 하단의 나무 받침과 왼쪽 뒤로 나온 짧은 지지부가 흙바닥에 닿아 있다. 긴 틀은 그 받침 위에서 기울어진 상태이며, 끈은 세로대에 묶여 늘어지고 일부는 바닥 가까이에 놓인다. 지지가 전혀 없는 부유 상태로 보이지는 않는다."
       }
      ],
      "all_candidates_fail": false
     },
     "normalized": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 8,
        "verdict_ko": "비스듬한 부감과 상단의 빈 마루, 열린 대문 주변의 장소 연속성이 지시에 더 충실하지만, 지게는 작은 중경의 서사적 중심치고 다소 크다."
       },
       {
        "label": "A",
        "score": 6,
        "verdict_ko": "빈 마당과 홀로 놓인 지게는 구현했으나, 낮은 시점과 크게 드러난 지붕들이 부감·여백 중심 구도를 약화하고 이전 장소와의 연결도 덜 명확하다."
       }
      ],
      "all_candidates_fail": false,
      "readings": [
       {
        "label": "B",
        "direction": "사람이나 시선, 무기, 이동체는 없다. 지게의 멜빵이 달린 면이 카메라 쪽을 비스듬히 향하고, 뒤쪽 버팀대는 화면 오른쪽으로 벌어져 있다. 카메라는 지게와 흙바닥을 위에서 내려다본다.",
        "built_space": "상단 왼쪽에 긴 목조 마루 한 줄과 격자문들이 있고, 마루 앞에는 직사각형 디딤돌 두 개가 보인다. 상단 중앙에는 기와를 얹은 돌담, 오른쪽에는 열린 대문 한 곳과 그 양옆의 닫힌 판문 두 곳이 있다. 참고 이미지의 마루·짧은 돌담·열린 대문 관계와 재료가 잘 이어진다. 추가로 보이는 디딤돌은 더 넓게 드러난 마루 구간에 있어 중복 시설로 단정할 수 없다. 마당에는 사람이 없고 반사면도 없다.",
        "entities": "나무 지게 한 개가 화면 중앙 부근에 있으며, 나무 세로대와 가로대, 등판, 굵은 섬유 멜빵이 보인다. 빈 흙마당과 먼 마루가 모두 포함된다. 사람·얼굴·읽을 수 있는 글자는 없다. 목재와 흙의 마모, 낮은 채도의 따뜻한 오후 빛은 요구와 부합한다.",
        "hard_violations": [],
        "physics": "지게의 앞쪽 두 다리와 오른쪽 뒤 버팀대 끝이 흙바닥에 닿아 삼각형 지지 구조를 이룬다. 멜빵은 위쪽 나무대에 걸려 아래로 처지고, 지게의 긴 그림자도 지면에 이어진다. 공중에 떠 있거나 지지 없이 매달린 물체는 없다."
       },
       {
        "label": "A",
        "direction": "사람이나 시선, 무기, 이동체는 없다. 지게의 판재 면과 끈이 카메라를 향해 비스듬히 드러나고, 긴 세로대는 화면 왼쪽으로 기울어 올라간다. 바닥은 내려다보이지만 A보다 시선이 낮아 지게를 위에서 관찰하는 효과가 약하다.",
        "built_space": "오른쪽에는 목조 마루 한 줄과 여러 기둥, 디딤돌 한 개가 보인다. 중앙 배경에는 긴 기와 돌담이 이어지고 왼쪽 끝에는 닫힌 양짝 문 한 곳이 있다. 담 뒤로 여러 지붕이 크게 드러난다. 전통 목재·회벽·돌담 재료는 참고와 유사하지만, 참고의 열린 대문과 인접 건물 관계는 이 화면에서 확인되지 않는다. 마루 가장자리는 상단 오른쪽에만 놓이고 지붕과 하늘이 상당한 면적을 차지한다. 사람과 반사면은 없다.",
        "entities": "나무 지게 한 개에 긴 세로대 두 개, 가로대, 판재 등판과 매달린 끈들이 보인다. 주변은 빈 흙마당이고 오른쪽 마루에도 사람이 없다. 얼굴이나 읽을 수 있는 글자는 없다. 지게와 건축 재료는 실물처럼 보이며 따뜻한 오후 광선도 유지된다.",
        "hard_violations": [],
        "physics": "지게 하단의 나무 받침과 왼쪽 뒤로 나온 짧은 지지부가 흙바닥에 닿아 있다. 긴 틀은 그 받침 위에서 기울어진 상태이며, 끈은 세로대에 묶여 늘어지고 일부는 바닥 가까이에 놓인다. 지지가 전혀 없는 부유 상태로 보이지는 않는다."
       }
      ]
     },
     "ok": true
    }
   ],
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "slot_winner_match": true,
   "slot_winner": {
    "gemini-pro": "B",
    "gpt-high": "B"
   },
   "route": "cross_slot_agree"
  },
  "dual": {
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "normalized": {
    "A": 1.35,
    "B": 2.0
   },
   "adjusted": {
    "A": 1.35,
    "B": 2.0
   },
   "violations": {},
   "per_model_winner": {
    "gemini-pro": "B",
    "gpt-high": "B"
   },
   "agreed": true
  },
  "totals": {
   "B": 2000,
   "A": 1350
  },
  "selected": "B",
  "ranking": [
   "B",
   "A"
  ],
  "verdicts": [
   {
    "label": "B",
    "score": 2000,
    "verdict_ko": "레퍼런스 이미지의 공간 구조(왼쪽의 마루, 오른쪽의 대문)와 바닥의 끌린 자국까지 완벽하게 유지하면서, 지시된 대로 마당 한가운데 덩그러니 놓인 지게를 인물 없이 훌륭하게 묘사했습니다."
   },
   {
    "label": "A",
    "score": 1350,
    "verdict_ko": "지게와 빈 마당을 표현했으나, 레퍼런스와 비교해 건축물의 위치(마루가 오른쪽, 돌담이 왼쪽으로 변경됨)가 달라져 장소의 연속성이 떨어지며 지게의 형태도 다소 이질적입니다."
   }
  ],
  "refs": [
   {
    "label": "PREVIOUS SHOT STILL — a visually related earlier shot of this same place: the location's look, materials, fixed features and lighting mood are LOCKED to this photo; never copy its camera framing. The people visible in that photo are NOT in this shot: never carry their faces, bodies or clothing onto anyone here. Each person in THIS shot is defined solely by the PEOPLE list, their own CHARACTER REFERENCE images and their wardrobe notes. If this photo shows a character who cannot move (dead or unconscious), that character's exact pose, position and orientation are ALSO LOCKED — treat their body as a fixed prop of the set that only the camera moves around.",
    "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/db00517b-165b-4005-a287-f28936b59b95/images/272bd7af-9294-4490-98c6-88a40e9deacc/scene/recipe/S7sh10_sel.png",
    "asset_id": "f8cfdf72-1c84-478e-b1f2-755fba463c0d",
    "role": "prev_still"
   }
  ],
  "critique_skipped": true,
  "shot_run_produced": false,
  "shot_run_uid": "06aa1407-044e-7d80-9246-934e56e885fd",
  "ref_mode": "prev만 (배경 전용·공유 계획)",
  "share_plan": {
   "ref_plan": "prev",
   "prev_anchor_tag": "S7sh10"
  },
  "lane_policy": "share_plan_prev_bgonly"
 },
 "S7sh11::cine": {
  "applied": true,
  "attempted_at": "2026-09-09T01:21:24.732643+00:00",
  "fingerprint": "ce7edc9c1235aaf9432e291895e99474e16ff8fc3f5bf332911f7268c6a87e02",
  "fingerprint_version": 2,
  "provider": "mai",
  "endpoint": "openrouter/chat-completions",
  "model": "microsoft/mai-image-2.6",
  "pack": "24.202608252115",
  "source_file": "S7sh11_sel.png",
  "source_sha256": "1d0fedbba67a52b18f1316a34724928ffcf676b25bb190f1cd7666b950699c09",
  "file": "S7sh11_cine.png",
  "staged_sha256": "03af9ceb836154f218daf133bcdf61ba49e5ea25c546b21ff9ab4f807117474c",
  "latency_ms": 24183
 },
 "S8sh5::signage": {
  "fp": "de0f42e79f73c0da",
  "inscriptions": [],
  "cues": [],
  "dropped": []
 },
 "era_assess::50040b24a0d6ba36": {
  "subjects": [
   {
    "subject_native": "20세기 초 전라도 초가 농가의 장독대",
    "search_terms_native": [
     "전라도 초가 장독대",
     "구한말 농가 옹기 장독",
     "전통 초가집 마당 장독대"
    ],
    "language_lock_native": "다른 언어로 번역하거나 외래 검색어를 덧붙이지 말고 오직 한국어로만 검색하십시오.",
    "reason_ko": "AI는 기와 한옥이나 현대 민속촌의 정갈한 형태를 재현하기 쉬우나, 20세기 초 남도 서민 농가의 장독대는 투박한 옹기와 낮은 돌받침으로 구성됩니다."
   }
  ],
  "subject_text": "안규홍의 집 마당·사립문 안쪽·장독대\n사립문으로 둘러진 소박한 흙마당. 한쪽에 장독대와 옹기들이 모여 있고, 부엌과 방의 출입구가 마당에 면한다.",
  "identity": "canonical",
  "scope_id": "L05",
  "scope_role": "location_exterior",
  "scope_sha": "90e0100827204b21"
 },
 "era_fail::1063bd2a0f0b0713": {
  "stage": "research",
  "subject": "20세기 초 전라도 초가 농가의 장독대",
  "terms": [
   "전라도 초가 장독대",
   "구한말 농가 옹기 장독",
   "전통 초가집 마당 장독대"
  ],
  "status": "no_usable",
  "queries": [
   [
    "전라도 초가 농가 장독대 민속마을"
   ]
  ],
  "candidate_urls": [
   "https://ojsfile.ohmynews.com/STD_IMG_FILE/2026/0115/IE003571206_STD.jpg",
   "https://www.regent-holidays.co.uk/upload-files/product-excursions-gallery/discover-suwon-korean-folk-village_3_korean-folk-village.jpg",
   "https://seoulmateskorea.files.wordpress.com/2013/10/5b331-p1040227.jpg",
   "https://gongu.copyright.or.kr/gongu/wrt/cmmn/wrtFileImageView.do?filePath=L2Rpc2sxL25ld2RhdGEvMjAxNC8yMS9DTFM2L2RpZ2lfMTEwMjExNzhfMDEyMDE0MTAxMA%3D%3D&thumbAt=Y&thumbSe=b_tbumb&wrtSn=11021178&wrtTy=10006"
  ],
  "coarse": {
   "eligible": [],
   "chosen_index": 0,
   "reason": "종류·보임·기준을 다 만족하는 후보가 없다 — 이 라운드에선 안 고른다",
   "single_judge": true,
   "rejected_judges": {}
  },
  "verdicts": [
   {
    "index": 1,
    "object_type_match": "yes",
    "visible": true,
    "criteria_match": "unsure",
    "similarity": 80
   },
   {
    "index": 2,
    "object_type_match": "yes",
    "visible": true,
    "criteria_match": "unsure",
    "similarity": 100
   },
   {
    "index": 3,
    "object_type_match": "yes",
    "visible": true,
    "criteria_match": "unsure",
    "similarity": 30
   },
   {
    "index": 4,
    "object_type_match": "yes",
    "visible": true,
    "criteria_match": "unsure",
    "similarity": 100
   }
  ],
  "chosen_reason_ko": "",
  "attempts": 5
 },
 "S8sh5::bgfirst_bg": {
  "input_fingerprint": "d422e1e26f44bf41",
  "prompt": "Create the EMPTY BACKGROUND PLATE for one film shot — NO PEOPLE, no figures, no body parts, no sketch lines, no arrows anywhere.\n\"Empty\" means no people only: KEEP the location's inherent occupants and stock that define the place — animals in an animal shelter, pen or farm, goods in a market, moored boats in a harbour — unless the shot text explicitly removes them.\nThe FIRST attached image is a thin-line storyboard sketch: use ONLY its camera angle, horizon, perspective and the placement/size of buildings and set masses — ignore the sketched people and arrows entirely. The SECOND attached image (LOCATION PHOTOGRAPH) is the real place: take its architecture, materials, signage and fixed features, and RE-PROJECT them into the sketch's camera. If the photograph's camera differs from the sketch's, the sketch's camera wins.\nHUMAN-SCALE CALIBRATION: derive every structure's true size from human-scale elements — a door ≈ 2m, a window ≈ 1–1.5m wide, one storey ≈ 2.5–3m; never inflate a small structure or shrink a large one.\n\nSHOT TEXT this background must serve (Korean): 반쯤 열린 사립문 틈으로 어두운 표정의 달래가 마당에 들어선 전신.\n\nLOCATION (lock): The inner side of a half-open rustic wicket gate leading into a modest home's darkening outdoor yard.\n\nTIME OF DAY (lock): dusk.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: wide shot\n- KEY BACKGROUND ELEMENTS: 사립문 (Half-open as 달래 enters) — The gate is seen obliquely from its courtyard side, leaving the entrance opening readable; used as Provides a narrow frame beside her full figure without masking her expression; 마당 (달래 is entering from the gate); used as Open ground ahead of her establishes the route toward the kitchen.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Subdued ambient light appropriate to the darkening setting preserves facial legibility with muted saturation and restrained contrast.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nRender ONE photorealistic empty location photograph, 16:9, that this shot can be staged inside later. No readable writing anywhere: surfaces that would carry writing may be present, but stage any wording out of legibility — an oblique angle, distance, shallow focus. No captions, watermarks or overlay text.\n\nThe THIRD attached image (STRUCTURE LOOK) is the identity source of the fixed structure at this location: its faces, openings, levels, materials and signage are truth. Where it conflicts with the LOCATION PHOTOGRAPH about the structure itself, the STRUCTURE LOOK wins; the photograph still governs the surroundings, time of day and lighting.",
  "effective_prompt": "Create the EMPTY BACKGROUND PLATE for one film shot — NO PEOPLE, no figures, no body parts, no sketch lines, no arrows anywhere.\n\"Empty\" means no people only: KEEP the location's inherent occupants and stock that define the place — animals in an animal shelter, pen or farm, goods in a market, moored boats in a harbour — unless the shot text explicitly removes them.\nThe FIRST attached image is a thin-line storyboard sketch: use ONLY its camera angle, horizon, perspective and the placement/size of buildings and set masses — ignore the sketched people and arrows entirely. The SECOND attached image (LOCATION PHOTOGRAPH) is the real place: take its architecture, materials, signage and fixed features, and RE-PROJECT them into the sketch's camera. If the photograph's camera differs from the sketch's, the sketch's camera wins.\nHUMAN-SCALE CALIBRATION: derive every structure's true size from human-scale elements — a door ≈ 2m, a window ≈ 1–1.5m wide, one storey ≈ 2.5–3m; never inflate a small structure or shrink a large one.\n\nSHOT TEXT this background must serve (Korean): 반쯤 열린 사립문 틈으로 어두운 표정의 달래가 마당에 들어선 전신.\n\nLOCATION (lock): The inner side of a half-open rustic wicket gate leading into a modest home's darkening outdoor yard.\n\nTIME OF DAY (lock): dusk.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: wide shot\n- KEY BACKGROUND ELEMENTS: 사립문 (Half-open as 달래 enters) — The gate is seen obliquely from its courtyard side, leaving the entrance opening readable; used as Provides a narrow frame beside her full figure without masking her expression; 마당 (달래 is entering from the gate); used as Open ground ahead of her establishes the route toward the kitchen.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Subdued ambient light appropriate to the darkening setting preserves facial legibility with muted saturation and restrained contrast.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nRender ONE photorealistic empty location photograph, 16:9, that this shot can be staged inside later. No readable writing anywhere: surfaces that would carry writing may be present, but stage any wording out of legibility — an oblique angle, distance, shallow focus. No captions, watermarks or overlay text.\n\nThe THIRD attached image (STRUCTURE LOOK) is the identity source of the fixed structure at this location: its faces, openings, levels, materials and signage are truth. Where it conflicts with the LOCATION PHOTOGRAPH about the structure itself, the STRUCTURE LOOK wins; the photograph still governs the surroundings, time of day and lighting.",
  "bg_path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/db00517b-165b-4005-a287-f28936b59b95/images/272bd7af-9294-4490-98c6-88a40e9deacc/scene/recipe/S8sh5__bgfirst_bg.png",
  "asset_id": "b99aae16-b8a5-4629-9e89-1289d0ed1546",
  "input_asset_ids": [
   "3d31da46-a00c-40fc-abe5-d1e93ac679d0",
   "7cc7dd4c-b9eb-4ed0-a5b4-ce58472b4a3f",
   "2dc88acf-a75f-4d7e-973d-10ab234660a8"
  ]
 },
 "S8sh5": {
  "input_fingerprint": "0b2134687335d75d",
  "prompt": "Create ONE FINAL photorealistic live-action film still of the moment below — early 20th-century Korea, Jeolla region; all people are Korean unless explicitly stated otherwise. TIME OF DAY (lock): dusk.\n\nSHOT TEXT (authoritative, Korean): 반쯤 열린 사립문 틈으로 어두운 표정의 달래가 마당에 들어선 전신.\n\nLOCATION (lock): The inner side of a half-open rustic wicket gate leading into a modest home's darkening outdoor yard. The shot takes place here — the attached LOCATION PHOTOGRAPH shows the exact spot.\n\nSTRUCTURE LOOK AUTHORITY: the attached STRUCTURE LOOK photograph is the identity of the fixed structure at this location — wherever that structure appears in the frame, its shape, proportions, openings, materials and colors are LOCKED to it. The LOCATION PHOTOGRAPH remains the authority for this shot's sub-space, surroundings, time of day and lighting. If the two conflict on the structure itself, the STRUCTURE LOOK photo wins; for everything else, the LOCATION PHOTOGRAPH wins.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: wide shot\n- KEY BACKGROUND ELEMENTS: 사립문 (Half-open as 달래 enters) — The gate is seen obliquely from its courtyard side, leaving the entrance opening readable; used as Provides a narrow frame beside her full figure without masking her expression; 마당 (달래 is entering from the gate); used as Open ground ahead of her establishes the route toward the kitchen.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Subdued ambient light appropriate to the darkening setting preserves facial legibility with muted saturation and restrained contrast.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The kitchen is dim; smoke rises from the unlit hearth, and the gate has opened. 달래: She has entered through the gate.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 달래 (한국인 여성, 젊은 성인, 자연스러운 검은 머리) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.",
  "roll_prompts": {
   "A": "Stage the shot. The FIRST attached image (SHOT BACKGROUND) is the finished empty background of this shot — keep it EXACTLY: its camera, perspective, architecture, lighting and every fixed feature stay untouched. The SECOND attached image (LAYOUT SKETCH) tells you ONLY where the people go: each sketched person's position, screen size, pose and the gaze/motion arrows. Ignore the sketch's background lines. The CHARACTER REFERENCE photographs show the real people.\nPlace the real people into the background at exactly the sketched positions, sizes and poses, following the arrow directions. No sketch lines or arrows may remain.\n\nCreate ONE FINAL photorealistic live-action film still of the moment below — early 20th-century Korea, Jeolla region; all people are Korean unless explicitly stated otherwise. TIME OF DAY (lock): dusk.\n\nSHOT TEXT (authoritative, Korean): 반쯤 열린 사립문 틈으로 어두운 표정의 달래가 마당에 들어선 전신.\n\nLOCATION (lock): The inner side of a half-open rustic wicket gate leading into a modest home's darkening outdoor yard. The shot takes place here — the FIRST attached image (SHOT BACKGROUND) is this exact place, already built: its ground, structures, horizon, materials and lighting are the finished truth of this location and must not be redesigned or replaced. No location photograph is attached — read the place from that image alone, and add no scenery, structure, vehicle or fixture that it does not already show. This lock governs the place only; the figures in the shot follow the staging and pose instructions.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: wide shot\n- KEY BACKGROUND ELEMENTS: 사립문 (Half-open as 달래 enters) — The gate is seen obliquely from its courtyard side, leaving the entrance opening readable; used as Provides a narrow frame beside her full figure without masking her expression; 마당 (달래 is entering from the gate); used as Open ground ahead of her establishes the route toward the kitchen.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Subdued ambient light appropriate to the darkening setting preserves facial legibility with muted saturation and restrained contrast.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The kitchen is dim; smoke rises from the unlit hearth, and the gate has opened. 달래: She has entered through the gate.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 달래 (한국인 여성, 젊은 성인, 자연스러운 검은 머리) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.",
   "B": "Create ONE FINAL photorealistic live-action film still of the moment below — early 20th-century Korea, Jeolla region; all people are Korean unless explicitly stated otherwise. TIME OF DAY (lock): dusk.\n\nSHOT TEXT (authoritative, Korean): 반쯤 열린 사립문 틈으로 어두운 표정의 달래가 마당에 들어선 전신.\n\nLOCATION (lock): The inner side of a half-open rustic wicket gate leading into a modest home's darkening outdoor yard. The shot takes place here — the attached LOCATION PHOTOGRAPH shows the exact spot.\n\nSTRUCTURE LOOK AUTHORITY: the attached STRUCTURE LOOK photograph is the identity of the fixed structure at this location — wherever that structure appears in the frame, its shape, proportions, openings, materials and colors are LOCKED to it. The LOCATION PHOTOGRAPH remains the authority for this shot's sub-space, surroundings, time of day and lighting. If the two conflict on the structure itself, the STRUCTURE LOOK photo wins; for everything else, the LOCATION PHOTOGRAPH wins.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: wide shot\n- KEY BACKGROUND ELEMENTS: 사립문 (Half-open as 달래 enters) — The gate is seen obliquely from its courtyard side, leaving the entrance opening readable; used as Provides a narrow frame beside her full figure without masking her expression; 마당 (달래 is entering from the gate); used as Open ground ahead of her establishes the route toward the kitchen.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Subdued ambient light appropriate to the darkening setting preserves facial legibility with muted saturation and restrained contrast.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The kitchen is dim; smoke rises from the unlit hearth, and the gate has opened. 달래: She has entered through the gate.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 달래 (한국인 여성, 젊은 성인, 자연스러운 검은 머리) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.\n\nCOMPOSITION VARIATION (this candidate only): keep the place, the people and the moment exactly as specified, and keep from the CAMERA & FRAME contract its subject scale, screen placement and key background elements exactly. For this candidate only, its camera angle and height are deliberately left open: choose a camera position distinctly different from the obvious one, as a film director picking a second setup on the same blocking. Stay true to any explicit framing the SHOT TEXT itself dictates."
  },
  "roll_refs": {
   "A": [
    {
     "label": "SHOT BACKGROUND",
     "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/db00517b-165b-4005-a287-f28936b59b95/images/272bd7af-9294-4490-98c6-88a40e9deacc/scene/recipe/S8sh5__bgfirst_bg.png",
     "asset_id": "b99aae16-b8a5-4629-9e89-1289d0ed1546",
     "role": "bgfirst_bg"
    },
    {
     "label": "LAYOUT SKETCH (people placement only)",
     "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/db00517b-165b-4005-a287-f28936b59b95/images/272bd7af-9294-4490-98c6-88a40e9deacc/conti/conti_S8sh5.png",
     "asset_id": "3d31da46-a00c-40fc-abe5-d1e93ac679d0",
     "role": "conti_light"
    },
    {
     "label": "CHARACTER REFERENCE — 달래: the exact person appearing in this shot; match face, hair and build exactly.",
     "path": "<bytes:879791>",
     "asset_id": "83024e69-b375-4f3e-98ec-820d0e2ff1f0",
     "role": "character_ref"
    }
   ],
   "B": [
    {
     "label": "LOCATION PHOTOGRAPH — the exact place of this shot: its spatial layout, surroundings, fixed features, time of day and lighting mood are spatial truth; stage the moment inside this place. If a STRUCTURE LOOK photograph is also attached, that photo wins for the fixed structure itself — this photograph wins for everything around it. Never copy its camera framing.",
     "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/db00517b-165b-4005-a287-f28936b59b95/episodes/272bd7af-9294-4490-98c6-88a40e9deacc/images/background_chain/L05B01.png",
     "asset_id": "7cc7dd4c-b9eb-4ed0-a5b4-ce58472b4a3f",
     "role": "location_plate"
    },
    {
     "label": "STRUCTURE LOOK — the confirmed photograph of the fixed structure at this location: wherever the structure appears in the frame, its shape, proportions, materials, colors and openings are LOCKED to this photo. Never copy its camera framing, time of day or lighting — the shot text and the LOCATION PHOTOGRAPH are the authorities for those.",
     "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/db00517b-165b-4005-a287-f28936b59b95/images/background_chain/seed_bg_thatched_house_sel.png",
     "asset_id": "2dc88acf-a75f-4d7e-973d-10ab234660a8",
     "role": "structure_seed_look"
    },
    {
     "label": "CHARACTER REFERENCE — 달래: the exact person appearing in this shot; match face, hair and build exactly.",
     "path": "<bytes:879791>",
     "asset_id": "83024e69-b375-4f3e-98ec-820d0e2ff1f0",
     "role": "character_ref"
    }
   ]
  },
  "shot_run_spend_attempt_count": 1,
  "judge_flip_skipped": "judge_fn_owns_order",
  "readings": [
   {
    "label": "A",
    "direction": "달래의 시선은 마당 앞쪽 바닥을 향하고 있으며, 문턱을 넘어 마당으로 향하는 방향성을 띰.",
    "built_space": "마당 안쪽에서 대문을 비스듬히 바라보는 앵글이며 우측 단 위에 3개의 장독이 배치됨. 단, 구조물 룩 레퍼런스(초가, 사립문) 대신 로케이션 사진의 기와지붕과 나무 판장문이 묘사됨.",
    "entities": "달래는 연두색 저고리, 붉은 고름, 검은 치마로 인물 레퍼런스와 정확히 일치함. 대문은 지문이 요구한 사립문이 아님.",
    "hard_violations": [],
    "physics": "땅을 딛고 걷는 자세가 바닥에 의해 자연스럽게 지지되고 있음."
   },
   {
    "label": "B",
    "direction": "달래의 시선은 정면 바닥을 향하며 대문 중앙을 통해 안으로 들어오고 있음.",
    "built_space": "초가지붕과 사립문이 정면으로 배치되었고 우측에 3개의 장독이 있음. 대문 밖 배경에 불이 피워진 건물이 보여 공간 배치의 모순이 있음.",
    "entities": "대문은 사립문과 초가 형태로 구조물 레퍼런스와 일치함. 달래의 치마가 회색으로 묘사되어 인물 레퍼런스와 불일치함.",
    "hard_violations": [
     "[gpt-high] 안마당과 주거 건물을 문 너머 배경에 두고 달래가 바깥 문턱으로 걸어오게 하여, 지정된 마당 안쪽 카메라와 마당 진입 동선을 반대로 배치했다."
    ],
    "physics": "두 발이 마당 바닥에 닿아 있어 서 있거나 걷는 자세가 자연스럽게 지지됨."
   }
  ],
  "cross_model_order": {
   "policy": "cross_model_order_v2_slot0_fwd_slot1_rev_parallel2",
   "slots": [
    {
     "model": "gemini-pro",
     "order": "forward",
     "display_to_canonical": {
      "A": "A",
      "B": "B"
     },
     "raw": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 6,
        "verdict_ko": "지시된 비스듬한 앵글과 마당의 공간감, 캐릭터의 복장(검은 치마)을 정확히 구현했으나, 강제된 사립문과 초가지붕 레퍼런스를 기와와 판장문으로 잘못 묘사했습니다."
       },
       {
        "label": "B",
        "score": 4,
        "verdict_ko": "사립문 구조는 레퍼런스를 따랐으나 정면 구도로 촬영되어 카메라 앵글 지시를 어겼으며, 치마 색상(회색) 불일치와 대문 밖에 불이 켜진 부엌이 있는 공간 모순이 발생했습니다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "달래의 시선은 마당 앞쪽 바닥을 향하고 있으며, 문턱을 넘어 마당으로 향하는 방향성을 띰.",
        "built_space": "마당 안쪽에서 대문을 비스듬히 바라보는 앵글이며 우측 단 위에 3개의 장독이 배치됨. 단, 구조물 룩 레퍼런스(초가, 사립문) 대신 로케이션 사진의 기와지붕과 나무 판장문이 묘사됨.",
        "entities": "달래는 연두색 저고리, 붉은 고름, 검은 치마로 인물 레퍼런스와 정확히 일치함. 대문은 지문이 요구한 사립문이 아님.",
        "hard_violations": [],
        "physics": "땅을 딛고 걷는 자세가 바닥에 의해 자연스럽게 지지되고 있음."
       },
       {
        "label": "B",
        "direction": "달래의 시선은 정면 바닥을 향하며 대문 중앙을 통해 안으로 들어오고 있음.",
        "built_space": "초가지붕과 사립문이 정면으로 배치되었고 우측에 3개의 장독이 있음. 대문 밖 배경에 불이 피워진 건물이 보여 공간 배치의 모순이 있음.",
        "entities": "대문은 사립문과 초가 형태로 구조물 레퍼런스와 일치함. 달래의 치마가 회색으로 묘사되어 인물 레퍼런스와 불일치함.",
        "hard_violations": [],
        "physics": "두 발이 마당 바닥에 닿아 있어 서 있거나 걷는 자세가 자연스럽게 지지됨."
       }
      ],
      "all_candidates_fail": false
     },
     "normalized": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 6,
        "verdict_ko": "지시된 비스듬한 앵글과 마당의 공간감, 캐릭터의 복장(검은 치마)을 정확히 구현했으나, 강제된 사립문과 초가지붕 레퍼런스를 기와와 판장문으로 잘못 묘사했습니다."
       },
       {
        "label": "B",
        "score": 4,
        "verdict_ko": "사립문 구조는 레퍼런스를 따랐으나 정면 구도로 촬영되어 카메라 앵글 지시를 어겼으며, 치마 색상(회색) 불일치와 대문 밖에 불이 켜진 부엌이 있는 공간 모순이 발생했습니다."
       }
      ],
      "all_candidates_fail": false,
      "readings": [
       {
        "label": "A",
        "direction": "달래의 시선은 마당 앞쪽 바닥을 향하고 있으며, 문턱을 넘어 마당으로 향하는 방향성을 띰.",
        "built_space": "마당 안쪽에서 대문을 비스듬히 바라보는 앵글이며 우측 단 위에 3개의 장독이 배치됨. 단, 구조물 룩 레퍼런스(초가, 사립문) 대신 로케이션 사진의 기와지붕과 나무 판장문이 묘사됨.",
        "entities": "달래는 연두색 저고리, 붉은 고름, 검은 치마로 인물 레퍼런스와 정확히 일치함. 대문은 지문이 요구한 사립문이 아님.",
        "hard_violations": [],
        "physics": "땅을 딛고 걷는 자세가 바닥에 의해 자연스럽게 지지되고 있음."
       },
       {
        "label": "B",
        "direction": "달래의 시선은 정면 바닥을 향하며 대문 중앙을 통해 안으로 들어오고 있음.",
        "built_space": "초가지붕과 사립문이 정면으로 배치되었고 우측에 3개의 장독이 있음. 대문 밖 배경에 불이 피워진 건물이 보여 공간 배치의 모순이 있음.",
        "entities": "대문은 사립문과 초가 형태로 구조물 레퍼런스와 일치함. 달래의 치마가 회색으로 묘사되어 인물 레퍼런스와 불일치함.",
        "hard_violations": [],
        "physics": "두 발이 마당 바닥에 닿아 있어 서 있거나 걷는 자세가 자연스럽게 지지됨."
       }
      ]
     },
     "ok": true
    },
    {
     "model": "gpt-high",
     "order": "reverse",
     "display_to_canonical": {
      "A": "B",
      "B": "A"
     },
     "raw": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 3,
        "verdict_ko": "초가 재료는 구조 참조에 가깝지만, 집과 마당을 등지고 문밖으로 나오는 구도여서 마당 안쪽에서 포착한 입장 장면이라는 핵심 연출을 위반한다."
       },
       {
        "label": "B",
        "score": 8,
        "verdict_ko": "마당 안쪽의 비스듬한 와이드 숏으로 달래의 전신과 입장 동선을 충실히 보여주지만, 기와 문간과 판자문은 우선 적용해야 할 구조 참조의 초가·엮은 문짝과 다르다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "달래는 시선을 앞쪽 땅으로 낮추고 카메라 쪽으로 걷는다. 뒤에 주거 건물과 안마당이 있으므로 이동은 마당 진입보다 문밖으로 나오는 방향으로 읽힌다. 문을 마당 쪽에서 비스듬히 보는 대신 거의 정면으로 바라본다.",
        "built_space": "중앙에 초가지붕 문간 하나와 양쪽 기둥, 왼쪽으로 열린 엮은 문짝이 보인다. 오른쪽 단 위에는 항아리 세 개가 있고, 문 너머에는 안마당과 주거 건물이 있다. 달래는 그 안마당에서 문턱 쪽으로 접근한다. 구조 참조의 초가와 흙·돌 재료는 반영했으나, 안마당이 문 뒤에 놓여 지정된 카메라의 안팎 관계가 뒤집혔다.",
        "entities": "젊은 한국인 여성으로 설정된 달래 한 명만 보이며, 검은 땋은 머리와 붉은 댕기는 참조와 유사하다. 얼굴은 어둡고 침울하며 전신이 보인다. 저고리는 탁한 회녹색이고 치마는 회갈색으로, 참조의 검은 치마와 다르다. 뒤쪽에 연기와 작은 불꽃이 보여 불이 꺼진 아궁이라는 조건에도 맞지 않는다. 읽을 수 있는 글자나 추가 인물은 없다.",
        "hard_violations": [
         "안마당과 주거 건물을 문 너머 배경에 두고 달래가 바깥 문턱으로 걸어오게 하여, 지정된 마당 안쪽 카메라와 마당 진입 동선을 반대로 배치했다."
        ],
        "physics": "앞발이 흙바닥에 닿고 다른 발이 뒤에서 몸을 지지하는 보행 자세로, 몸이 뜨지는 않는다. 손은 몸 옆에 자연스럽게 내려와 있다. 항아리는 단 위에 놓였고 문짝과 지붕은 문틀과 기둥에 연결되어 있어 지지 관계는 성립한다."
       },
       {
        "label": "B",
        "direction": "달래는 문을 뒤에 두고 화면 앞쪽의 열린 마당으로 한 걸음 들어온다. 시선은 진행 방향의 가까운 땅으로 내려가며 카메라를 응시하지 않는다. 부엌 자체는 보이지 않지만 앞쪽 빈 마당이 이후 이동 경로로 읽힌다.",
        "built_space": "카메라는 마당 안에서 문간을 비스듬히 본다. 문간 하나에 판자문 한 짝이 반쯤 열려 있고, 왼쪽 벽에는 격자창 하나, 오른쪽 긴 단에는 항아리 세 개가 있다. 달래는 열린 통로 바로 안쪽에 서서 앞으로 이동하며 문짝은 얼굴을 가리지 않는다. 위치 참조의 배치와는 잘 맞지만, 기와지붕과 단일 판자문은 구조 참조의 초가지붕 및 양쪽 엮은 문짝을 따르지 않았다.",
        "entities": "달래 한 명의 전신이 보이며 젊은 여성의 외형, 검은 땋은 머리, 붉은 댕기, 연녹색 저고리와 검은 치마가 인물 참조에 가깝다. 작은 화면 크기에서도 고개를 숙인 어두운 표정이 읽힌다. 문, 흙마당, 항아리 세 개가 보이고 문 뒤로 연기가 있으나 아궁이는 화면 밖이라 상태를 확인할 수 없다. 추가 인물이나 읽을 수 있는 글자는 없다.",
        "hard_violations": [],
        "physics": "앞발은 마당 바닥에 닿고 뒤쪽 발은 문턱 가까이에 남아 있어 진입 보행과 체중 이동이 가능하다. 팔과 옷자락도 중력에 맞게 내려온다. 문은 문틀에 연결되고 항아리 세 개는 단 위에 안정적으로 놓여 있으며 지지 없는 물체는 보이지 않는다."
       }
      ],
      "all_candidates_fail": false
     },
     "normalized": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 3,
        "verdict_ko": "초가 재료는 구조 참조에 가깝지만, 집과 마당을 등지고 문밖으로 나오는 구도여서 마당 안쪽에서 포착한 입장 장면이라는 핵심 연출을 위반한다."
       },
       {
        "label": "A",
        "score": 8,
        "verdict_ko": "마당 안쪽의 비스듬한 와이드 숏으로 달래의 전신과 입장 동선을 충실히 보여주지만, 기와 문간과 판자문은 우선 적용해야 할 구조 참조의 초가·엮은 문짝과 다르다."
       }
      ],
      "all_candidates_fail": false,
      "readings": [
       {
        "label": "B",
        "direction": "달래는 시선을 앞쪽 땅으로 낮추고 카메라 쪽으로 걷는다. 뒤에 주거 건물과 안마당이 있으므로 이동은 마당 진입보다 문밖으로 나오는 방향으로 읽힌다. 문을 마당 쪽에서 비스듬히 보는 대신 거의 정면으로 바라본다.",
        "built_space": "중앙에 초가지붕 문간 하나와 양쪽 기둥, 왼쪽으로 열린 엮은 문짝이 보인다. 오른쪽 단 위에는 항아리 세 개가 있고, 문 너머에는 안마당과 주거 건물이 있다. 달래는 그 안마당에서 문턱 쪽으로 접근한다. 구조 참조의 초가와 흙·돌 재료는 반영했으나, 안마당이 문 뒤에 놓여 지정된 카메라의 안팎 관계가 뒤집혔다.",
        "entities": "젊은 한국인 여성으로 설정된 달래 한 명만 보이며, 검은 땋은 머리와 붉은 댕기는 참조와 유사하다. 얼굴은 어둡고 침울하며 전신이 보인다. 저고리는 탁한 회녹색이고 치마는 회갈색으로, 참조의 검은 치마와 다르다. 뒤쪽에 연기와 작은 불꽃이 보여 불이 꺼진 아궁이라는 조건에도 맞지 않는다. 읽을 수 있는 글자나 추가 인물은 없다.",
        "hard_violations": [
         "안마당과 주거 건물을 문 너머 배경에 두고 달래가 바깥 문턱으로 걸어오게 하여, 지정된 마당 안쪽 카메라와 마당 진입 동선을 반대로 배치했다."
        ],
        "physics": "앞발이 흙바닥에 닿고 다른 발이 뒤에서 몸을 지지하는 보행 자세로, 몸이 뜨지는 않는다. 손은 몸 옆에 자연스럽게 내려와 있다. 항아리는 단 위에 놓였고 문짝과 지붕은 문틀과 기둥에 연결되어 있어 지지 관계는 성립한다."
       },
       {
        "label": "A",
        "direction": "달래는 문을 뒤에 두고 화면 앞쪽의 열린 마당으로 한 걸음 들어온다. 시선은 진행 방향의 가까운 땅으로 내려가며 카메라를 응시하지 않는다. 부엌 자체는 보이지 않지만 앞쪽 빈 마당이 이후 이동 경로로 읽힌다.",
        "built_space": "카메라는 마당 안에서 문간을 비스듬히 본다. 문간 하나에 판자문 한 짝이 반쯤 열려 있고, 왼쪽 벽에는 격자창 하나, 오른쪽 긴 단에는 항아리 세 개가 있다. 달래는 열린 통로 바로 안쪽에 서서 앞으로 이동하며 문짝은 얼굴을 가리지 않는다. 위치 참조의 배치와는 잘 맞지만, 기와지붕과 단일 판자문은 구조 참조의 초가지붕 및 양쪽 엮은 문짝을 따르지 않았다.",
        "entities": "달래 한 명의 전신이 보이며 젊은 여성의 외형, 검은 땋은 머리, 붉은 댕기, 연녹색 저고리와 검은 치마가 인물 참조에 가깝다. 작은 화면 크기에서도 고개를 숙인 어두운 표정이 읽힌다. 문, 흙마당, 항아리 세 개가 보이고 문 뒤로 연기가 있으나 아궁이는 화면 밖이라 상태를 확인할 수 없다. 추가 인물이나 읽을 수 있는 글자는 없다.",
        "hard_violations": [],
        "physics": "앞발은 마당 바닥에 닿고 뒤쪽 발은 문턱 가까이에 남아 있어 진입 보행과 체중 이동이 가능하다. 팔과 옷자락도 중력에 맞게 내려온다. 문은 문틀에 연결되고 항아리 세 개는 단 위에 안정적으로 놓여 있으며 지지 없는 물체는 보이지 않는다."
       }
      ]
     },
     "ok": true
    }
   ],
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "slot_winner_match": true,
   "slot_winner": {
    "gemini-pro": "A",
    "gpt-high": "A"
   },
   "route": "cross_slot_agree"
  },
  "dual": {
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "normalized": {
    "A": 2.0,
    "B": 1.042
   },
   "adjusted": {
    "A": 2.0,
    "B": 0.792
   },
   "violations": {
    "B": [
     "[gpt-high] 안마당과 주거 건물을 문 너머 배경에 두고 달래가 바깥 문턱으로 걸어오게 하여, 지정된 마당 안쪽 카메라와 마당 진입 동선을 반대로 배치했다."
    ]
   },
   "per_model_winner": {
    "gemini-pro": "A",
    "gpt-high": "A"
   },
   "agreed": true
  },
  "totals": {
   "A": 2000,
   "B": 792
  },
  "selected": "A",
  "ranking": [
   "A",
   "B"
  ],
  "verdicts": [
   {
    "label": "A",
    "score": 2000,
    "verdict_ko": "지시된 비스듬한 앵글과 마당의 공간감, 캐릭터의 복장(검은 치마)을 정확히 구현했으나, 강제된 사립문과 초가지붕 레퍼런스를 기와와 판장문으로 잘못 묘사했습니다."
   },
   {
    "label": "B",
    "score": 792,
    "verdict_ko": "사립문 구조는 레퍼런스를 따랐으나 정면 구도로 촬영되어 카메라 앵글 지시를 어겼으며, 치마 색상(회색) 불일치와 대문 밖에 불이 켜진 부엌이 있는 공간 모순이 발생했습니다.  ★위반: [gpt-high] 안마당과 주거 건물을 문 너머 배경에 두고 달래가 바깥 문턱으로 걸어오게 하여, 지정된 마당 안쪽 카메라와 마당 진입 동선을 반대로 배치했다."
   }
  ],
  "refs": [
   {
    "label": "LOCATION PHOTOGRAPH — the exact place of this shot: its spatial layout, surroundings, fixed features, time of day and lighting mood are spatial truth; stage the moment inside this place. If a STRUCTURE LOOK photograph is also attached, that photo wins for the fixed structure itself — this photograph wins for everything around it. Never copy its camera framing.",
    "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/db00517b-165b-4005-a287-f28936b59b95/episodes/272bd7af-9294-4490-98c6-88a40e9deacc/images/background_chain/L05B01.png",
    "asset_id": "7cc7dd4c-b9eb-4ed0-a5b4-ce58472b4a3f",
    "role": "location_plate"
   },
   {
    "label": "STRUCTURE LOOK — the confirmed photograph of the fixed structure at this location: wherever the structure appears in the frame, its shape, proportions, materials, colors and openings are LOCKED to this photo. Never copy its camera framing, time of day or lighting — the shot text and the LOCATION PHOTOGRAPH are the authorities for those.",
    "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/db00517b-165b-4005-a287-f28936b59b95/images/background_chain/seed_bg_thatched_house_sel.png",
    "asset_id": "2dc88acf-a75f-4d7e-973d-10ab234660a8",
    "role": "structure_seed_look"
   },
   {
    "label": "CHARACTER REFERENCE — 달래: the exact person appearing in this shot; match face, hair and build exactly.",
    "path": "<bytes:879791>",
    "asset_id": "83024e69-b375-4f3e-98ec-820d0e2ff1f0",
    "role": "character_ref"
   }
  ],
  "critique_skipped": true,
  "shot_run_produced": false,
  "shot_run_uid": "06aa1407-05d0-701d-9aba-8bb9692c9c52",
  "bgfirst": {
   "bg_path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/db00517b-165b-4005-a287-f28936b59b95/images/272bd7af-9294-4490-98c6-88a40e9deacc/scene/recipe/S8sh5__bgfirst_bg.png",
   "bg_asset_id": "b99aae16-b8a5-4629-9e89-1289d0ed1546",
   "bg_record_key": "S8sh5::bgfirst_bg",
   "chain_winner": true,
   "authority": "plate",
   "seed_attached": true
  },
  "ref_mode": "재투영 배경+콘티+엔티티 (2택1: 체인 승)",
  "share_plan": {
   "ref_plan": "background"
  },
  "lane_policy": "ab_select_ready"
 },
 "S8sh5::cine": {
  "applied": true,
  "attempted_at": "2026-09-09T01:25:15.616552+00:00",
  "fingerprint": "71f2a8b55b518fb1d40c938b954e63ec6a7e73b928938fff37ba562831a4fb43",
  "fingerprint_version": 2,
  "provider": "mai",
  "endpoint": "openrouter/chat-completions",
  "model": "microsoft/mai-image-2.6",
  "pack": "24.202608252115",
  "source_file": "S8sh5_sel.png",
  "source_sha256": "07620b7ce6798a5e950af8736ccd7e770f41b23719d3727536bf828c917364a9",
  "file": "S8sh5_cine.png",
  "staged_sha256": "cd6ce6437fdd025786bd89acf2aefaff4b1989ae3ff71873247881253ecbfe6f",
  "latency_ms": 26720
 },
 "S8sh6::signage": {
  "fp": "d97faf927df69247",
  "inscriptions": [],
  "cues": [],
  "dropped": []
 },
 "S8sh6::bgfirst_bg": {
  "input_fingerprint": "15fb35a33c33c0ff",
  "prompt": "Create the EMPTY BACKGROUND PLATE for one film shot — NO PEOPLE, no figures, no body parts, no sketch lines, no arrows anywhere.\n\"Empty\" means no people only: KEEP the location's inherent occupants and stock that define the place — animals in an animal shelter, pen or farm, goods in a market, moored boats in a harbour — unless the shot text explicitly removes them.\nThe FIRST attached image is a thin-line storyboard sketch: use ONLY its camera angle, horizon, perspective and the placement/size of buildings and set masses — ignore the sketched people and arrows entirely. The SECOND attached image (LOCATION PHOTOGRAPH) is the real place: take its architecture, materials, signage and fixed features, and RE-PROJECT them into the sketch's camera. If the photograph's camera differs from the sketch's, the sketch's camera wins.\nHUMAN-SCALE CALIBRATION: derive every structure's true size from human-scale elements — a door ≈ 2m, a window ≈ 1–1.5m wide, one storey ≈ 2.5–3m; never inflate a small structure or shrink a large one.\n\nSHOT TEXT this background must serve (Korean): 장독대 옆, 이관회를 걱정스레 올려다보는 달래의 흔들리는 눈빛 클로즈업.\n\nLOCATION (lock): Beside the earthenware storage jars in a modest home's outdoor yard at dusk or early night.\n\nTIME OF DAY (lock): dusk.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: close-up\n- KEY BACKGROUND ELEMENTS: 장독대 (Beside the two characters during their conversation); used as A soft peripheral fragment anchors the close-up to their new position without competing with her eyes.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Maintain the courtyard's subdued ambient illumination and muted saturation, keeping her eyes readable without a conspicuous lighting change.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nRender ONE photorealistic empty location photograph, 16:9, that this shot can be staged inside later. No readable writing anywhere: surfaces that would carry writing may be present, but stage any wording out of legibility — an oblique angle, distance, shallow focus. No captions, watermarks or overlay text.",
  "effective_prompt": "Create the EMPTY BACKGROUND PLATE for one film shot — NO PEOPLE, no figures, no body parts, no sketch lines, no arrows anywhere.\n\"Empty\" means no people only: KEEP the location's inherent occupants and stock that define the place — animals in an animal shelter, pen or farm, goods in a market, moored boats in a harbour — unless the shot text explicitly removes them.\nThe FIRST attached image is a thin-line storyboard sketch: use ONLY its camera angle, horizon, perspective and the placement/size of buildings and set masses — ignore the sketched people and arrows entirely. The SECOND attached image (LOCATION PHOTOGRAPH) is the real place: take its architecture, materials, signage and fixed features, and RE-PROJECT them into the sketch's camera. If the photograph's camera differs from the sketch's, the sketch's camera wins.\nHUMAN-SCALE CALIBRATION: derive every structure's true size from human-scale elements — a door ≈ 2m, a window ≈ 1–1.5m wide, one storey ≈ 2.5–3m; never inflate a small structure or shrink a large one.\n\nSHOT TEXT this background must serve (Korean): 장독대 옆, 이관회를 걱정스레 올려다보는 달래의 흔들리는 눈빛 클로즈업.\n\nLOCATION (lock): Beside the earthenware storage jars in a modest home's outdoor yard at dusk or early night.\n\nTIME OF DAY (lock): dusk.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: close-up\n- KEY BACKGROUND ELEMENTS: 장독대 (Beside the two characters during their conversation); used as A soft peripheral fragment anchors the close-up to their new position without competing with her eyes.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Maintain the courtyard's subdued ambient illumination and muted saturation, keeping her eyes readable without a conspicuous lighting change.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nRender ONE photorealistic empty location photograph, 16:9, that this shot can be staged inside later. No readable writing anywhere: surfaces that would carry writing may be present, but stage any wording out of legibility — an oblique angle, distance, shallow focus. No captions, watermarks or overlay text.",
  "bg_path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/db00517b-165b-4005-a287-f28936b59b95/images/272bd7af-9294-4490-98c6-88a40e9deacc/scene/recipe/S8sh6__bgfirst_bg.png",
  "asset_id": "0a0b9572-eaf9-40c3-9072-99d45cfed440",
  "input_asset_ids": [
   "849ed822-4ebd-4e1d-a09a-8e8abf015a41",
   "9d0732c6-18a5-4222-b339-11f5034a536b"
  ]
 },
 "S8sh6": {
  "input_fingerprint": "11c676d4aca99b11",
  "prompt": "Create ONE FINAL photorealistic live-action film still of the moment below — early 20th-century Korea, Jeolla region; all people are Korean unless explicitly stated otherwise. TIME OF DAY (lock): dusk.\n\nSHOT TEXT (authoritative, Korean): 장독대 옆, 이관회를 걱정스레 올려다보는 달래의 흔들리는 눈빛 클로즈업.\n\nLOCATION (lock): Beside the earthenware storage jars in a modest home's outdoor yard at dusk or early night. The shot takes place here — the attached LOCATION PHOTOGRAPH shows the exact spot.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: close-up\n- KEY BACKGROUND ELEMENTS: 장독대 (Beside the two characters during their conversation); used as A soft peripheral fragment anchors the close-up to their new position without competing with her eyes.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Maintain the courtyard's subdued ambient illumination and muted saturation, keeping her eyes readable without a conspicuous lighting change.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): Darkness settles around the jars and the open gate; the hearth has not been lit. 달래: She stands beside the jar terrace.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 달래 (한국인 여성, 젊은 성인, 자연스러운 검은 머리) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.",
  "roll_prompts": {
   "A": "Stage the shot. The FIRST attached image (SHOT BACKGROUND) is the finished empty background of this shot — keep it EXACTLY: its camera, perspective, architecture, lighting and every fixed feature stay untouched. The SECOND attached image (LAYOUT SKETCH) tells you ONLY where the people go: each sketched person's position, screen size, pose and the gaze/motion arrows. Ignore the sketch's background lines. The CHARACTER REFERENCE photographs show the real people.\nPlace the real people into the background at exactly the sketched positions, sizes and poses, following the arrow directions. No sketch lines or arrows may remain.\n\nCreate ONE FINAL photorealistic live-action film still of the moment below — early 20th-century Korea, Jeolla region; all people are Korean unless explicitly stated otherwise. TIME OF DAY (lock): dusk.\n\nSHOT TEXT (authoritative, Korean): 장독대 옆, 이관회를 걱정스레 올려다보는 달래의 흔들리는 눈빛 클로즈업.\n\nLOCATION (lock): Beside the earthenware storage jars in a modest home's outdoor yard at dusk or early night. The shot takes place here — the FIRST attached image (SHOT BACKGROUND) is this exact place, already built: its ground, structures, horizon, materials and lighting are the finished truth of this location and must not be redesigned or replaced. No location photograph is attached — read the place from that image alone, and add no scenery, structure, vehicle or fixture that it does not already show. This lock governs the place only; the figures in the shot follow the staging and pose instructions.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: close-up\n- KEY BACKGROUND ELEMENTS: 장독대 (Beside the two characters during their conversation); used as A soft peripheral fragment anchors the close-up to their new position without competing with her eyes.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Maintain the courtyard's subdued ambient illumination and muted saturation, keeping her eyes readable without a conspicuous lighting change.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): Darkness settles around the jars and the open gate; the hearth has not been lit. 달래: She stands beside the jar terrace.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 달래 (한국인 여성, 젊은 성인, 자연스러운 검은 머리) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.",
   "B": "Create ONE FINAL photorealistic live-action film still of the moment below — early 20th-century Korea, Jeolla region; all people are Korean unless explicitly stated otherwise. TIME OF DAY (lock): dusk.\n\nSHOT TEXT (authoritative, Korean): 장독대 옆, 이관회를 걱정스레 올려다보는 달래의 흔들리는 눈빛 클로즈업.\n\nLOCATION (lock): Beside the earthenware storage jars in a modest home's outdoor yard at dusk or early night. The shot takes place here — the attached LOCATION PHOTOGRAPH shows the exact spot.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: close-up\n- KEY BACKGROUND ELEMENTS: 장독대 (Beside the two characters during their conversation); used as A soft peripheral fragment anchors the close-up to their new position without competing with her eyes.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Maintain the courtyard's subdued ambient illumination and muted saturation, keeping her eyes readable without a conspicuous lighting change.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): Darkness settles around the jars and the open gate; the hearth has not been lit. 달래: She stands beside the jar terrace.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 달래 (한국인 여성, 젊은 성인, 자연스러운 검은 머리) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.\n\nCOMPOSITION VARIATION (this candidate only): keep the place, the people and the moment exactly as specified, and keep from the CAMERA & FRAME contract its subject scale, screen placement and key background elements exactly. For this candidate only, its camera angle and height are deliberately left open: choose a camera position distinctly different from the obvious one, as a film director picking a second setup on the same blocking. Stay true to any explicit framing the SHOT TEXT itself dictates."
  },
  "roll_refs": {
   "A": [
    {
     "label": "SHOT BACKGROUND",
     "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/db00517b-165b-4005-a287-f28936b59b95/images/272bd7af-9294-4490-98c6-88a40e9deacc/scene/recipe/S8sh6__bgfirst_bg.png",
     "asset_id": "0a0b9572-eaf9-40c3-9072-99d45cfed440",
     "role": "bgfirst_bg"
    },
    {
     "label": "LAYOUT SKETCH (people placement only)",
     "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/db00517b-165b-4005-a287-f28936b59b95/images/272bd7af-9294-4490-98c6-88a40e9deacc/conti/conti_S8sh6.png",
     "asset_id": "849ed822-4ebd-4e1d-a09a-8e8abf015a41",
     "role": "conti_light"
    },
    {
     "label": "CHARACTER REFERENCE — 달래: the exact person appearing in this shot; match face, hair and build exactly.",
     "path": "<bytes:879791>",
     "asset_id": "83024e69-b375-4f3e-98ec-820d0e2ff1f0",
     "role": "character_ref"
    }
   ],
   "B": [
    {
     "label": "LOCATION PHOTOGRAPH — the exact place of this shot: its architecture, materials, fixed features and lighting mood are spatial truth; stage the moment inside this place. Never copy its camera framing.",
     "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/db00517b-165b-4005-a287-f28936b59b95/episodes/272bd7af-9294-4490-98c6-88a40e9deacc/images/background_chain/L05B02.png",
     "asset_id": "9d0732c6-18a5-4222-b339-11f5034a536b",
     "role": "location_plate"
    },
    {
     "label": "CHARACTER REFERENCE — 달래: the exact person appearing in this shot; match face, hair and build exactly.",
     "path": "<bytes:879791>",
     "asset_id": "83024e69-b375-4f3e-98ec-820d0e2ff1f0",
     "role": "character_ref"
    }
   ]
  },
  "shot_run_spend_attempt_count": 1,
  "judge_flip_skipped": "judge_fn_owns_order",
  "readings": [
   {
    "label": "A",
    "direction": "여성이 우측 전경에 있는 남성의 뒷모습을 향해 시선을 위로 향하고 있습니다.",
    "built_space": "배경에 레퍼런스의 담장, 대문, 장독대가 보이나, 화면 중앙 전경에 배경에 있어야 할 장독 하나가 거대하게 확대되어 배치되어 있습니다.",
    "entities": "달래의 외모와 복장이 레퍼런스와 일치하며, 숏 텍스트의 맥락에 따라 이관회로 추정되는 남성의 뒷모습이 등장합니다.",
    "hard_violations": [
     "[gemini-pro] 지지대 없이 허공에 떠 있는 거대한 장독 (물리적으로 불가능한 배치)",
     "[gemini-pro] 배경 사물을 전경으로 거대하게 확대하지 말라는 프레이밍 지시 위반",
     "[gpt-high] 참조 장소의 장독 세 개를 여섯 개로 늘려 고정된 장독대 배치에 없는 물체를 추가했다.",
     "[gpt-high] 가시 인물을 달래로 한정한 인물 조건과 달리 오른쪽에 성인 남성의 신체를 노출했다."
    ],
    "physics": "화면 한가운데를 차지하는 큰 장독을 받치고 있는 바닥이나 손이 전혀 보이지 않아 허공에 떠 있는 상태입니다."
   },
   {
    "label": "B",
    "direction": "여성이 화면 밖의 대상(이관회)을 향해 걱정스러운 표정으로 위쪽을 응시하고 있습니다.",
    "built_space": "로케이션 레퍼런스에 제시된 흙벽, 기와, 장독대가 인물 뒤로 적절히 아웃포커싱되어 배치되어 있습니다.",
    "entities": "달래의 이목구비, 헤어스타일, 의상(옥색 저고리)이 캐릭터 레퍼런스와 정확히 일치하며 눈물 고인 눈빛이 묘사되었습니다.",
    "hard_violations": [],
    "physics": "인물이 땅에 서 있는 자세로 자연스럽게 지탱되고 있으며, 물리적으로 어색하거나 공중에 떠 있는 물체는 없습니다."
   }
  ],
  "cross_model_order": {
   "policy": "cross_model_order_v2_slot0_fwd_slot1_rev_parallel2",
   "slots": [
    {
     "model": "gemini-pro",
     "order": "forward",
     "display_to_canonical": {
      "A": "A",
      "B": "B"
     },
     "raw": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 10,
        "verdict_ko": "지시된 클로즈업 프레이밍과 얕은 심도의 장독대 배경을 완벽하게 구현하며, 걱정스레 올려다보는 달래의 눈빛을 훌륭하게 포착했습니다."
       },
       {
        "label": "A",
        "score": 0,
        "verdict_ko": "화면 중앙 전경에 아무런 지지대 없이 허공에 떠 있는 거대한 장독이 등장하여 물리적으로 불가능한 장면을 연출한 치명적인 오류가 있습니다."
       }
      ],
      "readings": [
       {
        "label": "B",
        "direction": "여성이 화면 밖의 대상(이관회)을 향해 걱정스러운 표정으로 위쪽을 응시하고 있습니다.",
        "built_space": "로케이션 레퍼런스에 제시된 흙벽, 기와, 장독대가 인물 뒤로 적절히 아웃포커싱되어 배치되어 있습니다.",
        "entities": "달래의 이목구비, 헤어스타일, 의상(옥색 저고리)이 캐릭터 레퍼런스와 정확히 일치하며 눈물 고인 눈빛이 묘사되었습니다.",
        "hard_violations": [],
        "physics": "인물이 땅에 서 있는 자세로 자연스럽게 지탱되고 있으며, 물리적으로 어색하거나 공중에 떠 있는 물체는 없습니다."
       },
       {
        "label": "A",
        "direction": "여성이 우측 전경에 있는 남성의 뒷모습을 향해 시선을 위로 향하고 있습니다.",
        "built_space": "배경에 레퍼런스의 담장, 대문, 장독대가 보이나, 화면 중앙 전경에 배경에 있어야 할 장독 하나가 거대하게 확대되어 배치되어 있습니다.",
        "entities": "달래의 외모와 복장이 레퍼런스와 일치하며, 숏 텍스트의 맥락에 따라 이관회로 추정되는 남성의 뒷모습이 등장합니다.",
        "hard_violations": [
         "지지대 없이 허공에 떠 있는 거대한 장독 (물리적으로 불가능한 배치)",
         "배경 사물을 전경으로 거대하게 확대하지 말라는 프레이밍 지시 위반"
        ],
        "physics": "화면 한가운데를 차지하는 큰 장독을 받치고 있는 바닥이나 손이 전혀 보이지 않아 허공에 떠 있는 상태입니다."
       }
      ],
      "all_candidates_fail": false
     },
     "normalized": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 10,
        "verdict_ko": "지시된 클로즈업 프레이밍과 얕은 심도의 장독대 배경을 완벽하게 구현하며, 걱정스레 올려다보는 달래의 눈빛을 훌륭하게 포착했습니다."
       },
       {
        "label": "A",
        "score": 0,
        "verdict_ko": "화면 중앙 전경에 아무런 지지대 없이 허공에 떠 있는 거대한 장독이 등장하여 물리적으로 불가능한 장면을 연출한 치명적인 오류가 있습니다."
       }
      ],
      "all_candidates_fail": false,
      "readings": [
       {
        "label": "B",
        "direction": "여성이 화면 밖의 대상(이관회)을 향해 걱정스러운 표정으로 위쪽을 응시하고 있습니다.",
        "built_space": "로케이션 레퍼런스에 제시된 흙벽, 기와, 장독대가 인물 뒤로 적절히 아웃포커싱되어 배치되어 있습니다.",
        "entities": "달래의 이목구비, 헤어스타일, 의상(옥색 저고리)이 캐릭터 레퍼런스와 정확히 일치하며 눈물 고인 눈빛이 묘사되었습니다.",
        "hard_violations": [],
        "physics": "인물이 땅에 서 있는 자세로 자연스럽게 지탱되고 있으며, 물리적으로 어색하거나 공중에 떠 있는 물체는 없습니다."
       },
       {
        "label": "A",
        "direction": "여성이 우측 전경에 있는 남성의 뒷모습을 향해 시선을 위로 향하고 있습니다.",
        "built_space": "배경에 레퍼런스의 담장, 대문, 장독대가 보이나, 화면 중앙 전경에 배경에 있어야 할 장독 하나가 거대하게 확대되어 배치되어 있습니다.",
        "entities": "달래의 외모와 복장이 레퍼런스와 일치하며, 숏 텍스트의 맥락에 따라 이관회로 추정되는 남성의 뒷모습이 등장합니다.",
        "hard_violations": [
         "지지대 없이 허공에 떠 있는 거대한 장독 (물리적으로 불가능한 배치)",
         "배경 사물을 전경으로 거대하게 확대하지 말라는 프레이밍 지시 위반"
        ],
        "physics": "화면 한가운데를 차지하는 큰 장독을 받치고 있는 바닥이나 손이 전혀 보이지 않아 허공에 떠 있는 상태입니다."
       }
      ]
     },
     "ok": true
    },
    {
     "model": "gpt-high",
     "order": "reverse",
     "display_to_canonical": {
      "A": "B",
      "B": "A"
     },
     "raw": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 9,
        "verdict_ko": "걱정스러운 상향 시선을 얼굴 클로즈업으로 담고, 세 장독을 흐린 주변 배경으로 남겨 핵심 연출과 황혼의 장소감을 충실히 구현했다."
       },
       {
        "label": "B",
        "score": 2,
        "verdict_ko": "남자를 올려다보는 방향은 맞지만, 허용되지 않은 남성 노출과 장독 증설에 더해 넓어진 구도와 전경의 큰 장독이 눈빛 클로즈업 지시를 훼손한다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "달래의 턱과 두 눈이 화면 왼쪽 위의 프레임 밖 상대를 향한다. 이관회 자체는 보이지 않지만, 자신보다 높은 위치의 상대를 걱정스럽게 올려다보는 시선으로 읽힌다. 찌푸린 눈썹과 젖은 눈가가 불안을 표현하며 렌즈를 응시하지 않는다.",
        "built_space": "왼쪽에 대문 기둥과 열린 통로 일부, 뒤에 기와를 얹은 흙담과 낮은 장독대가 보인다. 뚜껑 있는 장독은 왼쪽 하나, 얼굴 뒤 일부 하나, 오른쪽 가장자리 하나로 총 세 개가 식별되어 장소 참조와 부합한다. 달래는 장독대 앞 마당 쪽에 있으며, 장독과 담은 흐린 주변 배경으로 처리되었다.",
        "entities": "젊은 한국인 여성으로 설정된 달래 한 명만 보인다. 검은 가운데 가르마 머리, 뒤로 모은 머리, 옅은 녹색 저고리와 흰 깃이 참조에 부합한다. 얼굴의 세부 인상은 참조와 약간 다르지만 전반적인 연령과 외형은 가깝다. 땋은 머리의 전체 길이와 하의는 프레임 밖이므로 판단 대상이 아니다. 장독은 실제 유약 바른 옹기처럼 보이고 읽을 수 있는 글자는 없다.",
        "hard_violations": [],
        "physics": "머리는 목과 어깨에 자연스럽게 연결되고, 상체를 세운 채 턱을 조금 드는 자세가 가능하다. 발과 지면 접촉은 클로즈업 밖이므로 직접 확인할 수 없지만 공중에 뜬 징후는 없다. 장독은 장독대 위에 놓였으며, 뚜껑도 각 장독 입구에 얹혀 있다."
       },
       {
        "label": "B",
        "direction": "달래는 화면 오른쪽 위 남성의 얼굴 방향으로 두 눈과 턱을 들어 올린다. 남성도 얼굴을 달래 쪽으로 돌리고 있으나 눈은 프레임 밖이다. 걱정스러운 눈썹과 눈물이 보이고, 상대를 올려다보는 방향 자체는 명확하다.",
        "built_space": "왼쪽 격자창 하나, 중앙 열린 대문 하나, 기와지붕과 흙담, 오른쪽 장독대가 보여 참조 장소의 주요 구조는 유지된다. 그러나 장독대에 다섯 개의 장독이 보이고 달래 바로 뒤 전경에 큰 장독 하나가 추가되어, 참조의 세 개와 달리 총 여섯 개가 식별된다. 달래의 상반신과 남성의 어깨, 마당 상당 부분을 포함하는 구도여서 눈빛 중심 클로즈업보다 넓으며, 큰 전경 장독이 얼굴과 시각적으로 경쟁한다.",
        "entities": "달래는 젊은 한국인 여성으로 보이며 검은 가르마 머리와 어깨 앞으로 내린 땋은 머리, 녹색 저고리와 흰 깃이 참조에 대체로 부합한다. 오른쪽에는 짧은 검은 머리와 어두운 옷의 성인 남성이 귀·턱·목·어깨까지 보인다. 이관회 역할로 읽히지만, 가시 인물을 달래로 한정한 명시적 인물 조건과 충돌한다. 옹기와 의복은 물리적 재질로 표현되며 읽을 수 있는 글자는 없다.",
        "hard_violations": [
         "참조 장소의 장독 세 개를 여섯 개로 늘려 고정된 장독대 배치에 없는 물체를 추가했다.",
         "가시 인물을 달래로 한정한 인물 조건과 달리 오른쪽에 성인 남성의 신체를 노출했다."
        ],
        "physics": "달래의 목과 상체는 자연스럽게 이어지고, 가까이 선 키 큰 상대를 올려다보는 자세가 가능하다. 남성의 잘린 상체에도 부유나 불가능한 관절 징후는 없다. 두 사람의 발은 프레임 밖이다. 뒤쪽 장독들은 장독대에 놓여 있고, 큰 전경 장독의 바닥은 가려져 있어 지지 접촉을 직접 확인할 수는 없으나 떠 있다고 볼 근거는 없다."
       }
      ],
      "all_candidates_fail": false
     },
     "normalized": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 9,
        "verdict_ko": "걱정스러운 상향 시선을 얼굴 클로즈업으로 담고, 세 장독을 흐린 주변 배경으로 남겨 핵심 연출과 황혼의 장소감을 충실히 구현했다."
       },
       {
        "label": "A",
        "score": 2,
        "verdict_ko": "남자를 올려다보는 방향은 맞지만, 허용되지 않은 남성 노출과 장독 증설에 더해 넓어진 구도와 전경의 큰 장독이 눈빛 클로즈업 지시를 훼손한다."
       }
      ],
      "all_candidates_fail": false,
      "readings": [
       {
        "label": "B",
        "direction": "달래의 턱과 두 눈이 화면 왼쪽 위의 프레임 밖 상대를 향한다. 이관회 자체는 보이지 않지만, 자신보다 높은 위치의 상대를 걱정스럽게 올려다보는 시선으로 읽힌다. 찌푸린 눈썹과 젖은 눈가가 불안을 표현하며 렌즈를 응시하지 않는다.",
        "built_space": "왼쪽에 대문 기둥과 열린 통로 일부, 뒤에 기와를 얹은 흙담과 낮은 장독대가 보인다. 뚜껑 있는 장독은 왼쪽 하나, 얼굴 뒤 일부 하나, 오른쪽 가장자리 하나로 총 세 개가 식별되어 장소 참조와 부합한다. 달래는 장독대 앞 마당 쪽에 있으며, 장독과 담은 흐린 주변 배경으로 처리되었다.",
        "entities": "젊은 한국인 여성으로 설정된 달래 한 명만 보인다. 검은 가운데 가르마 머리, 뒤로 모은 머리, 옅은 녹색 저고리와 흰 깃이 참조에 부합한다. 얼굴의 세부 인상은 참조와 약간 다르지만 전반적인 연령과 외형은 가깝다. 땋은 머리의 전체 길이와 하의는 프레임 밖이므로 판단 대상이 아니다. 장독은 실제 유약 바른 옹기처럼 보이고 읽을 수 있는 글자는 없다.",
        "hard_violations": [],
        "physics": "머리는 목과 어깨에 자연스럽게 연결되고, 상체를 세운 채 턱을 조금 드는 자세가 가능하다. 발과 지면 접촉은 클로즈업 밖이므로 직접 확인할 수 없지만 공중에 뜬 징후는 없다. 장독은 장독대 위에 놓였으며, 뚜껑도 각 장독 입구에 얹혀 있다."
       },
       {
        "label": "A",
        "direction": "달래는 화면 오른쪽 위 남성의 얼굴 방향으로 두 눈과 턱을 들어 올린다. 남성도 얼굴을 달래 쪽으로 돌리고 있으나 눈은 프레임 밖이다. 걱정스러운 눈썹과 눈물이 보이고, 상대를 올려다보는 방향 자체는 명확하다.",
        "built_space": "왼쪽 격자창 하나, 중앙 열린 대문 하나, 기와지붕과 흙담, 오른쪽 장독대가 보여 참조 장소의 주요 구조는 유지된다. 그러나 장독대에 다섯 개의 장독이 보이고 달래 바로 뒤 전경에 큰 장독 하나가 추가되어, 참조의 세 개와 달리 총 여섯 개가 식별된다. 달래의 상반신과 남성의 어깨, 마당 상당 부분을 포함하는 구도여서 눈빛 중심 클로즈업보다 넓으며, 큰 전경 장독이 얼굴과 시각적으로 경쟁한다.",
        "entities": "달래는 젊은 한국인 여성으로 보이며 검은 가르마 머리와 어깨 앞으로 내린 땋은 머리, 녹색 저고리와 흰 깃이 참조에 대체로 부합한다. 오른쪽에는 짧은 검은 머리와 어두운 옷의 성인 남성이 귀·턱·목·어깨까지 보인다. 이관회 역할로 읽히지만, 가시 인물을 달래로 한정한 명시적 인물 조건과 충돌한다. 옹기와 의복은 물리적 재질로 표현되며 읽을 수 있는 글자는 없다.",
        "hard_violations": [
         "참조 장소의 장독 세 개를 여섯 개로 늘려 고정된 장독대 배치에 없는 물체를 추가했다.",
         "가시 인물을 달래로 한정한 인물 조건과 달리 오른쪽에 성인 남성의 신체를 노출했다."
        ],
        "physics": "달래의 목과 상체는 자연스럽게 이어지고, 가까이 선 키 큰 상대를 올려다보는 자세가 가능하다. 남성의 잘린 상체에도 부유나 불가능한 관절 징후는 없다. 두 사람의 발은 프레임 밖이다. 뒤쪽 장독들은 장독대에 놓여 있고, 큰 전경 장독의 바닥은 가려져 있어 지지 접촉을 직접 확인할 수는 없으나 떠 있다고 볼 근거는 없다."
       }
      ]
     },
     "ok": true
    }
   ],
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "slot_winner_match": true,
   "slot_winner": {
    "gemini-pro": "B",
    "gpt-high": "B"
   },
   "route": "cross_slot_agree"
  },
  "dual": {
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "normalized": {
    "A": 0.222,
    "B": 2.0
   },
   "adjusted": {
    "A": -0.028,
    "B": 2.0
   },
   "violations": {
    "A": [
     "[gemini-pro] 지지대 없이 허공에 떠 있는 거대한 장독 (물리적으로 불가능한 배치)",
     "[gemini-pro] 배경 사물을 전경으로 거대하게 확대하지 말라는 프레이밍 지시 위반",
     "[gpt-high] 참조 장소의 장독 세 개를 여섯 개로 늘려 고정된 장독대 배치에 없는 물체를 추가했다.",
     "[gpt-high] 가시 인물을 달래로 한정한 인물 조건과 달리 오른쪽에 성인 남성의 신체를 노출했다."
    ]
   },
   "per_model_winner": {
    "gemini-pro": "B",
    "gpt-high": "B"
   },
   "agreed": true
  },
  "totals": {
   "B": 2000,
   "A": -28
  },
  "selected": "B",
  "ranking": [
   "B",
   "A"
  ],
  "verdicts": [
   {
    "label": "B",
    "score": 2000,
    "verdict_ko": "지시된 클로즈업 프레이밍과 얕은 심도의 장독대 배경을 완벽하게 구현하며, 걱정스레 올려다보는 달래의 눈빛을 훌륭하게 포착했습니다."
   },
   {
    "label": "A",
    "score": -28,
    "verdict_ko": "화면 중앙 전경에 아무런 지지대 없이 허공에 떠 있는 거대한 장독이 등장하여 물리적으로 불가능한 장면을 연출한 치명적인 오류가 있습니다.  ★위반: [gemini-pro] 지지대 없이 허공에 떠 있는 거대한 장독 (물리적으로 불가능한 배치) / [gemini-pro] 배경 사물을 전경으로 거대하게 확대하지 말라는 프레이밍 지시 위반 / [gpt-high] 참조 장소의 장독 세 개를 여섯 개로 늘려 고정된 장독대 배치에 없는 물체를 추가했다. / [gpt-high] 가시 인물을 달래로 한정한 인물 조건과 달리 오른쪽에 성인 남성의 신체를 노출했다."
   }
  ],
  "refs": [
   {
    "label": "LOCATION PHOTOGRAPH — the exact place of this shot: its architecture, materials, fixed features and lighting mood are spatial truth; stage the moment inside this place. Never copy its camera framing.",
    "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/db00517b-165b-4005-a287-f28936b59b95/episodes/272bd7af-9294-4490-98c6-88a40e9deacc/images/background_chain/L05B02.png",
    "asset_id": "9d0732c6-18a5-4222-b339-11f5034a536b",
    "role": "location_plate"
   },
   {
    "label": "CHARACTER REFERENCE — 달래: the exact person appearing in this shot; match face, hair and build exactly.",
    "path": "<bytes:879791>",
    "asset_id": "83024e69-b375-4f3e-98ec-820d0e2ff1f0",
    "role": "character_ref"
   }
  ],
  "critique_skipped": true,
  "shot_run_produced": false,
  "shot_run_uid": "06aa140a-0a10-7b84-85bf-c194e842b5f4",
  "bgfirst": {
   "bg_path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/db00517b-165b-4005-a287-f28936b59b95/images/272bd7af-9294-4490-98c6-88a40e9deacc/scene/recipe/S8sh6__bgfirst_bg.png",
   "bg_asset_id": "0a0b9572-eaf9-40c3-9072-99d45cfed440",
   "bg_record_key": "S8sh6::bgfirst_bg",
   "chain_winner": false,
   "authority": "plate"
  },
  "ref_mode": "플레이트+엔티티 (2택1: 무콘티 승)",
  "share_plan": {
   "ref_plan": "background"
  }
 },
 "S8sh6::cine": {
  "applied": true,
  "attempted_at": "2026-09-09T01:27:55.461915+00:00",
  "fingerprint": "09c27cfcd3831efc4eb06e6c63c5273551d703b4d112dd9a6e4f09b9fc857c6b",
  "fingerprint_version": 2,
  "provider": "mai",
  "endpoint": "openrouter/chat-completions",
  "model": "microsoft/mai-image-2.6",
  "pack": "24.202608252115",
  "source_file": "S8sh6_sel.png",
  "source_sha256": "052d4ef40a929a2f04c37179c0246c5482da394f0cf18cad0c94432922608ff2",
  "file": "S8sh6_cine.png",
  "staged_sha256": "6d81642e3d9c13b3cd5dd6c66024f768f5a69239899b95a84765af1488d155ef",
  "latency_ms": 29427
 },
 "S8sh9::signage": {
  "fp": "76b5941deb5c5949",
  "inscriptions": [],
  "cues": [],
  "dropped": []
 },
 "S8sh9": {
  "input_fingerprint": "66e0c6a2f0d1b2b5",
  "prompt": "Create ONE FINAL photorealistic live-action film still of the moment below — early 20th-century Korea, Jeolla region; all people are Korean unless explicitly stated otherwise. TIME OF DAY (lock): dusk.\n\nSHOT TEXT (authoritative, Korean): 입을 굳게 다문 채 어둠이 깔린 사립문 밖을 뚫어져라 응시하는 이관회의 옆얼굴.\n\nLOCATION (lock): The outdoor standing space beside a home's storage jars, facing the dark area beyond the rustic wicket gate. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: close-up\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Keep the ambient illumination subdued against the darkness beyond the gate, preserving natural facial contrast and sober tension.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The people visible in that photo are NOT in this shot: never carry their faces, bodies or clothing onto anyone here. Each person in THIS shot is defined solely by the PEOPLE list, their own CHARACTER REFERENCE images and their wardrobe notes. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The yard and the area beyond the gate remain dark. 이관회: He stands beside the jar terrace, his forced smile gone.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 이관회 (한국인 남성, 청년 나이대, 자연스러운 검은 머리) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.",
  "roll_prompts": {
   "A": "Create ONE FINAL photorealistic live-action film still of the moment below — early 20th-century Korea, Jeolla region; all people are Korean unless explicitly stated otherwise. TIME OF DAY (lock): dusk.\n\nSHOT TEXT (authoritative, Korean): 입을 굳게 다문 채 어둠이 깔린 사립문 밖을 뚫어져라 응시하는 이관회의 옆얼굴.\n\nLOCATION (lock): The outdoor standing space beside a home's storage jars, facing the dark area beyond the rustic wicket gate. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: close-up\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Keep the ambient illumination subdued against the darkness beyond the gate, preserving natural facial contrast and sober tension.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The people visible in that photo are NOT in this shot: never carry their faces, bodies or clothing onto anyone here. Each person in THIS shot is defined solely by the PEOPLE list, their own CHARACTER REFERENCE images and their wardrobe notes. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The yard and the area beyond the gate remain dark. 이관회: He stands beside the jar terrace, his forced smile gone.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 이관회 (한국인 남성, 청년 나이대, 자연스러운 검은 머리) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.",
   "B": "Create ONE FINAL photorealistic live-action film still of the moment below — early 20th-century Korea, Jeolla region; all people are Korean unless explicitly stated otherwise. TIME OF DAY (lock): dusk.\n\nSHOT TEXT (authoritative, Korean): 입을 굳게 다문 채 어둠이 깔린 사립문 밖을 뚫어져라 응시하는 이관회의 옆얼굴.\n\nLOCATION (lock): The outdoor standing space beside a home's storage jars, facing the dark area beyond the rustic wicket gate. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: close-up\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Keep the ambient illumination subdued against the darkness beyond the gate, preserving natural facial contrast and sober tension.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The people visible in that photo are NOT in this shot: never carry their faces, bodies or clothing onto anyone here. Each person in THIS shot is defined solely by the PEOPLE list, their own CHARACTER REFERENCE images and their wardrobe notes. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The yard and the area beyond the gate remain dark. 이관회: He stands beside the jar terrace, his forced smile gone.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 이관회 (한국인 남성, 청년 나이대, 자연스러운 검은 머리) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.\n\nCOMPOSITION VARIATION (this candidate only): keep the place, the people and the moment exactly as specified, and keep from the CAMERA & FRAME contract its subject scale, screen placement and key background elements exactly. For this candidate only, its camera angle and height are deliberately left open: choose a camera position distinctly different from the obvious one, as a film director picking a second setup on the same blocking. Stay true to any explicit framing the SHOT TEXT itself dictates."
  },
  "shot_run_spend_attempt_count": 1,
  "readings": [
   {
    "label": "A",
    "direction": "인물이 화면 오른쪽의 문 너머 어두운 공간을 향해 시선을 고정하고 있음.",
    "built_space": "이전 컷의 흙담과 기와지붕, 단이 있는 장독대 대신 나무 판자 울타리와 맨땅에 놓인 항아리들이 배치되어 장소 연속성이 전혀 맞지 않음.",
    "entities": "이관회(한국인 남성, 청년, 짧은 검은 머리, 한복 의상)의 옆얼굴이 프롬프트와 레퍼런스에 맞게 묘사됨.",
    "hard_violations": [
     "[gpt-high] 지시문과 장소 참조에 없는 빨랫줄의 걸린 천 및 우측의 천으로 덮인 별도 물체를 눈에 띄는 소품으로 추가했다."
    ],
    "physics": "인물이 지면에 안정적으로 서서 올바른 자세를 유지하고 있음."
   },
   {
    "label": "B",
    "direction": "인물이 화면 오른쪽의 사립문 밖 어둠을 향해 시선을 고정하고 있음.",
    "built_space": "레퍼런스와 일치하는 흙담, 단이 있는 장독대 및 항아리들이 정확한 위치에 올바르게 배치됨.",
    "entities": "이관회(한국인 남성, 청년, 짧은 검은 머리, 한복 의상)의 옆얼굴이 프롬프트와 레퍼런스에 맞게 묘사됨.",
    "hard_violations": [
     "[gemini-pro] 화면 하단에 금지된 텍스트 자막(프롬프트 내용)이 그대로 노출됨 (leaked markers/diagrams/text)",
     "[gpt-high] 화면 하단에 읽을 수 있는 한국어 자막이 크게 삽입되어 무문자·무자막 조건을 직접 위반한다."
    ],
    "physics": "인물이 지면에 안정적으로 서서 올바른 자세를 유지하고 있음."
   }
  ],
  "cross_model_order": {
   "policy": "cross_model_order_v2_slot0_fwd_slot1_rev_parallel2",
   "slots": [
    {
     "model": "gemini-pro",
     "order": "forward",
     "display_to_canonical": {
      "A": "A",
      "B": "B"
     },
     "raw": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 4,
        "verdict_ko": "텍스트 자막이 없어 치명적 위반은 피했으나, 이전 컷에 고정된 흙담과 장독대 배경을 나무 울타리와 맨땅으로 완전히 다르게 그려 장소 연속성에서 크게 감점됨."
       },
       {
        "label": "B",
        "score": 0,
        "verdict_ko": "인물과 지정된 배경(장독대, 흙담)은 레퍼런스를 훌륭하게 따랐으나, 화면 하단에 프롬프트 텍스트가 그대로 자막으로 노출되어 치명적인 규정 위반으로 실격됨."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "인물이 화면 오른쪽의 문 너머 어두운 공간을 향해 시선을 고정하고 있음.",
        "built_space": "이전 컷의 흙담과 기와지붕, 단이 있는 장독대 대신 나무 판자 울타리와 맨땅에 놓인 항아리들이 배치되어 장소 연속성이 전혀 맞지 않음.",
        "entities": "이관회(한국인 남성, 청년, 짧은 검은 머리, 한복 의상)의 옆얼굴이 프롬프트와 레퍼런스에 맞게 묘사됨.",
        "hard_violations": [],
        "physics": "인물이 지면에 안정적으로 서서 올바른 자세를 유지하고 있음."
       },
       {
        "label": "B",
        "direction": "인물이 화면 오른쪽의 사립문 밖 어둠을 향해 시선을 고정하고 있음.",
        "built_space": "레퍼런스와 일치하는 흙담, 단이 있는 장독대 및 항아리들이 정확한 위치에 올바르게 배치됨.",
        "entities": "이관회(한국인 남성, 청년, 짧은 검은 머리, 한복 의상)의 옆얼굴이 프롬프트와 레퍼런스에 맞게 묘사됨.",
        "hard_violations": [
         "화면 하단에 금지된 텍스트 자막(프롬프트 내용)이 그대로 노출됨 (leaked markers/diagrams/text)"
        ],
        "physics": "인물이 지면에 안정적으로 서서 올바른 자세를 유지하고 있음."
       }
      ],
      "all_candidates_fail": false
     },
     "normalized": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 4,
        "verdict_ko": "텍스트 자막이 없어 치명적 위반은 피했으나, 이전 컷에 고정된 흙담과 장독대 배경을 나무 울타리와 맨땅으로 완전히 다르게 그려 장소 연속성에서 크게 감점됨."
       },
       {
        "label": "B",
        "score": 0,
        "verdict_ko": "인물과 지정된 배경(장독대, 흙담)은 레퍼런스를 훌륭하게 따랐으나, 화면 하단에 프롬프트 텍스트가 그대로 자막으로 노출되어 치명적인 규정 위반으로 실격됨."
       }
      ],
      "all_candidates_fail": false,
      "readings": [
       {
        "label": "A",
        "direction": "인물이 화면 오른쪽의 문 너머 어두운 공간을 향해 시선을 고정하고 있음.",
        "built_space": "이전 컷의 흙담과 기와지붕, 단이 있는 장독대 대신 나무 판자 울타리와 맨땅에 놓인 항아리들이 배치되어 장소 연속성이 전혀 맞지 않음.",
        "entities": "이관회(한국인 남성, 청년, 짧은 검은 머리, 한복 의상)의 옆얼굴이 프롬프트와 레퍼런스에 맞게 묘사됨.",
        "hard_violations": [],
        "physics": "인물이 지면에 안정적으로 서서 올바른 자세를 유지하고 있음."
       },
       {
        "label": "B",
        "direction": "인물이 화면 오른쪽의 사립문 밖 어둠을 향해 시선을 고정하고 있음.",
        "built_space": "레퍼런스와 일치하는 흙담, 단이 있는 장독대 및 항아리들이 정확한 위치에 올바르게 배치됨.",
        "entities": "이관회(한국인 남성, 청년, 짧은 검은 머리, 한복 의상)의 옆얼굴이 프롬프트와 레퍼런스에 맞게 묘사됨.",
        "hard_violations": [
         "화면 하단에 금지된 텍스트 자막(프롬프트 내용)이 그대로 노출됨 (leaked markers/diagrams/text)"
        ],
        "physics": "인물이 지면에 안정적으로 서서 올바른 자세를 유지하고 있음."
       }
      ]
     },
     "ok": true
    },
    {
     "model": "gpt-high",
     "order": "reverse",
     "display_to_canonical": {
      "A": "B",
      "B": "A"
     },
     "raw": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 2,
        "verdict_ko": "옆얼굴 클로즈업과 장소 재질은 비교적 충실하지만, 금지된 한글 자막이 크게 들어갔고 시선도 뒤쪽 사립문 밖을 향하지 않는다."
       },
       {
        "label": "B",
        "score": 3,
        "verdict_ko": "다문 입과 어두운 문밖은 구현했지만, 클로즈업보다 넓은 구도와 변경된 장소·추가 소품 때문에 충실도가 낮아 두 후보 모두 재생성이 필요하다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "남자는 입을 다물고 화면 오른쪽을 수평으로 응시한다. 문 개구부는 얼굴 뒤쪽 중앙에 있어, 보이는 시선은 그 문밖이 아니라 오른쪽 화면 밖으로 향한다. 무기나 이동 물체는 없다.",
        "built_space": "기와를 얹은 흙담, 중앙 뒤쪽 문 개구부 하나와 목재 문기둥, 낮은 장독대가 보인다. 항아리는 왼쪽 두 개와 오른쪽 두 개가 부분적으로 확인된다. 인물은 장독대 앞에 서 있는 것으로 읽히며, 기와·흙담·옹기 재질은 이전 장면과 비교적 잘 이어진다. 다만 문이 인물 뒤에 놓여 문밖을 바라보는 배치는 성립하지 않는다. 반사는 없다.",
        "entities": "검은 가르마 머리의 한국인 청년 남성 한 명이며 얼굴과 거친 갈색 저고리는 인물 참조에 대체로 부합한다. 입은 닫혀 있고 웃음은 없다. 이전 장면의 여성이나 추가 인물은 없다. 옹기와 전통 담장은 있으나, 화면 하단에 지시문의 한국어 문장이 읽을 수 있는 자막으로 들어가 있다.",
        "hard_violations": [
         "화면 하단에 읽을 수 있는 한국어 자막이 크게 삽입되어 무문자·무자막 조건을 직접 위반한다."
        ],
        "physics": "목과 어깨가 자연스럽게 연결된 직립 상반신이다. 발은 클로즈업 밖이므로 지면 접촉은 확인할 수 없지만 부유를 나타내는 징후는 없다. 항아리들은 장독대에 놓여 있고 뚜껑은 항아리 입구에 얹혀 있다. 지지 없는 물체나 불가능한 관절은 보이지 않는다."
       },
       {
        "label": "B",
        "direction": "남자는 입을 다문 채 화면 오른쪽을 응시한다. 어두운 문 개구부는 얼굴보다 오른쪽이지만 공간상 뒤쪽에 있어, 옆으로 뻗는 시선이 실제 문밖에 닿는지는 분명하지 않다. 렌즈를 보지는 않으며 무기나 이동 물체도 없다.",
        "built_space": "중앙에 지붕과 상부 살대가 있는 문 개구부 하나, 양쪽 목재 기둥, 오른쪽 널빤지 울타리, 왼쪽 집 벽과 창호가 보인다. 왼쪽과 인물 뒤에는 일부만 보이는 것을 포함해 최소 여섯 개의 항아리가 있다. 오른쪽에는 천을 건 줄과 천으로 덮인 별도 물체가 추가되어 있다. 참조의 기와 얹은 흙담과 단정한 장독대 대신 다른 마당 구조를 보여 장소 연속성이 약하다. 얼굴뿐 아니라 몸통과 넓은 마당까지 포함해 요구된 클로즈업보다 넓다. 반사는 없다.",
        "entities": "검은 가르마 머리의 한국인 청년 남성 한 명이며, 얼굴과 낡은 갈색 저고리는 인물 참조와 대체로 맞는다. 다문 입과 사라진 웃음도 보인다. 옹기와 어두운 문밖은 구현되어 있지만 참조에 없는 빨랫줄의 천과 우측 덮개 물체가 추가되었다. 추가 인물이나 읽을 수 있는 문자는 없다.",
        "hard_violations": [
         "지시문과 장소 참조에 없는 빨랫줄의 걸린 천 및 우측의 천으로 덮인 별도 물체를 눈에 띄는 소품으로 추가했다."
        ],
        "physics": "인물의 상반신은 자연스럽게 직립하고 있으며 하체는 프레임 밖이다. 부유나 불가능한 자세는 보이지 않는다. 항아리들은 바닥이나 낮은 받침에 놓이고 뚜껑은 항아리에 지지된다. 걸린 천은 줄에 걸쳐져 있으며 우측 덮개도 아래 물체를 따라 처져 있어, 추가 소품이라는 문제와 별개로 물리적 지지는 표현되어 있다."
       }
      ],
      "all_candidates_fail": true
     },
     "normalized": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 2,
        "verdict_ko": "옆얼굴 클로즈업과 장소 재질은 비교적 충실하지만, 금지된 한글 자막이 크게 들어갔고 시선도 뒤쪽 사립문 밖을 향하지 않는다."
       },
       {
        "label": "A",
        "score": 3,
        "verdict_ko": "다문 입과 어두운 문밖은 구현했지만, 클로즈업보다 넓은 구도와 변경된 장소·추가 소품 때문에 충실도가 낮아 두 후보 모두 재생성이 필요하다."
       }
      ],
      "all_candidates_fail": true,
      "readings": [
       {
        "label": "B",
        "direction": "남자는 입을 다물고 화면 오른쪽을 수평으로 응시한다. 문 개구부는 얼굴 뒤쪽 중앙에 있어, 보이는 시선은 그 문밖이 아니라 오른쪽 화면 밖으로 향한다. 무기나 이동 물체는 없다.",
        "built_space": "기와를 얹은 흙담, 중앙 뒤쪽 문 개구부 하나와 목재 문기둥, 낮은 장독대가 보인다. 항아리는 왼쪽 두 개와 오른쪽 두 개가 부분적으로 확인된다. 인물은 장독대 앞에 서 있는 것으로 읽히며, 기와·흙담·옹기 재질은 이전 장면과 비교적 잘 이어진다. 다만 문이 인물 뒤에 놓여 문밖을 바라보는 배치는 성립하지 않는다. 반사는 없다.",
        "entities": "검은 가르마 머리의 한국인 청년 남성 한 명이며 얼굴과 거친 갈색 저고리는 인물 참조에 대체로 부합한다. 입은 닫혀 있고 웃음은 없다. 이전 장면의 여성이나 추가 인물은 없다. 옹기와 전통 담장은 있으나, 화면 하단에 지시문의 한국어 문장이 읽을 수 있는 자막으로 들어가 있다.",
        "hard_violations": [
         "화면 하단에 읽을 수 있는 한국어 자막이 크게 삽입되어 무문자·무자막 조건을 직접 위반한다."
        ],
        "physics": "목과 어깨가 자연스럽게 연결된 직립 상반신이다. 발은 클로즈업 밖이므로 지면 접촉은 확인할 수 없지만 부유를 나타내는 징후는 없다. 항아리들은 장독대에 놓여 있고 뚜껑은 항아리 입구에 얹혀 있다. 지지 없는 물체나 불가능한 관절은 보이지 않는다."
       },
       {
        "label": "A",
        "direction": "남자는 입을 다문 채 화면 오른쪽을 응시한다. 어두운 문 개구부는 얼굴보다 오른쪽이지만 공간상 뒤쪽에 있어, 옆으로 뻗는 시선이 실제 문밖에 닿는지는 분명하지 않다. 렌즈를 보지는 않으며 무기나 이동 물체도 없다.",
        "built_space": "중앙에 지붕과 상부 살대가 있는 문 개구부 하나, 양쪽 목재 기둥, 오른쪽 널빤지 울타리, 왼쪽 집 벽과 창호가 보인다. 왼쪽과 인물 뒤에는 일부만 보이는 것을 포함해 최소 여섯 개의 항아리가 있다. 오른쪽에는 천을 건 줄과 천으로 덮인 별도 물체가 추가되어 있다. 참조의 기와 얹은 흙담과 단정한 장독대 대신 다른 마당 구조를 보여 장소 연속성이 약하다. 얼굴뿐 아니라 몸통과 넓은 마당까지 포함해 요구된 클로즈업보다 넓다. 반사는 없다.",
        "entities": "검은 가르마 머리의 한국인 청년 남성 한 명이며, 얼굴과 낡은 갈색 저고리는 인물 참조와 대체로 맞는다. 다문 입과 사라진 웃음도 보인다. 옹기와 어두운 문밖은 구현되어 있지만 참조에 없는 빨랫줄의 천과 우측 덮개 물체가 추가되었다. 추가 인물이나 읽을 수 있는 문자는 없다.",
        "hard_violations": [
         "지시문과 장소 참조에 없는 빨랫줄의 걸린 천 및 우측의 천으로 덮인 별도 물체를 눈에 띄는 소품으로 추가했다."
        ],
        "physics": "인물의 상반신은 자연스럽게 직립하고 있으며 하체는 프레임 밖이다. 부유나 불가능한 자세는 보이지 않는다. 항아리들은 바닥이나 낮은 받침에 놓이고 뚜껑은 항아리에 지지된다. 걸린 천은 줄에 걸쳐져 있으며 우측 덮개도 아래 물체를 따라 처져 있어, 추가 소품이라는 문제와 별개로 물리적 지지는 표현되어 있다."
       }
      ]
     },
     "ok": true
    }
   ],
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "slot_winner_match": true,
   "slot_winner": {
    "gemini-pro": "A",
    "gpt-high": "A"
   },
   "route": "cross_slot_agree"
  },
  "dual": {
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "normalized": {
    "A": 2.0,
    "B": 0.667
   },
   "adjusted": {
    "A": 1.75,
    "B": 0.417
   },
   "violations": {
    "B": [
     "[gemini-pro] 화면 하단에 금지된 텍스트 자막(프롬프트 내용)이 그대로 노출됨 (leaked markers/diagrams/text)",
     "[gpt-high] 화면 하단에 읽을 수 있는 한국어 자막이 크게 삽입되어 무문자·무자막 조건을 직접 위반한다."
    ],
    "A": [
     "[gpt-high] 지시문과 장소 참조에 없는 빨랫줄의 걸린 천 및 우측의 천으로 덮인 별도 물체를 눈에 띄는 소품으로 추가했다."
    ]
   },
   "per_model_winner": {
    "gemini-pro": "A",
    "gpt-high": "A"
   },
   "agreed": true
  },
  "totals": {
   "A": 1750,
   "B": 417
  },
  "selected": "A",
  "ranking": [
   "A",
   "B"
  ],
  "verdicts": [
   {
    "label": "A",
    "score": 1750,
    "verdict_ko": "텍스트 자막이 없어 치명적 위반은 피했으나, 이전 컷에 고정된 흙담과 장독대 배경을 나무 울타리와 맨땅으로 완전히 다르게 그려 장소 연속성에서 크게 감점됨.  ★위반: [gpt-high] 지시문과 장소 참조에 없는 빨랫줄의 걸린 천 및 우측의 천으로 덮인 별도 물체를 눈에 띄는 소품으로 추가했다."
   },
   {
    "label": "B",
    "score": 417,
    "verdict_ko": "인물과 지정된 배경(장독대, 흙담)은 레퍼런스를 훌륭하게 따랐으나, 화면 하단에 프롬프트 텍스트가 그대로 자막으로 노출되어 치명적인 규정 위반으로 실격됨.  ★위반: [gemini-pro] 화면 하단에 금지된 텍스트 자막(프롬프트 내용)이 그대로 노출됨 (leaked markers/diagrams/text) / [gpt-high] 화면 하단에 읽을 수 있는 한국어 자막이 크게 삽입되어 무문자·무자막 조건을 직접 위반한다."
   }
  ],
  "refs": [
   {
    "label": "PREVIOUS SHOT STILL — a visually related earlier shot of this same place: the location's look, materials, fixed features and lighting mood are LOCKED to this photo; never copy its camera framing. The people visible in that photo are NOT in this shot: never carry their faces, bodies or clothing onto anyone here. Each person in THIS shot is defined solely by the PEOPLE list, their own CHARACTER REFERENCE images and their wardrobe notes. If this photo shows a character who cannot move (dead or unconscious), that character's exact pose, position and orientation are ALSO LOCKED — treat their body as a fixed prop of the set that only the camera moves around.",
    "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/db00517b-165b-4005-a287-f28936b59b95/images/272bd7af-9294-4490-98c6-88a40e9deacc/scene/recipe/S8sh6_sel.png",
    "asset_id": "51c25e07-06ff-4d26-84da-aeae7d9b51d9",
    "role": "prev_still"
   },
   {
    "label": "CHARACTER REFERENCE — 이관회: the exact person appearing in this shot; match face, hair and build exactly.",
    "path": "<bytes:933540>",
    "asset_id": "851b450f-ac84-47fe-ac5a-03a56f63fcae",
    "role": "character_ref"
   }
  ],
  "critique_skipped": true,
  "shot_run_produced": false,
  "shot_run_uid": "06aa140a-0cd0-7c5e-8c15-3c7cfe4a42da",
  "ref_mode": "prev+엔티티",
  "share_plan": {
   "ref_plan": "prev",
   "prev_anchor_tag": "S8sh6"
  }
 },
 "S8sh9::cine": {
  "applied": true,
  "attempted_at": "2026-09-09T01:29:21.337498+00:00",
  "fingerprint": "27da5e7807c0969c8a828a4cff871774f6ce33dc9c678e5f099294c66f5c457d",
  "fingerprint_version": 2,
  "provider": "mai",
  "endpoint": "openrouter/chat-completions",
  "model": "microsoft/mai-image-2.6",
  "pack": "24.202608252115",
  "source_file": "S8sh9_sel.png",
  "source_sha256": "841b8a6bf24386d7d9d4d4b7ed10786d199eed58ad780c9ee995975010256713",
  "file": "S8sh9_cine.png",
  "staged_sha256": "5522dacd32cf576ae3963ffcfc71f2f0047cbb4c81a36a07a2488035fd84d586",
  "latency_ms": 21265
 },
 "S9sh1::signage": {
  "fp": "b42042dfa126301a",
  "inscriptions": [],
  "cues": [],
  "dropped": []
 },
 "era_assess::b22859c2ebd98e0c": {
  "subjects": [
   {
    "subject_native": "일제강점기 경찰 주재소 유치장 내부",
    "search_terms_native": [
     "일제강점기 주재소 유치장",
     "일제강점기 감옥 내부",
     "근대 경찰서 유치장",
     "서대문형무소 옥사 감방 내부"
    ],
    "language_lock_native": "반드시 한국어로만 검색해야 하며 다른 언어로 번역하거나 혼용하지 마십시오.",
    "reason_ko": "현대 서구식 교도소와 달리 일제강점기 주재소 특유의 목조 마루, 굵은 목재/철제 격자창 구조를 확인해야 합니다."
   }
  ],
  "subject_text": "보성 주재소 옥사 감방\n철창문으로 복도와 나뉜 어두운 감방. 벽과 맨바닥이 드러나고, 구석까지 짙은 그림자가 이어지는 단순한 공간이다.",
  "identity": "canonical",
  "scope_id": "L11",
  "scope_role": "location_interior",
  "scope_sha": "2641b9000b8db6af"
 },
 "era_fail::bb4bcad6b638c919": {
  "stage": "research",
  "subject": "일제강점기 경찰 주재소 유치장 내부",
  "terms": [
   "일제강점기 주재소 유치장",
   "일제강점기 감옥 내부",
   "근대 경찰서 유치장",
   "서대문형무소 옥사 감방 내부"
  ],
  "status": "no_usable",
  "queries": [
   [
    "일제강점기 경찰 주재소 유치장 내부 사진",
    "서대문형무소 역사관 옥사 감방 내부"
   ]
  ],
  "candidate_urls": [
   "https://pimg.mk.co.kr/meet/neds/2017/07/image_readbot_2017_504968_15011356852971980.jpg",
   "https://blog.kakaocdn.net/dna/bILR3V/btr2wtSPMlt/AAAAAAAAAAAAAAAAAAAAAJFG89qcRlh5UWB0mecHUSS0W9LydHKPkW8fLXOOipkZ/img.jpg?allow_ip=&allow_referer=&credential=yqXZFxpELC7KVnFOS48ylbz2pIh7yKj8&expires=1772290799&signature=KJk8JP%2FpdQFvc%2Bh4rCzSc1Coj6o%3D",
   "https://upload.wikimedia.org/wikipedia/commons/a/ac/Interrogation_room.jpg",
   "https://st5.depositphotos.com/35400196/68582/i/450/depositphotos_685821690-stock-photo-seodaemun-prison-history-hall-museum.jpg"
  ],
  "coarse": {
   "eligible": [],
   "chosen_index": 0,
   "reason": "종류·보임·기준을 다 만족하는 후보가 없다 — 이 라운드에선 안 고른다",
   "single_judge": true,
   "rejected_judges": {}
  },
  "verdicts": [
   {
    "index": 1,
    "object_type_match": "yes",
    "visible": true,
    "criteria_match": "unsure",
    "similarity": 90
   },
   {
    "index": 2,
    "object_type_match": "no",
    "visible": true,
    "criteria_match": "unsure",
    "similarity": 40
   },
   {
    "index": 3,
    "object_type_match": "no",
    "visible": true,
    "criteria_match": "unsure",
    "similarity": 40
   },
   {
    "index": 4,
    "object_type_match": "yes",
    "visible": true,
    "criteria_match": "unsure",
    "similarity": 95
   }
  ],
  "chosen_reason_ko": "",
  "attempts": 5
 },
 "S9sh1::bgfirst_bg": {
  "input_fingerprint": "a0a95306b64bc24d",
  "prompt": "Create the EMPTY BACKGROUND PLATE for one film shot — NO PEOPLE, no figures, no body parts, no sketch lines, no arrows anywhere.\n\"Empty\" means no people only: KEEP the location's inherent occupants and stock that define the place — animals in an animal shelter, pen or farm, goods in a market, moored boats in a harbour — unless the shot text explicitly removes them.\nThe FIRST attached image is a thin-line storyboard sketch: use ONLY its camera angle, horizon, perspective and the placement/size of buildings and set masses — ignore the sketched people and arrows entirely. The SECOND attached image (LOCATION PHOTOGRAPH) is the real place: take its architecture, materials, signage and fixed features, and RE-PROJECT them into the sketch's camera. If the photograph's camera differs from the sketch's, the sketch's camera wins.\nHUMAN-SCALE CALIBRATION: derive every structure's true size from human-scale elements — a door ≈ 2m, a window ≈ 1–1.5m wide, one storey ≈ 2.5–3m; never inflate a small structure or shrink a large one.\n\nSHOT TEXT this background must serve (Korean): 일본 순사가 휘두른 굵은 몽둥이가 안규홍의 뺨에 정통으로 격돌한 찰나.\n\nLOCATION (lock): Inside the police outpost's barred holding cell at night, in the open floor area used to restrain a prisoner; the light source is unspecified.\n\nTIME OF DAY (lock): night.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: close-up\n- KEY BACKGROUND ELEMENTS: 몽둥이 (In contact with 안규홍's cheek during the strike) — Crosses obliquely from the officer's side toward the visible cheek; used as Forms a narrow action line connecting the striking arm and the impact without exaggerated foreground scale; 감방 벽 (Encloses the interrogation space); used as A subdued background fragment keeps the blow spatially grounded.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Restrained ambient illumination appropriate to the nighttime cell keeps the impact readable without stylized highlights or heightened blood coloration.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nRender ONE photorealistic empty location photograph, 16:9, that this shot can be staged inside later. No readable writing anywhere: surfaces that would carry writing may be present, but stage any wording out of legibility — an oblique angle, distance, shallow focus. No captions, watermarks or overlay text.",
  "effective_prompt": "Create the EMPTY BACKGROUND PLATE for one film shot — NO PEOPLE, no figures, no body parts, no sketch lines, no arrows anywhere.\n\"Empty\" means no people only: KEEP the location's inherent occupants and stock that define the place — animals in an animal shelter, pen or farm, goods in a market, moored boats in a harbour — unless the shot text explicitly removes them.\nThe FIRST attached image is a thin-line storyboard sketch: use ONLY its camera angle, horizon, perspective and the placement/size of buildings and set masses — ignore the sketched people and arrows entirely. The SECOND attached image (LOCATION PHOTOGRAPH) is the real place: take its architecture, materials, signage and fixed features, and RE-PROJECT them into the sketch's camera. If the photograph's camera differs from the sketch's, the sketch's camera wins.\nHUMAN-SCALE CALIBRATION: derive every structure's true size from human-scale elements — a door ≈ 2m, a window ≈ 1–1.5m wide, one storey ≈ 2.5–3m; never inflate a small structure or shrink a large one.\n\nSHOT TEXT this background must serve (Korean): 일본 순사가 휘두른 굵은 몽둥이가 안규홍의 뺨에 정통으로 격돌한 찰나.\n\nLOCATION (lock): Inside the police outpost's barred holding cell at night, in the open floor area used to restrain a prisoner; the light source is unspecified.\n\nTIME OF DAY (lock): night.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: close-up\n- KEY BACKGROUND ELEMENTS: 몽둥이 (In contact with 안규홍's cheek during the strike) — Crosses obliquely from the officer's side toward the visible cheek; used as Forms a narrow action line connecting the striking arm and the impact without exaggerated foreground scale; 감방 벽 (Encloses the interrogation space); used as A subdued background fragment keeps the blow spatially grounded.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Restrained ambient illumination appropriate to the nighttime cell keeps the impact readable without stylized highlights or heightened blood coloration.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nRender ONE photorealistic empty location photograph, 16:9, that this shot can be staged inside later. No readable writing anywhere: surfaces that would carry writing may be present, but stage any wording out of legibility — an oblique angle, distance, shallow focus. No captions, watermarks or overlay text.",
  "bg_path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/db00517b-165b-4005-a287-f28936b59b95/images/272bd7af-9294-4490-98c6-88a40e9deacc/scene/recipe/S9sh1__bgfirst_bg.png",
  "asset_id": "2f626a3f-bfac-46d6-a478-76666e7dc34e",
  "input_asset_ids": [
   "22e137bb-feb9-4afc-9bbf-5052bc519354",
   "d577bc3d-221f-43c8-975f-f6d6ac63a63f"
  ]
 },
 "S9sh1": {
  "input_fingerprint": "93384f0efcf3a207",
  "prompt": "Create ONE FINAL photorealistic live-action film still of the moment below — early 20th-century Korea, Jeolla region; all people are Korean unless explicitly stated otherwise. TIME OF DAY (lock): night.\n\nSHOT TEXT (authoritative, Korean): 일본 순사가 휘두른 굵은 몽둥이가 안규홍의 뺨에 정통으로 격돌한 찰나.\n\nLOCATION (lock): Inside the police outpost's barred holding cell at night, in the open floor area used to restrain a prisoner; the light source is unspecified. The shot takes place here — the attached LOCATION PHOTOGRAPH shows the exact spot.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: close-up\n- KEY BACKGROUND ELEMENTS: 몽둥이 (In contact with 안규홍's cheek during the strike) — Crosses obliquely from the officer's side toward the visible cheek; used as Forms a narrow action line connecting the striking arm and the impact without exaggerated foreground scale; 감방 벽 (Encloses the interrogation space); used as A subdued background fragment keeps the blow spatially grounded.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Restrained ambient illumination appropriate to the nighttime cell keeps the impact readable without stylized highlights or heightened blood coloration.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): It is night inside the barred jail. 안규홍: He is under restraint with his arms forced behind his back and blood in his mouth.\n\nTHE HAND THAT IS DOING THIS: the object at the centre of this shot is being held, operated, read, aimed or handed over by 일본 순사 right now, so 일본 순사's hand — and as much of the wrist and forearm as the framing reaches — is in the frame, gripping or resting on that object exactly the way its real use requires. Match that hand to 일본 순사: its size, build, skin, age, grooming, sleeve and anything worn on it belong to that person and to no one else.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 안규홍 (한국인 남성, 청년 나이대, 검은 머리, 단단히 묶은 머리) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.",
  "roll_prompts": {
   "A": "Stage the shot. The FIRST attached image (SHOT BACKGROUND) is the finished empty background of this shot — keep it EXACTLY: its camera, perspective, architecture, lighting and every fixed feature stay untouched. The SECOND attached image (LAYOUT SKETCH) tells you ONLY where the people go: each sketched person's position, screen size, pose and the gaze/motion arrows. Ignore the sketch's background lines. The CHARACTER REFERENCE photographs show the real people.\nPlace the real people into the background at exactly the sketched positions, sizes and poses, following the arrow directions. No sketch lines or arrows may remain.\n\nCreate ONE FINAL photorealistic live-action film still of the moment below — early 20th-century Korea, Jeolla region; all people are Korean unless explicitly stated otherwise. TIME OF DAY (lock): night.\n\nSHOT TEXT (authoritative, Korean): 일본 순사가 휘두른 굵은 몽둥이가 안규홍의 뺨에 정통으로 격돌한 찰나.\n\nLOCATION (lock): Inside the police outpost's barred holding cell at night, in the open floor area used to restrain a prisoner; the light source is unspecified. The shot takes place here — the FIRST attached image (SHOT BACKGROUND) is this exact place, already built: its ground, structures, horizon, materials and lighting are the finished truth of this location and must not be redesigned or replaced. No location photograph is attached — read the place from that image alone, and add no scenery, structure, vehicle or fixture that it does not already show. This lock governs the place only; the figures in the shot follow the staging and pose instructions.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: close-up\n- KEY BACKGROUND ELEMENTS: 몽둥이 (In contact with 안규홍's cheek during the strike) — Crosses obliquely from the officer's side toward the visible cheek; used as Forms a narrow action line connecting the striking arm and the impact without exaggerated foreground scale; 감방 벽 (Encloses the interrogation space); used as A subdued background fragment keeps the blow spatially grounded.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Restrained ambient illumination appropriate to the nighttime cell keeps the impact readable without stylized highlights or heightened blood coloration.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): It is night inside the barred jail. 안규홍: He is under restraint with his arms forced behind his back and blood in his mouth.\n\nTHE HAND THAT IS DOING THIS: the object at the centre of this shot is being held, operated, read, aimed or handed over by 일본 순사 right now, so 일본 순사's hand — and as much of the wrist and forearm as the framing reaches — is in the frame, gripping or resting on that object exactly the way its real use requires. Match that hand to 일본 순사: its size, build, skin, age, grooming, sleeve and anything worn on it belong to that person and to no one else.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 안규홍 (한국인 남성, 청년 나이대, 검은 머리, 단단히 묶은 머리) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.",
   "B": "Create ONE FINAL photorealistic live-action film still of the moment below — early 20th-century Korea, Jeolla region; all people are Korean unless explicitly stated otherwise. TIME OF DAY (lock): night.\n\nSHOT TEXT (authoritative, Korean): 일본 순사가 휘두른 굵은 몽둥이가 안규홍의 뺨에 정통으로 격돌한 찰나.\n\nLOCATION (lock): Inside the police outpost's barred holding cell at night, in the open floor area used to restrain a prisoner; the light source is unspecified. The shot takes place here — the attached LOCATION PHOTOGRAPH shows the exact spot.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: close-up\n- KEY BACKGROUND ELEMENTS: 몽둥이 (In contact with 안규홍's cheek during the strike) — Crosses obliquely from the officer's side toward the visible cheek; used as Forms a narrow action line connecting the striking arm and the impact without exaggerated foreground scale; 감방 벽 (Encloses the interrogation space); used as A subdued background fragment keeps the blow spatially grounded.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Restrained ambient illumination appropriate to the nighttime cell keeps the impact readable without stylized highlights or heightened blood coloration.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): It is night inside the barred jail. 안규홍: He is under restraint with his arms forced behind his back and blood in his mouth.\n\nTHE HAND THAT IS DOING THIS: the object at the centre of this shot is being held, operated, read, aimed or handed over by 일본 순사 right now, so 일본 순사's hand — and as much of the wrist and forearm as the framing reaches — is in the frame, gripping or resting on that object exactly the way its real use requires. Match that hand to 일본 순사: its size, build, skin, age, grooming, sleeve and anything worn on it belong to that person and to no one else.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 안규홍 (한국인 남성, 청년 나이대, 검은 머리, 단단히 묶은 머리) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.\n\nCOMPOSITION VARIATION (this candidate only): keep the place, the people and the moment exactly as specified, and keep from the CAMERA & FRAME contract its subject scale, screen placement and key background elements exactly. For this candidate only, its camera angle and height are deliberately left open: choose a camera position distinctly different from the obvious one, as a film director picking a second setup on the same blocking. Stay true to any explicit framing the SHOT TEXT itself dictates."
  },
  "roll_refs": {
   "A": [
    {
     "label": "SHOT BACKGROUND",
     "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/db00517b-165b-4005-a287-f28936b59b95/images/272bd7af-9294-4490-98c6-88a40e9deacc/scene/recipe/S9sh1__bgfirst_bg.png",
     "asset_id": "2f626a3f-bfac-46d6-a478-76666e7dc34e",
     "role": "bgfirst_bg"
    },
    {
     "label": "LAYOUT SKETCH (people placement only)",
     "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/db00517b-165b-4005-a287-f28936b59b95/images/272bd7af-9294-4490-98c6-88a40e9deacc/conti/conti_S9sh1.png",
     "asset_id": "22e137bb-feb9-4afc-9bbf-5052bc519354",
     "role": "conti_light"
    },
    {
     "label": "CHARACTER REFERENCE — 안규홍: the exact person appearing in this shot; match face, hair and build exactly.",
     "path": "<bytes:1028384>",
     "asset_id": "3b5894db-ce55-4f34-b966-5e725c3aa952",
     "role": "character_ref"
    }
   ],
   "B": [
    {
     "label": "LOCATION PHOTOGRAPH — the exact place of this shot: its architecture, materials, fixed features and lighting mood are spatial truth; stage the moment inside this place. Never copy its camera framing.",
     "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/db00517b-165b-4005-a287-f28936b59b95/episodes/272bd7af-9294-4490-98c6-88a40e9deacc/images/background_chain/L11B01.png",
     "asset_id": "d577bc3d-221f-43c8-975f-f6d6ac63a63f",
     "role": "location_plate"
    },
    {
     "label": "CHARACTER REFERENCE — 안규홍: the exact person appearing in this shot; match face, hair and build exactly.",
     "path": "<bytes:1028384>",
     "asset_id": "3b5894db-ce55-4f34-b966-5e725c3aa952",
     "role": "character_ref"
    }
   ]
  },
  "shot_run_spend_attempt_count": 1,
  "judge_flip_skipped": "judge_fn_owns_order",
  "initial_roll_all_fail": true,
  "readings": [
   {
    "label": "A",
    "direction": "안규홍은 왼쪽 위를 향해 시선을 둠. 몽둥이가 뺨에 적중하는 타격의 방향성이나 액션이 아예 존재하지 않음.",
    "built_space": "감방 내부. 시멘트 벽과 열린 철창 문이 배경에 적절히 위치함.",
    "entities": "안규홍의 외모와 의상은 레퍼런스와 일치하나 입안에 피가 없음. 왼쪽 전경에 순사의 몸통과 빈 손이 보이며 몽둥이는 없음.",
    "hard_violations": [],
    "physics": "안규홍은 팔을 뒤로 한 채 무릎을 꿇고 상체를 바닥에 지탱하고 있음."
   },
   {
    "label": "B",
    "direction": "안규홍은 오른쪽을 응시함. 순사의 몽둥이는 뺨을 향해 휘둘러지지 않고 아래로 향한 채 들려 있음.",
    "built_space": "감방 내부. 배경 전체가 철창으로 둘러싸여 있으며 공간적 배치는 무난함.",
    "entities": "안규홍의 얼굴과 의상은 레퍼런스와 일치하나 입안의 피는 보이지 않음. 오른쪽 전경에 제복을 입은 순사의 팔과 몽둥이가 나타남.",
    "hard_violations": [],
    "physics": "안규홍은 팔을 뒤로 결박당한 채 바닥에 앉아 몸을 지탱함. 순사의 손이 몽둥이를 물리적으로 올바르게 쥐고 있음."
   }
  ],
  "cross_model_order": {
   "policy": "cross_model_order_v2_slot0_fwd_slot1_rev_parallel2",
   "slots": [
    {
     "model": "gemini-pro",
     "order": "forward",
     "display_to_canonical": {
      "A": "A",
      "B": "B"
     },
     "raw": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 4,
        "verdict_ko": "순사의 손에 몽둥이 소품은 존재하나, 몽둥이가 안규홍의 뺨에 적중하는 가장 중요한 찰나의 액션이 전혀 구현되지 않았습니다."
       },
       {
        "label": "A",
        "score": 3,
        "verdict_ko": "몽둥이가 뺨에 적중하는 핵심 액션이 누락되었으며, 순사의 손에 몽둥이조차 들려 있지 않아 제시문을 크게 벗어났습니다."
       }
      ],
      "readings": [
       {
        "label": "B",
        "direction": "안규홍은 오른쪽을 응시함. 순사의 몽둥이는 뺨을 향해 휘둘러지지 않고 아래로 향한 채 들려 있음.",
        "built_space": "감방 내부. 배경 전체가 철창으로 둘러싸여 있으며 공간적 배치는 무난함.",
        "entities": "안규홍의 얼굴과 의상은 레퍼런스와 일치하나 입안의 피는 보이지 않음. 오른쪽 전경에 제복을 입은 순사의 팔과 몽둥이가 나타남.",
        "hard_violations": [],
        "physics": "안규홍은 팔을 뒤로 결박당한 채 바닥에 앉아 몸을 지탱함. 순사의 손이 몽둥이를 물리적으로 올바르게 쥐고 있음."
       },
       {
        "label": "A",
        "direction": "안규홍은 왼쪽 위를 향해 시선을 둠. 몽둥이가 뺨에 적중하는 타격의 방향성이나 액션이 아예 존재하지 않음.",
        "built_space": "감방 내부. 시멘트 벽과 열린 철창 문이 배경에 적절히 위치함.",
        "entities": "안규홍의 외모와 의상은 레퍼런스와 일치하나 입안에 피가 없음. 왼쪽 전경에 순사의 몸통과 빈 손이 보이며 몽둥이는 없음.",
        "hard_violations": [],
        "physics": "안규홍은 팔을 뒤로 한 채 무릎을 꿇고 상체를 바닥에 지탱하고 있음."
       }
      ],
      "all_candidates_fail": true
     },
     "normalized": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 4,
        "verdict_ko": "순사의 손에 몽둥이 소품은 존재하나, 몽둥이가 안규홍의 뺨에 적중하는 가장 중요한 찰나의 액션이 전혀 구현되지 않았습니다."
       },
       {
        "label": "A",
        "score": 3,
        "verdict_ko": "몽둥이가 뺨에 적중하는 핵심 액션이 누락되었으며, 순사의 손에 몽둥이조차 들려 있지 않아 제시문을 크게 벗어났습니다."
       }
      ],
      "all_candidates_fail": true,
      "readings": [
       {
        "label": "B",
        "direction": "안규홍은 오른쪽을 응시함. 순사의 몽둥이는 뺨을 향해 휘둘러지지 않고 아래로 향한 채 들려 있음.",
        "built_space": "감방 내부. 배경 전체가 철창으로 둘러싸여 있으며 공간적 배치는 무난함.",
        "entities": "안규홍의 얼굴과 의상은 레퍼런스와 일치하나 입안의 피는 보이지 않음. 오른쪽 전경에 제복을 입은 순사의 팔과 몽둥이가 나타남.",
        "hard_violations": [],
        "physics": "안규홍은 팔을 뒤로 결박당한 채 바닥에 앉아 몸을 지탱함. 순사의 손이 몽둥이를 물리적으로 올바르게 쥐고 있음."
       },
       {
        "label": "A",
        "direction": "안규홍은 왼쪽 위를 향해 시선을 둠. 몽둥이가 뺨에 적중하는 타격의 방향성이나 액션이 아예 존재하지 않음.",
        "built_space": "감방 내부. 시멘트 벽과 열린 철창 문이 배경에 적절히 위치함.",
        "entities": "안규홍의 외모와 의상은 레퍼런스와 일치하나 입안에 피가 없음. 왼쪽 전경에 순사의 몸통과 빈 손이 보이며 몽둥이는 없음.",
        "hard_violations": [],
        "physics": "안규홍은 팔을 뒤로 한 채 무릎을 꿇고 상체를 바닥에 지탱하고 있음."
       }
      ]
     },
     "ok": true
    },
    {
     "model": "gpt-high",
     "order": "reverse",
     "display_to_canonical": {
      "A": "B",
      "B": "A"
     },
     "raw": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 3,
        "verdict_ko": "상대적으로 가까운 구도와 순사가 쥔 몽둥이는 맞지만, 몽둥이가 뺨과 떨어져 아래를 향하므로 핵심인 정통 격돌 순간을 구현하지 못했다."
       },
       {
        "label": "B",
        "score": 2,
        "verdict_ko": "등 뒤로 강제된 팔과 감방 내부는 읽히지만, 몽둥이가 아예 없고 허리까지 넓게 잡아 타격 순간과 클로즈업 지시 모두에서 벗어났다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "안규홍은 화면 오른쪽 위의 순사 쪽을 바라본다. 순사의 손에 들린 몽둥이는 오른쪽에서 왼쪽 아래로 내려가며 안규홍의 몸 옆 빈 공간을 향한다. 뺨에는 닿지 않고, 타격 팔과 뺨을 연결하는 사선도 형성하지 않는다.",
        "built_space": "뒤쪽에 세로 철봉과 가로 보강대가 있는 철창 면이 있고, 오른쪽에 빗장이 달린 철창문 한 개가 보인다. 왼쪽에는 거친 회벽과 어두운 하단 띠가 있다. 재료와 문 구성은 장소 참고와 유사하지만, 인물 뒤로 철창이 놓여 감방 내부의 열린 바닥에 있는지는 확정하기 어렵다. 반사면이나 중복 문은 보이지 않는다.",
        "entities": "안규홍은 젊은 동아시아계 남성으로 보이며, 검은 머리를 뒤로 단단히 묶었고 참고와 유사한 얼굴 및 때 묻은 옅은색 저고리를 갖췄다. 오른쪽에는 순사로 읽히는 제복 차림 인물의 몸통, 소매, 성인 손이 보인다. 얼굴이 없어 순사의 나이와 국적은 확인할 수 없다. 나무 몽둥이 한 개는 있으나 입가의 피는 보이지 않는다. 읽을 수 있는 글자는 없다.",
        "hard_violations": [],
        "physics": "몽둥이는 순사의 손이 손잡이를 감싸 쥐어 지지한다. 안규홍의 상체는 곧게 서 있고 충격으로 꺾이거나 돌아가는 반응이 없다. 아래팔과 하체는 화면 밖이라 구속 방식과 바닥 지지점은 확인할 수 없지만, 공중에 떠 있는 모습은 아니다. 보이는 위팔은 몸 옆으로 내려와 등 뒤로 강제된 상태가 뚜렷하지 않다."
       },
       {
        "label": "B",
        "direction": "안규홍은 고개를 숙이고 화면 왼쪽의 순사 쪽으로 눈을 돌린다. 순사의 보이는 손은 아래로 늘어져 있으며 아무것도 쥐지 않았다. 몽둥이가 없어 뺨을 향하는 타격선이나 접촉점 자체가 없다.",
        "built_space": "뒤에는 거친 회벽과 어두운 하단 띠가 있고, 오른쪽에 철창문 한 개와 빗장, 인접한 철창 부분이 보인다. 안규홍은 벽과 철창 사이의 빈 바닥 영역에 낮게 자리하며, 순사는 왼쪽에 일부만 들어온다. 참고 장소의 재료와 감방 내부 배치는 대체로 부합한다. 허리와 넓은 바닥까지 보여 요구된 클로즈업보다 넓다. 불가능한 반사나 중복된 문은 없다.",
        "entities": "안규홍은 젊은 동아시아계 남성으로 보이고, 참고와 유사한 얼굴, 뒤로 묶은 검은 머리, 때 묻은 옅은색 저고리를 갖췄다. 입술은 닫혀 있으며 피는 식별되지 않는다. 왼쪽에는 짙은 제복 소매와 성인 손이 보이지만 얼굴이 없어 순사의 나이와 국적은 확인할 수 없다. 핵심 소품인 굵은 몽둥이는 없다. 추가 인물이나 읽을 수 있는 글자는 보이지 않는다.",
        "hard_violations": [],
        "physics": "안규홍의 양팔은 등 뒤로 돌아가 있어 구속 상태와 양립하며, 몸통은 앞으로 기울어 있다. 손목과 하체는 잘려 실제 결박 및 무릎·발의 접지는 확인할 수 없지만, 떠 있다고 볼 근거는 없다. 순사의 빈손은 소매에서 자연스럽게 내려온다. 타격을 전달하는 물체나 뺨 접촉이 없으므로 이 자세를 몽둥이가 격돌한 순간으로 읽을 수 없다."
       }
      ],
      "all_candidates_fail": true
     },
     "normalized": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 3,
        "verdict_ko": "상대적으로 가까운 구도와 순사가 쥔 몽둥이는 맞지만, 몽둥이가 뺨과 떨어져 아래를 향하므로 핵심인 정통 격돌 순간을 구현하지 못했다."
       },
       {
        "label": "A",
        "score": 2,
        "verdict_ko": "등 뒤로 강제된 팔과 감방 내부는 읽히지만, 몽둥이가 아예 없고 허리까지 넓게 잡아 타격 순간과 클로즈업 지시 모두에서 벗어났다."
       }
      ],
      "all_candidates_fail": true,
      "readings": [
       {
        "label": "B",
        "direction": "안규홍은 화면 오른쪽 위의 순사 쪽을 바라본다. 순사의 손에 들린 몽둥이는 오른쪽에서 왼쪽 아래로 내려가며 안규홍의 몸 옆 빈 공간을 향한다. 뺨에는 닿지 않고, 타격 팔과 뺨을 연결하는 사선도 형성하지 않는다.",
        "built_space": "뒤쪽에 세로 철봉과 가로 보강대가 있는 철창 면이 있고, 오른쪽에 빗장이 달린 철창문 한 개가 보인다. 왼쪽에는 거친 회벽과 어두운 하단 띠가 있다. 재료와 문 구성은 장소 참고와 유사하지만, 인물 뒤로 철창이 놓여 감방 내부의 열린 바닥에 있는지는 확정하기 어렵다. 반사면이나 중복 문은 보이지 않는다.",
        "entities": "안규홍은 젊은 동아시아계 남성으로 보이며, 검은 머리를 뒤로 단단히 묶었고 참고와 유사한 얼굴 및 때 묻은 옅은색 저고리를 갖췄다. 오른쪽에는 순사로 읽히는 제복 차림 인물의 몸통, 소매, 성인 손이 보인다. 얼굴이 없어 순사의 나이와 국적은 확인할 수 없다. 나무 몽둥이 한 개는 있으나 입가의 피는 보이지 않는다. 읽을 수 있는 글자는 없다.",
        "hard_violations": [],
        "physics": "몽둥이는 순사의 손이 손잡이를 감싸 쥐어 지지한다. 안규홍의 상체는 곧게 서 있고 충격으로 꺾이거나 돌아가는 반응이 없다. 아래팔과 하체는 화면 밖이라 구속 방식과 바닥 지지점은 확인할 수 없지만, 공중에 떠 있는 모습은 아니다. 보이는 위팔은 몸 옆으로 내려와 등 뒤로 강제된 상태가 뚜렷하지 않다."
       },
       {
        "label": "A",
        "direction": "안규홍은 고개를 숙이고 화면 왼쪽의 순사 쪽으로 눈을 돌린다. 순사의 보이는 손은 아래로 늘어져 있으며 아무것도 쥐지 않았다. 몽둥이가 없어 뺨을 향하는 타격선이나 접촉점 자체가 없다.",
        "built_space": "뒤에는 거친 회벽과 어두운 하단 띠가 있고, 오른쪽에 철창문 한 개와 빗장, 인접한 철창 부분이 보인다. 안규홍은 벽과 철창 사이의 빈 바닥 영역에 낮게 자리하며, 순사는 왼쪽에 일부만 들어온다. 참고 장소의 재료와 감방 내부 배치는 대체로 부합한다. 허리와 넓은 바닥까지 보여 요구된 클로즈업보다 넓다. 불가능한 반사나 중복된 문은 없다.",
        "entities": "안규홍은 젊은 동아시아계 남성으로 보이고, 참고와 유사한 얼굴, 뒤로 묶은 검은 머리, 때 묻은 옅은색 저고리를 갖췄다. 입술은 닫혀 있으며 피는 식별되지 않는다. 왼쪽에는 짙은 제복 소매와 성인 손이 보이지만 얼굴이 없어 순사의 나이와 국적은 확인할 수 없다. 핵심 소품인 굵은 몽둥이는 없다. 추가 인물이나 읽을 수 있는 글자는 보이지 않는다.",
        "hard_violations": [],
        "physics": "안규홍의 양팔은 등 뒤로 돌아가 있어 구속 상태와 양립하며, 몸통은 앞으로 기울어 있다. 손목과 하체는 잘려 실제 결박 및 무릎·발의 접지는 확인할 수 없지만, 떠 있다고 볼 근거는 없다. 순사의 빈손은 소매에서 자연스럽게 내려온다. 타격을 전달하는 물체나 뺨 접촉이 없으므로 이 자세를 몽둥이가 격돌한 순간으로 읽을 수 없다."
       }
      ]
     },
     "ok": true
    }
   ],
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "slot_winner_match": true,
   "slot_winner": {
    "gemini-pro": "B",
    "gpt-high": "B"
   },
   "route": "cross_slot_agree"
  },
  "dual": {
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "normalized": {
    "A": 1.417,
    "B": 2.0
   },
   "adjusted": {
    "A": 1.417,
    "B": 2.0
   },
   "violations": {},
   "per_model_winner": {
    "gemini-pro": "B",
    "gpt-high": "B"
   },
   "agreed": true
  },
  "totals": {
   "B": 2000,
   "A": 1417
  },
  "selected": "B",
  "ranking": [
   "B",
   "A"
  ],
  "verdicts": [
   {
    "label": "B",
    "score": 2000,
    "verdict_ko": "순사의 손에 몽둥이 소품은 존재하나, 몽둥이가 안규홍의 뺨에 적중하는 가장 중요한 찰나의 액션이 전혀 구현되지 않았습니다."
   },
   {
    "label": "A",
    "score": 1417,
    "verdict_ko": "몽둥이가 뺨에 적중하는 핵심 액션이 누락되었으며, 순사의 손에 몽둥이조차 들려 있지 않아 제시문을 크게 벗어났습니다."
   }
  ],
  "refs": [
   {
    "label": "LOCATION PHOTOGRAPH — the exact place of this shot: its architecture, materials, fixed features and lighting mood are spatial truth; stage the moment inside this place. Never copy its camera framing.",
    "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/db00517b-165b-4005-a287-f28936b59b95/episodes/272bd7af-9294-4490-98c6-88a40e9deacc/images/background_chain/L11B01.png",
    "asset_id": "d577bc3d-221f-43c8-975f-f6d6ac63a63f",
    "role": "location_plate"
   },
   {
    "label": "CHARACTER REFERENCE — 안규홍: the exact person appearing in this shot; match face, hair and build exactly.",
    "path": "<bytes:1028384>",
    "asset_id": "3b5894db-ce55-4f34-b966-5e725c3aa952",
    "role": "character_ref"
   }
  ],
  "critique_skipped": true,
  "needs_reshoot": true,
  "shot_run_produced": false,
  "shot_run_uid": "06aa140a-0e47-7ebf-807f-22014b679b8e",
  "bgfirst": {
   "bg_path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/db00517b-165b-4005-a287-f28936b59b95/images/272bd7af-9294-4490-98c6-88a40e9deacc/scene/recipe/S9sh1__bgfirst_bg.png",
   "bg_asset_id": "2f626a3f-bfac-46d6-a478-76666e7dc34e",
   "bg_record_key": "S9sh1::bgfirst_bg",
   "chain_winner": false,
   "authority": "plate"
  },
  "ref_mode": "플레이트+엔티티 (2택1: 무콘티 승)",
  "share_plan": {
   "ref_plan": "background"
  }
 },
 "S9sh1::cine": {
  "applied": true,
  "attempted_at": "2026-09-09T01:34:08.574782+00:00",
  "fingerprint": "4953135ece403a12e2582f780ddeb4d57f3b522663eeaf0945a430492a46fae1",
  "fingerprint_version": 2,
  "provider": "mai",
  "endpoint": "openrouter/chat-completions",
  "model": "microsoft/mai-image-2.6",
  "pack": "24.202608252115",
  "source_file": "S9sh1_sel.png",
  "source_sha256": "2ea065ba4cbb79340635044381f4a930fb6b6cd8771973b451bb72b069371161",
  "file": "S9sh1_cine.png",
  "staged_sha256": "5064f9a2f5dd4d3264a87e0600dda02289bf16aedc8d53c533077d30bc519e75",
  "latency_ms": 25998
 },
 "S9sh11::signage": {
  "fp": "ef45e1c9bdf57f86",
  "inscriptions": [],
  "cues": [],
  "dropped": []
 },
 "S9sh11": {
  "input_fingerprint": "5fbc208d053579d2",
  "prompt": "Create ONE FINAL photorealistic live-action film still of the moment below — early 20th-century Korea, Jeolla region; all people are Korean unless explicitly stated otherwise. TIME OF DAY (lock): night.\n\nSHOT TEXT (authoritative, Korean): 굵은 쇠창살을 양손으로 꽉 쥔 채 복도 밖을 뚫어져라 내다보는 안규홍의 굳은 정면.\n\nLOCATION (lock): Just inside the police outpost's cell bars, facing the prison corridor that recedes into darkness at night. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: medium shot\n- KEY BACKGROUND ELEMENTS: 감방 쇠창살 (Closed, with 안규홍 gripping the bars in both hands) — Seen obliquely from the corridor, with his face visible between the bars; used as Creates narrow foreground divisions and makes his confinement physically legible.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Maintain subdued nighttime ambient illumination, naturalistic contrast, and enough separation to distinguish his hands and face from the bars.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The clothing, hair and overall look of 안규홍 — who appear both in that photo and in this shot — are LOCKED to that photo. Anyone else visible in that photo is NOT in this shot: never carry their face, body or clothing onto anyone here. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The cell door is shut; a long blood trail remains along the corridor floor. 안규홍: He stands at the bars after the beating, with an injured mouth and labored breathing.\n\nTHE HAND THAT IS DOING THIS: the object at the centre of this shot is being held, operated, read, aimed or handed over by 안규홍 right now, so 안규홍's hand — and as much of the wrist and forearm as the framing reaches — is in the frame, gripping or resting on that object exactly the way its real use requires. Match that hand to 안규홍: its size, build, skin, age, grooming, sleeve and anything worn on it belong to that person and to no one else.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 안규홍 (한국인 남성, 청년 나이대, 검은 머리, 단단히 묶은 머리) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.",
  "roll_prompts": {
   "A": "Create ONE FINAL photorealistic live-action film still of the moment below — early 20th-century Korea, Jeolla region; all people are Korean unless explicitly stated otherwise. TIME OF DAY (lock): night.\n\nSHOT TEXT (authoritative, Korean): 굵은 쇠창살을 양손으로 꽉 쥔 채 복도 밖을 뚫어져라 내다보는 안규홍의 굳은 정면.\n\nLOCATION (lock): Just inside the police outpost's cell bars, facing the prison corridor that recedes into darkness at night. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: medium shot\n- KEY BACKGROUND ELEMENTS: 감방 쇠창살 (Closed, with 안규홍 gripping the bars in both hands) — Seen obliquely from the corridor, with his face visible between the bars; used as Creates narrow foreground divisions and makes his confinement physically legible.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Maintain subdued nighttime ambient illumination, naturalistic contrast, and enough separation to distinguish his hands and face from the bars.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The clothing, hair and overall look of 안규홍 — who appear both in that photo and in this shot — are LOCKED to that photo. Anyone else visible in that photo is NOT in this shot: never carry their face, body or clothing onto anyone here. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The cell door is shut; a long blood trail remains along the corridor floor. 안규홍: He stands at the bars after the beating, with an injured mouth and labored breathing.\n\nTHE HAND THAT IS DOING THIS: the object at the centre of this shot is being held, operated, read, aimed or handed over by 안규홍 right now, so 안규홍's hand — and as much of the wrist and forearm as the framing reaches — is in the frame, gripping or resting on that object exactly the way its real use requires. Match that hand to 안규홍: its size, build, skin, age, grooming, sleeve and anything worn on it belong to that person and to no one else.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 안규홍 (한국인 남성, 청년 나이대, 검은 머리, 단단히 묶은 머리) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.",
   "B": "Create ONE FINAL photorealistic live-action film still of the moment below — early 20th-century Korea, Jeolla region; all people are Korean unless explicitly stated otherwise. TIME OF DAY (lock): night.\n\nSHOT TEXT (authoritative, Korean): 굵은 쇠창살을 양손으로 꽉 쥔 채 복도 밖을 뚫어져라 내다보는 안규홍의 굳은 정면.\n\nLOCATION (lock): Just inside the police outpost's cell bars, facing the prison corridor that recedes into darkness at night. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: medium shot\n- KEY BACKGROUND ELEMENTS: 감방 쇠창살 (Closed, with 안규홍 gripping the bars in both hands) — Seen obliquely from the corridor, with his face visible between the bars; used as Creates narrow foreground divisions and makes his confinement physically legible.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Maintain subdued nighttime ambient illumination, naturalistic contrast, and enough separation to distinguish his hands and face from the bars.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The clothing, hair and overall look of 안규홍 — who appear both in that photo and in this shot — are LOCKED to that photo. Anyone else visible in that photo is NOT in this shot: never carry their face, body or clothing onto anyone here. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The cell door is shut; a long blood trail remains along the corridor floor. 안규홍: He stands at the bars after the beating, with an injured mouth and labored breathing.\n\nTHE HAND THAT IS DOING THIS: the object at the centre of this shot is being held, operated, read, aimed or handed over by 안규홍 right now, so 안규홍's hand — and as much of the wrist and forearm as the framing reaches — is in the frame, gripping or resting on that object exactly the way its real use requires. Match that hand to 안규홍: its size, build, skin, age, grooming, sleeve and anything worn on it belong to that person and to no one else.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 안규홍 (한국인 남성, 청년 나이대, 검은 머리, 단단히 묶은 머리) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.\n\nCOMPOSITION VARIATION (this candidate only): keep the place, the people and the moment exactly as specified, and keep from the CAMERA & FRAME contract its subject scale, screen placement and key background elements exactly. For this candidate only, its camera angle and height are deliberately left open: choose a camera position distinctly different from the obvious one, as a film director picking a second setup on the same blocking. Stay true to any explicit framing the SHOT TEXT itself dictates."
  },
  "shot_run_spend_attempt_count": 1,
  "initial_roll_all_fail": true,
  "readings": [
   {
    "label": "A",
    "direction": "안규홍의 시선은 정면을 향해 렌즈 너머를 뚫어져라 응시하고 있다.",
    "built_space": "카메라는 복도에서 감방을 비스듬히 바라보는 구도이며, 좌측 전경에 쇠창살과 감방 내부가, 우측 원경에 핏자국이 있는 복도가 뻗어 있어 프롬프트가 요구한 공간 묘사에 부합한다.",
    "entities": "안규홍의 의상, 상처, 얼굴 특징은 레퍼런스와 일치하지만, 화면 중앙에 동일한 인물이 흐릿하게 한 명 더 복제되어 등장한다.",
    "hard_violations": [
     "[gemini-pro] 인물 복제 (Duplicated body)",
     "[gpt-high] 안규홍의 얼굴과 상체를 두 명으로 복제하여, 등장인물 한 명만 허용한 장면에 추가 인물을 만들었습니다."
    ],
    "physics": "본래 인물이 쇠창살을 쥔 손과 서 있는 자세는 자연스러우나, 복제된 두 번째 인물은 물리적인 지지 기반이나 공간적 맥락 없이 허공에 겹쳐져 있다."
   },
   {
    "label": "B",
    "direction": "안규홍은 쇠창살 너머로 정면 카메라 쪽을 날카롭게 응시하고 있다.",
    "built_space": "전면에 쇠창살이 렌더링되었으나, 인물 뒤편으로 핏자국이 길게 이어진 복도가 나타난다. 이는 카메라가 복도에서 감방을 보는 것이 아니라, 감방 안에서 복도를 내다보는 구조로 생성된 것으로 지정된 카메라 위치 및 공간 구조에 완전히 위배된다.",
    "entities": "안규홍 한 명만 명확하게 등장하며, 묶은 머리, 피 묻은 입술, 더러워진 의상 등 인물에 대한 세부 묘사가 지시사항과 잘 일치한다.",
    "hard_violations": [
     "[gemini-pro] 지정된 카메라 위치 및 공간 구조의 완전한 위배 (Rendering committed viewpoint from ground the structure does not have)",
     "[gpt-high] 양옆 감방 사이로 이어지는 복도와 혈흔을 인물 뒤에 배치하여, 감방 안에서 복도를 내다봐야 할 안규홍을 복도 쪽에 세운 역전된 공간 연출입니다."
    ],
    "physics": "양손으로 창살을 단단히 쥐고 있는 손의 형태와 체중이 실린 자세 등 물리적인 지지와 접촉은 자연스럽고 안정적이다."
   }
  ],
  "cross_model_order": {
   "policy": "cross_model_order_v2_slot0_fwd_slot1_rev_parallel2",
   "slots": [
    {
     "model": "gemini-pro",
     "order": "forward",
     "display_to_canonical": {
      "A": "A",
      "B": "B"
     },
     "raw": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 3,
        "verdict_ko": "인물의 묘사와 감정 표현은 뛰어나지만, 카메라가 복도에 있어야 한다는 지시를 무시하고 인물 뒤로 복도를 배치하여 공간 구조를 완전히 뒤바꾼 치명적 오류가 있습니다."
       },
       {
        "label": "A",
        "score": 2,
        "verdict_ko": "복도와 감방의 공간 배치 및 비스듬한 앵글은 지시사항을 잘 따랐으나, 메인 인물이 두 명으로 복제되어 나타나는 치명적인 하드 위반이 발생했습니다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "안규홍의 시선은 정면을 향해 렌즈 너머를 뚫어져라 응시하고 있다.",
        "built_space": "카메라는 복도에서 감방을 비스듬히 바라보는 구도이며, 좌측 전경에 쇠창살과 감방 내부가, 우측 원경에 핏자국이 있는 복도가 뻗어 있어 프롬프트가 요구한 공간 묘사에 부합한다.",
        "entities": "안규홍의 의상, 상처, 얼굴 특징은 레퍼런스와 일치하지만, 화면 중앙에 동일한 인물이 흐릿하게 한 명 더 복제되어 등장한다.",
        "hard_violations": [
         "인물 복제 (Duplicated body)"
        ],
        "physics": "본래 인물이 쇠창살을 쥔 손과 서 있는 자세는 자연스러우나, 복제된 두 번째 인물은 물리적인 지지 기반이나 공간적 맥락 없이 허공에 겹쳐져 있다."
       },
       {
        "label": "B",
        "direction": "안규홍은 쇠창살 너머로 정면 카메라 쪽을 날카롭게 응시하고 있다.",
        "built_space": "전면에 쇠창살이 렌더링되었으나, 인물 뒤편으로 핏자국이 길게 이어진 복도가 나타난다. 이는 카메라가 복도에서 감방을 보는 것이 아니라, 감방 안에서 복도를 내다보는 구조로 생성된 것으로 지정된 카메라 위치 및 공간 구조에 완전히 위배된다.",
        "entities": "안규홍 한 명만 명확하게 등장하며, 묶은 머리, 피 묻은 입술, 더러워진 의상 등 인물에 대한 세부 묘사가 지시사항과 잘 일치한다.",
        "hard_violations": [
         "지정된 카메라 위치 및 공간 구조의 완전한 위배 (Rendering committed viewpoint from ground the structure does not have)"
        ],
        "physics": "양손으로 창살을 단단히 쥐고 있는 손의 형태와 체중이 실린 자세 등 물리적인 지지와 접촉은 자연스럽고 안정적이다."
       }
      ],
      "all_candidates_fail": true
     },
     "normalized": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 3,
        "verdict_ko": "인물의 묘사와 감정 표현은 뛰어나지만, 카메라가 복도에 있어야 한다는 지시를 무시하고 인물 뒤로 복도를 배치하여 공간 구조를 완전히 뒤바꾼 치명적 오류가 있습니다."
       },
       {
        "label": "A",
        "score": 2,
        "verdict_ko": "복도와 감방의 공간 배치 및 비스듬한 앵글은 지시사항을 잘 따랐으나, 메인 인물이 두 명으로 복제되어 나타나는 치명적인 하드 위반이 발생했습니다."
       }
      ],
      "all_candidates_fail": true,
      "readings": [
       {
        "label": "A",
        "direction": "안규홍의 시선은 정면을 향해 렌즈 너머를 뚫어져라 응시하고 있다.",
        "built_space": "카메라는 복도에서 감방을 비스듬히 바라보는 구도이며, 좌측 전경에 쇠창살과 감방 내부가, 우측 원경에 핏자국이 있는 복도가 뻗어 있어 프롬프트가 요구한 공간 묘사에 부합한다.",
        "entities": "안규홍의 의상, 상처, 얼굴 특징은 레퍼런스와 일치하지만, 화면 중앙에 동일한 인물이 흐릿하게 한 명 더 복제되어 등장한다.",
        "hard_violations": [
         "인물 복제 (Duplicated body)"
        ],
        "physics": "본래 인물이 쇠창살을 쥔 손과 서 있는 자세는 자연스러우나, 복제된 두 번째 인물은 물리적인 지지 기반이나 공간적 맥락 없이 허공에 겹쳐져 있다."
       },
       {
        "label": "B",
        "direction": "안규홍은 쇠창살 너머로 정면 카메라 쪽을 날카롭게 응시하고 있다.",
        "built_space": "전면에 쇠창살이 렌더링되었으나, 인물 뒤편으로 핏자국이 길게 이어진 복도가 나타난다. 이는 카메라가 복도에서 감방을 보는 것이 아니라, 감방 안에서 복도를 내다보는 구조로 생성된 것으로 지정된 카메라 위치 및 공간 구조에 완전히 위배된다.",
        "entities": "안규홍 한 명만 명확하게 등장하며, 묶은 머리, 피 묻은 입술, 더러워진 의상 등 인물에 대한 세부 묘사가 지시사항과 잘 일치한다.",
        "hard_violations": [
         "지정된 카메라 위치 및 공간 구조의 완전한 위배 (Rendering committed viewpoint from ground the structure does not have)"
        ],
        "physics": "양손으로 창살을 단단히 쥐고 있는 손의 형태와 체중이 실린 자세 등 물리적인 지지와 접촉은 자연스럽고 안정적이다."
       }
      ]
     },
     "ok": true
    },
    {
     "model": "gpt-high",
     "order": "reverse",
     "display_to_canonical": {
      "A": "B",
      "B": "A"
     },
     "raw": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 3,
        "verdict_ko": "한 사람의 정면 미디엄 숏과 양손 파지는 맞지만, 감방 안이 아니라 복도를 등지고 창살 앞에 선 공간 배치가 결정적으로 어긋납니다."
       },
       {
        "label": "B",
        "score": 1,
        "verdict_ko": "복도 옆 감방이라는 배치는 더 명확하지만, 안규홍의 얼굴과 몸을 두 명으로 복제한 치명적 위반입니다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "안규홍은 고개를 조금 숙이고 눈을 들어 창살 너머 카메라 쪽을 노려봅니다. 두 손은 서로 다른 세로 창살을 감쌉니다. 다만 보이는 복도는 시선이 향하는 앞쪽이 아니라 인물 뒤로 이어져, 복도 밖을 내다보는 관계가 뒤집혀 있습니다.",
        "built_space": "전경에는 약 아홉 개의 세로 창살과 하단 가로대, 굵은 문틀이 보이고 열린 틈은 없습니다. 인물 뒤 중앙에는 바닥의 긴 혈흔을 따라 어두워지는 통로가 있으며, 통로 양옆에 추가 철창 구획이 보입니다. 왼쪽 뒤에는 작은 벽등 하나가 있습니다. 이 구조에서는 인물이 감방 내부보다 복도에 서 있고 카메라가 철창 안에서 그를 보는 것으로 읽힙니다. 낡은 금속과 회벽은 참고 장소와 유사하지만 공간의 앞뒤 관계는 맞지 않습니다.",
        "entities": "인물은 한 명뿐이며, 참고 인물과 유사한 젊은 한국인 남성의 외모, 뒤로 단단히 넘겨 묶은 검은 머리, 오염된 밝은색 저고리와 가슴 매듭이 보입니다. 입술에는 작은 상처가 있고 얼굴은 굳어 있습니다. 두 손과 소매는 같은 인물에게 자연스럽게 이어집니다. 쇠창살과 긴 혈흔은 있지만 혈흔의 공간적 위치가 잘못되었습니다. 읽을 수 있는 글자는 없습니다.",
        "hard_violations": [
         "양옆 감방 사이로 이어지는 복도와 혈흔을 인물 뒤에 배치하여, 감방 안에서 복도를 내다봐야 할 안규홍을 복도 쪽에 세운 역전된 공간 연출입니다."
        ],
        "physics": "양손의 손가락과 엄지가 각각 창살을 감싸고 손목과 팔뚝이 소매 안으로 이어져 파지는 물리적으로 가능합니다. 철창은 문틀과 가로대로 지지됩니다. 하체와 발은 미디엄 숏 밖이므로 지면 접촉은 확인할 수 없지만, 상체에 공중 부양이나 불가능한 자세의 증거는 없습니다. 혈흔은 바닥 표면에 놓여 있습니다."
       },
       {
        "label": "B",
        "direction": "왼쪽 인물과 그 오른쪽 뒤의 동일한 얼굴이 모두 창살 밖 카메라 방향을 바라봅니다. 왼쪽 인물은 양손으로 창살을 잡고, 오른쪽 인물도 보이는 한 손으로 다른 창살을 잡습니다. 오른쪽으로 뻗은 복도 바깥을 바라보는 방향 자체는 성립하지만, 시선의 주체가 한 사람이 아니라 두 명입니다.",
        "built_space": "전경에는 다섯 개가량의 주요 세로 창살과 오른쪽의 굵은 문틀이 보이며, 뒤로 가로 보강대와 철창 구획이 이어집니다. 두 인물은 왼쪽 감방 쪽에 있고, 오른쪽에는 긴 혈흔이 놓인 복도 바닥이 어둠 속으로 이어집니다. 복도에서 감방을 비스듬히 보는 구도는 A보다 명확합니다. 회벽과 낡은 철재는 참고 장소와 유사하나 조명은 더 밝고 황색이 강합니다. 두 번째 얼굴이 생길 만한 거울이나 반사면은 없습니다.",
        "entities": "동일한 젊은 한국인 남성으로 보이는 얼굴과 상체가 두 개 있습니다. 둘 다 참고의 안규홍을 닮은 얼굴, 뒤로 넘긴 검은 머리, 오염된 밝은색 저고리와 입가 상처를 지닙니다. 총 세 개의 창살을 쥔 손이 보이고 손등에도 피가 묻어 있습니다. 필요한 철창과 복도 혈흔은 있지만, 허용된 단 한 명의 안규홍이라는 조건을 위반합니다. 읽을 수 있는 글자는 없습니다.",
        "hard_violations": [
         "안규홍의 얼굴과 상체를 두 명으로 복제하여, 등장인물 한 명만 허용한 장면에 추가 인물을 만들었습니다."
        ],
        "physics": "보이는 세 손은 각각 창살에 접촉해 감싸고 있으며 철창은 고정 문틀로 지지됩니다. 두 상체가 겹쳐 일부 팔의 연결은 가려지지만, 보이는 부분만으로 공중에 뜬 몸이나 손을 단정할 수는 없습니다. 발은 프레임 밖이고 혈흔은 복도 바닥에 놓여 있습니다. 두 번째 인물은 반사상이 아니라 별도의 몸으로 묘사되어 있습니다."
       }
      ],
      "all_candidates_fail": true
     },
     "normalized": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 3,
        "verdict_ko": "한 사람의 정면 미디엄 숏과 양손 파지는 맞지만, 감방 안이 아니라 복도를 등지고 창살 앞에 선 공간 배치가 결정적으로 어긋납니다."
       },
       {
        "label": "A",
        "score": 1,
        "verdict_ko": "복도 옆 감방이라는 배치는 더 명확하지만, 안규홍의 얼굴과 몸을 두 명으로 복제한 치명적 위반입니다."
       }
      ],
      "all_candidates_fail": true,
      "readings": [
       {
        "label": "B",
        "direction": "안규홍은 고개를 조금 숙이고 눈을 들어 창살 너머 카메라 쪽을 노려봅니다. 두 손은 서로 다른 세로 창살을 감쌉니다. 다만 보이는 복도는 시선이 향하는 앞쪽이 아니라 인물 뒤로 이어져, 복도 밖을 내다보는 관계가 뒤집혀 있습니다.",
        "built_space": "전경에는 약 아홉 개의 세로 창살과 하단 가로대, 굵은 문틀이 보이고 열린 틈은 없습니다. 인물 뒤 중앙에는 바닥의 긴 혈흔을 따라 어두워지는 통로가 있으며, 통로 양옆에 추가 철창 구획이 보입니다. 왼쪽 뒤에는 작은 벽등 하나가 있습니다. 이 구조에서는 인물이 감방 내부보다 복도에 서 있고 카메라가 철창 안에서 그를 보는 것으로 읽힙니다. 낡은 금속과 회벽은 참고 장소와 유사하지만 공간의 앞뒤 관계는 맞지 않습니다.",
        "entities": "인물은 한 명뿐이며, 참고 인물과 유사한 젊은 한국인 남성의 외모, 뒤로 단단히 넘겨 묶은 검은 머리, 오염된 밝은색 저고리와 가슴 매듭이 보입니다. 입술에는 작은 상처가 있고 얼굴은 굳어 있습니다. 두 손과 소매는 같은 인물에게 자연스럽게 이어집니다. 쇠창살과 긴 혈흔은 있지만 혈흔의 공간적 위치가 잘못되었습니다. 읽을 수 있는 글자는 없습니다.",
        "hard_violations": [
         "양옆 감방 사이로 이어지는 복도와 혈흔을 인물 뒤에 배치하여, 감방 안에서 복도를 내다봐야 할 안규홍을 복도 쪽에 세운 역전된 공간 연출입니다."
        ],
        "physics": "양손의 손가락과 엄지가 각각 창살을 감싸고 손목과 팔뚝이 소매 안으로 이어져 파지는 물리적으로 가능합니다. 철창은 문틀과 가로대로 지지됩니다. 하체와 발은 미디엄 숏 밖이므로 지면 접촉은 확인할 수 없지만, 상체에 공중 부양이나 불가능한 자세의 증거는 없습니다. 혈흔은 바닥 표면에 놓여 있습니다."
       },
       {
        "label": "A",
        "direction": "왼쪽 인물과 그 오른쪽 뒤의 동일한 얼굴이 모두 창살 밖 카메라 방향을 바라봅니다. 왼쪽 인물은 양손으로 창살을 잡고, 오른쪽 인물도 보이는 한 손으로 다른 창살을 잡습니다. 오른쪽으로 뻗은 복도 바깥을 바라보는 방향 자체는 성립하지만, 시선의 주체가 한 사람이 아니라 두 명입니다.",
        "built_space": "전경에는 다섯 개가량의 주요 세로 창살과 오른쪽의 굵은 문틀이 보이며, 뒤로 가로 보강대와 철창 구획이 이어집니다. 두 인물은 왼쪽 감방 쪽에 있고, 오른쪽에는 긴 혈흔이 놓인 복도 바닥이 어둠 속으로 이어집니다. 복도에서 감방을 비스듬히 보는 구도는 A보다 명확합니다. 회벽과 낡은 철재는 참고 장소와 유사하나 조명은 더 밝고 황색이 강합니다. 두 번째 얼굴이 생길 만한 거울이나 반사면은 없습니다.",
        "entities": "동일한 젊은 한국인 남성으로 보이는 얼굴과 상체가 두 개 있습니다. 둘 다 참고의 안규홍을 닮은 얼굴, 뒤로 넘긴 검은 머리, 오염된 밝은색 저고리와 입가 상처를 지닙니다. 총 세 개의 창살을 쥔 손이 보이고 손등에도 피가 묻어 있습니다. 필요한 철창과 복도 혈흔은 있지만, 허용된 단 한 명의 안규홍이라는 조건을 위반합니다. 읽을 수 있는 글자는 없습니다.",
        "hard_violations": [
         "안규홍의 얼굴과 상체를 두 명으로 복제하여, 등장인물 한 명만 허용한 장면에 추가 인물을 만들었습니다."
        ],
        "physics": "보이는 세 손은 각각 창살에 접촉해 감싸고 있으며 철창은 고정 문틀로 지지됩니다. 두 상체가 겹쳐 일부 팔의 연결은 가려지지만, 보이는 부분만으로 공중에 뜬 몸이나 손을 단정할 수는 없습니다. 발은 프레임 밖이고 혈흔은 복도 바닥에 놓여 있습니다. 두 번째 인물은 반사상이 아니라 별도의 몸으로 묘사되어 있습니다."
       }
      ]
     },
     "ok": true
    }
   ],
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "slot_winner_match": true,
   "slot_winner": {
    "gemini-pro": "B",
    "gpt-high": "B"
   },
   "route": "cross_slot_agree"
  },
  "dual": {
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "normalized": {
    "A": 1.0,
    "B": 2.0
   },
   "adjusted": {
    "A": 0.75,
    "B": 1.75
   },
   "violations": {
    "A": [
     "[gemini-pro] 인물 복제 (Duplicated body)",
     "[gpt-high] 안규홍의 얼굴과 상체를 두 명으로 복제하여, 등장인물 한 명만 허용한 장면에 추가 인물을 만들었습니다."
    ],
    "B": [
     "[gemini-pro] 지정된 카메라 위치 및 공간 구조의 완전한 위배 (Rendering committed viewpoint from ground the structure does not have)",
     "[gpt-high] 양옆 감방 사이로 이어지는 복도와 혈흔을 인물 뒤에 배치하여, 감방 안에서 복도를 내다봐야 할 안규홍을 복도 쪽에 세운 역전된 공간 연출입니다."
    ]
   },
   "per_model_winner": {
    "gemini-pro": "B",
    "gpt-high": "B"
   },
   "agreed": true
  },
  "totals": {
   "B": 1750,
   "A": 750
  },
  "selected": "B",
  "ranking": [
   "B",
   "A"
  ],
  "verdicts": [
   {
    "label": "B",
    "score": 1750,
    "verdict_ko": "인물의 묘사와 감정 표현은 뛰어나지만, 카메라가 복도에 있어야 한다는 지시를 무시하고 인물 뒤로 복도를 배치하여 공간 구조를 완전히 뒤바꾼 치명적 오류가 있습니다.  ★위반: [gemini-pro] 지정된 카메라 위치 및 공간 구조의 완전한 위배 (Rendering committed viewpoint from ground the structure does not have) / [gpt-high] 양옆 감방 사이로 이어지는 복도와 혈흔을 인물 뒤에 배치하여, 감방 안에서 복도를 내다봐야 할 안규홍을 복도 쪽에 세운 역전된 공간 연출입니다."
   },
   {
    "label": "A",
    "score": 750,
    "verdict_ko": "복도와 감방의 공간 배치 및 비스듬한 앵글은 지시사항을 잘 따랐으나, 메인 인물이 두 명으로 복제되어 나타나는 치명적인 하드 위반이 발생했습니다.  ★위반: [gemini-pro] 인물 복제 (Duplicated body) / [gpt-high] 안규홍의 얼굴과 상체를 두 명으로 복제하여, 등장인물 한 명만 허용한 장면에 추가 인물을 만들었습니다."
   }
  ],
  "refs": [
   {
    "label": "PREVIOUS SHOT STILL — a visually related earlier shot of this same place: the location's look, materials, fixed features and lighting mood are LOCKED to this photo; never copy its camera framing. The clothing, hair and overall look of 안규홍 — who appear both in that photo and in this shot — are LOCKED to that photo. Anyone else visible in that photo is NOT in this shot: never carry their face, body or clothing onto anyone here. If this photo shows a character who cannot move (dead or unconscious), that character's exact pose, position and orientation are ALSO LOCKED — treat their body as a fixed prop of the set that only the camera moves around.",
    "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/db00517b-165b-4005-a287-f28936b59b95/images/272bd7af-9294-4490-98c6-88a40e9deacc/scene/recipe/S9sh1_sel.png",
    "asset_id": "51b372cd-1619-4f4f-9343-5dc5062dc1d7",
    "role": "prev_still"
   },
   {
    "label": "CHARACTER REFERENCE — 안규홍: the exact person appearing in this shot; match face, hair and build exactly.",
    "path": "<bytes:1028384>",
    "asset_id": "3b5894db-ce55-4f34-b966-5e725c3aa952",
    "role": "character_ref"
   }
  ],
  "critique_skipped": true,
  "needs_reshoot": true,
  "shot_run_produced": false,
  "shot_run_uid": "06aa140c-4914-74cb-a875-96c3e33b9c16",
  "ref_mode": "prev+엔티티",
  "share_plan": {
   "ref_plan": "prev",
   "prev_anchor_tag": "S9sh1"
  }
 },
 "S9sh11::cine": {
  "applied": true,
  "attempted_at": "2026-09-09T01:35:59.981640+00:00",
  "fingerprint": "f46144a7c6fa51eb240b7148cfe2dfd674ddb754e6f4491988c60d44a3212d34",
  "fingerprint_version": 2,
  "provider": "mai",
  "endpoint": "openrouter/chat-completions",
  "model": "microsoft/mai-image-2.6",
  "pack": "24.202608252115",
  "source_file": "S9sh11_sel.png",
  "source_sha256": "ae2828e2a0a8b81c8b0d248878348ac4b383c74c511198590f5e0ac465f12c34",
  "file": "S9sh11_cine.png",
  "staged_sha256": "c1a60571ffe84f7bbd82b7cf78186389e266d220194ded0e7deb382f9e9dafc9",
  "latency_ms": 25949
 },
 "S9sh15::signage": {
  "fp": "8f083f0db84f3ba6",
  "inscriptions": [],
  "cues": [],
  "dropped": []
 },
 "S9sh15": {
  "input_fingerprint": "30c637df52745d3f",
  "prompt": "Create ONE FINAL photorealistic live-action film still of the moment below — early 20th-century Korea, Jeolla region; all people are Korean unless explicitly stated otherwise. TIME OF DAY (lock): night.\n\nSHOT TEXT (authoritative, Korean): 열린 감방 문밖으로 한쪽 발을 내디딘 안규홍의 뒷모습을 뚫어져라 응시하는 염재보의 눈빛.\n\nLOCATION (lock): The corner of the police outpost's cell facing its open barred doorway and dark nighttime corridor. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: wide shot\n- FRAME LAYOUT: open cell doorway and visible threshold in the middle-right of the frame, background.\n- KEY BACKGROUND ELEMENTS: 감방 문 (Open while 안규홍 steps out) — The doorway is seen obliquely from inside the cell, with the threshold visible beneath his departing foot; used as Anchors the separation between the departing man and the man remaining inside; 염재보의 짚신 (The worn ties remain beside his paused hands); used as A small detail beside his seated body preserves the interrupted activity; 감방 구석 (Occupied by the seated 염재보); used as Provides the stationary counterweight to the open doorway.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Keep the cell's subdued nighttime illumination continuous, allowing both the seated observer and departing figure to remain readable without a new light cue.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The people visible in that photo are NOT in this shot: never carry their faces, bodies or clothing onto anyone here. Each person in THIS shot is defined solely by the PEOPLE list, their own CHARACTER REFERENCE images and their wardrobe notes. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The cell door is open for release; the long corridor blood trail remains. 염재보: He remains seated with his frayed straw-shoe ties in hand.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 염재보 (한국인 남성, 성인 나이대, 자연스러운 검은 머리) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.",
  "roll_prompts": {
   "A": "Create ONE FINAL photorealistic live-action film still of the moment below — early 20th-century Korea, Jeolla region; all people are Korean unless explicitly stated otherwise. TIME OF DAY (lock): night.\n\nSHOT TEXT (authoritative, Korean): 열린 감방 문밖으로 한쪽 발을 내디딘 안규홍의 뒷모습을 뚫어져라 응시하는 염재보의 눈빛.\n\nLOCATION (lock): The corner of the police outpost's cell facing its open barred doorway and dark nighttime corridor. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: wide shot\n- FRAME LAYOUT: open cell doorway and visible threshold in the middle-right of the frame, background.\n- KEY BACKGROUND ELEMENTS: 감방 문 (Open while 안규홍 steps out) — The doorway is seen obliquely from inside the cell, with the threshold visible beneath his departing foot; used as Anchors the separation between the departing man and the man remaining inside; 염재보의 짚신 (The worn ties remain beside his paused hands); used as A small detail beside his seated body preserves the interrupted activity; 감방 구석 (Occupied by the seated 염재보); used as Provides the stationary counterweight to the open doorway.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Keep the cell's subdued nighttime illumination continuous, allowing both the seated observer and departing figure to remain readable without a new light cue.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The people visible in that photo are NOT in this shot: never carry their faces, bodies or clothing onto anyone here. Each person in THIS shot is defined solely by the PEOPLE list, their own CHARACTER REFERENCE images and their wardrobe notes. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The cell door is open for release; the long corridor blood trail remains. 염재보: He remains seated with his frayed straw-shoe ties in hand.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 염재보 (한국인 남성, 성인 나이대, 자연스러운 검은 머리) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.",
   "B": "Create ONE FINAL photorealistic live-action film still of the moment below — early 20th-century Korea, Jeolla region; all people are Korean unless explicitly stated otherwise. TIME OF DAY (lock): night.\n\nSHOT TEXT (authoritative, Korean): 열린 감방 문밖으로 한쪽 발을 내디딘 안규홍의 뒷모습을 뚫어져라 응시하는 염재보의 눈빛.\n\nLOCATION (lock): The corner of the police outpost's cell facing its open barred doorway and dark nighttime corridor. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: wide shot\n- FRAME LAYOUT: open cell doorway and visible threshold in the middle-right of the frame, background.\n- KEY BACKGROUND ELEMENTS: 감방 문 (Open while 안규홍 steps out) — The doorway is seen obliquely from inside the cell, with the threshold visible beneath his departing foot; used as Anchors the separation between the departing man and the man remaining inside; 염재보의 짚신 (The worn ties remain beside his paused hands); used as A small detail beside his seated body preserves the interrupted activity; 감방 구석 (Occupied by the seated 염재보); used as Provides the stationary counterweight to the open doorway.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Keep the cell's subdued nighttime illumination continuous, allowing both the seated observer and departing figure to remain readable without a new light cue.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The people visible in that photo are NOT in this shot: never carry their faces, bodies or clothing onto anyone here. Each person in THIS shot is defined solely by the PEOPLE list, their own CHARACTER REFERENCE images and their wardrobe notes. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The cell door is open for release; the long corridor blood trail remains. 염재보: He remains seated with his frayed straw-shoe ties in hand.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 염재보 (한국인 남성, 성인 나이대, 자연스러운 검은 머리) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.\n\nCOMPOSITION VARIATION (this candidate only): keep the place, the people and the moment exactly as specified, and keep from the CAMERA & FRAME contract its subject scale, screen placement and key background elements exactly. For this candidate only, its camera angle and height are deliberately left open: choose a camera position distinctly different from the obvious one, as a film director picking a second setup on the same blocking. Stay true to any explicit framing the SHOT TEXT itself dictates."
  },
  "shot_run_spend_attempt_count": 1,
  "readings": [
   {
    "label": "A",
    "direction": "염재보의 시선이 감방 밖으로 나가는 안규홍의 뒷모습을 정확하게 향하고 있다.",
    "built_space": "이전 샷 레퍼런스와 동일하게 천장부터 바닥까지 이어지는 쇠창살, 하단이 칠해진 벽, 콘크리트 바닥이 구현되었으며, 쇠창살 문이 열려 있고 복도에는 핏자국이 유지되어 있다.",
    "entities": "염재보는 캐릭터 레퍼런스와 일치하는 외모와 누비옷을 입고 구석에 앉아 짚신 끈을 쥐고 있다. 안규홍은 흰옷을 입고 한쪽 발을 내디딘 뒷모습으로 등장한다.",
    "hard_violations": [],
    "physics": "염재보는 바닥에 안정적으로 앉아 체중을 지탱하고 있으며, 안규홍은 걷는 자세로 발이 바닥에 자연스럽게 닿아 있다."
   },
   {
    "label": "B",
    "direction": "염재보가 문 방향을 보고 있으나, 나가는 안규홍의 뒷모습을 뚫어져라 응시하기보다는 다소 시선이 아래로 향해 있다.",
    "built_space": "이전 샷 레퍼런스의 쇠창살 감방이 아닌 두꺼운 나무 문이 생성되었으며, 문밖 복도 천장에는 시대 배경에 맞지 않는 현대식 튜브형 형광등이 달려 있다.",
    "entities": "염재보가 구석에 앉아 짚신을 들고 있으나, 밖으로 나가는 안규홍이 염재보의 캐릭터 레퍼런스와 완전히 동일한 누비옷과 삿갓을 착용하고 있어 인물 간 의상이 중복되었다.",
    "hard_violations": [
     "[gemini-pro] 발명된 사물(시대적 배경에 어긋나는 현대식 형광등 조명 생성)",
     "[gemini-pro] 장소 고정(Location lock) 위반(쇠창살 대신 엉뚱한 나무 문과 열쇠구멍 생성)",
     "[gpt-high] 장소가 고정한 전면 쇠창살문을 대부분 불투명한 목재 판문으로 교체해 핵심 고정 구조를 바꿨다.",
     "[gpt-high] 복도에 밝은 현대식 관형 조명을 추가해 시대 설정과 새로운 조명 단서를 금지한 조건을 위반했다."
    ],
    "physics": "두 인물 모두 바닥에 지탱되어 서 있거나 앉아 있다."
   }
  ],
  "cross_model_order": {
   "policy": "cross_model_order_v2_slot0_fwd_slot1_rev_parallel2",
   "slots": [
    {
     "model": "gemini-pro",
     "order": "forward",
     "display_to_canonical": {
      "A": "A",
      "B": "B"
     },
     "raw": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 8,
        "verdict_ko": "레퍼런스에 맞는 쇠창살 감방과 복도의 핏자국을 정확히 구현했으며, 밖으로 나가는 인물을 응시하는 염재보의 시선과 자세를 훌륭하게 표현했습니다."
       },
       {
        "label": "B",
        "score": 2,
        "verdict_ko": "장소 고정 지시를 어기고 나무 문과 현대식 조명을 생성했으며, 두 인물의 의상이 동일하게 복제되어 프롬프트 충실도가 매우 낮습니다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "염재보의 시선이 감방 밖으로 나가는 안규홍의 뒷모습을 정확하게 향하고 있다.",
        "built_space": "이전 샷 레퍼런스와 동일하게 천장부터 바닥까지 이어지는 쇠창살, 하단이 칠해진 벽, 콘크리트 바닥이 구현되었으며, 쇠창살 문이 열려 있고 복도에는 핏자국이 유지되어 있다.",
        "entities": "염재보는 캐릭터 레퍼런스와 일치하는 외모와 누비옷을 입고 구석에 앉아 짚신 끈을 쥐고 있다. 안규홍은 흰옷을 입고 한쪽 발을 내디딘 뒷모습으로 등장한다.",
        "hard_violations": [],
        "physics": "염재보는 바닥에 안정적으로 앉아 체중을 지탱하고 있으며, 안규홍은 걷는 자세로 발이 바닥에 자연스럽게 닿아 있다."
       },
       {
        "label": "B",
        "direction": "염재보가 문 방향을 보고 있으나, 나가는 안규홍의 뒷모습을 뚫어져라 응시하기보다는 다소 시선이 아래로 향해 있다.",
        "built_space": "이전 샷 레퍼런스의 쇠창살 감방이 아닌 두꺼운 나무 문이 생성되었으며, 문밖 복도 천장에는 시대 배경에 맞지 않는 현대식 튜브형 형광등이 달려 있다.",
        "entities": "염재보가 구석에 앉아 짚신을 들고 있으나, 밖으로 나가는 안규홍이 염재보의 캐릭터 레퍼런스와 완전히 동일한 누비옷과 삿갓을 착용하고 있어 인물 간 의상이 중복되었다.",
        "hard_violations": [
         "발명된 사물(시대적 배경에 어긋나는 현대식 형광등 조명 생성)",
         "장소 고정(Location lock) 위반(쇠창살 대신 엉뚱한 나무 문과 열쇠구멍 생성)"
        ],
        "physics": "두 인물 모두 바닥에 지탱되어 서 있거나 앉아 있다."
       }
      ],
      "all_candidates_fail": false
     },
     "normalized": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 8,
        "verdict_ko": "레퍼런스에 맞는 쇠창살 감방과 복도의 핏자국을 정확히 구현했으며, 밖으로 나가는 인물을 응시하는 염재보의 시선과 자세를 훌륭하게 표현했습니다."
       },
       {
        "label": "B",
        "score": 2,
        "verdict_ko": "장소 고정 지시를 어기고 나무 문과 현대식 조명을 생성했으며, 두 인물의 의상이 동일하게 복제되어 프롬프트 충실도가 매우 낮습니다."
       }
      ],
      "all_candidates_fail": false,
      "readings": [
       {
        "label": "A",
        "direction": "염재보의 시선이 감방 밖으로 나가는 안규홍의 뒷모습을 정확하게 향하고 있다.",
        "built_space": "이전 샷 레퍼런스와 동일하게 천장부터 바닥까지 이어지는 쇠창살, 하단이 칠해진 벽, 콘크리트 바닥이 구현되었으며, 쇠창살 문이 열려 있고 복도에는 핏자국이 유지되어 있다.",
        "entities": "염재보는 캐릭터 레퍼런스와 일치하는 외모와 누비옷을 입고 구석에 앉아 짚신 끈을 쥐고 있다. 안규홍은 흰옷을 입고 한쪽 발을 내디딘 뒷모습으로 등장한다.",
        "hard_violations": [],
        "physics": "염재보는 바닥에 안정적으로 앉아 체중을 지탱하고 있으며, 안규홍은 걷는 자세로 발이 바닥에 자연스럽게 닿아 있다."
       },
       {
        "label": "B",
        "direction": "염재보가 문 방향을 보고 있으나, 나가는 안규홍의 뒷모습을 뚫어져라 응시하기보다는 다소 시선이 아래로 향해 있다.",
        "built_space": "이전 샷 레퍼런스의 쇠창살 감방이 아닌 두꺼운 나무 문이 생성되었으며, 문밖 복도 천장에는 시대 배경에 맞지 않는 현대식 튜브형 형광등이 달려 있다.",
        "entities": "염재보가 구석에 앉아 짚신을 들고 있으나, 밖으로 나가는 안규홍이 염재보의 캐릭터 레퍼런스와 완전히 동일한 누비옷과 삿갓을 착용하고 있어 인물 간 의상이 중복되었다.",
        "hard_violations": [
         "발명된 사물(시대적 배경에 어긋나는 현대식 형광등 조명 생성)",
         "장소 고정(Location lock) 위반(쇠창살 대신 엉뚱한 나무 문과 열쇠구멍 생성)"
        ],
        "physics": "두 인물 모두 바닥에 지탱되어 서 있거나 앉아 있다."
       }
      ]
     },
     "ok": true
    },
    {
     "model": "gpt-high",
     "order": "reverse",
     "display_to_canonical": {
      "A": "B",
      "B": "A"
     },
     "raw": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 3,
        "verdict_ko": "나가는 사람을 바라보는 관계는 보이지만, 쇠창살문을 목재문으로 바꾸고 현대식 조명을 추가했으며 안규홍도 이미 문턱을 벗어나 핵심 공간과 순간을 놓쳤습니다."
       },
       {
        "label": "B",
        "score": 8,
        "verdict_ko": "중앙 오른쪽의 열린 쇠창살문과 문턱을 넘는 발, 왼쪽에 앉아 응시하며 짚신 끈을 쥔 염재보를 가장 충실하게 구현했으나, 퇴장자의 복장은 이전 사진 인물의 복장과 겹칩니다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "왼쪽에 앉은 남자는 고개와 눈을 오른쪽 출입구의 퇴장자 쪽으로 향한다. 퇴장자는 카메라에 등을 보이고 복도 안쪽으로 걸어간다. 응시 대상은 대체로 맞지만, 퇴장자의 두 발이 이미 문턱 바깥에 있어 한쪽 발을 막 내디딘 순간보다 늦다.",
        "built_space": "왼쪽 감방 구석에 앉은 인물, 중앙 오른쪽 출입구 하나와 문턱 하나, 복도 바닥의 혈흔이 보인다. 열린 문짝은 대부분이 막힌 목재 판문이고 위쪽만 쇠창살이어서 참고 사진의 전면 쇠창살 구조와 다르다. 왼쪽 벽등 하나와 복도 천장의 밝은 관형 조명 하나가 보이며, 침구 더미와 용기들도 추가되어 있다. 반사는 없다.",
        "entities": "성인 동아시아계 남성 두 명이 보이며 추가 인체나 글자는 없다. 앉은 염재보는 검은 머리와 어두운 누빔옷을 갖췄지만, 길게 뒤로 묶은 머리와 얼굴은 인물 참고의 정확한 일치로 보기 어렵다. 손에는 해진 끈이 있고 앞에는 짚신 두 짝, 옆에는 삿갓이 놓여 있다. 오히려 떠나는 남자가 염재보 참고의 삿갓·어두운 옷·각반 조합을 그대로 갖춰 인물별 의상 배정이 혼동된다. 안규홍의 얼굴은 보이지 않는다.",
        "hard_violations": [
         "장소가 고정한 전면 쇠창살문을 대부분 불투명한 목재 판문으로 교체해 핵심 고정 구조를 바꿨다.",
         "복도에 밝은 현대식 관형 조명을 추가해 시대 설정과 새로운 조명 단서를 금지한 조건을 위반했다."
        ],
        "physics": "염재보는 엉덩이와 접은 다리를 바닥에 대고 앉아 있으며 양손으로 짚신 끈을 잡는다. 짚신과 삿갓은 바닥에 놓여 있다. 퇴장자는 오른발로 복도 바닥을 딛고 왼발을 들어 옮기는 보행 자세여서 지지가 성립한다. 문짝도 문틀에 연결되어 있고, 근거 없이 떠 있는 인체나 소품은 보이지 않는다."
       },
       {
        "label": "B",
        "direction": "왼쪽에 앉은 염재보의 얼굴과 눈은 오른쪽 출입구에 선 안규홍의 등 쪽을 향한다. 안규홍은 등을 카메라에 보인 채 어두운 복도 방향으로 이동한다. 문턱에 남은 뒤쪽 발과 바깥으로 옮긴 몸의 관계가 퇴장 순간을 보여준다.",
        "built_space": "감방 안쪽에서 본 와이드 구도이며, 왼쪽에는 앉은 염재보, 중앙 오른쪽에는 출입구 하나와 낮은 문턱 하나가 있다. 출입구 오른쪽의 쇠창살 문짝 하나가 감방 안으로 열려 있고 왼쪽은 고정 쇠창살 구획이다. 복도 양옆의 철창, 낡은 회벽과 어두운 하단 도장, 길게 이어지는 혈흔이 참고 장소와 잘 연결된다. 복도 벽등 하나 외에 천장등 하나가 보여 참고보다 조명이 늘어난 인상은 있다. 반사는 없다.",
        "entities": "성인 동아시아계 남성 두 명만 보인다. 염재보는 검은 머리, 어두운 누빔 상하의와 각반을 착용하고 짚신의 풀린 끈을 손에 쥐고 있다. 다만 얼굴과 묶은 머리는 인물 참고의 정확한 재현이라고 단정하기 어렵고 삿갓은 보이지 않는다. 안규홍은 얼굴이 가려진 뒷모습이지만 밝은색 낡은 옷과 묶은 머리가 이전 사진 속 인물의 복장·머리와 겹쳐, 이전 인물의 외형을 이식하지 말라는 조건에는 약점이 있다. 읽을 수 있는 글자나 표시물은 없다.",
        "hard_violations": [],
        "physics": "염재보는 엉덩이를 바닥에 두고 등을 벽에 기대며, 세운 무릎 가까이에서 양손으로 끈을 잡아 멈춘 작업 자세를 유지한다. 안규홍은 몸을 문밖으로 옮기면서 뒤쪽 발의 뒤꿈치를 들고 발끝을 문턱 부근에 둔 자세다. 앞쪽 발의 접지는 바지와 뒤쪽 다리에 일부 가려지지만 보행을 설명할 지지 관계가 있으며, 공중에 매달린 자세로 보이지 않는다. 열린 문짝은 오른쪽 문틀에 연결되어 있다."
       }
      ],
      "all_candidates_fail": false
     },
     "normalized": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 3,
        "verdict_ko": "나가는 사람을 바라보는 관계는 보이지만, 쇠창살문을 목재문으로 바꾸고 현대식 조명을 추가했으며 안규홍도 이미 문턱을 벗어나 핵심 공간과 순간을 놓쳤습니다."
       },
       {
        "label": "A",
        "score": 8,
        "verdict_ko": "중앙 오른쪽의 열린 쇠창살문과 문턱을 넘는 발, 왼쪽에 앉아 응시하며 짚신 끈을 쥔 염재보를 가장 충실하게 구현했으나, 퇴장자의 복장은 이전 사진 인물의 복장과 겹칩니다."
       }
      ],
      "all_candidates_fail": false,
      "readings": [
       {
        "label": "B",
        "direction": "왼쪽에 앉은 남자는 고개와 눈을 오른쪽 출입구의 퇴장자 쪽으로 향한다. 퇴장자는 카메라에 등을 보이고 복도 안쪽으로 걸어간다. 응시 대상은 대체로 맞지만, 퇴장자의 두 발이 이미 문턱 바깥에 있어 한쪽 발을 막 내디딘 순간보다 늦다.",
        "built_space": "왼쪽 감방 구석에 앉은 인물, 중앙 오른쪽 출입구 하나와 문턱 하나, 복도 바닥의 혈흔이 보인다. 열린 문짝은 대부분이 막힌 목재 판문이고 위쪽만 쇠창살이어서 참고 사진의 전면 쇠창살 구조와 다르다. 왼쪽 벽등 하나와 복도 천장의 밝은 관형 조명 하나가 보이며, 침구 더미와 용기들도 추가되어 있다. 반사는 없다.",
        "entities": "성인 동아시아계 남성 두 명이 보이며 추가 인체나 글자는 없다. 앉은 염재보는 검은 머리와 어두운 누빔옷을 갖췄지만, 길게 뒤로 묶은 머리와 얼굴은 인물 참고의 정확한 일치로 보기 어렵다. 손에는 해진 끈이 있고 앞에는 짚신 두 짝, 옆에는 삿갓이 놓여 있다. 오히려 떠나는 남자가 염재보 참고의 삿갓·어두운 옷·각반 조합을 그대로 갖춰 인물별 의상 배정이 혼동된다. 안규홍의 얼굴은 보이지 않는다.",
        "hard_violations": [
         "장소가 고정한 전면 쇠창살문을 대부분 불투명한 목재 판문으로 교체해 핵심 고정 구조를 바꿨다.",
         "복도에 밝은 현대식 관형 조명을 추가해 시대 설정과 새로운 조명 단서를 금지한 조건을 위반했다."
        ],
        "physics": "염재보는 엉덩이와 접은 다리를 바닥에 대고 앉아 있으며 양손으로 짚신 끈을 잡는다. 짚신과 삿갓은 바닥에 놓여 있다. 퇴장자는 오른발로 복도 바닥을 딛고 왼발을 들어 옮기는 보행 자세여서 지지가 성립한다. 문짝도 문틀에 연결되어 있고, 근거 없이 떠 있는 인체나 소품은 보이지 않는다."
       },
       {
        "label": "A",
        "direction": "왼쪽에 앉은 염재보의 얼굴과 눈은 오른쪽 출입구에 선 안규홍의 등 쪽을 향한다. 안규홍은 등을 카메라에 보인 채 어두운 복도 방향으로 이동한다. 문턱에 남은 뒤쪽 발과 바깥으로 옮긴 몸의 관계가 퇴장 순간을 보여준다.",
        "built_space": "감방 안쪽에서 본 와이드 구도이며, 왼쪽에는 앉은 염재보, 중앙 오른쪽에는 출입구 하나와 낮은 문턱 하나가 있다. 출입구 오른쪽의 쇠창살 문짝 하나가 감방 안으로 열려 있고 왼쪽은 고정 쇠창살 구획이다. 복도 양옆의 철창, 낡은 회벽과 어두운 하단 도장, 길게 이어지는 혈흔이 참고 장소와 잘 연결된다. 복도 벽등 하나 외에 천장등 하나가 보여 참고보다 조명이 늘어난 인상은 있다. 반사는 없다.",
        "entities": "성인 동아시아계 남성 두 명만 보인다. 염재보는 검은 머리, 어두운 누빔 상하의와 각반을 착용하고 짚신의 풀린 끈을 손에 쥐고 있다. 다만 얼굴과 묶은 머리는 인물 참고의 정확한 재현이라고 단정하기 어렵고 삿갓은 보이지 않는다. 안규홍은 얼굴이 가려진 뒷모습이지만 밝은색 낡은 옷과 묶은 머리가 이전 사진 속 인물의 복장·머리와 겹쳐, 이전 인물의 외형을 이식하지 말라는 조건에는 약점이 있다. 읽을 수 있는 글자나 표시물은 없다.",
        "hard_violations": [],
        "physics": "염재보는 엉덩이를 바닥에 두고 등을 벽에 기대며, 세운 무릎 가까이에서 양손으로 끈을 잡아 멈춘 작업 자세를 유지한다. 안규홍은 몸을 문밖으로 옮기면서 뒤쪽 발의 뒤꿈치를 들고 발끝을 문턱 부근에 둔 자세다. 앞쪽 발의 접지는 바지와 뒤쪽 다리에 일부 가려지지만 보행을 설명할 지지 관계가 있으며, 공중에 매달린 자세로 보이지 않는다. 열린 문짝은 오른쪽 문틀에 연결되어 있다."
       }
      ]
     },
     "ok": true
    }
   ],
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "slot_winner_match": true,
   "slot_winner": {
    "gemini-pro": "A",
    "gpt-high": "A"
   },
   "route": "cross_slot_agree"
  },
  "dual": {
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "normalized": {
    "A": 2.0,
    "B": 0.625
   },
   "adjusted": {
    "A": 2.0,
    "B": 0.375
   },
   "violations": {
    "B": [
     "[gemini-pro] 발명된 사물(시대적 배경에 어긋나는 현대식 형광등 조명 생성)",
     "[gemini-pro] 장소 고정(Location lock) 위반(쇠창살 대신 엉뚱한 나무 문과 열쇠구멍 생성)",
     "[gpt-high] 장소가 고정한 전면 쇠창살문을 대부분 불투명한 목재 판문으로 교체해 핵심 고정 구조를 바꿨다.",
     "[gpt-high] 복도에 밝은 현대식 관형 조명을 추가해 시대 설정과 새로운 조명 단서를 금지한 조건을 위반했다."
    ]
   },
   "per_model_winner": {
    "gemini-pro": "A",
    "gpt-high": "A"
   },
   "agreed": true
  },
  "totals": {
   "A": 2000,
   "B": 375
  },
  "selected": "A",
  "ranking": [
   "A",
   "B"
  ],
  "verdicts": [
   {
    "label": "A",
    "score": 2000,
    "verdict_ko": "레퍼런스에 맞는 쇠창살 감방과 복도의 핏자국을 정확히 구현했으며, 밖으로 나가는 인물을 응시하는 염재보의 시선과 자세를 훌륭하게 표현했습니다."
   },
   {
    "label": "B",
    "score": 375,
    "verdict_ko": "장소 고정 지시를 어기고 나무 문과 현대식 조명을 생성했으며, 두 인물의 의상이 동일하게 복제되어 프롬프트 충실도가 매우 낮습니다.  ★위반: [gemini-pro] 발명된 사물(시대적 배경에 어긋나는 현대식 형광등 조명 생성) / [gemini-pro] 장소 고정(Location lock) 위반(쇠창살 대신 엉뚱한 나무 문과 열쇠구멍 생성) / [gpt-high] 장소가 고정한 전면 쇠창살문을 대부분 불투명한 목재 판문으로 교체해 핵심 고정 구조를 바꿨다. / [gpt-high] 복도에 밝은 현대식 관형 조명을 추가해 시대 설정과 새로운 조명 단서를 금지한 조건을 위반했다."
   }
  ],
  "refs": [
   {
    "label": "PREVIOUS SHOT STILL — a visually related earlier shot of this same place: the location's look, materials, fixed features and lighting mood are LOCKED to this photo; never copy its camera framing. The people visible in that photo are NOT in this shot: never carry their faces, bodies or clothing onto anyone here. Each person in THIS shot is defined solely by the PEOPLE list, their own CHARACTER REFERENCE images and their wardrobe notes. If this photo shows a character who cannot move (dead or unconscious), that character's exact pose, position and orientation are ALSO LOCKED — treat their body as a fixed prop of the set that only the camera moves around.",
    "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/db00517b-165b-4005-a287-f28936b59b95/images/272bd7af-9294-4490-98c6-88a40e9deacc/scene/recipe/S9sh11_sel.png",
    "asset_id": "483704fe-d61c-4b00-b8f5-1d18a85dcc7d",
    "role": "prev_still"
   },
   {
    "label": "CHARACTER REFERENCE — 염재보: the exact person appearing in this shot; match face, hair and build exactly.",
    "path": "<bytes:964351>",
    "asset_id": "ab287289-d0b7-436d-8f70-68452f898f73",
    "role": "character_ref"
   }
  ],
  "critique_skipped": true,
  "shot_run_produced": false,
  "shot_run_uid": "06aa140c-4a99-7ab6-80e8-496715b30852",
  "ref_mode": "prev+엔티티",
  "share_plan": {
   "ref_plan": "prev",
   "prev_anchor_tag": "S9sh11"
  }
 },
 "S9sh15::cine": {
  "applied": true,
  "attempted_at": "2026-09-09T01:37:27.798353+00:00",
  "fingerprint": "7a8b25e73b7edbcbe00ba41d4970802be862bcfc9455828f51e638b9fe086f45",
  "fingerprint_version": 2,
  "provider": "mai",
  "endpoint": "openrouter/chat-completions",
  "model": "microsoft/mai-image-2.6",
  "pack": "24.202608252115",
  "source_file": "S9sh15_sel.png",
  "source_sha256": "83ed75afd3fc856f819a3c4d52100a3fe7dc810d4969bc95a9ad089989c6545e",
  "file": "S9sh15_cine.png",
  "staged_sha256": "484a89a1fb5c6d5c80bdb635dbec9063cc188923c28aeed11f44a53e194f4a5c",
  "latency_ms": 27630
 },
 "S10sh6::signage": {
  "fp": "54b6784c7bbba1f2",
  "inscriptions": [],
  "cues": [],
  "dropped": []
 },
 "S10sh6::bgfirst_bg": {
  "input_fingerprint": "68934950b3825432",
  "prompt": "Create the EMPTY BACKGROUND PLATE for one film shot — NO PEOPLE, no figures, no body parts, no sketch lines, no arrows anywhere.\n\"Empty\" means no people only: KEEP the location's inherent occupants and stock that define the place — animals in an animal shelter, pen or farm, goods in a market, moored boats in a harbour — unless the shot text explicitly removes them.\nThe FIRST attached image is a thin-line storyboard sketch: use ONLY its camera angle, horizon, perspective and the placement/size of buildings and set masses — ignore the sketched people and arrows entirely. The SECOND attached image (LOCATION PHOTOGRAPH) is the real place: take its architecture, materials, signage and fixed features, and RE-PROJECT them into the sketch's camera. If the photograph's camera differs from the sketch's, the sketch's camera wins.\nHUMAN-SCALE CALIBRATION: derive every structure's true size from human-scale elements — a door ≈ 2m, a window ≈ 1–1.5m wide, one storey ≈ 2.5–3m; never inflate a small structure or shrink a large one.\n\nSHOT TEXT this background must serve (Korean): 안규홍을 향해 고개를 돌린 채 원망스럽게 노려보는 조선인 청년의 붉어진 두 눈.\n\nLOCATION (lock): The damaged outdoor frontage of a straw-sandal stall in the daytime market, beside a ruined woven mat.\n\nTIME OF DAY (lock): day.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: close-up\n- KEY BACKGROUND ELEMENTS: 망가진 짚신전 (Damaged after the market violence) — Only an oblique fragment of the stall frontage remains behind the young man; used as Soft background context connects his accusation to the damage without distracting from his eyes.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Natural daytime illumination with muted saturation preserves the reddened eyes as a restrained emotional accent rather than a stylized color effect.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nRender ONE photorealistic empty location photograph, 16:9, that this shot can be staged inside later. No readable writing anywhere: surfaces that would carry writing may be present, but stage any wording out of legibility — an oblique angle, distance, shallow focus. No captions, watermarks or overlay text.",
  "effective_prompt": "Create the EMPTY BACKGROUND PLATE for one film shot — NO PEOPLE, no figures, no body parts, no sketch lines, no arrows anywhere.\n\"Empty\" means no people only: KEEP the location's inherent occupants and stock that define the place — animals in an animal shelter, pen or farm, goods in a market, moored boats in a harbour — unless the shot text explicitly removes them.\nThe FIRST attached image is a thin-line storyboard sketch: use ONLY its camera angle, horizon, perspective and the placement/size of buildings and set masses — ignore the sketched people and arrows entirely. The SECOND attached image (LOCATION PHOTOGRAPH) is the real place: take its architecture, materials, signage and fixed features, and RE-PROJECT them into the sketch's camera. If the photograph's camera differs from the sketch's, the sketch's camera wins.\nHUMAN-SCALE CALIBRATION: derive every structure's true size from human-scale elements — a door ≈ 2m, a window ≈ 1–1.5m wide, one storey ≈ 2.5–3m; never inflate a small structure or shrink a large one.\n\nSHOT TEXT this background must serve (Korean): 안규홍을 향해 고개를 돌린 채 원망스럽게 노려보는 조선인 청년의 붉어진 두 눈.\n\nLOCATION (lock): The damaged outdoor frontage of a straw-sandal stall in the daytime market, beside a ruined woven mat.\n\nTIME OF DAY (lock): day.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: close-up\n- KEY BACKGROUND ELEMENTS: 망가진 짚신전 (Damaged after the market violence) — Only an oblique fragment of the stall frontage remains behind the young man; used as Soft background context connects his accusation to the damage without distracting from his eyes.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Natural daytime illumination with muted saturation preserves the reddened eyes as a restrained emotional accent rather than a stylized color effect.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nRender ONE photorealistic empty location photograph, 16:9, that this shot can be staged inside later. No readable writing anywhere: surfaces that would carry writing may be present, but stage any wording out of legibility — an oblique angle, distance, shallow focus. No captions, watermarks or overlay text.",
  "bg_path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/db00517b-165b-4005-a287-f28936b59b95/images/272bd7af-9294-4490-98c6-88a40e9deacc/scene/recipe/S10sh6__bgfirst_bg.png",
  "asset_id": "525b41ef-ff22-4cae-9ff6-49fa3d7c6a24",
  "input_asset_ids": [
   "a9de656a-ac20-4988-b9cb-ecf760c07b21",
   "a82710e0-4099-4422-b0e7-8cd69a1606bb"
  ]
 },
 "S10sh6": {
  "input_fingerprint": "6fdae24e4ff94f21",
  "prompt": "Create ONE FINAL photorealistic live-action film still of the moment below — early 20th-century Korea, Jeolla region; all people are Korean unless explicitly stated otherwise. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 안규홍을 향해 고개를 돌린 채 원망스럽게 노려보는 조선인 청년의 붉어진 두 눈.\n\nLOCATION (lock): The damaged outdoor frontage of a straw-sandal stall in the daytime market, beside a ruined woven mat. The shot takes place here — the attached LOCATION PHOTOGRAPH shows the exact spot.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: close-up\n- KEY BACKGROUND ELEMENTS: 망가진 짚신전 (Damaged after the market violence) — Only an oblique fragment of the stall frontage remains behind the young man; used as Soft background context connects his accusation to the damage without distracting from his eyes.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Natural daytime illumination with muted saturation preserves the reddened eyes as a restrained emotional accent rather than a stylized color effect.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): Overturned stalls, torn mats and broken bowls litter the dirt; a trampled single straw shoe lies before the damaged shoe stall. 조선인 청년(짚신전): He stands at the damaged shoe stall, handling its ruined mat; his eyes are reddening.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 조선인 청년(짚신전) (한국인 남성, 청년 나이대, 자연스러운 검은 머리) — wearing: 젊은 청년의 체격에 맞는 수수한 무명 바지와 저고리. 짚신을 팔기 위해 움직임이 편한 복장. — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.",
  "roll_prompts": {
   "A": "Stage the shot. The FIRST attached image (SHOT BACKGROUND) is the finished empty background of this shot — keep it EXACTLY: its camera, perspective, architecture, lighting and every fixed feature stay untouched. The SECOND attached image (LAYOUT SKETCH) tells you ONLY where the people go: each sketched person's position, screen size, pose and the gaze/motion arrows. Ignore the sketch's background lines. The CHARACTER REFERENCE photographs show the real people.\nPlace the real people into the background at exactly the sketched positions, sizes and poses, following the arrow directions. No sketch lines or arrows may remain.\n\nCreate ONE FINAL photorealistic live-action film still of the moment below — early 20th-century Korea, Jeolla region; all people are Korean unless explicitly stated otherwise. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 안규홍을 향해 고개를 돌린 채 원망스럽게 노려보는 조선인 청년의 붉어진 두 눈.\n\nLOCATION (lock): The damaged outdoor frontage of a straw-sandal stall in the daytime market, beside a ruined woven mat. The shot takes place here — the FIRST attached image (SHOT BACKGROUND) is this exact place, already built: its ground, structures, horizon, materials and lighting are the finished truth of this location and must not be redesigned or replaced. No location photograph is attached — read the place from that image alone, and add no scenery, structure, vehicle or fixture that it does not already show. This lock governs the place only; the figures in the shot follow the staging and pose instructions.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: close-up\n- KEY BACKGROUND ELEMENTS: 망가진 짚신전 (Damaged after the market violence) — Only an oblique fragment of the stall frontage remains behind the young man; used as Soft background context connects his accusation to the damage without distracting from his eyes.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Natural daytime illumination with muted saturation preserves the reddened eyes as a restrained emotional accent rather than a stylized color effect.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): Overturned stalls, torn mats and broken bowls litter the dirt; a trampled single straw shoe lies before the damaged shoe stall. 조선인 청년(짚신전): He stands at the damaged shoe stall, handling its ruined mat; his eyes are reddening.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 조선인 청년(짚신전) (한국인 남성, 청년 나이대, 자연스러운 검은 머리) — wearing: 젊은 청년의 체격에 맞는 수수한 무명 바지와 저고리. 짚신을 팔기 위해 움직임이 편한 복장. — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.",
   "B": "Create ONE FINAL photorealistic live-action film still of the moment below — early 20th-century Korea, Jeolla region; all people are Korean unless explicitly stated otherwise. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 안규홍을 향해 고개를 돌린 채 원망스럽게 노려보는 조선인 청년의 붉어진 두 눈.\n\nLOCATION (lock): The damaged outdoor frontage of a straw-sandal stall in the daytime market, beside a ruined woven mat. The shot takes place here — the attached LOCATION PHOTOGRAPH shows the exact spot.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: close-up\n- KEY BACKGROUND ELEMENTS: 망가진 짚신전 (Damaged after the market violence) — Only an oblique fragment of the stall frontage remains behind the young man; used as Soft background context connects his accusation to the damage without distracting from his eyes.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Natural daytime illumination with muted saturation preserves the reddened eyes as a restrained emotional accent rather than a stylized color effect.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): Overturned stalls, torn mats and broken bowls litter the dirt; a trampled single straw shoe lies before the damaged shoe stall. 조선인 청년(짚신전): He stands at the damaged shoe stall, handling its ruined mat; his eyes are reddening.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 조선인 청년(짚신전) (한국인 남성, 청년 나이대, 자연스러운 검은 머리) — wearing: 젊은 청년의 체격에 맞는 수수한 무명 바지와 저고리. 짚신을 팔기 위해 움직임이 편한 복장. — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.\n\nCOMPOSITION VARIATION (this candidate only): keep the place, the people and the moment exactly as specified, and keep from the CAMERA & FRAME contract its subject scale, screen placement and key background elements exactly. For this candidate only, its camera angle and height are deliberately left open: choose a camera position distinctly different from the obvious one, as a film director picking a second setup on the same blocking. Stay true to any explicit framing the SHOT TEXT itself dictates."
  },
  "roll_refs": {
   "A": [
    {
     "label": "SHOT BACKGROUND",
     "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/db00517b-165b-4005-a287-f28936b59b95/images/272bd7af-9294-4490-98c6-88a40e9deacc/scene/recipe/S10sh6__bgfirst_bg.png",
     "asset_id": "525b41ef-ff22-4cae-9ff6-49fa3d7c6a24",
     "role": "bgfirst_bg"
    },
    {
     "label": "LAYOUT SKETCH (people placement only)",
     "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/db00517b-165b-4005-a287-f28936b59b95/images/272bd7af-9294-4490-98c6-88a40e9deacc/conti/conti_S10sh6.png",
     "asset_id": "a9de656a-ac20-4988-b9cb-ecf760c07b21",
     "role": "conti_light"
    }
   ],
   "B": [
    {
     "label": "LOCATION PHOTOGRAPH — the exact place of this shot: its architecture, materials, fixed features and lighting mood are spatial truth; stage the moment inside this place. Never copy its camera framing.",
     "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/db00517b-165b-4005-a287-f28936b59b95/images/272bd7af-9294-4490-98c6-88a40e9deacc/scene/recipe/groupbg_market_shoe_front_6546f9.png",
     "asset_id": "a82710e0-4099-4422-b0e7-8cd69a1606bb",
     "role": "bgfirst_group_bg"
    }
   ]
  },
  "shot_run_spend_attempt_count": 1,
  "judge_flip_skipped": "judge_fn_owns_order",
  "readings": [
   {
    "label": "A",
    "direction": "청년은 왼쪽 어깨 너머로 고개를 돌려 전경에 등지고 있는 인물을 향해 시선을 던지고 있습니다.",
    "built_space": "배경에 목조 구조물이 보이나 아웃포커싱되어 있으며, 레퍼런스의 짚신전 형태가 온전히 드러나지는 않습니다.",
    "entities": "청년은 붉어진 눈과 무명 옷을 입은 모습으로 잘 표현되었으나, 프롬프트에서 등장을 엄격히 금지한 추가 인물(전경의 인물)이 존재합니다.",
    "hard_violations": [
     "[gemini-pro] 허용되지 않은 추가 인물 등장 (프롬프트는 오직 조선인 청년 1명만 화면에 등장하도록 제한함)",
     "[gpt-high] 화면에 허용되지 않은 두 번째 남성의 뒤통수와 어깨를 전경에 추가했다."
    ],
    "physics": "청년은 서서 양손으로 부서진 멍석을 쥐고 있으며, 자세와 쥐고 있는 물체의 상태 모두 물리적으로 자연스럽습니다."
   },
   {
    "label": "B",
    "direction": "청년은 화면 밖 왼쪽을 향해 원망스러운 시선을 보내고 있습니다.",
    "built_space": "레퍼런스 사진과 동일한 형태의 짚신전이 화면 왼쪽에 위치하며, 배경의 시장 거리와 산맥의 배치도 정확히 일치합니다.",
    "entities": "청년의 눈이 붉게 충혈되어 있고 수수한 무명 옷을 입고 있으며, 명시된 대로 다른 인물은 등장하지 않습니다.",
    "hard_violations": [
     "[gpt-high] 서 있어야 한다는 지속 상태와 달리, 청년은 무릎을 가슴 가까이 올리고 낮게 웅크린 자세로 배치되어 있다."
    ],
    "physics": "청년은 바닥에 쪼그려 앉아 두 손을 다리 위에 얹고 있으며, 멍석과 짚신들은 바닥에 안정적으로 놓여 있습니다."
   }
  ],
  "cross_model_order": {
   "policy": "cross_model_order_v2_slot0_fwd_slot1_rev_parallel2",
   "slots": [
    {
     "model": "gemini-pro",
     "order": "forward",
     "display_to_canonical": {
      "A": "A",
      "B": "B"
     },
     "raw": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 5,
        "verdict_ko": "하드 비올레이션은 피했으나, 클로즈업 지시를 무시하고 배경을 선명하게 다 드러냈으며 서서 멍석을 만지는 대신 쪼그려 앉아 있어 프레이밍과 포즈 지시(우선순위 2, 5)를 크게 위반했습니다."
       },
       {
        "label": "A",
        "score": 2,
        "verdict_ko": "포즈와 행동은 프롬프트에 부합하나, 화면에 오직 청년 1명만 허용한다는 지시를 어기고 전경에 다른 인물을 등장시켜 최우선 기준인 하드 비올레이션을 범했습니다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "청년은 왼쪽 어깨 너머로 고개를 돌려 전경에 등지고 있는 인물을 향해 시선을 던지고 있습니다.",
        "built_space": "배경에 목조 구조물이 보이나 아웃포커싱되어 있으며, 레퍼런스의 짚신전 형태가 온전히 드러나지는 않습니다.",
        "entities": "청년은 붉어진 눈과 무명 옷을 입은 모습으로 잘 표현되었으나, 프롬프트에서 등장을 엄격히 금지한 추가 인물(전경의 인물)이 존재합니다.",
        "hard_violations": [
         "허용되지 않은 추가 인물 등장 (프롬프트는 오직 조선인 청년 1명만 화면에 등장하도록 제한함)"
        ],
        "physics": "청년은 서서 양손으로 부서진 멍석을 쥐고 있으며, 자세와 쥐고 있는 물체의 상태 모두 물리적으로 자연스럽습니다."
       },
       {
        "label": "B",
        "direction": "청년은 화면 밖 왼쪽을 향해 원망스러운 시선을 보내고 있습니다.",
        "built_space": "레퍼런스 사진과 동일한 형태의 짚신전이 화면 왼쪽에 위치하며, 배경의 시장 거리와 산맥의 배치도 정확히 일치합니다.",
        "entities": "청년의 눈이 붉게 충혈되어 있고 수수한 무명 옷을 입고 있으며, 명시된 대로 다른 인물은 등장하지 않습니다.",
        "hard_violations": [],
        "physics": "청년은 바닥에 쪼그려 앉아 두 손을 다리 위에 얹고 있으며, 멍석과 짚신들은 바닥에 안정적으로 놓여 있습니다."
       }
      ],
      "all_candidates_fail": false
     },
     "normalized": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 5,
        "verdict_ko": "하드 비올레이션은 피했으나, 클로즈업 지시를 무시하고 배경을 선명하게 다 드러냈으며 서서 멍석을 만지는 대신 쪼그려 앉아 있어 프레이밍과 포즈 지시(우선순위 2, 5)를 크게 위반했습니다."
       },
       {
        "label": "A",
        "score": 2,
        "verdict_ko": "포즈와 행동은 프롬프트에 부합하나, 화면에 오직 청년 1명만 허용한다는 지시를 어기고 전경에 다른 인물을 등장시켜 최우선 기준인 하드 비올레이션을 범했습니다."
       }
      ],
      "all_candidates_fail": false,
      "readings": [
       {
        "label": "A",
        "direction": "청년은 왼쪽 어깨 너머로 고개를 돌려 전경에 등지고 있는 인물을 향해 시선을 던지고 있습니다.",
        "built_space": "배경에 목조 구조물이 보이나 아웃포커싱되어 있으며, 레퍼런스의 짚신전 형태가 온전히 드러나지는 않습니다.",
        "entities": "청년은 붉어진 눈과 무명 옷을 입은 모습으로 잘 표현되었으나, 프롬프트에서 등장을 엄격히 금지한 추가 인물(전경의 인물)이 존재합니다.",
        "hard_violations": [
         "허용되지 않은 추가 인물 등장 (프롬프트는 오직 조선인 청년 1명만 화면에 등장하도록 제한함)"
        ],
        "physics": "청년은 서서 양손으로 부서진 멍석을 쥐고 있으며, 자세와 쥐고 있는 물체의 상태 모두 물리적으로 자연스럽습니다."
       },
       {
        "label": "B",
        "direction": "청년은 화면 밖 왼쪽을 향해 원망스러운 시선을 보내고 있습니다.",
        "built_space": "레퍼런스 사진과 동일한 형태의 짚신전이 화면 왼쪽에 위치하며, 배경의 시장 거리와 산맥의 배치도 정확히 일치합니다.",
        "entities": "청년의 눈이 붉게 충혈되어 있고 수수한 무명 옷을 입고 있으며, 명시된 대로 다른 인물은 등장하지 않습니다.",
        "hard_violations": [],
        "physics": "청년은 바닥에 쪼그려 앉아 두 손을 다리 위에 얹고 있으며, 멍석과 짚신들은 바닥에 안정적으로 놓여 있습니다."
       }
      ]
     },
     "ok": true
    },
    {
     "model": "gpt-high",
     "order": "reverse",
     "display_to_canonical": {
      "A": "B",
      "B": "A"
     },
     "raw": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 3,
        "verdict_ko": "청년 한 명의 붉어진 두 눈과 화면 밖을 향한 원망은 더 잘 드러나지만, 클로즈업보다 넓고 서 있어야 할 청년이 웅크려 있다."
       },
       {
        "label": "B",
        "score": 2,
        "verdict_ko": "망가진 멍석을 잡고 고개를 돌리는 동작은 맞지만, 금지된 전경 인물을 추가하고 눈의 클로즈업을 넓은 어깨너머 구도로 바꿨다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "청년의 얼굴과 두 눈은 화면 왼쪽 바깥을 향한다. 렌즈를 직접 응시하지 않으며 원망과 고통이 섞인 표정이다. 시선 끝의 사람은 보이지 않아 안규홍인지는 확인할 수 없지만, 화면 밖 안규홍을 보는 설정과 양립한다.",
        "built_space": "왼쪽에 짚신이 걸린 가로 진열대 세 단, 그 옆에 세로 직조 발 하나, 아래 목재 판매대가 보인다. 찢어진 멍석은 판매대 앞 흙바닥에 놓여 있다. 참고 장소의 목재·짚 진열 구조는 상당히 유지되지만, 점포 전면과 시장 길이 넓고 선명하게 노출되어 요구된 흐릿한 비스듬한 배경 조각에 그치지 않는다. 청년은 오른쪽 목재 기둥 앞에 몸을 낮추고 있다.",
        "entities": "자연스러운 검은 머리의 한국인 청년 남성으로 보이는 인물 한 명이다. 수수하고 때 묻은 밝은 무명 저고리와 바지가 보이며, 두 눈에는 정상적인 홍채와 동공, 붉어진 눈가가 있다. 짚신전, 찢어진 멍석, 흩어진 짚신과 나무 잔해가 보인다. 앞쪽에는 짚신 한 짝이 따로 놓여 있으나 짓밟힌 상태는 뚜렷하지 않다. 읽을 수 있는 글자나 추가 인물은 없다.",
        "hard_violations": [
         "서 있어야 한다는 지속 상태와 달리, 청년은 무릎을 가슴 가까이 올리고 낮게 웅크린 자세로 배치되어 있다."
        ],
        "physics": "상체 아래로 크게 굽힌 무릎과 하체가 이어져 웅크린 자세로 읽힌다. 발과 최종 지지점은 프레임 밖이지만 몸이 공중에 떠 있다는 징후는 없다. 멍석과 짚신은 흙바닥 또는 판매대에 놓이고, 진열된 짚신은 가로대에 걸려 있다. 손은 보이지 않으므로 멍석을 잡는 상태는 확인할 수 없다."
       },
       {
        "label": "B",
        "direction": "청년은 몸을 오른쪽 점포로 향한 채 고개를 왼쪽 뒤로 돌려 전경 인물 쪽을 노려본다. 붉어진 두 눈과 찌푸린 눈썹이 원망을 전달한다. 시선의 상대는 왼쪽 전경의 남성으로 읽히지만, 그가 안규홍임을 확인할 식별 정보는 없다.",
        "built_space": "오른쪽에 목재 기둥과 처마, 짚신 걸이 일부, 낮은 판매대 하나와 크게 찢어진 직조 발이 보인다. 왼쪽 뒤로 시장 길이 길게 드러난다. 청년은 점포 바로 앞에 서서 직조물의 테두리를 잡고 있어 위치와 사용 관계는 자연스럽다. 다만 점포와 멍석이 화면 대부분을 차지하고 청년의 허리 아래까지 보여, 작은 배경 조각만 남기는 클로즈업과 다르다.",
        "entities": "중앙 청년은 검은 머리의 한국인 젊은 남성으로 보이며, 수수한 회갈색 무명 저고리와 바지를 입고 있다. 붉어진 눈가 안에 정상적인 홍채와 동공이 보인다. 망가진 직조물과 짚신 진열, 바닥에 넘어진 용기가 있다. 왼쪽 전경에는 별도의 남성 뒤통수와 어깨가 크게 추가되어 허용된 청년 한 명이라는 인물 제한을 어긴다. 읽을 수 있는 글자는 없다.",
        "hard_violations": [
         "화면에 허용되지 않은 두 번째 남성의 뒤통수와 어깨를 전경에 추가했다."
        ],
        "physics": "청년의 몸통은 서 있는 하체로 자연스럽게 이어지며 발은 프레임 밖이다. 양손이 망가진 직조물의 단단한 테두리를 각각 잡고 있고, 직조물 윗부분은 점포 구조에 기대거나 걸쳐져 있어 지지가 보인다. 바닥의 용기는 흙바닥에, 짚신과 바구니는 판매대에 놓여 있다. 지지 없이 떠 있는 물체나 불가능한 관절 자세는 보이지 않는다."
       }
      ],
      "all_candidates_fail": true
     },
     "normalized": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 3,
        "verdict_ko": "청년 한 명의 붉어진 두 눈과 화면 밖을 향한 원망은 더 잘 드러나지만, 클로즈업보다 넓고 서 있어야 할 청년이 웅크려 있다."
       },
       {
        "label": "A",
        "score": 2,
        "verdict_ko": "망가진 멍석을 잡고 고개를 돌리는 동작은 맞지만, 금지된 전경 인물을 추가하고 눈의 클로즈업을 넓은 어깨너머 구도로 바꿨다."
       }
      ],
      "all_candidates_fail": true,
      "readings": [
       {
        "label": "B",
        "direction": "청년의 얼굴과 두 눈은 화면 왼쪽 바깥을 향한다. 렌즈를 직접 응시하지 않으며 원망과 고통이 섞인 표정이다. 시선 끝의 사람은 보이지 않아 안규홍인지는 확인할 수 없지만, 화면 밖 안규홍을 보는 설정과 양립한다.",
        "built_space": "왼쪽에 짚신이 걸린 가로 진열대 세 단, 그 옆에 세로 직조 발 하나, 아래 목재 판매대가 보인다. 찢어진 멍석은 판매대 앞 흙바닥에 놓여 있다. 참고 장소의 목재·짚 진열 구조는 상당히 유지되지만, 점포 전면과 시장 길이 넓고 선명하게 노출되어 요구된 흐릿한 비스듬한 배경 조각에 그치지 않는다. 청년은 오른쪽 목재 기둥 앞에 몸을 낮추고 있다.",
        "entities": "자연스러운 검은 머리의 한국인 청년 남성으로 보이는 인물 한 명이다. 수수하고 때 묻은 밝은 무명 저고리와 바지가 보이며, 두 눈에는 정상적인 홍채와 동공, 붉어진 눈가가 있다. 짚신전, 찢어진 멍석, 흩어진 짚신과 나무 잔해가 보인다. 앞쪽에는 짚신 한 짝이 따로 놓여 있으나 짓밟힌 상태는 뚜렷하지 않다. 읽을 수 있는 글자나 추가 인물은 없다.",
        "hard_violations": [
         "서 있어야 한다는 지속 상태와 달리, 청년은 무릎을 가슴 가까이 올리고 낮게 웅크린 자세로 배치되어 있다."
        ],
        "physics": "상체 아래로 크게 굽힌 무릎과 하체가 이어져 웅크린 자세로 읽힌다. 발과 최종 지지점은 프레임 밖이지만 몸이 공중에 떠 있다는 징후는 없다. 멍석과 짚신은 흙바닥 또는 판매대에 놓이고, 진열된 짚신은 가로대에 걸려 있다. 손은 보이지 않으므로 멍석을 잡는 상태는 확인할 수 없다."
       },
       {
        "label": "A",
        "direction": "청년은 몸을 오른쪽 점포로 향한 채 고개를 왼쪽 뒤로 돌려 전경 인물 쪽을 노려본다. 붉어진 두 눈과 찌푸린 눈썹이 원망을 전달한다. 시선의 상대는 왼쪽 전경의 남성으로 읽히지만, 그가 안규홍임을 확인할 식별 정보는 없다.",
        "built_space": "오른쪽에 목재 기둥과 처마, 짚신 걸이 일부, 낮은 판매대 하나와 크게 찢어진 직조 발이 보인다. 왼쪽 뒤로 시장 길이 길게 드러난다. 청년은 점포 바로 앞에 서서 직조물의 테두리를 잡고 있어 위치와 사용 관계는 자연스럽다. 다만 점포와 멍석이 화면 대부분을 차지하고 청년의 허리 아래까지 보여, 작은 배경 조각만 남기는 클로즈업과 다르다.",
        "entities": "중앙 청년은 검은 머리의 한국인 젊은 남성으로 보이며, 수수한 회갈색 무명 저고리와 바지를 입고 있다. 붉어진 눈가 안에 정상적인 홍채와 동공이 보인다. 망가진 직조물과 짚신 진열, 바닥에 넘어진 용기가 있다. 왼쪽 전경에는 별도의 남성 뒤통수와 어깨가 크게 추가되어 허용된 청년 한 명이라는 인물 제한을 어긴다. 읽을 수 있는 글자는 없다.",
        "hard_violations": [
         "화면에 허용되지 않은 두 번째 남성의 뒤통수와 어깨를 전경에 추가했다."
        ],
        "physics": "청년의 몸통은 서 있는 하체로 자연스럽게 이어지며 발은 프레임 밖이다. 양손이 망가진 직조물의 단단한 테두리를 각각 잡고 있고, 직조물 윗부분은 점포 구조에 기대거나 걸쳐져 있어 지지가 보인다. 바닥의 용기는 흙바닥에, 짚신과 바구니는 판매대에 놓여 있다. 지지 없이 떠 있는 물체나 불가능한 관절 자세는 보이지 않는다."
       }
      ]
     },
     "ok": true
    }
   ],
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "slot_winner_match": true,
   "slot_winner": {
    "gemini-pro": "B",
    "gpt-high": "B"
   },
   "route": "cross_slot_agree"
  },
  "dual": {
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "normalized": {
    "A": 1.067,
    "B": 2.0
   },
   "adjusted": {
    "A": 0.817,
    "B": 1.75
   },
   "violations": {
    "A": [
     "[gemini-pro] 허용되지 않은 추가 인물 등장 (프롬프트는 오직 조선인 청년 1명만 화면에 등장하도록 제한함)",
     "[gpt-high] 화면에 허용되지 않은 두 번째 남성의 뒤통수와 어깨를 전경에 추가했다."
    ],
    "B": [
     "[gpt-high] 서 있어야 한다는 지속 상태와 달리, 청년은 무릎을 가슴 가까이 올리고 낮게 웅크린 자세로 배치되어 있다."
    ]
   },
   "per_model_winner": {
    "gemini-pro": "B",
    "gpt-high": "B"
   },
   "agreed": true
  },
  "totals": {
   "B": 1750,
   "A": 817
  },
  "selected": "B",
  "ranking": [
   "B",
   "A"
  ],
  "verdicts": [
   {
    "label": "B",
    "score": 1750,
    "verdict_ko": "하드 비올레이션은 피했으나, 클로즈업 지시를 무시하고 배경을 선명하게 다 드러냈으며 서서 멍석을 만지는 대신 쪼그려 앉아 있어 프레이밍과 포즈 지시(우선순위 2, 5)를 크게 위반했습니다.  ★위반: [gpt-high] 서 있어야 한다는 지속 상태와 달리, 청년은 무릎을 가슴 가까이 올리고 낮게 웅크린 자세로 배치되어 있다."
   },
   {
    "label": "A",
    "score": 817,
    "verdict_ko": "포즈와 행동은 프롬프트에 부합하나, 화면에 오직 청년 1명만 허용한다는 지시를 어기고 전경에 다른 인물을 등장시켜 최우선 기준인 하드 비올레이션을 범했습니다.  ★위반: [gemini-pro] 허용되지 않은 추가 인물 등장 (프롬프트는 오직 조선인 청년 1명만 화면에 등장하도록 제한함) / [gpt-high] 화면에 허용되지 않은 두 번째 남성의 뒤통수와 어깨를 전경에 추가했다."
   }
  ],
  "refs": [
   {
    "label": "LOCATION PHOTOGRAPH — the exact place of this shot: its architecture, materials, fixed features and lighting mood are spatial truth; stage the moment inside this place. Never copy its camera framing.",
    "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/db00517b-165b-4005-a287-f28936b59b95/images/272bd7af-9294-4490-98c6-88a40e9deacc/scene/recipe/groupbg_market_shoe_front_6546f9.png",
    "asset_id": "a82710e0-4099-4422-b0e7-8cd69a1606bb",
    "role": "bgfirst_group_bg"
   }
  ],
  "critique_skipped": true,
  "shot_run_produced": false,
  "shot_run_uid": "06aa140c-4c0e-7df6-9b92-38edfb0037d7",
  "bgfirst": {
   "bg_path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/db00517b-165b-4005-a287-f28936b59b95/images/272bd7af-9294-4490-98c6-88a40e9deacc/scene/recipe/S10sh6__bgfirst_bg.png",
   "bg_asset_id": "525b41ef-ff22-4cae-9ff6-49fa3d7c6a24",
   "bg_record_key": "S10sh6::bgfirst_bg",
   "chain_winner": false,
   "authority": "groupbg",
   "group_key": "market_shoe_front",
   "groupbg_asset_id": "a82710e0-4099-4422-b0e7-8cd69a1606bb"
  },
  "ref_mode": "그룹 배경+엔티티 (2택1: 무콘티 승)",
  "share_plan": {
   "ref_plan": "background"
  }
 },
 "S10sh6::cine": {
  "applied": true,
  "attempted_at": "2026-09-09T01:41:06.100768+00:00",
  "fingerprint": "a421e16d4f7c64621dbceb4d9dbe9b554bd6ba74fb5a71da1a805964a2666ea8",
  "fingerprint_version": 2,
  "provider": "mai",
  "endpoint": "openrouter/chat-completions",
  "model": "microsoft/mai-image-2.6",
  "pack": "24.202608252115",
  "source_file": "S10sh6_sel.png",
  "source_sha256": "8fdc491754e2ae060a817ac136de70ec3d1eb304443cbb40e4b826198dccc8de",
  "file": "S10sh6_cine.png",
  "staged_sha256": "329bf0235265b33f2bb47f3081be43949f16226a2f3d5d85335aee40af85e136",
  "latency_ms": 27330
 },
 "S10sh13::signage": {
  "fp": "529c972b38e73468",
  "inscriptions": [],
  "cues": [],
  "dropped": []
 },
 "S10sh13": {
  "input_fingerprint": "b70f0306ad9ab776",
  "prompt": "Create ONE FINAL photorealistic live-action film still of the moment below — early 20th-century Korea, Jeolla region; all people are Korean unless explicitly stated otherwise. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 박제현의 손이 안규홍의 넓은 어깨를 꽉 움켜쥔 찰나.\n\nLOCATION (lock): The dirt standing space before the wrecked straw-sandal stall in a subdued daytime market. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: close-up\n- KEY BACKGROUND ELEMENTS: 부서진 장터 좌판 (Broken in the aftermath of the market violence) — A small oblique section appears beyond the two men's shoulders; used as Provides subdued environmental context behind the human contact.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Maintain the same restrained daylight and muted saturation, using natural contrast to reveal the pressure of the hand without dramatized highlights.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The people visible in that photo are NOT in this shot: never carry their faces, bodies or clothing onto anyone here. Each person in THIS shot is defined solely by the PEOPLE list, their own CHARACTER REFERENCE images and their wardrobe notes. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): Broken bowls and torn mats remain around the damaged stalls; the trampled straw shoe is still on the dirt. 안규홍: He stands with his head lowered and his hand clenched, fingertips trembling. 박제현: He stands in the market with his dopo gathered closed and one hand extended at shoulder height.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 안규홍 (한국인 남성, 청년 나이대, 검은 머리, 단단히 묶은 머리); 박제현 (한국인 남성, 중년 이상 나이대, 자연스러운 연령대의 얼굴) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.",
  "roll_prompts": {
   "A": "Create ONE FINAL photorealistic live-action film still of the moment below — early 20th-century Korea, Jeolla region; all people are Korean unless explicitly stated otherwise. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 박제현의 손이 안규홍의 넓은 어깨를 꽉 움켜쥔 찰나.\n\nLOCATION (lock): The dirt standing space before the wrecked straw-sandal stall in a subdued daytime market. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: close-up\n- KEY BACKGROUND ELEMENTS: 부서진 장터 좌판 (Broken in the aftermath of the market violence) — A small oblique section appears beyond the two men's shoulders; used as Provides subdued environmental context behind the human contact.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Maintain the same restrained daylight and muted saturation, using natural contrast to reveal the pressure of the hand without dramatized highlights.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The people visible in that photo are NOT in this shot: never carry their faces, bodies or clothing onto anyone here. Each person in THIS shot is defined solely by the PEOPLE list, their own CHARACTER REFERENCE images and their wardrobe notes. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): Broken bowls and torn mats remain around the damaged stalls; the trampled straw shoe is still on the dirt. 안규홍: He stands with his head lowered and his hand clenched, fingertips trembling. 박제현: He stands in the market with his dopo gathered closed and one hand extended at shoulder height.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 안규홍 (한국인 남성, 청년 나이대, 검은 머리, 단단히 묶은 머리); 박제현 (한국인 남성, 중년 이상 나이대, 자연스러운 연령대의 얼굴) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.",
   "B": "Create ONE FINAL photorealistic live-action film still of the moment below — early 20th-century Korea, Jeolla region; all people are Korean unless explicitly stated otherwise. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 박제현의 손이 안규홍의 넓은 어깨를 꽉 움켜쥔 찰나.\n\nLOCATION (lock): The dirt standing space before the wrecked straw-sandal stall in a subdued daytime market. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: close-up\n- KEY BACKGROUND ELEMENTS: 부서진 장터 좌판 (Broken in the aftermath of the market violence) — A small oblique section appears beyond the two men's shoulders; used as Provides subdued environmental context behind the human contact.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Maintain the same restrained daylight and muted saturation, using natural contrast to reveal the pressure of the hand without dramatized highlights.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The people visible in that photo are NOT in this shot: never carry their faces, bodies or clothing onto anyone here. Each person in THIS shot is defined solely by the PEOPLE list, their own CHARACTER REFERENCE images and their wardrobe notes. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): Broken bowls and torn mats remain around the damaged stalls; the trampled straw shoe is still on the dirt. 안규홍: He stands with his head lowered and his hand clenched, fingertips trembling. 박제현: He stands in the market with his dopo gathered closed and one hand extended at shoulder height.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 안규홍 (한국인 남성, 청년 나이대, 검은 머리, 단단히 묶은 머리); 박제현 (한국인 남성, 중년 이상 나이대, 자연스러운 연령대의 얼굴) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.\n\nCOMPOSITION VARIATION (this candidate only): keep the place, the people and the moment exactly as specified, and keep from the CAMERA & FRAME contract its subject scale, screen placement and key background elements exactly. For this candidate only, its camera angle and height are deliberately left open: choose a camera position distinctly different from the obvious one, as a film director picking a second setup on the same blocking. Stay true to any explicit framing the SHOT TEXT itself dictates."
  },
  "shot_run_spend_attempt_count": 1,
  "readings": [
   {
    "label": "A",
    "direction": "안규홍은 고개를 숙이고 시선은 프레임 밖 바닥을 향하고 있습니다. 박제현의 오른손은 안규홍의 왼쪽 어깨를 향해 뻗어 정확히 닿아 있으며 단단히 움켜쥐고 있습니다.",
    "built_space": "부서진 장터 좌판 앞 흙바닥입니다. 화면 왼쪽에 짚신이 진열된 나무 좌판 구조물이 있고, 바닥에는 찢어진 돗자리와 짚신들이 놓여 있습니다. 뒤쪽으로는 나무통 등 이전 컷에서 제시된 고정 배경 요소와 위치 및 비례가 정확히 일치합니다.",
    "entities": "안규홍이 등장하나 캐릭터 레퍼런스의 '단단히 묶은 머리'가 아닌 이전 컷 인물의 짧은 머리와 얼굴을 그대로 가져왔습니다. 박제현은 어깨를 잡은 손과 몸통 일부로 등장하며, 레퍼런스의 밝은색 도포 대신 검은색 의상을 입고 있습니다.",
    "hard_violations": [],
    "physics": "안규홍은 서 있는 자세로 오른손을 가슴 높이로 들어 주먹을 쥐고 있으며, 이는 스스로의 팔 근육으로 지탱됩니다. 박제현의 오른손은 안규홍의 어깨를 꽉 쥐어 압박감이 표현되었고, 바닥의 돗자리와 짚신은 중력에 맞게 자연스럽게 놓여 있습니다."
   },
   {
    "label": "B",
    "direction": "안규홍은 고개를 숙이고 시선을 아래로 향하고 있습니다. 오른쪽 인물은 안규홍을 바라보며 왼손을 뻗어 안규홍의 오른쪽 어깨에 올렸고, 왼쪽 전경의 추가 인물은 안규홍을 향해 서 있습니다.",
    "built_space": "부서진 좌판 앞 흙바닥입니다. 왼쪽에 짚신이 걸린 구조물이 보이고, 바닥에 찢어진 돗자리와 짚신이 있으며 뒷배경에 큰 솥 등 장터 구조물들이 위치해 있습니다.",
    "entities": "안규홍은 단단히 묶은 머리를 하고 있어 캐릭터 레퍼런스와 일치합니다. 그러나 프레임 좌측 전경에 지시되지 않은 제3의 인물이 추가로 등장하며, 오른쪽의 박제현 추정 인물 역시 레퍼런스와 달리 검은색 옷을 입고 있습니다.",
    "hard_violations": [
     "[gemini-pro] invented people (프레임 왼쪽 전경에 프롬프트에 존재하지 않는 제3의 인물이 등장함)",
     "[gpt-high] 지정된 안규홍과 박제현 외에 왼쪽 전경에 세 번째 남성의 머리와 몸통이 추가되어 있습니다."
    ],
    "physics": "안규홍은 서서 오른손 주먹을 쥐고 팔을 아래로 늘어뜨리고 있습니다. 오른쪽 인물의 왼손은 안규홍의 어깨에 올려져 지탱되며, 배경의 짚신과 돗자리는 바닥에 정상적으로 위치합니다."
   }
  ],
  "cross_model_order": {
   "policy": "cross_model_order_v2_slot0_fwd_slot1_rev_parallel2",
   "slots": [
    {
     "model": "gemini-pro",
     "order": "forward",
     "display_to_canonical": {
      "A": "A",
      "B": "B"
     },
     "raw": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 7,
        "verdict_ko": "안규홍의 외모와 헤어스타일이 레퍼런스가 아닌 이전 컷 인물을 따라갔고 박제현의 의상 색상이 다르지만, 요구된 클로즈업 구도 내에서 어깨를 강하게 움켜쥐는 동작을 완벽하게 연출했습니다."
       },
       {
        "label": "B",
        "score": 3,
        "verdict_ko": "안규홍의 캐릭터 레퍼런스를 잘 반영했으나, 프롬프트에 지시되지 않은 제3의 인물이 화면 왼쪽에 추가로 등장하는 치명적인 오류(Hard Violation)가 있습니다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "안규홍은 고개를 숙이고 시선은 프레임 밖 바닥을 향하고 있습니다. 박제현의 오른손은 안규홍의 왼쪽 어깨를 향해 뻗어 정확히 닿아 있으며 단단히 움켜쥐고 있습니다.",
        "built_space": "부서진 장터 좌판 앞 흙바닥입니다. 화면 왼쪽에 짚신이 진열된 나무 좌판 구조물이 있고, 바닥에는 찢어진 돗자리와 짚신들이 놓여 있습니다. 뒤쪽으로는 나무통 등 이전 컷에서 제시된 고정 배경 요소와 위치 및 비례가 정확히 일치합니다.",
        "entities": "안규홍이 등장하나 캐릭터 레퍼런스의 '단단히 묶은 머리'가 아닌 이전 컷 인물의 짧은 머리와 얼굴을 그대로 가져왔습니다. 박제현은 어깨를 잡은 손과 몸통 일부로 등장하며, 레퍼런스의 밝은색 도포 대신 검은색 의상을 입고 있습니다.",
        "hard_violations": [],
        "physics": "안규홍은 서 있는 자세로 오른손을 가슴 높이로 들어 주먹을 쥐고 있으며, 이는 스스로의 팔 근육으로 지탱됩니다. 박제현의 오른손은 안규홍의 어깨를 꽉 쥐어 압박감이 표현되었고, 바닥의 돗자리와 짚신은 중력에 맞게 자연스럽게 놓여 있습니다."
       },
       {
        "label": "B",
        "direction": "안규홍은 고개를 숙이고 시선을 아래로 향하고 있습니다. 오른쪽 인물은 안규홍을 바라보며 왼손을 뻗어 안규홍의 오른쪽 어깨에 올렸고, 왼쪽 전경의 추가 인물은 안규홍을 향해 서 있습니다.",
        "built_space": "부서진 좌판 앞 흙바닥입니다. 왼쪽에 짚신이 걸린 구조물이 보이고, 바닥에 찢어진 돗자리와 짚신이 있으며 뒷배경에 큰 솥 등 장터 구조물들이 위치해 있습니다.",
        "entities": "안규홍은 단단히 묶은 머리를 하고 있어 캐릭터 레퍼런스와 일치합니다. 그러나 프레임 좌측 전경에 지시되지 않은 제3의 인물이 추가로 등장하며, 오른쪽의 박제현 추정 인물 역시 레퍼런스와 달리 검은색 옷을 입고 있습니다.",
        "hard_violations": [
         "invented people (프레임 왼쪽 전경에 프롬프트에 존재하지 않는 제3의 인물이 등장함)"
        ],
        "physics": "안규홍은 서서 오른손 주먹을 쥐고 팔을 아래로 늘어뜨리고 있습니다. 오른쪽 인물의 왼손은 안규홍의 어깨에 올려져 지탱되며, 배경의 짚신과 돗자리는 바닥에 정상적으로 위치합니다."
       }
      ],
      "all_candidates_fail": false
     },
     "normalized": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 7,
        "verdict_ko": "안규홍의 외모와 헤어스타일이 레퍼런스가 아닌 이전 컷 인물을 따라갔고 박제현의 의상 색상이 다르지만, 요구된 클로즈업 구도 내에서 어깨를 강하게 움켜쥐는 동작을 완벽하게 연출했습니다."
       },
       {
        "label": "B",
        "score": 3,
        "verdict_ko": "안규홍의 캐릭터 레퍼런스를 잘 반영했으나, 프롬프트에 지시되지 않은 제3의 인물이 화면 왼쪽에 추가로 등장하는 치명적인 오류(Hard Violation)가 있습니다."
       }
      ],
      "all_candidates_fail": false,
      "readings": [
       {
        "label": "A",
        "direction": "안규홍은 고개를 숙이고 시선은 프레임 밖 바닥을 향하고 있습니다. 박제현의 오른손은 안규홍의 왼쪽 어깨를 향해 뻗어 정확히 닿아 있으며 단단히 움켜쥐고 있습니다.",
        "built_space": "부서진 장터 좌판 앞 흙바닥입니다. 화면 왼쪽에 짚신이 진열된 나무 좌판 구조물이 있고, 바닥에는 찢어진 돗자리와 짚신들이 놓여 있습니다. 뒤쪽으로는 나무통 등 이전 컷에서 제시된 고정 배경 요소와 위치 및 비례가 정확히 일치합니다.",
        "entities": "안규홍이 등장하나 캐릭터 레퍼런스의 '단단히 묶은 머리'가 아닌 이전 컷 인물의 짧은 머리와 얼굴을 그대로 가져왔습니다. 박제현은 어깨를 잡은 손과 몸통 일부로 등장하며, 레퍼런스의 밝은색 도포 대신 검은색 의상을 입고 있습니다.",
        "hard_violations": [],
        "physics": "안규홍은 서 있는 자세로 오른손을 가슴 높이로 들어 주먹을 쥐고 있으며, 이는 스스로의 팔 근육으로 지탱됩니다. 박제현의 오른손은 안규홍의 어깨를 꽉 쥐어 압박감이 표현되었고, 바닥의 돗자리와 짚신은 중력에 맞게 자연스럽게 놓여 있습니다."
       },
       {
        "label": "B",
        "direction": "안규홍은 고개를 숙이고 시선을 아래로 향하고 있습니다. 오른쪽 인물은 안규홍을 바라보며 왼손을 뻗어 안규홍의 오른쪽 어깨에 올렸고, 왼쪽 전경의 추가 인물은 안규홍을 향해 서 있습니다.",
        "built_space": "부서진 좌판 앞 흙바닥입니다. 왼쪽에 짚신이 걸린 구조물이 보이고, 바닥에 찢어진 돗자리와 짚신이 있으며 뒷배경에 큰 솥 등 장터 구조물들이 위치해 있습니다.",
        "entities": "안규홍은 단단히 묶은 머리를 하고 있어 캐릭터 레퍼런스와 일치합니다. 그러나 프레임 좌측 전경에 지시되지 않은 제3의 인물이 추가로 등장하며, 오른쪽의 박제현 추정 인물 역시 레퍼런스와 달리 검은색 옷을 입고 있습니다.",
        "hard_violations": [
         "invented people (프레임 왼쪽 전경에 프롬프트에 존재하지 않는 제3의 인물이 등장함)"
        ],
        "physics": "안규홍은 서서 오른손 주먹을 쥐고 팔을 아래로 늘어뜨리고 있습니다. 오른쪽 인물의 왼손은 안규홍의 어깨에 올려져 지탱되며, 배경의 짚신과 돗자리는 바닥에 정상적으로 위치합니다."
       }
      ]
     },
     "ok": true
    },
    {
     "model": "gpt-high",
     "order": "reverse",
     "display_to_canonical": {
      "A": "B",
      "B": "A"
     },
     "raw": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 1,
        "verdict_ko": "불필요한 세 번째 인물이 등장하는 실격 오류가 있으며, 어깨 접촉보다 넓은 인물·장터 구도와 박제현의 잘못된 의상이 지시를 어깁니다."
       },
       {
        "label": "B",
        "score": 4,
        "verdict_ko": "두 사람의 어깨 움켜쥐기를 더 가깝게 보여주지만, 안규홍 대신 이전 장면 인물의 얼굴·짧은 머리를 재현하고 박제현의 도포 색도 틀렸습니다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "중앙 안규홍은 고개와 시선을 아래로 내리고 있습니다. 오른쪽 남성의 손은 안규홍의 어깨 바깥쪽에 닿아 손가락으로 감싸며, 해당 남성의 얼굴도 안규홍 쪽을 향합니다. 왼쪽의 추가 남성 역시 중앙 인물을 향하지만 눈은 보이지 않습니다.",
        "built_space": "중앙 인물 양옆에 남성이 한 명씩 있어 총 세 사람이 좌판 앞을 차지합니다. 뒤에는 파손된 목제 좌판 한 구역, 크게 찢어진 멍석 한 장, 기대어진 직조 판 한 개, 나무통 한 개, 큰 검은 용기 한 개와 흙바닥의 짚신이 보입니다. 장소의 재질과 파손 흔적은 이전 장면과 유사하지만, 배경이 작은 비스듬한 단편에 그치지 않고 화면 왼쪽을 크게 차지합니다. 중앙 인물도 허벅지 부근까지 보여 요구한 접촉 중심 클로즈업보다 넓습니다.",
        "entities": "총 세 명의 한국인 남성으로 보이는 인물이 있어 허용된 두 사람을 초과합니다. 중앙 청년은 검게 묶은 머리, 때 탄 밝은 저고리와 허리띠가 안규홍 참고에 대체로 부합합니다. 어깨를 잡는 오른쪽 남성은 짧은 머리와 짙은 도포 차림으로, 박제현 참고의 갓과 연한 크림색 도포에 맞지 않습니다. 왼쪽 검은 복장의 남성은 지정되지 않은 인물입니다. 깨진 그릇류, 찢어진 멍석과 짚신이 보이며 읽을 수 있는 글자는 없습니다.",
        "hard_violations": [
         "지정된 안규홍과 박제현 외에 왼쪽 전경에 세 번째 남성의 머리와 몸통이 추가되어 있습니다."
        ],
        "physics": "어깨를 잡는 손은 오른쪽 남성의 소매와 팔에 연결되어 있고 손가락이 어깨 곡면에 접촉합니다. 다만 옷감을 깊게 움켜쥔 압력보다 넓게 얹어 감싼 동작에 가깝습니다. 안규홍의 주먹은 팔 끝에 자연스럽게 연결됩니다. 발은 프레임 밖이므로 지면 접촉은 확인할 수 없지만 공중에 뜬 몸은 보이지 않습니다. 멍석과 짚신은 흙바닥에, 용기들은 바닥이나 좌판에 지지되어 있습니다."
       },
       {
        "label": "B",
        "direction": "앞쪽 남성은 얼굴을 오른쪽 아래로 기울이고 시선도 그 방향으로 내립니다. 뒤쪽 인물의 손은 앞쪽 남성의 어깨 윗부분을 정확히 겨냥해 옷감을 움켜쥡니다. 앞쪽 남성은 자신의 가슴 앞에 주먹을 들고 있으며, 뒤쪽 인물의 얼굴과 시선은 프레임 밖입니다.",
        "built_space": "앞쪽 남성 한 명과 그 뒤에서 손을 뻗은 인물 한 명만 보입니다. 뒤에는 짚신 진열대 한 구역, 파손된 좌판, 찢어진 멍석 한 장, 기대어진 직조 판 한 개, 나무통 한 개와 큰 검은 용기 한 개가 있습니다. 흙바닥에는 짚신 한 짝이 보입니다. 이전 장소의 목재·직조물·잔해는 유지되지만 배경이 화면 왼쪽의 상당 부분을 차지해 요구된 작은 배경 단편보다 큽니다. A보다 가까운 상반신 구도이나 접촉 부위 외에 가슴과 주먹까지 강조됩니다.",
        "entities": "인물 수는 두 명으로 맞습니다. 그러나 앞쪽 남성은 안규홍 참고의 얼굴과 단단히 묶은 머리가 아니라, 이전 장면 인물과 매우 유사한 얼굴 및 짧게 자른 머리를 보입니다. 밝은 낡은 저고리만 안규홍 의상과 대체로 맞습니다. 뒤쪽 인물은 남성의 손과 몸통만 보여 얼굴·머리·나이를 확인할 수 없으며, 보이는 도포는 박제현 참고의 크림색이 아닌 짙은 남색입니다. 찢어진 멍석, 그릇 잔해와 짚신은 있으며 읽을 수 있는 글자는 없습니다.",
        "hard_violations": [],
        "physics": "뒤쪽 인물의 손은 굽힌 팔과 소매에 연결되고, 말아 쥔 손가락이 어깨 위 옷감을 잡아 주름을 만듭니다. 어깨 접촉과 압력은 물리적으로 가능합니다. 앞쪽 주먹도 손목과 굽힌 팔에 정상적으로 연결됩니다. 앞쪽 남성은 몸을 숙였지만 하체가 잘려 서 있는지 앉아 있는지는 확정할 수 없습니다. 부유하는 신체나 물체는 보이지 않으며 배경 잔해는 지면과 좌판에 놓여 있습니다."
       }
      ],
      "all_candidates_fail": true
     },
     "normalized": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 1,
        "verdict_ko": "불필요한 세 번째 인물이 등장하는 실격 오류가 있으며, 어깨 접촉보다 넓은 인물·장터 구도와 박제현의 잘못된 의상이 지시를 어깁니다."
       },
       {
        "label": "A",
        "score": 4,
        "verdict_ko": "두 사람의 어깨 움켜쥐기를 더 가깝게 보여주지만, 안규홍 대신 이전 장면 인물의 얼굴·짧은 머리를 재현하고 박제현의 도포 색도 틀렸습니다."
       }
      ],
      "all_candidates_fail": true,
      "readings": [
       {
        "label": "B",
        "direction": "중앙 안규홍은 고개와 시선을 아래로 내리고 있습니다. 오른쪽 남성의 손은 안규홍의 어깨 바깥쪽에 닿아 손가락으로 감싸며, 해당 남성의 얼굴도 안규홍 쪽을 향합니다. 왼쪽의 추가 남성 역시 중앙 인물을 향하지만 눈은 보이지 않습니다.",
        "built_space": "중앙 인물 양옆에 남성이 한 명씩 있어 총 세 사람이 좌판 앞을 차지합니다. 뒤에는 파손된 목제 좌판 한 구역, 크게 찢어진 멍석 한 장, 기대어진 직조 판 한 개, 나무통 한 개, 큰 검은 용기 한 개와 흙바닥의 짚신이 보입니다. 장소의 재질과 파손 흔적은 이전 장면과 유사하지만, 배경이 작은 비스듬한 단편에 그치지 않고 화면 왼쪽을 크게 차지합니다. 중앙 인물도 허벅지 부근까지 보여 요구한 접촉 중심 클로즈업보다 넓습니다.",
        "entities": "총 세 명의 한국인 남성으로 보이는 인물이 있어 허용된 두 사람을 초과합니다. 중앙 청년은 검게 묶은 머리, 때 탄 밝은 저고리와 허리띠가 안규홍 참고에 대체로 부합합니다. 어깨를 잡는 오른쪽 남성은 짧은 머리와 짙은 도포 차림으로, 박제현 참고의 갓과 연한 크림색 도포에 맞지 않습니다. 왼쪽 검은 복장의 남성은 지정되지 않은 인물입니다. 깨진 그릇류, 찢어진 멍석과 짚신이 보이며 읽을 수 있는 글자는 없습니다.",
        "hard_violations": [
         "지정된 안규홍과 박제현 외에 왼쪽 전경에 세 번째 남성의 머리와 몸통이 추가되어 있습니다."
        ],
        "physics": "어깨를 잡는 손은 오른쪽 남성의 소매와 팔에 연결되어 있고 손가락이 어깨 곡면에 접촉합니다. 다만 옷감을 깊게 움켜쥔 압력보다 넓게 얹어 감싼 동작에 가깝습니다. 안규홍의 주먹은 팔 끝에 자연스럽게 연결됩니다. 발은 프레임 밖이므로 지면 접촉은 확인할 수 없지만 공중에 뜬 몸은 보이지 않습니다. 멍석과 짚신은 흙바닥에, 용기들은 바닥이나 좌판에 지지되어 있습니다."
       },
       {
        "label": "A",
        "direction": "앞쪽 남성은 얼굴을 오른쪽 아래로 기울이고 시선도 그 방향으로 내립니다. 뒤쪽 인물의 손은 앞쪽 남성의 어깨 윗부분을 정확히 겨냥해 옷감을 움켜쥡니다. 앞쪽 남성은 자신의 가슴 앞에 주먹을 들고 있으며, 뒤쪽 인물의 얼굴과 시선은 프레임 밖입니다.",
        "built_space": "앞쪽 남성 한 명과 그 뒤에서 손을 뻗은 인물 한 명만 보입니다. 뒤에는 짚신 진열대 한 구역, 파손된 좌판, 찢어진 멍석 한 장, 기대어진 직조 판 한 개, 나무통 한 개와 큰 검은 용기 한 개가 있습니다. 흙바닥에는 짚신 한 짝이 보입니다. 이전 장소의 목재·직조물·잔해는 유지되지만 배경이 화면 왼쪽의 상당 부분을 차지해 요구된 작은 배경 단편보다 큽니다. A보다 가까운 상반신 구도이나 접촉 부위 외에 가슴과 주먹까지 강조됩니다.",
        "entities": "인물 수는 두 명으로 맞습니다. 그러나 앞쪽 남성은 안규홍 참고의 얼굴과 단단히 묶은 머리가 아니라, 이전 장면 인물과 매우 유사한 얼굴 및 짧게 자른 머리를 보입니다. 밝은 낡은 저고리만 안규홍 의상과 대체로 맞습니다. 뒤쪽 인물은 남성의 손과 몸통만 보여 얼굴·머리·나이를 확인할 수 없으며, 보이는 도포는 박제현 참고의 크림색이 아닌 짙은 남색입니다. 찢어진 멍석, 그릇 잔해와 짚신은 있으며 읽을 수 있는 글자는 없습니다.",
        "hard_violations": [],
        "physics": "뒤쪽 인물의 손은 굽힌 팔과 소매에 연결되고, 말아 쥔 손가락이 어깨 위 옷감을 잡아 주름을 만듭니다. 어깨 접촉과 압력은 물리적으로 가능합니다. 앞쪽 주먹도 손목과 굽힌 팔에 정상적으로 연결됩니다. 앞쪽 남성은 몸을 숙였지만 하체가 잘려 서 있는지 앉아 있는지는 확정할 수 없습니다. 부유하는 신체나 물체는 보이지 않으며 배경 잔해는 지면과 좌판에 놓여 있습니다."
       }
      ]
     },
     "ok": true
    }
   ],
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "slot_winner_match": true,
   "slot_winner": {
    "gemini-pro": "A",
    "gpt-high": "A"
   },
   "route": "cross_slot_agree"
  },
  "dual": {
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "normalized": {
    "A": 2.0,
    "B": 0.679
   },
   "adjusted": {
    "A": 2.0,
    "B": 0.429
   },
   "violations": {
    "B": [
     "[gemini-pro] invented people (프레임 왼쪽 전경에 프롬프트에 존재하지 않는 제3의 인물이 등장함)",
     "[gpt-high] 지정된 안규홍과 박제현 외에 왼쪽 전경에 세 번째 남성의 머리와 몸통이 추가되어 있습니다."
    ]
   },
   "per_model_winner": {
    "gemini-pro": "A",
    "gpt-high": "A"
   },
   "agreed": true
  },
  "totals": {
   "A": 2000,
   "B": 429
  },
  "selected": "A",
  "ranking": [
   "A",
   "B"
  ],
  "verdicts": [
   {
    "label": "A",
    "score": 2000,
    "verdict_ko": "안규홍의 외모와 헤어스타일이 레퍼런스가 아닌 이전 컷 인물을 따라갔고 박제현의 의상 색상이 다르지만, 요구된 클로즈업 구도 내에서 어깨를 강하게 움켜쥐는 동작을 완벽하게 연출했습니다."
   },
   {
    "label": "B",
    "score": 429,
    "verdict_ko": "안규홍의 캐릭터 레퍼런스를 잘 반영했으나, 프롬프트에 지시되지 않은 제3의 인물이 화면 왼쪽에 추가로 등장하는 치명적인 오류(Hard Violation)가 있습니다.  ★위반: [gemini-pro] invented people (프레임 왼쪽 전경에 프롬프트에 존재하지 않는 제3의 인물이 등장함) / [gpt-high] 지정된 안규홍과 박제현 외에 왼쪽 전경에 세 번째 남성의 머리와 몸통이 추가되어 있습니다."
   }
  ],
  "refs": [
   {
    "label": "PREVIOUS SHOT STILL — a visually related earlier shot of this same place: the location's look, materials, fixed features and lighting mood are LOCKED to this photo; never copy its camera framing. The people visible in that photo are NOT in this shot: never carry their faces, bodies or clothing onto anyone here. Each person in THIS shot is defined solely by the PEOPLE list, their own CHARACTER REFERENCE images and their wardrobe notes. If this photo shows a character who cannot move (dead or unconscious), that character's exact pose, position and orientation are ALSO LOCKED — treat their body as a fixed prop of the set that only the camera moves around.",
    "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/db00517b-165b-4005-a287-f28936b59b95/images/272bd7af-9294-4490-98c6-88a40e9deacc/scene/recipe/S10sh6_sel.png",
    "asset_id": "436233a6-ac99-438d-b4c2-f4b4611eda40",
    "role": "prev_still"
   },
   {
    "label": "CHARACTER REFERENCE — 안규홍: the exact person appearing in this shot; match face, hair and build exactly.",
    "path": "<bytes:1028384>",
    "asset_id": "3b5894db-ce55-4f34-b966-5e725c3aa952",
    "role": "character_ref"
   },
   {
    "label": "CHARACTER REFERENCE — 박제현: the exact person appearing in this shot; match face, hair and build exactly.",
    "path": "<bytes:870304>",
    "asset_id": "02a56dab-2641-421d-89df-af22b5de63df",
    "role": "character_ref"
   }
  ],
  "critique_skipped": true,
  "shot_run_produced": false,
  "shot_run_uid": "06aa140c-5013-78de-8a0c-b6235b3743b2",
  "ref_mode": "prev+엔티티",
  "share_plan": {
   "ref_plan": "prev",
   "prev_anchor_tag": "S10sh6"
  }
 },
 "S10sh13::cine": {
  "applied": true,
  "attempted_at": "2026-09-09T01:43:22.025136+00:00",
  "fingerprint": "2374e4f63e7e7ff069d0a56ea6cfa058b2ee1719f9ac364873ed82b99cb5b418",
  "fingerprint_version": 2,
  "provider": "mai",
  "endpoint": "openrouter/chat-completions",
  "model": "microsoft/mai-image-2.6",
  "pack": "24.202608252115",
  "source_file": "S10sh13_sel.png",
  "source_sha256": "5d2d1d50a6e4d4e9938b397f9098b18f0e090ec2940966f33fd16b6228b0c8f8",
  "file": "S10sh13_cine.png",
  "staged_sha256": "2e689d1329d24cd4101cf555ecccdfcc84d65ead2abf601afbc1eefcc647ac3a",
  "latency_ms": 25795
 },
 "S10sh15::signage": {
  "fp": "928af7f177f457d6",
  "inscriptions": [],
  "cues": [],
  "dropped": []
 },
 "S10sh15": {
  "input_fingerprint": "587793f9bd6b5776",
  "prompt": "Create ONE FINAL photorealistic live-action film still of the moment below — early 20th-century Korea, Jeolla region; all people are Korean unless explicitly stated otherwise. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 박살 난 짚신전 앞 텅 빈 장터에 홀로 우두커니 서 있는 안규홍의 외로운 전신.\n\nLOCATION (lock): The deserted frontage of a broken straw-sandal stall, opening onto the market's dirt lane in daylight. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: wide shot\n- KEY BACKGROUND ELEMENTS: 부서진 짚신전 (Damaged, with 안규홍 remaining alone immediately before it) — The frontage is seen obliquely from above behind 안규홍; used as Occupies a limited portion of the upper frame as the tangible source of his isolation; 장터 흙바닥 (The visible ground surrounding 안규홍 is unoccupied); used as Negative space separates his complete figure from the edges of the frame; 깨진 그릇 (Broken pieces lie scattered on the ground); used as Small peripheral details sustain the aftermath without overwhelming the lone figure.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Unchanged natural daylight and restrained saturation expose the damaged market plainly, without turning the isolation into a theatrical lighting effect.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The clothing, hair and overall look of 안규홍 — who appear both in that photo and in this shot — are LOCKED to that photo. Anyone else visible in that photo is NOT in this shot: never carry their face, body or clothing onto anyone here. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The shoe stall remains wrecked, with a trampled shoe on the dirt and broken bowls and mats around it. 안규홍: He stands motionless before the wrecked shoe stall, without a carrying frame.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 안규홍 (한국인 남성, 청년 나이대, 검은 머리, 단단히 묶은 머리) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.",
  "roll_prompts": {
   "A": "Create ONE FINAL photorealistic live-action film still of the moment below — early 20th-century Korea, Jeolla region; all people are Korean unless explicitly stated otherwise. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 박살 난 짚신전 앞 텅 빈 장터에 홀로 우두커니 서 있는 안규홍의 외로운 전신.\n\nLOCATION (lock): The deserted frontage of a broken straw-sandal stall, opening onto the market's dirt lane in daylight. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: wide shot\n- KEY BACKGROUND ELEMENTS: 부서진 짚신전 (Damaged, with 안규홍 remaining alone immediately before it) — The frontage is seen obliquely from above behind 안규홍; used as Occupies a limited portion of the upper frame as the tangible source of his isolation; 장터 흙바닥 (The visible ground surrounding 안규홍 is unoccupied); used as Negative space separates his complete figure from the edges of the frame; 깨진 그릇 (Broken pieces lie scattered on the ground); used as Small peripheral details sustain the aftermath without overwhelming the lone figure.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Unchanged natural daylight and restrained saturation expose the damaged market plainly, without turning the isolation into a theatrical lighting effect.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The clothing, hair and overall look of 안규홍 — who appear both in that photo and in this shot — are LOCKED to that photo. Anyone else visible in that photo is NOT in this shot: never carry their face, body or clothing onto anyone here. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The shoe stall remains wrecked, with a trampled shoe on the dirt and broken bowls and mats around it. 안규홍: He stands motionless before the wrecked shoe stall, without a carrying frame.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 안규홍 (한국인 남성, 청년 나이대, 검은 머리, 단단히 묶은 머리) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.",
   "B": "Create ONE FINAL photorealistic live-action film still of the moment below — early 20th-century Korea, Jeolla region; all people are Korean unless explicitly stated otherwise. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 박살 난 짚신전 앞 텅 빈 장터에 홀로 우두커니 서 있는 안규홍의 외로운 전신.\n\nLOCATION (lock): The deserted frontage of a broken straw-sandal stall, opening onto the market's dirt lane in daylight. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: wide shot\n- KEY BACKGROUND ELEMENTS: 부서진 짚신전 (Damaged, with 안규홍 remaining alone immediately before it) — The frontage is seen obliquely from above behind 안규홍; used as Occupies a limited portion of the upper frame as the tangible source of his isolation; 장터 흙바닥 (The visible ground surrounding 안규홍 is unoccupied); used as Negative space separates his complete figure from the edges of the frame; 깨진 그릇 (Broken pieces lie scattered on the ground); used as Small peripheral details sustain the aftermath without overwhelming the lone figure.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Unchanged natural daylight and restrained saturation expose the damaged market plainly, without turning the isolation into a theatrical lighting effect.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The clothing, hair and overall look of 안규홍 — who appear both in that photo and in this shot — are LOCKED to that photo. Anyone else visible in that photo is NOT in this shot: never carry their face, body or clothing onto anyone here. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The shoe stall remains wrecked, with a trampled shoe on the dirt and broken bowls and mats around it. 안규홍: He stands motionless before the wrecked shoe stall, without a carrying frame.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 안규홍 (한국인 남성, 청년 나이대, 검은 머리, 단단히 묶은 머리) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.\n\nCOMPOSITION VARIATION (this candidate only): keep the place, the people and the moment exactly as specified, and keep from the CAMERA & FRAME contract its subject scale, screen placement and key background elements exactly. For this candidate only, its camera angle and height are deliberately left open: choose a camera position distinctly different from the obvious one, as a film director picking a second setup on the same blocking. Stay true to any explicit framing the SHOT TEXT itself dictates."
  },
  "shot_run_spend_attempt_count": 1,
  "readings": [
   {
    "label": "A",
    "direction": "안규홍은 카메라를 등지고 파괴된 짚신전을 향해 서 있으며, 카메라는 지시문대로 그의 뒤에서 비스듬히 위에서 내려다보고 있음.",
    "built_space": "장터의 흙바닥. 왼쪽 배경에는 이전 샷과 구조가 일치하는 낮은 평상 형태의 짚신전이 무너져 있으며, 찢어진 짚석과 깨진 그릇 조각들이 바닥에 흩어져 있음. 오른쪽 배경에는 이전 샷의 커다란 검은 솥이 올바르게 위치함.",
    "entities": "안규홍은 뒷모습만 보이나 지정된 흰색 의상과 체형이 레퍼런스와 일치함. 짚신전은 지시문에 맞게 심하게 파손된 상태임. 바닥의 깨진 그릇 파편들도 확인됨.",
    "hard_violations": [],
    "physics": "안규홍은 두 발로 흙바닥을 단단히 딛고 서 있으며, 무너진 목재, 짚석, 그릇 파편 등 모든 배경 물체들이 중력에 맞게 바닥과 평상 위에 안정적으로 놓여 있음."
   },
   {
    "label": "B",
    "direction": "안규홍은 정면의 카메라를 멍하니 응시하고 있으며, 짚신전은 그의 뒤에 위치함. 카메라는 안규홍의 정면에 위치함.",
    "built_space": "장터 흙바닥. 안규홍 뒤로 짚신전이 위치하나, 이전 샷에 존재하지 않던 높고 온전한 3단 나무 선반이 추가되어 있으며 그 위에 짚신들이 가지런히 정렬되어 있음. 바닥에는 찢어진 짚석과 깨진 그릇 조각이 있음.",
    "entities": "안규홍의 얼굴과 의상은 레퍼런스 이미지와 완벽히 일치함. 하지만 짚신전은 '박살 난'이라는 텍스트 지시와 달리 전혀 파손되지 않은 온전한 선반 형태를 유지하고 있음.",
    "hard_violations": [
     "[gemini-pro] 이전 샷에 없던 3단 나무 선반을 화면에 임의로 생성하여 로케이션 고정 지시를 위반함.",
     "[gemini-pro] '부서진 짚신전(Damaged)'이라는 텍스트 지시와 완전히 상반되게 물건이 가지런히 놓인 멀쩡한 선반을 그림."
    ],
    "physics": "안규홍은 바닥에 정상적으로 서 있으며, 바닥의 짚석과 그릇 파편들도 흙바닥 위에 안정적으로 지탱되어 있음."
   }
  ],
  "cross_model_order": {
   "policy": "cross_model_order_v2_slot0_fwd_slot1_rev_parallel2",
   "slots": [
    {
     "model": "gemini-pro",
     "order": "forward",
     "display_to_canonical": {
      "A": "A",
      "B": "B"
     },
     "raw": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 8,
        "verdict_ko": "안규홍의 뒤에서 내려다보는 카메라 앵글을 정확히 따랐으며, 이전 샷과 일치하는 구조의 짚신전이 무너진 채 방치된 모습을 충실하게 구현하여 프롬프트의 의도를 완벽히 살렸습니다."
       },
       {
        "label": "B",
        "score": 4,
        "verdict_ko": "캐릭터의 얼굴 묘사는 레퍼런스와 일치하나, 카메라 위치 지시(안규홍의 뒤에서)를 어겼으며, 이전 샷에 없던 멀쩡한 3단 선반을 만들어내어 '박살 난 짚신전'이라는 핵심 배경 조건을 위반했습니다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "안규홍은 카메라를 등지고 파괴된 짚신전을 향해 서 있으며, 카메라는 지시문대로 그의 뒤에서 비스듬히 위에서 내려다보고 있음.",
        "built_space": "장터의 흙바닥. 왼쪽 배경에는 이전 샷과 구조가 일치하는 낮은 평상 형태의 짚신전이 무너져 있으며, 찢어진 짚석과 깨진 그릇 조각들이 바닥에 흩어져 있음. 오른쪽 배경에는 이전 샷의 커다란 검은 솥이 올바르게 위치함.",
        "entities": "안규홍은 뒷모습만 보이나 지정된 흰색 의상과 체형이 레퍼런스와 일치함. 짚신전은 지시문에 맞게 심하게 파손된 상태임. 바닥의 깨진 그릇 파편들도 확인됨.",
        "hard_violations": [],
        "physics": "안규홍은 두 발로 흙바닥을 단단히 딛고 서 있으며, 무너진 목재, 짚석, 그릇 파편 등 모든 배경 물체들이 중력에 맞게 바닥과 평상 위에 안정적으로 놓여 있음."
       },
       {
        "label": "B",
        "direction": "안규홍은 정면의 카메라를 멍하니 응시하고 있으며, 짚신전은 그의 뒤에 위치함. 카메라는 안규홍의 정면에 위치함.",
        "built_space": "장터 흙바닥. 안규홍 뒤로 짚신전이 위치하나, 이전 샷에 존재하지 않던 높고 온전한 3단 나무 선반이 추가되어 있으며 그 위에 짚신들이 가지런히 정렬되어 있음. 바닥에는 찢어진 짚석과 깨진 그릇 조각이 있음.",
        "entities": "안규홍의 얼굴과 의상은 레퍼런스 이미지와 완벽히 일치함. 하지만 짚신전은 '박살 난'이라는 텍스트 지시와 달리 전혀 파손되지 않은 온전한 선반 형태를 유지하고 있음.",
        "hard_violations": [
         "이전 샷에 없던 3단 나무 선반을 화면에 임의로 생성하여 로케이션 고정 지시를 위반함.",
         "'부서진 짚신전(Damaged)'이라는 텍스트 지시와 완전히 상반되게 물건이 가지런히 놓인 멀쩡한 선반을 그림."
        ],
        "physics": "안규홍은 바닥에 정상적으로 서 있으며, 바닥의 짚석과 그릇 파편들도 흙바닥 위에 안정적으로 지탱되어 있음."
       }
      ],
      "all_candidates_fail": false
     },
     "normalized": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 8,
        "verdict_ko": "안규홍의 뒤에서 내려다보는 카메라 앵글을 정확히 따랐으며, 이전 샷과 일치하는 구조의 짚신전이 무너진 채 방치된 모습을 충실하게 구현하여 프롬프트의 의도를 완벽히 살렸습니다."
       },
       {
        "label": "B",
        "score": 4,
        "verdict_ko": "캐릭터의 얼굴 묘사는 레퍼런스와 일치하나, 카메라 위치 지시(안규홍의 뒤에서)를 어겼으며, 이전 샷에 없던 멀쩡한 3단 선반을 만들어내어 '박살 난 짚신전'이라는 핵심 배경 조건을 위반했습니다."
       }
      ],
      "all_candidates_fail": false,
      "readings": [
       {
        "label": "A",
        "direction": "안규홍은 카메라를 등지고 파괴된 짚신전을 향해 서 있으며, 카메라는 지시문대로 그의 뒤에서 비스듬히 위에서 내려다보고 있음.",
        "built_space": "장터의 흙바닥. 왼쪽 배경에는 이전 샷과 구조가 일치하는 낮은 평상 형태의 짚신전이 무너져 있으며, 찢어진 짚석과 깨진 그릇 조각들이 바닥에 흩어져 있음. 오른쪽 배경에는 이전 샷의 커다란 검은 솥이 올바르게 위치함.",
        "entities": "안규홍은 뒷모습만 보이나 지정된 흰색 의상과 체형이 레퍼런스와 일치함. 짚신전은 지시문에 맞게 심하게 파손된 상태임. 바닥의 깨진 그릇 파편들도 확인됨.",
        "hard_violations": [],
        "physics": "안규홍은 두 발로 흙바닥을 단단히 딛고 서 있으며, 무너진 목재, 짚석, 그릇 파편 등 모든 배경 물체들이 중력에 맞게 바닥과 평상 위에 안정적으로 놓여 있음."
       },
       {
        "label": "B",
        "direction": "안규홍은 정면의 카메라를 멍하니 응시하고 있으며, 짚신전은 그의 뒤에 위치함. 카메라는 안규홍의 정면에 위치함.",
        "built_space": "장터 흙바닥. 안규홍 뒤로 짚신전이 위치하나, 이전 샷에 존재하지 않던 높고 온전한 3단 나무 선반이 추가되어 있으며 그 위에 짚신들이 가지런히 정렬되어 있음. 바닥에는 찢어진 짚석과 깨진 그릇 조각이 있음.",
        "entities": "안규홍의 얼굴과 의상은 레퍼런스 이미지와 완벽히 일치함. 하지만 짚신전은 '박살 난'이라는 텍스트 지시와 달리 전혀 파손되지 않은 온전한 선반 형태를 유지하고 있음.",
        "hard_violations": [
         "이전 샷에 없던 3단 나무 선반을 화면에 임의로 생성하여 로케이션 고정 지시를 위반함.",
         "'부서진 짚신전(Damaged)'이라는 텍스트 지시와 완전히 상반되게 물건이 가지런히 놓인 멀쩡한 선반을 그림."
        ],
        "physics": "안규홍은 바닥에 정상적으로 서 있으며, 바닥의 짚석과 그릇 파편들도 흙바닥 위에 안정적으로 지탱되어 있음."
       }
      ]
     },
     "ok": true
    },
    {
     "model": "gpt-high",
     "order": "reverse",
     "display_to_canonical": {
      "A": "B",
      "B": "A"
     },
     "raw": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 5,
        "verdict_ko": "텅 빈 장터의 외로운 전신과 참고 장소의 재질은 살렸지만, 인물을 앞에서 찍어 명시된 후방 사선 부감 구도를 놓쳤다."
       },
       {
        "label": "B",
        "score": 8,
        "verdict_ko": "안규홍 뒤쪽 위에서 부서진 짚신전을 바라보는 전신 구도가 핵심 지시에 맞지만, 인물과 점포의 화면 비중이 다소 크고 햇빛이 참고보다 강하다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "안규홍의 가슴과 얼굴이 카메라 쪽으로 드러나며, 시선은 화면 왼쪽 아래 흙바닥을 향한다. 짚신전은 그의 뒤쪽 왼편에 있어, 카메라가 그의 뒤에서 점포를 바라봐야 한다는 지시와 반대다. 무기나 손에 든 지향성 물체는 없다.",
        "built_space": "왼쪽에 목조 짚신전 하나가 있고, 진열대는 상·중·하 세 높이로 보인다. 세로 기둥들이 진열대를 지지하며, 오른쪽 낮은 단에는 원통형 통 하나, 그 뒤에는 큰 검은 대야 하나가 있다. 기대 놓은 짜임판과 찢어진 멍석, 흙길은 이전 장면의 주요 재료와 배치를 이어간다. 인물은 점포 바로 앞 흙길에 홀로 서지만, 점포가 화면 왼쪽의 상당 부분을 차지하여 상단의 제한된 배경이라는 요구에는 덜 맞는다. 반사상은 없다.",
        "entities": "인물은 한 명뿐이며 한국인 청년 남성 설정에 부합하는 외형이다. 검은 짧은 머리와 때 묻은 미색 저고리, 허리띠, 바지, 짚신이 보이고 지게는 없다. 머리는 이전 장면과 유사하지만 단단히 묶은 머리는 확인되지 않는다. 얼굴은 이전 장면 인물과 대체로 유사하다. 짚신 진열, 바닥에 놓인 찌그러진 짚신, 찢어진 멍석, 주변의 깨진 그릇 조각이 모두 보인다. 다른 사람이나 읽을 수 있는 글자는 없다.",
        "hard_violations": [],
        "physics": "두 짚신 바닥이 흙길에 닿아 체중을 지지하고, 양팔은 몸 옆으로 내려와 있다. 발을 조금 벌린 정지 자세로 실제 사람이 취할 수 있다. 멍석은 땅과 쓰러진 목재에 걸쳐 있고, 통과 대야는 각각 단과 받침에 놓여 있다. 짚신과 그릇 파편도 지면에 놓여 있으며, 근거 없이 떠 있는 물체는 없다."
       },
       {
        "label": "B",
        "direction": "안규홍은 등을 카메라에 보이고 부서진 짚신전 쪽을 향한다. 눈은 보이지 않아 정확한 시선은 확인할 수 없지만, 머리와 몸의 방향은 앞쪽 점포 잔해를 향한다. 카메라는 그의 뒤쪽 위에서 사선으로 내려다보므로 지정된 관찰 방향에 맞는다. 손에 든 물체는 없다.",
        "built_space": "상단 왼쪽에 목조 짚신전 하나와 세로 기둥들, 짚신을 놓은 긴 낮은 진열대가 있다. 뒤에는 짠 발이 있고, 오른쪽 낮은 단에 원통형 통 하나, 그 뒤 받침 위에 큰 검은 대야 하나가 보인다. 기대 놓은 짜임판과 찢어진 멍석이 이전 장소의 주요 특징을 유지한다. 안규홍은 점포 앞 흙길 중앙에 서고, 전신 주변으로 사람이 없는 흙바닥이 이어진다. 다만 점포와 잔해가 상단의 상당한 면적을 차지한다. 반사상은 없다.",
        "entities": "등을 보인 인물 한 명만 있으며, 청년 남성의 체격과 검은 머리, 때 묻은 미색 저고리와 바지, 천 허리띠, 짚신이 참고의 외형을 따른다. 얼굴과 민족적 세부 외형은 후면 구도라 판별할 수 없다. 단단히 묶은 머리는 뚜렷하지 않고 이전 장면처럼 짧게 보인다. 지게는 없으며, 짚신전의 짚신들과 바닥의 찌그러진 짚신, 찢어진 멍석, 깨진 그릇과 파편이 보인다. 다른 인물이나 읽을 수 있는 글자는 없다. 자연광의 낮이지만 참고보다 직사광과 그림자가 강하다.",
        "hard_violations": [],
        "physics": "양발이 흙바닥에 닿아 서 있는 몸을 지지하고, 팔과 손은 자연스럽게 몸 옆에 내려와 있다. 움직임 없이 서 있는 자세에 물리적 모순은 없다. 찢어진 멍석과 목재는 지면에 걸쳐 놓였고, 진열품은 선반에, 통과 대야는 단과 받침에 지지된다. 그릇과 파편은 땅에 놓여 있으며, 지지 없이 공중에 떠 있는 대상은 없다."
       }
      ],
      "all_candidates_fail": false
     },
     "normalized": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 5,
        "verdict_ko": "텅 빈 장터의 외로운 전신과 참고 장소의 재질은 살렸지만, 인물을 앞에서 찍어 명시된 후방 사선 부감 구도를 놓쳤다."
       },
       {
        "label": "A",
        "score": 8,
        "verdict_ko": "안규홍 뒤쪽 위에서 부서진 짚신전을 바라보는 전신 구도가 핵심 지시에 맞지만, 인물과 점포의 화면 비중이 다소 크고 햇빛이 참고보다 강하다."
       }
      ],
      "all_candidates_fail": false,
      "readings": [
       {
        "label": "B",
        "direction": "안규홍의 가슴과 얼굴이 카메라 쪽으로 드러나며, 시선은 화면 왼쪽 아래 흙바닥을 향한다. 짚신전은 그의 뒤쪽 왼편에 있어, 카메라가 그의 뒤에서 점포를 바라봐야 한다는 지시와 반대다. 무기나 손에 든 지향성 물체는 없다.",
        "built_space": "왼쪽에 목조 짚신전 하나가 있고, 진열대는 상·중·하 세 높이로 보인다. 세로 기둥들이 진열대를 지지하며, 오른쪽 낮은 단에는 원통형 통 하나, 그 뒤에는 큰 검은 대야 하나가 있다. 기대 놓은 짜임판과 찢어진 멍석, 흙길은 이전 장면의 주요 재료와 배치를 이어간다. 인물은 점포 바로 앞 흙길에 홀로 서지만, 점포가 화면 왼쪽의 상당 부분을 차지하여 상단의 제한된 배경이라는 요구에는 덜 맞는다. 반사상은 없다.",
        "entities": "인물은 한 명뿐이며 한국인 청년 남성 설정에 부합하는 외형이다. 검은 짧은 머리와 때 묻은 미색 저고리, 허리띠, 바지, 짚신이 보이고 지게는 없다. 머리는 이전 장면과 유사하지만 단단히 묶은 머리는 확인되지 않는다. 얼굴은 이전 장면 인물과 대체로 유사하다. 짚신 진열, 바닥에 놓인 찌그러진 짚신, 찢어진 멍석, 주변의 깨진 그릇 조각이 모두 보인다. 다른 사람이나 읽을 수 있는 글자는 없다.",
        "hard_violations": [],
        "physics": "두 짚신 바닥이 흙길에 닿아 체중을 지지하고, 양팔은 몸 옆으로 내려와 있다. 발을 조금 벌린 정지 자세로 실제 사람이 취할 수 있다. 멍석은 땅과 쓰러진 목재에 걸쳐 있고, 통과 대야는 각각 단과 받침에 놓여 있다. 짚신과 그릇 파편도 지면에 놓여 있으며, 근거 없이 떠 있는 물체는 없다."
       },
       {
        "label": "A",
        "direction": "안규홍은 등을 카메라에 보이고 부서진 짚신전 쪽을 향한다. 눈은 보이지 않아 정확한 시선은 확인할 수 없지만, 머리와 몸의 방향은 앞쪽 점포 잔해를 향한다. 카메라는 그의 뒤쪽 위에서 사선으로 내려다보므로 지정된 관찰 방향에 맞는다. 손에 든 물체는 없다.",
        "built_space": "상단 왼쪽에 목조 짚신전 하나와 세로 기둥들, 짚신을 놓은 긴 낮은 진열대가 있다. 뒤에는 짠 발이 있고, 오른쪽 낮은 단에 원통형 통 하나, 그 뒤 받침 위에 큰 검은 대야 하나가 보인다. 기대 놓은 짜임판과 찢어진 멍석이 이전 장소의 주요 특징을 유지한다. 안규홍은 점포 앞 흙길 중앙에 서고, 전신 주변으로 사람이 없는 흙바닥이 이어진다. 다만 점포와 잔해가 상단의 상당한 면적을 차지한다. 반사상은 없다.",
        "entities": "등을 보인 인물 한 명만 있으며, 청년 남성의 체격과 검은 머리, 때 묻은 미색 저고리와 바지, 천 허리띠, 짚신이 참고의 외형을 따른다. 얼굴과 민족적 세부 외형은 후면 구도라 판별할 수 없다. 단단히 묶은 머리는 뚜렷하지 않고 이전 장면처럼 짧게 보인다. 지게는 없으며, 짚신전의 짚신들과 바닥의 찌그러진 짚신, 찢어진 멍석, 깨진 그릇과 파편이 보인다. 다른 인물이나 읽을 수 있는 글자는 없다. 자연광의 낮이지만 참고보다 직사광과 그림자가 강하다.",
        "hard_violations": [],
        "physics": "양발이 흙바닥에 닿아 서 있는 몸을 지지하고, 팔과 손은 자연스럽게 몸 옆에 내려와 있다. 움직임 없이 서 있는 자세에 물리적 모순은 없다. 찢어진 멍석과 목재는 지면에 걸쳐 놓였고, 진열품은 선반에, 통과 대야는 단과 받침에 지지된다. 그릇과 파편은 땅에 놓여 있으며, 지지 없이 공중에 떠 있는 대상은 없다."
       }
      ]
     },
     "ok": true
    }
   ],
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "slot_winner_match": true,
   "slot_winner": {
    "gemini-pro": "A",
    "gpt-high": "A"
   },
   "route": "cross_slot_agree"
  },
  "dual": {
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "normalized": {
    "A": 2.0,
    "B": 1.125
   },
   "adjusted": {
    "A": 2.0,
    "B": 0.875
   },
   "violations": {
    "B": [
     "[gemini-pro] 이전 샷에 없던 3단 나무 선반을 화면에 임의로 생성하여 로케이션 고정 지시를 위반함.",
     "[gemini-pro] '부서진 짚신전(Damaged)'이라는 텍스트 지시와 완전히 상반되게 물건이 가지런히 놓인 멀쩡한 선반을 그림."
    ]
   },
   "per_model_winner": {
    "gemini-pro": "A",
    "gpt-high": "A"
   },
   "agreed": true
  },
  "totals": {
   "A": 2000,
   "B": 875
  },
  "selected": "A",
  "ranking": [
   "A",
   "B"
  ],
  "verdicts": [
   {
    "label": "A",
    "score": 2000,
    "verdict_ko": "안규홍의 뒤에서 내려다보는 카메라 앵글을 정확히 따랐으며, 이전 샷과 일치하는 구조의 짚신전이 무너진 채 방치된 모습을 충실하게 구현하여 프롬프트의 의도를 완벽히 살렸습니다."
   },
   {
    "label": "B",
    "score": 875,
    "verdict_ko": "캐릭터의 얼굴 묘사는 레퍼런스와 일치하나, 카메라 위치 지시(안규홍의 뒤에서)를 어겼으며, 이전 샷에 없던 멀쩡한 3단 선반을 만들어내어 '박살 난 짚신전'이라는 핵심 배경 조건을 위반했습니다.  ★위반: [gemini-pro] 이전 샷에 없던 3단 나무 선반을 화면에 임의로 생성하여 로케이션 고정 지시를 위반함. / [gemini-pro] '부서진 짚신전(Damaged)'이라는 텍스트 지시와 완전히 상반되게 물건이 가지런히 놓인 멀쩡한 선반을 그림."
   }
  ],
  "refs": [
   {
    "label": "PREVIOUS SHOT STILL — a visually related earlier shot of this same place: the location's look, materials, fixed features and lighting mood are LOCKED to this photo; never copy its camera framing. The clothing, hair and overall look of 안규홍 — who appear both in that photo and in this shot — are LOCKED to that photo. Anyone else visible in that photo is NOT in this shot: never carry their face, body or clothing onto anyone here. If this photo shows a character who cannot move (dead or unconscious), that character's exact pose, position and orientation are ALSO LOCKED — treat their body as a fixed prop of the set that only the camera moves around.",
    "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/db00517b-165b-4005-a287-f28936b59b95/images/272bd7af-9294-4490-98c6-88a40e9deacc/scene/recipe/S10sh13_sel.png",
    "asset_id": "76c32d45-3ad2-49bf-a826-fae42cef8fc1",
    "role": "prev_still"
   },
   {
    "label": "CHARACTER REFERENCE — 안규홍: the exact person appearing in this shot; match face, hair and build exactly.",
    "path": "<bytes:1028384>",
    "asset_id": "3b5894db-ce55-4f34-b966-5e725c3aa952",
    "role": "character_ref"
   }
  ],
  "critique_skipped": true,
  "shot_run_produced": false,
  "shot_run_uid": "06aa140c-51a4-7b23-9afc-b6f7d5fc7b6f",
  "ref_mode": "prev+엔티티",
  "share_plan": {
   "ref_plan": "prev",
   "prev_anchor_tag": "S10sh13"
  },
  "lane_policy": "ab_select_bypass:prev"
 },
 "S10sh15::cine": {
  "applied": true,
  "attempted_at": "2026-09-09T01:45:54.578952+00:00",
  "fingerprint": "a221788ecd58df985032a6a5f2216419e256b4c143446956aac3edfb8dbfffa2",
  "fingerprint_version": 2,
  "provider": "mai",
  "endpoint": "openrouter/chat-completions",
  "model": "microsoft/mai-image-2.6",
  "pack": "24.202608252115",
  "source_file": "S10sh15_sel.png",
  "source_sha256": "a16265b38f513a82badbd1aacce2933918f2b3019278d9d5153bf76b47202dcc",
  "file": "S10sh15_cine.png",
  "staged_sha256": "625f78942ff7b61ac56ba4170e47937de9e26ce124c49f23306f39f9b0d71ebf",
  "latency_ms": 25134
 },
 "S11sh4::signage": {
  "fp": "f06486d42770a986",
  "inscriptions": [],
  "cues": [],
  "dropped": []
 },
 "era_assess::5ff7c99c44cbb155": {
  "subjects": [
   {
    "subject_native": "20세기 초 전라도 민가 온돌방 내부",
    "search_terms_native": [
     "구한말 초가집 방 내부",
     "일제강점기 농가 온돌방",
     "전라도 민가 온돌방 옛사진",
     "전통 초가집 방 내부"
    ],
    "language_lock_native": "반드시 한국어로만 검색어를 작성하고 다른 언어를 섞거나 번역하지 마십시오.",
    "reason_ko": "현대 개량 한옥이나 일본식 목조 가옥 구조와 혼동되어 당대 전라도 서민 가옥의 소박한 흙벽과 창호문 구조가 왜곡되기 쉽습니다."
   }
  ],
  "subject_text": "안규홍의 집 방\n바닥에 이불을 펴는 소박한 온돌방. 벽 가까이에 침구가 놓이고, 바깥으로 연결되는 방문이 있다.",
  "identity": "canonical",
  "scope_id": "L04",
  "scope_role": "location_interior",
  "scope_sha": "013e56667e3d283d"
 },
 "era_fail::ca5d321de0a6d8f1": {
  "stage": "research",
  "subject": "20세기 초 전라도 민가 온돌방 내부",
  "terms": [
   "구한말 초가집 방 내부",
   "일제강점기 농가 온돌방",
   "전라도 민가 온돌방 옛사진",
   "전통 초가집 방 내부"
  ],
  "status": "no_usable",
  "queries": [
   [
    "전라도 민가 온돌방 내부 옛사진",
    "국립민속박물관 초가집 방 내부 일제강점기"
   ],
   [
    "\"유리건판\" \"방 내부\"",
    "\"초가\" \"방안\" \"사진\"",
    "\"전라도\" \"민가\" \"내부\" \"사진\""
   ],
   [
    "\"온돌방\" \"옛사진\"",
    "\"농가 내부\" \"사진\"",
    "\"민가 내부\" \"유리건판\"",
    "\"초가집 내부\" \"사진\""
   ]
  ],
  "candidate_urls": [
   "https://minio.nculture.org/amsweb-opt/multimedia_assets/228/72628/66822/c/%EC%A0%84%EB%B6%81-%EB%B6%80%EC%95%88%EA%B5%B0-%EB%B6%80%EC%95%88-%EA%B9%80%EC%83%81%EB%A7%8C%EA%B0%80%EC%98%A5-%EC%95%88%EC%B1%84-%EB%82%B4%EB%B6%80-%281%29-medium-size.jpg",
   "https://ims.ev.mu/medium_shutterstock_2246685559_38e6a84a1c/medium_shutterstock_2246685559_38e6a84a1c.jpg"
  ],
  "coarse": {
   "eligible": [],
   "chosen_index": 0,
   "reason": "종류·보임·기준을 다 만족하는 후보가 없다 — 이 라운드에선 안 고른다",
   "single_judge": true,
   "rejected_judges": {}
  },
  "verdicts": [
   {
    "index": 1,
    "object_type_match": "no",
    "visible": true,
    "criteria_match": "unsure",
    "similarity": 20
   },
   {
    "index": 2,
    "object_type_match": "no",
    "visible": true,
    "criteria_match": "unsure",
    "similarity": 30
   }
  ],
  "chosen_reason_ko": "",
  "attempts": 5
 },
 "S11sh4::bgfirst_bg": {
  "input_fingerprint": "489365d4a32c4d9d",
  "prompt": "Create the EMPTY BACKGROUND PLATE for one film shot — NO PEOPLE, no figures, no body parts, no sketch lines, no arrows anywhere.\n\"Empty\" means no people only: KEEP the location's inherent occupants and stock that define the place — animals in an animal shelter, pen or farm, goods in a market, moored boats in a harbour — unless the shot text explicitly removes them.\nThe FIRST attached image is a thin-line storyboard sketch: use ONLY its camera angle, horizon, perspective and the placement/size of buildings and set masses — ignore the sketched people and arrows entirely. The SECOND attached image (LOCATION PHOTOGRAPH) is the real place: take its architecture, materials, signage and fixed features, and RE-PROJECT them into the sketch's camera. If the photograph's camera differs from the sketch's, the sketch's camera wins.\nHUMAN-SCALE CALIBRATION: derive every structure's true size from human-scale elements — a door ≈ 2m, a window ≈ 1–1.5m wide, one storey ≈ 2.5–3m; never inflate a small structure or shrink a large one.\n\nSHOT TEXT this background must serve (Korean): 안규홍의 이마가 방바닥에 납작하게 닿은 채 엎드려 있는 전신.\n\nLOCATION (lock): The bare bowing space on the floor inside a modest home's room, washed in bluish dawn light.\n\nTIME OF DAY (lock): dawn, bluish natural light.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: wide shot\n- KEY BACKGROUND ELEMENTS: 방바닥 (안규홍's forehead rests against it during the deep bow); used as The visible contact beneath his head makes the depth and duration of the bow legible; 방 벽 (Supports the seated mother); used as Anchors her position beyond the bowed figure; 어머니의 이불 (Wrapped around her waist); used as A limited shape around her seated body preserves the domestic context; 작은 봇짐 (Packed for 안규홍's departure); used as Remains small beside the bow, connecting the farewell to the imminent journey.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: The screenplay's bluish dawn light remains soft and restrained, maintaining natural contrast and quiet emotional gravity.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nRender ONE photorealistic empty location photograph, 16:9, that this shot can be staged inside later. No readable writing anywhere: surfaces that would carry writing may be present, but stage any wording out of legibility — an oblique angle, distance, shallow focus. No captions, watermarks or overlay text.",
  "effective_prompt": "Create the EMPTY BACKGROUND PLATE for one film shot — NO PEOPLE, no figures, no body parts, no sketch lines, no arrows anywhere.\n\"Empty\" means no people only: KEEP the location's inherent occupants and stock that define the place — animals in an animal shelter, pen or farm, goods in a market, moored boats in a harbour — unless the shot text explicitly removes them.\nThe FIRST attached image is a thin-line storyboard sketch: use ONLY its camera angle, horizon, perspective and the placement/size of buildings and set masses — ignore the sketched people and arrows entirely. The SECOND attached image (LOCATION PHOTOGRAPH) is the real place: take its architecture, materials, signage and fixed features, and RE-PROJECT them into the sketch's camera. If the photograph's camera differs from the sketch's, the sketch's camera wins.\nHUMAN-SCALE CALIBRATION: derive every structure's true size from human-scale elements — a door ≈ 2m, a window ≈ 1–1.5m wide, one storey ≈ 2.5–3m; never inflate a small structure or shrink a large one.\n\nSHOT TEXT this background must serve (Korean): 안규홍의 이마가 방바닥에 납작하게 닿은 채 엎드려 있는 전신.\n\nLOCATION (lock): The bare bowing space on the floor inside a modest home's room, washed in bluish dawn light.\n\nTIME OF DAY (lock): dawn, bluish natural light.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: wide shot\n- KEY BACKGROUND ELEMENTS: 방바닥 (안규홍's forehead rests against it during the deep bow); used as The visible contact beneath his head makes the depth and duration of the bow legible; 방 벽 (Supports the seated mother); used as Anchors her position beyond the bowed figure; 어머니의 이불 (Wrapped around her waist); used as A limited shape around her seated body preserves the domestic context; 작은 봇짐 (Packed for 안규홍's departure); used as Remains small beside the bow, connecting the farewell to the imminent journey.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: The screenplay's bluish dawn light remains soft and restrained, maintaining natural contrast and quiet emotional gravity.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nRender ONE photorealistic empty location photograph, 16:9, that this shot can be staged inside later. No readable writing anywhere: surfaces that would carry writing may be present, but stage any wording out of legibility — an oblique angle, distance, shallow focus. No captions, watermarks or overlay text.",
  "bg_path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/db00517b-165b-4005-a287-f28936b59b95/images/272bd7af-9294-4490-98c6-88a40e9deacc/scene/recipe/S11sh4__bgfirst_bg.png",
  "asset_id": "4390caec-8bb9-4035-bcc2-6749e1eaca05",
  "input_asset_ids": [
   "336d2b17-9092-4c9b-946a-d0463a1330f3",
   "8aaff0f1-df32-4a26-92ab-4f112c48d1af"
  ]
 },
 "S11sh4": {
  "input_fingerprint": "9d2be09e8cafcbfa",
  "prompt": "Create ONE FINAL photorealistic live-action film still of the moment below — early 20th-century Korea, Jeolla region; all people are Korean unless explicitly stated otherwise. TIME OF DAY (lock): dawn, bluish natural light.\n\nSHOT TEXT (authoritative, Korean): 안규홍의 이마가 방바닥에 납작하게 닿은 채 엎드려 있는 전신.\n\nLOCATION (lock): The bare bowing space on the floor inside a modest home's room, washed in bluish dawn light. The shot takes place here — the attached LOCATION PHOTOGRAPH shows the exact spot.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: wide shot\n- KEY BACKGROUND ELEMENTS: 방바닥 (안규홍's forehead rests against it during the deep bow); used as The visible contact beneath his head makes the depth and duration of the bow legible; 방 벽 (Supports the seated mother); used as Anchors her position beyond the bowed figure; 어머니의 이불 (Wrapped around her waist); used as A limited shape around her seated body preserves the domestic context; 작은 봇짐 (Packed for 안규홍's departure); used as Remains small beside the bow, connecting the farewell to the imminent journey.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: The screenplay's bluish dawn light remains soft and restrained, maintaining natural contrast and quiet emotional gravity.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): Bluish dawn light fills the house; the small travel bundle has been tied. 안규홍: He is fully prostrate with his forehead against the room floor.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 안규홍 (한국인 남성, 청년 나이대, 검은 머리, 단단히 묶은 머리) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.",
  "roll_prompts": {
   "A": "Stage the shot. The FIRST attached image (SHOT BACKGROUND) is the finished empty background of this shot — keep it EXACTLY: its camera, perspective, architecture, lighting and every fixed feature stay untouched. The SECOND attached image (LAYOUT SKETCH) tells you ONLY where the people go: each sketched person's position, screen size, pose and the gaze/motion arrows. Ignore the sketch's background lines. The CHARACTER REFERENCE photographs show the real people.\nPlace the real people into the background at exactly the sketched positions, sizes and poses, following the arrow directions. No sketch lines or arrows may remain.\n\nCreate ONE FINAL photorealistic live-action film still of the moment below — early 20th-century Korea, Jeolla region; all people are Korean unless explicitly stated otherwise. TIME OF DAY (lock): dawn, bluish natural light.\n\nSHOT TEXT (authoritative, Korean): 안규홍의 이마가 방바닥에 납작하게 닿은 채 엎드려 있는 전신.\n\nLOCATION (lock): The bare bowing space on the floor inside a modest home's room, washed in bluish dawn light. The shot takes place here — the FIRST attached image (SHOT BACKGROUND) is this exact place, already built: its ground, structures, horizon, materials and lighting are the finished truth of this location and must not be redesigned or replaced. No location photograph is attached — read the place from that image alone, and add no scenery, structure, vehicle or fixture that it does not already show. This lock governs the place only; the figures in the shot follow the staging and pose instructions.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: wide shot\n- KEY BACKGROUND ELEMENTS: 방바닥 (안규홍's forehead rests against it during the deep bow); used as The visible contact beneath his head makes the depth and duration of the bow legible; 방 벽 (Supports the seated mother); used as Anchors her position beyond the bowed figure; 어머니의 이불 (Wrapped around her waist); used as A limited shape around her seated body preserves the domestic context; 작은 봇짐 (Packed for 안규홍's departure); used as Remains small beside the bow, connecting the farewell to the imminent journey.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: The screenplay's bluish dawn light remains soft and restrained, maintaining natural contrast and quiet emotional gravity.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): Bluish dawn light fills the house; the small travel bundle has been tied. 안규홍: He is fully prostrate with his forehead against the room floor.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 안규홍 (한국인 남성, 청년 나이대, 검은 머리, 단단히 묶은 머리) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.",
   "B": "Create ONE FINAL photorealistic live-action film still of the moment below — early 20th-century Korea, Jeolla region; all people are Korean unless explicitly stated otherwise. TIME OF DAY (lock): dawn, bluish natural light.\n\nSHOT TEXT (authoritative, Korean): 안규홍의 이마가 방바닥에 납작하게 닿은 채 엎드려 있는 전신.\n\nLOCATION (lock): The bare bowing space on the floor inside a modest home's room, washed in bluish dawn light. The shot takes place here — the attached LOCATION PHOTOGRAPH shows the exact spot.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: wide shot\n- KEY BACKGROUND ELEMENTS: 방바닥 (안규홍's forehead rests against it during the deep bow); used as The visible contact beneath his head makes the depth and duration of the bow legible; 방 벽 (Supports the seated mother); used as Anchors her position beyond the bowed figure; 어머니의 이불 (Wrapped around her waist); used as A limited shape around her seated body preserves the domestic context; 작은 봇짐 (Packed for 안규홍's departure); used as Remains small beside the bow, connecting the farewell to the imminent journey.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: The screenplay's bluish dawn light remains soft and restrained, maintaining natural contrast and quiet emotional gravity.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): Bluish dawn light fills the house; the small travel bundle has been tied. 안규홍: He is fully prostrate with his forehead against the room floor.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 안규홍 (한국인 남성, 청년 나이대, 검은 머리, 단단히 묶은 머리) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.\n\nCOMPOSITION VARIATION (this candidate only): keep the place, the people and the moment exactly as specified, and keep from the CAMERA & FRAME contract its subject scale, screen placement and key background elements exactly. For this candidate only, its camera angle and height are deliberately left open: choose a camera position distinctly different from the obvious one, as a film director picking a second setup on the same blocking. Stay true to any explicit framing the SHOT TEXT itself dictates."
  },
  "roll_refs": {
   "A": [
    {
     "label": "SHOT BACKGROUND",
     "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/db00517b-165b-4005-a287-f28936b59b95/images/272bd7af-9294-4490-98c6-88a40e9deacc/scene/recipe/S11sh4__bgfirst_bg.png",
     "asset_id": "4390caec-8bb9-4035-bcc2-6749e1eaca05",
     "role": "bgfirst_bg"
    },
    {
     "label": "LAYOUT SKETCH (people placement only)",
     "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/db00517b-165b-4005-a287-f28936b59b95/images/272bd7af-9294-4490-98c6-88a40e9deacc/conti/conti_S11sh4.png",
     "asset_id": "336d2b17-9092-4c9b-946a-d0463a1330f3",
     "role": "conti_light"
    },
    {
     "label": "CHARACTER REFERENCE — 안규홍: the exact person appearing in this shot; match face, hair and build exactly.",
     "path": "<bytes:1028384>",
     "asset_id": "3b5894db-ce55-4f34-b966-5e725c3aa952",
     "role": "character_ref"
    }
   ],
   "B": [
    {
     "label": "LOCATION PHOTOGRAPH — the exact place of this shot: its architecture, materials, fixed features and lighting mood are spatial truth; stage the moment inside this place. Never copy its camera framing.",
     "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/db00517b-165b-4005-a287-f28936b59b95/episodes/272bd7af-9294-4490-98c6-88a40e9deacc/images/background_chain/L04B02.png",
     "asset_id": "8aaff0f1-df32-4a26-92ab-4f112c48d1af",
     "role": "location_plate"
    },
    {
     "label": "CHARACTER REFERENCE — 안규홍: the exact person appearing in this shot; match face, hair and build exactly.",
     "path": "<bytes:1028384>",
     "asset_id": "3b5894db-ce55-4f34-b966-5e725c3aa952",
     "role": "character_ref"
    }
   ]
  },
  "shot_run_spend_attempt_count": 1,
  "judge_flip_skipped": "judge_fn_owns_order",
  "readings": [
   {
    "label": "A",
    "direction": "안규홍은 방바닥을 향해 이마가 닿을 듯이 엎드려 절하고 있으며, 어머니는 엎드린 인물 쪽으로 시선을 약간 숙이고 있다.",
    "built_space": "레퍼런스와 동일한 흙벽, 바닥, 나무 문틀, 장롱이 올바른 위치에 배치되어 있으며, 인물들이 이 공간의 구조에 알맞게 자리를 잡고 있다.",
    "entities": "레퍼런스의 복장과 묶은 머리를 한 안규홍, 벽에 기댄 채 이불을 덮고 앉아있는 어머니, 그리고 바닥에 놓인 작은 봇짐이 모두 정확하게 표현되었다.",
    "hard_violations": [
     "[gpt-high] 숏 텍스트와 인물 제한이 허용하지 않은 여성을 추가했다."
    ],
    "physics": "안규홍은 바닥에 두 무릎과 손을 짚고 체중을 지탱하며 안정적으로 엎드려 있고, 어머니 역시 바닥에 자연스럽게 앉아 자세를 유지하고 있다."
   },
   {
    "label": "B",
    "direction": "바닥을 향해 엎드린 머리 외에도, 허리 쪽에서 뒤를 돌아보는 방향으로 또 다른 머리가 튀어나와 있다.",
    "built_space": "방 내부 구조와 주요 가구들의 배치는 레퍼런스 사진의 형태를 대체로 잘 반영하고 있다.",
    "entities": "안규홍의 복장을 입고 있으나 한 몸에 머리가 두 개 달린 기괴한 형상의 인물이 생성되었으며, 앉아있는 인물과 봇짐도 존재한다.",
    "hard_violations": [
     "[gemini-pro] physically impossible anatomy (엎드린 인물의 엉덩이/허리 부근에 두 번째 머리가 달려 있음)",
     "[gpt-high] 절하는 남성의 머리가 두 개이고 몸통·팔다리가 서로 겹쳐 융합된 형태여서, 인물 중복 및 불가능한 인체 구조에 해당한다.",
     "[gpt-high] 숏 텍스트와 인물 제한이 허용하지 않은 여성을 추가했다."
    ],
    "physics": "엎드린 인물의 신체가 물리적으로 불가능하게 결합되어 있으며, 한 몸에 두 개의 머리와 비정상적인 팔다리 구조가 나타난다."
   }
  ],
  "cross_model_order": {
   "policy": "cross_model_order_v2_slot0_fwd_slot1_rev_parallel2",
   "slots": [
    {
     "model": "gemini-pro",
     "order": "forward",
     "display_to_canonical": {
      "A": "A",
      "B": "B"
     },
     "raw": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 9,
        "verdict_ko": "샷 텍스트가 요구하는 안규홍의 엎드린 자세와 배경 요소(앉아있는 어머니, 봇짐)를 지정된 방 구조 안에 사실적이고 자연스럽게 구현했습니다."
       },
       {
        "label": "B",
        "score": 0,
        "verdict_ko": "엎드린 안규홍의 몸에 머리가 두 개 생성되는 등 물리적으로 불가능한 기괴한 신체 구조가 나타나 완전히 실격입니다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "안규홍은 방바닥을 향해 이마가 닿을 듯이 엎드려 절하고 있으며, 어머니는 엎드린 인물 쪽으로 시선을 약간 숙이고 있다.",
        "built_space": "레퍼런스와 동일한 흙벽, 바닥, 나무 문틀, 장롱이 올바른 위치에 배치되어 있으며, 인물들이 이 공간의 구조에 알맞게 자리를 잡고 있다.",
        "entities": "레퍼런스의 복장과 묶은 머리를 한 안규홍, 벽에 기댄 채 이불을 덮고 앉아있는 어머니, 그리고 바닥에 놓인 작은 봇짐이 모두 정확하게 표현되었다.",
        "hard_violations": [],
        "physics": "안규홍은 바닥에 두 무릎과 손을 짚고 체중을 지탱하며 안정적으로 엎드려 있고, 어머니 역시 바닥에 자연스럽게 앉아 자세를 유지하고 있다."
       },
       {
        "label": "B",
        "direction": "바닥을 향해 엎드린 머리 외에도, 허리 쪽에서 뒤를 돌아보는 방향으로 또 다른 머리가 튀어나와 있다.",
        "built_space": "방 내부 구조와 주요 가구들의 배치는 레퍼런스 사진의 형태를 대체로 잘 반영하고 있다.",
        "entities": "안규홍의 복장을 입고 있으나 한 몸에 머리가 두 개 달린 기괴한 형상의 인물이 생성되었으며, 앉아있는 인물과 봇짐도 존재한다.",
        "hard_violations": [
         "physically impossible anatomy (엎드린 인물의 엉덩이/허리 부근에 두 번째 머리가 달려 있음)"
        ],
        "physics": "엎드린 인물의 신체가 물리적으로 불가능하게 결합되어 있으며, 한 몸에 두 개의 머리와 비정상적인 팔다리 구조가 나타난다."
       }
      ],
      "all_candidates_fail": false
     },
     "normalized": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 9,
        "verdict_ko": "샷 텍스트가 요구하는 안규홍의 엎드린 자세와 배경 요소(앉아있는 어머니, 봇짐)를 지정된 방 구조 안에 사실적이고 자연스럽게 구현했습니다."
       },
       {
        "label": "B",
        "score": 0,
        "verdict_ko": "엎드린 안규홍의 몸에 머리가 두 개 생성되는 등 물리적으로 불가능한 기괴한 신체 구조가 나타나 완전히 실격입니다."
       }
      ],
      "all_candidates_fail": false,
      "readings": [
       {
        "label": "A",
        "direction": "안규홍은 방바닥을 향해 이마가 닿을 듯이 엎드려 절하고 있으며, 어머니는 엎드린 인물 쪽으로 시선을 약간 숙이고 있다.",
        "built_space": "레퍼런스와 동일한 흙벽, 바닥, 나무 문틀, 장롱이 올바른 위치에 배치되어 있으며, 인물들이 이 공간의 구조에 알맞게 자리를 잡고 있다.",
        "entities": "레퍼런스의 복장과 묶은 머리를 한 안규홍, 벽에 기댄 채 이불을 덮고 앉아있는 어머니, 그리고 바닥에 놓인 작은 봇짐이 모두 정확하게 표현되었다.",
        "hard_violations": [],
        "physics": "안규홍은 바닥에 두 무릎과 손을 짚고 체중을 지탱하며 안정적으로 엎드려 있고, 어머니 역시 바닥에 자연스럽게 앉아 자세를 유지하고 있다."
       },
       {
        "label": "B",
        "direction": "바닥을 향해 엎드린 머리 외에도, 허리 쪽에서 뒤를 돌아보는 방향으로 또 다른 머리가 튀어나와 있다.",
        "built_space": "방 내부 구조와 주요 가구들의 배치는 레퍼런스 사진의 형태를 대체로 잘 반영하고 있다.",
        "entities": "안규홍의 복장을 입고 있으나 한 몸에 머리가 두 개 달린 기괴한 형상의 인물이 생성되었으며, 앉아있는 인물과 봇짐도 존재한다.",
        "hard_violations": [
         "physically impossible anatomy (엎드린 인물의 엉덩이/허리 부근에 두 번째 머리가 달려 있음)"
        ],
        "physics": "엎드린 인물의 신체가 물리적으로 불가능하게 결합되어 있으며, 한 몸에 두 개의 머리와 비정상적인 팔다리 구조가 나타난다."
       }
      ]
     },
     "ok": true
    },
    {
     "model": "gpt-high",
     "order": "reverse",
     "display_to_canonical": {
      "A": "B",
      "B": "A"
     },
     "raw": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 1,
        "verdict_ko": "방과 새벽빛은 유사하지만, 절하는 남자의 머리와 신체가 중복·융합되어 전신 엎드림을 정상적인 인체로 구현하지 못했고 허용되지 않은 여성도 등장한다."
       },
       {
        "label": "B",
        "score": 4,
        "verdict_ko": "전신 와이드숏에서 바닥에 이마를 댄 절과 신체 지지는 자연스럽지만, 안규홍만 허용하는 인물 제한과 달리 어머니를 추가해 최종 사용에는 부적합하다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "전경에는 바닥을 향해 숙인 남성 머리가 두 개 보이며 하나는 왼쪽, 다른 하나는 오른쪽 아래를 향한다. 두 머리 모두 눈은 바닥 쪽으로 가려져 있다. 뒤쪽 여성은 전경의 남성을 내려다본다. 여성 쪽으로 한 사람이 절하는 명료한 방향 관계가 아니다.",
        "built_space": "오른쪽 출입구 하나, 뒤쪽 높은 장 하나와 그 위 바구니 하나, 낮은 서랍장 하나, 왼쪽 낮은 함 하나, 왼쪽 침구 한 벌이 보인다. 흙빛 바닥과 목재 골조의 벽은 장소 사진과 대체로 일치한다. 여성은 왼쪽 침구 위에 앉아 있으며 등은 벽에 닿지 않아 배경 지시의 벽에 기대앉은 위치와 다르다. 작은 봇짐 하나는 오른쪽 문 옆 바닥에 있다.",
        "entities": "안규홍에 해당하는 밝은색 저고리·바지, 허리끈, 짚신, 검은 묶음머리가 보이지만 남성 머리가 두 개여서 단일 인물의 정체성과 전신이 성립하지 않는다. 얼굴은 숙여져 정확한 얼굴 일치는 판단하기 어렵다. 한국인 중년 여성으로 보이는 인물이 이불을 하체에 두르고 추가되어 있다. 배경 설명에는 어머니가 언급되지만, 우선하는 숏 텍스트와 인물 제한은 안규홍만 허용한다. 묶인 작은 봇짐과 방바닥은 있으며 읽을 수 있는 글자는 없다.",
        "hard_violations": [
         "절하는 남성의 머리가 두 개이고 몸통·팔다리가 서로 겹쳐 융합된 형태여서, 인물 중복 및 불가능한 인체 구조에 해당한다.",
         "숏 텍스트와 인물 제한이 허용하지 않은 여성을 추가했다."
        ],
        "physics": "손바닥과 팔뚝, 접힌 다리 및 짚신은 바닥에 닿아 있어 공중에 떠 있는 몸은 아니다. 그러나 두 머리가 하나의 겹친 몸통 덩어리에 연결되는 구조는 한 사람이 무릎을 꿇고 절하는 동작으로 만들 수 없다. 여성은 침구와 바닥에 앉아 지지되고, 봇짐과 가구도 바닥에 놓여 있다."
       },
       {
        "label": "B",
        "direction": "남성은 화면 오른쪽의 여성 방향으로 몸을 접어 절하고, 얼굴과 이마는 바로 아래 방바닥을 향한다. 이마 부근의 바닥 접촉이 보인다. 여성은 아래쪽의 절하는 남성을 바라본다.",
        "built_space": "오른쪽 출입구 하나, 뒤쪽 높은 장 하나와 상단 바구니 하나, 낮은 서랍장 하나, 왼쪽 낮은 함 하나, 왼쪽 바닥 침구 한 벌이 보인다. 장소 사진의 벽 재질, 목재 골조, 바닥과 주요 가구 배치가 잘 유지된다. 여성은 오른쪽 벽에 등을 대고 바닥에 앉아 있고, 봇짐 하나가 여성 옆 문 가까이에 놓여 있다. 남성의 전신과 머리 아래 바닥을 포함하는 와이드숏이다.",
        "entities": "남성 한 명은 청년으로 보이며 검은 묶음머리, 낡은 밝은색 저고리와 바지, 허리끈 및 짚신이 인물 참조와 대체로 맞는다. 얼굴이 바닥을 향하므로 정확한 얼굴 일치는 확인할 수 없다. 회색 머리의 한국인 노년 여성으로 보이는 인물이 허리와 다리에 이불을 두르고 등장한다. 어머니의 배경 묘사에는 맞지만, 숏 텍스트가 정하는 안규홍 단독 등장 제한에는 어긋난다. 작은 묶인 봇짐이 있으며 읽을 수 있는 글자는 없다.",
        "hard_violations": [
         "숏 텍스트와 인물 제한이 허용하지 않은 여성을 추가했다."
        ],
        "physics": "남성은 무릎과 정강이를 바닥에 접고 앉아 상체를 낮췄으며, 팔뚝과 손 및 이마 부근이 바닥에 닿는다. 체중을 지탱하는 접점이 있고 깊은 절로 만들 수 있는 자세다. 여성의 엉덩이와 접힌 다리는 바닥에 지지되며 등 뒤에는 벽이 있다. 이불은 무릎 위에 내려앉고 봇짐과 가구는 바닥에 놓여 있어 지지 없는 부유는 없다."
       }
      ],
      "all_candidates_fail": true
     },
     "normalized": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 1,
        "verdict_ko": "방과 새벽빛은 유사하지만, 절하는 남자의 머리와 신체가 중복·융합되어 전신 엎드림을 정상적인 인체로 구현하지 못했고 허용되지 않은 여성도 등장한다."
       },
       {
        "label": "A",
        "score": 4,
        "verdict_ko": "전신 와이드숏에서 바닥에 이마를 댄 절과 신체 지지는 자연스럽지만, 안규홍만 허용하는 인물 제한과 달리 어머니를 추가해 최종 사용에는 부적합하다."
       }
      ],
      "all_candidates_fail": true,
      "readings": [
       {
        "label": "B",
        "direction": "전경에는 바닥을 향해 숙인 남성 머리가 두 개 보이며 하나는 왼쪽, 다른 하나는 오른쪽 아래를 향한다. 두 머리 모두 눈은 바닥 쪽으로 가려져 있다. 뒤쪽 여성은 전경의 남성을 내려다본다. 여성 쪽으로 한 사람이 절하는 명료한 방향 관계가 아니다.",
        "built_space": "오른쪽 출입구 하나, 뒤쪽 높은 장 하나와 그 위 바구니 하나, 낮은 서랍장 하나, 왼쪽 낮은 함 하나, 왼쪽 침구 한 벌이 보인다. 흙빛 바닥과 목재 골조의 벽은 장소 사진과 대체로 일치한다. 여성은 왼쪽 침구 위에 앉아 있으며 등은 벽에 닿지 않아 배경 지시의 벽에 기대앉은 위치와 다르다. 작은 봇짐 하나는 오른쪽 문 옆 바닥에 있다.",
        "entities": "안규홍에 해당하는 밝은색 저고리·바지, 허리끈, 짚신, 검은 묶음머리가 보이지만 남성 머리가 두 개여서 단일 인물의 정체성과 전신이 성립하지 않는다. 얼굴은 숙여져 정확한 얼굴 일치는 판단하기 어렵다. 한국인 중년 여성으로 보이는 인물이 이불을 하체에 두르고 추가되어 있다. 배경 설명에는 어머니가 언급되지만, 우선하는 숏 텍스트와 인물 제한은 안규홍만 허용한다. 묶인 작은 봇짐과 방바닥은 있으며 읽을 수 있는 글자는 없다.",
        "hard_violations": [
         "절하는 남성의 머리가 두 개이고 몸통·팔다리가 서로 겹쳐 융합된 형태여서, 인물 중복 및 불가능한 인체 구조에 해당한다.",
         "숏 텍스트와 인물 제한이 허용하지 않은 여성을 추가했다."
        ],
        "physics": "손바닥과 팔뚝, 접힌 다리 및 짚신은 바닥에 닿아 있어 공중에 떠 있는 몸은 아니다. 그러나 두 머리가 하나의 겹친 몸통 덩어리에 연결되는 구조는 한 사람이 무릎을 꿇고 절하는 동작으로 만들 수 없다. 여성은 침구와 바닥에 앉아 지지되고, 봇짐과 가구도 바닥에 놓여 있다."
       },
       {
        "label": "A",
        "direction": "남성은 화면 오른쪽의 여성 방향으로 몸을 접어 절하고, 얼굴과 이마는 바로 아래 방바닥을 향한다. 이마 부근의 바닥 접촉이 보인다. 여성은 아래쪽의 절하는 남성을 바라본다.",
        "built_space": "오른쪽 출입구 하나, 뒤쪽 높은 장 하나와 상단 바구니 하나, 낮은 서랍장 하나, 왼쪽 낮은 함 하나, 왼쪽 바닥 침구 한 벌이 보인다. 장소 사진의 벽 재질, 목재 골조, 바닥과 주요 가구 배치가 잘 유지된다. 여성은 오른쪽 벽에 등을 대고 바닥에 앉아 있고, 봇짐 하나가 여성 옆 문 가까이에 놓여 있다. 남성의 전신과 머리 아래 바닥을 포함하는 와이드숏이다.",
        "entities": "남성 한 명은 청년으로 보이며 검은 묶음머리, 낡은 밝은색 저고리와 바지, 허리끈 및 짚신이 인물 참조와 대체로 맞는다. 얼굴이 바닥을 향하므로 정확한 얼굴 일치는 확인할 수 없다. 회색 머리의 한국인 노년 여성으로 보이는 인물이 허리와 다리에 이불을 두르고 등장한다. 어머니의 배경 묘사에는 맞지만, 숏 텍스트가 정하는 안규홍 단독 등장 제한에는 어긋난다. 작은 묶인 봇짐이 있으며 읽을 수 있는 글자는 없다.",
        "hard_violations": [
         "숏 텍스트와 인물 제한이 허용하지 않은 여성을 추가했다."
        ],
        "physics": "남성은 무릎과 정강이를 바닥에 접고 앉아 상체를 낮췄으며, 팔뚝과 손 및 이마 부근이 바닥에 닿는다. 체중을 지탱하는 접점이 있고 깊은 절로 만들 수 있는 자세다. 여성의 엉덩이와 접힌 다리는 바닥에 지지되며 등 뒤에는 벽이 있다. 이불은 무릎 위에 내려앉고 봇짐과 가구는 바닥에 놓여 있어 지지 없는 부유는 없다."
       }
      ]
     },
     "ok": true
    }
   ],
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "slot_winner_match": true,
   "slot_winner": {
    "gemini-pro": "A",
    "gpt-high": "A"
   },
   "route": "cross_slot_agree"
  },
  "dual": {
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "normalized": {
    "A": 2.0,
    "B": 0.25
   },
   "adjusted": {
    "A": 1.75,
    "B": 0.0
   },
   "violations": {
    "B": [
     "[gemini-pro] physically impossible anatomy (엎드린 인물의 엉덩이/허리 부근에 두 번째 머리가 달려 있음)",
     "[gpt-high] 절하는 남성의 머리가 두 개이고 몸통·팔다리가 서로 겹쳐 융합된 형태여서, 인물 중복 및 불가능한 인체 구조에 해당한다.",
     "[gpt-high] 숏 텍스트와 인물 제한이 허용하지 않은 여성을 추가했다."
    ],
    "A": [
     "[gpt-high] 숏 텍스트와 인물 제한이 허용하지 않은 여성을 추가했다."
    ]
   },
   "per_model_winner": {
    "gemini-pro": "A",
    "gpt-high": "A"
   },
   "agreed": true
  },
  "totals": {
   "A": 1750,
   "B": 0
  },
  "selected": "A",
  "ranking": [
   "A",
   "B"
  ],
  "verdicts": [
   {
    "label": "A",
    "score": 1750,
    "verdict_ko": "샷 텍스트가 요구하는 안규홍의 엎드린 자세와 배경 요소(앉아있는 어머니, 봇짐)를 지정된 방 구조 안에 사실적이고 자연스럽게 구현했습니다.  ★위반: [gpt-high] 숏 텍스트와 인물 제한이 허용하지 않은 여성을 추가했다."
   },
   {
    "label": "B",
    "score": 0,
    "verdict_ko": "엎드린 안규홍의 몸에 머리가 두 개 생성되는 등 물리적으로 불가능한 기괴한 신체 구조가 나타나 완전히 실격입니다.  ★위반: [gemini-pro] physically impossible anatomy (엎드린 인물의 엉덩이/허리 부근에 두 번째 머리가 달려 있음) / [gpt-high] 절하는 남성의 머리가 두 개이고 몸통·팔다리가 서로 겹쳐 융합된 형태여서, 인물 중복 및 불가능한 인체 구조에 해당한다. / [gpt-high] 숏 텍스트와 인물 제한이 허용하지 않은 여성을 추가했다."
   }
  ],
  "refs": [
   {
    "label": "LOCATION PHOTOGRAPH — the exact place of this shot: its architecture, materials, fixed features and lighting mood are spatial truth; stage the moment inside this place. Never copy its camera framing.",
    "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/db00517b-165b-4005-a287-f28936b59b95/episodes/272bd7af-9294-4490-98c6-88a40e9deacc/images/background_chain/L04B02.png",
    "asset_id": "8aaff0f1-df32-4a26-92ab-4f112c48d1af",
    "role": "location_plate"
   },
   {
    "label": "CHARACTER REFERENCE — 안규홍: the exact person appearing in this shot; match face, hair and build exactly.",
    "path": "<bytes:1028384>",
    "asset_id": "3b5894db-ce55-4f34-b966-5e725c3aa952",
    "role": "character_ref"
   }
  ],
  "critique_skipped": true,
  "shot_run_produced": false,
  "shot_run_uid": "06aa140c-5327-7967-ab84-cb20711fb78c",
  "bgfirst": {
   "bg_path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/db00517b-165b-4005-a287-f28936b59b95/images/272bd7af-9294-4490-98c6-88a40e9deacc/scene/recipe/S11sh4__bgfirst_bg.png",
   "bg_asset_id": "4390caec-8bb9-4035-bcc2-6749e1eaca05",
   "bg_record_key": "S11sh4::bgfirst_bg",
   "chain_winner": true,
   "authority": "plate"
  },
  "ref_mode": "재투영 배경+콘티+엔티티 (2택1: 체인 승)",
  "share_plan": {
   "ref_plan": "background"
  }
 },
 "S11sh4::cine": {
  "applied": true,
  "attempted_at": "2026-09-09T01:50:07.199897+00:00",
  "fingerprint": "0cd4e88375f5b1f5fa370546ddeab0dd84872a1ba84cb31434f57001f44a96e0",
  "fingerprint_version": 2,
  "provider": "mai",
  "endpoint": "openrouter/chat-completions",
  "model": "microsoft/mai-image-2.6",
  "pack": "24.202608252115",
  "source_file": "S11sh4_sel.png",
  "source_sha256": "88605fbab0e5070561f874db0ffa9863182808262a458a3071cfa67066b3f88a",
  "file": "S11sh4_cine.png",
  "staged_sha256": "4c47ad255a5d0b832e8296ae6ee590cafffd1fcaa7564e097f43fc98d5a4ba0c",
  "latency_ms": 29252
 },
 "S11sh9::signage": {
  "fp": "fccb35051e71cd7a",
  "inscriptions": [],
  "cues": [],
  "dropped": []
 },
 "S11sh9::bgfirst_bg": {
  "input_fingerprint": "6a02699fab4212a6",
  "prompt": "Create the EMPTY BACKGROUND PLATE for one film shot — NO PEOPLE, no figures, no body parts, no sketch lines, no arrows anywhere.\n\"Empty\" means no people only: KEEP the location's inherent occupants and stock that define the place — animals in an animal shelter, pen or farm, goods in a market, moored boats in a harbour — unless the shot text explicitly removes them.\nThe FIRST attached image is a thin-line storyboard sketch: use ONLY its camera angle, horizon, perspective and the placement/size of buildings and set masses — ignore the sketched people and arrows entirely. The SECOND attached image (LOCATION PHOTOGRAPH) is the real place: take its architecture, materials, signage and fixed features, and RE-PROJECT them into the sketch's camera. If the photograph's camera differs from the sketch's, the sketch's camera wins.\nHUMAN-SCALE CALIBRATION: derive every structure's true size from human-scale elements — a door ≈ 2m, a window ≈ 1–1.5m wide, one storey ≈ 2.5–3m; never inflate a small structure or shrink a large one.\n\nSHOT TEXT this background must serve (Korean): 안규홍의 크고 굵은 손바닥 위로 옥가락지를 억지로 꽉 누르고 있는 달래의 두 손 클로즈업.\n\nLOCATION (lock): Immediately outside the rustic wicket gate of a modest home, where a farewell takes place at dawn.\n\nTIME OF DAY (lock): dawn, bluish natural light.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: close-up\n- KEY BACKGROUND ELEMENTS: 옥가락지 (Being pressed into 안규홍's open palm by 달래's two hands) — Seen obliquely from above between the pressing fingers and receiving palm; used as Small central narrative detail, kept visible without enlarging it beyond its scale relative to the hands.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Bluish dawn light gently models the hands with muted saturation and naturalistic contrast, preserving the restrained warmth of the gesture.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nRender ONE photorealistic empty location photograph, 16:9, that this shot can be staged inside later. No readable writing anywhere: surfaces that would carry writing may be present, but stage any wording out of legibility — an oblique angle, distance, shallow focus. No captions, watermarks or overlay text.",
  "effective_prompt": "Create the EMPTY BACKGROUND PLATE for one film shot — NO PEOPLE, no figures, no body parts, no sketch lines, no arrows anywhere.\n\"Empty\" means no people only: KEEP the location's inherent occupants and stock that define the place — animals in an animal shelter, pen or farm, goods in a market, moored boats in a harbour — unless the shot text explicitly removes them.\nThe FIRST attached image is a thin-line storyboard sketch: use ONLY its camera angle, horizon, perspective and the placement/size of buildings and set masses — ignore the sketched people and arrows entirely. The SECOND attached image (LOCATION PHOTOGRAPH) is the real place: take its architecture, materials, signage and fixed features, and RE-PROJECT them into the sketch's camera. If the photograph's camera differs from the sketch's, the sketch's camera wins.\nHUMAN-SCALE CALIBRATION: derive every structure's true size from human-scale elements — a door ≈ 2m, a window ≈ 1–1.5m wide, one storey ≈ 2.5–3m; never inflate a small structure or shrink a large one.\n\nSHOT TEXT this background must serve (Korean): 안규홍의 크고 굵은 손바닥 위로 옥가락지를 억지로 꽉 누르고 있는 달래의 두 손 클로즈업.\n\nLOCATION (lock): Immediately outside the rustic wicket gate of a modest home, where a farewell takes place at dawn.\n\nTIME OF DAY (lock): dawn, bluish natural light.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: close-up\n- KEY BACKGROUND ELEMENTS: 옥가락지 (Being pressed into 안규홍's open palm by 달래's two hands) — Seen obliquely from above between the pressing fingers and receiving palm; used as Small central narrative detail, kept visible without enlarging it beyond its scale relative to the hands.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Bluish dawn light gently models the hands with muted saturation and naturalistic contrast, preserving the restrained warmth of the gesture.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nRender ONE photorealistic empty location photograph, 16:9, that this shot can be staged inside later. No readable writing anywhere: surfaces that would carry writing may be present, but stage any wording out of legibility — an oblique angle, distance, shallow focus. No captions, watermarks or overlay text.",
  "bg_path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/db00517b-165b-4005-a287-f28936b59b95/images/272bd7af-9294-4490-98c6-88a40e9deacc/scene/recipe/S11sh9__bgfirst_bg.png",
  "asset_id": "a9bf211b-d971-40e5-b48e-ee6b6f16bc5f",
  "input_asset_ids": [
   "36b911e5-76fa-4c1f-a0a3-4f1a6003cd32",
   "ad856e2d-b926-4109-933d-7db2c7884222"
  ]
 },
 "S11sh9": {
  "input_fingerprint": "9fde479963177626",
  "prompt": "Create ONE FINAL photorealistic live-action film still of the moment below — early 20th-century Korea, Jeolla region; all people are Korean unless explicitly stated otherwise. TIME OF DAY (lock): dawn, bluish natural light.\n\nSHOT TEXT (authoritative, Korean): 안규홍의 크고 굵은 손바닥 위로 옥가락지를 억지로 꽉 누르고 있는 달래의 두 손 클로즈업.\n\nLOCATION (lock): Immediately outside the rustic wicket gate of a modest home, where a farewell takes place at dawn. The shot takes place here — the attached LOCATION PHOTOGRAPH shows the exact spot.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: close-up\n- KEY BACKGROUND ELEMENTS: 옥가락지 (Being pressed into 안규홍's open palm by 달래's two hands) — Seen obliquely from above between the pressing fingers and receiving palm; used as Small central narrative detail, kept visible without enlarging it beyond its scale relative to the hands.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Bluish dawn light gently models the hands with muted saturation and naturalistic contrast, preserving the restrained warmth of the gesture.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The room door and the outer gate are closed behind the departure; dawn light remains bluish. 안규홍: He stands outside with his travel bundle and an open palm receiving a small jade ring. 달래: She stands outside the gate with her hands extended in the ring handover.\n\nTHE HAND THAT IS DOING THIS: the object at the centre of this shot is being held, operated, read, aimed or handed over by 달래 right now, so 달래's hand — and as much of the wrist and forearm as the framing reaches — is in the frame, gripping or resting on that object exactly the way its real use requires. Match that hand to 달래: its size, build, skin, age, grooming, sleeve and anything worn on it belong to that person and to no one else.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 안규홍 (한국인 남성, 청년 나이대, 검은 머리, 단단히 묶은 머리); 달래 (한국인 여성, 젊은 성인, 자연스러운 검은 머리) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.",
  "roll_prompts": {
   "A": "Stage the shot. The FIRST attached image (SHOT BACKGROUND) is the finished empty background of this shot — keep it EXACTLY: its camera, perspective, architecture, lighting and every fixed feature stay untouched. The SECOND attached image (LAYOUT SKETCH) tells you ONLY where the people go: each sketched person's position, screen size, pose and the gaze/motion arrows. Ignore the sketch's background lines. The CHARACTER REFERENCE photographs show the real people.\nPlace the real people into the background at exactly the sketched positions, sizes and poses, following the arrow directions. No sketch lines or arrows may remain.\n\nCreate ONE FINAL photorealistic live-action film still of the moment below — early 20th-century Korea, Jeolla region; all people are Korean unless explicitly stated otherwise. TIME OF DAY (lock): dawn, bluish natural light.\n\nSHOT TEXT (authoritative, Korean): 안규홍의 크고 굵은 손바닥 위로 옥가락지를 억지로 꽉 누르고 있는 달래의 두 손 클로즈업.\n\nLOCATION (lock): Immediately outside the rustic wicket gate of a modest home, where a farewell takes place at dawn. The shot takes place here — the FIRST attached image (SHOT BACKGROUND) is this exact place, already built: its ground, structures, horizon, materials and lighting are the finished truth of this location and must not be redesigned or replaced. No location photograph is attached — read the place from that image alone, and add no scenery, structure, vehicle or fixture that it does not already show. This lock governs the place only; the figures in the shot follow the staging and pose instructions.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: close-up\n- KEY BACKGROUND ELEMENTS: 옥가락지 (Being pressed into 안규홍's open palm by 달래's two hands) — Seen obliquely from above between the pressing fingers and receiving palm; used as Small central narrative detail, kept visible without enlarging it beyond its scale relative to the hands.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Bluish dawn light gently models the hands with muted saturation and naturalistic contrast, preserving the restrained warmth of the gesture.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The room door and the outer gate are closed behind the departure; dawn light remains bluish. 안규홍: He stands outside with his travel bundle and an open palm receiving a small jade ring. 달래: She stands outside the gate with her hands extended in the ring handover.\n\nTHE HAND THAT IS DOING THIS: the object at the centre of this shot is being held, operated, read, aimed or handed over by 달래 right now, so 달래's hand — and as much of the wrist and forearm as the framing reaches — is in the frame, gripping or resting on that object exactly the way its real use requires. Match that hand to 달래: its size, build, skin, age, grooming, sleeve and anything worn on it belong to that person and to no one else.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 안규홍 (한국인 남성, 청년 나이대, 검은 머리, 단단히 묶은 머리); 달래 (한국인 여성, 젊은 성인, 자연스러운 검은 머리) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.",
   "B": "Create ONE FINAL photorealistic live-action film still of the moment below — early 20th-century Korea, Jeolla region; all people are Korean unless explicitly stated otherwise. TIME OF DAY (lock): dawn, bluish natural light.\n\nSHOT TEXT (authoritative, Korean): 안규홍의 크고 굵은 손바닥 위로 옥가락지를 억지로 꽉 누르고 있는 달래의 두 손 클로즈업.\n\nLOCATION (lock): Immediately outside the rustic wicket gate of a modest home, where a farewell takes place at dawn. The shot takes place here — the attached LOCATION PHOTOGRAPH shows the exact spot.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: close-up\n- KEY BACKGROUND ELEMENTS: 옥가락지 (Being pressed into 안규홍's open palm by 달래's two hands) — Seen obliquely from above between the pressing fingers and receiving palm; used as Small central narrative detail, kept visible without enlarging it beyond its scale relative to the hands.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Bluish dawn light gently models the hands with muted saturation and naturalistic contrast, preserving the restrained warmth of the gesture.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The room door and the outer gate are closed behind the departure; dawn light remains bluish. 안규홍: He stands outside with his travel bundle and an open palm receiving a small jade ring. 달래: She stands outside the gate with her hands extended in the ring handover.\n\nTHE HAND THAT IS DOING THIS: the object at the centre of this shot is being held, operated, read, aimed or handed over by 달래 right now, so 달래's hand — and as much of the wrist and forearm as the framing reaches — is in the frame, gripping or resting on that object exactly the way its real use requires. Match that hand to 달래: its size, build, skin, age, grooming, sleeve and anything worn on it belong to that person and to no one else.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 안규홍 (한국인 남성, 청년 나이대, 검은 머리, 단단히 묶은 머리); 달래 (한국인 여성, 젊은 성인, 자연스러운 검은 머리) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.\n\nCOMPOSITION VARIATION (this candidate only): keep the place, the people and the moment exactly as specified, and keep from the CAMERA & FRAME contract its subject scale, screen placement and key background elements exactly. For this candidate only, its camera angle and height are deliberately left open: choose a camera position distinctly different from the obvious one, as a film director picking a second setup on the same blocking. Stay true to any explicit framing the SHOT TEXT itself dictates."
  },
  "roll_refs": {
   "A": [
    {
     "label": "SHOT BACKGROUND",
     "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/db00517b-165b-4005-a287-f28936b59b95/images/272bd7af-9294-4490-98c6-88a40e9deacc/scene/recipe/S11sh9__bgfirst_bg.png",
     "asset_id": "a9bf211b-d971-40e5-b48e-ee6b6f16bc5f",
     "role": "bgfirst_bg"
    },
    {
     "label": "LAYOUT SKETCH (people placement only)",
     "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/db00517b-165b-4005-a287-f28936b59b95/images/272bd7af-9294-4490-98c6-88a40e9deacc/conti/conti_S11sh9.png",
     "asset_id": "36b911e5-76fa-4c1f-a0a3-4f1a6003cd32",
     "role": "conti_light"
    },
    {
     "label": "CHARACTER REFERENCE — 안규홍: the exact person appearing in this shot; match face, hair and build exactly.",
     "path": "<bytes:1028384>",
     "asset_id": "3b5894db-ce55-4f34-b966-5e725c3aa952",
     "role": "character_ref"
    },
    {
     "label": "CHARACTER REFERENCE — 달래: the exact person appearing in this shot; match face, hair and build exactly.",
     "path": "<bytes:879791>",
     "asset_id": "83024e69-b375-4f3e-98ec-820d0e2ff1f0",
     "role": "character_ref"
    }
   ],
   "B": [
    {
     "label": "LOCATION PHOTOGRAPH — the exact place of this shot: its architecture, materials, fixed features and lighting mood are spatial truth; stage the moment inside this place. Never copy its camera framing.",
     "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/db00517b-165b-4005-a287-f28936b59b95/episodes/272bd7af-9294-4490-98c6-88a40e9deacc/images/background_chain/L04B01.png",
     "asset_id": "ad856e2d-b926-4109-933d-7db2c7884222",
     "role": "location_plate"
    },
    {
     "label": "CHARACTER REFERENCE — 안규홍: the exact person appearing in this shot; match face, hair and build exactly.",
     "path": "<bytes:1028384>",
     "asset_id": "3b5894db-ce55-4f34-b966-5e725c3aa952",
     "role": "character_ref"
    },
    {
     "label": "CHARACTER REFERENCE — 달래: the exact person appearing in this shot; match face, hair and build exactly.",
     "path": "<bytes:879791>",
     "asset_id": "83024e69-b375-4f3e-98ec-820d0e2ff1f0",
     "role": "character_ref"
    }
   ]
  },
  "shot_run_spend_attempt_count": 1,
  "judge_flip_skipped": "judge_fn_owns_order",
  "readings": [
   {
    "label": "A",
    "direction": "달래의 손이 왼쪽에서 안규홍의 손(오른쪽)으로 향하며, 옥가락지를 안규홍의 펴진 손바닥 위에 올려놓으려 하고 있습니다.",
    "built_space": "배경 왼쪽에 참고 사진의 흙담집과 사립문이, 오른쪽에는 산길이 위치해 있어 인물들이 문 밖에 서 있는 설정에 부합합니다.",
    "entities": "안규홍의 거친 손과 흰색 무명 소매, 달래의 옥색 저고리 소매가 확인되나 옥가락지의 형태가 매우 얇게 왜곡되었습니다.",
    "hard_violations": [
     "[gemini-pro] 소매는 하나뿐인데 맨살의 또 다른 팔이 기형적으로 겹쳐서 나오거나 하나의 팔에서 두 손이 파생되는 등 물리적으로 불가능한 신체 구조가 나타납니다."
    ],
    "physics": "두 손이 가락지를 잡고 다른 손 위에 얹는 자세를 취하려 하나, 구조적으로 불가능한 팔뚝의 형태로 인해 신체가 물리적 지지력을 상실했습니다."
   },
   {
    "label": "B",
    "direction": "오른쪽에서 뻗어 나온 달래의 두 손이 왼쪽의 안규홍의 손바닥 위로 옥가락지를 강하게 누르고 있습니다.",
    "built_space": "배경 중앙 뒤편에 사립문의 질감이 아웃포커스되어 배치되어 있으며, 두 인물이 문 밖에서 마주 선 동선을 정확하게 반영합니다.",
    "entities": "안규홍의 크고 투박한 손바닥과 흰색 소매, 달래의 작고 섬세한 두 손과 옥색 소매, 그리고 옥가락지의 질감과 형태가 모두 지시문 및 참고 사진과 일치합니다.",
    "hard_violations": [],
    "physics": "달래의 한 손이 옥가락지를 쥐고 다른 한 손이 그 위를 덮어 힘을 가하는 행동이 자연스럽게 연출되었으며, 안규홍의 손바닥이 이를 안정적으로 받치고 있습니다."
   }
  ],
  "cross_model_order": {
   "policy": "cross_model_order_v2_slot0_fwd_slot1_rev_parallel2",
   "slots": [
    {
     "model": "gemini-pro",
     "order": "forward",
     "display_to_canonical": {
      "A": "A",
      "B": "B"
     },
     "raw": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 10,
        "verdict_ko": "지시된 클로즈업 샷과 캐릭터의 소매, 옥가락지 등 디테일을 완벽하게 살렸으며 두 손으로 가락지를 누르는 행동 묘사가 자연스럽습니다."
       },
       {
        "label": "A",
        "score": 2,
        "verdict_ko": "달래의 팔이 하나의 소매에서 두 개의 손으로 기형적으로 분리되어 나타나는 치명적인 신체 구조 오류가 있습니다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "달래의 손이 왼쪽에서 안규홍의 손(오른쪽)으로 향하며, 옥가락지를 안규홍의 펴진 손바닥 위에 올려놓으려 하고 있습니다.",
        "built_space": "배경 왼쪽에 참고 사진의 흙담집과 사립문이, 오른쪽에는 산길이 위치해 있어 인물들이 문 밖에 서 있는 설정에 부합합니다.",
        "entities": "안규홍의 거친 손과 흰색 무명 소매, 달래의 옥색 저고리 소매가 확인되나 옥가락지의 형태가 매우 얇게 왜곡되었습니다.",
        "hard_violations": [
         "소매는 하나뿐인데 맨살의 또 다른 팔이 기형적으로 겹쳐서 나오거나 하나의 팔에서 두 손이 파생되는 등 물리적으로 불가능한 신체 구조가 나타납니다."
        ],
        "physics": "두 손이 가락지를 잡고 다른 손 위에 얹는 자세를 취하려 하나, 구조적으로 불가능한 팔뚝의 형태로 인해 신체가 물리적 지지력을 상실했습니다."
       },
       {
        "label": "B",
        "direction": "오른쪽에서 뻗어 나온 달래의 두 손이 왼쪽의 안규홍의 손바닥 위로 옥가락지를 강하게 누르고 있습니다.",
        "built_space": "배경 중앙 뒤편에 사립문의 질감이 아웃포커스되어 배치되어 있으며, 두 인물이 문 밖에서 마주 선 동선을 정확하게 반영합니다.",
        "entities": "안규홍의 크고 투박한 손바닥과 흰색 소매, 달래의 작고 섬세한 두 손과 옥색 소매, 그리고 옥가락지의 질감과 형태가 모두 지시문 및 참고 사진과 일치합니다.",
        "hard_violations": [],
        "physics": "달래의 한 손이 옥가락지를 쥐고 다른 한 손이 그 위를 덮어 힘을 가하는 행동이 자연스럽게 연출되었으며, 안규홍의 손바닥이 이를 안정적으로 받치고 있습니다."
       }
      ],
      "all_candidates_fail": false
     },
     "normalized": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 10,
        "verdict_ko": "지시된 클로즈업 샷과 캐릭터의 소매, 옥가락지 등 디테일을 완벽하게 살렸으며 두 손으로 가락지를 누르는 행동 묘사가 자연스럽습니다."
       },
       {
        "label": "A",
        "score": 2,
        "verdict_ko": "달래의 팔이 하나의 소매에서 두 개의 손으로 기형적으로 분리되어 나타나는 치명적인 신체 구조 오류가 있습니다."
       }
      ],
      "all_candidates_fail": false,
      "readings": [
       {
        "label": "A",
        "direction": "달래의 손이 왼쪽에서 안규홍의 손(오른쪽)으로 향하며, 옥가락지를 안규홍의 펴진 손바닥 위에 올려놓으려 하고 있습니다.",
        "built_space": "배경 왼쪽에 참고 사진의 흙담집과 사립문이, 오른쪽에는 산길이 위치해 있어 인물들이 문 밖에 서 있는 설정에 부합합니다.",
        "entities": "안규홍의 거친 손과 흰색 무명 소매, 달래의 옥색 저고리 소매가 확인되나 옥가락지의 형태가 매우 얇게 왜곡되었습니다.",
        "hard_violations": [
         "소매는 하나뿐인데 맨살의 또 다른 팔이 기형적으로 겹쳐서 나오거나 하나의 팔에서 두 손이 파생되는 등 물리적으로 불가능한 신체 구조가 나타납니다."
        ],
        "physics": "두 손이 가락지를 잡고 다른 손 위에 얹는 자세를 취하려 하나, 구조적으로 불가능한 팔뚝의 형태로 인해 신체가 물리적 지지력을 상실했습니다."
       },
       {
        "label": "B",
        "direction": "오른쪽에서 뻗어 나온 달래의 두 손이 왼쪽의 안규홍의 손바닥 위로 옥가락지를 강하게 누르고 있습니다.",
        "built_space": "배경 중앙 뒤편에 사립문의 질감이 아웃포커스되어 배치되어 있으며, 두 인물이 문 밖에서 마주 선 동선을 정확하게 반영합니다.",
        "entities": "안규홍의 크고 투박한 손바닥과 흰색 소매, 달래의 작고 섬세한 두 손과 옥색 소매, 그리고 옥가락지의 질감과 형태가 모두 지시문 및 참고 사진과 일치합니다.",
        "hard_violations": [],
        "physics": "달래의 한 손이 옥가락지를 쥐고 다른 한 손이 그 위를 덮어 힘을 가하는 행동이 자연스럽게 연출되었으며, 안규홍의 손바닥이 이를 안정적으로 받치고 있습니다."
       }
      ]
     },
     "ok": true
    },
    {
     "model": "gpt-high",
     "order": "reverse",
     "display_to_canonical": {
      "A": "B",
      "B": "A"
     },
     "raw": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 8,
        "verdict_ko": "손 클로즈업과 장소·새벽빛은 맞지만, 달래가 한 손으로 가락지를 세워 놓고 다른 손은 손바닥에 얹어 두어 ‘두 손으로 억지로 꽉 누르는’ 동작이 약하다."
       },
       {
        "label": "B",
        "score": 9,
        "verdict_ko": "달래의 두 손이 큰 손바닥 중앙의 작은 옥가락지로 모여 눌러 주는 순간을 잡아, 요구한 클로즈업과 강한 전달 동작을 더 충실하게 구현한다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "왼쪽에서 나온 안규홍의 손바닥은 위로 열려 있고, 오른쪽에서 나온 달래의 두 손은 그 손바닥 중앙을 향한다. 한 손의 손끝은 옥가락지를 잡고 아래로 향하지만, 다른 손의 손가락은 가락지보다 받는 손의 손가락 쪽에 놓여 있다. 시선은 프레임 밖이다.",
        "built_space": "손 뒤로 닫힌 나뭇가지 사립문 한 짝과 굵은 문기둥 두 개가 보이며, 왼쪽 아래에는 엮은 울타리, 뒤에는 흙벽, 오른쪽에는 흙길이 있다. 두 사람의 팔은 문 바깥 전경에서 마주 뻗어 있다. 참조 장소의 재료와 배치에 부합하며, 방 문은 보이지 않아 닫힘 여부를 판단할 수 없다. 반사면이나 중복 설비는 없다.",
        "entities": "거칠고 넓은 남성 손 하나와 상대적으로 매끈하고 가는 여성 손 두 개가 보인다. 남성의 낡은 미색 소매와 여성의 옅은 녹색 소매는 각 인물 참조와 부합한다. 얼굴과 머리카락은 보이지 않아 정확한 인물 정체성·민족성·나이는 확인할 수 없다. 작은 짙은 녹색 고리 하나가 실제 손가락 가락지 크기로 보이나, 비스듬히 서 있어 구멍은 일부만 드러난다. 보따리는 프레임 밖이며 추가 인물이나 읽을 수 있는 글자는 없다.",
        "hard_violations": [],
        "physics": "가락지의 아래쪽은 안규홍의 손바닥에 닿고 양옆은 달래의 손끝에 잡혀 있어 지지가 명확하다. 세 손 모두 손목과 소매로 이어지며 접촉 관계도 가능하다. 다만 가락지를 세워 잡는 손과 손바닥에 편하게 얹힌 다른 손의 조합은 강하게 함께 누르기보다 조심스럽게 건네는 동작에 가깝다."
       },
       {
        "label": "B",
        "direction": "오른쪽에서 나온 안규홍의 큰 손바닥이 위로 열리고, 왼쪽에서 나온 달래의 두 손이 손바닥 중앙의 옥가락지를 향해 모인다. 굽힌 손가락과 아래로 향한 손끝이 가락지 및 바로 주변을 누르는 방향이다. 가락지의 구멍은 위에서 비스듬히 보이며, 시선은 프레임 밖이다.",
        "built_space": "배경 왼쪽에 닫힌 사립문 한 짝과 굵은 문기둥 두 개, 왼쪽 아래에 나뭇가지 울타리가 있다. 뒤쪽 집은 흙벽과 기와지붕, 격자창 하나를 갖추고 오른쪽으로 흙길이 이어져 참조 장소와 잘 맞는다. 두 사람의 팔은 문 바깥 전경에서 만난다. 방 문은 프레임에 없으며 불가능한 반사나 중복 설비는 없다.",
        "entities": "큰 거친 남성 손 하나, 젊은 여성에 부합하는 가는 손 두 개, 작은 녹색 옥가락지 하나가 보인다. 낡은 미색 남성 소매와 옅은 녹색 여성 소매는 참조 의상과 맞는다. 가락지는 둥근 구멍과 두께, 은은한 반사가 있는 실제 장신구로 읽히며 손 대비 크기도 적절하다. 얼굴·머리·보따리는 구도 밖이므로 확인 대상이 아니다. 추가 인물이나 읽을 수 있는 글자는 없다.",
        "hard_violations": [],
        "physics": "가락지는 펼친 손바닥 위에 누워 지지되고, 달래의 손끝이 윗면과 가장자리에 닿는다. 다른 손도 가까이 굽혀 받는 손 위를 누르고 있어 두 손이 함께 힘을 싣는 동작이 읽힌다. 손과 팔의 연결 및 접촉은 물리적으로 가능하며 떠 있는 물체는 없다."
       }
      ],
      "all_candidates_fail": false
     },
     "normalized": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 8,
        "verdict_ko": "손 클로즈업과 장소·새벽빛은 맞지만, 달래가 한 손으로 가락지를 세워 놓고 다른 손은 손바닥에 얹어 두어 ‘두 손으로 억지로 꽉 누르는’ 동작이 약하다."
       },
       {
        "label": "A",
        "score": 9,
        "verdict_ko": "달래의 두 손이 큰 손바닥 중앙의 작은 옥가락지로 모여 눌러 주는 순간을 잡아, 요구한 클로즈업과 강한 전달 동작을 더 충실하게 구현한다."
       }
      ],
      "all_candidates_fail": false,
      "readings": [
       {
        "label": "B",
        "direction": "왼쪽에서 나온 안규홍의 손바닥은 위로 열려 있고, 오른쪽에서 나온 달래의 두 손은 그 손바닥 중앙을 향한다. 한 손의 손끝은 옥가락지를 잡고 아래로 향하지만, 다른 손의 손가락은 가락지보다 받는 손의 손가락 쪽에 놓여 있다. 시선은 프레임 밖이다.",
        "built_space": "손 뒤로 닫힌 나뭇가지 사립문 한 짝과 굵은 문기둥 두 개가 보이며, 왼쪽 아래에는 엮은 울타리, 뒤에는 흙벽, 오른쪽에는 흙길이 있다. 두 사람의 팔은 문 바깥 전경에서 마주 뻗어 있다. 참조 장소의 재료와 배치에 부합하며, 방 문은 보이지 않아 닫힘 여부를 판단할 수 없다. 반사면이나 중복 설비는 없다.",
        "entities": "거칠고 넓은 남성 손 하나와 상대적으로 매끈하고 가는 여성 손 두 개가 보인다. 남성의 낡은 미색 소매와 여성의 옅은 녹색 소매는 각 인물 참조와 부합한다. 얼굴과 머리카락은 보이지 않아 정확한 인물 정체성·민족성·나이는 확인할 수 없다. 작은 짙은 녹색 고리 하나가 실제 손가락 가락지 크기로 보이나, 비스듬히 서 있어 구멍은 일부만 드러난다. 보따리는 프레임 밖이며 추가 인물이나 읽을 수 있는 글자는 없다.",
        "hard_violations": [],
        "physics": "가락지의 아래쪽은 안규홍의 손바닥에 닿고 양옆은 달래의 손끝에 잡혀 있어 지지가 명확하다. 세 손 모두 손목과 소매로 이어지며 접촉 관계도 가능하다. 다만 가락지를 세워 잡는 손과 손바닥에 편하게 얹힌 다른 손의 조합은 강하게 함께 누르기보다 조심스럽게 건네는 동작에 가깝다."
       },
       {
        "label": "A",
        "direction": "오른쪽에서 나온 안규홍의 큰 손바닥이 위로 열리고, 왼쪽에서 나온 달래의 두 손이 손바닥 중앙의 옥가락지를 향해 모인다. 굽힌 손가락과 아래로 향한 손끝이 가락지 및 바로 주변을 누르는 방향이다. 가락지의 구멍은 위에서 비스듬히 보이며, 시선은 프레임 밖이다.",
        "built_space": "배경 왼쪽에 닫힌 사립문 한 짝과 굵은 문기둥 두 개, 왼쪽 아래에 나뭇가지 울타리가 있다. 뒤쪽 집은 흙벽과 기와지붕, 격자창 하나를 갖추고 오른쪽으로 흙길이 이어져 참조 장소와 잘 맞는다. 두 사람의 팔은 문 바깥 전경에서 만난다. 방 문은 프레임에 없으며 불가능한 반사나 중복 설비는 없다.",
        "entities": "큰 거친 남성 손 하나, 젊은 여성에 부합하는 가는 손 두 개, 작은 녹색 옥가락지 하나가 보인다. 낡은 미색 남성 소매와 옅은 녹색 여성 소매는 참조 의상과 맞는다. 가락지는 둥근 구멍과 두께, 은은한 반사가 있는 실제 장신구로 읽히며 손 대비 크기도 적절하다. 얼굴·머리·보따리는 구도 밖이므로 확인 대상이 아니다. 추가 인물이나 읽을 수 있는 글자는 없다.",
        "hard_violations": [],
        "physics": "가락지는 펼친 손바닥 위에 누워 지지되고, 달래의 손끝이 윗면과 가장자리에 닿는다. 다른 손도 가까이 굽혀 받는 손 위를 누르고 있어 두 손이 함께 힘을 싣는 동작이 읽힌다. 손과 팔의 연결 및 접촉은 물리적으로 가능하며 떠 있는 물체는 없다."
       }
      ]
     },
     "ok": true
    }
   ],
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "slot_winner_match": false,
   "slot_winner": {
    "gemini-pro": "B",
    "gpt-high": "A"
   },
   "route": "cross_slot_combined"
  },
  "dual": {
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "normalized": {
    "A": 1.2,
    "B": 1.889
   },
   "adjusted": {
    "A": 0.95,
    "B": 1.889
   },
   "violations": {
    "A": [
     "[gemini-pro] 소매는 하나뿐인데 맨살의 또 다른 팔이 기형적으로 겹쳐서 나오거나 하나의 팔에서 두 손이 파생되는 등 물리적으로 불가능한 신체 구조가 나타납니다."
    ]
   },
   "per_model_winner": {
    "gemini-pro": "B",
    "gpt-high": "A"
   },
   "agreed": false
  },
  "totals": {
   "B": 1889,
   "A": 950
  },
  "selected": "B",
  "ranking": [
   "B",
   "A"
  ],
  "verdicts": [
   {
    "label": "B",
    "score": 1889,
    "verdict_ko": "지시된 클로즈업 샷과 캐릭터의 소매, 옥가락지 등 디테일을 완벽하게 살렸으며 두 손으로 가락지를 누르는 행동 묘사가 자연스럽습니다."
   },
   {
    "label": "A",
    "score": 950,
    "verdict_ko": "달래의 팔이 하나의 소매에서 두 개의 손으로 기형적으로 분리되어 나타나는 치명적인 신체 구조 오류가 있습니다.  ★위반: [gemini-pro] 소매는 하나뿐인데 맨살의 또 다른 팔이 기형적으로 겹쳐서 나오거나 하나의 팔에서 두 손이 파생되는 등 물리적으로 불가능한 신체 구조가 나타납니다."
   }
  ],
  "refs": [
   {
    "label": "LOCATION PHOTOGRAPH — the exact place of this shot: its architecture, materials, fixed features and lighting mood are spatial truth; stage the moment inside this place. Never copy its camera framing.",
    "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/db00517b-165b-4005-a287-f28936b59b95/episodes/272bd7af-9294-4490-98c6-88a40e9deacc/images/background_chain/L04B01.png",
    "asset_id": "ad856e2d-b926-4109-933d-7db2c7884222",
    "role": "location_plate"
   },
   {
    "label": "CHARACTER REFERENCE — 안규홍: the exact person appearing in this shot; match face, hair and build exactly.",
    "path": "<bytes:1028384>",
    "asset_id": "3b5894db-ce55-4f34-b966-5e725c3aa952",
    "role": "character_ref"
   },
   {
    "label": "CHARACTER REFERENCE — 달래: the exact person appearing in this shot; match face, hair and build exactly.",
    "path": "<bytes:879791>",
    "asset_id": "83024e69-b375-4f3e-98ec-820d0e2ff1f0",
    "role": "character_ref"
   }
  ],
  "critique_skipped": true,
  "shot_run_produced": false,
  "shot_run_uid": "06aa140f-0c53-7187-94a0-11972ff322ba",
  "bgfirst": {
   "bg_path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/db00517b-165b-4005-a287-f28936b59b95/images/272bd7af-9294-4490-98c6-88a40e9deacc/scene/recipe/S11sh9__bgfirst_bg.png",
   "bg_asset_id": "a9bf211b-d971-40e5-b48e-ee6b6f16bc5f",
   "bg_record_key": "S11sh9::bgfirst_bg",
   "chain_winner": false,
   "authority": "plate"
  },
  "ref_mode": "플레이트+엔티티 (2택1: 무콘티 승)",
  "share_plan": {
   "ref_plan": "background"
  }
 },
 "S11sh9::cine": {
  "applied": true,
  "attempted_at": "2026-09-09T01:53:44.537818+00:00",
  "fingerprint": "8f15dd7b4288629f57ed52ef5a2b7d8eac4e8b0a352e6d98bce4839a52aa2ff0",
  "fingerprint_version": 2,
  "provider": "mai",
  "endpoint": "openrouter/chat-completions",
  "model": "microsoft/mai-image-2.6",
  "pack": "24.202608252115",
  "source_file": "S11sh9_sel.png",
  "source_sha256": "26521d0d0ee6814268760c125125071d74193b230cc40663362512c824d3587f",
  "file": "S11sh9_cine.png",
  "staged_sha256": "45b97b10fd224c74993186ecf8257981dc3c78caf74dd0f351e60d47f9fe736f",
  "latency_ms": 25446
 },
 "S11sh16::signage": {
  "fp": "f1ef95bc46fc092f",
  "inscriptions": [],
  "cues": [],
  "dropped": []
 },
 "S11sh16": {
  "input_fingerprint": "e85494d9e07ff0dd",
  "prompt": "Create ONE FINAL photorealistic live-action film still of the moment below — early 20th-century Korea, Jeolla region; all people are Korean unless explicitly stated otherwise. TIME OF DAY (lock): dawn, bluish natural light.\n\nSHOT TEXT (authoritative, Korean): 안개가 깔린 산길을 향해 성큼 한쪽 발을 내디딘 mid-stride 순간의 안규홍의 뒷모습.\n\nLOCATION (lock): The outdoor path beyond a modest home's wicket gate, turning toward the misty mountain approach at dawn. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: wide shot\n- KEY BACKGROUND ELEMENTS: 산으로 향하는 길 (The departure route continues toward the mountain) — Recedes diagonally from beneath 안규홍 toward the upper-right background; used as Provides space ahead of the stepping foot and a continuous direction of departure.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: The established bluish dawn light remains subdued and naturalistic, letting separation register without an abrupt change in contrast.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The clothing, hair and overall look of 안규홍 — who appear both in that photo and in this shot — are LOCKED to that photo. Anyone else visible in that photo is NOT in this shot: never carry their face, body or clothing onto anyone here. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): Bluish dawn continues outside the closed gate. 안규홍: He carries his tied travel bundle and has tucked the jade ring inside his clothing.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 안규홍 (한국인 남성, 청년 나이대, 검은 머리, 단단히 묶은 머리) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.",
  "roll_prompts": {
   "A": "Create ONE FINAL photorealistic live-action film still of the moment below — early 20th-century Korea, Jeolla region; all people are Korean unless explicitly stated otherwise. TIME OF DAY (lock): dawn, bluish natural light.\n\nSHOT TEXT (authoritative, Korean): 안개가 깔린 산길을 향해 성큼 한쪽 발을 내디딘 mid-stride 순간의 안규홍의 뒷모습.\n\nLOCATION (lock): The outdoor path beyond a modest home's wicket gate, turning toward the misty mountain approach at dawn. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: wide shot\n- KEY BACKGROUND ELEMENTS: 산으로 향하는 길 (The departure route continues toward the mountain) — Recedes diagonally from beneath 안규홍 toward the upper-right background; used as Provides space ahead of the stepping foot and a continuous direction of departure.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: The established bluish dawn light remains subdued and naturalistic, letting separation register without an abrupt change in contrast.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The clothing, hair and overall look of 안규홍 — who appear both in that photo and in this shot — are LOCKED to that photo. Anyone else visible in that photo is NOT in this shot: never carry their face, body or clothing onto anyone here. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): Bluish dawn continues outside the closed gate. 안규홍: He carries his tied travel bundle and has tucked the jade ring inside his clothing.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 안규홍 (한국인 남성, 청년 나이대, 검은 머리, 단단히 묶은 머리) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.",
   "B": "Create ONE FINAL photorealistic live-action film still of the moment below — early 20th-century Korea, Jeolla region; all people are Korean unless explicitly stated otherwise. TIME OF DAY (lock): dawn, bluish natural light.\n\nSHOT TEXT (authoritative, Korean): 안개가 깔린 산길을 향해 성큼 한쪽 발을 내디딘 mid-stride 순간의 안규홍의 뒷모습.\n\nLOCATION (lock): The outdoor path beyond a modest home's wicket gate, turning toward the misty mountain approach at dawn. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: wide shot\n- KEY BACKGROUND ELEMENTS: 산으로 향하는 길 (The departure route continues toward the mountain) — Recedes diagonally from beneath 안규홍 toward the upper-right background; used as Provides space ahead of the stepping foot and a continuous direction of departure.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: The established bluish dawn light remains subdued and naturalistic, letting separation register without an abrupt change in contrast.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The clothing, hair and overall look of 안규홍 — who appear both in that photo and in this shot — are LOCKED to that photo. Anyone else visible in that photo is NOT in this shot: never carry their face, body or clothing onto anyone here. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): Bluish dawn continues outside the closed gate. 안규홍: He carries his tied travel bundle and has tucked the jade ring inside his clothing.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 안규홍 (한국인 남성, 청년 나이대, 검은 머리, 단단히 묶은 머리) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.\n\nCOMPOSITION VARIATION (this candidate only): keep the place, the people and the moment exactly as specified, and keep from the CAMERA & FRAME contract its subject scale, screen placement and key background elements exactly. For this candidate only, its camera angle and height are deliberately left open: choose a camera position distinctly different from the obvious one, as a film director picking a second setup on the same blocking. Stay true to any explicit framing the SHOT TEXT itself dictates."
  },
  "shot_run_spend_attempt_count": 1,
  "readings": [
   {
    "label": "A",
    "direction": "인물은 카메라를 등지고 화면 우측 상단으로 이어지는 산길을 향하고 있음.",
    "built_space": "화면 좌측에 나뭇가지로 엮은 사립문과 흙벽돌 집이 있으며, 안개가 깔린 흙길이 산을 향해 대각선으로 뻗어 있어 이전 샷의 장소와 일치함.",
    "entities": "안규홍의 뒷모습. 거친 흰색 옷과 짚신, 묶은 머리 등 레퍼런스의 특징을 잘 반영함. 등에 봇짐을 지고 있음.",
    "hard_violations": [
     "[gemini-pro] 어깨나 몸을 감싸는 끈이 전혀 없이 등에 봇짐이 마술처럼 붙어 있어 물리적으로 불가능함 (Unsupported object)",
     "[gpt-high] 오른쪽 전경에 참조에 없는 판재 문·울타리와 기둥을 추가하여, 고정되어야 하는 사립문 밖 장소의 설비와 공간 구성을 바꿨다."
    ],
    "physics": "왼발을 뒤로 들어 올려 걷는 순간(mid-stride)을 표현했으나, 등에 있는 봇짐을 지탱하는 어떠한 끈이나 구조물도 존재하지 않아 허공에 떠 있는 상태임."
   },
   {
    "label": "B",
    "direction": "인물은 카메라를 등지고 우측 대각선 방향의 안개 낀 산길을 향해 나아가고 있음.",
    "built_space": "좌측의 사립문과 흙벽, 우측으로 굽어지는 산길 등 레퍼런스에서 설정된 장소의 고정 요소들이 정확히 배치됨.",
    "entities": "안규홍의 뒷모습. 의상, 묶은 머리, 짚신 등 캐릭터 레퍼런스와 일치하며 봇짐을 멘 상태임.",
    "hard_violations": [],
    "physics": "왼발이 땅에서 떨어져 앞으로 나아가는 자연스러운 보행 동작이 묘사됨. 봇짐은 인물의 오른쪽 어깨를 가로지르는 천 끈으로 확실하게 지탱되고 있음."
   }
  ],
  "cross_model_order": {
   "policy": "cross_model_order_v2_slot0_fwd_slot1_rev_parallel2",
   "slots": [
    {
     "model": "gemini-pro",
     "order": "forward",
     "display_to_canonical": {
      "A": "A",
      "B": "B"
     },
     "raw": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 9,
        "verdict_ko": "지시된 장소에서 산길을 향해 성큼 걷는 안규홍의 뒷모습과 봇짐의 형태를 자연스럽고 정확하게 구현했습니다."
       },
       {
        "label": "A",
        "score": 4,
        "verdict_ko": "등에 멘 봇짐을 고정하는 끈이 전혀 보이지 않아 봇짐이 공중에 떠 있는 물리적 오류(Hard Violation)로 인해 크게 감점되었습니다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "인물은 카메라를 등지고 화면 우측 상단으로 이어지는 산길을 향하고 있음.",
        "built_space": "화면 좌측에 나뭇가지로 엮은 사립문과 흙벽돌 집이 있으며, 안개가 깔린 흙길이 산을 향해 대각선으로 뻗어 있어 이전 샷의 장소와 일치함.",
        "entities": "안규홍의 뒷모습. 거친 흰색 옷과 짚신, 묶은 머리 등 레퍼런스의 특징을 잘 반영함. 등에 봇짐을 지고 있음.",
        "hard_violations": [
         "어깨나 몸을 감싸는 끈이 전혀 없이 등에 봇짐이 마술처럼 붙어 있어 물리적으로 불가능함 (Unsupported object)"
        ],
        "physics": "왼발을 뒤로 들어 올려 걷는 순간(mid-stride)을 표현했으나, 등에 있는 봇짐을 지탱하는 어떠한 끈이나 구조물도 존재하지 않아 허공에 떠 있는 상태임."
       },
       {
        "label": "B",
        "direction": "인물은 카메라를 등지고 우측 대각선 방향의 안개 낀 산길을 향해 나아가고 있음.",
        "built_space": "좌측의 사립문과 흙벽, 우측으로 굽어지는 산길 등 레퍼런스에서 설정된 장소의 고정 요소들이 정확히 배치됨.",
        "entities": "안규홍의 뒷모습. 의상, 묶은 머리, 짚신 등 캐릭터 레퍼런스와 일치하며 봇짐을 멘 상태임.",
        "hard_violations": [],
        "physics": "왼발이 땅에서 떨어져 앞으로 나아가는 자연스러운 보행 동작이 묘사됨. 봇짐은 인물의 오른쪽 어깨를 가로지르는 천 끈으로 확실하게 지탱되고 있음."
       }
      ],
      "all_candidates_fail": false
     },
     "normalized": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 9,
        "verdict_ko": "지시된 장소에서 산길을 향해 성큼 걷는 안규홍의 뒷모습과 봇짐의 형태를 자연스럽고 정확하게 구현했습니다."
       },
       {
        "label": "A",
        "score": 4,
        "verdict_ko": "등에 멘 봇짐을 고정하는 끈이 전혀 보이지 않아 봇짐이 공중에 떠 있는 물리적 오류(Hard Violation)로 인해 크게 감점되었습니다."
       }
      ],
      "all_candidates_fail": false,
      "readings": [
       {
        "label": "A",
        "direction": "인물은 카메라를 등지고 화면 우측 상단으로 이어지는 산길을 향하고 있음.",
        "built_space": "화면 좌측에 나뭇가지로 엮은 사립문과 흙벽돌 집이 있으며, 안개가 깔린 흙길이 산을 향해 대각선으로 뻗어 있어 이전 샷의 장소와 일치함.",
        "entities": "안규홍의 뒷모습. 거친 흰색 옷과 짚신, 묶은 머리 등 레퍼런스의 특징을 잘 반영함. 등에 봇짐을 지고 있음.",
        "hard_violations": [
         "어깨나 몸을 감싸는 끈이 전혀 없이 등에 봇짐이 마술처럼 붙어 있어 물리적으로 불가능함 (Unsupported object)"
        ],
        "physics": "왼발을 뒤로 들어 올려 걷는 순간(mid-stride)을 표현했으나, 등에 있는 봇짐을 지탱하는 어떠한 끈이나 구조물도 존재하지 않아 허공에 떠 있는 상태임."
       },
       {
        "label": "B",
        "direction": "인물은 카메라를 등지고 우측 대각선 방향의 안개 낀 산길을 향해 나아가고 있음.",
        "built_space": "좌측의 사립문과 흙벽, 우측으로 굽어지는 산길 등 레퍼런스에서 설정된 장소의 고정 요소들이 정확히 배치됨.",
        "entities": "안규홍의 뒷모습. 의상, 묶은 머리, 짚신 등 캐릭터 레퍼런스와 일치하며 봇짐을 멘 상태임.",
        "hard_violations": [],
        "physics": "왼발이 땅에서 떨어져 앞으로 나아가는 자연스러운 보행 동작이 묘사됨. 봇짐은 인물의 오른쪽 어깨를 가로지르는 천 끈으로 확실하게 지탱되고 있음."
       }
      ]
     },
     "ok": true
    },
    {
     "model": "gpt-high",
     "order": "reverse",
     "display_to_canonical": {
      "A": "B",
      "B": "A"
     },
     "raw": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 8,
        "verdict_ko": "닫힌 사립문 밖에서 산길로 떠나는 뒷모습과 장소 연속성을 잘 지켰으나, 성큼 내딛는 보폭과 우상향 출발 동선은 다소 약하다."
       },
       {
        "label": "B",
        "score": 4,
        "verdict_ko": "우상향 산길과 와이드 구도는 명확하지만, 오른쪽 전경에 참조에 없는 판재 문·울타리 구조물을 추가해 고정된 장소를 바꿨다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "남성은 카메라에 등을 보이고 머리와 몸을 앞쪽 산길로 향한다. 눈은 보이지 않아 정확한 시선은 확인할 수 없다. 발밑 길은 인물 오른쪽으로 이어진 뒤 안개 속에서 굽어지므로 산으로 떠나는 방향은 성립하지만, 우상단으로 곧게 이어지는 대각선은 덜 선명하다. 겨누거나 가리키는 물건은 없다.",
        "built_space": "왼쪽에 닫힌 나뭇가지 사립문 한 개와 굵은 양쪽 문기둥, 낮은 엮은 울타리, 흙벽과 지붕 일부가 보인다. 남성은 문 바깥 흙길에 서 있으며 문이나 울타리와 겹쳐 관통하지 않는다. 참조의 거친 목재와 흙벽, 길 가장자리 관목이 유지된다. 지면 가까운 뒤쪽 시점의 전신 구도로, 인물이 화면 높이를 비교적 크게 차지한다. 반사면은 없다.",
        "entities": "보이는 사람은 안규홍으로 설정된 젊은 성인 남성 한 명뿐이다. 뒷모습이라 얼굴 동일성과 정확한 나이·민족성은 직접 검증할 수 없지만, 검은 묶음머리와 체격은 참조에 부합한다. 때 묻은 밝은색 저고리와 바지, 발목 매듭과 짚신이 보인다. 등에 매듭지은 천 보따리 한 개를 메고 있으며 옥가락지는 노출되지 않는다. 안개 낀 산길과 푸른 새벽빛이 있고, 읽을 수 있는 글자나 추가 인물은 없다.",
        "hard_violations": [],
        "physics": "오른발이 흙길에 닿아 체중을 지탱하고 왼발은 뒤에서 들려 밑창을 드러낸다. 지지발과 다리의 연결이 자연스러워 실제 보행 중간 순간으로 가능하다. 다만 큰 보폭보다는 평범한 걸음에 가깝다. 보따리는 어깨를 넘는 천 끈으로 지지되고 등에 닿아 있어 떠 있지 않다."
       },
       {
        "label": "B",
        "direction": "남성의 등과 머리는 앞쪽 길을 향하고, 길은 발밑에서 오른쪽 위 안개 낀 산으로 연속해서 이어진다. 출발 방향과 길의 목적지가 A보다 명료하다. 눈은 뒤로 가려져 정확한 시선은 확인되지 않으며, 겨누는 물건은 없다.",
        "built_space": "왼쪽에는 닫힌 나뭇가지 사립문 한 개, 양쪽 문기둥, 엮은 울타리와 흙집이 있다. 그런데 오른쪽 전경에 각재 기둥과 수평·수직 판재로 만든 별도의 문 또는 울타리 구조물이 크게 추가되어 있다. 참조에서 그쪽은 관목과 길 가장자리로 보이며 이러한 설비가 없다. 남성은 기존 사립문 밖 길에 있고, 전신과 산길을 함께 담은 와이드 구도는 적절하다. 반사면은 없다.",
        "entities": "젊은 성인 남성 한 명만 등장하며, 검게 묶은 머리와 밝은색 낡은 저고리·바지, 발목 매듭과 짚신은 참조와 대체로 맞는다. 얼굴은 보이지 않아 얼굴 동일성이나 민족성을 직접 확인할 수 없다. 매듭지은 천 보따리 한 개를 등에 메고 있고 옥가락지는 보이지 않는다. 산과 안개, 차분한 푸른 새벽빛은 충족한다. 글자와 추가 인물은 없지만 참조에 없는 오른쪽 목재 구조물이 추가되었다.",
        "hard_violations": [
         "오른쪽 전경에 참조에 없는 판재 문·울타리와 기둥을 추가하여, 고정되어야 하는 사립문 밖 장소의 설비와 공간 구성을 바꿨다."
        ],
        "physics": "오른발은 지면을 딛고 왼발은 뒤로 들려 있어 지지점이 분명한 보행 자세다. 몸이 공중에 떠 있거나 팔다리가 불가능하게 연결된 부분은 없다. 보폭은 성큼 내딛기보다는 보통 걸음에 가깝다. 보따리는 양쪽 어깨로 이어지는 끈이 무게를 받으며 등에 기대어 있다."
       }
      ],
      "all_candidates_fail": false
     },
     "normalized": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 8,
        "verdict_ko": "닫힌 사립문 밖에서 산길로 떠나는 뒷모습과 장소 연속성을 잘 지켰으나, 성큼 내딛는 보폭과 우상향 출발 동선은 다소 약하다."
       },
       {
        "label": "A",
        "score": 4,
        "verdict_ko": "우상향 산길과 와이드 구도는 명확하지만, 오른쪽 전경에 참조에 없는 판재 문·울타리 구조물을 추가해 고정된 장소를 바꿨다."
       }
      ],
      "all_candidates_fail": false,
      "readings": [
       {
        "label": "B",
        "direction": "남성은 카메라에 등을 보이고 머리와 몸을 앞쪽 산길로 향한다. 눈은 보이지 않아 정확한 시선은 확인할 수 없다. 발밑 길은 인물 오른쪽으로 이어진 뒤 안개 속에서 굽어지므로 산으로 떠나는 방향은 성립하지만, 우상단으로 곧게 이어지는 대각선은 덜 선명하다. 겨누거나 가리키는 물건은 없다.",
        "built_space": "왼쪽에 닫힌 나뭇가지 사립문 한 개와 굵은 양쪽 문기둥, 낮은 엮은 울타리, 흙벽과 지붕 일부가 보인다. 남성은 문 바깥 흙길에 서 있으며 문이나 울타리와 겹쳐 관통하지 않는다. 참조의 거친 목재와 흙벽, 길 가장자리 관목이 유지된다. 지면 가까운 뒤쪽 시점의 전신 구도로, 인물이 화면 높이를 비교적 크게 차지한다. 반사면은 없다.",
        "entities": "보이는 사람은 안규홍으로 설정된 젊은 성인 남성 한 명뿐이다. 뒷모습이라 얼굴 동일성과 정확한 나이·민족성은 직접 검증할 수 없지만, 검은 묶음머리와 체격은 참조에 부합한다. 때 묻은 밝은색 저고리와 바지, 발목 매듭과 짚신이 보인다. 등에 매듭지은 천 보따리 한 개를 메고 있으며 옥가락지는 노출되지 않는다. 안개 낀 산길과 푸른 새벽빛이 있고, 읽을 수 있는 글자나 추가 인물은 없다.",
        "hard_violations": [],
        "physics": "오른발이 흙길에 닿아 체중을 지탱하고 왼발은 뒤에서 들려 밑창을 드러낸다. 지지발과 다리의 연결이 자연스러워 실제 보행 중간 순간으로 가능하다. 다만 큰 보폭보다는 평범한 걸음에 가깝다. 보따리는 어깨를 넘는 천 끈으로 지지되고 등에 닿아 있어 떠 있지 않다."
       },
       {
        "label": "A",
        "direction": "남성의 등과 머리는 앞쪽 길을 향하고, 길은 발밑에서 오른쪽 위 안개 낀 산으로 연속해서 이어진다. 출발 방향과 길의 목적지가 A보다 명료하다. 눈은 뒤로 가려져 정확한 시선은 확인되지 않으며, 겨누는 물건은 없다.",
        "built_space": "왼쪽에는 닫힌 나뭇가지 사립문 한 개, 양쪽 문기둥, 엮은 울타리와 흙집이 있다. 그런데 오른쪽 전경에 각재 기둥과 수평·수직 판재로 만든 별도의 문 또는 울타리 구조물이 크게 추가되어 있다. 참조에서 그쪽은 관목과 길 가장자리로 보이며 이러한 설비가 없다. 남성은 기존 사립문 밖 길에 있고, 전신과 산길을 함께 담은 와이드 구도는 적절하다. 반사면은 없다.",
        "entities": "젊은 성인 남성 한 명만 등장하며, 검게 묶은 머리와 밝은색 낡은 저고리·바지, 발목 매듭과 짚신은 참조와 대체로 맞는다. 얼굴은 보이지 않아 얼굴 동일성이나 민족성을 직접 확인할 수 없다. 매듭지은 천 보따리 한 개를 등에 메고 있고 옥가락지는 보이지 않는다. 산과 안개, 차분한 푸른 새벽빛은 충족한다. 글자와 추가 인물은 없지만 참조에 없는 오른쪽 목재 구조물이 추가되었다.",
        "hard_violations": [
         "오른쪽 전경에 참조에 없는 판재 문·울타리와 기둥을 추가하여, 고정되어야 하는 사립문 밖 장소의 설비와 공간 구성을 바꿨다."
        ],
        "physics": "오른발은 지면을 딛고 왼발은 뒤로 들려 있어 지지점이 분명한 보행 자세다. 몸이 공중에 떠 있거나 팔다리가 불가능하게 연결된 부분은 없다. 보폭은 성큼 내딛기보다는 보통 걸음에 가깝다. 보따리는 양쪽 어깨로 이어지는 끈이 무게를 받으며 등에 기대어 있다."
       }
      ]
     },
     "ok": true
    }
   ],
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "slot_winner_match": true,
   "slot_winner": {
    "gemini-pro": "B",
    "gpt-high": "B"
   },
   "route": "cross_slot_agree"
  },
  "dual": {
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "normalized": {
    "A": 0.944,
    "B": 2.0
   },
   "adjusted": {
    "A": 0.694,
    "B": 2.0
   },
   "violations": {
    "A": [
     "[gemini-pro] 어깨나 몸을 감싸는 끈이 전혀 없이 등에 봇짐이 마술처럼 붙어 있어 물리적으로 불가능함 (Unsupported object)",
     "[gpt-high] 오른쪽 전경에 참조에 없는 판재 문·울타리와 기둥을 추가하여, 고정되어야 하는 사립문 밖 장소의 설비와 공간 구성을 바꿨다."
    ]
   },
   "per_model_winner": {
    "gemini-pro": "B",
    "gpt-high": "B"
   },
   "agreed": true
  },
  "totals": {
   "B": 2000,
   "A": 694
  },
  "selected": "B",
  "ranking": [
   "B",
   "A"
  ],
  "verdicts": [
   {
    "label": "B",
    "score": 2000,
    "verdict_ko": "지시된 장소에서 산길을 향해 성큼 걷는 안규홍의 뒷모습과 봇짐의 형태를 자연스럽고 정확하게 구현했습니다."
   },
   {
    "label": "A",
    "score": 694,
    "verdict_ko": "등에 멘 봇짐을 고정하는 끈이 전혀 보이지 않아 봇짐이 공중에 떠 있는 물리적 오류(Hard Violation)로 인해 크게 감점되었습니다.  ★위반: [gemini-pro] 어깨나 몸을 감싸는 끈이 전혀 없이 등에 봇짐이 마술처럼 붙어 있어 물리적으로 불가능함 (Unsupported object) / [gpt-high] 오른쪽 전경에 참조에 없는 판재 문·울타리와 기둥을 추가하여, 고정되어야 하는 사립문 밖 장소의 설비와 공간 구성을 바꿨다."
   }
  ],
  "refs": [
   {
    "label": "PREVIOUS SHOT STILL — a visually related earlier shot of this same place: the location's look, materials, fixed features and lighting mood are LOCKED to this photo; never copy its camera framing. The clothing, hair and overall look of 안규홍 — who appear both in that photo and in this shot — are LOCKED to that photo. Anyone else visible in that photo is NOT in this shot: never carry their face, body or clothing onto anyone here. If this photo shows a character who cannot move (dead or unconscious), that character's exact pose, position and orientation are ALSO LOCKED — treat their body as a fixed prop of the set that only the camera moves around.",
    "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/db00517b-165b-4005-a287-f28936b59b95/images/272bd7af-9294-4490-98c6-88a40e9deacc/scene/recipe/S11sh9_sel.png",
    "asset_id": "e390f3bf-6449-4a91-800a-11a851b2a6d7",
    "role": "prev_still"
   },
   {
    "label": "CHARACTER REFERENCE — 안규홍: the exact person appearing in this shot; match face, hair and build exactly.",
    "path": "<bytes:1028384>",
    "asset_id": "3b5894db-ce55-4f34-b966-5e725c3aa952",
    "role": "character_ref"
   }
  ],
  "critique_skipped": true,
  "shot_run_produced": false,
  "shot_run_uid": "06aa140f-0f41-7248-88ca-d6fca9e58070",
  "ref_mode": "prev+엔티티",
  "share_plan": {
   "ref_plan": "prev",
   "prev_anchor_tag": "S11sh9"
  }
 },
 "S11sh16::cine": {
  "applied": true,
  "attempted_at": "2026-09-09T01:54:56.505945+00:00",
  "fingerprint": "5330a32e3ca77abf2d31b642ad3b56bd40241be66953f40db7fd6d575ecdece0",
  "fingerprint_version": 2,
  "provider": "mai",
  "endpoint": "openrouter/chat-completions",
  "model": "microsoft/mai-image-2.6",
  "pack": "24.202608252115",
  "source_file": "S11sh16_sel.png",
  "source_sha256": "ec7033689e57be24e45a0a06fd4561b8ac869d8d4722b3c183603a002acba53f",
  "file": "S11sh16_cine.png",
  "staged_sha256": "287b70c02ac79df26c9623985dfc5bac6150d5d7010b5fcccc903c349000b98d",
  "latency_ms": 24393
 },
 "S12sh1::signage": {
  "fp": "467250bc9886a3ed",
  "inscriptions": [],
  "cues": [],
  "dropped": []
 },
 "era_assess::4571bb01c1f7ad91": {
  "subjects": [],
  "subject_text": "논둑길\n논 사이로 좁게 이어지는 흙 논둑길. 낮은 논둑의 선이 산 아래로 이어지고, 새벽에는 엷은 빛과 안개가 깔린다.",
  "identity": "canonical",
  "scope_id": "L13",
  "scope_role": "location_exterior",
  "scope_sha": "d896f91c73afb0ca"
 },
 "groupbg::paddy_dyke": {
  "input_fingerprint": "d191aac2fed16cb9",
  "meta": {
   "model": "gpt-image-2",
   "size": "1536x864",
   "pack": "11.202607220237",
   "contract": "bgfirst_full_v3",
   "group_sig": {
    "key": "paddy_dyke",
    "tags": [
     "S12sh1"
    ]
   },
   "context_sig": "fd5ec5c9cee0f0a9"
  },
  "prompt": "Create ONE empty live-action location background photograph — NO PEOPLE, no figures, no body parts, no silhouettes, no shadows or reflections of people anywhere.\n\"Empty\" means no people only: KEEP the location's inherent occupants and stock that define the place — animals in an animal shelter, pen or farm, goods and produce in a market, moored boats in a harbour — unless the shot text explicitly removes them.\nThe FIRST attached image is a thin-line storyboard sketch of a shot that happens at this location: use it ONLY as spatial evidence — what this place contains, how its ground, structures and landmarks are arranged and proportioned. Ignore the sketched people and arrows entirely, and do NOT copy its line style: render a fully photographic, physically plausible real place that fits THE LOCATION text below.\nHUMAN-SCALE CALIBRATION: derive every structure's true size from human-scale elements — a door ≈ 2m, a window ≈ 1–1.5m wide, one storey ≈ 2.5–3m; never inflate a small structure or shrink a large one.\n\nTHE LOCATION — early 20th-century Korea, Jeolla region; all people are Korean unless explicitly stated otherwise: An exposed earthen footpath along rice paddies, enveloped in dense dawn fog.\n\nLOCATION DETAIL (descriptive evidence about this place from the production's location records — evidence, not a staging order): use it to understand what kind of place this is and what it permanently contains. Carry over ONLY its enduring physical features — layout, structures, ground and materials, fixed equipment and installations. Do NOT treat any lighting, weather or time-of-day wording, momentary object states, depicted events or subjective impressions in it as fixed facts of the place; the TIME OF DAY lock below is the sole authority for time and lighting.\n논둑길: 안개가 자욱하게 깔린 새벽, 양옆으로 농경지가 펼쳐진 다져진 흙 논둑길. (특징: 자욱한 안개; 일자로 뻗은 흙 논둑; 양옆의 빈 논밭 지형)\n\nSCENE EVIDENCE (verbatim quotes from the screenplay about this place — treat them as evidence of what the location physically contains and looks like; stage the PLACE those moments happen in, but do NOT depict the momentary actions, people or staged props themselves):\n- 안개가 자욱한 새벽 논둑길.\n규홍이 봇짐을 어깨에 메고 묵묵히 걷는다.\n\nTIME OF DAY (lock): dawn, dense fog.\n\nRender ONE photorealistic empty location photograph, 16:9, neutral enough that every shot of this place can be staged from it later. No readable writing anywhere: surfaces that would carry writing may be present, but stage any wording out of legibility — an oblique angle, distance, shallow focus. No captions, watermarks or overlay text.",
  "effective_prompt": "Create ONE empty live-action location background photograph — NO PEOPLE, no figures, no body parts, no silhouettes, no shadows or reflections of people anywhere.\n\"Empty\" means no people only: KEEP the location's inherent occupants and stock that define the place — animals in an animal shelter, pen or farm, goods and produce in a market, moored boats in a harbour — unless the shot text explicitly removes them.\nThe FIRST attached image is a thin-line storyboard sketch of a shot that happens at this location: use it ONLY as spatial evidence — what this place contains, how its ground, structures and landmarks are arranged and proportioned. Ignore the sketched people and arrows entirely, and do NOT copy its line style: render a fully photographic, physically plausible real place that fits THE LOCATION text below.\nHUMAN-SCALE CALIBRATION: derive every structure's true size from human-scale elements — a door ≈ 2m, a window ≈ 1–1.5m wide, one storey ≈ 2.5–3m; never inflate a small structure or shrink a large one.\n\nTHE LOCATION — early 20th-century Korea, Jeolla region; all people are Korean unless explicitly stated otherwise: An exposed earthen footpath along rice paddies, enveloped in dense dawn fog.\n\nLOCATION DETAIL (descriptive evidence about this place from the production's location records — evidence, not a staging order): use it to understand what kind of place this is and what it permanently contains. Carry over ONLY its enduring physical features — layout, structures, ground and materials, fixed equipment and installations. Do NOT treat any lighting, weather or time-of-day wording, momentary object states, depicted events or subjective impressions in it as fixed facts of the place; the TIME OF DAY lock below is the sole authority for time and lighting.\n논둑길: 안개가 자욱하게 깔린 새벽, 양옆으로 농경지가 펼쳐진 다져진 흙 논둑길. (특징: 자욱한 안개; 일자로 뻗은 흙 논둑; 양옆의 빈 논밭 지형)\n\nSCENE EVIDENCE (verbatim quotes from the screenplay about this place — treat them as evidence of what the location physically contains and looks like; stage the PLACE those moments happen in, but do NOT depict the momentary actions, people or staged props themselves):\n- 안개가 자욱한 새벽 논둑길.\n규홍이 봇짐을 어깨에 메고 묵묵히 걷는다.\n\nTIME OF DAY (lock): dawn, dense fog.\n\nRender ONE photorealistic empty location photograph, 16:9, neutral enough that every shot of this place can be staged from it later. No readable writing anywhere: surfaces that would carry writing may be present, but stage any wording out of legibility — an oblique angle, distance, shallow focus. No captions, watermarks or overlay text.",
  "bg_path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/db00517b-165b-4005-a287-f28936b59b95/images/272bd7af-9294-4490-98c6-88a40e9deacc/scene/recipe/groupbg_paddy_dyke_98913c.png",
  "asset_id": "768752bf-2da6-43b2-9b4e-2428397a7a10",
  "input_asset_ids": [
   "338fb4b6-2ab4-4852-b958-6a51925690f3"
  ],
  "origin_tag": "S12sh1",
  "place_text": "An exposed earthen footpath along rice paddies, enveloped in dense dawn fog.",
  "origin_inputs": {
   "place_text": "An exposed earthen footpath along rice paddies, enveloped in dense dawn fog.",
   "time_of_day_en": "dawn, dense fog",
   "conti_asset_id": "338fb4b6-2ab4-4852-b958-6a51925690f3"
  }
 },
 "S12sh1::bgfirst_bg": {
  "input_fingerprint": "db57c38873f7834e",
  "prompt": "Create the EMPTY BACKGROUND PLATE for one film shot — NO PEOPLE, no figures, no body parts, no sketch lines, no arrows anywhere.\n\"Empty\" means no people only: KEEP the location's inherent occupants and stock that define the place — animals in an animal shelter, pen or farm, goods in a market, moored boats in a harbour — unless the shot text explicitly removes them.\nThe FIRST attached image is a thin-line storyboard sketch: use ONLY its camera angle, horizon, perspective and the placement/size of buildings and set masses — ignore the sketched people and arrows entirely. The SECOND attached image (LOCATION PHOTOGRAPH) is the real place: take its architecture, materials, signage and fixed features, and RE-PROJECT them into the sketch's camera. If the photograph's camera differs from the sketch's, the sketch's camera wins.\nHUMAN-SCALE CALIBRATION: derive every structure's true size from human-scale elements — a door ≈ 2m, a window ≈ 1–1.5m wide, one storey ≈ 2.5–3m; never inflate a small structure or shrink a large one.\n\nSHOT TEXT this background must serve (Korean): 짙은 안개가 자욱하게 깔린 새벽 논둑길, 봇짐을 진 채 앞으로 한 발을 뻗어 성큼 내디딘 미드 액션 찰나의 안규홍 전신.\n\nLOCATION (lock): An exposed earthen footpath along rice paddies, enveloped in dense dawn fog.\n\nTIME OF DAY (lock): dawn, dense fog.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: wide shot\n- KEY BACKGROUND ELEMENTS: 논둑길 (Fog obscures the farther reaches of the path; no following figure is yet visible) — The path passes diagonally through the frame with a visible section behind 안규홍; used as Establishes the solitary walk while reserving space for the later arrival; 안규홍의 봇짐 (Carried on his shoulder while he walks) — Its rear and outer side are visible beside his upper back; used as A modest silhouette detail anchoring the departure continuity.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Dawn light diffused by the explicitly established fog lowers distant contrast while keeping 안규홍's figure legible in a muted palette.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nRender ONE photorealistic empty location photograph, 16:9, that this shot can be staged inside later. No readable writing anywhere: surfaces that would carry writing may be present, but stage any wording out of legibility — an oblique angle, distance, shallow focus. No captions, watermarks or overlay text.",
  "effective_prompt": "Create the EMPTY BACKGROUND PLATE for one film shot — NO PEOPLE, no figures, no body parts, no sketch lines, no arrows anywhere.\n\"Empty\" means no people only: KEEP the location's inherent occupants and stock that define the place — animals in an animal shelter, pen or farm, goods in a market, moored boats in a harbour — unless the shot text explicitly removes them.\nThe FIRST attached image is a thin-line storyboard sketch: use ONLY its camera angle, horizon, perspective and the placement/size of buildings and set masses — ignore the sketched people and arrows entirely. The SECOND attached image (LOCATION PHOTOGRAPH) is the real place: take its architecture, materials, signage and fixed features, and RE-PROJECT them into the sketch's camera. If the photograph's camera differs from the sketch's, the sketch's camera wins.\nHUMAN-SCALE CALIBRATION: derive every structure's true size from human-scale elements — a door ≈ 2m, a window ≈ 1–1.5m wide, one storey ≈ 2.5–3m; never inflate a small structure or shrink a large one.\n\nSHOT TEXT this background must serve (Korean): 짙은 안개가 자욱하게 깔린 새벽 논둑길, 봇짐을 진 채 앞으로 한 발을 뻗어 성큼 내디딘 미드 액션 찰나의 안규홍 전신.\n\nLOCATION (lock): An exposed earthen footpath along rice paddies, enveloped in dense dawn fog.\n\nTIME OF DAY (lock): dawn, dense fog.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: wide shot\n- KEY BACKGROUND ELEMENTS: 논둑길 (Fog obscures the farther reaches of the path; no following figure is yet visible) — The path passes diagonally through the frame with a visible section behind 안규홍; used as Establishes the solitary walk while reserving space for the later arrival; 안규홍의 봇짐 (Carried on his shoulder while he walks) — Its rear and outer side are visible beside his upper back; used as A modest silhouette detail anchoring the departure continuity.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Dawn light diffused by the explicitly established fog lowers distant contrast while keeping 안규홍's figure legible in a muted palette.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nRender ONE photorealistic empty location photograph, 16:9, that this shot can be staged inside later. No readable writing anywhere: surfaces that would carry writing may be present, but stage any wording out of legibility — an oblique angle, distance, shallow focus. No captions, watermarks or overlay text.",
  "bg_path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/db00517b-165b-4005-a287-f28936b59b95/images/272bd7af-9294-4490-98c6-88a40e9deacc/scene/recipe/S12sh1__bgfirst_bg.png",
  "asset_id": "096965c6-37fe-4feb-8cac-99ce4b2541bf",
  "input_asset_ids": [
   "338fb4b6-2ab4-4852-b958-6a51925690f3",
   "768752bf-2da6-43b2-9b4e-2428397a7a10"
  ]
 },
 "S12sh1": {
  "input_fingerprint": "0110c1036be90f31",
  "prompt": "Create ONE FINAL photorealistic live-action film still of the moment below — early 20th-century Korea, Jeolla region; all people are Korean unless explicitly stated otherwise. TIME OF DAY (lock): dawn, dense fog.\n\nSHOT TEXT (authoritative, Korean): 짙은 안개가 자욱하게 깔린 새벽 논둑길, 봇짐을 진 채 앞으로 한 발을 뻗어 성큼 내디딘 미드 액션 찰나의 안규홍 전신.\n\nLOCATION (lock): An exposed earthen footpath along rice paddies, enveloped in dense dawn fog. The shot takes place here — the attached LOCATION PHOTOGRAPH shows the exact spot.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: wide shot\n- KEY BACKGROUND ELEMENTS: 논둑길 (Fog obscures the farther reaches of the path; no following figure is yet visible) — The path passes diagonally through the frame with a visible section behind 안규홍; used as Establishes the solitary walk while reserving space for the later arrival; 안규홍의 봇짐 (Carried on his shoulder while he walks) — Its rear and outer side are visible beside his upper back; used as A modest silhouette detail anchoring the departure continuity.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Dawn light diffused by the explicitly established fog lowers distant contrast while keeping 안규홍's figure legible in a muted palette.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): Dense dawn fog blankets the rice-field embankment path. 안규홍: He walks with a travel bundle on his shoulder and the jade ring concealed inside his clothing.\n\nTHE HAND THAT IS DOING THIS: the object at the centre of this shot is being held, operated, read, aimed or handed over by 안규홍 right now, so 안규홍's hand — and as much of the wrist and forearm as the framing reaches — is in the frame, gripping or resting on that object exactly the way its real use requires. Match that hand to 안규홍: its size, build, skin, age, grooming, sleeve and anything worn on it belong to that person and to no one else.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 안규홍 (한국인 남성, 청년 나이대, 검은 머리, 단단히 묶은 머리) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.",
  "roll_prompts": {
   "A": "Stage the shot. The FIRST attached image (SHOT BACKGROUND) is the finished empty background of this shot — keep it EXACTLY: its camera, perspective, architecture, lighting and every fixed feature stay untouched. The SECOND attached image (LAYOUT SKETCH) tells you ONLY where the people go: each sketched person's position, screen size, pose and the gaze/motion arrows. Ignore the sketch's background lines. The CHARACTER REFERENCE photographs show the real people.\nPlace the real people into the background at exactly the sketched positions, sizes and poses, following the arrow directions. No sketch lines or arrows may remain.\n\nCreate ONE FINAL photorealistic live-action film still of the moment below — early 20th-century Korea, Jeolla region; all people are Korean unless explicitly stated otherwise. TIME OF DAY (lock): dawn, dense fog.\n\nSHOT TEXT (authoritative, Korean): 짙은 안개가 자욱하게 깔린 새벽 논둑길, 봇짐을 진 채 앞으로 한 발을 뻗어 성큼 내디딘 미드 액션 찰나의 안규홍 전신.\n\nLOCATION (lock): An exposed earthen footpath along rice paddies, enveloped in dense dawn fog. The shot takes place here — the FIRST attached image (SHOT BACKGROUND) is this exact place, already built: its ground, structures, horizon, materials and lighting are the finished truth of this location and must not be redesigned or replaced. No location photograph is attached — read the place from that image alone, and add no scenery, structure, vehicle or fixture that it does not already show. This lock governs the place only; the figures in the shot follow the staging and pose instructions.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: wide shot\n- KEY BACKGROUND ELEMENTS: 논둑길 (Fog obscures the farther reaches of the path; no following figure is yet visible) — The path passes diagonally through the frame with a visible section behind 안규홍; used as Establishes the solitary walk while reserving space for the later arrival; 안규홍의 봇짐 (Carried on his shoulder while he walks) — Its rear and outer side are visible beside his upper back; used as A modest silhouette detail anchoring the departure continuity.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Dawn light diffused by the explicitly established fog lowers distant contrast while keeping 안규홍's figure legible in a muted palette.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): Dense dawn fog blankets the rice-field embankment path. 안규홍: He walks with a travel bundle on his shoulder and the jade ring concealed inside his clothing.\n\nTHE HAND THAT IS DOING THIS: the object at the centre of this shot is being held, operated, read, aimed or handed over by 안규홍 right now, so 안규홍's hand — and as much of the wrist and forearm as the framing reaches — is in the frame, gripping or resting on that object exactly the way its real use requires. Match that hand to 안규홍: its size, build, skin, age, grooming, sleeve and anything worn on it belong to that person and to no one else.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 안규홍 (한국인 남성, 청년 나이대, 검은 머리, 단단히 묶은 머리) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.",
   "B": "Create ONE FINAL photorealistic live-action film still of the moment below — early 20th-century Korea, Jeolla region; all people are Korean unless explicitly stated otherwise. TIME OF DAY (lock): dawn, dense fog.\n\nSHOT TEXT (authoritative, Korean): 짙은 안개가 자욱하게 깔린 새벽 논둑길, 봇짐을 진 채 앞으로 한 발을 뻗어 성큼 내디딘 미드 액션 찰나의 안규홍 전신.\n\nLOCATION (lock): An exposed earthen footpath along rice paddies, enveloped in dense dawn fog. The shot takes place here — the attached LOCATION PHOTOGRAPH shows the exact spot.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: wide shot\n- KEY BACKGROUND ELEMENTS: 논둑길 (Fog obscures the farther reaches of the path; no following figure is yet visible) — The path passes diagonally through the frame with a visible section behind 안규홍; used as Establishes the solitary walk while reserving space for the later arrival; 안규홍의 봇짐 (Carried on his shoulder while he walks) — Its rear and outer side are visible beside his upper back; used as A modest silhouette detail anchoring the departure continuity.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Dawn light diffused by the explicitly established fog lowers distant contrast while keeping 안규홍's figure legible in a muted palette.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): Dense dawn fog blankets the rice-field embankment path. 안규홍: He walks with a travel bundle on his shoulder and the jade ring concealed inside his clothing.\n\nTHE HAND THAT IS DOING THIS: the object at the centre of this shot is being held, operated, read, aimed or handed over by 안규홍 right now, so 안규홍's hand — and as much of the wrist and forearm as the framing reaches — is in the frame, gripping or resting on that object exactly the way its real use requires. Match that hand to 안규홍: its size, build, skin, age, grooming, sleeve and anything worn on it belong to that person and to no one else.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 안규홍 (한국인 남성, 청년 나이대, 검은 머리, 단단히 묶은 머리) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.\n\nCOMPOSITION VARIATION (this candidate only): keep the place, the people and the moment exactly as specified, and keep from the CAMERA & FRAME contract its subject scale, screen placement and key background elements exactly. For this candidate only, its camera angle and height are deliberately left open: choose a camera position distinctly different from the obvious one, as a film director picking a second setup on the same blocking. Stay true to any explicit framing the SHOT TEXT itself dictates."
  },
  "roll_refs": {
   "A": [
    {
     "label": "SHOT BACKGROUND",
     "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/db00517b-165b-4005-a287-f28936b59b95/images/272bd7af-9294-4490-98c6-88a40e9deacc/scene/recipe/S12sh1__bgfirst_bg.png",
     "asset_id": "096965c6-37fe-4feb-8cac-99ce4b2541bf",
     "role": "bgfirst_bg"
    },
    {
     "label": "LAYOUT SKETCH (people placement only)",
     "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/db00517b-165b-4005-a287-f28936b59b95/images/272bd7af-9294-4490-98c6-88a40e9deacc/conti/conti_S12sh1.png",
     "asset_id": "338fb4b6-2ab4-4852-b958-6a51925690f3",
     "role": "conti_light"
    },
    {
     "label": "CHARACTER REFERENCE — 안규홍: the exact person appearing in this shot; match face, hair and build exactly.",
     "path": "<bytes:1028384>",
     "asset_id": "3b5894db-ce55-4f34-b966-5e725c3aa952",
     "role": "character_ref"
    },
    {
     "label": "PROP REFERENCE — 보자기 꾸러미·봇짐: the exact object appearing in this shot; match its look, material and wear exactly.",
     "path": "<bytes:1465734>",
     "asset_id": "3ca92b24-68b4-4883-b851-cabc06c00d1a",
     "role": "prop_ref"
    }
   ],
   "B": [
    {
     "label": "LOCATION PHOTOGRAPH — the exact place of this shot: its architecture, materials, fixed features and lighting mood are spatial truth; stage the moment inside this place. Never copy its camera framing.",
     "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/db00517b-165b-4005-a287-f28936b59b95/images/272bd7af-9294-4490-98c6-88a40e9deacc/scene/recipe/groupbg_paddy_dyke_98913c.png",
     "asset_id": "768752bf-2da6-43b2-9b4e-2428397a7a10",
     "role": "bgfirst_group_bg"
    },
    {
     "label": "CHARACTER REFERENCE — 안규홍: the exact person appearing in this shot; match face, hair and build exactly.",
     "path": "<bytes:1028384>",
     "asset_id": "3b5894db-ce55-4f34-b966-5e725c3aa952",
     "role": "character_ref"
    },
    {
     "label": "PROP REFERENCE — 보자기 꾸러미·봇짐: the exact object appearing in this shot; match its look, material and wear exactly.",
     "path": "<bytes:1465734>",
     "asset_id": "3ca92b24-68b4-4883-b851-cabc06c00d1a",
     "role": "prop_ref"
    }
   ]
  },
  "shot_run_spend_attempt_count": 1,
  "judge_flip_skipped": "judge_fn_owns_order",
  "readings": [
   {
    "label": "A",
    "direction": "시선이 카메라 정면을 향함.",
    "built_space": "레퍼런스와 일치하는 안개 낀 논둑길 중앙에 서 있음.",
    "entities": "안규홍의 얼굴, 의상, 짚신 일치. 양쪽 어깨 뒤로 녹색 봇짐을 메고 끈을 쥐고 있음.",
    "hard_violations": [],
    "physics": "양발이 땅에 닿아 있으며, 보폭이 좁고 정적인 일반적인 걷기 자세임."
   },
   {
    "label": "B",
    "direction": "시선과 몸의 진행 방향이 프레임 우측을 향하고 있음.",
    "built_space": "레퍼런스와 일치하는 안개 낀 논둑길. 인물이 대각선 길 위에 자연스럽게 배치됨.",
    "entities": "안규홍의 얼굴(측면), 의상, 짚신 일치. 녹색 봇짐을 등에 메고 있어 후면과 측면이 잘 보임.",
    "hard_violations": [],
    "physics": "앞으로 다리를 길게 뻗은 보행 찰나의 역동적인 자세로, 발의 지지와 무게 중심이 자연스러움."
   }
  ],
  "cross_model_order": {
   "policy": "cross_model_order_v2_slot0_fwd_slot1_rev_parallel2",
   "slots": [
    {
     "model": "gemini-pro",
     "order": "forward",
     "display_to_canonical": {
      "A": "A",
      "B": "B"
     },
     "raw": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 10,
        "verdict_ko": "'앞으로 한 발을 뻗어 성큼 내디딘 미드 액션 찰나'라는 핵심 지시를 측면 구도를 통해 가장 정확하고 역동적으로 구현했습니다."
       },
       {
        "label": "A",
        "score": 6,
        "verdict_ko": "카메라를 정면으로 응시하는 정적인 걷기 자세로 연출되어 '성큼 내디딘 미드 액션'이라는 프롬프트의 요구를 충족하지 못했습니다."
       }
      ],
      "readings": [
       {
        "label": "B",
        "direction": "시선과 몸의 진행 방향이 프레임 우측을 향하고 있음.",
        "built_space": "레퍼런스와 일치하는 안개 낀 논둑길. 인물이 대각선 길 위에 자연스럽게 배치됨.",
        "entities": "안규홍의 얼굴(측면), 의상, 짚신 일치. 녹색 봇짐을 등에 메고 있어 후면과 측면이 잘 보임.",
        "hard_violations": [],
        "physics": "앞으로 다리를 길게 뻗은 보행 찰나의 역동적인 자세로, 발의 지지와 무게 중심이 자연스러움."
       },
       {
        "label": "A",
        "direction": "시선이 카메라 정면을 향함.",
        "built_space": "레퍼런스와 일치하는 안개 낀 논둑길 중앙에 서 있음.",
        "entities": "안규홍의 얼굴, 의상, 짚신 일치. 양쪽 어깨 뒤로 녹색 봇짐을 메고 끈을 쥐고 있음.",
        "hard_violations": [],
        "physics": "양발이 땅에 닿아 있으며, 보폭이 좁고 정적인 일반적인 걷기 자세임."
       }
      ],
      "all_candidates_fail": false
     },
     "normalized": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 10,
        "verdict_ko": "'앞으로 한 발을 뻗어 성큼 내디딘 미드 액션 찰나'라는 핵심 지시를 측면 구도를 통해 가장 정확하고 역동적으로 구현했습니다."
       },
       {
        "label": "A",
        "score": 6,
        "verdict_ko": "카메라를 정면으로 응시하는 정적인 걷기 자세로 연출되어 '성큼 내디딘 미드 액션'이라는 프롬프트의 요구를 충족하지 못했습니다."
       }
      ],
      "all_candidates_fail": false,
      "readings": [
       {
        "label": "B",
        "direction": "시선과 몸의 진행 방향이 프레임 우측을 향하고 있음.",
        "built_space": "레퍼런스와 일치하는 안개 낀 논둑길. 인물이 대각선 길 위에 자연스럽게 배치됨.",
        "entities": "안규홍의 얼굴(측면), 의상, 짚신 일치. 녹색 봇짐을 등에 메고 있어 후면과 측면이 잘 보임.",
        "hard_violations": [],
        "physics": "앞으로 다리를 길게 뻗은 보행 찰나의 역동적인 자세로, 발의 지지와 무게 중심이 자연스러움."
       },
       {
        "label": "A",
        "direction": "시선이 카메라 정면을 향함.",
        "built_space": "레퍼런스와 일치하는 안개 낀 논둑길 중앙에 서 있음.",
        "entities": "안규홍의 얼굴, 의상, 짚신 일치. 양쪽 어깨 뒤로 녹색 봇짐을 메고 끈을 쥐고 있음.",
        "hard_violations": [],
        "physics": "양발이 땅에 닿아 있으며, 보폭이 좁고 정적인 일반적인 걷기 자세임."
       }
      ]
     },
     "ok": true
    },
    {
     "model": "gpt-high",
     "order": "reverse",
     "display_to_canonical": {
      "A": "B",
      "B": "A"
     },
     "raw": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 6,
        "verdict_ko": "큰 보폭의 전신 동작과 봇짐 옆면은 잘 보이지만, 논둑길을 따라가기보다 가로지르는 진행 방향이며 봇짐을 잡아야 할 손도 떨어져 있습니다."
       },
       {
        "label": "B",
        "score": 8,
        "verdict_ko": "안개 낀 논둑길을 따라 홀로 걷는 전신과 봇짐 끈을 잡은 손이 요구에 더 충실하지만, 성큼 뻗는 보폭은 약하고 봇짐은 참조보다 큽니다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "얼굴과 시선, 앞으로 뻗은 발이 모두 화면 오른쪽을 향합니다. 논둑길은 전경에서 오른쪽 뒤 안개 속으로 이어져, 인물은 길을 따라 걷기보다 그 폭을 가로질러 걷는 방향으로 보입니다.",
        "built_space": "흙 논둑길 하나, 왼쪽 가장자리의 좁은 배수로 하나, 양쪽 논과 왼쪽 먼 건물·나무 윤곽이 보여 장소 참조와 잘 맞습니다. 인물 뒤로 길이 남아 있고 후속 인물은 없습니다. 전신은 들어오지만 화면 높이 대부분을 차지해 넓은 환경을 보여주는 구도보다는 다소 빽빽합니다.",
        "entities": "젊은 한국인 남성으로 보이는 인물 한 명이며 검게 묶은 머리, 마른 체격, 때 묻은 밝은 저고리와 바지, 허리띠와 짚신이 참조와 대체로 맞습니다. 청록색 천 봇짐 하나의 뒤·바깥 면이 등 옆에 보입니다. 옥가락지는 노출되지 않았고 읽을 수 있는 글자도 없습니다. 양손은 모두 봇짐이나 끈을 잡지 않습니다.",
        "hard_violations": [],
        "physics": "뒤쪽 짚신의 앞부분이 길 가장자리 지면에 닿아 몸을 지지하고, 앞발은 착지 직전으로 들려 있습니다. 보폭은 과장되어 보이지만 지지 없는 공중 부양으로 단정할 상태는 아닙니다. 봇짐은 어깨에 걸린 끈으로 지지되므로 떠 있는 물체는 아니지만, 손으로 잡으라는 명시 조건은 충족하지 않습니다."
       },
       {
        "label": "B",
        "direction": "몸과 시선은 카메라 쪽 전방을 향하고, 인물은 뒤쪽 안개에서 전경으로 이어지는 논둑길을 따라 걸어옵니다. 시선이 렌즈에 가까워 다소 정면 응시처럼 보이지만 진행 방향과는 일치합니다.",
        "built_space": "흙길 하나와 왼쪽 배수로 하나, 좌우 논, 안개에 흐려진 왼쪽 건물과 나무가 참조의 공간 구성을 유지합니다. 인물은 길 위에 있고 뒤쪽 길도 충분히 드러납니다. 다른 사람은 없으며, 전신 주변에 논과 안개를 남긴 넓은 구도가 구현되었습니다.",
        "entities": "젊은 한국인 남성으로 보이는 인물 한 명으로, 얼굴과 검게 묶은 머리, 체격 및 낡은 밝은 한복과 짚신이 인물 참조에 가깝습니다. 오른손이 청록색 봇짐 끈을 잡고 있습니다. 봇짐은 등 뒤 양옆으로 드러나지만 참조보다 크고 넓어 소박한 실루엣이라는 요구에는 덜 맞습니다. 옥가락지는 보이지 않으며 읽을 수 있는 글자는 없습니다.",
        "hard_violations": [],
        "physics": "화면 왼쪽의 앞선 발이 흙길에 붙어 체중을 지지하고, 반대쪽 발은 무릎을 굽힌 채 앞으로 옮기는 중입니다. 실제 보행으로 가능한 자세이나 한 발을 멀리 뻗어 성큼 내딛는 순간은 약합니다. 봇짐은 어깨와 등을 따라 걸리고 오른손이 끈을 붙잡아 무게를 지지합니다."
       }
      ],
      "all_candidates_fail": false
     },
     "normalized": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 6,
        "verdict_ko": "큰 보폭의 전신 동작과 봇짐 옆면은 잘 보이지만, 논둑길을 따라가기보다 가로지르는 진행 방향이며 봇짐을 잡아야 할 손도 떨어져 있습니다."
       },
       {
        "label": "A",
        "score": 8,
        "verdict_ko": "안개 낀 논둑길을 따라 홀로 걷는 전신과 봇짐 끈을 잡은 손이 요구에 더 충실하지만, 성큼 뻗는 보폭은 약하고 봇짐은 참조보다 큽니다."
       }
      ],
      "all_candidates_fail": false,
      "readings": [
       {
        "label": "B",
        "direction": "얼굴과 시선, 앞으로 뻗은 발이 모두 화면 오른쪽을 향합니다. 논둑길은 전경에서 오른쪽 뒤 안개 속으로 이어져, 인물은 길을 따라 걷기보다 그 폭을 가로질러 걷는 방향으로 보입니다.",
        "built_space": "흙 논둑길 하나, 왼쪽 가장자리의 좁은 배수로 하나, 양쪽 논과 왼쪽 먼 건물·나무 윤곽이 보여 장소 참조와 잘 맞습니다. 인물 뒤로 길이 남아 있고 후속 인물은 없습니다. 전신은 들어오지만 화면 높이 대부분을 차지해 넓은 환경을 보여주는 구도보다는 다소 빽빽합니다.",
        "entities": "젊은 한국인 남성으로 보이는 인물 한 명이며 검게 묶은 머리, 마른 체격, 때 묻은 밝은 저고리와 바지, 허리띠와 짚신이 참조와 대체로 맞습니다. 청록색 천 봇짐 하나의 뒤·바깥 면이 등 옆에 보입니다. 옥가락지는 노출되지 않았고 읽을 수 있는 글자도 없습니다. 양손은 모두 봇짐이나 끈을 잡지 않습니다.",
        "hard_violations": [],
        "physics": "뒤쪽 짚신의 앞부분이 길 가장자리 지면에 닿아 몸을 지지하고, 앞발은 착지 직전으로 들려 있습니다. 보폭은 과장되어 보이지만 지지 없는 공중 부양으로 단정할 상태는 아닙니다. 봇짐은 어깨에 걸린 끈으로 지지되므로 떠 있는 물체는 아니지만, 손으로 잡으라는 명시 조건은 충족하지 않습니다."
       },
       {
        "label": "A",
        "direction": "몸과 시선은 카메라 쪽 전방을 향하고, 인물은 뒤쪽 안개에서 전경으로 이어지는 논둑길을 따라 걸어옵니다. 시선이 렌즈에 가까워 다소 정면 응시처럼 보이지만 진행 방향과는 일치합니다.",
        "built_space": "흙길 하나와 왼쪽 배수로 하나, 좌우 논, 안개에 흐려진 왼쪽 건물과 나무가 참조의 공간 구성을 유지합니다. 인물은 길 위에 있고 뒤쪽 길도 충분히 드러납니다. 다른 사람은 없으며, 전신 주변에 논과 안개를 남긴 넓은 구도가 구현되었습니다.",
        "entities": "젊은 한국인 남성으로 보이는 인물 한 명으로, 얼굴과 검게 묶은 머리, 체격 및 낡은 밝은 한복과 짚신이 인물 참조에 가깝습니다. 오른손이 청록색 봇짐 끈을 잡고 있습니다. 봇짐은 등 뒤 양옆으로 드러나지만 참조보다 크고 넓어 소박한 실루엣이라는 요구에는 덜 맞습니다. 옥가락지는 보이지 않으며 읽을 수 있는 글자는 없습니다.",
        "hard_violations": [],
        "physics": "화면 왼쪽의 앞선 발이 흙길에 붙어 체중을 지지하고, 반대쪽 발은 무릎을 굽힌 채 앞으로 옮기는 중입니다. 실제 보행으로 가능한 자세이나 한 발을 멀리 뻗어 성큼 내딛는 순간은 약합니다. 봇짐은 어깨와 등을 따라 걸리고 오른손이 끈을 붙잡아 무게를 지지합니다."
       }
      ]
     },
     "ok": true
    }
   ],
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "slot_winner_match": false,
   "slot_winner": {
    "gemini-pro": "B",
    "gpt-high": "A"
   },
   "route": "cross_slot_combined"
  },
  "dual": {
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "normalized": {
    "A": 1.6,
    "B": 1.75
   },
   "adjusted": {
    "A": 1.6,
    "B": 1.75
   },
   "violations": {},
   "per_model_winner": {
    "gemini-pro": "B",
    "gpt-high": "A"
   },
   "agreed": false
  },
  "totals": {
   "B": 1750,
   "A": 1600
  },
  "selected": "B",
  "ranking": [
   "B",
   "A"
  ],
  "verdicts": [
   {
    "label": "B",
    "score": 1750,
    "verdict_ko": "'앞으로 한 발을 뻗어 성큼 내디딘 미드 액션 찰나'라는 핵심 지시를 측면 구도를 통해 가장 정확하고 역동적으로 구현했습니다."
   },
   {
    "label": "A",
    "score": 1600,
    "verdict_ko": "카메라를 정면으로 응시하는 정적인 걷기 자세로 연출되어 '성큼 내디딘 미드 액션'이라는 프롬프트의 요구를 충족하지 못했습니다."
   }
  ],
  "refs": [
   {
    "label": "LOCATION PHOTOGRAPH — the exact place of this shot: its architecture, materials, fixed features and lighting mood are spatial truth; stage the moment inside this place. Never copy its camera framing.",
    "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/db00517b-165b-4005-a287-f28936b59b95/images/272bd7af-9294-4490-98c6-88a40e9deacc/scene/recipe/groupbg_paddy_dyke_98913c.png",
    "asset_id": "768752bf-2da6-43b2-9b4e-2428397a7a10",
    "role": "bgfirst_group_bg"
   },
   {
    "label": "CHARACTER REFERENCE — 안규홍: the exact person appearing in this shot; match face, hair and build exactly.",
    "path": "<bytes:1028384>",
    "asset_id": "3b5894db-ce55-4f34-b966-5e725c3aa952",
    "role": "character_ref"
   },
   {
    "label": "PROP REFERENCE — 보자기 꾸러미·봇짐: the exact object appearing in this shot; match its look, material and wear exactly.",
    "path": "<bytes:1465734>",
    "asset_id": "3ca92b24-68b4-4883-b851-cabc06c00d1a",
    "role": "prop_ref"
   }
  ],
  "critique_skipped": true,
  "shot_run_produced": false,
  "shot_run_uid": "06aa140f-10df-717b-b833-e046d83121ba",
  "bgfirst": {
   "bg_path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/db00517b-165b-4005-a287-f28936b59b95/images/272bd7af-9294-4490-98c6-88a40e9deacc/scene/recipe/S12sh1__bgfirst_bg.png",
   "bg_asset_id": "096965c6-37fe-4feb-8cac-99ce4b2541bf",
   "bg_record_key": "S12sh1::bgfirst_bg",
   "chain_winner": false,
   "authority": "groupbg",
   "group_key": "paddy_dyke",
   "groupbg_asset_id": "768752bf-2da6-43b2-9b4e-2428397a7a10"
  },
  "ref_mode": "그룹 배경+엔티티 (2택1: 무콘티 승)",
  "share_plan": {
   "ref_plan": "background"
  }
 },
 "S12sh1::cine": {
  "applied": true,
  "attempted_at": "2026-09-09T01:59:05.494429+00:00",
  "fingerprint": "c2c4f9079da3cac230ff64f8dd53b129e0b837187b5259843b0f4841701924b8",
  "fingerprint_version": 2,
  "provider": "mai",
  "endpoint": "openrouter/chat-completions",
  "model": "microsoft/mai-image-2.6",
  "pack": "24.202608252115",
  "source_file": "S12sh1_sel.png",
  "source_sha256": "9e7a673345ec58e4d24940c5b824f3579e1d2cc139948913f6ac6640a4c75c8a",
  "file": "S12sh1_cine.png",
  "staged_sha256": "cf019b076c19bcbd441d6b42bce4a7467d2d17d441439b2aaa91571222b14183",
  "latency_ms": 25328
 },
 "S12sh5::signage": {
  "fp": "d862c2020546f443",
  "inscriptions": [],
  "cues": [],
  "dropped": []
 },
 "groupbg::tree_fork": {
  "input_fingerprint": "9c015f18a6acd84f",
  "meta": {
   "model": "gpt-image-2",
   "size": "1536x864",
   "pack": "11.202607220237",
   "contract": "bgfirst_full_v3",
   "group_sig": {
    "key": "tree_fork",
    "tags": [
     "S12sh10",
     "S12sh5"
    ]
   },
   "context_sig": "eccd06b9813603e5"
  },
  "prompt": "Create ONE empty live-action location background photograph — NO PEOPLE, no figures, no body parts, no silhouettes, no shadows or reflections of people anywhere.\n\"Empty\" means no people only: KEEP the location's inherent occupants and stock that define the place — animals in an animal shelter, pen or farm, goods and produce in a market, moored boats in a harbour — unless the shot text explicitly removes them.\nThe FIRST attached image is a thin-line storyboard sketch of a shot that happens at this location: use it ONLY as spatial evidence — what this place contains, how its ground, structures and landmarks are arranged and proportioned. Ignore the sketched people and arrows entirely, and do NOT copy its line style: render a fully photographic, physically plausible real place that fits THE LOCATION text below.\nHUMAN-SCALE CALIBRATION: derive every structure's true size from human-scale elements — a door ≈ 2m, a window ≈ 1–1.5m wide, one storey ≈ 2.5–3m; never inflate a small structure or shrink a large one.\n\nTHE LOCATION — early 20th-century Korea, Jeolla region; all people are Korean unless explicitly stated otherwise: Beneath a large village guardian tree at the outdoor fork below the mountain, where the paddy route meets the mountain path.\n\nLOCATION DETAIL (descriptive evidence about this place from the production's location records — evidence, not a staging order): use it to understand what kind of place this is and what it permanently contains. Carry over ONLY its enduring physical features — layout, structures, ground and materials, fixed equipment and installations. Do NOT treat any lighting, weather or time-of-day wording, momentary object states, depicted events or subjective impressions in it as fixed facts of the place; the TIME OF DAY lock below is the sole authority for time and lighting.\n논둑길: 안개가 자욱하게 깔린 새벽, 양옆으로 농경지가 펼쳐진 다져진 흙 논둑길. (특징: 자욱한 안개; 일자로 뻗은 흙 논둑; 양옆의 빈 논밭 지형)\n\nSCENE EVIDENCE (verbatim quotes from the screenplay about this place — treat them as evidence of what the location physically contains and looks like; stage the PLACE those moments happen in, but do NOT depict the momentary actions, people or staged props themselves):\n- 산 아래 당산나무 갈림길.\n나무 밑에 앉아 짚신 끈을 고쳐 매던 사내 하나가 인기척에 고개를 든다.\n- 재보가 짚신 끈을 조여 매고 일어난다.\n규홍과 관회가 다시 산길을 오른다.\n잠시 뒤—\n재보도 말없이 그 뒤를 따른다.\n\nTIME OF DAY (lock): dawn, dense fog.\n\nRender ONE photorealistic empty location photograph, 16:9, neutral enough that every shot of this place can be staged from it later. No readable writing anywhere: surfaces that would carry writing may be present, but stage any wording out of legibility — an oblique angle, distance, shallow focus. No captions, watermarks or overlay text.",
  "effective_prompt": "Create ONE empty live-action location background photograph — NO PEOPLE, no figures, no body parts, no silhouettes, no shadows or reflections of people anywhere.\n\"Empty\" means no people only: KEEP the location's inherent occupants and stock that define the place — animals in an animal shelter, pen or farm, goods and produce in a market, moored boats in a harbour — unless the shot text explicitly removes them.\nThe FIRST attached image is a thin-line storyboard sketch of a shot that happens at this location: use it ONLY as spatial evidence — what this place contains, how its ground, structures and landmarks are arranged and proportioned. Ignore the sketched people and arrows entirely, and do NOT copy its line style: render a fully photographic, physically plausible real place that fits THE LOCATION text below.\nHUMAN-SCALE CALIBRATION: derive every structure's true size from human-scale elements — a door ≈ 2m, a window ≈ 1–1.5m wide, one storey ≈ 2.5–3m; never inflate a small structure or shrink a large one.\n\nTHE LOCATION — early 20th-century Korea, Jeolla region; all people are Korean unless explicitly stated otherwise: Beneath a large village guardian tree at the outdoor fork below the mountain, where the paddy route meets the mountain path.\n\nLOCATION DETAIL (descriptive evidence about this place from the production's location records — evidence, not a staging order): use it to understand what kind of place this is and what it permanently contains. Carry over ONLY its enduring physical features — layout, structures, ground and materials, fixed equipment and installations. Do NOT treat any lighting, weather or time-of-day wording, momentary object states, depicted events or subjective impressions in it as fixed facts of the place; the TIME OF DAY lock below is the sole authority for time and lighting.\n논둑길: 안개가 자욱하게 깔린 새벽, 양옆으로 농경지가 펼쳐진 다져진 흙 논둑길. (특징: 자욱한 안개; 일자로 뻗은 흙 논둑; 양옆의 빈 논밭 지형)\n\nSCENE EVIDENCE (verbatim quotes from the screenplay about this place — treat them as evidence of what the location physically contains and looks like; stage the PLACE those moments happen in, but do NOT depict the momentary actions, people or staged props themselves):\n- 산 아래 당산나무 갈림길.\n나무 밑에 앉아 짚신 끈을 고쳐 매던 사내 하나가 인기척에 고개를 든다.\n- 재보가 짚신 끈을 조여 매고 일어난다.\n규홍과 관회가 다시 산길을 오른다.\n잠시 뒤—\n재보도 말없이 그 뒤를 따른다.\n\nTIME OF DAY (lock): dawn, dense fog.\n\nRender ONE photorealistic empty location photograph, 16:9, neutral enough that every shot of this place can be staged from it later. No readable writing anywhere: surfaces that would carry writing may be present, but stage any wording out of legibility — an oblique angle, distance, shallow focus. No captions, watermarks or overlay text.",
  "bg_path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/db00517b-165b-4005-a287-f28936b59b95/images/272bd7af-9294-4490-98c6-88a40e9deacc/scene/recipe/groupbg_tree_fork_ffe6c5.png",
  "asset_id": "3544664f-8fc9-4835-9ea1-26b67592da92",
  "input_asset_ids": [
   "4fd769e1-1085-4232-a26a-c52a6c64d715"
  ],
  "origin_tag": "S12sh5",
  "place_text": "Beneath a large village guardian tree at the outdoor fork below the mountain, where the paddy route meets the mountain path.",
  "origin_inputs": {
   "place_text": "Beneath a large village guardian tree at the outdoor fork below the mountain, where the paddy route meets the mountain path.",
   "time_of_day_en": "dawn, dense fog",
   "conti_asset_id": "4fd769e1-1085-4232-a26a-c52a6c64d715"
  }
 },
 "S12sh5::bgfirst_bg": {
  "input_fingerprint": "2106f13d9baec260",
  "prompt": "Create the EMPTY BACKGROUND PLATE for one film shot — NO PEOPLE, no figures, no body parts, no sketch lines, no arrows anywhere.\n\"Empty\" means no people only: KEEP the location's inherent occupants and stock that define the place — animals in an animal shelter, pen or farm, goods in a market, moored boats in a harbour — unless the shot text explicitly removes them.\nThe FIRST attached image is a thin-line storyboard sketch: use ONLY its camera angle, horizon, perspective and the placement/size of buildings and set masses — ignore the sketched people and arrows entirely. The SECOND attached image (LOCATION PHOTOGRAPH) is the real place: take its architecture, materials, signage and fixed features, and RE-PROJECT them into the sketch's camera. If the photograph's camera differs from the sketch's, the sketch's camera wins.\nHUMAN-SCALE CALIBRATION: derive every structure's true size from human-scale elements — a door ≈ 2m, a window ≈ 1–1.5m wide, one storey ≈ 2.5–3m; never inflate a small structure or shrink a large one.\n\nSHOT TEXT this background must serve (Korean): 거대한 당산나무 아래 쪼그려 앉아 짚신 끈을 잡은 채 고개를 확 쳐든 염재보의 날카로운 눈빛.\n\nLOCATION (lock): Beneath a large village guardian tree at the outdoor fork below the mountain, where the paddy route meets the mountain path.\n\nTIME OF DAY (lock): dawn, dense fog.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: medium shot\n- KEY BACKGROUND ELEMENTS: 당산나무 (염재보 is crouched beneath the tree at the mountain-foot fork) — Only a section behind his crouched body is visible; used as Locates his pause without letting the tree overwhelm his face and hands; 짚신 끈 (Held in his hands during an interrupted adjustment) — The held ends are visible obliquely at the lower frame edge; used as Preserves evidence of the interrupted action beneath the alert expression.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Natural dawn ambience gives his lifted face readable, restrained contrast without adding an unestablished light source or extending the field-path fog into this location.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nRender ONE photorealistic empty location photograph, 16:9, that this shot can be staged inside later. No readable writing anywhere: surfaces that would carry writing may be present, but stage any wording out of legibility — an oblique angle, distance, shallow focus. No captions, watermarks or overlay text.",
  "effective_prompt": "Create the EMPTY BACKGROUND PLATE for one film shot — NO PEOPLE, no figures, no body parts, no sketch lines, no arrows anywhere.\n\"Empty\" means no people only: KEEP the location's inherent occupants and stock that define the place — animals in an animal shelter, pen or farm, goods in a market, moored boats in a harbour — unless the shot text explicitly removes them.\nThe FIRST attached image is a thin-line storyboard sketch: use ONLY its camera angle, horizon, perspective and the placement/size of buildings and set masses — ignore the sketched people and arrows entirely. The SECOND attached image (LOCATION PHOTOGRAPH) is the real place: take its architecture, materials, signage and fixed features, and RE-PROJECT them into the sketch's camera. If the photograph's camera differs from the sketch's, the sketch's camera wins.\nHUMAN-SCALE CALIBRATION: derive every structure's true size from human-scale elements — a door ≈ 2m, a window ≈ 1–1.5m wide, one storey ≈ 2.5–3m; never inflate a small structure or shrink a large one.\n\nSHOT TEXT this background must serve (Korean): 거대한 당산나무 아래 쪼그려 앉아 짚신 끈을 잡은 채 고개를 확 쳐든 염재보의 날카로운 눈빛.\n\nLOCATION (lock): Beneath a large village guardian tree at the outdoor fork below the mountain, where the paddy route meets the mountain path.\n\nTIME OF DAY (lock): dawn, dense fog.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: medium shot\n- KEY BACKGROUND ELEMENTS: 당산나무 (염재보 is crouched beneath the tree at the mountain-foot fork) — Only a section behind his crouched body is visible; used as Locates his pause without letting the tree overwhelm his face and hands; 짚신 끈 (Held in his hands during an interrupted adjustment) — The held ends are visible obliquely at the lower frame edge; used as Preserves evidence of the interrupted action beneath the alert expression.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Natural dawn ambience gives his lifted face readable, restrained contrast without adding an unestablished light source or extending the field-path fog into this location.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nRender ONE photorealistic empty location photograph, 16:9, that this shot can be staged inside later. No readable writing anywhere: surfaces that would carry writing may be present, but stage any wording out of legibility — an oblique angle, distance, shallow focus. No captions, watermarks or overlay text.",
  "bg_path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/db00517b-165b-4005-a287-f28936b59b95/images/272bd7af-9294-4490-98c6-88a40e9deacc/scene/recipe/S12sh5__bgfirst_bg.png",
  "asset_id": "a7592293-5b1b-4289-be1c-0895c83122eb",
  "input_asset_ids": [
   "4fd769e1-1085-4232-a26a-c52a6c64d715",
   "3544664f-8fc9-4835-9ea1-26b67592da92"
  ]
 },
 "S12sh5": {
  "input_fingerprint": "f65eee71ce7f41b8",
  "prompt": "Create ONE FINAL photorealistic live-action film still of the moment below — early 20th-century Korea, Jeolla region; all people are Korean unless explicitly stated otherwise. TIME OF DAY (lock): dawn, dense fog.\n\nSHOT TEXT (authoritative, Korean): 거대한 당산나무 아래 쪼그려 앉아 짚신 끈을 잡은 채 고개를 확 쳐든 염재보의 날카로운 눈빛.\n\nLOCATION (lock): Beneath a large village guardian tree at the outdoor fork below the mountain, where the paddy route meets the mountain path. The shot takes place here — the attached LOCATION PHOTOGRAPH shows the exact spot.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: medium shot\n- KEY BACKGROUND ELEMENTS: 당산나무 (염재보 is crouched beneath the tree at the mountain-foot fork) — Only a section behind his crouched body is visible; used as Locates his pause without letting the tree overwhelm his face and hands; 짚신 끈 (Held in his hands during an interrupted adjustment) — The held ends are visible obliquely at the lower frame edge; used as Preserves evidence of the interrupted action beneath the alert expression.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Natural dawn ambience gives his lifted face readable, restrained contrast without adding an unestablished light source or extending the field-path fog into this location.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): Dawn fog surrounds the fork beneath the sacred tree. 염재보: He sits under the tree holding his straw-shoe ties; his confiscated gun is absent.\n\nTHE HAND THAT IS DOING THIS: the object at the centre of this shot is being held, operated, read, aimed or handed over by 염재보 right now, so 염재보's hand — and as much of the wrist and forearm as the framing reaches — is in the frame, gripping or resting on that object exactly the way its real use requires. Match that hand to 염재보: its size, build, skin, age, grooming, sleeve and anything worn on it belong to that person and to no one else.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 염재보 (한국인 남성, 성인 나이대, 자연스러운 검은 머리) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.",
  "roll_prompts": {
   "A": "Stage the shot. The FIRST attached image (SHOT BACKGROUND) is the finished empty background of this shot — keep it EXACTLY: its camera, perspective, architecture, lighting and every fixed feature stay untouched. The SECOND attached image (LAYOUT SKETCH) tells you ONLY where the people go: each sketched person's position, screen size, pose and the gaze/motion arrows. Ignore the sketch's background lines. The CHARACTER REFERENCE photographs show the real people.\nPlace the real people into the background at exactly the sketched positions, sizes and poses, following the arrow directions. No sketch lines or arrows may remain.\n\nCreate ONE FINAL photorealistic live-action film still of the moment below — early 20th-century Korea, Jeolla region; all people are Korean unless explicitly stated otherwise. TIME OF DAY (lock): dawn, dense fog.\n\nSHOT TEXT (authoritative, Korean): 거대한 당산나무 아래 쪼그려 앉아 짚신 끈을 잡은 채 고개를 확 쳐든 염재보의 날카로운 눈빛.\n\nLOCATION (lock): Beneath a large village guardian tree at the outdoor fork below the mountain, where the paddy route meets the mountain path. The shot takes place here — the FIRST attached image (SHOT BACKGROUND) is this exact place, already built: its ground, structures, horizon, materials and lighting are the finished truth of this location and must not be redesigned or replaced. No location photograph is attached — read the place from that image alone, and add no scenery, structure, vehicle or fixture that it does not already show. This lock governs the place only; the figures in the shot follow the staging and pose instructions.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: medium shot\n- KEY BACKGROUND ELEMENTS: 당산나무 (염재보 is crouched beneath the tree at the mountain-foot fork) — Only a section behind his crouched body is visible; used as Locates his pause without letting the tree overwhelm his face and hands; 짚신 끈 (Held in his hands during an interrupted adjustment) — The held ends are visible obliquely at the lower frame edge; used as Preserves evidence of the interrupted action beneath the alert expression.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Natural dawn ambience gives his lifted face readable, restrained contrast without adding an unestablished light source or extending the field-path fog into this location.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): Dawn fog surrounds the fork beneath the sacred tree. 염재보: He sits under the tree holding his straw-shoe ties; his confiscated gun is absent.\n\nTHE HAND THAT IS DOING THIS: the object at the centre of this shot is being held, operated, read, aimed or handed over by 염재보 right now, so 염재보's hand — and as much of the wrist and forearm as the framing reaches — is in the frame, gripping or resting on that object exactly the way its real use requires. Match that hand to 염재보: its size, build, skin, age, grooming, sleeve and anything worn on it belong to that person and to no one else.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 염재보 (한국인 남성, 성인 나이대, 자연스러운 검은 머리) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.",
   "B": "Create ONE FINAL photorealistic live-action film still of the moment below — early 20th-century Korea, Jeolla region; all people are Korean unless explicitly stated otherwise. TIME OF DAY (lock): dawn, dense fog.\n\nSHOT TEXT (authoritative, Korean): 거대한 당산나무 아래 쪼그려 앉아 짚신 끈을 잡은 채 고개를 확 쳐든 염재보의 날카로운 눈빛.\n\nLOCATION (lock): Beneath a large village guardian tree at the outdoor fork below the mountain, where the paddy route meets the mountain path. The shot takes place here — the attached LOCATION PHOTOGRAPH shows the exact spot.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: medium shot\n- KEY BACKGROUND ELEMENTS: 당산나무 (염재보 is crouched beneath the tree at the mountain-foot fork) — Only a section behind his crouched body is visible; used as Locates his pause without letting the tree overwhelm his face and hands; 짚신 끈 (Held in his hands during an interrupted adjustment) — The held ends are visible obliquely at the lower frame edge; used as Preserves evidence of the interrupted action beneath the alert expression.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Natural dawn ambience gives his lifted face readable, restrained contrast without adding an unestablished light source or extending the field-path fog into this location.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): Dawn fog surrounds the fork beneath the sacred tree. 염재보: He sits under the tree holding his straw-shoe ties; his confiscated gun is absent.\n\nTHE HAND THAT IS DOING THIS: the object at the centre of this shot is being held, operated, read, aimed or handed over by 염재보 right now, so 염재보's hand — and as much of the wrist and forearm as the framing reaches — is in the frame, gripping or resting on that object exactly the way its real use requires. Match that hand to 염재보: its size, build, skin, age, grooming, sleeve and anything worn on it belong to that person and to no one else.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 염재보 (한국인 남성, 성인 나이대, 자연스러운 검은 머리) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.\n\nCOMPOSITION VARIATION (this candidate only): keep the place, the people and the moment exactly as specified, and keep from the CAMERA & FRAME contract its subject scale, screen placement and key background elements exactly. For this candidate only, its camera angle and height are deliberately left open: choose a camera position distinctly different from the obvious one, as a film director picking a second setup on the same blocking. Stay true to any explicit framing the SHOT TEXT itself dictates."
  },
  "roll_refs": {
   "A": [
    {
     "label": "SHOT BACKGROUND",
     "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/db00517b-165b-4005-a287-f28936b59b95/images/272bd7af-9294-4490-98c6-88a40e9deacc/scene/recipe/S12sh5__bgfirst_bg.png",
     "asset_id": "a7592293-5b1b-4289-be1c-0895c83122eb",
     "role": "bgfirst_bg"
    },
    {
     "label": "LAYOUT SKETCH (people placement only)",
     "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/db00517b-165b-4005-a287-f28936b59b95/images/272bd7af-9294-4490-98c6-88a40e9deacc/conti/conti_S12sh5.png",
     "asset_id": "4fd769e1-1085-4232-a26a-c52a6c64d715",
     "role": "conti_light"
    },
    {
     "label": "CHARACTER REFERENCE — 염재보: the exact person appearing in this shot; match face, hair and build exactly.",
     "path": "<bytes:964351>",
     "asset_id": "ab287289-d0b7-436d-8f70-68452f898f73",
     "role": "character_ref"
    }
   ],
   "B": [
    {
     "label": "LOCATION PHOTOGRAPH — the exact place of this shot: its architecture, materials, fixed features and lighting mood are spatial truth; stage the moment inside this place. Never copy its camera framing.",
     "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/db00517b-165b-4005-a287-f28936b59b95/images/272bd7af-9294-4490-98c6-88a40e9deacc/scene/recipe/groupbg_tree_fork_ffe6c5.png",
     "asset_id": "3544664f-8fc9-4835-9ea1-26b67592da92",
     "role": "bgfirst_group_bg"
    },
    {
     "label": "CHARACTER REFERENCE — 염재보: the exact person appearing in this shot; match face, hair and build exactly.",
     "path": "<bytes:964351>",
     "asset_id": "ab287289-d0b7-436d-8f70-68452f898f73",
     "role": "character_ref"
    }
   ]
  },
  "shot_run_spend_attempt_count": 1,
  "judge_flip_skipped": "judge_fn_owns_order",
  "readings": [
   {
    "label": "A",
    "direction": "고개를 들고 화면 왼쪽 바깥을 주시하고 있습니다.",
    "built_space": "안개가 자욱한 산 아래 갈림길과 거대한 당산나무가 레퍼런스와 유사하게 배치되어 있습니다.",
    "entities": "염재보(레퍼런스와 일치하는 복장과 삿갓 착용), 손에 쥐어진 짚신 끈이 확인되나, 배경 안개 속에 프롬프트에 없는 두 명의 인물이 등장합니다.",
    "hard_violations": [
     "[gemini-pro] 프롬프트에 명시되지 않은 추가 인물(배경의 2명) 생성",
     "[gemini-pro] 지정된 샷 크기(미디엄 샷) 위반",
     "[gpt-high] 논길에 원문이 허용하지 않은 추가 인물 두 명을 배치했습니다."
    ],
    "physics": "땅에 쪼그려 앉아 있으며, 손은 짚신 끈을 물리적으로 쥐고 지탱하고 있습니다."
   },
   {
    "label": "B",
    "direction": "고개를 확 쳐들고 화면 오른쪽 위를 날카롭게 주시하고 있습니다.",
    "built_space": "안개 낀 당산나무 아래 흙바닥에 위치해 있으며, 인물 뒤로 나무 기둥이 지시대로 배치되어 있습니다.",
    "entities": "염재보(얼굴과 복장이 레퍼런스와 일치함), 하단 프레임 경계에 위치한 짚신과 손에 쥐어진 끈이 정확히 묘사되었습니다.",
    "hard_violations": [],
    "physics": "바닥에 안정적으로 쪼그려 앉은 자세를 유지하며, 양손으로 짚신 끈을 자연스럽게 쥐고 있습니다."
   }
  ],
  "cross_model_order": {
   "policy": "cross_model_order_v2_slot0_fwd_slot1_rev_parallel2",
   "slots": [
    {
     "model": "gemini-pro",
     "order": "forward",
     "display_to_canonical": {
      "A": "A",
      "B": "B"
     },
     "raw": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 9,
        "verdict_ko": "지정된 미디엄 샷 프레이밍을 준수하며, 고개를 확 쳐든 인물의 표정과 짚신 끈을 쥔 손의 액션을 프롬프트에 맞게 훌륭히 구현했습니다."
       },
       {
        "label": "A",
        "score": 2,
        "verdict_ko": "지정되지 않은 인물들이 배경에 등장하는 치명적인 오류가 있으며, 프레이밍 또한 요청된 미디엄 샷보다 훨씬 넓게 잡혔습니다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "고개를 들고 화면 왼쪽 바깥을 주시하고 있습니다.",
        "built_space": "안개가 자욱한 산 아래 갈림길과 거대한 당산나무가 레퍼런스와 유사하게 배치되어 있습니다.",
        "entities": "염재보(레퍼런스와 일치하는 복장과 삿갓 착용), 손에 쥐어진 짚신 끈이 확인되나, 배경 안개 속에 프롬프트에 없는 두 명의 인물이 등장합니다.",
        "hard_violations": [
         "프롬프트에 명시되지 않은 추가 인물(배경의 2명) 생성",
         "지정된 샷 크기(미디엄 샷) 위반"
        ],
        "physics": "땅에 쪼그려 앉아 있으며, 손은 짚신 끈을 물리적으로 쥐고 지탱하고 있습니다."
       },
       {
        "label": "B",
        "direction": "고개를 확 쳐들고 화면 오른쪽 위를 날카롭게 주시하고 있습니다.",
        "built_space": "안개 낀 당산나무 아래 흙바닥에 위치해 있으며, 인물 뒤로 나무 기둥이 지시대로 배치되어 있습니다.",
        "entities": "염재보(얼굴과 복장이 레퍼런스와 일치함), 하단 프레임 경계에 위치한 짚신과 손에 쥐어진 끈이 정확히 묘사되었습니다.",
        "hard_violations": [],
        "physics": "바닥에 안정적으로 쪼그려 앉은 자세를 유지하며, 양손으로 짚신 끈을 자연스럽게 쥐고 있습니다."
       }
      ],
      "all_candidates_fail": false
     },
     "normalized": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 9,
        "verdict_ko": "지정된 미디엄 샷 프레이밍을 준수하며, 고개를 확 쳐든 인물의 표정과 짚신 끈을 쥔 손의 액션을 프롬프트에 맞게 훌륭히 구현했습니다."
       },
       {
        "label": "A",
        "score": 2,
        "verdict_ko": "지정되지 않은 인물들이 배경에 등장하는 치명적인 오류가 있으며, 프레이밍 또한 요청된 미디엄 샷보다 훨씬 넓게 잡혔습니다."
       }
      ],
      "all_candidates_fail": false,
      "readings": [
       {
        "label": "A",
        "direction": "고개를 들고 화면 왼쪽 바깥을 주시하고 있습니다.",
        "built_space": "안개가 자욱한 산 아래 갈림길과 거대한 당산나무가 레퍼런스와 유사하게 배치되어 있습니다.",
        "entities": "염재보(레퍼런스와 일치하는 복장과 삿갓 착용), 손에 쥐어진 짚신 끈이 확인되나, 배경 안개 속에 프롬프트에 없는 두 명의 인물이 등장합니다.",
        "hard_violations": [
         "프롬프트에 명시되지 않은 추가 인물(배경의 2명) 생성",
         "지정된 샷 크기(미디엄 샷) 위반"
        ],
        "physics": "땅에 쪼그려 앉아 있으며, 손은 짚신 끈을 물리적으로 쥐고 지탱하고 있습니다."
       },
       {
        "label": "B",
        "direction": "고개를 확 쳐들고 화면 오른쪽 위를 날카롭게 주시하고 있습니다.",
        "built_space": "안개 낀 당산나무 아래 흙바닥에 위치해 있으며, 인물 뒤로 나무 기둥이 지시대로 배치되어 있습니다.",
        "entities": "염재보(얼굴과 복장이 레퍼런스와 일치함), 하단 프레임 경계에 위치한 짚신과 손에 쥐어진 끈이 정확히 묘사되었습니다.",
        "hard_violations": [],
        "physics": "바닥에 안정적으로 쪼그려 앉은 자세를 유지하며, 양손으로 짚신 끈을 자연스럽게 쥐고 있습니다."
       }
      ]
     },
     "ok": true
    },
    {
     "model": "gpt-high",
     "order": "reverse",
     "display_to_canonical": {
      "A": "B",
      "B": "A"
     },
     "raw": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 7,
        "verdict_ko": "끈을 잡은 손을 하단에 두고 고개를 치켜든 순간을 더 충실히 구현했지만, 참조의 삿갓 착용과 정확한 나무·갈림길 지형은 일치하지 않습니다."
       },
       {
        "label": "B",
        "score": 2,
        "verdict_ko": "참조의 복장과 논길 주변 지형은 가깝지만, 금지된 인물 두 명을 추가했고 고개를 확 쳐든 순간보다 그들을 옆으로 바라보는 장면이 되어 탈락합니다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "염재보는 숙인 몸에서 얼굴을 위로 들고 화면 오른쪽 위의 프레임 밖을 응시합니다. 대상은 보이지 않으며, 지정된 응시 대상이 없는 원문의 경계하는 순간과 맞습니다. 양손은 아래쪽 신발 부근의 끈을 향하고 끈 끝을 붙잡고 있습니다. 무기는 없습니다.",
        "built_space": "인물 바로 뒤에 큰 나무 한 그루의 줄기와 굵게 갈라진 가지, 지면으로 뻗은 뿌리가 보입니다. 인물은 뿌리 앞 흙바닥에 쪼그려 있습니다. 참조의 낮은 돌 경계와 논길·산길의 분기는 이 화면에서 확인되지 않습니다. 나무의 낮고 굵은 분지 형태도 참조와 다르며, 줄기와 가지가 배경을 상당히 차지합니다. 인공 시설이나 반사는 없습니다.",
        "entities": "보이는 사람은 성인 한국인 남성으로 표현된 염재보 한 명뿐이며, 검은 머리와 정상적인 눈, 어두운 누비 저고리와 바지, 갈색 다리 감개가 보입니다. 얼굴은 참조와 대체로 유사하지만 완전히 같은 인상은 아니며, 참조의 삿갓은 머리에 없습니다. 양손에 짚 재질 끈이 있고 하단에는 신발 일부가 보입니다. 오른쪽 아래 짚 엮음 물체는 잘려 있어 종류를 확정하기 어렵습니다. 총과 읽을 수 있는 글자는 없습니다. 차가운 새벽빛과 짙은 배경 안개가 표현되어 있습니다.",
        "hard_violations": [],
        "physics": "무릎을 깊이 굽힌 몸은 하단에 보이는 두 발의 지면 접촉으로 지지됩니다. 상체를 앞으로 기울인 채 목과 얼굴만 들어 올리는 자세가 가능하며, 끈 조정을 중단한 동작으로 읽힙니다. 양손 손가락이 끈을 실제로 집고 있고, 풀린 끈 끝은 아래로 처집니다. 지지 없이 떠 있는 몸이나 물체는 보이지 않습니다."
       },
       {
        "label": "B",
        "direction": "염재보의 얼굴과 눈은 화면 왼쪽 논길에 있는 두 사람 쪽을 향합니다. 얼굴은 손보다 들어 올려져 있지만, 위로 확 치켜든 동작보다는 수평에 가까운 옆보기입니다. 배경 두 사람은 카메라와 염재보가 있는 전경 쪽을 향해 서 있거나 다가오는 모습입니다. 양손은 무릎 앞의 짚끈을 붙잡고 있으며 무기는 없습니다.",
        "built_space": "화면 오른쪽에 큰 나무 한 그루의 줄기와 뿌리가 있고, 염재보는 그 바로 앞에 쪼그려 있습니다. 왼쪽으로 논 사이 흙길 하나가 이어지며, 중앙에는 낮은 돌 경계가 있습니다. 참조의 논과 숲 가장자리 배치는 비교적 가깝지만, 오르막 산길은 인물과 나무에 가려 분명하지 않습니다. 배경을 넓게 열어 길과 추가 인물들을 보여주므로 얼굴과 하단의 손에 집중하는 지정 구도에서 벗어납니다.",
        "entities": "전경의 성인 한국인 남성은 검은 머리, 삿갓, 어두운 누비옷과 바지, 갈색 다리 감개를 갖춰 참조 복장과 대체로 맞습니다. 얼굴은 측면으로 보이며 눈은 정상적인 사람의 눈입니다. 짚신 한 짝이 지면에 놓이고 그 끈을 양손으로 잡고 있습니다. 그러나 논길에 성인 남성으로 보이는 인물 두 명이 추가되어, 염재보만 등장해야 한다는 조건을 위반합니다. 총이나 읽을 수 있는 글자는 없으며 새벽 안개가 보입니다.",
        "hard_violations": [
         "논길에 원문이 허용하지 않은 추가 인물 두 명을 배치했습니다."
        ],
        "physics": "염재보는 깊이 굽힌 두 다리와 지면에 닿은 발로 몸을 지탱하며 나무뿌리 앞에 낮게 앉아 있습니다. 손은 짚끈을 잡고 있고 끈이 연결된 짚신은 흙바닥에 놓여 있어 지지가 분명합니다. 추가 인물 두 명도 발이 길에 닿아 있는 것으로 보이며, 공중에 떠 있는 몸이나 물체는 없습니다."
       }
      ],
      "all_candidates_fail": false
     },
     "normalized": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 7,
        "verdict_ko": "끈을 잡은 손을 하단에 두고 고개를 치켜든 순간을 더 충실히 구현했지만, 참조의 삿갓 착용과 정확한 나무·갈림길 지형은 일치하지 않습니다."
       },
       {
        "label": "A",
        "score": 2,
        "verdict_ko": "참조의 복장과 논길 주변 지형은 가깝지만, 금지된 인물 두 명을 추가했고 고개를 확 쳐든 순간보다 그들을 옆으로 바라보는 장면이 되어 탈락합니다."
       }
      ],
      "all_candidates_fail": false,
      "readings": [
       {
        "label": "B",
        "direction": "염재보는 숙인 몸에서 얼굴을 위로 들고 화면 오른쪽 위의 프레임 밖을 응시합니다. 대상은 보이지 않으며, 지정된 응시 대상이 없는 원문의 경계하는 순간과 맞습니다. 양손은 아래쪽 신발 부근의 끈을 향하고 끈 끝을 붙잡고 있습니다. 무기는 없습니다.",
        "built_space": "인물 바로 뒤에 큰 나무 한 그루의 줄기와 굵게 갈라진 가지, 지면으로 뻗은 뿌리가 보입니다. 인물은 뿌리 앞 흙바닥에 쪼그려 있습니다. 참조의 낮은 돌 경계와 논길·산길의 분기는 이 화면에서 확인되지 않습니다. 나무의 낮고 굵은 분지 형태도 참조와 다르며, 줄기와 가지가 배경을 상당히 차지합니다. 인공 시설이나 반사는 없습니다.",
        "entities": "보이는 사람은 성인 한국인 남성으로 표현된 염재보 한 명뿐이며, 검은 머리와 정상적인 눈, 어두운 누비 저고리와 바지, 갈색 다리 감개가 보입니다. 얼굴은 참조와 대체로 유사하지만 완전히 같은 인상은 아니며, 참조의 삿갓은 머리에 없습니다. 양손에 짚 재질 끈이 있고 하단에는 신발 일부가 보입니다. 오른쪽 아래 짚 엮음 물체는 잘려 있어 종류를 확정하기 어렵습니다. 총과 읽을 수 있는 글자는 없습니다. 차가운 새벽빛과 짙은 배경 안개가 표현되어 있습니다.",
        "hard_violations": [],
        "physics": "무릎을 깊이 굽힌 몸은 하단에 보이는 두 발의 지면 접촉으로 지지됩니다. 상체를 앞으로 기울인 채 목과 얼굴만 들어 올리는 자세가 가능하며, 끈 조정을 중단한 동작으로 읽힙니다. 양손 손가락이 끈을 실제로 집고 있고, 풀린 끈 끝은 아래로 처집니다. 지지 없이 떠 있는 몸이나 물체는 보이지 않습니다."
       },
       {
        "label": "A",
        "direction": "염재보의 얼굴과 눈은 화면 왼쪽 논길에 있는 두 사람 쪽을 향합니다. 얼굴은 손보다 들어 올려져 있지만, 위로 확 치켜든 동작보다는 수평에 가까운 옆보기입니다. 배경 두 사람은 카메라와 염재보가 있는 전경 쪽을 향해 서 있거나 다가오는 모습입니다. 양손은 무릎 앞의 짚끈을 붙잡고 있으며 무기는 없습니다.",
        "built_space": "화면 오른쪽에 큰 나무 한 그루의 줄기와 뿌리가 있고, 염재보는 그 바로 앞에 쪼그려 있습니다. 왼쪽으로 논 사이 흙길 하나가 이어지며, 중앙에는 낮은 돌 경계가 있습니다. 참조의 논과 숲 가장자리 배치는 비교적 가깝지만, 오르막 산길은 인물과 나무에 가려 분명하지 않습니다. 배경을 넓게 열어 길과 추가 인물들을 보여주므로 얼굴과 하단의 손에 집중하는 지정 구도에서 벗어납니다.",
        "entities": "전경의 성인 한국인 남성은 검은 머리, 삿갓, 어두운 누비옷과 바지, 갈색 다리 감개를 갖춰 참조 복장과 대체로 맞습니다. 얼굴은 측면으로 보이며 눈은 정상적인 사람의 눈입니다. 짚신 한 짝이 지면에 놓이고 그 끈을 양손으로 잡고 있습니다. 그러나 논길에 성인 남성으로 보이는 인물 두 명이 추가되어, 염재보만 등장해야 한다는 조건을 위반합니다. 총이나 읽을 수 있는 글자는 없으며 새벽 안개가 보입니다.",
        "hard_violations": [
         "논길에 원문이 허용하지 않은 추가 인물 두 명을 배치했습니다."
        ],
        "physics": "염재보는 깊이 굽힌 두 다리와 지면에 닿은 발로 몸을 지탱하며 나무뿌리 앞에 낮게 앉아 있습니다. 손은 짚끈을 잡고 있고 끈이 연결된 짚신은 흙바닥에 놓여 있어 지지가 분명합니다. 추가 인물 두 명도 발이 길에 닿아 있는 것으로 보이며, 공중에 떠 있는 몸이나 물체는 없습니다."
       }
      ]
     },
     "ok": true
    }
   ],
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "slot_winner_match": true,
   "slot_winner": {
    "gemini-pro": "B",
    "gpt-high": "B"
   },
   "route": "cross_slot_agree"
  },
  "dual": {
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "normalized": {
    "A": 0.508,
    "B": 2.0
   },
   "adjusted": {
    "A": 0.258,
    "B": 2.0
   },
   "violations": {
    "A": [
     "[gemini-pro] 프롬프트에 명시되지 않은 추가 인물(배경의 2명) 생성",
     "[gemini-pro] 지정된 샷 크기(미디엄 샷) 위반",
     "[gpt-high] 논길에 원문이 허용하지 않은 추가 인물 두 명을 배치했습니다."
    ]
   },
   "per_model_winner": {
    "gemini-pro": "B",
    "gpt-high": "B"
   },
   "agreed": true
  },
  "totals": {
   "B": 2000,
   "A": 258
  },
  "selected": "B",
  "ranking": [
   "B",
   "A"
  ],
  "verdicts": [
   {
    "label": "B",
    "score": 2000,
    "verdict_ko": "지정된 미디엄 샷 프레이밍을 준수하며, 고개를 확 쳐든 인물의 표정과 짚신 끈을 쥔 손의 액션을 프롬프트에 맞게 훌륭히 구현했습니다."
   },
   {
    "label": "A",
    "score": 258,
    "verdict_ko": "지정되지 않은 인물들이 배경에 등장하는 치명적인 오류가 있으며, 프레이밍 또한 요청된 미디엄 샷보다 훨씬 넓게 잡혔습니다.  ★위반: [gemini-pro] 프롬프트에 명시되지 않은 추가 인물(배경의 2명) 생성 / [gemini-pro] 지정된 샷 크기(미디엄 샷) 위반 / [gpt-high] 논길에 원문이 허용하지 않은 추가 인물 두 명을 배치했습니다."
   }
  ],
  "refs": [
   {
    "label": "LOCATION PHOTOGRAPH — the exact place of this shot: its architecture, materials, fixed features and lighting mood are spatial truth; stage the moment inside this place. Never copy its camera framing.",
    "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/db00517b-165b-4005-a287-f28936b59b95/images/272bd7af-9294-4490-98c6-88a40e9deacc/scene/recipe/groupbg_tree_fork_ffe6c5.png",
    "asset_id": "3544664f-8fc9-4835-9ea1-26b67592da92",
    "role": "bgfirst_group_bg"
   },
   {
    "label": "CHARACTER REFERENCE — 염재보: the exact person appearing in this shot; match face, hair and build exactly.",
    "path": "<bytes:964351>",
    "asset_id": "ab287289-d0b7-436d-8f70-68452f898f73",
    "role": "character_ref"
   }
  ],
  "critique_skipped": true,
  "shot_run_produced": false,
  "shot_run_uid": "06aa140f-14f8-7f79-ba7b-0ef474ea2f54",
  "bgfirst": {
   "bg_path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/db00517b-165b-4005-a287-f28936b59b95/images/272bd7af-9294-4490-98c6-88a40e9deacc/scene/recipe/S12sh5__bgfirst_bg.png",
   "bg_asset_id": "a7592293-5b1b-4289-be1c-0895c83122eb",
   "bg_record_key": "S12sh5::bgfirst_bg",
   "chain_winner": false,
   "authority": "groupbg",
   "group_key": "tree_fork",
   "groupbg_asset_id": "3544664f-8fc9-4835-9ea1-26b67592da92"
  },
  "ref_mode": "그룹 배경+엔티티 (2택1: 무콘티 승)",
  "share_plan": {
   "ref_plan": "background"
  }
 },
 "S12sh5::cine": {
  "applied": true,
  "attempted_at": "2026-09-09T02:03:33.400461+00:00",
  "fingerprint": "aa9435120e2df58e533ef38ceb8ff9f2f6841f04631d1a91129cad4e4d7ed14b",
  "fingerprint_version": 2,
  "provider": "mai",
  "endpoint": "openrouter/chat-completions",
  "model": "microsoft/mai-image-2.6",
  "pack": "24.202608252115",
  "source_file": "S12sh5_sel.png",
  "source_sha256": "dccfe88fd87098d5145b6b72ce2d9737a25181e139e834be972cd8452054e224",
  "file": "S12sh5_cine.png",
  "staged_sha256": "2602b14c5c2917ed8965e1e2a67194d7988d305092609608865fc0f78644277d",
  "latency_ms": 24546
 },
 "S12sh10::signage": {
  "fp": "260442245c983046",
  "inscriptions": [],
  "cues": [],
  "dropped": []
 },
 "S12sh10::bgfirst_bg": {
  "input_fingerprint": "aeeeae5441716546",
  "prompt": "Turn the attached storyboard sketch into a photorealistic film still of\nits LOCATION ONLY, keeping the figures exactly as they are.\n\nKEEP EXACTLY: the camera framing, the horizon, and where every element\nsits in the frame. Each mannequin figure stays a plain grey featureless\nmannequin standing in the very same spot, at the same size, in the same\npose, turned the same way — do not turn them into people, do not move,\nrotate, mirror or re-pose them, do not add or remove figures.\n\nBUILD PHOTOREALISTICALLY: everything that is not a figure — ground,\nsurfacing, structures, vegetation, sky, water, distance. The bare lines\nof the sketch are a layout guide; replace them with the real materials,\ndepth and lighting of the place described below.\n\nSHOT TEXT this background must serve (Korean): 멀어지는 두 사람의 등 뒤로 거리를 둔 채 산길을 향해 한 발을 뻗은 미드 액션 상태인 염재보의 뒷모습.\n\nLOCATION (lock): The ascending mountain path just beyond the guardian-tree fork, trailing the route taken by two travelers.\n\nTIME OF DAY (lock): dawn, dense fog.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: wide shot\n- FRAME LAYOUT: Uninterrupted uphill path separating 염재보 from the departing pair in the middle-center of the frame, midground.\n- KEY BACKGROUND ELEMENTS: 산길 (The pair is walking uphill, with 염재보 beginning to follow after a delay) — Recedes from the lower-right area toward the upper-left background; used as Its visible empty interval makes the delayed following action unmistakable.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Subdued natural dawn ambience maintains sober continuity, with sufficient tonal separation to read all three figures and the gap between them.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHINGS AT THIS PLACE:\nThe site follows a connected route from a paddy embankment path to a mountain path and a junction at the mountain foot. Paddy plots directly border both sides of the embankment path. Beyond the fields, the mountain path connects to the junction and continues uphill beyond it. A separate branch leaves the junction to one side; its destination is unspecified. The village guardian tree stands beside the junction, opposite the branch opening, with its base adjoining the accessible junction ground rather than obstructing the route. For the site plan, place the field-path approach toward the bottom and the uphill continuation toward the top; these are drawing directions, not compass bearings. Preserve an open path connection through the junction and accessible ground beside the tree.\n- path: Paddy embankment path\n- fields: Flanking paddy plots\n- junction: Mountain-foot path junction\n- tree: Mountain-foot village guardian tree\n- path: Lateral junction branch\n- path: Mountain approach and uphill path\n\nWORLD FACTS (creator-confirmed — always true):\n- Region (real-world reference): 한반도 전라남도 보성 일대의 농촌·산지·장터 및 인근 해안권.\n- Era: 대한제국 말기, 20세기 초 항일 의병 활동기. 정확한 극중 연도는 명시되지 않음.\n- 20세기 초 대한제국 말기의 생활상과 물질문화를 기준으로 삼는다. 조선 전기의 사극 양식이나 일제강점기 후기의 복식·장비를 혼합하지 않으며, 명시되지 않은 정확한 연도는 임의로 확정하지 않는다.\n- 화승총과 근대식 소총은 발화·장전 구조와 외형을 구별하고, 기관총은 해당 시대에 부합하는 형식으로 제한한다. 농기구는 실제 작업 도구의 크기와 재질을 유지한다. 운송·기록·생활 설비는 지게, 수레, 말, 종이 장부 등 원문 근거를 따르며 현대 차량·전자기기·전술 장비를 추가하지 않는다.\n\nTHINGS THAT LIVE AT THIS PLACE (what they are, not where they go): the\nlist above names what permanently belongs to this site. Where each one\nsits in this frame is decided by the attached sketch alone — do not add\nor reposition anything on the strength of the list, build only what the\nsketch already has lines for, and do not invent facilities that are not\nlisted.\n\nNo readable writing anywhere: surfaces that would carry writing may be\npresent, but stage any wording out of legibility — an oblique angle,\ndistance, shallow focus. No captions, watermarks or overlay text,\nand none of the sketch's diagram symbols, lines or markers anywhere in\nthe image.",
  "effective_prompt": "Turn the attached storyboard sketch into a photorealistic film still of\nits LOCATION ONLY, keeping the figures exactly as they are.\n\nKEEP EXACTLY: the camera framing, the horizon, and where every element\nsits in the frame. Each mannequin figure stays a plain grey featureless\nmannequin standing in the very same spot, at the same size, in the same\npose, turned the same way — do not turn them into people, do not move,\nrotate, mirror or re-pose them, do not add or remove figures.\n\nBUILD PHOTOREALISTICALLY: everything that is not a figure — ground,\nsurfacing, structures, vegetation, sky, water, distance. The bare lines\nof the sketch are a layout guide; replace them with the real materials,\ndepth and lighting of the place described below.\n\nSHOT TEXT this background must serve (Korean): 멀어지는 두 사람의 등 뒤로 거리를 둔 채 산길을 향해 한 발을 뻗은 미드 액션 상태인 염재보의 뒷모습.\n\nLOCATION (lock): The ascending mountain path just beyond the guardian-tree fork, trailing the route taken by two travelers.\n\nTIME OF DAY (lock): dawn, dense fog.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: wide shot\n- FRAME LAYOUT: Uninterrupted uphill path separating 염재보 from the departing pair in the middle-center of the frame, midground.\n- KEY BACKGROUND ELEMENTS: 산길 (The pair is walking uphill, with 염재보 beginning to follow after a delay) — Recedes from the lower-right area toward the upper-left background; used as Its visible empty interval makes the delayed following action unmistakable.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Subdued natural dawn ambience maintains sober continuity, with sufficient tonal separation to read all three figures and the gap between them.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHINGS AT THIS PLACE:\nThe site follows a connected route from a paddy embankment path to a mountain path and a junction at the mountain foot. Paddy plots directly border both sides of the embankment path. Beyond the fields, the mountain path connects to the junction and continues uphill beyond it. A separate branch leaves the junction to one side; its destination is unspecified. The village guardian tree stands beside the junction, opposite the branch opening, with its base adjoining the accessible junction ground rather than obstructing the route. For the site plan, place the field-path approach toward the bottom and the uphill continuation toward the top; these are drawing directions, not compass bearings. Preserve an open path connection through the junction and accessible ground beside the tree.\n- path: Paddy embankment path\n- fields: Flanking paddy plots\n- junction: Mountain-foot path junction\n- tree: Mountain-foot village guardian tree\n- path: Lateral junction branch\n- path: Mountain approach and uphill path\n\nWORLD FACTS (creator-confirmed — always true):\n- Region (real-world reference): 한반도 전라남도 보성 일대의 농촌·산지·장터 및 인근 해안권.\n- Era: 대한제국 말기, 20세기 초 항일 의병 활동기. 정확한 극중 연도는 명시되지 않음.\n- 20세기 초 대한제국 말기의 생활상과 물질문화를 기준으로 삼는다. 조선 전기의 사극 양식이나 일제강점기 후기의 복식·장비를 혼합하지 않으며, 명시되지 않은 정확한 연도는 임의로 확정하지 않는다.\n- 화승총과 근대식 소총은 발화·장전 구조와 외형을 구별하고, 기관총은 해당 시대에 부합하는 형식으로 제한한다. 농기구는 실제 작업 도구의 크기와 재질을 유지한다. 운송·기록·생활 설비는 지게, 수레, 말, 종이 장부 등 원문 근거를 따르며 현대 차량·전자기기·전술 장비를 추가하지 않는다.\n\nTHINGS THAT LIVE AT THIS PLACE (what they are, not where they go): the\nlist above names what permanently belongs to this site. Where each one\nsits in this frame is decided by the attached sketch alone — do not add\nor reposition anything on the strength of the list, build only what the\nsketch already has lines for, and do not invent facilities that are not\nlisted.\n\nNo readable writing anywhere: surfaces that would carry writing may be\npresent, but stage any wording out of legibility — an oblique angle,\ndistance, shallow focus. No captions, watermarks or overlay text,\nand none of the sketch's diagram symbols, lines or markers anywhere in\nthe image.",
  "bg_path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/db00517b-165b-4005-a287-f28936b59b95/images/272bd7af-9294-4490-98c6-88a40e9deacc/scene/recipe/S12sh10__bgfirst_bg.png",
  "asset_id": "6f415a70-e5cd-4a43-94f9-ff5d1f8f8a16",
  "input_asset_ids": [
   "d286f973-862c-4243-8f09-c550226664ee"
  ]
 },
 "S12sh10": {
  "input_fingerprint": "18ef3fb948111b08",
  "prompt": "Create ONE FINAL photorealistic live-action film still of the moment below — early 20th-century Korea, Jeolla region; all people are Korean unless explicitly stated otherwise. TIME OF DAY (lock): dawn, dense fog.\n\nSHOT TEXT (authoritative, Korean): 멀어지는 두 사람의 등 뒤로 거리를 둔 채 산길을 향해 한 발을 뻗은 미드 액션 상태인 염재보의 뒷모습.\n\nLOCATION (lock): The ascending mountain path just beyond the guardian-tree fork, trailing the route taken by two travelers. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nFRAMING SCALE (follow exactly — this alone decides how much of the frame the subject fills; the storyboard sketch supplies where the figures stand and which way they face, never this scale):\n- FRAMING SCALE: wide shot\nMatch this shot size exactly. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Subdued natural dawn ambience maintains sober continuity, with sufficient tonal separation to read all three figures and the gap between them.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The clothing, hair and overall look of 염재보 — who appear both in that photo and in this shot — are LOCKED to that photo. Anyone else visible in that photo is NOT in this shot: never carry their face, body or clothing onto anyone here. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): Dawn fog persists along the mountain path beyond the tree fork. 염재보: He has tightened his straw-shoe ties and now follows on foot without a gun. 안규홍: He ascends with his shoulder bundle and concealed jade ring. 이관회: He ascends carrying his travel bundle.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 안규홍 (한국인 남성, 청년 나이대, 검은 머리, 단단히 묶은 머리); 이관회 (한국인 남성, 청년 나이대, 자연스러운 검은 머리); 염재보 (한국인 남성, 성인 나이대, 자연스러운 검은 머리) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.",
  "roll_prompts": {
   "A": "Replace every grey or outlined mannequin figure in the FIRST attached\nimage with the real people the CHARACTER REFERENCE images show, and\noutput a photorealistic film still. Each mannequin becomes the character\nwhose POSE CANON and the shot text match that figure's pose and\nposition; never repeat one person across two mannequins. If no CHARACTER\nREFERENCE image is attached for a figure, still replace it with a\nplausible real person consistent with the text. Every mannequin becomes\na person — erasing a figure, or leaving one out, is not a replacement,\nand the number of figures never changes.\n\nKEEP EXACTLY: the background of the first image, the camera framing,\nand each mannequin's position, size, pose and the direction it is\nturned. A person must be exactly where their mannequin was, at the same\nscale, in the same pose, facing the same way — never mirrored, never\nre-staged, never straightened up, stood upright or otherwise re-posed.\nNo mannequin, grey figure or sketch line may remain anywhere in the\noutput.\n\nCreate ONE FINAL photorealistic live-action film still of the moment below — early 20th-century Korea, Jeolla region; all people are Korean unless explicitly stated otherwise. TIME OF DAY (lock): dawn, dense fog.\n\nSHOT TEXT (authoritative, Korean): 멀어지는 두 사람의 등 뒤로 거리를 둔 채 산길을 향해 한 발을 뻗은 미드 액션 상태인 염재보의 뒷모습.\n\nLOCATION (lock): The ascending mountain path just beyond the guardian-tree fork, trailing the route taken by two travelers. The shot takes place here — the FIRST attached image (SHOT BACKGROUND) is this exact place, already built: its ground, structures, horizon, materials and lighting are the finished truth of this location and must not be redesigned or replaced. No location photograph is attached — read the place from that image alone, and add no scenery, structure, vehicle or fixture that it does not already show. This lock governs the place only; the figures in the shot follow the staging and pose instructions.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: wide shot\n- FRAME LAYOUT: Uninterrupted uphill path separating 염재보 from the departing pair in the middle-center of the frame, midground.\n- KEY BACKGROUND ELEMENTS: 산길 (The pair is walking uphill, with 염재보 beginning to follow after a delay) — Recedes from the lower-right area toward the upper-left background; used as Its visible empty interval makes the delayed following action unmistakable.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Subdued natural dawn ambience maintains sober continuity, with sufficient tonal separation to read all three figures and the gap between them.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): Dawn fog persists along the mountain path beyond the tree fork. 염재보: He has tightened his straw-shoe ties and now follows on foot without a gun. 안규홍: He ascends with his shoulder bundle and concealed jade ring. 이관회: He ascends carrying his travel bundle.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 안규홍 (한국인 남성, 청년 나이대, 검은 머리, 단단히 묶은 머리); 이관회 (한국인 남성, 청년 나이대, 자연스러운 검은 머리); 염재보 (한국인 남성, 성인 나이대, 자연스러운 검은 머리) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.",
   "B": "Replace every grey or outlined mannequin figure in the FIRST attached\nimage with the real people the CHARACTER REFERENCE images show, and\noutput a photorealistic film still. Each mannequin becomes the character\nwhose POSE CANON and the shot text match that figure's pose and\nposition; never repeat one person across two mannequins. If no CHARACTER\nREFERENCE image is attached for a figure, still replace it with a\nplausible real person consistent with the text. Every mannequin becomes\na person — erasing a figure, or leaving one out, is not a replacement,\nand the number of figures never changes.\n\nKEEP EXACTLY: the background of the first image, the camera framing,\nand each mannequin's position, size, pose and the direction it is\nturned. A person must be exactly where their mannequin was, at the same\nscale, in the same pose, facing the same way — never mirrored, never\nre-staged, never straightened up, stood upright or otherwise re-posed.\nNo mannequin, grey figure or sketch line may remain anywhere in the\noutput.\n\nCreate ONE FINAL photorealistic live-action film still of the moment below — early 20th-century Korea, Jeolla region; all people are Korean unless explicitly stated otherwise. TIME OF DAY (lock): dawn, dense fog.\n\nSHOT TEXT (authoritative, Korean): 멀어지는 두 사람의 등 뒤로 거리를 둔 채 산길을 향해 한 발을 뻗은 미드 액션 상태인 염재보의 뒷모습.\n\nLOCATION (lock): The ascending mountain path just beyond the guardian-tree fork, trailing the route taken by two travelers. The shot takes place here — the FIRST attached image (SHOT BACKGROUND) is this exact place, already built: its ground, structures, horizon, materials and lighting are the finished truth of this location and must not be redesigned or replaced. No location photograph is attached — read the place from that image alone, and add no scenery, structure, vehicle or fixture that it does not already show. This lock governs the place only; the figures in the shot follow the staging and pose instructions.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: wide shot\n- FRAME LAYOUT: Uninterrupted uphill path separating 염재보 from the departing pair in the middle-center of the frame, midground.\n- KEY BACKGROUND ELEMENTS: 산길 (The pair is walking uphill, with 염재보 beginning to follow after a delay) — Recedes from the lower-right area toward the upper-left background; used as Its visible empty interval makes the delayed following action unmistakable.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Subdued natural dawn ambience maintains sober continuity, with sufficient tonal separation to read all three figures and the gap between them.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): Dawn fog persists along the mountain path beyond the tree fork. 염재보: He has tightened his straw-shoe ties and now follows on foot without a gun. 안규홍: He ascends with his shoulder bundle and concealed jade ring. 이관회: He ascends carrying his travel bundle.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 안규홍 (한국인 남성, 청년 나이대, 검은 머리, 단단히 묶은 머리); 이관회 (한국인 남성, 청년 나이대, 자연스러운 검은 머리); 염재보 (한국인 남성, 성인 나이대, 자연스러운 검은 머리) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.\n\nCOMPOSITION VARIATION (this candidate only): keep the place, the people and the moment exactly as specified, and keep from the CAMERA & FRAME contract its subject scale, screen placement and key background elements exactly. For this candidate only, its camera angle and height are deliberately left open: choose a camera position distinctly different from the obvious one, as a film director picking a second setup on the same blocking. Stay true to any explicit framing the SHOT TEXT itself dictates."
  },
  "roll_refs": {
   "A": [
    {
     "label": "SHOT BACKGROUND",
     "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/db00517b-165b-4005-a287-f28936b59b95/images/272bd7af-9294-4490-98c6-88a40e9deacc/scene/recipe/S12sh10__bgfirst_bg.png",
     "asset_id": "6f415a70-e5cd-4a43-94f9-ff5d1f8f8a16",
     "role": "bgfirst_bg"
    },
    {
     "label": "CHARACTER REFERENCE — 안규홍: the exact person appearing in this shot; match face, hair and build exactly.",
     "path": "<bytes:1028384>",
     "asset_id": "3b5894db-ce55-4f34-b966-5e725c3aa952",
     "role": "character_ref"
    },
    {
     "label": "CHARACTER REFERENCE — 이관회: the exact person appearing in this shot; match face, hair and build exactly.",
     "path": "<bytes:933540>",
     "asset_id": "851b450f-ac84-47fe-ac5a-03a56f63fcae",
     "role": "character_ref"
    },
    {
     "label": "CHARACTER REFERENCE — 염재보: the exact person appearing in this shot; match face, hair and build exactly.",
     "path": "<bytes:964351>",
     "asset_id": "ab287289-d0b7-436d-8f70-68452f898f73",
     "role": "character_ref"
    },
    {
     "label": "PROP REFERENCE — 보자기 꾸러미·봇짐: the exact object appearing in this shot; match its look, material and wear exactly.",
     "path": "<bytes:1465734>",
     "asset_id": "3ca92b24-68b4-4883-b851-cabc06c00d1a",
     "role": "prop_ref"
    }
   ],
   "B": [
    {
     "label": "SHOT BACKGROUND",
     "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/db00517b-165b-4005-a287-f28936b59b95/images/272bd7af-9294-4490-98c6-88a40e9deacc/scene/recipe/S12sh10__bgfirst_bg.png",
     "asset_id": "6f415a70-e5cd-4a43-94f9-ff5d1f8f8a16",
     "role": "bgfirst_bg"
    },
    {
     "label": "CHARACTER REFERENCE — 안규홍: the exact person appearing in this shot; match face, hair and build exactly.",
     "path": "<bytes:1028384>",
     "asset_id": "3b5894db-ce55-4f34-b966-5e725c3aa952",
     "role": "character_ref"
    },
    {
     "label": "CHARACTER REFERENCE — 이관회: the exact person appearing in this shot; match face, hair and build exactly.",
     "path": "<bytes:933540>",
     "asset_id": "851b450f-ac84-47fe-ac5a-03a56f63fcae",
     "role": "character_ref"
    },
    {
     "label": "CHARACTER REFERENCE — 염재보: the exact person appearing in this shot; match face, hair and build exactly.",
     "path": "<bytes:964351>",
     "asset_id": "ab287289-d0b7-436d-8f70-68452f898f73",
     "role": "character_ref"
    },
    {
     "label": "PROP REFERENCE — 보자기 꾸러미·봇짐: the exact object appearing in this shot; match its look, material and wear exactly.",
     "path": "<bytes:1465734>",
     "asset_id": "3ca92b24-68b4-4883-b851-cabc06c00d1a",
     "role": "prop_ref"
    }
   ]
  },
  "shot_run_spend_attempt_count": 1,
  "readings": [
   {
    "label": "A",
    "direction": "세 인물 모두 안개 낀 산길 위쪽을 향해 걷고 있으며, 염재보는 앞서가는 두 사람의 등 뒤를 향해 시선과 방향을 두고 있음.",
    "built_space": "갈림길, 흙길, 왼편의 나무와 오른편의 논 등 이전 샷에 설정된 배경 요소들이 정확한 위치에 유지됨.",
    "entities": "안규홍(왼쪽)은 흰옷과 묶은 머리, 레퍼런스와 일치하는 녹색 봇짐을 메고 있음. 이관회(오른쪽)는 갈색 옷을 입고 다른 꾸러미를 들고 있음. 염재보는 의상 레퍼런스와 일치하는 누비옷을 입었으나, 삿갓 대신 챙이 없는 검은 두건을 착용함.",
    "hard_violations": [],
    "physics": "모든 인물이 지면을 딛고 자연스럽게 걷고 있으며, 염재보 역시 땅을 차고 나가는 걷기 액션이 물리적으로 타당함."
   },
   {
    "label": "B",
    "direction": "세 인물 모두 안개 낀 오르막길을 향해 전진하고 있으며 올바른 방향을 향함.",
    "built_space": "이전 샷의 배경인 나무, 흙길 갈림길, 논 등의 위치와 형태가 적절히 유지됨.",
    "entities": "염재보가 삿갓을 전혀 쓰지 않고 맨머리 상태임. 이관회(왼쪽, 갈색 옷)가 녹색 봇짐을 매고 있고, 안규홍(오른쪽, 흰 옷)이 일반 꾸러미를 메고 있어 두 인물의 소품 배치가 뒤바뀜.",
    "hard_violations": [],
    "physics": "인물들의 걷는 자세와 발이 땅에 닿아 있는 등 체중 지지가 정상적이며 떠 있는 요소가 없음."
   }
  ],
  "cross_model_order": {
   "policy": "cross_model_order_v2_slot0_fwd_slot1_rev_parallel2",
   "slots": [
    {
     "model": "gemini-pro",
     "order": "forward",
     "display_to_canonical": {
      "A": "A",
      "B": "B"
     },
     "raw": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 8,
        "verdict_ko": "염재보의 삿갓이 검은 두건으로 대체된 점은 아쉬우나, 안규홍(흰 옷, 녹색 봇짐)과 이관회(갈색 옷)의 의상 및 소품 배치가 프롬프트와 정확히 일치하여 훌륭합니다."
       },
       {
        "label": "B",
        "score": 6,
        "verdict_ko": "염재보가 모자를 쓰지 않았으며, 안규홍이 메어야 할 녹색 봇짐을 이관회가 메고 있어 인물 간 소품 배치가 어긋났습니다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "세 인물 모두 안개 낀 산길 위쪽을 향해 걷고 있으며, 염재보는 앞서가는 두 사람의 등 뒤를 향해 시선과 방향을 두고 있음.",
        "built_space": "갈림길, 흙길, 왼편의 나무와 오른편의 논 등 이전 샷에 설정된 배경 요소들이 정확한 위치에 유지됨.",
        "entities": "안규홍(왼쪽)은 흰옷과 묶은 머리, 레퍼런스와 일치하는 녹색 봇짐을 메고 있음. 이관회(오른쪽)는 갈색 옷을 입고 다른 꾸러미를 들고 있음. 염재보는 의상 레퍼런스와 일치하는 누비옷을 입었으나, 삿갓 대신 챙이 없는 검은 두건을 착용함.",
        "hard_violations": [],
        "physics": "모든 인물이 지면을 딛고 자연스럽게 걷고 있으며, 염재보 역시 땅을 차고 나가는 걷기 액션이 물리적으로 타당함."
       },
       {
        "label": "B",
        "direction": "세 인물 모두 안개 낀 오르막길을 향해 전진하고 있으며 올바른 방향을 향함.",
        "built_space": "이전 샷의 배경인 나무, 흙길 갈림길, 논 등의 위치와 형태가 적절히 유지됨.",
        "entities": "염재보가 삿갓을 전혀 쓰지 않고 맨머리 상태임. 이관회(왼쪽, 갈색 옷)가 녹색 봇짐을 매고 있고, 안규홍(오른쪽, 흰 옷)이 일반 꾸러미를 메고 있어 두 인물의 소품 배치가 뒤바뀜.",
        "hard_violations": [],
        "physics": "인물들의 걷는 자세와 발이 땅에 닿아 있는 등 체중 지지가 정상적이며 떠 있는 요소가 없음."
       }
      ],
      "all_candidates_fail": false
     },
     "normalized": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 8,
        "verdict_ko": "염재보의 삿갓이 검은 두건으로 대체된 점은 아쉬우나, 안규홍(흰 옷, 녹색 봇짐)과 이관회(갈색 옷)의 의상 및 소품 배치가 프롬프트와 정확히 일치하여 훌륭합니다."
       },
       {
        "label": "B",
        "score": 6,
        "verdict_ko": "염재보가 모자를 쓰지 않았으며, 안규홍이 메어야 할 녹색 봇짐을 이관회가 메고 있어 인물 간 소품 배치가 어긋났습니다."
       }
      ],
      "all_candidates_fail": false,
      "readings": [
       {
        "label": "A",
        "direction": "세 인물 모두 안개 낀 산길 위쪽을 향해 걷고 있으며, 염재보는 앞서가는 두 사람의 등 뒤를 향해 시선과 방향을 두고 있음.",
        "built_space": "갈림길, 흙길, 왼편의 나무와 오른편의 논 등 이전 샷에 설정된 배경 요소들이 정확한 위치에 유지됨.",
        "entities": "안규홍(왼쪽)은 흰옷과 묶은 머리, 레퍼런스와 일치하는 녹색 봇짐을 메고 있음. 이관회(오른쪽)는 갈색 옷을 입고 다른 꾸러미를 들고 있음. 염재보는 의상 레퍼런스와 일치하는 누비옷을 입었으나, 삿갓 대신 챙이 없는 검은 두건을 착용함.",
        "hard_violations": [],
        "physics": "모든 인물이 지면을 딛고 자연스럽게 걷고 있으며, 염재보 역시 땅을 차고 나가는 걷기 액션이 물리적으로 타당함."
       },
       {
        "label": "B",
        "direction": "세 인물 모두 안개 낀 오르막길을 향해 전진하고 있으며 올바른 방향을 향함.",
        "built_space": "이전 샷의 배경인 나무, 흙길 갈림길, 논 등의 위치와 형태가 적절히 유지됨.",
        "entities": "염재보가 삿갓을 전혀 쓰지 않고 맨머리 상태임. 이관회(왼쪽, 갈색 옷)가 녹색 봇짐을 매고 있고, 안규홍(오른쪽, 흰 옷)이 일반 꾸러미를 메고 있어 두 인물의 소품 배치가 뒤바뀜.",
        "hard_violations": [],
        "physics": "인물들의 걷는 자세와 발이 땅에 닿아 있는 등 체중 지지가 정상적이며 떠 있는 요소가 없음."
       }
      ]
     },
     "ok": true
    },
    {
     "model": "gpt-high",
     "order": "reverse",
     "display_to_canonical": {
      "A": "B",
      "B": "A"
     },
     "raw": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 8,
        "verdict_ko": "와이드 뒷모습과 추종 간격, 오르막 보행을 충실히 구현하고 안규홍의 묶은 머리와 청록색 봇짐도 잘 맞지만, 염재보의 삿갓이 두건으로 바뀌었다."
       },
       {
        "label": "A",
        "score": 7,
        "verdict_ko": "장소와 와이드 구도, 염재보의 지지발이 있는 보행 순간은 맞지만, 삿갓이 없고 안규홍의 머리와 봇짐 형태가 참고보다 덜 충실하다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "세 사람 모두 카메라에 등을 보이고 화면 중앙 위쪽의 오르막 산길을 향한다. 오른쪽 전경의 염재보는 앞선 두 사람 쪽으로 몸과 머리를 돌리고 따라간다. 얼굴이 가려져 정확한 눈동자 방향은 확인되지 않는다. 총이나 겨냥하는 물체는 없다.",
        "built_space": "건물은 없으며, 왼쪽 당산나무 한 그루, 그 아래 왼쪽으로 갈라지는 샛길, 중앙의 오르막 흙길, 오른쪽 물 댄 논이 참고와 같은 배치를 이룬다. 세 사람 모두 흙길 위에 있고 염재보는 오른쪽 전경에서 두 여행자와 간격을 둔다. 전신과 주변 지형을 담은 와이드 구도이며, 새벽 안개와 길의 상승 방향도 유지된다.",
        "entities": "실제 사람으로 묘사된 남성 세 명만 보인다. 뒷모습이어서 한국인 여부와 정확한 얼굴 일치는 확인할 수 없지만 외형과 체격은 지정된 성인·청년 남성 설정에 어긋나지 않는다. 앞쪽 이관회는 검은 머리, 갈색 저고리, 짧은 밝은 바지와 짚신을 착용한다. 가운데 안규홍은 밝은 저고리와 긴 바지를 입었으나 머리가 단단히 묶였는지는 불명확하다. 염재보는 검은 누비옷과 각반을 착용하고 총이 없지만 참고의 삿갓이 빠졌다. 여행자마다 봇짐 하나가 있으며 이관회의 청록색 봇짐은 참고 색상에 가깝고 안규홍의 회갈색 봇짐은 다르다. 숨긴 옥반지는 드러나지 않으며 읽을 수 있는 글자도 없다.",
        "hard_violations": [],
        "physics": "염재보의 왼발은 흙길에 닿아 체중을 지지하고 오른발은 뒤에서 뒤꿈치가 들린 보행 단계다. 앞선 두 사람도 각각 지면에 닿은 발로 지지되며 공중에 뜬 몸은 없다. 봇짐은 어깨를 지나는 끈으로 등에 매달려 있고 천이 아래로 처진다. 팔의 굽힘과 다리 자세는 걷는 동작으로 가능하다."
       },
       {
        "label": "B",
        "direction": "앞선 두 여행자는 중앙 위쪽 오르막으로 멀어지고 오른쪽 전경의 염재보도 같은 길과 두 사람을 향해 등을 보인다. 머리 방향 역시 진행 경로를 향하며 카메라를 돌아보지 않는다. 눈동자는 보이지 않는다. 무기나 별도의 지향성 소품은 없다.",
        "built_space": "왼쪽 당산나무 한 그루와 갈림길, 중앙의 좁아지며 올라가는 흙길, 오른쪽 논과 논둑이 참고 배치를 유지한다. 염재보는 오른쪽 전경, 두 여행자는 중앙의 서로 다른 깊이에 놓여 세 사람 사이의 거리가 읽힌다. 모두 길 위에 있으며 논 안에 잘못 들어선 사람은 없다. 와이드 촬영 범위와 낮은 새벽빛, 짙은 안개가 유지된다.",
        "entities": "남성 세 명이 구체적인 인체와 의복으로 표현되어 있다. 얼굴은 뒤로 가려져 정확한 신원과 민족성은 직접 검증할 수 없다. 선두 안규홍의 묶은 검은 머리, 밝은 저고리와 긴 바지가 참고에 가깝다. 가운데 이관회의 짧은 검은 머리, 갈색 저고리와 짧은 밝은 바지도 맞는다. 염재보의 검은 누비옷과 각반은 일치하지만 삿갓 대신 검은 두건을 썼다. 안규홍의 청록색 어깨 봇짐은 참고 색상과 재질에 가깝고, 이관회의 회갈색 봇짐은 참고와 색상·형태가 다르다. 총과 노출된 옥반지는 없고 읽을 수 있는 글자도 없다.",
        "hard_violations": [],
        "physics": "염재보는 앞으로 놓인 왼발로 땅을 지지하고 뒤쪽 오른발의 뒤꿈치를 들어 다음 걸음으로 넘어간다. 두 여행자 역시 지면에 닿은 발이 있으며 선두의 들린 발은 정상적인 보행 중간 단계다. 두 봇짐은 어깨에 걸린 끈으로 지지되어 등과 옆구리에 닿는다. 지지 없이 떠 있는 사람이나 물체는 없다."
       }
      ],
      "all_candidates_fail": false
     },
     "normalized": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 8,
        "verdict_ko": "와이드 뒷모습과 추종 간격, 오르막 보행을 충실히 구현하고 안규홍의 묶은 머리와 청록색 봇짐도 잘 맞지만, 염재보의 삿갓이 두건으로 바뀌었다."
       },
       {
        "label": "B",
        "score": 7,
        "verdict_ko": "장소와 와이드 구도, 염재보의 지지발이 있는 보행 순간은 맞지만, 삿갓이 없고 안규홍의 머리와 봇짐 형태가 참고보다 덜 충실하다."
       }
      ],
      "all_candidates_fail": false,
      "readings": [
       {
        "label": "B",
        "direction": "세 사람 모두 카메라에 등을 보이고 화면 중앙 위쪽의 오르막 산길을 향한다. 오른쪽 전경의 염재보는 앞선 두 사람 쪽으로 몸과 머리를 돌리고 따라간다. 얼굴이 가려져 정확한 눈동자 방향은 확인되지 않는다. 총이나 겨냥하는 물체는 없다.",
        "built_space": "건물은 없으며, 왼쪽 당산나무 한 그루, 그 아래 왼쪽으로 갈라지는 샛길, 중앙의 오르막 흙길, 오른쪽 물 댄 논이 참고와 같은 배치를 이룬다. 세 사람 모두 흙길 위에 있고 염재보는 오른쪽 전경에서 두 여행자와 간격을 둔다. 전신과 주변 지형을 담은 와이드 구도이며, 새벽 안개와 길의 상승 방향도 유지된다.",
        "entities": "실제 사람으로 묘사된 남성 세 명만 보인다. 뒷모습이어서 한국인 여부와 정확한 얼굴 일치는 확인할 수 없지만 외형과 체격은 지정된 성인·청년 남성 설정에 어긋나지 않는다. 앞쪽 이관회는 검은 머리, 갈색 저고리, 짧은 밝은 바지와 짚신을 착용한다. 가운데 안규홍은 밝은 저고리와 긴 바지를 입었으나 머리가 단단히 묶였는지는 불명확하다. 염재보는 검은 누비옷과 각반을 착용하고 총이 없지만 참고의 삿갓이 빠졌다. 여행자마다 봇짐 하나가 있으며 이관회의 청록색 봇짐은 참고 색상에 가깝고 안규홍의 회갈색 봇짐은 다르다. 숨긴 옥반지는 드러나지 않으며 읽을 수 있는 글자도 없다.",
        "hard_violations": [],
        "physics": "염재보의 왼발은 흙길에 닿아 체중을 지지하고 오른발은 뒤에서 뒤꿈치가 들린 보행 단계다. 앞선 두 사람도 각각 지면에 닿은 발로 지지되며 공중에 뜬 몸은 없다. 봇짐은 어깨를 지나는 끈으로 등에 매달려 있고 천이 아래로 처진다. 팔의 굽힘과 다리 자세는 걷는 동작으로 가능하다."
       },
       {
        "label": "A",
        "direction": "앞선 두 여행자는 중앙 위쪽 오르막으로 멀어지고 오른쪽 전경의 염재보도 같은 길과 두 사람을 향해 등을 보인다. 머리 방향 역시 진행 경로를 향하며 카메라를 돌아보지 않는다. 눈동자는 보이지 않는다. 무기나 별도의 지향성 소품은 없다.",
        "built_space": "왼쪽 당산나무 한 그루와 갈림길, 중앙의 좁아지며 올라가는 흙길, 오른쪽 논과 논둑이 참고 배치를 유지한다. 염재보는 오른쪽 전경, 두 여행자는 중앙의 서로 다른 깊이에 놓여 세 사람 사이의 거리가 읽힌다. 모두 길 위에 있으며 논 안에 잘못 들어선 사람은 없다. 와이드 촬영 범위와 낮은 새벽빛, 짙은 안개가 유지된다.",
        "entities": "남성 세 명이 구체적인 인체와 의복으로 표현되어 있다. 얼굴은 뒤로 가려져 정확한 신원과 민족성은 직접 검증할 수 없다. 선두 안규홍의 묶은 검은 머리, 밝은 저고리와 긴 바지가 참고에 가깝다. 가운데 이관회의 짧은 검은 머리, 갈색 저고리와 짧은 밝은 바지도 맞는다. 염재보의 검은 누비옷과 각반은 일치하지만 삿갓 대신 검은 두건을 썼다. 안규홍의 청록색 어깨 봇짐은 참고 색상과 재질에 가깝고, 이관회의 회갈색 봇짐은 참고와 색상·형태가 다르다. 총과 노출된 옥반지는 없고 읽을 수 있는 글자도 없다.",
        "hard_violations": [],
        "physics": "염재보는 앞으로 놓인 왼발로 땅을 지지하고 뒤쪽 오른발의 뒤꿈치를 들어 다음 걸음으로 넘어간다. 두 여행자 역시 지면에 닿은 발이 있으며 선두의 들린 발은 정상적인 보행 중간 단계다. 두 봇짐은 어깨에 걸린 끈으로 지지되어 등과 옆구리에 닿는다. 지지 없이 떠 있는 사람이나 물체는 없다."
       }
      ]
     },
     "ok": true
    }
   ],
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "slot_winner_match": true,
   "slot_winner": {
    "gemini-pro": "A",
    "gpt-high": "A"
   },
   "route": "cross_slot_agree"
  },
  "dual": {
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "normalized": {
    "A": 2.0,
    "B": 1.625
   },
   "adjusted": {
    "A": 2.0,
    "B": 1.625
   },
   "violations": {},
   "per_model_winner": {
    "gemini-pro": "A",
    "gpt-high": "A"
   },
   "agreed": true
  },
  "totals": {
   "A": 2000,
   "B": 1625
  },
  "selected": "A",
  "ranking": [
   "A",
   "B"
  ],
  "verdicts": [
   {
    "label": "A",
    "score": 2000,
    "verdict_ko": "염재보의 삿갓이 검은 두건으로 대체된 점은 아쉬우나, 안규홍(흰 옷, 녹색 봇짐)과 이관회(갈색 옷)의 의상 및 소품 배치가 프롬프트와 정확히 일치하여 훌륭합니다."
   },
   {
    "label": "B",
    "score": 1625,
    "verdict_ko": "염재보가 모자를 쓰지 않았으며, 안규홍이 메어야 할 녹색 봇짐을 이관회가 메고 있어 인물 간 소품 배치가 어긋났습니다."
   }
  ],
  "refs": [
   {
    "label": "SHOT BACKGROUND",
    "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/db00517b-165b-4005-a287-f28936b59b95/images/272bd7af-9294-4490-98c6-88a40e9deacc/scene/recipe/S12sh10__bgfirst_bg.png",
    "asset_id": "6f415a70-e5cd-4a43-94f9-ff5d1f8f8a16",
    "role": "bgfirst_bg"
   },
   {
    "label": "CHARACTER REFERENCE — 안규홍: the exact person appearing in this shot; match face, hair and build exactly.",
    "path": "<bytes:1028384>",
    "asset_id": "3b5894db-ce55-4f34-b966-5e725c3aa952",
    "role": "character_ref"
   },
   {
    "label": "CHARACTER REFERENCE — 이관회: the exact person appearing in this shot; match face, hair and build exactly.",
    "path": "<bytes:933540>",
    "asset_id": "851b450f-ac84-47fe-ac5a-03a56f63fcae",
    "role": "character_ref"
   },
   {
    "label": "CHARACTER REFERENCE — 염재보: the exact person appearing in this shot; match face, hair and build exactly.",
    "path": "<bytes:964351>",
    "asset_id": "ab287289-d0b7-436d-8f70-68452f898f73",
    "role": "character_ref"
   },
   {
    "label": "PROP REFERENCE — 보자기 꾸러미·봇짐: the exact object appearing in this shot; match its look, material and wear exactly.",
    "path": "<bytes:1465734>",
    "asset_id": "3ca92b24-68b4-4883-b851-cabc06c00d1a",
    "role": "prop_ref"
   }
  ],
  "critique_skipped": true,
  "shot_run_produced": false,
  "shot_run_uid": "06aa140f-18ff-7c97-a868-7631e8154e65",
  "bgfirst": {
   "bg_path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/db00517b-165b-4005-a287-f28936b59b95/images/272bd7af-9294-4490-98c6-88a40e9deacc/scene/recipe/S12sh10__bgfirst_bg.png",
   "bg_asset_id": "6f415a70-e5cd-4a43-94f9-ff5d1f8f8a16",
   "bg_record_key": "S12sh10::bgfirst_bg",
   "chain_winner": true,
   "authority": "lane_conti_only"
  },
  "ref_mode": "lane(map_marker): 스케치+prev+엔티티",
  "share_plan": {
   "ref_plan": "prev",
   "prev_anchor_tag": "S12sh5"
  }
 },
 "S12sh10::cine": {
  "applied": true,
  "attempted_at": "2026-09-09T02:06:52.642768+00:00",
  "fingerprint": "8866d190eec43ed396bf6edade723118aa28696d970f813b2bed402df1ab627a",
  "fingerprint_version": 2,
  "provider": "mai",
  "endpoint": "openrouter/chat-completions",
  "model": "microsoft/mai-image-2.6",
  "pack": "24.202608252115",
  "source_file": "S12sh10_sel.png",
  "source_sha256": "39e5afc2399efee76de9496b02bde628f931b98179022e94b33c0c9c812dc390",
  "file": "S12sh10_cine.png",
  "staged_sha256": "d02bebc23721d6d34cdad86ee575cb13df02ab732cd6a711c58f8b0e221ca8d5",
  "latency_ms": 25920
 },
 "S13sh3::signage": {
  "fp": "689fb6bc3c6e2278",
  "inscriptions": [],
  "cues": [],
  "dropped": []
 },
 "S13sh3": {
  "input_fingerprint": "baa0f8ac1b7d9e78",
  "prompt": "Create ONE FINAL photorealistic live-action film still of the moment below — early 20th-century Korea, Jeolla region; all people are Korean unless explicitly stated otherwise. TIME OF DAY (lock): morning.\n\nSHOT TEXT (authoritative, Korean): 양쪽의 빽빽한 대나무 틈새로 검은 총구 두 개가 숲길 중앙을 향해 일제히 튀어나온 찰나.\n\nLOCATION (lock): Both bamboo-lined edges of a narrow outdoor forest path, with opposing gaps opening toward the center of the trail. The shot takes place here — the LOCATION text above is the only authority for this place — no location photograph is attached. Build the place strictly from that text and the shot text, inventing nothing beyond them.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: wide shot\n- FRAME LAYOUT: Left bamboo opening with an inward-projecting muzzle in the middle-left of the frame, midground; Right bamboo opening with the opposing inward-projecting muzzle in the middle-right of the frame, midground.\n- KEY BACKGROUND ELEMENTS: 길 양쪽 대나무숲 (Dense bamboo conceals the gunmen on opposite sides of the narrow path) — Two flanking masses are viewed diagonally from above, with openings facing the path; used as Occludes the gunmen while permitting separate, unobstructed views of both muzzles; 양쪽 총구 (Both have emerged and are aimed at the travelers on the path) — Seen predominantly from the side, extending inward from opposing frame edges and angling toward the lower central path; used as Creates converging threat lines without exaggerated foreground scale; 좁은 숲길 (The descending route is covered from both sides) — Runs through the center beneath the elevated, diagonal viewpoint; used as Provides the spatial reference that makes the two-sided ambush intelligible.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Natural morning ambience retains sober contrast and muted saturation, making the emerging muzzles readable without theatrical lighting.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): A narrow morning path runs through dense, wind-stirred bamboo and rocks.\n\nNO PEOPLE IN THIS SHOT: the shot text shows only the place and its state — no living person appears in frame, and no face, and no figure standing, sitting or passing through. Do not invent someone to occupy the space.\n\nOne thing this rule does NOT forbid: the hand or arm of whoever is doing what the shot text describes. If the moment is an object being held, carried, operated, read, aimed, worn or handed over, the hand doing it belongs in frame, resting on or gripping that object exactly as its real use requires. An object photographed floating unsupported, with the hand that must be holding it erased, is a worse failure than the person you were avoiding — render the hand.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.",
  "roll_prompts": {
   "A": "Create ONE FINAL photorealistic live-action film still of the moment below — early 20th-century Korea, Jeolla region; all people are Korean unless explicitly stated otherwise. TIME OF DAY (lock): morning.\n\nSHOT TEXT (authoritative, Korean): 양쪽의 빽빽한 대나무 틈새로 검은 총구 두 개가 숲길 중앙을 향해 일제히 튀어나온 찰나.\n\nLOCATION (lock): Both bamboo-lined edges of a narrow outdoor forest path, with opposing gaps opening toward the center of the trail. The shot takes place here — the LOCATION text above is the only authority for this place — no location photograph is attached. Build the place strictly from that text and the shot text, inventing nothing beyond them.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: wide shot\n- FRAME LAYOUT: Left bamboo opening with an inward-projecting muzzle in the middle-left of the frame, midground; Right bamboo opening with the opposing inward-projecting muzzle in the middle-right of the frame, midground.\n- KEY BACKGROUND ELEMENTS: 길 양쪽 대나무숲 (Dense bamboo conceals the gunmen on opposite sides of the narrow path) — Two flanking masses are viewed diagonally from above, with openings facing the path; used as Occludes the gunmen while permitting separate, unobstructed views of both muzzles; 양쪽 총구 (Both have emerged and are aimed at the travelers on the path) — Seen predominantly from the side, extending inward from opposing frame edges and angling toward the lower central path; used as Creates converging threat lines without exaggerated foreground scale; 좁은 숲길 (The descending route is covered from both sides) — Runs through the center beneath the elevated, diagonal viewpoint; used as Provides the spatial reference that makes the two-sided ambush intelligible.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Natural morning ambience retains sober contrast and muted saturation, making the emerging muzzles readable without theatrical lighting.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): A narrow morning path runs through dense, wind-stirred bamboo and rocks.\n\nNO PEOPLE IN THIS SHOT: the shot text shows only the place and its state — no living person appears in frame, and no face, and no figure standing, sitting or passing through. Do not invent someone to occupy the space.\n\nOne thing this rule does NOT forbid: the hand or arm of whoever is doing what the shot text describes. If the moment is an object being held, carried, operated, read, aimed, worn or handed over, the hand doing it belongs in frame, resting on or gripping that object exactly as its real use requires. An object photographed floating unsupported, with the hand that must be holding it erased, is a worse failure than the person you were avoiding — render the hand.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.",
   "B": "Create ONE FINAL photorealistic live-action film still of the moment below — early 20th-century Korea, Jeolla region; all people are Korean unless explicitly stated otherwise. TIME OF DAY (lock): morning.\n\nSHOT TEXT (authoritative, Korean): 양쪽의 빽빽한 대나무 틈새로 검은 총구 두 개가 숲길 중앙을 향해 일제히 튀어나온 찰나.\n\nLOCATION (lock): Both bamboo-lined edges of a narrow outdoor forest path, with opposing gaps opening toward the center of the trail. The shot takes place here — the LOCATION text above is the only authority for this place — no location photograph is attached. Build the place strictly from that text and the shot text, inventing nothing beyond them.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: wide shot\n- FRAME LAYOUT: Left bamboo opening with an inward-projecting muzzle in the middle-left of the frame, midground; Right bamboo opening with the opposing inward-projecting muzzle in the middle-right of the frame, midground.\n- KEY BACKGROUND ELEMENTS: 길 양쪽 대나무숲 (Dense bamboo conceals the gunmen on opposite sides of the narrow path) — Two flanking masses are viewed diagonally from above, with openings facing the path; used as Occludes the gunmen while permitting separate, unobstructed views of both muzzles; 양쪽 총구 (Both have emerged and are aimed at the travelers on the path) — Seen predominantly from the side, extending inward from opposing frame edges and angling toward the lower central path; used as Creates converging threat lines without exaggerated foreground scale; 좁은 숲길 (The descending route is covered from both sides) — Runs through the center beneath the elevated, diagonal viewpoint; used as Provides the spatial reference that makes the two-sided ambush intelligible.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Natural morning ambience retains sober contrast and muted saturation, making the emerging muzzles readable without theatrical lighting.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): A narrow morning path runs through dense, wind-stirred bamboo and rocks.\n\nNO PEOPLE IN THIS SHOT: the shot text shows only the place and its state — no living person appears in frame, and no face, and no figure standing, sitting or passing through. Do not invent someone to occupy the space.\n\nOne thing this rule does NOT forbid: the hand or arm of whoever is doing what the shot text describes. If the moment is an object being held, carried, operated, read, aimed, worn or handed over, the hand doing it belongs in frame, resting on or gripping that object exactly as its real use requires. An object photographed floating unsupported, with the hand that must be holding it erased, is a worse failure than the person you were avoiding — render the hand.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.\n\nCOMPOSITION VARIATION (this candidate only): keep the place, the people and the moment exactly as specified, and keep from the CAMERA & FRAME contract its subject scale, screen placement and key background elements exactly. For this candidate only, its camera angle and height are deliberately left open: choose a camera position distinctly different from the obvious one, as a film director picking a second setup on the same blocking. Stay true to any explicit framing the SHOT TEXT itself dictates."
  },
  "shot_run_spend_attempt_count": 1,
  "readings": [
   {
    "label": "A",
    "direction": "양쪽 대나무 숲에서 돌출된 두 소총의 총구가 숲길 중앙 하단을 정확히 향하고 있음.",
    "built_space": "좁은 흙길 양옆으로 빽빽한 대나무 숲이 묘사되어 있으며, 위에서 내려다보는 대각선 구도를 잘 따름.",
    "entities": "대나무숲, 좁은 길, 두 개의 소총, 총을 쥔 손. 단, 왼쪽 대나무 틈으로 지시문에 금지된 사람의 얼굴(모자)과 몸통 일부가 보임.",
    "hard_violations": [
     "[gpt-high] 오른쪽 가장자리의 대나무 잎 뒤로 사수의 얼굴 일부가 노출되어, 얼굴은 보이지 않고 손·팔만 허용한다는 명시 조건을 위반한다."
    ],
    "physics": "양쪽 소총 모두 사람의 두 손과 팔에 의해 쥐어져 물리적으로 올바르게 지탱되고 있음."
   },
   {
    "label": "B",
    "direction": "양쪽 숲에서 나온 두 소총의 총구가 숲길 중앙을 겨냥하고 있음.",
    "built_space": "좁고 경사진 흙길을 양옆의 빽빽한 대나무 숲이 감싸고 있는 형태가 잘 구현됨.",
    "entities": "대나무숲, 좁은 길, 두 개의 소총. 오른쪽 총에는 손가락 일부가 보이지만, 왼쪽 총 주변에는 아무도 없음.",
    "hard_violations": [
     "[gemini-pro] 왼쪽 총을 쥐고 있는 손이나 지지대가 전혀 없어 무기가 허공에 떠 있음"
    ],
    "physics": "오른쪽 총은 손으로 쥐고 있으나, 왼쪽 총은 잡고 있는 손이나 지지물 없이 대나무 잎 사이에 떠 있음."
   }
  ],
  "cross_model_order": {
   "policy": "cross_model_order_v2_slot0_fwd_slot1_rev_parallel2",
   "slots": [
    {
     "model": "gemini-pro",
     "order": "forward",
     "display_to_canonical": {
      "A": "A",
      "B": "B"
     },
     "raw": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 8,
        "verdict_ko": "총을 쥐는 손을 묘사하여 물리적 오류를 피했으나, 화면에 등장하지 않아야 할 인물의 얼굴과 몸통이 왼쪽 틈새로 노출된 점이 감점 요인입니다."
       },
       {
        "label": "B",
        "score": 3,
        "verdict_ko": "인물을 배제하려다 왼쪽 총을 쥐어야 할 손마저 지워버려, 총이 허공에 떠 있는 치명적인 물리적 오류를 범했습니다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "양쪽 대나무 숲에서 돌출된 두 소총의 총구가 숲길 중앙 하단을 정확히 향하고 있음.",
        "built_space": "좁은 흙길 양옆으로 빽빽한 대나무 숲이 묘사되어 있으며, 위에서 내려다보는 대각선 구도를 잘 따름.",
        "entities": "대나무숲, 좁은 길, 두 개의 소총, 총을 쥔 손. 단, 왼쪽 대나무 틈으로 지시문에 금지된 사람의 얼굴(모자)과 몸통 일부가 보임.",
        "hard_violations": [],
        "physics": "양쪽 소총 모두 사람의 두 손과 팔에 의해 쥐어져 물리적으로 올바르게 지탱되고 있음."
       },
       {
        "label": "B",
        "direction": "양쪽 숲에서 나온 두 소총의 총구가 숲길 중앙을 겨냥하고 있음.",
        "built_space": "좁고 경사진 흙길을 양옆의 빽빽한 대나무 숲이 감싸고 있는 형태가 잘 구현됨.",
        "entities": "대나무숲, 좁은 길, 두 개의 소총. 오른쪽 총에는 손가락 일부가 보이지만, 왼쪽 총 주변에는 아무도 없음.",
        "hard_violations": [
         "왼쪽 총을 쥐고 있는 손이나 지지대가 전혀 없어 무기가 허공에 떠 있음"
        ],
        "physics": "오른쪽 총은 손으로 쥐고 있으나, 왼쪽 총은 잡고 있는 손이나 지지물 없이 대나무 잎 사이에 떠 있음."
       }
      ],
      "all_candidates_fail": false
     },
     "normalized": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 8,
        "verdict_ko": "총을 쥐는 손을 묘사하여 물리적 오류를 피했으나, 화면에 등장하지 않아야 할 인물의 얼굴과 몸통이 왼쪽 틈새로 노출된 점이 감점 요인입니다."
       },
       {
        "label": "B",
        "score": 3,
        "verdict_ko": "인물을 배제하려다 왼쪽 총을 쥐어야 할 손마저 지워버려, 총이 허공에 떠 있는 치명적인 물리적 오류를 범했습니다."
       }
      ],
      "all_candidates_fail": false,
      "readings": [
       {
        "label": "A",
        "direction": "양쪽 대나무 숲에서 돌출된 두 소총의 총구가 숲길 중앙 하단을 정확히 향하고 있음.",
        "built_space": "좁은 흙길 양옆으로 빽빽한 대나무 숲이 묘사되어 있으며, 위에서 내려다보는 대각선 구도를 잘 따름.",
        "entities": "대나무숲, 좁은 길, 두 개의 소총, 총을 쥔 손. 단, 왼쪽 대나무 틈으로 지시문에 금지된 사람의 얼굴(모자)과 몸통 일부가 보임.",
        "hard_violations": [],
        "physics": "양쪽 소총 모두 사람의 두 손과 팔에 의해 쥐어져 물리적으로 올바르게 지탱되고 있음."
       },
       {
        "label": "B",
        "direction": "양쪽 숲에서 나온 두 소총의 총구가 숲길 중앙을 겨냥하고 있음.",
        "built_space": "좁고 경사진 흙길을 양옆의 빽빽한 대나무 숲이 감싸고 있는 형태가 잘 구현됨.",
        "entities": "대나무숲, 좁은 길, 두 개의 소총. 오른쪽 총에는 손가락 일부가 보이지만, 왼쪽 총 주변에는 아무도 없음.",
        "hard_violations": [
         "왼쪽 총을 쥐고 있는 손이나 지지대가 전혀 없어 무기가 허공에 떠 있음"
        ],
        "physics": "오른쪽 총은 손으로 쥐고 있으나, 왼쪽 총은 잡고 있는 손이나 지지물 없이 대나무 잎 사이에 떠 있음."
       }
      ]
     },
     "ok": true
    },
    {
     "model": "gpt-high",
     "order": "reverse",
     "display_to_canonical": {
      "A": "B",
      "B": "A"
     },
     "raw": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 6,
        "verdict_ko": "양쪽 대나무에 사수를 감추고 총구를 비교적 작게 분리해 보여 주지만, 높은 사선 시점과 길 하단 중앙으로 내려가는 조준 방향은 부족하다."
       },
       {
        "label": "B",
        "score": 3,
        "verdict_ko": "총을 받치는 손은 분명하지만 오른쪽 얼굴 일부가 노출되고, 길게 강조된 총과 낮은 시점이 지정된 중경 배치에서 벗어난다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "왼쪽 총구는 오른쪽으로 조금 내려가고 오른쪽 총구는 왼쪽으로 조금 내려간다. 두 총열의 연장선은 화면 중앙의 길 위쪽 공간으로 모이며, 요구된 하단 중앙의 길을 향하는 각도는 약하다. 표적이나 시선을 확인할 얼굴은 보이지 않는다.",
        "built_space": "중앙의 좁은 흙길 하나와 양쪽의 빽빽한 대나무 군락, 길 가장자리의 돌이 보인다. 양쪽 틈에서 총구가 하나씩 돌출되어 서로 가리지 않는다. 인공 구조물이나 고정 설비는 없다. 카메라는 길을 따라 거의 수평으로 바라보며, 두 군락을 위에서 비스듬히 내려다보는 배치는 아니다.",
        "entities": "검은 금속 총열 두 개와 목재 총몸 일부, 대나무, 낙엽 깔린 흙길, 바위가 보인다. 얼굴이나 전신은 드러나지 않는다. 가장자리의 손과 소매 일부만으로 인물의 민족·나이·성별을 판별할 수 없다. 총은 시대극용 장총으로 읽히며 현대 장비나 읽을 수 있는 글자는 없다. 자연광과 절제된 색감은 아침 설정에 어긋나지 않는다.",
        "hard_violations": [],
        "physics": "두 총의 뒤쪽은 대나무 뒤로 이어지고, 왼쪽 가장자리와 오른쪽 총몸 부근에 손 일부가 보인다. 총을 든 사수가 식생에 가려진 상태로 읽히며 공중에 떠 있는 물체로 보이지 않는다. 늘어진 멜빵도 중력 방향에 맞는다. 정지 화면에서 동시 돌출의 속도나 대나무의 바람 움직임은 확인하기 어렵다."
       },
       {
        "label": "B",
        "direction": "왼쪽 총은 오른쪽 아래로, 오른쪽 총은 왼쪽 아래로 얕게 기울어 있다. 조준선은 중앙의 먼 길 위 공간으로 모이고, 하단 중앙의 가까운 길을 겨누는 위협선은 뚜렷하지 않다. 오른쪽 가장자리의 얼굴 일부에서는 시선 방향을 판독할 수 없다.",
        "built_space": "좁은 흙길 하나가 대나무 두 군락 사이로 이어지고 양쪽 가장자리에 돌이 놓여 있다. 총은 양쪽 틈에서 하나씩 나오지만 긴 총몸까지 크게 노출되어 중경의 작은 총구보다 전경의 장총이 강조된다. 카메라는 높은 사선 부감보다 길 높이에서 정면을 보는 시점에 가깝다. 인공 설비나 반사는 없다.",
        "entities": "목재와 검은 금속으로 된 장총 두 정, 총을 쥔 손과 소매, 대나무와 돌길이 보인다. 오른쪽 끝의 잎 뒤에는 얼굴 일부로 읽히는 피부색 윤곽이 드러나 있어 손·팔만 허용한 조건을 벗어난다. 인물의 세부 신원은 확인할 수 없다. 읽을 수 있는 글자나 현대적인 소품은 없고 자연광은 아침 장면과 양립한다.",
        "hard_violations": [
         "오른쪽 가장자리의 대나무 잎 뒤로 사수의 얼굴 일부가 노출되어, 얼굴은 보이지 않고 손·팔만 허용한다는 명시 조건을 위반한다."
        ],
        "physics": "왼쪽 총은 총몸 아래의 손과 뒤쪽 손으로 지지되고, 오른쪽 총도 아래에서 받치는 손이 분명하다. 총의 무게를 받는 접촉이 있어 부유 문제는 없다. 멜빵은 아래로 처져 있으며 손과 총의 관계도 물리적으로 가능하다. 다만 정지된 조준 자세여서 일제히 튀어나온 동작 자체는 강하게 드러나지 않는다."
       }
      ],
      "all_candidates_fail": true
     },
     "normalized": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 6,
        "verdict_ko": "양쪽 대나무에 사수를 감추고 총구를 비교적 작게 분리해 보여 주지만, 높은 사선 시점과 길 하단 중앙으로 내려가는 조준 방향은 부족하다."
       },
       {
        "label": "A",
        "score": 3,
        "verdict_ko": "총을 받치는 손은 분명하지만 오른쪽 얼굴 일부가 노출되고, 길게 강조된 총과 낮은 시점이 지정된 중경 배치에서 벗어난다."
       }
      ],
      "all_candidates_fail": true,
      "readings": [
       {
        "label": "B",
        "direction": "왼쪽 총구는 오른쪽으로 조금 내려가고 오른쪽 총구는 왼쪽으로 조금 내려간다. 두 총열의 연장선은 화면 중앙의 길 위쪽 공간으로 모이며, 요구된 하단 중앙의 길을 향하는 각도는 약하다. 표적이나 시선을 확인할 얼굴은 보이지 않는다.",
        "built_space": "중앙의 좁은 흙길 하나와 양쪽의 빽빽한 대나무 군락, 길 가장자리의 돌이 보인다. 양쪽 틈에서 총구가 하나씩 돌출되어 서로 가리지 않는다. 인공 구조물이나 고정 설비는 없다. 카메라는 길을 따라 거의 수평으로 바라보며, 두 군락을 위에서 비스듬히 내려다보는 배치는 아니다.",
        "entities": "검은 금속 총열 두 개와 목재 총몸 일부, 대나무, 낙엽 깔린 흙길, 바위가 보인다. 얼굴이나 전신은 드러나지 않는다. 가장자리의 손과 소매 일부만으로 인물의 민족·나이·성별을 판별할 수 없다. 총은 시대극용 장총으로 읽히며 현대 장비나 읽을 수 있는 글자는 없다. 자연광과 절제된 색감은 아침 설정에 어긋나지 않는다.",
        "hard_violations": [],
        "physics": "두 총의 뒤쪽은 대나무 뒤로 이어지고, 왼쪽 가장자리와 오른쪽 총몸 부근에 손 일부가 보인다. 총을 든 사수가 식생에 가려진 상태로 읽히며 공중에 떠 있는 물체로 보이지 않는다. 늘어진 멜빵도 중력 방향에 맞는다. 정지 화면에서 동시 돌출의 속도나 대나무의 바람 움직임은 확인하기 어렵다."
       },
       {
        "label": "A",
        "direction": "왼쪽 총은 오른쪽 아래로, 오른쪽 총은 왼쪽 아래로 얕게 기울어 있다. 조준선은 중앙의 먼 길 위 공간으로 모이고, 하단 중앙의 가까운 길을 겨누는 위협선은 뚜렷하지 않다. 오른쪽 가장자리의 얼굴 일부에서는 시선 방향을 판독할 수 없다.",
        "built_space": "좁은 흙길 하나가 대나무 두 군락 사이로 이어지고 양쪽 가장자리에 돌이 놓여 있다. 총은 양쪽 틈에서 하나씩 나오지만 긴 총몸까지 크게 노출되어 중경의 작은 총구보다 전경의 장총이 강조된다. 카메라는 높은 사선 부감보다 길 높이에서 정면을 보는 시점에 가깝다. 인공 설비나 반사는 없다.",
        "entities": "목재와 검은 금속으로 된 장총 두 정, 총을 쥔 손과 소매, 대나무와 돌길이 보인다. 오른쪽 끝의 잎 뒤에는 얼굴 일부로 읽히는 피부색 윤곽이 드러나 있어 손·팔만 허용한 조건을 벗어난다. 인물의 세부 신원은 확인할 수 없다. 읽을 수 있는 글자나 현대적인 소품은 없고 자연광은 아침 장면과 양립한다.",
        "hard_violations": [
         "오른쪽 가장자리의 대나무 잎 뒤로 사수의 얼굴 일부가 노출되어, 얼굴은 보이지 않고 손·팔만 허용한다는 명시 조건을 위반한다."
        ],
        "physics": "왼쪽 총은 총몸 아래의 손과 뒤쪽 손으로 지지되고, 오른쪽 총도 아래에서 받치는 손이 분명하다. 총의 무게를 받는 접촉이 있어 부유 문제는 없다. 멜빵은 아래로 처져 있으며 손과 총의 관계도 물리적으로 가능하다. 다만 정지된 조준 자세여서 일제히 튀어나온 동작 자체는 강하게 드러나지 않는다."
       }
      ]
     },
     "ok": true
    }
   ],
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "slot_winner_match": false,
   "slot_winner": {
    "gemini-pro": "A",
    "gpt-high": "B"
   },
   "route": "cross_slot_combined"
  },
  "dual": {
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "normalized": {
    "A": 1.5,
    "B": 1.375
   },
   "adjusted": {
    "A": 1.25,
    "B": 1.125
   },
   "violations": {
    "B": [
     "[gemini-pro] 왼쪽 총을 쥐고 있는 손이나 지지대가 전혀 없어 무기가 허공에 떠 있음"
    ],
    "A": [
     "[gpt-high] 오른쪽 가장자리의 대나무 잎 뒤로 사수의 얼굴 일부가 노출되어, 얼굴은 보이지 않고 손·팔만 허용한다는 명시 조건을 위반한다."
    ]
   },
   "per_model_winner": {
    "gemini-pro": "A",
    "gpt-high": "B"
   },
   "agreed": false
  },
  "totals": {
   "A": 1250,
   "B": 1125
  },
  "selected": "A",
  "ranking": [
   "A",
   "B"
  ],
  "verdicts": [
   {
    "label": "A",
    "score": 1250,
    "verdict_ko": "총을 쥐는 손을 묘사하여 물리적 오류를 피했으나, 화면에 등장하지 않아야 할 인물의 얼굴과 몸통이 왼쪽 틈새로 노출된 점이 감점 요인입니다.  ★위반: [gpt-high] 오른쪽 가장자리의 대나무 잎 뒤로 사수의 얼굴 일부가 노출되어, 얼굴은 보이지 않고 손·팔만 허용한다는 명시 조건을 위반한다."
   },
   {
    "label": "B",
    "score": 1125,
    "verdict_ko": "인물을 배제하려다 왼쪽 총을 쥐어야 할 손마저 지워버려, 총이 허공에 떠 있는 치명적인 물리적 오류를 범했습니다.  ★위반: [gemini-pro] 왼쪽 총을 쥐고 있는 손이나 지지대가 전혀 없어 무기가 허공에 떠 있음"
   }
  ],
  "refs": [],
  "critique_skipped": true,
  "shot_run_produced": false,
  "shot_run_uid": "06aa140f-1c06-7f47-bd79-99304e88362d",
  "ref_mode": "플레이트만 (배경 전용)",
  "share_plan": {
   "ref_plan": "background"
  }
 },
 "S13sh3::cine": {
  "applied": true,
  "attempted_at": "2026-09-09T02:08:04.358933+00:00",
  "fingerprint": "e3637f62339c15da019adee36441ffa23679f60a31ce0adb43394692ef397e38",
  "fingerprint_version": 2,
  "provider": "mai",
  "endpoint": "openrouter/chat-completions",
  "model": "microsoft/mai-image-2.6",
  "pack": "24.202608252115",
  "source_file": "S13sh3_sel.png",
  "source_sha256": "7550f53e60550873b7d43153d4fbe02d9304696fa953c92cabe06872ea33e1a7",
  "file": "S13sh3_cine.png",
  "staged_sha256": "6979b97d8951eec1aa27451030abb64147fe48db7efc7a491f6d324dcfdb22df",
  "latency_ms": 25699
 },
 "S13sh9::signage": {
  "fp": "9b23737534e6e559",
  "inscriptions": [],
  "cues": [],
  "dropped": []
 },
 "S13sh9": {
  "input_fingerprint": "928b3c601b126f35",
  "prompt": "Create ONE FINAL photorealistic live-action film still of the moment below — early 20th-century Korea, Jeolla region; all people are Korean unless explicitly stated otherwise. TIME OF DAY (lock): morning.\n\nSHOT TEXT (authoritative, Korean): 서로 마주 선 염재보와 복만이 억센 힘으로 상대의 어깨를 꽉 끌어안은 상체 클로즈업.\n\nLOCATION (lock): The narrow bamboo-forest path beside the rocks that conceal the ambush position. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: close-up\n- KEY BACKGROUND ELEMENTS: 대나무숲 (Borders the path where the two men embrace) — Visible in narrow, out-of-focus sections beyond their shoulders; used as Maintains location continuity without distracting from the hands and bodily contact; 복만의 낡은 두루마기 (Worn by 복만 and gripped at the shoulder during the embrace) — The shoulder and upper side face the oblique camera; used as Makes the force of 염재보's grasp legible through the compressed garment.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Keep the established natural morning contrast and muted palette, allowing warmth to come from physical contact rather than a lighting change.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The people visible in that photo are NOT in this shot: never carry their faces, bodies or clothing onto anyone here. Each person in THIS shot is defined solely by the PEOPLE list, their own CHARACTER REFERENCE images and their wardrobe notes. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): Dense bamboo and rocks enclose the narrow path. 염재보: He stands with his arms raised in an embrace, still without a gun. 복만: He wears a worn durumagi and rough beard, with a matchlock slung diagonally over his shoulder.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 염재보 (한국인 남성, 성인 나이대, 자연스러운 검은 머리); 복만 (한국인 남성, 성인 나이대, 거칠게 자란 수염, 자연스러운 검은 머리) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.",
  "roll_prompts": {
   "A": "Create ONE FINAL photorealistic live-action film still of the moment below — early 20th-century Korea, Jeolla region; all people are Korean unless explicitly stated otherwise. TIME OF DAY (lock): morning.\n\nSHOT TEXT (authoritative, Korean): 서로 마주 선 염재보와 복만이 억센 힘으로 상대의 어깨를 꽉 끌어안은 상체 클로즈업.\n\nLOCATION (lock): The narrow bamboo-forest path beside the rocks that conceal the ambush position. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: close-up\n- KEY BACKGROUND ELEMENTS: 대나무숲 (Borders the path where the two men embrace) — Visible in narrow, out-of-focus sections beyond their shoulders; used as Maintains location continuity without distracting from the hands and bodily contact; 복만의 낡은 두루마기 (Worn by 복만 and gripped at the shoulder during the embrace) — The shoulder and upper side face the oblique camera; used as Makes the force of 염재보's grasp legible through the compressed garment.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Keep the established natural morning contrast and muted palette, allowing warmth to come from physical contact rather than a lighting change.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The people visible in that photo are NOT in this shot: never carry their faces, bodies or clothing onto anyone here. Each person in THIS shot is defined solely by the PEOPLE list, their own CHARACTER REFERENCE images and their wardrobe notes. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): Dense bamboo and rocks enclose the narrow path. 염재보: He stands with his arms raised in an embrace, still without a gun. 복만: He wears a worn durumagi and rough beard, with a matchlock slung diagonally over his shoulder.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 염재보 (한국인 남성, 성인 나이대, 자연스러운 검은 머리); 복만 (한국인 남성, 성인 나이대, 거칠게 자란 수염, 자연스러운 검은 머리) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.",
   "B": "Create ONE FINAL photorealistic live-action film still of the moment below — early 20th-century Korea, Jeolla region; all people are Korean unless explicitly stated otherwise. TIME OF DAY (lock): morning.\n\nSHOT TEXT (authoritative, Korean): 서로 마주 선 염재보와 복만이 억센 힘으로 상대의 어깨를 꽉 끌어안은 상체 클로즈업.\n\nLOCATION (lock): The narrow bamboo-forest path beside the rocks that conceal the ambush position. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: close-up\n- KEY BACKGROUND ELEMENTS: 대나무숲 (Borders the path where the two men embrace) — Visible in narrow, out-of-focus sections beyond their shoulders; used as Maintains location continuity without distracting from the hands and bodily contact; 복만의 낡은 두루마기 (Worn by 복만 and gripped at the shoulder during the embrace) — The shoulder and upper side face the oblique camera; used as Makes the force of 염재보's grasp legible through the compressed garment.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Keep the established natural morning contrast and muted palette, allowing warmth to come from physical contact rather than a lighting change.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The people visible in that photo are NOT in this shot: never carry their faces, bodies or clothing onto anyone here. Each person in THIS shot is defined solely by the PEOPLE list, their own CHARACTER REFERENCE images and their wardrobe notes. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): Dense bamboo and rocks enclose the narrow path. 염재보: He stands with his arms raised in an embrace, still without a gun. 복만: He wears a worn durumagi and rough beard, with a matchlock slung diagonally over his shoulder.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 염재보 (한국인 남성, 성인 나이대, 자연스러운 검은 머리); 복만 (한국인 남성, 성인 나이대, 거칠게 자란 수염, 자연스러운 검은 머리) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.\n\nCOMPOSITION VARIATION (this candidate only): keep the place, the people and the moment exactly as specified, and keep from the CAMERA & FRAME contract its subject scale, screen placement and key background elements exactly. For this candidate only, its camera angle and height are deliberately left open: choose a camera position distinctly different from the obvious one, as a film director picking a second setup on the same blocking. Stay true to any explicit framing the SHOT TEXT itself dictates."
  },
  "shot_run_spend_attempt_count": 1,
  "readings": [
   {
    "label": "A",
    "direction": "카메라는 좁은 대나무 숲길에서 서로 껴안고 있는 두 남자를 향하고 있으며, 복만은 비스듬히 카메라 쪽을 향하고 있음.",
    "built_space": "레퍼런스와 일치하는 바위와 대나무가 있는 좁은 숲길 배경이 심도 얕게 적절히 구현됨.",
    "entities": "염재보(파란색 누비옷, 등을 보임)와 복만(거친 수염, 낡은 회색 두루마기, 얼굴 측면이 보임)의 외모와 의상이 잘 일치함. 복만의 화승총은 샷에 나타나지 않음.",
    "hard_violations": [],
    "physics": "두 사람이 서로를 억센 힘으로 끌어안고 지탱하는 자세와 각 손의 위치, 옷깃이 눌리는 물리적 압박감이 자연스럽게 묘사됨."
   },
   {
    "label": "B",
    "direction": "카메라는 대나무 숲길에서 포옹하는 두 남자를 향하며, 염재보가 카메라 정면 쪽을 향하고 있음.",
    "built_space": "바위와 대나무가 있는 좁은 산길 배경이 잘 묘사됨.",
    "entities": "염재보(수염 없음)와 복만(등을 보임, 화승총을 멤)이 등장하지만, 두 인물의 의상이 서로 뒤바뀌고 섞이는 심각한 오류가 발생함.",
    "hard_violations": [
     "[gemini-pro] physically impossible anatomy or staging (등을 보인 인물의 몸통은 회색 옷이나 카메라를 향해 뻗은 팔은 파란색 누비옷을 입고 있으며, 정면을 향한 인물은 상대방의 회색 옷을 입고 있는 등 불가능한 의상 구조)"
    ],
    "physics": "두 사람의 포옹 자세 자체는 지탱되고 있으나, 팔과 몸통의 의상이 물리적으로 연결될 수 없는 형태로 결합됨."
   }
  ],
  "cross_model_order": {
   "policy": "cross_model_order_v2_slot0_fwd_slot1_rev_parallel2",
   "slots": [
    {
     "model": "gemini-pro",
     "order": "forward",
     "display_to_canonical": {
      "A": "A",
      "B": "B"
     },
     "raw": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 8,
        "verdict_ko": "프롬프트에 명시된 두 인물의 의상과 외모를 정확히 반영하여 자연스럽게 포옹하는 상체 클로즈업을 훌륭하게 구현했으나, 복만의 화승총이 누락되었습니다."
       },
       {
        "label": "B",
        "score": 2,
        "verdict_ko": "염재보와 복만의 의상이 서로 뒤바뀌고 한 인물의 몸에 두 가지 다른 옷감이 섞여 나오는 등 물리적으로 불가능한 의상 오류가 발생했습니다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "카메라는 좁은 대나무 숲길에서 서로 껴안고 있는 두 남자를 향하고 있으며, 복만은 비스듬히 카메라 쪽을 향하고 있음.",
        "built_space": "레퍼런스와 일치하는 바위와 대나무가 있는 좁은 숲길 배경이 심도 얕게 적절히 구현됨.",
        "entities": "염재보(파란색 누비옷, 등을 보임)와 복만(거친 수염, 낡은 회색 두루마기, 얼굴 측면이 보임)의 외모와 의상이 잘 일치함. 복만의 화승총은 샷에 나타나지 않음.",
        "hard_violations": [],
        "physics": "두 사람이 서로를 억센 힘으로 끌어안고 지탱하는 자세와 각 손의 위치, 옷깃이 눌리는 물리적 압박감이 자연스럽게 묘사됨."
       },
       {
        "label": "B",
        "direction": "카메라는 대나무 숲길에서 포옹하는 두 남자를 향하며, 염재보가 카메라 정면 쪽을 향하고 있음.",
        "built_space": "바위와 대나무가 있는 좁은 산길 배경이 잘 묘사됨.",
        "entities": "염재보(수염 없음)와 복만(등을 보임, 화승총을 멤)이 등장하지만, 두 인물의 의상이 서로 뒤바뀌고 섞이는 심각한 오류가 발생함.",
        "hard_violations": [
         "physically impossible anatomy or staging (등을 보인 인물의 몸통은 회색 옷이나 카메라를 향해 뻗은 팔은 파란색 누비옷을 입고 있으며, 정면을 향한 인물은 상대방의 회색 옷을 입고 있는 등 불가능한 의상 구조)"
        ],
        "physics": "두 사람의 포옹 자세 자체는 지탱되고 있으나, 팔과 몸통의 의상이 물리적으로 연결될 수 없는 형태로 결합됨."
       }
      ],
      "all_candidates_fail": false
     },
     "normalized": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 8,
        "verdict_ko": "프롬프트에 명시된 두 인물의 의상과 외모를 정확히 반영하여 자연스럽게 포옹하는 상체 클로즈업을 훌륭하게 구현했으나, 복만의 화승총이 누락되었습니다."
       },
       {
        "label": "B",
        "score": 2,
        "verdict_ko": "염재보와 복만의 의상이 서로 뒤바뀌고 한 인물의 몸에 두 가지 다른 옷감이 섞여 나오는 등 물리적으로 불가능한 의상 오류가 발생했습니다."
       }
      ],
      "all_candidates_fail": false,
      "readings": [
       {
        "label": "A",
        "direction": "카메라는 좁은 대나무 숲길에서 서로 껴안고 있는 두 남자를 향하고 있으며, 복만은 비스듬히 카메라 쪽을 향하고 있음.",
        "built_space": "레퍼런스와 일치하는 바위와 대나무가 있는 좁은 숲길 배경이 심도 얕게 적절히 구현됨.",
        "entities": "염재보(파란색 누비옷, 등을 보임)와 복만(거친 수염, 낡은 회색 두루마기, 얼굴 측면이 보임)의 외모와 의상이 잘 일치함. 복만의 화승총은 샷에 나타나지 않음.",
        "hard_violations": [],
        "physics": "두 사람이 서로를 억센 힘으로 끌어안고 지탱하는 자세와 각 손의 위치, 옷깃이 눌리는 물리적 압박감이 자연스럽게 묘사됨."
       },
       {
        "label": "B",
        "direction": "카메라는 대나무 숲길에서 포옹하는 두 남자를 향하며, 염재보가 카메라 정면 쪽을 향하고 있음.",
        "built_space": "바위와 대나무가 있는 좁은 산길 배경이 잘 묘사됨.",
        "entities": "염재보(수염 없음)와 복만(등을 보임, 화승총을 멤)이 등장하지만, 두 인물의 의상이 서로 뒤바뀌고 섞이는 심각한 오류가 발생함.",
        "hard_violations": [
         "physically impossible anatomy or staging (등을 보인 인물의 몸통은 회색 옷이나 카메라를 향해 뻗은 팔은 파란색 누비옷을 입고 있으며, 정면을 향한 인물은 상대방의 회색 옷을 입고 있는 등 불가능한 의상 구조)"
        ],
        "physics": "두 사람의 포옹 자세 자체는 지탱되고 있으나, 팔과 몸통의 의상이 물리적으로 연결될 수 없는 형태로 결합됨."
       }
      ]
     },
     "ok": true
    },
    {
     "model": "gpt-high",
     "order": "reverse",
     "display_to_canonical": {
      "A": "B",
      "B": "A"
     },
     "raw": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 7,
        "verdict_ko": "힘껏 끌어안는 접촉과 장소는 맞지만, 배경과 몸통을 더 넓게 보여 주어 어깨와 손 중심의 상체 클로즈업 요구에서는 B보다 약하다."
       },
       {
        "label": "B",
        "score": 9,
        "verdict_ko": "상체를 밀착해 자른 구도에서 서로의 어깨를 감싼 팔과 옷을 압박하는 손이 크게 보여, 지정된 클로즈업과 강한 포옹 동작을 더 충실히 구현한다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "오른쪽 염재보는 얼굴을 왼쪽 복만의 어깨에 묻고 눈을 아래로 내린다. 복만은 카메라에 등을 보인 채 염재보 쪽으로 머리와 몸을 돌려 밀착한다. 염재보의 손은 복만의 어깨 뒤와 등 옆을 누르고, 복만의 팔은 염재보의 어깨를 감싼다. 복만의 등 왼쪽 총은 개머리판이 위로 향하고 총구 쪽은 아래 화면 밖으로 이어지며, 사람을 겨누지 않는다.",
        "built_space": "인공 구조물이나 고정 설비는 없다. 두 사람 뒤로 좁은 흙길, 양옆의 빽빽한 대나무, 왼쪽 큰 바위와 오른쪽 여러 바위가 보인다. 참조의 숲길 재질과 자연광은 이어지지만, 배경이 어깨 너머의 좁고 흐린 부분에 한정되지 않고 상당히 넓게 드러난다.",
        "entities": "성인 동아시아계 남성 두 명만 보이며 한국인 설정과 외관상 충돌하지 않는다. 염재보의 검은 머리, 수염 없는 얼굴과 짙은 누빔옷은 대체로 참조에 부합하지만 참조의 삿갓은 없고 짧은 머리가 드러난다. 복만은 낡고 해진 갈색 두루마기를 입었으며, 얼굴과 수염은 뒤돌아 있어 확인할 수 없다. 참조의 모자 대신 상투와 머리띠가 보인다. 복만 쪽에는 목제 장총과 멜빵이 있으나 화승식 기구의 세부는 확정하기 어렵다. 염재보에게 총은 없으며 읽을 수 있는 글자나 추가 인물도 없다.",
        "hard_violations": [],
        "physics": "두 사람의 팔이 상대 몸을 감싸고 손바닥과 손가락이 옷에 닿아 있어 포옹의 접촉과 압박이 성립한다. 총은 복만의 등에 붙어 있고 멜빵이 보여 휴대 상태를 설명한다. 하체와 발은 화면 밖이므로 지면 접촉을 직접 확인할 수 없지만, 상체는 직립 포옹으로 읽히며 부유하거나 지지 없이 떨어진 물체는 없다."
       },
       {
        "label": "B",
        "direction": "왼쪽 염재보는 등을 카메라에 보이며 오른쪽 복만에게 얼굴과 가슴을 밀착한다. 복만의 얼굴은 염재보의 머리 옆을 향하지만 눈이 위쪽에서 잘려 정확한 시선 대상은 확인되지 않는다. 염재보의 보이는 손은 복만의 어깨 윗부분을 잡고, 복만의 두 손은 염재보의 어깨 뒤와 등 옆을 눌러 감싼다. 보이는 총구나 별도의 지향성 소품은 없다.",
        "built_space": "인공 구조물이나 고정 설비는 없다. 왼쪽 뒤로 좁은 흙길과 흐린 대나무가 이어지고, 오른쪽에는 대나무 줄기와 바위 무리가 보인다. 두 사람의 상체가 화면 대부분을 차지하며 머리 윗부분까지 잘려 손과 어깨 접촉이 중심이 된다. 참조의 바위 곁 대숲길과 절제된 아침 자연광에 부합한다.",
        "entities": "성인 동아시아계 남성 두 명만 있으며 한국인 설정과 외관상 충돌하지 않는다. 염재보는 검은 머리와 짙은 누빔옷이 보이지만 얼굴은 가려져 정확한 동일인 여부를 확인하기 어렵다. 복만의 검은 거친 수염과 낡은 회갈색 두루마기는 설정에 맞으며, 보이는 얼굴은 참조보다 다소 거칠고 나이 들어 보인다. 머리 윗부분이 잘려 참조 모자의 유지 여부는 판정할 수 없다. 복만의 총과 멜빵은 보이지 않지만 등 뒤가 가려진 구도이므로 소지 상태를 단정할 수 없다. 추가 인물이나 읽을 수 있는 글자는 없다.",
        "hard_violations": [],
        "physics": "복만의 양손은 염재보의 등에 실제로 닿고, 염재보의 팔은 복만의 어깨를 돌아 손으로 붙잡는다. 손가락 아래 옷감의 눌림과 팔 사이의 밀착이 강한 포옹을 뒷받침한다. 보이는 세 손은 두 인물의 팔과 자연스럽게 연결되며 추가 신체로 보이지 않는다. 발은 클로즈업 밖이지만 상체는 서로 기대어 선 자세이고, 공중에 떠 있거나 지지 없는 물체는 없다."
       }
      ],
      "all_candidates_fail": false
     },
     "normalized": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 7,
        "verdict_ko": "힘껏 끌어안는 접촉과 장소는 맞지만, 배경과 몸통을 더 넓게 보여 주어 어깨와 손 중심의 상체 클로즈업 요구에서는 B보다 약하다."
       },
       {
        "label": "A",
        "score": 9,
        "verdict_ko": "상체를 밀착해 자른 구도에서 서로의 어깨를 감싼 팔과 옷을 압박하는 손이 크게 보여, 지정된 클로즈업과 강한 포옹 동작을 더 충실히 구현한다."
       }
      ],
      "all_candidates_fail": false,
      "readings": [
       {
        "label": "B",
        "direction": "오른쪽 염재보는 얼굴을 왼쪽 복만의 어깨에 묻고 눈을 아래로 내린다. 복만은 카메라에 등을 보인 채 염재보 쪽으로 머리와 몸을 돌려 밀착한다. 염재보의 손은 복만의 어깨 뒤와 등 옆을 누르고, 복만의 팔은 염재보의 어깨를 감싼다. 복만의 등 왼쪽 총은 개머리판이 위로 향하고 총구 쪽은 아래 화면 밖으로 이어지며, 사람을 겨누지 않는다.",
        "built_space": "인공 구조물이나 고정 설비는 없다. 두 사람 뒤로 좁은 흙길, 양옆의 빽빽한 대나무, 왼쪽 큰 바위와 오른쪽 여러 바위가 보인다. 참조의 숲길 재질과 자연광은 이어지지만, 배경이 어깨 너머의 좁고 흐린 부분에 한정되지 않고 상당히 넓게 드러난다.",
        "entities": "성인 동아시아계 남성 두 명만 보이며 한국인 설정과 외관상 충돌하지 않는다. 염재보의 검은 머리, 수염 없는 얼굴과 짙은 누빔옷은 대체로 참조에 부합하지만 참조의 삿갓은 없고 짧은 머리가 드러난다. 복만은 낡고 해진 갈색 두루마기를 입었으며, 얼굴과 수염은 뒤돌아 있어 확인할 수 없다. 참조의 모자 대신 상투와 머리띠가 보인다. 복만 쪽에는 목제 장총과 멜빵이 있으나 화승식 기구의 세부는 확정하기 어렵다. 염재보에게 총은 없으며 읽을 수 있는 글자나 추가 인물도 없다.",
        "hard_violations": [],
        "physics": "두 사람의 팔이 상대 몸을 감싸고 손바닥과 손가락이 옷에 닿아 있어 포옹의 접촉과 압박이 성립한다. 총은 복만의 등에 붙어 있고 멜빵이 보여 휴대 상태를 설명한다. 하체와 발은 화면 밖이므로 지면 접촉을 직접 확인할 수 없지만, 상체는 직립 포옹으로 읽히며 부유하거나 지지 없이 떨어진 물체는 없다."
       },
       {
        "label": "A",
        "direction": "왼쪽 염재보는 등을 카메라에 보이며 오른쪽 복만에게 얼굴과 가슴을 밀착한다. 복만의 얼굴은 염재보의 머리 옆을 향하지만 눈이 위쪽에서 잘려 정확한 시선 대상은 확인되지 않는다. 염재보의 보이는 손은 복만의 어깨 윗부분을 잡고, 복만의 두 손은 염재보의 어깨 뒤와 등 옆을 눌러 감싼다. 보이는 총구나 별도의 지향성 소품은 없다.",
        "built_space": "인공 구조물이나 고정 설비는 없다. 왼쪽 뒤로 좁은 흙길과 흐린 대나무가 이어지고, 오른쪽에는 대나무 줄기와 바위 무리가 보인다. 두 사람의 상체가 화면 대부분을 차지하며 머리 윗부분까지 잘려 손과 어깨 접촉이 중심이 된다. 참조의 바위 곁 대숲길과 절제된 아침 자연광에 부합한다.",
        "entities": "성인 동아시아계 남성 두 명만 있으며 한국인 설정과 외관상 충돌하지 않는다. 염재보는 검은 머리와 짙은 누빔옷이 보이지만 얼굴은 가려져 정확한 동일인 여부를 확인하기 어렵다. 복만의 검은 거친 수염과 낡은 회갈색 두루마기는 설정에 맞으며, 보이는 얼굴은 참조보다 다소 거칠고 나이 들어 보인다. 머리 윗부분이 잘려 참조 모자의 유지 여부는 판정할 수 없다. 복만의 총과 멜빵은 보이지 않지만 등 뒤가 가려진 구도이므로 소지 상태를 단정할 수 없다. 추가 인물이나 읽을 수 있는 글자는 없다.",
        "hard_violations": [],
        "physics": "복만의 양손은 염재보의 등에 실제로 닿고, 염재보의 팔은 복만의 어깨를 돌아 손으로 붙잡는다. 손가락 아래 옷감의 눌림과 팔 사이의 밀착이 강한 포옹을 뒷받침한다. 보이는 세 손은 두 인물의 팔과 자연스럽게 연결되며 추가 신체로 보이지 않는다. 발은 클로즈업 밖이지만 상체는 서로 기대어 선 자세이고, 공중에 떠 있거나 지지 없는 물체는 없다."
       }
      ]
     },
     "ok": true
    }
   ],
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "slot_winner_match": true,
   "slot_winner": {
    "gemini-pro": "A",
    "gpt-high": "A"
   },
   "route": "cross_slot_agree"
  },
  "dual": {
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "normalized": {
    "A": 2.0,
    "B": 1.028
   },
   "adjusted": {
    "A": 2.0,
    "B": 0.778
   },
   "violations": {
    "B": [
     "[gemini-pro] physically impossible anatomy or staging (등을 보인 인물의 몸통은 회색 옷이나 카메라를 향해 뻗은 팔은 파란색 누비옷을 입고 있으며, 정면을 향한 인물은 상대방의 회색 옷을 입고 있는 등 불가능한 의상 구조)"
    ]
   },
   "per_model_winner": {
    "gemini-pro": "A",
    "gpt-high": "A"
   },
   "agreed": true
  },
  "totals": {
   "A": 2000,
   "B": 778
  },
  "selected": "A",
  "ranking": [
   "A",
   "B"
  ],
  "verdicts": [
   {
    "label": "A",
    "score": 2000,
    "verdict_ko": "프롬프트에 명시된 두 인물의 의상과 외모를 정확히 반영하여 자연스럽게 포옹하는 상체 클로즈업을 훌륭하게 구현했으나, 복만의 화승총이 누락되었습니다."
   },
   {
    "label": "B",
    "score": 778,
    "verdict_ko": "염재보와 복만의 의상이 서로 뒤바뀌고 한 인물의 몸에 두 가지 다른 옷감이 섞여 나오는 등 물리적으로 불가능한 의상 오류가 발생했습니다.  ★위반: [gemini-pro] physically impossible anatomy or staging (등을 보인 인물의 몸통은 회색 옷이나 카메라를 향해 뻗은 팔은 파란색 누비옷을 입고 있으며, 정면을 향한 인물은 상대방의 회색 옷을 입고 있는 등 불가능한 의상 구조)"
   }
  ],
  "refs": [
   {
    "label": "PREVIOUS SHOT STILL — a visually related earlier shot of this same place: the location's look, materials, fixed features and lighting mood are LOCKED to this photo; never copy its camera framing. The people visible in that photo are NOT in this shot: never carry their faces, bodies or clothing onto anyone here. Each person in THIS shot is defined solely by the PEOPLE list, their own CHARACTER REFERENCE images and their wardrobe notes. If this photo shows a character who cannot move (dead or unconscious), that character's exact pose, position and orientation are ALSO LOCKED — treat their body as a fixed prop of the set that only the camera moves around.",
    "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/db00517b-165b-4005-a287-f28936b59b95/images/272bd7af-9294-4490-98c6-88a40e9deacc/scene/recipe/S13sh3_sel.png",
    "asset_id": "51105773-a133-4cde-8acd-7871393153f3",
    "role": "prev_still"
   },
   {
    "label": "CHARACTER REFERENCE — 염재보: the exact person appearing in this shot; match face, hair and build exactly.",
    "path": "<bytes:964351>",
    "asset_id": "ab287289-d0b7-436d-8f70-68452f898f73",
    "role": "character_ref"
   },
   {
    "label": "CHARACTER REFERENCE — 복만: the exact person appearing in this shot; match face, hair and build exactly.",
    "path": "<bytes:1027479>",
    "asset_id": "b7e69570-d15d-4241-b5ea-4cb8d4ece1cf",
    "role": "character_ref"
   }
  ],
  "critique_skipped": true,
  "shot_run_produced": false,
  "shot_run_uid": "06aa140f-1d95-7d70-b542-110f49f88fc8",
  "ref_mode": "prev+엔티티",
  "share_plan": {
   "ref_plan": "prev",
   "prev_anchor_tag": "S13sh3"
  }
 },
 "S13sh9::cine": {
  "applied": true,
  "attempted_at": "2026-09-09T02:09:52.542901+00:00",
  "fingerprint": "3dfd5f9c00d962ca5a0c3ebd512cb7e9784d0fd271e7cf76a51ce060a6711e49",
  "fingerprint_version": 2,
  "provider": "mai",
  "endpoint": "openrouter/chat-completions",
  "model": "microsoft/mai-image-2.6",
  "pack": "24.202608252115",
  "source_file": "S13sh9_sel.png",
  "source_sha256": "e439f655a57b4278553c0fc41684105f6a564a9f60e5fcfbd60b2b440365ead8",
  "file": "S13sh9_cine.png",
  "staged_sha256": "b3c75da4967099dc2e11d2e2fd509b44eccd96df18377de9aef39236cb4b575a",
  "latency_ms": 27740
 },
 "S13sh10::signage": {
  "fp": "bd18e4c5877427a7",
  "inscriptions": [],
  "cues": [],
  "dropped": []
 },
 "S13sh10": {
  "input_fingerprint": "43076f633d0fe4b2",
  "prompt": "Create ONE FINAL photorealistic live-action film still of the moment below — early 20th-century Korea, Jeolla region; all people are Korean unless explicitly stated otherwise. TIME OF DAY (lock): morning.\n\nSHOT TEXT (authoritative, Korean): 어깨를 안은 염재보의 등 너머로 서 있는 안규홍을 매섭게 쏘아보는 복만의 굳은 눈빛 클로즈업.\n\nLOCATION (lock): The outdoor meeting point on the bamboo trail beside the rocky ambush cover. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: close-up\n- KEY BACKGROUND ELEMENTS: 대나무숲 (Remains beyond the embracing men) — A limited, soft section is visible beside 복만's head; used as Retains environmental continuity without adding a misleading background figure.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Unchanged natural morning light keeps 복만's tightening eyes readable, so the return of suspicion is carried by expression rather than an artificial contrast shift.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The clothing, hair and overall look of 복만, 염재보 — who appear both in that photo and in this shot — are LOCKED to that photo. Anyone else visible in that photo is NOT in this shot: never carry their face, body or clothing onto anyone here. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The rocky bamboo path remains enclosed by dense vegetation. 복만: He retains his worn durumagi, rough beard and diagonally slung matchlock; his expression has hardened. 염재보: He stands near the rocks after the greeting, without a gun.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 염재보 (한국인 남성, 성인 나이대, 자연스러운 검은 머리); 복만 (한국인 남성, 성인 나이대, 거칠게 자란 수염, 자연스러운 검은 머리) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.",
  "roll_prompts": {
   "A": "Create ONE FINAL photorealistic live-action film still of the moment below — early 20th-century Korea, Jeolla region; all people are Korean unless explicitly stated otherwise. TIME OF DAY (lock): morning.\n\nSHOT TEXT (authoritative, Korean): 어깨를 안은 염재보의 등 너머로 서 있는 안규홍을 매섭게 쏘아보는 복만의 굳은 눈빛 클로즈업.\n\nLOCATION (lock): The outdoor meeting point on the bamboo trail beside the rocky ambush cover. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: close-up\n- KEY BACKGROUND ELEMENTS: 대나무숲 (Remains beyond the embracing men) — A limited, soft section is visible beside 복만's head; used as Retains environmental continuity without adding a misleading background figure.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Unchanged natural morning light keeps 복만's tightening eyes readable, so the return of suspicion is carried by expression rather than an artificial contrast shift.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The clothing, hair and overall look of 복만, 염재보 — who appear both in that photo and in this shot — are LOCKED to that photo. Anyone else visible in that photo is NOT in this shot: never carry their face, body or clothing onto anyone here. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The rocky bamboo path remains enclosed by dense vegetation. 복만: He retains his worn durumagi, rough beard and diagonally slung matchlock; his expression has hardened. 염재보: He stands near the rocks after the greeting, without a gun.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 염재보 (한국인 남성, 성인 나이대, 자연스러운 검은 머리); 복만 (한국인 남성, 성인 나이대, 거칠게 자란 수염, 자연스러운 검은 머리) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.",
   "B": "Create ONE FINAL photorealistic live-action film still of the moment below — early 20th-century Korea, Jeolla region; all people are Korean unless explicitly stated otherwise. TIME OF DAY (lock): morning.\n\nSHOT TEXT (authoritative, Korean): 어깨를 안은 염재보의 등 너머로 서 있는 안규홍을 매섭게 쏘아보는 복만의 굳은 눈빛 클로즈업.\n\nLOCATION (lock): The outdoor meeting point on the bamboo trail beside the rocky ambush cover. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: close-up\n- KEY BACKGROUND ELEMENTS: 대나무숲 (Remains beyond the embracing men) — A limited, soft section is visible beside 복만's head; used as Retains environmental continuity without adding a misleading background figure.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Unchanged natural morning light keeps 복만's tightening eyes readable, so the return of suspicion is carried by expression rather than an artificial contrast shift.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The clothing, hair and overall look of 복만, 염재보 — who appear both in that photo and in this shot — are LOCKED to that photo. Anyone else visible in that photo is NOT in this shot: never carry their face, body or clothing onto anyone here. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The rocky bamboo path remains enclosed by dense vegetation. 복만: He retains his worn durumagi, rough beard and diagonally slung matchlock; his expression has hardened. 염재보: He stands near the rocks after the greeting, without a gun.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 염재보 (한국인 남성, 성인 나이대, 자연스러운 검은 머리); 복만 (한국인 남성, 성인 나이대, 거칠게 자란 수염, 자연스러운 검은 머리) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.\n\nCOMPOSITION VARIATION (this candidate only): keep the place, the people and the moment exactly as specified, and keep from the CAMERA & FRAME contract its subject scale, screen placement and key background elements exactly. For this candidate only, its camera angle and height are deliberately left open: choose a camera position distinctly different from the obvious one, as a film director picking a second setup on the same blocking. Stay true to any explicit framing the SHOT TEXT itself dictates."
  },
  "shot_run_spend_attempt_count": 1,
  "readings": [
   {
    "label": "A",
    "direction": "복만의 시선이 화면 우측을 향하고 있으나, 뚜렷한 대상을 쏘아보는 느낌이 부족하고 다소 모호하게 처리됨.",
    "built_space": "야외 대나무숲을 배경으로 두 사람이 서 있는 공간이 적절히 구성됨.",
    "entities": "복만(수염, 두루마기)과 염재보(누비옷)가 지시대로 등장하며, 프롬프트의 지시대로 추가 배경 인물은 없음.",
    "hard_violations": [
     "[gemini-pro] physically impossible anatomy (염재보의 왼쪽 어깨에 얹힌 손의 손가락이 6개이며 좌우 구조가 뒤바뀐 기형적인 형태임)",
     "[gemini-pro] unsupported object (복만의 등 뒤에 있는 화승총을 지탱하는 어깨끈이나 손이 전혀 보이지 않아 무기가 허공에 떠 있음)"
    ],
    "physics": "화승총이 어떠한 지지대(끈 등)도 없이 허공에 떠 있으며, 포옹 상태에서 어깨에 얹혀진 손의 뼈대와 방향이 인체공학적으로 완전히 어긋남."
   },
   {
    "label": "B",
    "direction": "복만이 화면 우측 배경에 서 있는 인물(안규홍)을 향해 굳고 매서운 눈빛으로 정확히 시선을 고정하고 있음.",
    "built_space": "대나무숲 배경 속에서 포옹하는 두 사람과 뒤편의 인물이 공간적으로 무리 없이 배치됨.",
    "entities": "복만과 염재보의 외형(수염, 누비옷 등)이 잘 묘사되었으나, 프롬프트가 엄격히 추가를 금지한 제3의 인물(안규홍)이 배경에 등장함.",
    "hard_violations": [
     "[gemini-pro] a person placed where the staging does not put them (배경 인물 추가를 명시적으로 금지한 프롬프트의 'PEOPLE' 및 'KEY BACKGROUND ELEMENTS' 제약을 어기고 제3의 인물이 배치됨)",
     "[gpt-high] 허용 인물이 복만과 염재보로 제한되고 오해를 낳는 배경 인물도 명시적으로 금지되었는데, 오른쪽 배경에 세 번째 남성의 얼굴과 상체를 추가했다."
    ],
    "physics": "복만의 가슴을 가로지르는 가죽 끈이 화승총을 안정적으로 지탱하고 있으며, 염재보가 복만의 어깨를 짚고 있는 손의 형태와 무게감도 자연스러움."
   }
  ],
  "cross_model_order": {
   "policy": "cross_model_order_v2_slot0_fwd_slot1_rev_parallel2",
   "slots": [
    {
     "model": "gemini-pro",
     "order": "forward",
     "display_to_canonical": {
      "A": "A",
      "B": "B"
     },
     "raw": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 4,
        "verdict_ko": "총기를 지탱하는 물리적 구조와 매서운 표정 연기는 훌륭하게 구현되었으나, 프롬프트에서 명시적으로 추가를 금지한 제3의 인물이 배경에 등장하여 치명적인 제약 위반이 발생했습니다."
       },
       {
        "label": "A",
        "score": 2,
        "verdict_ko": "해부학적으로 불가능하게 일그러진 손의 형태와 아무것도 지탱하지 않은 채 공중에 떠 있는 화승총 등 심각한 물리적 오류가 있어 사용할 수 없습니다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "복만의 시선이 화면 우측을 향하고 있으나, 뚜렷한 대상을 쏘아보는 느낌이 부족하고 다소 모호하게 처리됨.",
        "built_space": "야외 대나무숲을 배경으로 두 사람이 서 있는 공간이 적절히 구성됨.",
        "entities": "복만(수염, 두루마기)과 염재보(누비옷)가 지시대로 등장하며, 프롬프트의 지시대로 추가 배경 인물은 없음.",
        "hard_violations": [
         "physically impossible anatomy (염재보의 왼쪽 어깨에 얹힌 손의 손가락이 6개이며 좌우 구조가 뒤바뀐 기형적인 형태임)",
         "unsupported object (복만의 등 뒤에 있는 화승총을 지탱하는 어깨끈이나 손이 전혀 보이지 않아 무기가 허공에 떠 있음)"
        ],
        "physics": "화승총이 어떠한 지지대(끈 등)도 없이 허공에 떠 있으며, 포옹 상태에서 어깨에 얹혀진 손의 뼈대와 방향이 인체공학적으로 완전히 어긋남."
       },
       {
        "label": "B",
        "direction": "복만이 화면 우측 배경에 서 있는 인물(안규홍)을 향해 굳고 매서운 눈빛으로 정확히 시선을 고정하고 있음.",
        "built_space": "대나무숲 배경 속에서 포옹하는 두 사람과 뒤편의 인물이 공간적으로 무리 없이 배치됨.",
        "entities": "복만과 염재보의 외형(수염, 누비옷 등)이 잘 묘사되었으나, 프롬프트가 엄격히 추가를 금지한 제3의 인물(안규홍)이 배경에 등장함.",
        "hard_violations": [
         "a person placed where the staging does not put them (배경 인물 추가를 명시적으로 금지한 프롬프트의 'PEOPLE' 및 'KEY BACKGROUND ELEMENTS' 제약을 어기고 제3의 인물이 배치됨)"
        ],
        "physics": "복만의 가슴을 가로지르는 가죽 끈이 화승총을 안정적으로 지탱하고 있으며, 염재보가 복만의 어깨를 짚고 있는 손의 형태와 무게감도 자연스러움."
       }
      ],
      "all_candidates_fail": true
     },
     "normalized": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 4,
        "verdict_ko": "총기를 지탱하는 물리적 구조와 매서운 표정 연기는 훌륭하게 구현되었으나, 프롬프트에서 명시적으로 추가를 금지한 제3의 인물이 배경에 등장하여 치명적인 제약 위반이 발생했습니다."
       },
       {
        "label": "A",
        "score": 2,
        "verdict_ko": "해부학적으로 불가능하게 일그러진 손의 형태와 아무것도 지탱하지 않은 채 공중에 떠 있는 화승총 등 심각한 물리적 오류가 있어 사용할 수 없습니다."
       }
      ],
      "all_candidates_fail": true,
      "readings": [
       {
        "label": "A",
        "direction": "복만의 시선이 화면 우측을 향하고 있으나, 뚜렷한 대상을 쏘아보는 느낌이 부족하고 다소 모호하게 처리됨.",
        "built_space": "야외 대나무숲을 배경으로 두 사람이 서 있는 공간이 적절히 구성됨.",
        "entities": "복만(수염, 두루마기)과 염재보(누비옷)가 지시대로 등장하며, 프롬프트의 지시대로 추가 배경 인물은 없음.",
        "hard_violations": [
         "physically impossible anatomy (염재보의 왼쪽 어깨에 얹힌 손의 손가락이 6개이며 좌우 구조가 뒤바뀐 기형적인 형태임)",
         "unsupported object (복만의 등 뒤에 있는 화승총을 지탱하는 어깨끈이나 손이 전혀 보이지 않아 무기가 허공에 떠 있음)"
        ],
        "physics": "화승총이 어떠한 지지대(끈 등)도 없이 허공에 떠 있으며, 포옹 상태에서 어깨에 얹혀진 손의 뼈대와 방향이 인체공학적으로 완전히 어긋남."
       },
       {
        "label": "B",
        "direction": "복만이 화면 우측 배경에 서 있는 인물(안규홍)을 향해 굳고 매서운 눈빛으로 정확히 시선을 고정하고 있음.",
        "built_space": "대나무숲 배경 속에서 포옹하는 두 사람과 뒤편의 인물이 공간적으로 무리 없이 배치됨.",
        "entities": "복만과 염재보의 외형(수염, 누비옷 등)이 잘 묘사되었으나, 프롬프트가 엄격히 추가를 금지한 제3의 인물(안규홍)이 배경에 등장함.",
        "hard_violations": [
         "a person placed where the staging does not put them (배경 인물 추가를 명시적으로 금지한 프롬프트의 'PEOPLE' 및 'KEY BACKGROUND ELEMENTS' 제약을 어기고 제3의 인물이 배치됨)"
        ],
        "physics": "복만의 가슴을 가로지르는 가죽 끈이 화승총을 안정적으로 지탱하고 있으며, 염재보가 복만의 어깨를 짚고 있는 손의 형태와 무게감도 자연스러움."
       }
      ]
     },
     "ok": true
    },
    {
     "model": "gpt-high",
     "order": "reverse",
     "display_to_canonical": {
      "A": "B",
      "B": "A"
     },
     "raw": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 8,
        "verdict_ko": "염재보의 등 너머 복만의 굳은 눈빛을 클로즈업하고 추가 인물을 배제해 핵심 구도를 지켰으나, 시선이 렌즈 가까이를 향하고 대나무 배경의 비중이 다소 크다."
       },
       {
        "label": "A",
        "score": 2,
        "verdict_ko": "눈빛의 긴장과 밀착 구도는 맞지만, 금지된 세 번째 인물을 배경에 추가했으며 복만의 시선도 그 인물이 있는 방향과 맞지 않는다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "복만의 눈은 염재보의 머리 옆을 지나 화면 왼쪽의 화면 밖 지점을 향한다. 오른쪽 뒤에 보이는 세 번째 남성을 바라보는 방향은 아니다. 염재보의 눈은 뒤통수에 가려져 확인할 수 없고, 배경 남성은 두 사람 쪽을 향한다. 총은 오른쪽 위로 비스듬히 올라가며 사람을 겨누는 자세가 아니다.",
        "built_space": "인공 구조물이나 고정 설비는 없다. 왼쪽과 오른쪽에 대나무가 있고 오른쪽 아래에는 바위 일부가 보여 이전 장소의 재료와 식생이 이어진다. 염재보는 왼쪽 전경, 복만은 중앙 오른쪽에 밀착해 있으나, 두 사람 너머의 제한된 대나무 배경이어야 할 공간에 세 번째 남성이 서 있다.",
        "entities": "전경에는 성인 한국인 남성으로 표현된 염재보의 검은 머리와 낡은 청회색 누비옷, 복만의 검은 머리와 거친 수염 및 해진 회갈색 두루마기가 보인다. 이전 스틸의 복장과 대체로 일치하지만 복만의 머리는 더 흐트러져 있다. 목재와 금속으로 된 장총 및 어깨끈이 보이고 염재보에게 별도 총은 없다. 허용된 두 사람 외에 검은 머리띠를 한 성인 남성의 얼굴과 상체가 추가되었다. 읽을 수 있는 글자는 없다.",
        "hard_violations": [
         "허용 인물이 복만과 염재보로 제한되고 오해를 낳는 배경 인물도 명시적으로 금지되었는데, 오른쪽 배경에 세 번째 남성의 얼굴과 상체를 추가했다."
        ],
        "physics": "염재보의 팔은 복만의 어깨로 이어지고 손이 어깨에 얹혀 있어 포옹의 접촉이 자연스럽다. 복만의 장총은 어깨를 가로지르는 끈과 몸 뒤의 위치로 휴대 상태가 설명된다. 하체는 프레임 밖이며 공중에 뜬 몸이나 지지 없이 떠 있는 물체는 보이지 않는다."
       },
       {
        "label": "B",
        "direction": "복만은 염재보의 어깨 너머 카메라 가까운 정면의 화면 밖 지점을 응시한다. 안규홍은 보이지 않아 표적의 실제 위치를 확인할 수 없지만, 화면 밖에 선 상대를 노려보는 배치는 가능하다. 염재보의 얼굴과 시선은 가려져 있다. 장총은 오른쪽 위를 향한 대각선 휴대 방향이며 조준 동작은 없다.",
        "built_space": "인공 구조물과 고정 설비는 없다. 왼쪽 전경의 염재보와 중앙의 복만 뒤로 대나무가 이어지고 오른쪽 아래에 바위 한 덩어리의 일부가 보인다. 이전 스틸의 바위 옆 대나무 길과 양립하며 추가 인물이나 반사는 없다. 다만 머리 옆에 제한적으로 보여야 할 대나무가 오른쪽 배경에서 비교적 넓고 선명하게 드러난다.",
        "entities": "보이는 인물은 성인 한국인 남성으로 표현된 복만과 염재보 두 명뿐이다. 염재보의 검은 머리, 머리띠와 해진 청회색 누비옷, 복만의 거친 수염과 검은 머리 및 낡은 회갈색 두루마기가 이전 스틸과 대체로 맞는다. 염재보는 뒷모습이라 얼굴 동일성은 확인할 수 없다. 장총 한 정이 대각선으로 보이며 염재보의 별도 무기는 없다. 안규홍은 화면 밖에 있고 읽을 수 있는 문자는 없다.",
        "hard_violations": [],
        "physics": "복만의 손은 염재보의 어깨에, 염재보의 손은 복만의 어깨에 닿아 있어 서로 어깨를 안은 자세가 성립한다. 장총의 아래쪽은 두 몸과 팔에 가려져 정확한 끈 연결이나 지지점은 확인되지 않지만, 몸에 기대어 대각선으로 휴대한 배치이며 떠 있다고 볼 근거는 없다. 하체는 클로즈업 밖으로 잘려 있고 보이는 상체와 팔에는 불가능한 관절이나 접촉이 없다."
       }
      ],
      "all_candidates_fail": false
     },
     "normalized": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 8,
        "verdict_ko": "염재보의 등 너머 복만의 굳은 눈빛을 클로즈업하고 추가 인물을 배제해 핵심 구도를 지켰으나, 시선이 렌즈 가까이를 향하고 대나무 배경의 비중이 다소 크다."
       },
       {
        "label": "B",
        "score": 2,
        "verdict_ko": "눈빛의 긴장과 밀착 구도는 맞지만, 금지된 세 번째 인물을 배경에 추가했으며 복만의 시선도 그 인물이 있는 방향과 맞지 않는다."
       }
      ],
      "all_candidates_fail": false,
      "readings": [
       {
        "label": "B",
        "direction": "복만의 눈은 염재보의 머리 옆을 지나 화면 왼쪽의 화면 밖 지점을 향한다. 오른쪽 뒤에 보이는 세 번째 남성을 바라보는 방향은 아니다. 염재보의 눈은 뒤통수에 가려져 확인할 수 없고, 배경 남성은 두 사람 쪽을 향한다. 총은 오른쪽 위로 비스듬히 올라가며 사람을 겨누는 자세가 아니다.",
        "built_space": "인공 구조물이나 고정 설비는 없다. 왼쪽과 오른쪽에 대나무가 있고 오른쪽 아래에는 바위 일부가 보여 이전 장소의 재료와 식생이 이어진다. 염재보는 왼쪽 전경, 복만은 중앙 오른쪽에 밀착해 있으나, 두 사람 너머의 제한된 대나무 배경이어야 할 공간에 세 번째 남성이 서 있다.",
        "entities": "전경에는 성인 한국인 남성으로 표현된 염재보의 검은 머리와 낡은 청회색 누비옷, 복만의 검은 머리와 거친 수염 및 해진 회갈색 두루마기가 보인다. 이전 스틸의 복장과 대체로 일치하지만 복만의 머리는 더 흐트러져 있다. 목재와 금속으로 된 장총 및 어깨끈이 보이고 염재보에게 별도 총은 없다. 허용된 두 사람 외에 검은 머리띠를 한 성인 남성의 얼굴과 상체가 추가되었다. 읽을 수 있는 글자는 없다.",
        "hard_violations": [
         "허용 인물이 복만과 염재보로 제한되고 오해를 낳는 배경 인물도 명시적으로 금지되었는데, 오른쪽 배경에 세 번째 남성의 얼굴과 상체를 추가했다."
        ],
        "physics": "염재보의 팔은 복만의 어깨로 이어지고 손이 어깨에 얹혀 있어 포옹의 접촉이 자연스럽다. 복만의 장총은 어깨를 가로지르는 끈과 몸 뒤의 위치로 휴대 상태가 설명된다. 하체는 프레임 밖이며 공중에 뜬 몸이나 지지 없이 떠 있는 물체는 보이지 않는다."
       },
       {
        "label": "A",
        "direction": "복만은 염재보의 어깨 너머 카메라 가까운 정면의 화면 밖 지점을 응시한다. 안규홍은 보이지 않아 표적의 실제 위치를 확인할 수 없지만, 화면 밖에 선 상대를 노려보는 배치는 가능하다. 염재보의 얼굴과 시선은 가려져 있다. 장총은 오른쪽 위를 향한 대각선 휴대 방향이며 조준 동작은 없다.",
        "built_space": "인공 구조물과 고정 설비는 없다. 왼쪽 전경의 염재보와 중앙의 복만 뒤로 대나무가 이어지고 오른쪽 아래에 바위 한 덩어리의 일부가 보인다. 이전 스틸의 바위 옆 대나무 길과 양립하며 추가 인물이나 반사는 없다. 다만 머리 옆에 제한적으로 보여야 할 대나무가 오른쪽 배경에서 비교적 넓고 선명하게 드러난다.",
        "entities": "보이는 인물은 성인 한국인 남성으로 표현된 복만과 염재보 두 명뿐이다. 염재보의 검은 머리, 머리띠와 해진 청회색 누비옷, 복만의 거친 수염과 검은 머리 및 낡은 회갈색 두루마기가 이전 스틸과 대체로 맞는다. 염재보는 뒷모습이라 얼굴 동일성은 확인할 수 없다. 장총 한 정이 대각선으로 보이며 염재보의 별도 무기는 없다. 안규홍은 화면 밖에 있고 읽을 수 있는 문자는 없다.",
        "hard_violations": [],
        "physics": "복만의 손은 염재보의 어깨에, 염재보의 손은 복만의 어깨에 닿아 있어 서로 어깨를 안은 자세가 성립한다. 장총의 아래쪽은 두 몸과 팔에 가려져 정확한 끈 연결이나 지지점은 확인되지 않지만, 몸에 기대어 대각선으로 휴대한 배치이며 떠 있다고 볼 근거는 없다. 하체는 클로즈업 밖으로 잘려 있고 보이는 상체와 팔에는 불가능한 관절이나 접촉이 없다."
       }
      ]
     },
     "ok": true
    }
   ],
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "slot_winner_match": false,
   "slot_winner": {
    "gemini-pro": "B",
    "gpt-high": "A"
   },
   "route": "cross_slot_combined"
  },
  "dual": {
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "normalized": {
    "A": 1.5,
    "B": 1.25
   },
   "adjusted": {
    "A": 1.25,
    "B": 1.0
   },
   "violations": {
    "A": [
     "[gemini-pro] physically impossible anatomy (염재보의 왼쪽 어깨에 얹힌 손의 손가락이 6개이며 좌우 구조가 뒤바뀐 기형적인 형태임)",
     "[gemini-pro] unsupported object (복만의 등 뒤에 있는 화승총을 지탱하는 어깨끈이나 손이 전혀 보이지 않아 무기가 허공에 떠 있음)"
    ],
    "B": [
     "[gemini-pro] a person placed where the staging does not put them (배경 인물 추가를 명시적으로 금지한 프롬프트의 'PEOPLE' 및 'KEY BACKGROUND ELEMENTS' 제약을 어기고 제3의 인물이 배치됨)",
     "[gpt-high] 허용 인물이 복만과 염재보로 제한되고 오해를 낳는 배경 인물도 명시적으로 금지되었는데, 오른쪽 배경에 세 번째 남성의 얼굴과 상체를 추가했다."
    ]
   },
   "per_model_winner": {
    "gemini-pro": "B",
    "gpt-high": "A"
   },
   "agreed": false
  },
  "totals": {
   "B": 1000,
   "A": 1250
  },
  "selected": "A",
  "ranking": [
   "A",
   "B"
  ],
  "verdicts": [
   {
    "label": "B",
    "score": 1000,
    "verdict_ko": "총기를 지탱하는 물리적 구조와 매서운 표정 연기는 훌륭하게 구현되었으나, 프롬프트에서 명시적으로 추가를 금지한 제3의 인물이 배경에 등장하여 치명적인 제약 위반이 발생했습니다.  ★위반: [gemini-pro] a person placed where the staging does not put them (배경 인물 추가를 명시적으로 금지한 프롬프트의 'PEOPLE' 및 'KEY BACKGROUND ELEMENTS' 제약을 어기고 제3의 인물이 배치됨) / [gpt-high] 허용 인물이 복만과 염재보로 제한되고 오해를 낳는 배경 인물도 명시적으로 금지되었는데, 오른쪽 배경에 세 번째 남성의 얼굴과 상체를 추가했다."
   },
   {
    "label": "A",
    "score": 1250,
    "verdict_ko": "해부학적으로 불가능하게 일그러진 손의 형태와 아무것도 지탱하지 않은 채 공중에 떠 있는 화승총 등 심각한 물리적 오류가 있어 사용할 수 없습니다.  ★위반: [gemini-pro] physically impossible anatomy (염재보의 왼쪽 어깨에 얹힌 손의 손가락이 6개이며 좌우 구조가 뒤바뀐 기형적인 형태임) / [gemini-pro] unsupported object (복만의 등 뒤에 있는 화승총을 지탱하는 어깨끈이나 손이 전혀 보이지 않아 무기가 허공에 떠 있음)"
   }
  ],
  "refs": [
   {
    "label": "PREVIOUS SHOT STILL — a visually related earlier shot of this same place: the location's look, materials, fixed features and lighting mood are LOCKED to this photo; never copy its camera framing. The clothing, hair and overall look of 복만, 염재보 — who appear both in that photo and in this shot — are LOCKED to that photo. Anyone else visible in that photo is NOT in this shot: never carry their face, body or clothing onto anyone here. If this photo shows a character who cannot move (dead or unconscious), that character's exact pose, position and orientation are ALSO LOCKED — treat their body as a fixed prop of the set that only the camera moves around.",
    "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/db00517b-165b-4005-a287-f28936b59b95/images/272bd7af-9294-4490-98c6-88a40e9deacc/scene/recipe/S13sh9_sel.png",
    "asset_id": "7f488b68-96b3-4a83-836c-bac8fcc4ead4",
    "role": "prev_still"
   },
   {
    "label": "CHARACTER REFERENCE — 염재보: the exact person appearing in this shot; match face, hair and build exactly.",
    "path": "<bytes:964351>",
    "asset_id": "ab287289-d0b7-436d-8f70-68452f898f73",
    "role": "character_ref"
   },
   {
    "label": "CHARACTER REFERENCE — 복만: the exact person appearing in this shot; match face, hair and build exactly.",
    "path": "<bytes:1027479>",
    "asset_id": "b7e69570-d15d-4241-b5ea-4cb8d4ece1cf",
    "role": "character_ref"
   }
  ],
  "critique_skipped": true,
  "shot_run_produced": false,
  "shot_run_uid": "06aa140f-1f30-7ee4-af9a-0e68b64f9b59",
  "ref_mode": "prev+엔티티",
  "share_plan": {
   "ref_plan": "prev",
   "prev_anchor_tag": "S13sh9"
  }
 },
 "S13sh10::cine": {
  "applied": true,
  "attempted_at": "2026-09-09T02:12:14.912095+00:00",
  "fingerprint": "3eb5b70ac2b04e32d8eeaca96afb2929a0cda28844dcb2d3f27e8e6cc358a182",
  "fingerprint_version": 2,
  "provider": "mai",
  "endpoint": "openrouter/chat-completions",
  "model": "microsoft/mai-image-2.6",
  "pack": "24.202608252115",
  "source_file": "S13sh10_sel.png",
  "source_sha256": "bde30d5ccf347a47d3fbd1612680cc0d814b4ee188ae395d30ac088ae9b57449",
  "file": "S13sh10_cine.png",
  "staged_sha256": "ff538e0d49989218522dcd53d369d2a9252dceb436cd2e1f65c40ff7ac9104d0",
  "latency_ms": 27229
 },
 "S14sh1::signage": {
  "fp": "f735f820b6a05f25",
  "inscriptions": [],
  "cues": [],
  "dropped": []
 },
 "S14sh1": {
  "input_fingerprint": "3a7d5bb401fd820a",
  "prompt": "Create ONE FINAL photorealistic live-action film still of the moment below — early 20th-century Korea, Jeolla region; all people are Korean unless explicitly stated otherwise. TIME OF DAY (lock): late afternoon.\n\nSHOT TEXT (authoritative, Korean): 낡은 천과 거적을 엮어 만든 임시 막사들 주변으로 짐들이 어수선하게 쌓여 있는 전경.\n\nLOCATION (lock): The outdoor shelter cluster of a temporary lakeside camp, with cloth-and-mat huts and tied loads on open ground before a bamboo windbreak. The shot takes place here — the attached LOCATION STRUCTURE PHOTOGRAPH is the single authority for this exact place — its fixed structure and permanent site details are LOCKED to it. No separate location photograph exists for this place. Build everything else strictly from the location text above and the shot text; the layout sketch (when attached) governs framing and placement only, and the shot text governs time of day, lighting and action.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: wide shot\n- KEY BACKGROUND ELEMENTS: 임시 막사들 (Several temporary shelters are assembled from old cloth and woven mats) — Their upper surfaces and oblique sides are visible from the elevated path-end position; used as Forms a broken middle-ground band rather than a uniform row; 등짐과 곡식 자루들 (Tied for departure and grouped irregularly around the shelters) — Viewed from above and to one side, with separate bundles retaining natural differences in placement; used as Foreground and midground evidence that the camp can move quickly; 자갈밭과 대지 (Extend around the lakeside camp); used as Creates readable intervals between the scattered shelters and supplies; 호수 (The broad lake extends beyond the camp) — Visible beyond the shelters across the upper background; no reflected subject is emphasized; used as Provides geographic depth and the camp's lakeside context.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Natural late-afternoon light, restrained saturation, and balanced contrast preserve the worn camp textures without romanticizing the refuge.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): Late-afternoon light falls over lakeside gravel, makeshift cloth-and-mat shelters and bundles tied ready for departure; thin smoke rises from a stone hearth.\n\nNO PEOPLE IN THIS SHOT: the shot text shows only the place and its state — no living person appears in frame, and no face, and no figure standing, sitting or passing through. Do not invent someone to occupy the space.\n\nOne thing this rule does NOT forbid: the hand or arm of whoever is doing what the shot text describes. If the moment is an object being held, carried, operated, read, aimed, worn or handed over, the hand doing it belongs in frame, resting on or gripping that object exactly as its real use requires. An object photographed floating unsupported, with the hand that must be holding it erased, is a worse failure than the person you were avoiding — render the hand.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.",
  "roll_prompts": {
   "A": "Create ONE FINAL photorealistic live-action film still of the moment below — early 20th-century Korea, Jeolla region; all people are Korean unless explicitly stated otherwise. TIME OF DAY (lock): late afternoon.\n\nSHOT TEXT (authoritative, Korean): 낡은 천과 거적을 엮어 만든 임시 막사들 주변으로 짐들이 어수선하게 쌓여 있는 전경.\n\nLOCATION (lock): The outdoor shelter cluster of a temporary lakeside camp, with cloth-and-mat huts and tied loads on open ground before a bamboo windbreak. The shot takes place here — the attached LOCATION STRUCTURE PHOTOGRAPH is the single authority for this exact place — its fixed structure and permanent site details are LOCKED to it. No separate location photograph exists for this place. Build everything else strictly from the location text above and the shot text; the layout sketch (when attached) governs framing and placement only, and the shot text governs time of day, lighting and action.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: wide shot\n- KEY BACKGROUND ELEMENTS: 임시 막사들 (Several temporary shelters are assembled from old cloth and woven mats) — Their upper surfaces and oblique sides are visible from the elevated path-end position; used as Forms a broken middle-ground band rather than a uniform row; 등짐과 곡식 자루들 (Tied for departure and grouped irregularly around the shelters) — Viewed from above and to one side, with separate bundles retaining natural differences in placement; used as Foreground and midground evidence that the camp can move quickly; 자갈밭과 대지 (Extend around the lakeside camp); used as Creates readable intervals between the scattered shelters and supplies; 호수 (The broad lake extends beyond the camp) — Visible beyond the shelters across the upper background; no reflected subject is emphasized; used as Provides geographic depth and the camp's lakeside context.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Natural late-afternoon light, restrained saturation, and balanced contrast preserve the worn camp textures without romanticizing the refuge.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): Late-afternoon light falls over lakeside gravel, makeshift cloth-and-mat shelters and bundles tied ready for departure; thin smoke rises from a stone hearth.\n\nNO PEOPLE IN THIS SHOT: the shot text shows only the place and its state — no living person appears in frame, and no face, and no figure standing, sitting or passing through. Do not invent someone to occupy the space.\n\nOne thing this rule does NOT forbid: the hand or arm of whoever is doing what the shot text describes. If the moment is an object being held, carried, operated, read, aimed, worn or handed over, the hand doing it belongs in frame, resting on or gripping that object exactly as its real use requires. An object photographed floating unsupported, with the hand that must be holding it erased, is a worse failure than the person you were avoiding — render the hand.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.",
   "B": "Create ONE FINAL photorealistic live-action film still of the moment below — early 20th-century Korea, Jeolla region; all people are Korean unless explicitly stated otherwise. TIME OF DAY (lock): late afternoon.\n\nSHOT TEXT (authoritative, Korean): 낡은 천과 거적을 엮어 만든 임시 막사들 주변으로 짐들이 어수선하게 쌓여 있는 전경.\n\nLOCATION (lock): The outdoor shelter cluster of a temporary lakeside camp, with cloth-and-mat huts and tied loads on open ground before a bamboo windbreak. The shot takes place here — the attached LOCATION STRUCTURE PHOTOGRAPH is the single authority for this exact place — its fixed structure and permanent site details are LOCKED to it. No separate location photograph exists for this place. Build everything else strictly from the location text above and the shot text; the layout sketch (when attached) governs framing and placement only, and the shot text governs time of day, lighting and action.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: wide shot\n- KEY BACKGROUND ELEMENTS: 임시 막사들 (Several temporary shelters are assembled from old cloth and woven mats) — Their upper surfaces and oblique sides are visible from the elevated path-end position; used as Forms a broken middle-ground band rather than a uniform row; 등짐과 곡식 자루들 (Tied for departure and grouped irregularly around the shelters) — Viewed from above and to one side, with separate bundles retaining natural differences in placement; used as Foreground and midground evidence that the camp can move quickly; 자갈밭과 대지 (Extend around the lakeside camp); used as Creates readable intervals between the scattered shelters and supplies; 호수 (The broad lake extends beyond the camp) — Visible beyond the shelters across the upper background; no reflected subject is emphasized; used as Provides geographic depth and the camp's lakeside context.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Natural late-afternoon light, restrained saturation, and balanced contrast preserve the worn camp textures without romanticizing the refuge.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): Late-afternoon light falls over lakeside gravel, makeshift cloth-and-mat shelters and bundles tied ready for departure; thin smoke rises from a stone hearth.\n\nNO PEOPLE IN THIS SHOT: the shot text shows only the place and its state — no living person appears in frame, and no face, and no figure standing, sitting or passing through. Do not invent someone to occupy the space.\n\nOne thing this rule does NOT forbid: the hand or arm of whoever is doing what the shot text describes. If the moment is an object being held, carried, operated, read, aimed, worn or handed over, the hand doing it belongs in frame, resting on or gripping that object exactly as its real use requires. An object photographed floating unsupported, with the hand that must be holding it erased, is a worse failure than the person you were avoiding — render the hand.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.\n\nCOMPOSITION VARIATION (this candidate only): keep the place, the people and the moment exactly as specified, and keep from the CAMERA & FRAME contract its subject scale, screen placement and key background elements exactly. For this candidate only, its camera angle and height are deliberately left open: choose a camera position distinctly different from the obvious one, as a film director picking a second setup on the same blocking. Stay true to any explicit framing the SHOT TEXT itself dictates."
  },
  "shot_run_spend_attempt_count": 1,
  "readings": [
   {
    "label": "A",
    "direction": "조준하거나 시선이 향하는 인물 및 사물이 존재하지 않음.",
    "built_space": "호숫가 자갈밭에 5채의 막사가 배치됨. 배경의 막사 3채는 레퍼런스의 배치를 흉내 냈으나 호수를 등지는 방향으로 지어져 원본의 공간 방향(호수를 향함)과 모순됨. 중앙에 돌 화덕이 놓여 있음.",
    "entities": "호수, 대나무 숲, 막사 등 주요 요소는 프롬프트에 부합함. 그러나 전경 좌측 짐 무더기 속에 20세기 초 한국 배경과 완전히 어긋나는 현대식 캔버스 배낭과 파란색 침낭이 명확하게 묘사됨.",
    "hard_violations": [
     "[gpt-high] 고정된 장소 사진의 세 막사 외에 두 동을 추가하여, 동일 장소여야 한다는 구조 고정을 위반했습니다."
    ],
    "physics": "모든 막사, 짐, 화덕이 지면에 안정적으로 놓여 있으며, 화덕에서 피어오르는 연기도 자연스러움."
   },
   {
    "label": "B",
    "direction": "조준하거나 시선이 향하는 인물 및 사물이 존재하지 않음.",
    "built_space": "상단 좌측의 호수를 배경으로 자갈밭 위에 약 5채의 천과 거적 막사가 불규칙하게 배치됨. 우측에 대나무 숲이 위치하고 중앙에 돌 화덕이 있음. 레퍼런스의 나뭇가지 지붕 디테일은 천 덮개로 대체됨.",
    "entities": "막사, 묶인 짐들, 호수, 자갈밭, 대나무 방풍림 모두 지시된 대로 묘사됨. 짚으로 엮은 밧줄, 가마니, 도자기 등 20세기 초 한국의 시대상을 정확히 반영하는 소품들이 배치됨.",
    "hard_violations": [
     "[gpt-high] 고정된 장소 사진에 있는 세 동을 넘어 막사를 여섯 동으로 증설하여, 장소의 구조물 수와 배치를 변경했습니다."
    ],
    "physics": "짐 무더기와 막사들이 자갈밭 지면에 안정적으로 밀착되어 있으며, 돌 화덕에서 발생하는 연기의 흐름이 물리적으로 타당함."
   }
  ],
  "cross_model_order": {
   "policy": "cross_model_order_v2_slot0_fwd_slot1_rev_parallel2",
   "slots": [
    {
     "model": "gemini-pro",
     "order": "forward",
     "display_to_canonical": {
      "A": "A",
      "B": "B"
     },
     "raw": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 8,
        "verdict_ko": "늦은 오후의 자연광과 지정된 하이앵글 구도를 잘 살렸으며, 새끼줄과 가마니 등 20세기 초 시대상에 맞는 소품들을 사실적인 재질감으로 훌륭하게 구현했습니다."
       },
       {
        "label": "A",
        "score": 4,
        "verdict_ko": "구도와 조명은 무난하나, 20세기 초 시대 배경에 전혀 맞지 않는 현대식 등산 배낭과 나일론 침낭이 전경에 두드러지게 등장하여 소품 및 시대적 지시를 크게 위반했습니다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "조준하거나 시선이 향하는 인물 및 사물이 존재하지 않음.",
        "built_space": "호숫가 자갈밭에 5채의 막사가 배치됨. 배경의 막사 3채는 레퍼런스의 배치를 흉내 냈으나 호수를 등지는 방향으로 지어져 원본의 공간 방향(호수를 향함)과 모순됨. 중앙에 돌 화덕이 놓여 있음.",
        "entities": "호수, 대나무 숲, 막사 등 주요 요소는 프롬프트에 부합함. 그러나 전경 좌측 짐 무더기 속에 20세기 초 한국 배경과 완전히 어긋나는 현대식 캔버스 배낭과 파란색 침낭이 명확하게 묘사됨.",
        "hard_violations": [],
        "physics": "모든 막사, 짐, 화덕이 지면에 안정적으로 놓여 있으며, 화덕에서 피어오르는 연기도 자연스러움."
       },
       {
        "label": "B",
        "direction": "조준하거나 시선이 향하는 인물 및 사물이 존재하지 않음.",
        "built_space": "상단 좌측의 호수를 배경으로 자갈밭 위에 약 5채의 천과 거적 막사가 불규칙하게 배치됨. 우측에 대나무 숲이 위치하고 중앙에 돌 화덕이 있음. 레퍼런스의 나뭇가지 지붕 디테일은 천 덮개로 대체됨.",
        "entities": "막사, 묶인 짐들, 호수, 자갈밭, 대나무 방풍림 모두 지시된 대로 묘사됨. 짚으로 엮은 밧줄, 가마니, 도자기 등 20세기 초 한국의 시대상을 정확히 반영하는 소품들이 배치됨.",
        "hard_violations": [],
        "physics": "짐 무더기와 막사들이 자갈밭 지면에 안정적으로 밀착되어 있으며, 돌 화덕에서 발생하는 연기의 흐름이 물리적으로 타당함."
       }
      ],
      "all_candidates_fail": false
     },
     "normalized": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 8,
        "verdict_ko": "늦은 오후의 자연광과 지정된 하이앵글 구도를 잘 살렸으며, 새끼줄과 가마니 등 20세기 초 시대상에 맞는 소품들을 사실적인 재질감으로 훌륭하게 구현했습니다."
       },
       {
        "label": "A",
        "score": 4,
        "verdict_ko": "구도와 조명은 무난하나, 20세기 초 시대 배경에 전혀 맞지 않는 현대식 등산 배낭과 나일론 침낭이 전경에 두드러지게 등장하여 소품 및 시대적 지시를 크게 위반했습니다."
       }
      ],
      "all_candidates_fail": false,
      "readings": [
       {
        "label": "A",
        "direction": "조준하거나 시선이 향하는 인물 및 사물이 존재하지 않음.",
        "built_space": "호숫가 자갈밭에 5채의 막사가 배치됨. 배경의 막사 3채는 레퍼런스의 배치를 흉내 냈으나 호수를 등지는 방향으로 지어져 원본의 공간 방향(호수를 향함)과 모순됨. 중앙에 돌 화덕이 놓여 있음.",
        "entities": "호수, 대나무 숲, 막사 등 주요 요소는 프롬프트에 부합함. 그러나 전경 좌측 짐 무더기 속에 20세기 초 한국 배경과 완전히 어긋나는 현대식 캔버스 배낭과 파란색 침낭이 명확하게 묘사됨.",
        "hard_violations": [],
        "physics": "모든 막사, 짐, 화덕이 지면에 안정적으로 놓여 있으며, 화덕에서 피어오르는 연기도 자연스러움."
       },
       {
        "label": "B",
        "direction": "조준하거나 시선이 향하는 인물 및 사물이 존재하지 않음.",
        "built_space": "상단 좌측의 호수를 배경으로 자갈밭 위에 약 5채의 천과 거적 막사가 불규칙하게 배치됨. 우측에 대나무 숲이 위치하고 중앙에 돌 화덕이 있음. 레퍼런스의 나뭇가지 지붕 디테일은 천 덮개로 대체됨.",
        "entities": "막사, 묶인 짐들, 호수, 자갈밭, 대나무 방풍림 모두 지시된 대로 묘사됨. 짚으로 엮은 밧줄, 가마니, 도자기 등 20세기 초 한국의 시대상을 정확히 반영하는 소품들이 배치됨.",
        "hard_violations": [],
        "physics": "짐 무더기와 막사들이 자갈밭 지면에 안정적으로 밀착되어 있으며, 돌 화덕에서 발생하는 연기의 흐름이 물리적으로 타당함."
       }
      ]
     },
     "ok": true
    },
    {
     "model": "gpt-high",
     "order": "reverse",
     "display_to_canonical": {
      "A": "B",
      "B": "A"
     },
     "raw": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 4,
        "verdict_ko": "높은 위치의 와이드 숏과 호수 배경은 맞지만, 기준의 세 막사를 여섯 동으로 늘렸고 지붕 구조도 크게 바꾸어 장소 고정을 위반합니다."
       },
       {
        "label": "B",
        "score": 5,
        "verdict_ko": "막사를 다섯 동으로 늘린 장소 위반은 있으나, 기준의 거적 벽·나뭇가지 지붕과 출발 준비된 등짐·곡식 자루를 A보다 충실히 재현합니다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "카메라는 높은 위치에서 막사의 지붕과 비스듬한 옆면을 내려다봅니다. 호수는 막사 너머 화면 왼쪽 상단에 펼쳐지고, 중앙 돌 화덕의 연기는 위로 올라갑니다. 사람의 시선이나 겨냥하는 물체는 없습니다.",
        "built_space": "앞쪽 좌우 두 동, 왼쪽 호숫가 한 동, 중앙 한 동, 중앙 뒤 천막 한 동, 오른쪽 뒤 한 동으로 총 여섯 막사가 보입니다. 기준 사진의 세 동과 수가 다르며, 대부분의 지붕을 천과 돗자리 패치로 바꾸었습니다. 대나무 숲은 오른쪽 뒤에 있고 자갈과 흙 공터가 막사 사이로 이어집니다. 돌 화덕 한 곳이 중앙 오른쪽에 있으며, 불가능한 반사는 보이지 않습니다.",
        "entities": "낡은 천과 거적 막사, 호수, 대나무, 자갈밭, 묶은 짐, 자루와 돌 화덕이 모두 보입니다. 전경에는 밧줄 뭉치와 말린 돗자리가 두드러지고, 명확한 지게 등짐은 B보다 덜 드러납니다. 사람이나 얼굴, 읽을 수 있는 글자는 없습니다. 빛은 늦은 오후의 자연광으로 읽힙니다.",
        "hard_violations": [
         "고정된 장소 사진에 있는 세 동을 넘어 막사를 여섯 동으로 증설하여, 장소의 구조물 수와 배치를 변경했습니다."
        ],
        "physics": "막사 지붕은 나무 골조와 기둥에 받쳐져 있고 천은 골조 위로 처져 있습니다. 자루, 밧줄, 돗자리와 용기는 땅에 놓이거나 다른 짐에 기대어 있으며, 돌 화덕에서 연기가 올라옵니다. 지지 없이 떠 있는 물체나 신체는 없습니다."
       },
       {
        "label": "B",
        "direction": "카메라는 높은 진입부 쪽에서 공터와 막사 지붕을 비스듬히 내려다봅니다. 넓은 호수가 막사 너머 상단에 보이고, 중앙 화덕에서 나온 연기는 위쪽으로 퍼집니다. 지게들은 짐에 기대어 있으며, 시선이나 무기의 조준 방향을 확인할 대상은 없습니다.",
        "built_space": "전경 좌우 두 동, 중앙 한 동, 뒤쪽 좌우 두 동으로 총 다섯 막사가 보입니다. 기준 사진의 세 동보다 많지만, 중앙과 오른쪽 뒤 막사의 거적 벽, 외부 나무 골조, 거친 나뭇가지 지붕은 기준 구조에 더 가깝습니다. 중앙 화덕은 한 곳이며, 막사 사이에 흙 공터가 있고 전경에는 자갈이 깔려 있습니다. 대나무 숲은 오른쪽 뒤, 호수는 상단에 배치되어 있습니다.",
        "entities": "낡은 천과 거적을 쓴 막사, 끈으로 묶은 보따리, 곡식 자루로 읽히는 큰 자루, 지게 등짐, 자갈밭, 호수, 대나무와 연기 나는 돌 화덕이 보입니다. 짐은 좌우 막사 옆과 전경에 서로 다른 형태로 모여 있어 출발 준비 상태가 명확합니다. 사람, 얼굴, 읽을 수 있는 글자는 없고 자연광의 색과 대비도 절제되어 있습니다.",
        "hard_violations": [
         "고정된 장소 사진의 세 막사 외에 두 동을 추가하여, 동일 장소여야 한다는 구조 고정을 위반했습니다."
        ],
        "physics": "지게의 아래 끝은 지면에 닿고 위쪽은 짐 더미에 기대어 있습니다. 자루와 보따리는 땅 또는 아래쪽 짐에 받쳐져 있고, 막사 골조가 벽과 지붕을 지탱합니다. 화덕의 돌과 용기도 지지면이 보이며 연기는 화덕에서 발생합니다. 지지 없이 떠 있는 물체는 없습니다."
       }
      ],
      "all_candidates_fail": true
     },
     "normalized": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 4,
        "verdict_ko": "높은 위치의 와이드 숏과 호수 배경은 맞지만, 기준의 세 막사를 여섯 동으로 늘렸고 지붕 구조도 크게 바꾸어 장소 고정을 위반합니다."
       },
       {
        "label": "A",
        "score": 5,
        "verdict_ko": "막사를 다섯 동으로 늘린 장소 위반은 있으나, 기준의 거적 벽·나뭇가지 지붕과 출발 준비된 등짐·곡식 자루를 A보다 충실히 재현합니다."
       }
      ],
      "all_candidates_fail": true,
      "readings": [
       {
        "label": "B",
        "direction": "카메라는 높은 위치에서 막사의 지붕과 비스듬한 옆면을 내려다봅니다. 호수는 막사 너머 화면 왼쪽 상단에 펼쳐지고, 중앙 돌 화덕의 연기는 위로 올라갑니다. 사람의 시선이나 겨냥하는 물체는 없습니다.",
        "built_space": "앞쪽 좌우 두 동, 왼쪽 호숫가 한 동, 중앙 한 동, 중앙 뒤 천막 한 동, 오른쪽 뒤 한 동으로 총 여섯 막사가 보입니다. 기준 사진의 세 동과 수가 다르며, 대부분의 지붕을 천과 돗자리 패치로 바꾸었습니다. 대나무 숲은 오른쪽 뒤에 있고 자갈과 흙 공터가 막사 사이로 이어집니다. 돌 화덕 한 곳이 중앙 오른쪽에 있으며, 불가능한 반사는 보이지 않습니다.",
        "entities": "낡은 천과 거적 막사, 호수, 대나무, 자갈밭, 묶은 짐, 자루와 돌 화덕이 모두 보입니다. 전경에는 밧줄 뭉치와 말린 돗자리가 두드러지고, 명확한 지게 등짐은 B보다 덜 드러납니다. 사람이나 얼굴, 읽을 수 있는 글자는 없습니다. 빛은 늦은 오후의 자연광으로 읽힙니다.",
        "hard_violations": [
         "고정된 장소 사진에 있는 세 동을 넘어 막사를 여섯 동으로 증설하여, 장소의 구조물 수와 배치를 변경했습니다."
        ],
        "physics": "막사 지붕은 나무 골조와 기둥에 받쳐져 있고 천은 골조 위로 처져 있습니다. 자루, 밧줄, 돗자리와 용기는 땅에 놓이거나 다른 짐에 기대어 있으며, 돌 화덕에서 연기가 올라옵니다. 지지 없이 떠 있는 물체나 신체는 없습니다."
       },
       {
        "label": "A",
        "direction": "카메라는 높은 진입부 쪽에서 공터와 막사 지붕을 비스듬히 내려다봅니다. 넓은 호수가 막사 너머 상단에 보이고, 중앙 화덕에서 나온 연기는 위쪽으로 퍼집니다. 지게들은 짐에 기대어 있으며, 시선이나 무기의 조준 방향을 확인할 대상은 없습니다.",
        "built_space": "전경 좌우 두 동, 중앙 한 동, 뒤쪽 좌우 두 동으로 총 다섯 막사가 보입니다. 기준 사진의 세 동보다 많지만, 중앙과 오른쪽 뒤 막사의 거적 벽, 외부 나무 골조, 거친 나뭇가지 지붕은 기준 구조에 더 가깝습니다. 중앙 화덕은 한 곳이며, 막사 사이에 흙 공터가 있고 전경에는 자갈이 깔려 있습니다. 대나무 숲은 오른쪽 뒤, 호수는 상단에 배치되어 있습니다.",
        "entities": "낡은 천과 거적을 쓴 막사, 끈으로 묶은 보따리, 곡식 자루로 읽히는 큰 자루, 지게 등짐, 자갈밭, 호수, 대나무와 연기 나는 돌 화덕이 보입니다. 짐은 좌우 막사 옆과 전경에 서로 다른 형태로 모여 있어 출발 준비 상태가 명확합니다. 사람, 얼굴, 읽을 수 있는 글자는 없고 자연광의 색과 대비도 절제되어 있습니다.",
        "hard_violations": [
         "고정된 장소 사진의 세 막사 외에 두 동을 추가하여, 동일 장소여야 한다는 구조 고정을 위반했습니다."
        ],
        "physics": "지게의 아래 끝은 지면에 닿고 위쪽은 짐 더미에 기대어 있습니다. 자루와 보따리는 땅 또는 아래쪽 짐에 받쳐져 있고, 막사 골조가 벽과 지붕을 지탱합니다. 화덕의 돌과 용기도 지지면이 보이며 연기는 화덕에서 발생합니다. 지지 없이 떠 있는 물체는 없습니다."
       }
      ]
     },
     "ok": true
    }
   ],
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "slot_winner_match": false,
   "slot_winner": {
    "gemini-pro": "B",
    "gpt-high": "A"
   },
   "route": "cross_slot_combined"
  },
  "dual": {
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "normalized": {
    "A": 1.5,
    "B": 1.8
   },
   "adjusted": {
    "A": 1.25,
    "B": 1.55
   },
   "violations": {
    "B": [
     "[gpt-high] 고정된 장소 사진에 있는 세 동을 넘어 막사를 여섯 동으로 증설하여, 장소의 구조물 수와 배치를 변경했습니다."
    ],
    "A": [
     "[gpt-high] 고정된 장소 사진의 세 막사 외에 두 동을 추가하여, 동일 장소여야 한다는 구조 고정을 위반했습니다."
    ]
   },
   "per_model_winner": {
    "gemini-pro": "B",
    "gpt-high": "A"
   },
   "agreed": false
  },
  "totals": {
   "B": 1550,
   "A": 1250
  },
  "selected": "B",
  "ranking": [
   "B",
   "A"
  ],
  "verdicts": [
   {
    "label": "B",
    "score": 1550,
    "verdict_ko": "늦은 오후의 자연광과 지정된 하이앵글 구도를 잘 살렸으며, 새끼줄과 가마니 등 20세기 초 시대상에 맞는 소품들을 사실적인 재질감으로 훌륭하게 구현했습니다.  ★위반: [gpt-high] 고정된 장소 사진에 있는 세 동을 넘어 막사를 여섯 동으로 증설하여, 장소의 구조물 수와 배치를 변경했습니다."
   },
   {
    "label": "A",
    "score": 1250,
    "verdict_ko": "구도와 조명은 무난하나, 20세기 초 시대 배경에 전혀 맞지 않는 현대식 등산 배낭과 나일론 침낭이 전경에 두드러지게 등장하여 소품 및 시대적 지시를 크게 위반했습니다.  ★위반: [gpt-high] 고정된 장소 사진의 세 막사 외에 두 동을 추가하여, 동일 장소여야 한다는 구조 고정을 위반했습니다."
   }
  ],
  "refs": [
   {
    "label": "LOCATION STRUCTURE PHOTOGRAPH — the confirmed photograph of this exact place and its fixed structure: it is the SINGLE authority for the location, the structure's shape, proportions, materials, colors, openings and every permanent site detail. Never copy its camera framing, time of day or lighting — the shot text is the authority for those.",
    "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/db00517b-165b-4005-a287-f28936b59b95/images/background_chain/seed_bg_lakeside_camp_sel.png",
    "asset_id": "4d7542af-9887-40c1-ad92-e627c929288c",
    "role": "location_seed_bg"
   }
  ],
  "critique_skipped": true,
  "shot_run_produced": false,
  "shot_run_uid": "06aa140f-20c2-75f2-9566-9ace91af7b21",
  "ref_mode": "seed-bg만 (배경 전용)",
  "share_plan": {
   "ref_plan": "background"
  },
  "lane_policy": "ab_select_bypass:bg_only"
 },
 "S14sh1::cine": {
  "applied": true,
  "attempted_at": "2026-09-09T02:14:16.603319+00:00",
  "fingerprint": "bc347836a0553a8cc05ade7fca6ecd44cf7107689ab1d692f62719523fa224b2",
  "fingerprint_version": 2,
  "provider": "mai",
  "endpoint": "openrouter/chat-completions",
  "model": "microsoft/mai-image-2.6",
  "pack": "24.202608252115",
  "source_file": "S14sh1_sel.png",
  "source_sha256": "3e5b1ff95f290f9ec8b47ea9b83c664dcea159552243b84e31e4ac899b7bb553",
  "file": "S14sh1_cine.png",
  "staged_sha256": "337cfd5bc2eaa8a8b638d1be3190287d3d63a28de8bfca945f4d76190000d291",
  "latency_ms": 24221
 },
 "S14sh9::signage": {
  "fp": "524527e398ef103f",
  "inscriptions": [],
  "cues": [],
  "dropped": []
 },
 "era_assess::283e3a2cd754e918": {
  "subjects": [],
  "subject_text": "보성 호숫가 강용언 임시거점 자갈밭·공터·막사 앞\n넓은 호수와 자갈밭 앞에 낡은 천과 거적으로 엮은 막사가 모여 있다. 뒤로 대나무숲이 서고, 돌 화덕과 큰 무쇠솥이 놓인다.",
  "identity": "canonical",
  "scope_id": "L16",
  "scope_role": "location_exterior",
  "scope_sha": "ebb6fbd8bebad8c6"
 },
 "S14sh9::bgfirst_bg": {
  "input_fingerprint": "a82c30d92b1dd6e9",
  "prompt": "Create the EMPTY BACKGROUND PLATE for one film shot — NO PEOPLE, no figures, no body parts, no sketch lines, no arrows anywhere.\n\"Empty\" means no people only: KEEP the location's inherent occupants and stock that define the place — animals in an animal shelter, pen or farm, goods in a market, moored boats in a harbour — unless the shot text explicitly removes them.\nThe FIRST attached image is a thin-line storyboard sketch: use ONLY its camera angle, horizon, perspective and the placement/size of buildings and set masses — ignore the sketched people and arrows entirely. The SECOND attached image (LOCATION PHOTOGRAPH) is the real place: take its architecture, materials, signage and fixed features, and RE-PROJECT them into the sketch's camera. If the photograph's camera differs from the sketch's, the sketch's camera wins.\nHUMAN-SCALE CALIBRATION: derive every structure's true size from human-scale elements — a door ≈ 2m, a window ≈ 1–1.5m wide, one storey ≈ 2.5–3m; never inflate a small structure or shrink a large one.\n\nSHOT TEXT this background must serve (Korean): 중앙의 낡은 천막을 젖힌 채 밖으로 걸음을 내디딘 강용언의 당당한 전신.\n\nLOCATION (lock): The exterior entrance of the central rear tent in a temporary lakeside camp, opening toward the communal cooking area.\n\nTIME OF DAY (lock): late afternoon.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: wide shot\n- KEY BACKGROUND ELEMENTS: 중앙 천막 입구 (Its old cloth is held aside as 강용언 steps out) — The entrance and one adjoining side are seen obliquely behind his body; used as Frames the emergence without enlarging the tent cloth into an overpowering foreground object; 강용언의 칼 (Worn at his side as he exits) — Seen obliquely alongside his moving body; used as A proportionate character detail within the full-body silhouette.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Natural late-afternoon light gives 강용언's sun-darkened face and worn surroundings grounded definition, with restrained contrast rather than heroic spotlighting.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nRender ONE photorealistic empty location photograph, 16:9, that this shot can be staged inside later. No readable writing anywhere: surfaces that would carry writing may be present, but stage any wording out of legibility — an oblique angle, distance, shallow focus. No captions, watermarks or overlay text.",
  "effective_prompt": "Create the EMPTY BACKGROUND PLATE for one film shot — NO PEOPLE, no figures, no body parts, no sketch lines, no arrows anywhere.\n\"Empty\" means no people only: KEEP the location's inherent occupants and stock that define the place — animals in an animal shelter, pen or farm, goods in a market, moored boats in a harbour — unless the shot text explicitly removes them.\nThe FIRST attached image is a thin-line storyboard sketch: use ONLY its camera angle, horizon, perspective and the placement/size of buildings and set masses — ignore the sketched people and arrows entirely. The SECOND attached image (LOCATION PHOTOGRAPH) is the real place: take its architecture, materials, signage and fixed features, and RE-PROJECT them into the sketch's camera. If the photograph's camera differs from the sketch's, the sketch's camera wins.\nHUMAN-SCALE CALIBRATION: derive every structure's true size from human-scale elements — a door ≈ 2m, a window ≈ 1–1.5m wide, one storey ≈ 2.5–3m; never inflate a small structure or shrink a large one.\n\nSHOT TEXT this background must serve (Korean): 중앙의 낡은 천막을 젖힌 채 밖으로 걸음을 내디딘 강용언의 당당한 전신.\n\nLOCATION (lock): The exterior entrance of the central rear tent in a temporary lakeside camp, opening toward the communal cooking area.\n\nTIME OF DAY (lock): late afternoon.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: wide shot\n- KEY BACKGROUND ELEMENTS: 중앙 천막 입구 (Its old cloth is held aside as 강용언 steps out) — The entrance and one adjoining side are seen obliquely behind his body; used as Frames the emergence without enlarging the tent cloth into an overpowering foreground object; 강용언의 칼 (Worn at his side as he exits) — Seen obliquely alongside his moving body; used as A proportionate character detail within the full-body silhouette.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Natural late-afternoon light gives 강용언's sun-darkened face and worn surroundings grounded definition, with restrained contrast rather than heroic spotlighting.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nRender ONE photorealistic empty location photograph, 16:9, that this shot can be staged inside later. No readable writing anywhere: surfaces that would carry writing may be present, but stage any wording out of legibility — an oblique angle, distance, shallow focus. No captions, watermarks or overlay text.",
  "bg_path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/db00517b-165b-4005-a287-f28936b59b95/images/272bd7af-9294-4490-98c6-88a40e9deacc/scene/recipe/S14sh9__bgfirst_bg.png",
  "asset_id": "e4d2d435-263a-4691-bfaa-947714f65db4",
  "input_asset_ids": [
   "eb1859c6-e579-43d9-9942-d45d4fe3b7ba",
   "4d7542af-9887-40c1-ad92-e627c929288c"
  ]
 },
 "S14sh9": {
  "input_fingerprint": "fd03b44f29964615",
  "prompt": "Create ONE FINAL photorealistic live-action film still of the moment below — early 20th-century Korea, Jeolla region; all people are Korean unless explicitly stated otherwise. TIME OF DAY (lock): late afternoon.\n\nSHOT TEXT (authoritative, Korean): 중앙의 낡은 천막을 젖힌 채 밖으로 걸음을 내디딘 강용언의 당당한 전신.\n\nLOCATION (lock): The exterior entrance of the central rear tent in a temporary lakeside camp, opening toward the communal cooking area. The shot takes place here — the attached LOCATION STRUCTURE PHOTOGRAPH is the single authority for this exact place — its fixed structure and permanent site details are LOCKED to it. No separate location photograph exists for this place. Build everything else strictly from the location text above and the shot text; the layout sketch (when attached) governs framing and placement only, and the shot text governs time of day, lighting and action.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: wide shot\n- KEY BACKGROUND ELEMENTS: 중앙 천막 입구 (Its old cloth is held aside as 강용언 steps out) — The entrance and one adjoining side are seen obliquely behind his body; used as Frames the emergence without enlarging the tent cloth into an overpowering foreground object; 강용언의 칼 (Worn at his side as he exits) — Seen obliquely alongside his moving body; used as A proportionate character detail within the full-body silhouette.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Natural late-afternoon light gives 강용언's sun-darkened face and worn surroundings grounded definition, with restrained contrast rather than heroic spotlighting.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): Thin potato porridge boils in the large iron pot below the rock; individual earthenware bowls have been distributed on the cleared mat. 강용언: He emerges from the central tent with a sword at his waist and a sun-darkened face.\n\nTHE HAND THAT IS DOING THIS: the object at the centre of this shot is being held, operated, read, aimed or handed over by 강용언 right now, so 강용언's hand — and as much of the wrist and forearm as the framing reaches — is in the frame, gripping or resting on that object exactly the way its real use requires. Match that hand to 강용언: its size, build, skin, age, grooming, sleeve and anything worn on it belong to that person and to no one else.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 강용언 (한국인 남성, 성인 나이대, 햇볕에 그을린 얼굴, 자연스러운 검은 머리) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.",
  "roll_prompts": {
   "A": "Stage the shot. The FIRST attached image (SHOT BACKGROUND) is the finished empty background of this shot — keep it EXACTLY: its camera, perspective, architecture, lighting and every fixed feature stay untouched. The SECOND attached image (LAYOUT SKETCH) tells you ONLY where the people go: each sketched person's position, screen size, pose and the gaze/motion arrows. Ignore the sketch's background lines. The CHARACTER REFERENCE photographs show the real people.\nPlace the real people into the background at exactly the sketched positions, sizes and poses, following the arrow directions. No sketch lines or arrows may remain.\n\nCreate ONE FINAL photorealistic live-action film still of the moment below — early 20th-century Korea, Jeolla region; all people are Korean unless explicitly stated otherwise. TIME OF DAY (lock): late afternoon.\n\nSHOT TEXT (authoritative, Korean): 중앙의 낡은 천막을 젖힌 채 밖으로 걸음을 내디딘 강용언의 당당한 전신.\n\nLOCATION (lock): The exterior entrance of the central rear tent in a temporary lakeside camp, opening toward the communal cooking area. The shot takes place here — the FIRST attached image (SHOT BACKGROUND) is this exact place, already built: its ground, structures, horizon, materials and lighting are the finished truth of this location and must not be redesigned or replaced. No location photograph is attached — read the place from that image alone, and add no scenery, structure, vehicle or fixture that it does not already show. This lock governs the place only; the figures in the shot follow the staging and pose instructions.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: wide shot\n- KEY BACKGROUND ELEMENTS: 중앙 천막 입구 (Its old cloth is held aside as 강용언 steps out) — The entrance and one adjoining side are seen obliquely behind his body; used as Frames the emergence without enlarging the tent cloth into an overpowering foreground object; 강용언의 칼 (Worn at his side as he exits) — Seen obliquely alongside his moving body; used as A proportionate character detail within the full-body silhouette.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Natural late-afternoon light gives 강용언's sun-darkened face and worn surroundings grounded definition, with restrained contrast rather than heroic spotlighting.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): Thin potato porridge boils in the large iron pot below the rock; individual earthenware bowls have been distributed on the cleared mat. 강용언: He emerges from the central tent with a sword at his waist and a sun-darkened face.\n\nTHE HAND THAT IS DOING THIS: the object at the centre of this shot is being held, operated, read, aimed or handed over by 강용언 right now, so 강용언's hand — and as much of the wrist and forearm as the framing reaches — is in the frame, gripping or resting on that object exactly the way its real use requires. Match that hand to 강용언: its size, build, skin, age, grooming, sleeve and anything worn on it belong to that person and to no one else.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 강용언 (한국인 남성, 성인 나이대, 햇볕에 그을린 얼굴, 자연스러운 검은 머리) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.",
   "B": "Create ONE FINAL photorealistic live-action film still of the moment below — early 20th-century Korea, Jeolla region; all people are Korean unless explicitly stated otherwise. TIME OF DAY (lock): late afternoon.\n\nSHOT TEXT (authoritative, Korean): 중앙의 낡은 천막을 젖힌 채 밖으로 걸음을 내디딘 강용언의 당당한 전신.\n\nLOCATION (lock): The exterior entrance of the central rear tent in a temporary lakeside camp, opening toward the communal cooking area. The shot takes place here — the attached LOCATION STRUCTURE PHOTOGRAPH is the single authority for this exact place — its fixed structure and permanent site details are LOCKED to it. No separate location photograph exists for this place. Build everything else strictly from the location text above and the shot text; the layout sketch (when attached) governs framing and placement only, and the shot text governs time of day, lighting and action.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: wide shot\n- KEY BACKGROUND ELEMENTS: 중앙 천막 입구 (Its old cloth is held aside as 강용언 steps out) — The entrance and one adjoining side are seen obliquely behind his body; used as Frames the emergence without enlarging the tent cloth into an overpowering foreground object; 강용언의 칼 (Worn at his side as he exits) — Seen obliquely alongside his moving body; used as A proportionate character detail within the full-body silhouette.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Natural late-afternoon light gives 강용언's sun-darkened face and worn surroundings grounded definition, with restrained contrast rather than heroic spotlighting.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): Thin potato porridge boils in the large iron pot below the rock; individual earthenware bowls have been distributed on the cleared mat. 강용언: He emerges from the central tent with a sword at his waist and a sun-darkened face.\n\nTHE HAND THAT IS DOING THIS: the object at the centre of this shot is being held, operated, read, aimed or handed over by 강용언 right now, so 강용언's hand — and as much of the wrist and forearm as the framing reaches — is in the frame, gripping or resting on that object exactly the way its real use requires. Match that hand to 강용언: its size, build, skin, age, grooming, sleeve and anything worn on it belong to that person and to no one else.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 강용언 (한국인 남성, 성인 나이대, 햇볕에 그을린 얼굴, 자연스러운 검은 머리) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.\n\nCOMPOSITION VARIATION (this candidate only): keep the place, the people and the moment exactly as specified, and keep from the CAMERA & FRAME contract its subject scale, screen placement and key background elements exactly. For this candidate only, its camera angle and height are deliberately left open: choose a camera position distinctly different from the obvious one, as a film director picking a second setup on the same blocking. Stay true to any explicit framing the SHOT TEXT itself dictates."
  },
  "roll_refs": {
   "A": [
    {
     "label": "SHOT BACKGROUND",
     "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/db00517b-165b-4005-a287-f28936b59b95/images/272bd7af-9294-4490-98c6-88a40e9deacc/scene/recipe/S14sh9__bgfirst_bg.png",
     "asset_id": "e4d2d435-263a-4691-bfaa-947714f65db4",
     "role": "bgfirst_bg"
    },
    {
     "label": "LAYOUT SKETCH (people placement only)",
     "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/db00517b-165b-4005-a287-f28936b59b95/images/272bd7af-9294-4490-98c6-88a40e9deacc/conti/conti_S14sh9.png",
     "asset_id": "eb1859c6-e579-43d9-9942-d45d4fe3b7ba",
     "role": "conti_light"
    },
    {
     "label": "CHARACTER REFERENCE — 강용언: the exact person appearing in this shot; match face, hair and build exactly.",
     "path": "<bytes:822588>",
     "asset_id": "1a75090f-1eb9-4224-a132-de7a4b454b73",
     "role": "character_ref"
    },
    {
     "label": "PROP REFERENCE — 칼: the exact object appearing in this shot; match its look, material and wear exactly.",
     "path": "<bytes:780122>",
     "asset_id": "82614014-6091-423c-89ca-1e15d209de72",
     "role": "prop_ref"
    }
   ],
   "B": [
    {
     "label": "LOCATION STRUCTURE PHOTOGRAPH — the confirmed photograph of this exact place and its fixed structure: it is the SINGLE authority for the location, the structure's shape, proportions, materials, colors, openings and every permanent site detail. Never copy its camera framing, time of day or lighting — the shot text is the authority for those.",
     "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/db00517b-165b-4005-a287-f28936b59b95/images/background_chain/seed_bg_lakeside_camp_sel.png",
     "asset_id": "4d7542af-9887-40c1-ad92-e627c929288c",
     "role": "location_seed_bg"
    },
    {
     "label": "CHARACTER REFERENCE — 강용언: the exact person appearing in this shot; match face, hair and build exactly.",
     "path": "<bytes:822588>",
     "asset_id": "1a75090f-1eb9-4224-a132-de7a4b454b73",
     "role": "character_ref"
    },
    {
     "label": "PROP REFERENCE — 칼: the exact object appearing in this shot; match its look, material and wear exactly.",
     "path": "<bytes:780122>",
     "asset_id": "82614014-6091-423c-89ca-1e15d209de72",
     "role": "prop_ref"
    }
   ]
  },
  "shot_run_spend_attempt_count": 1,
  "judge_flip_skipped": "judge_fn_owns_order",
  "readings": [
   {
    "label": "A",
    "direction": "강용언은 화면의 왼쪽 바깥을 응시하며 걷고 있다.",
    "built_space": "화면 오른쪽에 기준 사진과 유사한 천막이 배치되어 있고, 왼쪽에 물가가 있다. 솥은 돌멩이들 위에 놓여 있고, 그릇들은 돗자리 위에 배열되어 있다.",
    "entities": "강용언, 천막, 가마솥, 그릇과 돗자리가 등장한다. 강용언의 복장과 칼은 레퍼런스와 일치한다.",
    "hard_violations": [
     "[gemini-pro] 강용언이 천막을 젖히고 있어야 하나, 그의 두 손은 허리춤과 아래를 향해 있고 천막 안쪽에서 주인이 없는 정체불명의 손 하나가 튀어나와 천막 천을 붙잡고 있음 (발명된 사람/물리적으로 불가능한 신체 및 연출).",
     "[gpt-high] 강용언의 두 손이 각각 몸 옆과 칼에 보이는 동시에 천을 잡는 별도의 손이 출입구에 추가되어 있다. 이는 허용되지 않은 다른 사람의 신체 일부이거나 인물의 세 번째 손에 해당한다."
    ],
    "physics": "강용언의 발은 땅에 안정적으로 닿아 걸음을 내딛는 무게감을 보여주나, 천막을 잡고 있는 안쪽의 손은 연결된 몸이나 지지대 없이 허공에 떠 있는 상태이다."
   },
   {
    "label": "B",
    "direction": "강용언은 카메라 정면을 당당하게 응시하며 천막 밖으로 걸음을 내디디고 있다.",
    "built_space": "중앙 천막이 정면을 향해 있으며, 오른편에 강물이 흐르고 왼편에 다른 천막이 보인다. 가마솥과 불 피운 흔적, 돗자리와 그릇들이 프롬프트의 지시대로 배치되어 있다.",
    "entities": "강용언, 천막, 솥, 그릇과 돗자리 모두 지시된 대로 명확하게 식별되며, 복장과 모자, 차고 있는 칼 역시 레퍼런스의 디테일과 잘 맞는다.",
    "hard_violations": [],
    "physics": "강용언이 오른손으로 천막의 천을 자연스럽게 쥐어 젖히고 있으며, 왼손은 칼을 안정적으로 잡고 있다. 내딛는 발과 몸의 중심 이동이 바닥에 온전히 지지되어 있다."
   }
  ],
  "cross_model_order": {
   "policy": "cross_model_order_v2_slot0_fwd_slot1_rev_parallel2",
   "slots": [
    {
     "model": "gemini-pro",
     "order": "forward",
     "display_to_canonical": {
      "A": "A",
      "B": "B"
     },
     "raw": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 2,
        "verdict_ko": "프롬프트는 강용언이 천막을 젖히며 나오는 모습을 요구했으나, 그의 두 손은 아래로 내려가 있고 천막 안쪽에서 정체불명의 손이 천을 잡고 있는 치명적인 신체/연출 오류가 있습니다."
       },
       {
        "label": "B",
        "score": 9,
        "verdict_ko": "강용언이 직접 오른손으로 천막을 젖히며 걸어 나오는 모습을 완벽하게 구현했으며, 와이드 샷의 구도 안에서 요구된 소품과 배경 요소들을 충실하게 배치했습니다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "강용언은 화면의 왼쪽 바깥을 응시하며 걷고 있다.",
        "built_space": "화면 오른쪽에 기준 사진과 유사한 천막이 배치되어 있고, 왼쪽에 물가가 있다. 솥은 돌멩이들 위에 놓여 있고, 그릇들은 돗자리 위에 배열되어 있다.",
        "entities": "강용언, 천막, 가마솥, 그릇과 돗자리가 등장한다. 강용언의 복장과 칼은 레퍼런스와 일치한다.",
        "hard_violations": [
         "강용언이 천막을 젖히고 있어야 하나, 그의 두 손은 허리춤과 아래를 향해 있고 천막 안쪽에서 주인이 없는 정체불명의 손 하나가 튀어나와 천막 천을 붙잡고 있음 (발명된 사람/물리적으로 불가능한 신체 및 연출)."
        ],
        "physics": "강용언의 발은 땅에 안정적으로 닿아 걸음을 내딛는 무게감을 보여주나, 천막을 잡고 있는 안쪽의 손은 연결된 몸이나 지지대 없이 허공에 떠 있는 상태이다."
       },
       {
        "label": "B",
        "direction": "강용언은 카메라 정면을 당당하게 응시하며 천막 밖으로 걸음을 내디디고 있다.",
        "built_space": "중앙 천막이 정면을 향해 있으며, 오른편에 강물이 흐르고 왼편에 다른 천막이 보인다. 가마솥과 불 피운 흔적, 돗자리와 그릇들이 프롬프트의 지시대로 배치되어 있다.",
        "entities": "강용언, 천막, 솥, 그릇과 돗자리 모두 지시된 대로 명확하게 식별되며, 복장과 모자, 차고 있는 칼 역시 레퍼런스의 디테일과 잘 맞는다.",
        "hard_violations": [],
        "physics": "강용언이 오른손으로 천막의 천을 자연스럽게 쥐어 젖히고 있으며, 왼손은 칼을 안정적으로 잡고 있다. 내딛는 발과 몸의 중심 이동이 바닥에 온전히 지지되어 있다."
       }
      ],
      "all_candidates_fail": false
     },
     "normalized": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 2,
        "verdict_ko": "프롬프트는 강용언이 천막을 젖히며 나오는 모습을 요구했으나, 그의 두 손은 아래로 내려가 있고 천막 안쪽에서 정체불명의 손이 천을 잡고 있는 치명적인 신체/연출 오류가 있습니다."
       },
       {
        "label": "B",
        "score": 9,
        "verdict_ko": "강용언이 직접 오른손으로 천막을 젖히며 걸어 나오는 모습을 완벽하게 구현했으며, 와이드 샷의 구도 안에서 요구된 소품과 배경 요소들을 충실하게 배치했습니다."
       }
      ],
      "all_candidates_fail": false,
      "readings": [
       {
        "label": "A",
        "direction": "강용언은 화면의 왼쪽 바깥을 응시하며 걷고 있다.",
        "built_space": "화면 오른쪽에 기준 사진과 유사한 천막이 배치되어 있고, 왼쪽에 물가가 있다. 솥은 돌멩이들 위에 놓여 있고, 그릇들은 돗자리 위에 배열되어 있다.",
        "entities": "강용언, 천막, 가마솥, 그릇과 돗자리가 등장한다. 강용언의 복장과 칼은 레퍼런스와 일치한다.",
        "hard_violations": [
         "강용언이 천막을 젖히고 있어야 하나, 그의 두 손은 허리춤과 아래를 향해 있고 천막 안쪽에서 주인이 없는 정체불명의 손 하나가 튀어나와 천막 천을 붙잡고 있음 (발명된 사람/물리적으로 불가능한 신체 및 연출)."
        ],
        "physics": "강용언의 발은 땅에 안정적으로 닿아 걸음을 내딛는 무게감을 보여주나, 천막을 잡고 있는 안쪽의 손은 연결된 몸이나 지지대 없이 허공에 떠 있는 상태이다."
       },
       {
        "label": "B",
        "direction": "강용언은 카메라 정면을 당당하게 응시하며 천막 밖으로 걸음을 내디디고 있다.",
        "built_space": "중앙 천막이 정면을 향해 있으며, 오른편에 강물이 흐르고 왼편에 다른 천막이 보인다. 가마솥과 불 피운 흔적, 돗자리와 그릇들이 프롬프트의 지시대로 배치되어 있다.",
        "entities": "강용언, 천막, 솥, 그릇과 돗자리 모두 지시된 대로 명확하게 식별되며, 복장과 모자, 차고 있는 칼 역시 레퍼런스의 디테일과 잘 맞는다.",
        "hard_violations": [],
        "physics": "강용언이 오른손으로 천막의 천을 자연스럽게 쥐어 젖히고 있으며, 왼손은 칼을 안정적으로 잡고 있다. 내딛는 발과 몸의 중심 이동이 바닥에 온전히 지지되어 있다."
       }
      ]
     },
     "ok": true
    },
    {
     "model": "gpt-high",
     "order": "reverse",
     "display_to_canonical": {
      "A": "B",
      "B": "A"
     },
     "raw": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 8,
        "verdict_ko": "강용언이 직접 천을 젖히며 밖으로 내딛는 전신과 늦은 오후를 구현했으나, 입구의 사선 구도가 약하고 칼집의 재질·색이 참조와 다르다."
       },
       {
        "label": "B",
        "score": 2,
        "verdict_ko": "입구와 옆면을 비스듬히 담은 전신 구도는 맞지만, 강용언의 두 손과 별개로 천을 잡는 손이 추가되어 인물 제한과 핵심 동작을 위반한다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "강용언은 입구에서 카메라 쪽 마당으로 걸어 나오며 시선도 거의 정면 바깥을 향한다. 오른손은 화면 왼쪽으로 천을 당겨 통로를 열고, 왼손은 허리의 칼 부근에 있다. 칼집은 왼쪽 허리에서 화면 오른쪽 아래로 향하며 누구를 겨누지 않는다.",
        "built_space": "중앙의 큰 갈대벽·목재 골조 건물 한 채와 왼쪽 뒤의 작은 건물 한 채가 보인다. 주 건물의 출입구는 하나이며 인물 바로 뒤에 놓인다. 입구 면이 거의 정면이고 오른쪽 옆면은 좁게 보여, 요구한 사선 구도는 약하다. 대숲·자갈밭·호숫가 재료는 참조와 통하지만 지붕 비례와 주변 배치는 정확히 일치하지 않는다. 왼쪽 멍석에는 그릇 다섯 개, 오른쪽 돌 화덕에는 큰 솥 하나가 있다. 천은 인물을 압도하는 전경 물체로 확대되지 않았다.",
        "entities": "성인 한국인 남성으로 표현된 인물 한 명만 있으며, 얼굴과 체격, 검은 전립과 장식, 청회색 군복, 붉은 허리띠, 행전과 짚신은 인물 참조에 대체로 맞는다. 얼굴의 그을림은 강하지 않다. 허리에 찬 칼은 보이지만 갈색 칼집과 단순한 장식은 소품 참조의 검은 칼집·금속 장식과 다르다. 낡은 천과 토기 그릇, 철솥은 식별되지만 솥이 덮여 있어 묽은 감자죽과 끓는 상태는 확인할 수 없다. 읽을 수 있는 글자는 없다.",
        "hard_violations": [],
        "physics": "앞으로 내디딘 오른발이 지면에 닿아 몸을 지지하고 왼발은 뒤에서 따라 나오는 보행 순간이다. 오른손이 천 가장자리를 실제로 잡고 있으며 천은 상단 부착부와 손 사이에서 접혀 아래로 늘어진다. 칼은 허리띠에 걸리고 왼손이 칼 부근을 받친다. 솥은 화덕 돌 위에 놓이고 그릇은 멍석에 놓여 있다. 지지 없이 뜬 신체나 물체는 보이지 않는다."
       },
       {
        "label": "B",
        "direction": "강용언은 입구에서 화면 왼쪽 앞의 호숫가 마당으로 걸으며 그 방향을 바라본다. 자신의 오른손은 몸 옆으로 내려가 있고 왼손은 허리의 칼을 잡는다. 그런데 화면 오른쪽의 천은 이 두 손과 별개인 손이 잡아 젖히고 있어, 강용언이 직접 천을 젖힌다는 동작이 성립하지 않는다. 칼집은 허리에서 화면 오른쪽 아래로 향한다.",
        "built_space": "화면 오른쪽 주 건물 한 채와 왼쪽 뒤 작은 건물 한 채가 보인다. 주 건물은 출입구 하나와 인접 벽면이 사선으로 드러나 구도 지시에 가깝다. 인물은 문턱 앞 자갈밭으로 나와 있다. 참조보다 높게 드러난 목재 바닥과 문턱, 주변 건물의 개구부 구성은 차이가 있다. 조리 공간에는 돌에 받친 솥 하나, 멍석 하나와 여러 토기 그릇, 별도의 목재 걸이 하나와 항아리 하나가 보인다. 호수는 전경, 대숲은 후경에 있다.",
        "entities": "주 인물의 성인 남성 외형과 전립, 청회색 옷, 붉은 허리띠, 행전·짚신은 참조와 대체로 맞는다. 그러나 주 인물 외에 출입구 천 가장자리에서 별도의 손이 보인다. 칼은 허리에 있지만 참조의 검은 칼집과 금속 장식보다 갈색의 단순한 형태로 표현됐다. 큰 철솥과 토기 그릇은 있으나 감자죽의 내용물이나 끓는 상태는 뚜렷하지 않다. 읽을 수 있는 글자는 없다.",
        "hard_violations": [
         "강용언의 두 손이 각각 몸 옆과 칼에 보이는 동시에 천을 잡는 별도의 손이 출입구에 추가되어 있다. 이는 허용되지 않은 다른 사람의 신체 일부이거나 인물의 세 번째 손에 해당한다."
        ],
        "physics": "강용언은 앞선 오른발로 지면을 딛고 왼발을 뒤에서 들어 옮기는 자세여서 보행 자체는 가능하다. 칼은 허리와 왼손으로 지지된다. 천은 상단과 추가된 손에 의해 지지되지만, 그 손을 강용언의 팔에 연결할 수 없다. 솥은 돌에, 그릇은 멍석에, 항아리와 목재 걸이는 지면에 놓여 있다. 문제는 부유가 아니라 천을 조작하는 추가 신체다."
       }
      ],
      "all_candidates_fail": false
     },
     "normalized": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 8,
        "verdict_ko": "강용언이 직접 천을 젖히며 밖으로 내딛는 전신과 늦은 오후를 구현했으나, 입구의 사선 구도가 약하고 칼집의 재질·색이 참조와 다르다."
       },
       {
        "label": "A",
        "score": 2,
        "verdict_ko": "입구와 옆면을 비스듬히 담은 전신 구도는 맞지만, 강용언의 두 손과 별개로 천을 잡는 손이 추가되어 인물 제한과 핵심 동작을 위반한다."
       }
      ],
      "all_candidates_fail": false,
      "readings": [
       {
        "label": "B",
        "direction": "강용언은 입구에서 카메라 쪽 마당으로 걸어 나오며 시선도 거의 정면 바깥을 향한다. 오른손은 화면 왼쪽으로 천을 당겨 통로를 열고, 왼손은 허리의 칼 부근에 있다. 칼집은 왼쪽 허리에서 화면 오른쪽 아래로 향하며 누구를 겨누지 않는다.",
        "built_space": "중앙의 큰 갈대벽·목재 골조 건물 한 채와 왼쪽 뒤의 작은 건물 한 채가 보인다. 주 건물의 출입구는 하나이며 인물 바로 뒤에 놓인다. 입구 면이 거의 정면이고 오른쪽 옆면은 좁게 보여, 요구한 사선 구도는 약하다. 대숲·자갈밭·호숫가 재료는 참조와 통하지만 지붕 비례와 주변 배치는 정확히 일치하지 않는다. 왼쪽 멍석에는 그릇 다섯 개, 오른쪽 돌 화덕에는 큰 솥 하나가 있다. 천은 인물을 압도하는 전경 물체로 확대되지 않았다.",
        "entities": "성인 한국인 남성으로 표현된 인물 한 명만 있으며, 얼굴과 체격, 검은 전립과 장식, 청회색 군복, 붉은 허리띠, 행전과 짚신은 인물 참조에 대체로 맞는다. 얼굴의 그을림은 강하지 않다. 허리에 찬 칼은 보이지만 갈색 칼집과 단순한 장식은 소품 참조의 검은 칼집·금속 장식과 다르다. 낡은 천과 토기 그릇, 철솥은 식별되지만 솥이 덮여 있어 묽은 감자죽과 끓는 상태는 확인할 수 없다. 읽을 수 있는 글자는 없다.",
        "hard_violations": [],
        "physics": "앞으로 내디딘 오른발이 지면에 닿아 몸을 지지하고 왼발은 뒤에서 따라 나오는 보행 순간이다. 오른손이 천 가장자리를 실제로 잡고 있으며 천은 상단 부착부와 손 사이에서 접혀 아래로 늘어진다. 칼은 허리띠에 걸리고 왼손이 칼 부근을 받친다. 솥은 화덕 돌 위에 놓이고 그릇은 멍석에 놓여 있다. 지지 없이 뜬 신체나 물체는 보이지 않는다."
       },
       {
        "label": "A",
        "direction": "강용언은 입구에서 화면 왼쪽 앞의 호숫가 마당으로 걸으며 그 방향을 바라본다. 자신의 오른손은 몸 옆으로 내려가 있고 왼손은 허리의 칼을 잡는다. 그런데 화면 오른쪽의 천은 이 두 손과 별개인 손이 잡아 젖히고 있어, 강용언이 직접 천을 젖힌다는 동작이 성립하지 않는다. 칼집은 허리에서 화면 오른쪽 아래로 향한다.",
        "built_space": "화면 오른쪽 주 건물 한 채와 왼쪽 뒤 작은 건물 한 채가 보인다. 주 건물은 출입구 하나와 인접 벽면이 사선으로 드러나 구도 지시에 가깝다. 인물은 문턱 앞 자갈밭으로 나와 있다. 참조보다 높게 드러난 목재 바닥과 문턱, 주변 건물의 개구부 구성은 차이가 있다. 조리 공간에는 돌에 받친 솥 하나, 멍석 하나와 여러 토기 그릇, 별도의 목재 걸이 하나와 항아리 하나가 보인다. 호수는 전경, 대숲은 후경에 있다.",
        "entities": "주 인물의 성인 남성 외형과 전립, 청회색 옷, 붉은 허리띠, 행전·짚신은 참조와 대체로 맞는다. 그러나 주 인물 외에 출입구 천 가장자리에서 별도의 손이 보인다. 칼은 허리에 있지만 참조의 검은 칼집과 금속 장식보다 갈색의 단순한 형태로 표현됐다. 큰 철솥과 토기 그릇은 있으나 감자죽의 내용물이나 끓는 상태는 뚜렷하지 않다. 읽을 수 있는 글자는 없다.",
        "hard_violations": [
         "강용언의 두 손이 각각 몸 옆과 칼에 보이는 동시에 천을 잡는 별도의 손이 출입구에 추가되어 있다. 이는 허용되지 않은 다른 사람의 신체 일부이거나 인물의 세 번째 손에 해당한다."
        ],
        "physics": "강용언은 앞선 오른발로 지면을 딛고 왼발을 뒤에서 들어 옮기는 자세여서 보행 자체는 가능하다. 칼은 허리와 왼손으로 지지된다. 천은 상단과 추가된 손에 의해 지지되지만, 그 손을 강용언의 팔에 연결할 수 없다. 솥은 돌에, 그릇은 멍석에, 항아리와 목재 걸이는 지면에 놓여 있다. 문제는 부유가 아니라 천을 조작하는 추가 신체다."
       }
      ]
     },
     "ok": true
    }
   ],
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "slot_winner_match": true,
   "slot_winner": {
    "gemini-pro": "B",
    "gpt-high": "B"
   },
   "route": "cross_slot_agree"
  },
  "dual": {
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "normalized": {
    "A": 0.472,
    "B": 2.0
   },
   "adjusted": {
    "A": 0.222,
    "B": 2.0
   },
   "violations": {
    "A": [
     "[gemini-pro] 강용언이 천막을 젖히고 있어야 하나, 그의 두 손은 허리춤과 아래를 향해 있고 천막 안쪽에서 주인이 없는 정체불명의 손 하나가 튀어나와 천막 천을 붙잡고 있음 (발명된 사람/물리적으로 불가능한 신체 및 연출).",
     "[gpt-high] 강용언의 두 손이 각각 몸 옆과 칼에 보이는 동시에 천을 잡는 별도의 손이 출입구에 추가되어 있다. 이는 허용되지 않은 다른 사람의 신체 일부이거나 인물의 세 번째 손에 해당한다."
    ]
   },
   "per_model_winner": {
    "gemini-pro": "B",
    "gpt-high": "B"
   },
   "agreed": true
  },
  "totals": {
   "A": 222,
   "B": 2000
  },
  "selected": "B",
  "ranking": [
   "B",
   "A"
  ],
  "verdicts": [
   {
    "label": "A",
    "score": 222,
    "verdict_ko": "프롬프트는 강용언이 천막을 젖히며 나오는 모습을 요구했으나, 그의 두 손은 아래로 내려가 있고 천막 안쪽에서 정체불명의 손이 천을 잡고 있는 치명적인 신체/연출 오류가 있습니다.  ★위반: [gemini-pro] 강용언이 천막을 젖히고 있어야 하나, 그의 두 손은 허리춤과 아래를 향해 있고 천막 안쪽에서 주인이 없는 정체불명의 손 하나가 튀어나와 천막 천을 붙잡고 있음 (발명된 사람/물리적으로 불가능한 신체 및 연출). / [gpt-high] 강용언의 두 손이 각각 몸 옆과 칼에 보이는 동시에 천을 잡는 별도의 손이 출입구에 추가되어 있다. 이는 허용되지 않은 다른 사람의 신체 일부이거나 인물의 세 번째 손에 해당한다."
   },
   {
    "label": "B",
    "score": 2000,
    "verdict_ko": "강용언이 직접 오른손으로 천막을 젖히며 걸어 나오는 모습을 완벽하게 구현했으며, 와이드 샷의 구도 안에서 요구된 소품과 배경 요소들을 충실하게 배치했습니다."
   }
  ],
  "refs": [
   {
    "label": "LOCATION STRUCTURE PHOTOGRAPH — the confirmed photograph of this exact place and its fixed structure: it is the SINGLE authority for the location, the structure's shape, proportions, materials, colors, openings and every permanent site detail. Never copy its camera framing, time of day or lighting — the shot text is the authority for those.",
    "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/db00517b-165b-4005-a287-f28936b59b95/images/background_chain/seed_bg_lakeside_camp_sel.png",
    "asset_id": "4d7542af-9887-40c1-ad92-e627c929288c",
    "role": "location_seed_bg"
   },
   {
    "label": "CHARACTER REFERENCE — 강용언: the exact person appearing in this shot; match face, hair and build exactly.",
    "path": "<bytes:822588>",
    "asset_id": "1a75090f-1eb9-4224-a132-de7a4b454b73",
    "role": "character_ref"
   },
   {
    "label": "PROP REFERENCE — 칼: the exact object appearing in this shot; match its look, material and wear exactly.",
    "path": "<bytes:780122>",
    "asset_id": "82614014-6091-423c-89ca-1e15d209de72",
    "role": "prop_ref"
   }
  ],
  "critique_skipped": true,
  "shot_run_produced": false,
  "shot_run_uid": "06aa140f-2260-7ee8-b43f-36b3210beff3",
  "bgfirst": {
   "bg_path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/db00517b-165b-4005-a287-f28936b59b95/images/272bd7af-9294-4490-98c6-88a40e9deacc/scene/recipe/S14sh9__bgfirst_bg.png",
   "bg_asset_id": "e4d2d435-263a-4691-bfaa-947714f65db4",
   "bg_record_key": "S14sh9::bgfirst_bg",
   "chain_winner": false,
   "authority": "seed_bg"
  },
  "ref_mode": "seed-bg+엔티티 (2택1: 무콘티 승)",
  "share_plan": {
   "ref_plan": "background"
  },
  "lane_policy": "ab_select_ready"
 },
 "S14sh9::cine": {
  "applied": true,
  "attempted_at": "2026-09-09T02:17:40.082704+00:00",
  "fingerprint": "a1ccbab48e4d204a686415872dafeba1f48b6aea870b0ee67137dec09d8382a4",
  "fingerprint_version": 2,
  "provider": "mai",
  "endpoint": "openrouter/chat-completions",
  "model": "microsoft/mai-image-2.6",
  "pack": "24.202608252115",
  "source_file": "S14sh9_sel.png",
  "source_sha256": "cb1cd4224d5dc7de90326a3ab07fccdf357724f2ce4d08349a0615917e14cc93",
  "file": "S14sh9_cine.png",
  "staged_sha256": "20a3b1b2440f9ba0956562459f882ff6a48a3229e5bee7b61353cafa3a2a27cb",
  "latency_ms": 25144
 },
 "S14sh15::signage": {
  "fp": "ffe4ed6f63cbc7eb",
  "inscriptions": [],
  "cues": [],
  "dropped": []
 },
 "groupbg::lake_cooking_circle": {
  "input_fingerprint": "444c370da2ff9997",
  "meta": {
   "model": "gpt-image-2",
   "size": "1536x864",
   "pack": "11.202607220237",
   "contract": "bgfirst_full_v3",
   "group_sig": {
    "key": "lake_cooking_circle",
    "tags": [
     "S14sh15",
     "S15sh1",
     "S15sh12",
     "S15sh8",
     "S18sh1"
    ]
   },
   "context_sig": "4844ca4f302cb993"
  },
  "prompt": "Create ONE empty live-action location background photograph — NO PEOPLE, no figures, no body parts, no silhouettes, no shadows or reflections of people anywhere.\n\"Empty\" means no people only: KEEP the location's inherent occupants and stock that define the place — animals in an animal shelter, pen or farm, goods and produce in a market, moored boats in a harbour — unless the shot text explicitly removes them.\nThe FIRST attached image is a thin-line storyboard sketch of a shot that happens at this location: use it ONLY as spatial evidence — what this place contains, how its ground, structures and landmarks are arranged and proportioned. Ignore the sketched people and arrows entirely, and do NOT copy its line style: render a fully photographic, physically plausible real place that fits THE LOCATION text below.\nHUMAN-SCALE CALIBRATION: derive every structure's true size from human-scale elements — a door ≈ 2m, a window ≈ 1–1.5m wide, one storey ≈ 2.5–3m; never inflate a small structure or shrink a large one.\n\nTHE LOCATION — early 20th-century Korea, Jeolla region; all people are Korean unless explicitly stated otherwise: The outdoor eating place on a woven mat beside the large iron pot below a rock in the lakeside camp.\n\nLOCATION DETAIL (descriptive evidence about this place from the production's location records — evidence, not a staging order): use it to understand what kind of place this is and what it permanently contains. Carry over ONLY its enduring physical features — layout, structures, ground and materials, fixed equipment and installations. Do NOT treat any lighting, weather or time-of-day wording, momentary object states, depicted events or subjective impressions in it as fixed facts of the place; the TIME OF DAY lock below is the sole authority for time and lighting.\n보성 호숫가 강용언 임시거점 자갈밭·공터·막사 앞: 잿빛 호숫가에 펼쳐진 자갈밭과 흙 공터. 대나무 숲을 등지고 있으며, 낡은 천과 거적을 엮어 만든 조잡한 임시 막사 여러 채, 큰 무쇠솥이 얹힌 돌 화덕이 배치되어 있다. (특징: 넓은 잿빛 호수면; 자갈과 흙이 섞인 공터; 낡은 천, 짚, 거적으로 거칠게 엮은 임시 천막/막사; 불이 타오르는 돌 화덕; 끓고 있는 큰 무쇠 가마솥; 쌓여 있는 짐 꾸러미와 곡식 자루)\n\nSCENE EVIDENCE (verbatim quotes from the screenplay about this place — treat them as evidence of what the location physically contains and looks like; stage the PLACE those moments happen in, but do NOT depict the momentary actions, people or staged props themselves):\n- 바위 아래 큰 무쇠솥. 감자를 으깨 넣은 묽은 죽이 끓고 있다.\n- 주변에 다시 웃음. 강용언이 규홍 맞은편에 털썩 앉는다.\n- 규홍과 관회, 재보를 비롯한 의병들이 솥 주위에 둘러앉아 있다.\n- 빈 솥 주변에서 굶주림에 지쳐 널브러져 있던 의병들.\n\nTIME OF DAY (lock): late afternoon.\n\nRender ONE photorealistic empty location photograph, 16:9, neutral enough that every shot of this place can be staged from it later. No readable writing anywhere: surfaces that would carry writing may be present, but stage any wording out of legibility — an oblique angle, distance, shallow focus. No captions, watermarks or overlay text.",
  "effective_prompt": "Create ONE empty live-action location background photograph — NO PEOPLE, no figures, no body parts, no silhouettes, no shadows or reflections of people anywhere.\n\"Empty\" means no people only: KEEP the location's inherent occupants and stock that define the place — animals in an animal shelter, pen or farm, goods and produce in a market, moored boats in a harbour — unless the shot text explicitly removes them.\nThe FIRST attached image is a thin-line storyboard sketch of a shot that happens at this location: use it ONLY as spatial evidence — what this place contains, how its ground, structures and landmarks are arranged and proportioned. Ignore the sketched people and arrows entirely, and do NOT copy its line style: render a fully photographic, physically plausible real place that fits THE LOCATION text below.\nHUMAN-SCALE CALIBRATION: derive every structure's true size from human-scale elements — a door ≈ 2m, a window ≈ 1–1.5m wide, one storey ≈ 2.5–3m; never inflate a small structure or shrink a large one.\n\nTHE LOCATION — early 20th-century Korea, Jeolla region; all people are Korean unless explicitly stated otherwise: The outdoor eating place on a woven mat beside the large iron pot below a rock in the lakeside camp.\n\nLOCATION DETAIL (descriptive evidence about this place from the production's location records — evidence, not a staging order): use it to understand what kind of place this is and what it permanently contains. Carry over ONLY its enduring physical features — layout, structures, ground and materials, fixed equipment and installations. Do NOT treat any lighting, weather or time-of-day wording, momentary object states, depicted events or subjective impressions in it as fixed facts of the place; the TIME OF DAY lock below is the sole authority for time and lighting.\n보성 호숫가 강용언 임시거점 자갈밭·공터·막사 앞: 잿빛 호숫가에 펼쳐진 자갈밭과 흙 공터. 대나무 숲을 등지고 있으며, 낡은 천과 거적을 엮어 만든 조잡한 임시 막사 여러 채, 큰 무쇠솥이 얹힌 돌 화덕이 배치되어 있다. (특징: 넓은 잿빛 호수면; 자갈과 흙이 섞인 공터; 낡은 천, 짚, 거적으로 거칠게 엮은 임시 천막/막사; 불이 타오르는 돌 화덕; 끓고 있는 큰 무쇠 가마솥; 쌓여 있는 짐 꾸러미와 곡식 자루)\n\nSCENE EVIDENCE (verbatim quotes from the screenplay about this place — treat them as evidence of what the location physically contains and looks like; stage the PLACE those moments happen in, but do NOT depict the momentary actions, people or staged props themselves):\n- 바위 아래 큰 무쇠솥. 감자를 으깨 넣은 묽은 죽이 끓고 있다.\n- 주변에 다시 웃음. 강용언이 규홍 맞은편에 털썩 앉는다.\n- 규홍과 관회, 재보를 비롯한 의병들이 솥 주위에 둘러앉아 있다.\n- 빈 솥 주변에서 굶주림에 지쳐 널브러져 있던 의병들.\n\nTIME OF DAY (lock): late afternoon.\n\nRender ONE photorealistic empty location photograph, 16:9, neutral enough that every shot of this place can be staged from it later. No readable writing anywhere: surfaces that would carry writing may be present, but stage any wording out of legibility — an oblique angle, distance, shallow focus. No captions, watermarks or overlay text.",
  "bg_path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/db00517b-165b-4005-a287-f28936b59b95/images/272bd7af-9294-4490-98c6-88a40e9deacc/scene/recipe/groupbg_lake_cooking_circle_fb9493.png",
  "asset_id": "f7b5ab04-6778-439a-b1ab-10ec63e47377",
  "input_asset_ids": [
   "12517b8f-b307-4223-99ed-8ba78794f29b"
  ],
  "origin_tag": "S14sh15",
  "place_text": "The outdoor eating place on a woven mat beside the large iron pot below a rock in the lakeside camp.",
  "origin_inputs": {
   "place_text": "The outdoor eating place on a woven mat beside the large iron pot below a rock in the lakeside camp.",
   "time_of_day_en": "late afternoon",
   "conti_asset_id": "12517b8f-b307-4223-99ed-8ba78794f29b"
  }
 },
 "S14sh15::bgfirst_bg": {
  "input_fingerprint": "f1d1c86407965c98",
  "prompt": "Create the EMPTY BACKGROUND PLATE for one film shot — NO PEOPLE, no figures, no body parts, no sketch lines, no arrows anywhere.\n\"Empty\" means no people only: KEEP the location's inherent occupants and stock that define the place — animals in an animal shelter, pen or farm, goods in a market, moored boats in a harbour — unless the shot text explicitly removes them.\nThe FIRST attached image is a thin-line storyboard sketch: use ONLY its camera angle, horizon, perspective and the placement/size of buildings and set masses — ignore the sketched people and arrows entirely. The SECOND attached image (LOCATION PHOTOGRAPH) is the real place: take its architecture, materials, signage and fixed features, and RE-PROJECT them into the sketch's camera. If the photograph's camera differs from the sketch's, the sketch's camera wins.\nHUMAN-SCALE CALIBRATION: derive every structure's true size from human-scale elements — a door ≈ 2m, a window ≈ 1–1.5m wide, one storey ≈ 2.5–3m; never inflate a small structure or shrink a large one.\n\nSHOT TEXT this background must serve (Korean): 고개를 가볍게 숙인 채 단단한 눈빛으로 강용언을 마주 보는 안규홍의 정면.\n\nLOCATION (lock): The outdoor eating place on a woven mat beside the large iron pot below a rock in the lakeside camp.\n\nTIME OF DAY (lock): late afternoon.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: close-up\n- KEY BACKGROUND ELEMENTS: 임시 막사 (Constructed from old cloth and woven mats) — Only a soft, partial exterior is visible behind 안규홍; used as Provides restrained camp context without competing with his expression.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Late-afternoon daylight retains naturalistic contrast and muted earth tones, preserving warmth in the exchange without softening 안규홍's resolve.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nRender ONE photorealistic empty location photograph, 16:9, that this shot can be staged inside later. No readable writing anywhere: surfaces that would carry writing may be present, but stage any wording out of legibility — an oblique angle, distance, shallow focus. No captions, watermarks or overlay text.",
  "effective_prompt": "Create the EMPTY BACKGROUND PLATE for one film shot — NO PEOPLE, no figures, no body parts, no sketch lines, no arrows anywhere.\n\"Empty\" means no people only: KEEP the location's inherent occupants and stock that define the place — animals in an animal shelter, pen or farm, goods in a market, moored boats in a harbour — unless the shot text explicitly removes them.\nThe FIRST attached image is a thin-line storyboard sketch: use ONLY its camera angle, horizon, perspective and the placement/size of buildings and set masses — ignore the sketched people and arrows entirely. The SECOND attached image (LOCATION PHOTOGRAPH) is the real place: take its architecture, materials, signage and fixed features, and RE-PROJECT them into the sketch's camera. If the photograph's camera differs from the sketch's, the sketch's camera wins.\nHUMAN-SCALE CALIBRATION: derive every structure's true size from human-scale elements — a door ≈ 2m, a window ≈ 1–1.5m wide, one storey ≈ 2.5–3m; never inflate a small structure or shrink a large one.\n\nSHOT TEXT this background must serve (Korean): 고개를 가볍게 숙인 채 단단한 눈빛으로 강용언을 마주 보는 안규홍의 정면.\n\nLOCATION (lock): The outdoor eating place on a woven mat beside the large iron pot below a rock in the lakeside camp.\n\nTIME OF DAY (lock): late afternoon.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: close-up\n- KEY BACKGROUND ELEMENTS: 임시 막사 (Constructed from old cloth and woven mats) — Only a soft, partial exterior is visible behind 안규홍; used as Provides restrained camp context without competing with his expression.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Late-afternoon daylight retains naturalistic contrast and muted earth tones, preserving warmth in the exchange without softening 안규홍's resolve.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nRender ONE photorealistic empty location photograph, 16:9, that this shot can be staged inside later. No readable writing anywhere: surfaces that would carry writing may be present, but stage any wording out of legibility — an oblique angle, distance, shallow focus. No captions, watermarks or overlay text.",
  "bg_path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/db00517b-165b-4005-a287-f28936b59b95/images/272bd7af-9294-4490-98c6-88a40e9deacc/scene/recipe/S14sh15__bgfirst_bg.png",
  "asset_id": "56970c7a-2622-49e3-b157-df2dfde2252a",
  "input_asset_ids": [
   "12517b8f-b307-4223-99ed-8ba78794f29b",
   "f7b5ab04-6778-439a-b1ab-10ec63e47377"
  ]
 },
 "S14sh15": {
  "input_fingerprint": "cde92d8f543ff239",
  "prompt": "Create ONE FINAL photorealistic live-action film still of the moment below — early 20th-century Korea, Jeolla region; all people are Korean unless explicitly stated otherwise. TIME OF DAY (lock): late afternoon.\n\nSHOT TEXT (authoritative, Korean): 고개를 가볍게 숙인 채 단단한 눈빛으로 강용언을 마주 보는 안규홍의 정면.\n\nLOCATION (lock): The outdoor eating place on a woven mat beside the large iron pot below a rock in the lakeside camp. The shot takes place here — the attached LOCATION PHOTOGRAPH shows the exact spot.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: close-up\n- KEY BACKGROUND ELEMENTS: 임시 막사 (Constructed from old cloth and woven mats) — Only a soft, partial exterior is visible behind 안규홍; used as Provides restrained camp context without competing with his expression.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Late-afternoon daylight retains naturalistic contrast and muted earth tones, preserving warmth in the exchange without softening 안규홍's resolve.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The porridge pot, distributed earthenware bowls and cleared mat remain beside the lakeside shelters; packed supplies stay tied. 안규홍: He sits with a partially eaten bowl of porridge and the jade ring concealed in his clothing.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 안규홍 (한국인 남성, 청년 나이대, 검은 머리, 단단히 묶은 머리) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.",
  "roll_prompts": {
   "A": "Stage the shot. The FIRST attached image (SHOT BACKGROUND) is the finished empty background of this shot — keep it EXACTLY: its camera, perspective, architecture, lighting and every fixed feature stay untouched. The SECOND attached image (LAYOUT SKETCH) tells you ONLY where the people go: each sketched person's position, screen size, pose and the gaze/motion arrows. Ignore the sketch's background lines. The CHARACTER REFERENCE photographs show the real people.\nPlace the real people into the background at exactly the sketched positions, sizes and poses, following the arrow directions. No sketch lines or arrows may remain.\n\nCreate ONE FINAL photorealistic live-action film still of the moment below — early 20th-century Korea, Jeolla region; all people are Korean unless explicitly stated otherwise. TIME OF DAY (lock): late afternoon.\n\nSHOT TEXT (authoritative, Korean): 고개를 가볍게 숙인 채 단단한 눈빛으로 강용언을 마주 보는 안규홍의 정면.\n\nLOCATION (lock): The outdoor eating place on a woven mat beside the large iron pot below a rock in the lakeside camp. The shot takes place here — the FIRST attached image (SHOT BACKGROUND) is this exact place, already built: its ground, structures, horizon, materials and lighting are the finished truth of this location and must not be redesigned or replaced. No location photograph is attached — read the place from that image alone, and add no scenery, structure, vehicle or fixture that it does not already show. This lock governs the place only; the figures in the shot follow the staging and pose instructions.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: close-up\n- KEY BACKGROUND ELEMENTS: 임시 막사 (Constructed from old cloth and woven mats) — Only a soft, partial exterior is visible behind 안규홍; used as Provides restrained camp context without competing with his expression.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Late-afternoon daylight retains naturalistic contrast and muted earth tones, preserving warmth in the exchange without softening 안규홍's resolve.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The porridge pot, distributed earthenware bowls and cleared mat remain beside the lakeside shelters; packed supplies stay tied. 안규홍: He sits with a partially eaten bowl of porridge and the jade ring concealed in his clothing.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 안규홍 (한국인 남성, 청년 나이대, 검은 머리, 단단히 묶은 머리) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.",
   "B": "Create ONE FINAL photorealistic live-action film still of the moment below — early 20th-century Korea, Jeolla region; all people are Korean unless explicitly stated otherwise. TIME OF DAY (lock): late afternoon.\n\nSHOT TEXT (authoritative, Korean): 고개를 가볍게 숙인 채 단단한 눈빛으로 강용언을 마주 보는 안규홍의 정면.\n\nLOCATION (lock): The outdoor eating place on a woven mat beside the large iron pot below a rock in the lakeside camp. The shot takes place here — the attached LOCATION PHOTOGRAPH shows the exact spot.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: close-up\n- KEY BACKGROUND ELEMENTS: 임시 막사 (Constructed from old cloth and woven mats) — Only a soft, partial exterior is visible behind 안규홍; used as Provides restrained camp context without competing with his expression.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Late-afternoon daylight retains naturalistic contrast and muted earth tones, preserving warmth in the exchange without softening 안규홍's resolve.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The porridge pot, distributed earthenware bowls and cleared mat remain beside the lakeside shelters; packed supplies stay tied. 안규홍: He sits with a partially eaten bowl of porridge and the jade ring concealed in his clothing.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 안규홍 (한국인 남성, 청년 나이대, 검은 머리, 단단히 묶은 머리) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.\n\nCOMPOSITION VARIATION (this candidate only): keep the place, the people and the moment exactly as specified, and keep from the CAMERA & FRAME contract its subject scale, screen placement and key background elements exactly. For this candidate only, its camera angle and height are deliberately left open: choose a camera position distinctly different from the obvious one, as a film director picking a second setup on the same blocking. Stay true to any explicit framing the SHOT TEXT itself dictates."
  },
  "roll_refs": {
   "A": [
    {
     "label": "SHOT BACKGROUND",
     "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/db00517b-165b-4005-a287-f28936b59b95/images/272bd7af-9294-4490-98c6-88a40e9deacc/scene/recipe/S14sh15__bgfirst_bg.png",
     "asset_id": "56970c7a-2622-49e3-b157-df2dfde2252a",
     "role": "bgfirst_bg"
    },
    {
     "label": "LAYOUT SKETCH (people placement only)",
     "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/db00517b-165b-4005-a287-f28936b59b95/images/272bd7af-9294-4490-98c6-88a40e9deacc/conti/conti_S14sh15.png",
     "asset_id": "12517b8f-b307-4223-99ed-8ba78794f29b",
     "role": "conti_light"
    },
    {
     "label": "CHARACTER REFERENCE — 안규홍: the exact person appearing in this shot; match face, hair and build exactly.",
     "path": "<bytes:1028384>",
     "asset_id": "3b5894db-ce55-4f34-b966-5e725c3aa952",
     "role": "character_ref"
    }
   ],
   "B": [
    {
     "label": "LOCATION PHOTOGRAPH — the exact place of this shot: its architecture, materials, fixed features and lighting mood are spatial truth; stage the moment inside this place. Never copy its camera framing.",
     "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/db00517b-165b-4005-a287-f28936b59b95/images/272bd7af-9294-4490-98c6-88a40e9deacc/scene/recipe/groupbg_lake_cooking_circle_fb9493.png",
     "asset_id": "f7b5ab04-6778-439a-b1ab-10ec63e47377",
     "role": "bgfirst_group_bg"
    },
    {
     "label": "CHARACTER REFERENCE — 안규홍: the exact person appearing in this shot; match face, hair and build exactly.",
     "path": "<bytes:1028384>",
     "asset_id": "3b5894db-ce55-4f34-b966-5e725c3aa952",
     "role": "character_ref"
    }
   ]
  },
  "shot_run_spend_attempt_count": 1,
  "judge_flip_skipped": "judge_fn_owns_order",
  "readings": [
   {
    "label": "A",
    "direction": "카메라 앞 가상의 상대(강용언)를 향해 똑바로 시선을 고정하고 있음.",
    "built_space": "호숫가 지대. 요구된 클로즈업을 무시한 넓은 화각으로 배경의 솥, 막사, 멍석이 모두 뚜렷하게 보임.",
    "entities": "안규홍의 외모는 일치하나, 우측 바위 위에 위치 레퍼런스에 존재하지 않는 임의의 커다란 항아리가 배치됨.",
    "hard_violations": [
     "[gemini-pro] 요구된 클로즈업(close-up) 화각을 무시하고 전신/미디엄 샷으로 렌더링함",
     "[gemini-pro] 위치 레퍼런스에 없는 대형 항아리를 화면 우측에 임의로 생성함"
    ],
    "physics": "멍석 위에 책상다리로 앉아 그릇과 숟가락을 들고 있음."
   },
   {
    "label": "B",
    "direction": "고개를 약간 숙인 채 정면(강용언)을 향해 단단한 눈빛을 보내고 있음.",
    "built_space": "호숫가 캠프. 지시문대로 배경의 임시 막사와 솥이 부드러운 아웃포커싱으로 배치되어 인물에 집중하게 함.",
    "entities": "안규홍의 인상, 헤어스타일, 의상이 레퍼런스와 일치하며 클로즈업으로 잘 표현됨.",
    "hard_violations": [],
    "physics": "화면 밖 바닥에 앉아 있는 안정적인 상체 자세가 확인됨."
   }
  ],
  "cross_model_order": {
   "policy": "cross_model_order_v2_slot0_fwd_slot1_rev_parallel2",
   "slots": [
    {
     "model": "gemini-pro",
     "order": "forward",
     "display_to_canonical": {
      "A": "A",
      "B": "B"
     },
     "raw": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 4,
        "verdict_ko": "클로즈업 프레이밍 지시를 어기고 화각을 넓혔으며, 위치 레퍼런스에 없는 사물(대형 항아리)을 추가하여 감점됨."
       },
       {
        "label": "B",
        "score": 7,
        "verdict_ko": "지정된 클로즈업 화각을 완벽히 준수하고 부드러운 배경 속에서 안규홍의 단단한 눈빛과 표정을 훌륭하게 연출함."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "카메라 앞 가상의 상대(강용언)를 향해 똑바로 시선을 고정하고 있음.",
        "built_space": "호숫가 지대. 요구된 클로즈업을 무시한 넓은 화각으로 배경의 솥, 막사, 멍석이 모두 뚜렷하게 보임.",
        "entities": "안규홍의 외모는 일치하나, 우측 바위 위에 위치 레퍼런스에 존재하지 않는 임의의 커다란 항아리가 배치됨.",
        "hard_violations": [
         "요구된 클로즈업(close-up) 화각을 무시하고 전신/미디엄 샷으로 렌더링함",
         "위치 레퍼런스에 없는 대형 항아리를 화면 우측에 임의로 생성함"
        ],
        "physics": "멍석 위에 책상다리로 앉아 그릇과 숟가락을 들고 있음."
       },
       {
        "label": "B",
        "direction": "고개를 약간 숙인 채 정면(강용언)을 향해 단단한 눈빛을 보내고 있음.",
        "built_space": "호숫가 캠프. 지시문대로 배경의 임시 막사와 솥이 부드러운 아웃포커싱으로 배치되어 인물에 집중하게 함.",
        "entities": "안규홍의 인상, 헤어스타일, 의상이 레퍼런스와 일치하며 클로즈업으로 잘 표현됨.",
        "hard_violations": [],
        "physics": "화면 밖 바닥에 앉아 있는 안정적인 상체 자세가 확인됨."
       }
      ],
      "all_candidates_fail": false
     },
     "normalized": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 4,
        "verdict_ko": "클로즈업 프레이밍 지시를 어기고 화각을 넓혔으며, 위치 레퍼런스에 없는 사물(대형 항아리)을 추가하여 감점됨."
       },
       {
        "label": "B",
        "score": 7,
        "verdict_ko": "지정된 클로즈업 화각을 완벽히 준수하고 부드러운 배경 속에서 안규홍의 단단한 눈빛과 표정을 훌륭하게 연출함."
       }
      ],
      "all_candidates_fail": false,
      "readings": [
       {
        "label": "A",
        "direction": "카메라 앞 가상의 상대(강용언)를 향해 똑바로 시선을 고정하고 있음.",
        "built_space": "호숫가 지대. 요구된 클로즈업을 무시한 넓은 화각으로 배경의 솥, 막사, 멍석이 모두 뚜렷하게 보임.",
        "entities": "안규홍의 외모는 일치하나, 우측 바위 위에 위치 레퍼런스에 존재하지 않는 임의의 커다란 항아리가 배치됨.",
        "hard_violations": [
         "요구된 클로즈업(close-up) 화각을 무시하고 전신/미디엄 샷으로 렌더링함",
         "위치 레퍼런스에 없는 대형 항아리를 화면 우측에 임의로 생성함"
        ],
        "physics": "멍석 위에 책상다리로 앉아 그릇과 숟가락을 들고 있음."
       },
       {
        "label": "B",
        "direction": "고개를 약간 숙인 채 정면(강용언)을 향해 단단한 눈빛을 보내고 있음.",
        "built_space": "호숫가 캠프. 지시문대로 배경의 임시 막사와 솥이 부드러운 아웃포커싱으로 배치되어 인물에 집중하게 함.",
        "entities": "안규홍의 인상, 헤어스타일, 의상이 레퍼런스와 일치하며 클로즈업으로 잘 표현됨.",
        "hard_violations": [],
        "physics": "화면 밖 바닥에 앉아 있는 안정적인 상체 자세가 확인됨."
       }
      ]
     },
     "ok": true
    },
    {
     "model": "gpt-high",
     "order": "reverse",
     "display_to_canonical": {
      "A": "B",
      "B": "A"
     },
     "raw": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 6,
        "verdict_ko": "고개를 조금 숙인 정면과 단단한 눈빛은 더 충실하지만, 허리까지 담아 지정된 클로즈업보다 넓고 막사도 지나치게 많이 보인다."
       },
       {
        "label": "B",
        "score": 3,
        "verdict_ko": "앉은 자세와 죽그릇은 자연스럽지만, 발까지 보이는 전신 구도로 넓혀 핵심인 얼굴 클로즈업과 절제된 부분 배경을 놓쳤다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "얼굴은 정면이고 턱을 조금 내린 채 눈은 카메라 부근의 정면 상대를 향한다. 화면 밖 강용언과 마주 보는 구도로 해석할 수 있으나, 강용언 자체는 보이지 않아 정확한 시선 도착점은 확인할 수 없다. 무기나 방향성 있는 도구는 없다.",
        "built_space": "왼쪽에 호수와 돌 해안, 오른쪽 뒤에 천 지붕과 엮은 벽을 가진 막사군이 보인다. 검은 뚜껑 솥 하나가 바위 사이에 있고, 돗자리 한 장의 일부와 묶인 짐들이 보인다. 참조 장소의 주요 배치는 유지하지만 막사 외부와 주변 풍경을 상당히 넓게 보여 주어, 부드러운 부분 배경만 요구한 구도에는 못 미친다. 인물은 중앙에서 허리까지 담겼다.",
        "entities": "청년 동아시아계 남성 한 명으로 안규홍의 인물 설정과 대체로 맞는다. 검은 머리를 뒤로 당긴 형태, 얼굴 윤곽, 때 묻은 미색 저고리와 허리띠가 참조와 가깝다. 뒤쪽 머리 매듭은 직접 확인되지 않는다. 강용언이나 다른 사람은 추가되지 않았다. 죽그릇과 옥가락지는 화면에 드러나지 않으며, 이 크롭에서 부재를 단정할 수 없다. 읽을 수 있는 글자는 없다.",
        "hard_violations": [],
        "physics": "머리와 목, 어깨와 몸통의 연결이 자연스럽고 팔은 아래쪽 화면 밖으로 이어진다. 골반과 지면 접점이 잘려 앉은 자세의 지지를 직접 확인할 수는 없지만, 공중에 떠 있다는 증거는 없다. 솥은 바위 화덕에 놓여 있고 천과 엮은 벽은 막사 기둥에 지지되어 있다."
       },
       {
        "label": "B",
        "direction": "턱을 조금 숙이고 눈은 정면의 카메라 가까운 상대를 응시한다. 화면 밖 강용언을 바라보는 배치와 양립하지만 상대 위치는 직접 확인되지 않는다. 오른손의 숟가락은 죽그릇 위에서 수평으로 들려 있으며 입으로 향하는 순간은 아니다.",
        "built_space": "인물이 전경 돗자리 위에 가부좌로 앉아 있고 뒤쪽에도 별도의 돗자리 부분이 보인다. 오른쪽 뒤 바위 화덕에는 검은 뚜껑 솥 하나가 있으며, 천막 지붕과 엮은 벽, 지지대, 묶인 짐들이 넓고 비교적 선명하게 드러난다. 오른쪽 바위 위에는 큰 갈색 토기 용기 하나가 추가로 두드러진다. 호숫가 장소는 참조와 가깝지만 얼굴보다 전신과 캠프 설명에 많은 화면을 할애한다.",
        "entities": "청년 동아시아계 남성 한 명이며 검은 머리의 단단한 매듭, 미색의 낡은 저고리와 바지, 허리띠와 짚신이 보인다. 얼굴과 의복은 안규홍 참조에 대체로 부합한다. 손에 든 죽그릇 하나, 앞쪽 돗자리 위 죽그릇 하나, 오른쪽의 큰 토기 용기 하나가 보이며, 마지막 용기는 식사용 그릇보다 훨씬 크다. 옥가락지는 노출되지 않았다. 다른 사람이나 읽을 수 있는 글자는 없다.",
        "hard_violations": [],
        "physics": "골반과 접힌 다리는 돗자리에 지지되어 있어 앉은 자세가 성립한다. 왼손은 죽그릇 아래와 옆을 받치고 오른손은 숟가락 손잡이를 잡는다. 앞쪽 그릇은 돗자리 위에, 큰 토기 용기는 바위 위에 놓여 있으며 솥은 화덕 돌에 지지된다. 지지 없이 떠 있는 신체나 물건은 보이지 않는다."
       }
      ],
      "all_candidates_fail": true
     },
     "normalized": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 6,
        "verdict_ko": "고개를 조금 숙인 정면과 단단한 눈빛은 더 충실하지만, 허리까지 담아 지정된 클로즈업보다 넓고 막사도 지나치게 많이 보인다."
       },
       {
        "label": "A",
        "score": 3,
        "verdict_ko": "앉은 자세와 죽그릇은 자연스럽지만, 발까지 보이는 전신 구도로 넓혀 핵심인 얼굴 클로즈업과 절제된 부분 배경을 놓쳤다."
       }
      ],
      "all_candidates_fail": true,
      "readings": [
       {
        "label": "B",
        "direction": "얼굴은 정면이고 턱을 조금 내린 채 눈은 카메라 부근의 정면 상대를 향한다. 화면 밖 강용언과 마주 보는 구도로 해석할 수 있으나, 강용언 자체는 보이지 않아 정확한 시선 도착점은 확인할 수 없다. 무기나 방향성 있는 도구는 없다.",
        "built_space": "왼쪽에 호수와 돌 해안, 오른쪽 뒤에 천 지붕과 엮은 벽을 가진 막사군이 보인다. 검은 뚜껑 솥 하나가 바위 사이에 있고, 돗자리 한 장의 일부와 묶인 짐들이 보인다. 참조 장소의 주요 배치는 유지하지만 막사 외부와 주변 풍경을 상당히 넓게 보여 주어, 부드러운 부분 배경만 요구한 구도에는 못 미친다. 인물은 중앙에서 허리까지 담겼다.",
        "entities": "청년 동아시아계 남성 한 명으로 안규홍의 인물 설정과 대체로 맞는다. 검은 머리를 뒤로 당긴 형태, 얼굴 윤곽, 때 묻은 미색 저고리와 허리띠가 참조와 가깝다. 뒤쪽 머리 매듭은 직접 확인되지 않는다. 강용언이나 다른 사람은 추가되지 않았다. 죽그릇과 옥가락지는 화면에 드러나지 않으며, 이 크롭에서 부재를 단정할 수 없다. 읽을 수 있는 글자는 없다.",
        "hard_violations": [],
        "physics": "머리와 목, 어깨와 몸통의 연결이 자연스럽고 팔은 아래쪽 화면 밖으로 이어진다. 골반과 지면 접점이 잘려 앉은 자세의 지지를 직접 확인할 수는 없지만, 공중에 떠 있다는 증거는 없다. 솥은 바위 화덕에 놓여 있고 천과 엮은 벽은 막사 기둥에 지지되어 있다."
       },
       {
        "label": "A",
        "direction": "턱을 조금 숙이고 눈은 정면의 카메라 가까운 상대를 응시한다. 화면 밖 강용언을 바라보는 배치와 양립하지만 상대 위치는 직접 확인되지 않는다. 오른손의 숟가락은 죽그릇 위에서 수평으로 들려 있으며 입으로 향하는 순간은 아니다.",
        "built_space": "인물이 전경 돗자리 위에 가부좌로 앉아 있고 뒤쪽에도 별도의 돗자리 부분이 보인다. 오른쪽 뒤 바위 화덕에는 검은 뚜껑 솥 하나가 있으며, 천막 지붕과 엮은 벽, 지지대, 묶인 짐들이 넓고 비교적 선명하게 드러난다. 오른쪽 바위 위에는 큰 갈색 토기 용기 하나가 추가로 두드러진다. 호숫가 장소는 참조와 가깝지만 얼굴보다 전신과 캠프 설명에 많은 화면을 할애한다.",
        "entities": "청년 동아시아계 남성 한 명이며 검은 머리의 단단한 매듭, 미색의 낡은 저고리와 바지, 허리띠와 짚신이 보인다. 얼굴과 의복은 안규홍 참조에 대체로 부합한다. 손에 든 죽그릇 하나, 앞쪽 돗자리 위 죽그릇 하나, 오른쪽의 큰 토기 용기 하나가 보이며, 마지막 용기는 식사용 그릇보다 훨씬 크다. 옥가락지는 노출되지 않았다. 다른 사람이나 읽을 수 있는 글자는 없다.",
        "hard_violations": [],
        "physics": "골반과 접힌 다리는 돗자리에 지지되어 있어 앉은 자세가 성립한다. 왼손은 죽그릇 아래와 옆을 받치고 오른손은 숟가락 손잡이를 잡는다. 앞쪽 그릇은 돗자리 위에, 큰 토기 용기는 바위 위에 놓여 있으며 솥은 화덕 돌에 지지된다. 지지 없이 떠 있는 신체나 물건은 보이지 않는다."
       }
      ]
     },
     "ok": true
    }
   ],
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "slot_winner_match": true,
   "slot_winner": {
    "gemini-pro": "B",
    "gpt-high": "B"
   },
   "route": "cross_slot_agree"
  },
  "dual": {
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "normalized": {
    "A": 1.071,
    "B": 2.0
   },
   "adjusted": {
    "A": 0.821,
    "B": 2.0
   },
   "violations": {
    "A": [
     "[gemini-pro] 요구된 클로즈업(close-up) 화각을 무시하고 전신/미디엄 샷으로 렌더링함",
     "[gemini-pro] 위치 레퍼런스에 없는 대형 항아리를 화면 우측에 임의로 생성함"
    ]
   },
   "per_model_winner": {
    "gemini-pro": "B",
    "gpt-high": "B"
   },
   "agreed": true
  },
  "totals": {
   "A": 821,
   "B": 2000
  },
  "selected": "B",
  "ranking": [
   "B",
   "A"
  ],
  "verdicts": [
   {
    "label": "A",
    "score": 821,
    "verdict_ko": "클로즈업 프레이밍 지시를 어기고 화각을 넓혔으며, 위치 레퍼런스에 없는 사물(대형 항아리)을 추가하여 감점됨.  ★위반: [gemini-pro] 요구된 클로즈업(close-up) 화각을 무시하고 전신/미디엄 샷으로 렌더링함 / [gemini-pro] 위치 레퍼런스에 없는 대형 항아리를 화면 우측에 임의로 생성함"
   },
   {
    "label": "B",
    "score": 2000,
    "verdict_ko": "지정된 클로즈업 화각을 완벽히 준수하고 부드러운 배경 속에서 안규홍의 단단한 눈빛과 표정을 훌륭하게 연출함."
   }
  ],
  "refs": [
   {
    "label": "LOCATION PHOTOGRAPH — the exact place of this shot: its architecture, materials, fixed features and lighting mood are spatial truth; stage the moment inside this place. Never copy its camera framing.",
    "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/db00517b-165b-4005-a287-f28936b59b95/images/272bd7af-9294-4490-98c6-88a40e9deacc/scene/recipe/groupbg_lake_cooking_circle_fb9493.png",
    "asset_id": "f7b5ab04-6778-439a-b1ab-10ec63e47377",
    "role": "bgfirst_group_bg"
   },
   {
    "label": "CHARACTER REFERENCE — 안규홍: the exact person appearing in this shot; match face, hair and build exactly.",
    "path": "<bytes:1028384>",
    "asset_id": "3b5894db-ce55-4f34-b966-5e725c3aa952",
    "role": "character_ref"
   }
  ],
  "critique_skipped": true,
  "shot_run_produced": false,
  "shot_run_uid": "06aa140f-2552-7125-9490-fc57bf567882",
  "bgfirst": {
   "bg_path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/db00517b-165b-4005-a287-f28936b59b95/images/272bd7af-9294-4490-98c6-88a40e9deacc/scene/recipe/S14sh15__bgfirst_bg.png",
   "bg_asset_id": "56970c7a-2622-49e3-b157-df2dfde2252a",
   "bg_record_key": "S14sh15::bgfirst_bg",
   "chain_winner": false,
   "authority": "groupbg",
   "group_key": "lake_cooking_circle",
   "groupbg_asset_id": "f7b5ab04-6778-439a-b1ab-10ec63e47377"
  },
  "ref_mode": "그룹 배경+엔티티 (2택1: 무콘티 승)",
  "share_plan": {
   "ref_plan": "background"
  }
 },
 "S14sh15::cine": {
  "applied": true,
  "attempted_at": "2026-09-09T02:22:24.058047+00:00",
  "fingerprint": "ce419c41bb8a0b7ab13958559b337ab61e46877e98c88f01a46b23fe6bee2e75",
  "fingerprint_version": 2,
  "provider": "mai",
  "endpoint": "openrouter/chat-completions",
  "model": "microsoft/mai-image-2.6",
  "pack": "24.202608252115",
  "source_file": "S14sh15_sel.png",
  "source_sha256": "3db0655e03dcb48f14df181e26e039fddcfe86a4420912a4aae0b1cc836c4411",
  "file": "S14sh15_cine.png",
  "staged_sha256": "c41d2dbb18ccf0cb65aab432a96367cf73a4c3498636bb734f71a23087744ecc",
  "latency_ms": 34273
 },
 "S15sh1::signage": {
  "fp": "649fb39a325c35bb",
  "inscriptions": [],
  "cues": [],
  "dropped": []
 },
 "S15sh1::bgfirst_bg": {
  "input_fingerprint": "50e931aefa09bb24",
  "prompt": "Create the EMPTY BACKGROUND PLATE for one film shot — NO PEOPLE, no figures, no body parts, no sketch lines, no arrows anywhere.\n\"Empty\" means no people only: KEEP the location's inherent occupants and stock that define the place — animals in an animal shelter, pen or farm, goods in a market, moored boats in a harbour — unless the shot text explicitly removes them.\nThe FIRST attached image is a thin-line storyboard sketch: use ONLY its camera angle, horizon, perspective and the placement/size of buildings and set masses — ignore the sketched people and arrows entirely. The SECOND attached image (LOCATION PHOTOGRAPH) is the real place: take its architecture, materials, signage and fixed features, and RE-PROJECT them into the sketch's camera. If the photograph's camera differs from the sketch's, the sketch's camera wins.\nHUMAN-SCALE CALIBRATION: derive every structure's true size from human-scale elements — a door ≈ 2m, a window ≈ 1–1.5m wide, one storey ≈ 2.5–3m; never inflate a small structure or shrink a large one.\n\nSHOT TEXT this background must serve (Korean): 붉은 모닥불이 타오르는 어두운 호숫가 옆, 무쇠솥을 둥글게 둘러싸고 앉은 조선인 의병들의 왁자지껄한 전경.\n\nLOCATION (lock): The outdoor communal eating circle around a large iron pot at the lakeside camp, lit by red campfires after dark.\n\nTIME OF DAY (lock): night.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: wide shot\n- KEY BACKGROUND ELEMENTS: 큰 무쇠솥 (Chicken and potato broth is boiling inside) — The open top and near side are visible from the elevated camera; used as Organizes the seated circle without becoming oversized; 모닥불 (Firewood is burning beside the gathering); used as Provides visible points of activity around the circle; 호수 (Visible beyond the lakeside gathering); used as Leaves a broad environmental margin above the clustered diners.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Red-burning campfires illuminate faces against the descending night, balancing convivial warmth with subdued saturation and naturally deep surrounding darkness.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nRender ONE photorealistic empty location photograph, 16:9, that this shot can be staged inside later. No readable writing anywhere: surfaces that would carry writing may be present, but stage any wording out of legibility — an oblique angle, distance, shallow focus. No captions, watermarks or overlay text.",
  "effective_prompt": "Create the EMPTY BACKGROUND PLATE for one film shot — NO PEOPLE, no figures, no body parts, no sketch lines, no arrows anywhere.\n\"Empty\" means no people only: KEEP the location's inherent occupants and stock that define the place — animals in an animal shelter, pen or farm, goods in a market, moored boats in a harbour — unless the shot text explicitly removes them.\nThe FIRST attached image is a thin-line storyboard sketch: use ONLY its camera angle, horizon, perspective and the placement/size of buildings and set masses — ignore the sketched people and arrows entirely. The SECOND attached image (LOCATION PHOTOGRAPH) is the real place: take its architecture, materials, signage and fixed features, and RE-PROJECT them into the sketch's camera. If the photograph's camera differs from the sketch's, the sketch's camera wins.\nHUMAN-SCALE CALIBRATION: derive every structure's true size from human-scale elements — a door ≈ 2m, a window ≈ 1–1.5m wide, one storey ≈ 2.5–3m; never inflate a small structure or shrink a large one.\n\nSHOT TEXT this background must serve (Korean): 붉은 모닥불이 타오르는 어두운 호숫가 옆, 무쇠솥을 둥글게 둘러싸고 앉은 조선인 의병들의 왁자지껄한 전경.\n\nLOCATION (lock): The outdoor communal eating circle around a large iron pot at the lakeside camp, lit by red campfires after dark.\n\nTIME OF DAY (lock): night.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: wide shot\n- KEY BACKGROUND ELEMENTS: 큰 무쇠솥 (Chicken and potato broth is boiling inside) — The open top and near side are visible from the elevated camera; used as Organizes the seated circle without becoming oversized; 모닥불 (Firewood is burning beside the gathering); used as Provides visible points of activity around the circle; 호수 (Visible beyond the lakeside gathering); used as Leaves a broad environmental margin above the clustered diners.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Red-burning campfires illuminate faces against the descending night, balancing convivial warmth with subdued saturation and naturally deep surrounding darkness.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nRender ONE photorealistic empty location photograph, 16:9, that this shot can be staged inside later. No readable writing anywhere: surfaces that would carry writing may be present, but stage any wording out of legibility — an oblique angle, distance, shallow focus. No captions, watermarks or overlay text.",
  "bg_path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/db00517b-165b-4005-a287-f28936b59b95/images/272bd7af-9294-4490-98c6-88a40e9deacc/scene/recipe/S15sh1__bgfirst_bg.png",
  "asset_id": "7f647561-3d51-485b-adc7-5e7d3269fe0a",
  "input_asset_ids": [
   "aa28beca-dd5c-4a15-a739-4a4397077ace",
   "f7b5ab04-6778-439a-b1ab-10ec63e47377"
  ]
 },
 "S15sh1": {
  "input_fingerprint": "a266ff5b3205e459",
  "prompt": "Create ONE FINAL photorealistic live-action film still of the moment below — early 20th-century Korea, Jeolla region; all people are Korean unless explicitly stated otherwise. TIME OF DAY (lock): night.\n\nSHOT TEXT (authoritative, Korean): 붉은 모닥불이 타오르는 어두운 호숫가 옆, 무쇠솥을 둥글게 둘러싸고 앉은 조선인 의병들의 왁자지껄한 전경.\n\nLOCATION (lock): The outdoor communal eating circle around a large iron pot at the lakeside camp, lit by red campfires after dark. The shot takes place here — the attached LOCATION PHOTOGRAPH shows the exact spot.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: wide shot\n- KEY BACKGROUND ELEMENTS: 큰 무쇠솥 (Chicken and potato broth is boiling inside) — The open top and near side are visible from the elevated camera; used as Organizes the seated circle without becoming oversized; 모닥불 (Firewood is burning beside the gathering); used as Provides visible points of activity around the circle; 호수 (Visible beyond the lakeside gathering); used as Leaves a broad environmental margin above the clustered diners.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Red-burning campfires illuminate faces against the descending night, balancing convivial warmth with subdued saturation and naturally deep surrounding darkness.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): Campfires light the dark lakeshore; the iron pot now boils broth containing two chickens and potatoes rather than porridge.\n\nNo people appear unless the shot text itself says so.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.",
  "roll_prompts": {
   "A": "Stage the shot. The FIRST attached image (SHOT BACKGROUND) is the finished empty background of this shot — keep it EXACTLY: its camera, perspective, architecture, lighting and every fixed feature stay untouched. The SECOND attached image (LAYOUT SKETCH) tells you ONLY where the people go: each sketched person's position, screen size, pose and the gaze/motion arrows. Ignore the sketch's background lines. The CHARACTER REFERENCE photographs show the real people.\nPlace the real people into the background at exactly the sketched positions, sizes and poses, following the arrow directions. No sketch lines or arrows may remain.\n\nCreate ONE FINAL photorealistic live-action film still of the moment below — early 20th-century Korea, Jeolla region; all people are Korean unless explicitly stated otherwise. TIME OF DAY (lock): night.\n\nSHOT TEXT (authoritative, Korean): 붉은 모닥불이 타오르는 어두운 호숫가 옆, 무쇠솥을 둥글게 둘러싸고 앉은 조선인 의병들의 왁자지껄한 전경.\n\nLOCATION (lock): The outdoor communal eating circle around a large iron pot at the lakeside camp, lit by red campfires after dark. The shot takes place here — the FIRST attached image (SHOT BACKGROUND) is this exact place, already built: its ground, structures, horizon, materials and lighting are the finished truth of this location and must not be redesigned or replaced. No location photograph is attached — read the place from that image alone, and add no scenery, structure, vehicle or fixture that it does not already show. This lock governs the place only; the figures in the shot follow the staging and pose instructions.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: wide shot\n- KEY BACKGROUND ELEMENTS: 큰 무쇠솥 (Chicken and potato broth is boiling inside) — The open top and near side are visible from the elevated camera; used as Organizes the seated circle without becoming oversized; 모닥불 (Firewood is burning beside the gathering); used as Provides visible points of activity around the circle; 호수 (Visible beyond the lakeside gathering); used as Leaves a broad environmental margin above the clustered diners.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Red-burning campfires illuminate faces against the descending night, balancing convivial warmth with subdued saturation and naturally deep surrounding darkness.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): Campfires light the dark lakeshore; the iron pot now boils broth containing two chickens and potatoes rather than porridge.\n\nNo people appear unless the shot text itself says so.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.",
   "B": "Create ONE FINAL photorealistic live-action film still of the moment below — early 20th-century Korea, Jeolla region; all people are Korean unless explicitly stated otherwise. TIME OF DAY (lock): night.\n\nSHOT TEXT (authoritative, Korean): 붉은 모닥불이 타오르는 어두운 호숫가 옆, 무쇠솥을 둥글게 둘러싸고 앉은 조선인 의병들의 왁자지껄한 전경.\n\nLOCATION (lock): The outdoor communal eating circle around a large iron pot at the lakeside camp, lit by red campfires after dark. The shot takes place here — the attached LOCATION PHOTOGRAPH shows the exact spot.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: wide shot\n- KEY BACKGROUND ELEMENTS: 큰 무쇠솥 (Chicken and potato broth is boiling inside) — The open top and near side are visible from the elevated camera; used as Organizes the seated circle without becoming oversized; 모닥불 (Firewood is burning beside the gathering); used as Provides visible points of activity around the circle; 호수 (Visible beyond the lakeside gathering); used as Leaves a broad environmental margin above the clustered diners.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Red-burning campfires illuminate faces against the descending night, balancing convivial warmth with subdued saturation and naturally deep surrounding darkness.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): Campfires light the dark lakeshore; the iron pot now boils broth containing two chickens and potatoes rather than porridge.\n\nNo people appear unless the shot text itself says so.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.\n\nCOMPOSITION VARIATION (this candidate only): keep the place, the people and the moment exactly as specified, and keep from the CAMERA & FRAME contract its subject scale, screen placement and key background elements exactly. For this candidate only, its camera angle and height are deliberately left open: choose a camera position distinctly different from the obvious one, as a film director picking a second setup on the same blocking. Stay true to any explicit framing the SHOT TEXT itself dictates."
  },
  "roll_refs": {
   "A": [
    {
     "label": "SHOT BACKGROUND",
     "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/db00517b-165b-4005-a287-f28936b59b95/images/272bd7af-9294-4490-98c6-88a40e9deacc/scene/recipe/S15sh1__bgfirst_bg.png",
     "asset_id": "7f647561-3d51-485b-adc7-5e7d3269fe0a",
     "role": "bgfirst_bg"
    },
    {
     "label": "LAYOUT SKETCH (people placement only)",
     "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/db00517b-165b-4005-a287-f28936b59b95/images/272bd7af-9294-4490-98c6-88a40e9deacc/conti/conti_S15sh1.png",
     "asset_id": "aa28beca-dd5c-4a15-a739-4a4397077ace",
     "role": "conti_light"
    },
    {
     "label": "PROP REFERENCE — 무쇠솥: the exact object appearing in this shot; match its look, material and wear exactly.",
     "path": "<bytes:1131910>",
     "asset_id": "8510d902-c12d-45d6-9f5c-8a90c9a902b3",
     "role": "prop_ref"
    }
   ],
   "B": [
    {
     "label": "LOCATION PHOTOGRAPH — the exact place of this shot: its architecture, materials, fixed features and lighting mood are spatial truth; stage the moment inside this place. Never copy its camera framing.",
     "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/db00517b-165b-4005-a287-f28936b59b95/images/272bd7af-9294-4490-98c6-88a40e9deacc/scene/recipe/groupbg_lake_cooking_circle_fb9493.png",
     "asset_id": "f7b5ab04-6778-439a-b1ab-10ec63e47377",
     "role": "bgfirst_group_bg"
    },
    {
     "label": "PROP REFERENCE — 무쇠솥: the exact object appearing in this shot; match its look, material and wear exactly.",
     "path": "<bytes:1131910>",
     "asset_id": "8510d902-c12d-45d6-9f5c-8a90c9a902b3",
     "role": "prop_ref"
    }
   ]
  },
  "shot_run_spend_attempt_count": 1,
  "judge_flip_skipped": "judge_fn_owns_order",
  "readings": [
   {
    "label": "A",
    "direction": "인물들의 시선이 각자의 그릇, 솥 내부, 혹은 서로를 향해 자연스럽게 교차하고 있으며, 우측의 한 남성이 솥을 향해 국자를 뻗고 있음.",
    "built_space": "호숫가 야영지. 좌측에는 호수가 펼쳐져 있고, 우측에는 레퍼런스 이미지에 있는 대나무 숲과 천막 구조물이 올바르게 배치됨. 중앙의 무쇠솥은 레퍼런스처럼 돌무더기 위에 안정적으로 거치되어 있음.",
    "entities": "전통 복장의 조선인들(남성 및 여성 1명 포함), 모닥불, 돌 위에 놓인 무쇠솥. 솥 안에는 감자와 함께 지침(2마리)보다 많은 수의 닭이 끓고 있음.",
    "hard_violations": [],
    "physics": "모든 인물이 바닥에 안정적으로 앉아 있으며, 각자가 들고 있는 그릇이나 국자는 손에 의해 올바르게 지탱되고 있음."
   },
   {
    "label": "B",
    "direction": "대다수의 인물들이 국자를 들고 있는 중앙의 남성이나 무쇠솥 쪽으로 시선을 집중하고 있음.",
    "built_space": "호숫가 야영지이나, 레퍼런스에 존재하던 우측의 천막 구조물이 누락되었고 호수가 배경 전체를 차지하도록 지형이 변형됨. 무쇠솥 위로 임의의 나무 삼각대 구조물이 설치되어 있음.",
    "entities": "전통 복장의 조선인 남성들, 모닥불, 무쇠솥. 무쇠솥에는 레퍼런스에 없는 금속 손잡이가 달려 삼각대에 연결되어 있으며 솥뚜껑이 바닥에 놓여 있음. 솥 안에는 닭과 감자가 보임.",
    "hard_violations": [
     "[gemini-pro] 로케이션 레퍼런스에 없는 삼각대 구조물 임의 추가",
     "[gemini-pro] 로케이션 레퍼런스의 주요 지형 및 구조물(우측 천막) 누락 및 변경"
    ],
    "physics": "인물들은 돗자리나 바닥에 앉아 있고, 솥은 바닥의 돌과 삼각대에 의해 지지되고 있으며 국자와 그릇은 인물들의 손에 들려 있음."
   }
  ],
  "cross_model_order": {
   "policy": "cross_model_order_v2_slot0_fwd_slot1_rev_parallel2",
   "slots": [
    {
     "model": "gemini-pro",
     "order": "forward",
     "display_to_canonical": {
      "A": "A",
      "B": "B"
     },
     "raw": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 8,
        "verdict_ko": "지정된 로케이션의 지형(좌측 호수, 우측 천막 구조물 등)과 무쇠솥이 돌 위에 놓인 방식을 정확히 구현했으나, 솥 안에 묘사된 닭의 마릿수가 지침(두 마리)을 초과한 점이 아쉽습니다."
       },
       {
        "label": "B",
        "score": 4,
        "verdict_ko": "로케이션 레퍼런스의 공간 구조(우측 천막 부재, 배경 전체가 호수로 변경됨)를 무시하였으며, 무쇠솥 위에 임의로 삼각대 거치대를 추가하여 공간 및 소품 일치도에서 크게 감점되었습니다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "인물들의 시선이 각자의 그릇, 솥 내부, 혹은 서로를 향해 자연스럽게 교차하고 있으며, 우측의 한 남성이 솥을 향해 국자를 뻗고 있음.",
        "built_space": "호숫가 야영지. 좌측에는 호수가 펼쳐져 있고, 우측에는 레퍼런스 이미지에 있는 대나무 숲과 천막 구조물이 올바르게 배치됨. 중앙의 무쇠솥은 레퍼런스처럼 돌무더기 위에 안정적으로 거치되어 있음.",
        "entities": "전통 복장의 조선인들(남성 및 여성 1명 포함), 모닥불, 돌 위에 놓인 무쇠솥. 솥 안에는 감자와 함께 지침(2마리)보다 많은 수의 닭이 끓고 있음.",
        "hard_violations": [],
        "physics": "모든 인물이 바닥에 안정적으로 앉아 있으며, 각자가 들고 있는 그릇이나 국자는 손에 의해 올바르게 지탱되고 있음."
       },
       {
        "label": "B",
        "direction": "대다수의 인물들이 국자를 들고 있는 중앙의 남성이나 무쇠솥 쪽으로 시선을 집중하고 있음.",
        "built_space": "호숫가 야영지이나, 레퍼런스에 존재하던 우측의 천막 구조물이 누락되었고 호수가 배경 전체를 차지하도록 지형이 변형됨. 무쇠솥 위로 임의의 나무 삼각대 구조물이 설치되어 있음.",
        "entities": "전통 복장의 조선인 남성들, 모닥불, 무쇠솥. 무쇠솥에는 레퍼런스에 없는 금속 손잡이가 달려 삼각대에 연결되어 있으며 솥뚜껑이 바닥에 놓여 있음. 솥 안에는 닭과 감자가 보임.",
        "hard_violations": [
         "로케이션 레퍼런스에 없는 삼각대 구조물 임의 추가",
         "로케이션 레퍼런스의 주요 지형 및 구조물(우측 천막) 누락 및 변경"
        ],
        "physics": "인물들은 돗자리나 바닥에 앉아 있고, 솥은 바닥의 돌과 삼각대에 의해 지지되고 있으며 국자와 그릇은 인물들의 손에 들려 있음."
       }
      ],
      "all_candidates_fail": false
     },
     "normalized": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 8,
        "verdict_ko": "지정된 로케이션의 지형(좌측 호수, 우측 천막 구조물 등)과 무쇠솥이 돌 위에 놓인 방식을 정확히 구현했으나, 솥 안에 묘사된 닭의 마릿수가 지침(두 마리)을 초과한 점이 아쉽습니다."
       },
       {
        "label": "B",
        "score": 4,
        "verdict_ko": "로케이션 레퍼런스의 공간 구조(우측 천막 부재, 배경 전체가 호수로 변경됨)를 무시하였으며, 무쇠솥 위에 임의로 삼각대 거치대를 추가하여 공간 및 소품 일치도에서 크게 감점되었습니다."
       }
      ],
      "all_candidates_fail": false,
      "readings": [
       {
        "label": "A",
        "direction": "인물들의 시선이 각자의 그릇, 솥 내부, 혹은 서로를 향해 자연스럽게 교차하고 있으며, 우측의 한 남성이 솥을 향해 국자를 뻗고 있음.",
        "built_space": "호숫가 야영지. 좌측에는 호수가 펼쳐져 있고, 우측에는 레퍼런스 이미지에 있는 대나무 숲과 천막 구조물이 올바르게 배치됨. 중앙의 무쇠솥은 레퍼런스처럼 돌무더기 위에 안정적으로 거치되어 있음.",
        "entities": "전통 복장의 조선인들(남성 및 여성 1명 포함), 모닥불, 돌 위에 놓인 무쇠솥. 솥 안에는 감자와 함께 지침(2마리)보다 많은 수의 닭이 끓고 있음.",
        "hard_violations": [],
        "physics": "모든 인물이 바닥에 안정적으로 앉아 있으며, 각자가 들고 있는 그릇이나 국자는 손에 의해 올바르게 지탱되고 있음."
       },
       {
        "label": "B",
        "direction": "대다수의 인물들이 국자를 들고 있는 중앙의 남성이나 무쇠솥 쪽으로 시선을 집중하고 있음.",
        "built_space": "호숫가 야영지이나, 레퍼런스에 존재하던 우측의 천막 구조물이 누락되었고 호수가 배경 전체를 차지하도록 지형이 변형됨. 무쇠솥 위로 임의의 나무 삼각대 구조물이 설치되어 있음.",
        "entities": "전통 복장의 조선인 남성들, 모닥불, 무쇠솥. 무쇠솥에는 레퍼런스에 없는 금속 손잡이가 달려 삼각대에 연결되어 있으며 솥뚜껑이 바닥에 놓여 있음. 솥 안에는 닭과 감자가 보임.",
        "hard_violations": [
         "로케이션 레퍼런스에 없는 삼각대 구조물 임의 추가",
         "로케이션 레퍼런스의 주요 지형 및 구조물(우측 천막) 누락 및 변경"
        ],
        "physics": "인물들은 돗자리나 바닥에 앉아 있고, 솥은 바닥의 돌과 삼각대에 의해 지지되고 있으며 국자와 그릇은 인물들의 손에 들려 있음."
       }
      ]
     },
     "ok": true
    },
    {
     "model": "gpt-high",
     "order": "reverse",
     "display_to_canonical": {
      "A": "B",
      "B": "A"
     },
     "raw": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 6,
        "verdict_ko": "웃고 떠드는 의병들의 원형 모임은 잘 표현했지만, 인물이 전경을 빽빽하게 채우고 참조 장소의 천막과 생활 공간이 사라져 지정된 전경과 장소 재현에서 밀립니다."
       },
       {
        "label": "B",
        "score": 8,
        "verdict_ko": "솥을 둘러싼 모임과 호수의 여백, 오른쪽 천막·대숲을 함께 담아 구도와 장소에 더 충실하지만, 솥이 다소 크고 왁자지껄한 분위기는 약합니다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "앞쪽 인물들은 중앙의 솥과 국자를 든 남자 또는 맞은편 동료를 바라보며 웃는다. 국자는 손에 들려 솥 바로 위에 있고, 다른 손의 그릇은 위로 열려 있다. 주변 인물들의 숟가락과 젓가락도 식기 쪽을 향한다. 뒤쪽에는 중앙 모임이 아니라 별도 모닥불 주변을 바라보는 작은 무리들이 있다.",
        "built_space": "왼쪽과 상단에 호수와 산이 펼쳐지고 오른쪽에 대숲이 있다. 중앙 솥 아래 불 한 곳과 주변 모닥불 세 곳이 보인다. 중앙에는 돌 화덕, 삼각대 하나와 매단 사슬, 솥 옆에 세운 뚜껑 하나가 있다. 인물들은 돗자리와 자갈 바닥에 여러 겹으로 둘러앉아 있다. 참조 장소를 특징짓는 오른쪽 천막, 갈대 가림막과 짐 꾸러미는 보이지 않는다. 약간 높은 시점에서 솥의 내부와 앞면은 보이지만, 가까운 인물들이 하단에서 잘리고 모임이 화면을 조밀하게 채운다.",
        "entities": "수십 명의 인물이 있으며, 대체로 성인부터 중장년의 한국인 남성으로 읽히는 외모에 상투·머리띠·갓과 회갈색 전통 복장을 갖췄다. 특정 인물의 신원은 지정되지 않았으며, 의병이라는 신분 자체를 확정할 표지는 뚜렷하지 않다. 검은 솥 하나에 닭 형태와 감자, 국물과 김이 보이지만 닭 두 마리를 명확히 분리해 세기는 어렵다. 솥은 참조와 같은 검은 금속 재질이지만 테두리와 매다는 장치의 형태가 다르다. 붉은 주황색 장작불, 호수와 야간 환경은 있다. 읽을 수 있는 글자는 없다.",
        "hard_violations": [],
        "physics": "앉은 인물들의 엉덩이와 접힌 다리는 땅이나 돗자리에 닿아 있고, 국자와 그릇은 손으로 지지된다. 솥은 돌 화덕 위에 놓이고 삼각대의 사슬도 솥의 손잡이 쪽으로 이어진다. 뚜껑은 바닥과 화덕 돌에 기대어 있다. 장작은 땅에 놓여 타며 열기와 김이 위로 오른다. 지지 없이 떠 있는 인물이나 물체는 보이지 않는다."
       },
       {
        "label": "B",
        "direction": "둘러앉은 인물들은 자기 그릇, 중앙 솥 또는 옆 사람을 바라본다. 솥 오른쪽 남자는 시선을 국물에 두고 국자를 솥 안으로 뻗으며, 다른 인물들은 숟가락을 그릇이나 입 쪽으로 향한다. 식사 동작의 방향은 자연스럽지만, 여러 사람이 서로 크게 웃고 떠드는 장면보다는 차분히 먹는 순간으로 읽힌다.",
        "built_space": "성인 인물 13명이 중앙 솥 하나를 중심으로 원형으로 앉아 있다. 왼쪽에는 호수와 바위 해안, 상단에는 산과 밤하늘의 여백이 있고 오른쪽에는 대숲과 천막 시설이 이어진다. 천막의 기둥, 천 지붕, 세운 갈대 가림막과 묶은 짐이 참조 장소의 구성을 상당히 유지한다. 모임 양옆에 모닥불 두 곳이 있으며, 솥은 중앙의 돌 받침 위에 놓인다. 높은 시점의 와이드숏으로 솥의 열린 윗면과 가까운 측면이 모두 보인다. 다만 솥의 지름은 주변 사람들에 비해 상당히 커서 참조보다 과장되어 보인다.",
        "entities": "인물들은 한국인 성인 남성으로 읽히는 외모에 상투와 소박한 회색·베이지색 전통 복장을 갖췄으며, 일부는 중년으로 보인다. 의병 신분을 확정할 무장은 보이지 않지만 시대 복장은 부합한다. 검은 무쇠솥 하나에 두 마리로 구분되는 닭 형태와 여러 감자, 국물과 김이 보인다. 솥의 양옆 고리 손잡이와 둥근 몸체는 참조 소품에 가깝지만, 크기와 몸통의 턱 형태는 완전히 같지 않다. 장작 모닥불, 호수, 천막과 대숲이 있고 밤의 어둠도 유지된다. 읽을 수 있는 글자는 없다.",
        "hard_violations": [],
        "physics": "인물들은 흙바닥이나 깔개에 엉덩이와 접은 다리를 대고 앉아 있으며, 식기와 국자는 손으로 잡고 있다. 솥의 무게는 둘레의 돌들이 받치고, 통과 바구니 및 장작도 지면에 놓여 있다. 천막은 기둥과 가로대가 지지한다. 솥에서 김은 올라오지만 솥 바로 아래의 불꽃은 보이지 않아 가열원은 직접 확인되지 않는다. 이것만으로 물리적 불가능을 단정할 수는 없으며, 지지 없이 떠 있는 물체나 인물은 없다."
       }
      ],
      "all_candidates_fail": false
     },
     "normalized": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 6,
        "verdict_ko": "웃고 떠드는 의병들의 원형 모임은 잘 표현했지만, 인물이 전경을 빽빽하게 채우고 참조 장소의 천막과 생활 공간이 사라져 지정된 전경과 장소 재현에서 밀립니다."
       },
       {
        "label": "A",
        "score": 8,
        "verdict_ko": "솥을 둘러싼 모임과 호수의 여백, 오른쪽 천막·대숲을 함께 담아 구도와 장소에 더 충실하지만, 솥이 다소 크고 왁자지껄한 분위기는 약합니다."
       }
      ],
      "all_candidates_fail": false,
      "readings": [
       {
        "label": "B",
        "direction": "앞쪽 인물들은 중앙의 솥과 국자를 든 남자 또는 맞은편 동료를 바라보며 웃는다. 국자는 손에 들려 솥 바로 위에 있고, 다른 손의 그릇은 위로 열려 있다. 주변 인물들의 숟가락과 젓가락도 식기 쪽을 향한다. 뒤쪽에는 중앙 모임이 아니라 별도 모닥불 주변을 바라보는 작은 무리들이 있다.",
        "built_space": "왼쪽과 상단에 호수와 산이 펼쳐지고 오른쪽에 대숲이 있다. 중앙 솥 아래 불 한 곳과 주변 모닥불 세 곳이 보인다. 중앙에는 돌 화덕, 삼각대 하나와 매단 사슬, 솥 옆에 세운 뚜껑 하나가 있다. 인물들은 돗자리와 자갈 바닥에 여러 겹으로 둘러앉아 있다. 참조 장소를 특징짓는 오른쪽 천막, 갈대 가림막과 짐 꾸러미는 보이지 않는다. 약간 높은 시점에서 솥의 내부와 앞면은 보이지만, 가까운 인물들이 하단에서 잘리고 모임이 화면을 조밀하게 채운다.",
        "entities": "수십 명의 인물이 있으며, 대체로 성인부터 중장년의 한국인 남성으로 읽히는 외모에 상투·머리띠·갓과 회갈색 전통 복장을 갖췄다. 특정 인물의 신원은 지정되지 않았으며, 의병이라는 신분 자체를 확정할 표지는 뚜렷하지 않다. 검은 솥 하나에 닭 형태와 감자, 국물과 김이 보이지만 닭 두 마리를 명확히 분리해 세기는 어렵다. 솥은 참조와 같은 검은 금속 재질이지만 테두리와 매다는 장치의 형태가 다르다. 붉은 주황색 장작불, 호수와 야간 환경은 있다. 읽을 수 있는 글자는 없다.",
        "hard_violations": [],
        "physics": "앉은 인물들의 엉덩이와 접힌 다리는 땅이나 돗자리에 닿아 있고, 국자와 그릇은 손으로 지지된다. 솥은 돌 화덕 위에 놓이고 삼각대의 사슬도 솥의 손잡이 쪽으로 이어진다. 뚜껑은 바닥과 화덕 돌에 기대어 있다. 장작은 땅에 놓여 타며 열기와 김이 위로 오른다. 지지 없이 떠 있는 인물이나 물체는 보이지 않는다."
       },
       {
        "label": "A",
        "direction": "둘러앉은 인물들은 자기 그릇, 중앙 솥 또는 옆 사람을 바라본다. 솥 오른쪽 남자는 시선을 국물에 두고 국자를 솥 안으로 뻗으며, 다른 인물들은 숟가락을 그릇이나 입 쪽으로 향한다. 식사 동작의 방향은 자연스럽지만, 여러 사람이 서로 크게 웃고 떠드는 장면보다는 차분히 먹는 순간으로 읽힌다.",
        "built_space": "성인 인물 13명이 중앙 솥 하나를 중심으로 원형으로 앉아 있다. 왼쪽에는 호수와 바위 해안, 상단에는 산과 밤하늘의 여백이 있고 오른쪽에는 대숲과 천막 시설이 이어진다. 천막의 기둥, 천 지붕, 세운 갈대 가림막과 묶은 짐이 참조 장소의 구성을 상당히 유지한다. 모임 양옆에 모닥불 두 곳이 있으며, 솥은 중앙의 돌 받침 위에 놓인다. 높은 시점의 와이드숏으로 솥의 열린 윗면과 가까운 측면이 모두 보인다. 다만 솥의 지름은 주변 사람들에 비해 상당히 커서 참조보다 과장되어 보인다.",
        "entities": "인물들은 한국인 성인 남성으로 읽히는 외모에 상투와 소박한 회색·베이지색 전통 복장을 갖췄으며, 일부는 중년으로 보인다. 의병 신분을 확정할 무장은 보이지 않지만 시대 복장은 부합한다. 검은 무쇠솥 하나에 두 마리로 구분되는 닭 형태와 여러 감자, 국물과 김이 보인다. 솥의 양옆 고리 손잡이와 둥근 몸체는 참조 소품에 가깝지만, 크기와 몸통의 턱 형태는 완전히 같지 않다. 장작 모닥불, 호수, 천막과 대숲이 있고 밤의 어둠도 유지된다. 읽을 수 있는 글자는 없다.",
        "hard_violations": [],
        "physics": "인물들은 흙바닥이나 깔개에 엉덩이와 접은 다리를 대고 앉아 있으며, 식기와 국자는 손으로 잡고 있다. 솥의 무게는 둘레의 돌들이 받치고, 통과 바구니 및 장작도 지면에 놓여 있다. 천막은 기둥과 가로대가 지지한다. 솥에서 김은 올라오지만 솥 바로 아래의 불꽃은 보이지 않아 가열원은 직접 확인되지 않는다. 이것만으로 물리적 불가능을 단정할 수는 없으며, 지지 없이 떠 있는 물체나 인물은 없다."
       }
      ]
     },
     "ok": true
    }
   ],
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "slot_winner_match": true,
   "slot_winner": {
    "gemini-pro": "A",
    "gpt-high": "A"
   },
   "route": "cross_slot_agree"
  },
  "dual": {
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "normalized": {
    "A": 2.0,
    "B": 1.25
   },
   "adjusted": {
    "A": 2.0,
    "B": 1.0
   },
   "violations": {
    "B": [
     "[gemini-pro] 로케이션 레퍼런스에 없는 삼각대 구조물 임의 추가",
     "[gemini-pro] 로케이션 레퍼런스의 주요 지형 및 구조물(우측 천막) 누락 및 변경"
    ]
   },
   "per_model_winner": {
    "gemini-pro": "A",
    "gpt-high": "A"
   },
   "agreed": true
  },
  "totals": {
   "A": 2000,
   "B": 1000
  },
  "selected": "A",
  "ranking": [
   "A",
   "B"
  ],
  "verdicts": [
   {
    "label": "A",
    "score": 2000,
    "verdict_ko": "지정된 로케이션의 지형(좌측 호수, 우측 천막 구조물 등)과 무쇠솥이 돌 위에 놓인 방식을 정확히 구현했으나, 솥 안에 묘사된 닭의 마릿수가 지침(두 마리)을 초과한 점이 아쉽습니다."
   },
   {
    "label": "B",
    "score": 1000,
    "verdict_ko": "로케이션 레퍼런스의 공간 구조(우측 천막 부재, 배경 전체가 호수로 변경됨)를 무시하였으며, 무쇠솥 위에 임의로 삼각대 거치대를 추가하여 공간 및 소품 일치도에서 크게 감점되었습니다.  ★위반: [gemini-pro] 로케이션 레퍼런스에 없는 삼각대 구조물 임의 추가 / [gemini-pro] 로케이션 레퍼런스의 주요 지형 및 구조물(우측 천막) 누락 및 변경"
   }
  ],
  "refs": [
   {
    "label": "LOCATION PHOTOGRAPH — the exact place of this shot: its architecture, materials, fixed features and lighting mood are spatial truth; stage the moment inside this place. Never copy its camera framing.",
    "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/db00517b-165b-4005-a287-f28936b59b95/images/272bd7af-9294-4490-98c6-88a40e9deacc/scene/recipe/groupbg_lake_cooking_circle_fb9493.png",
    "asset_id": "f7b5ab04-6778-439a-b1ab-10ec63e47377",
    "role": "bgfirst_group_bg"
   },
   {
    "label": "PROP REFERENCE — 무쇠솥: the exact object appearing in this shot; match its look, material and wear exactly.",
    "path": "<bytes:1131910>",
    "asset_id": "8510d902-c12d-45d6-9f5c-8a90c9a902b3",
    "role": "prop_ref"
   }
  ],
  "critique_skipped": true,
  "shot_run_produced": false,
  "shot_run_uid": "06aa140f-2982-73d9-83f0-59473a05b91f",
  "bgfirst": {
   "bg_path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/db00517b-165b-4005-a287-f28936b59b95/images/272bd7af-9294-4490-98c6-88a40e9deacc/scene/recipe/S15sh1__bgfirst_bg.png",
   "bg_asset_id": "7f647561-3d51-485b-adc7-5e7d3269fe0a",
   "bg_record_key": "S15sh1::bgfirst_bg",
   "chain_winner": true,
   "authority": "groupbg",
   "group_key": "lake_cooking_circle",
   "groupbg_asset_id": "f7b5ab04-6778-439a-b1ab-10ec63e47377"
  },
  "ref_mode": "재투영 배경+콘티+엔티티 (2택1: 체인 승)",
  "share_plan": {
   "ref_plan": "background"
  }
 },
 "S15sh1::cine": {
  "applied": true,
  "attempted_at": "2026-09-09T02:25:44.156695+00:00",
  "fingerprint": "48957ee406832cd201ddeadff0931a9167c87a05399ae3c7f804043f59db4f1e",
  "fingerprint_version": 2,
  "provider": "mai",
  "endpoint": "openrouter/chat-completions",
  "model": "microsoft/mai-image-2.6",
  "pack": "24.202608252115",
  "source_file": "S15sh1_sel.png",
  "source_sha256": "3eb80836b545fb9ededd267e674ce92beade39dab92a208630002c67a7b3bc23",
  "file": "S15sh1_cine.png",
  "staged_sha256": "04ee44c91852361411a38e731515ec9163f6fb725ab36259c97d5ad4af025d30",
  "latency_ms": 25628
 },
 "S15sh8::signage": {
  "fp": "a7db927be0c1e238",
  "inscriptions": [],
  "cues": [],
  "dropped": []
 },
 "S15sh8": {
  "input_fingerprint": "7c2ed79222b0f99a",
  "prompt": "Create ONE FINAL photorealistic live-action film still of the moment below — early 20th-century Korea, Jeolla region; all people are Korean unless explicitly stated otherwise. TIME OF DAY (lock): night.\n\nSHOT TEXT (authoritative, Korean): 먹던 고기를 내려놓은 채 대장 막사를 향해 등을 보인 강용언의 뒷모습을 뚫어져라 응시하는 안규홍의 눈빛.\n\nLOCATION (lock): The pot-side eating area of the lakeside camp, looking across the outdoor fire toward the commander's tent. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: close-up\n- KEY BACKGROUND ELEMENTS: 먹던 닭고기 (Lowered as 안규홍 pauses eating); used as A small fragment at the lower edge makes the interrupted meal legible.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Existing campfire light retains its restrained warmth and natural falloff, with no lighting shift to exaggerate 안규홍's quiet attention.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The people visible in that photo are NOT in this shot: never carry their faces, bodies or clothing onto anyone here. Each person in THIS shot is defined solely by the PEOPLE list, their own CHARACTER REFERENCE images and their wardrobe notes. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nPREVIOUS STILL USAGE (follow exactly — what to take from the attached still and what to exclude; it governs place and objects only, never who is in this shot or how the camera sees them): Take the iron cooking pot, gravel, earthenware bowls, and warm firelight against the dark lakeshore. Exclude the departing leader from the eating group and any airborne sparks unique to the reference instant.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): Chicken-and-potato broth continues boiling beside the campfires; the makeshift shelters remain in darkness. 안규홍: He remains seated with chicken left in his bowl and the jade ring concealed in his clothing.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 안규홍 (한국인 남성, 청년 나이대, 검은 머리, 단단히 묶은 머리) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.",
  "roll_prompts": {
   "A": "Create ONE FINAL photorealistic live-action film still of the moment below — early 20th-century Korea, Jeolla region; all people are Korean unless explicitly stated otherwise. TIME OF DAY (lock): night.\n\nSHOT TEXT (authoritative, Korean): 먹던 고기를 내려놓은 채 대장 막사를 향해 등을 보인 강용언의 뒷모습을 뚫어져라 응시하는 안규홍의 눈빛.\n\nLOCATION (lock): The pot-side eating area of the lakeside camp, looking across the outdoor fire toward the commander's tent. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: close-up\n- KEY BACKGROUND ELEMENTS: 먹던 닭고기 (Lowered as 안규홍 pauses eating); used as A small fragment at the lower edge makes the interrupted meal legible.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Existing campfire light retains its restrained warmth and natural falloff, with no lighting shift to exaggerate 안규홍's quiet attention.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The people visible in that photo are NOT in this shot: never carry their faces, bodies or clothing onto anyone here. Each person in THIS shot is defined solely by the PEOPLE list, their own CHARACTER REFERENCE images and their wardrobe notes. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nPREVIOUS STILL USAGE (follow exactly — what to take from the attached still and what to exclude; it governs place and objects only, never who is in this shot or how the camera sees them): Take the iron cooking pot, gravel, earthenware bowls, and warm firelight against the dark lakeshore. Exclude the departing leader from the eating group and any airborne sparks unique to the reference instant.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): Chicken-and-potato broth continues boiling beside the campfires; the makeshift shelters remain in darkness. 안규홍: He remains seated with chicken left in his bowl and the jade ring concealed in his clothing.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 안규홍 (한국인 남성, 청년 나이대, 검은 머리, 단단히 묶은 머리) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.",
   "B": "Create ONE FINAL photorealistic live-action film still of the moment below — early 20th-century Korea, Jeolla region; all people are Korean unless explicitly stated otherwise. TIME OF DAY (lock): night.\n\nSHOT TEXT (authoritative, Korean): 먹던 고기를 내려놓은 채 대장 막사를 향해 등을 보인 강용언의 뒷모습을 뚫어져라 응시하는 안규홍의 눈빛.\n\nLOCATION (lock): The pot-side eating area of the lakeside camp, looking across the outdoor fire toward the commander's tent. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: close-up\n- KEY BACKGROUND ELEMENTS: 먹던 닭고기 (Lowered as 안규홍 pauses eating); used as A small fragment at the lower edge makes the interrupted meal legible.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Existing campfire light retains its restrained warmth and natural falloff, with no lighting shift to exaggerate 안규홍's quiet attention.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The people visible in that photo are NOT in this shot: never carry their faces, bodies or clothing onto anyone here. Each person in THIS shot is defined solely by the PEOPLE list, their own CHARACTER REFERENCE images and their wardrobe notes. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nPREVIOUS STILL USAGE (follow exactly — what to take from the attached still and what to exclude; it governs place and objects only, never who is in this shot or how the camera sees them): Take the iron cooking pot, gravel, earthenware bowls, and warm firelight against the dark lakeshore. Exclude the departing leader from the eating group and any airborne sparks unique to the reference instant.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): Chicken-and-potato broth continues boiling beside the campfires; the makeshift shelters remain in darkness. 안규홍: He remains seated with chicken left in his bowl and the jade ring concealed in his clothing.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 안규홍 (한국인 남성, 청년 나이대, 검은 머리, 단단히 묶은 머리) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.\n\nCOMPOSITION VARIATION (this candidate only): keep the place, the people and the moment exactly as specified, and keep from the CAMERA & FRAME contract its subject scale, screen placement and key background elements exactly. For this candidate only, its camera angle and height are deliberately left open: choose a camera position distinctly different from the obvious one, as a film director picking a second setup on the same blocking. Stay true to any explicit framing the SHOT TEXT itself dictates."
  },
  "shot_run_spend_attempt_count": 1,
  "readings": [
   {
    "label": "A",
    "direction": "안규홍의 시선은 화면 우측 전경에 있는 인물의 측면 및 뒷모습을 향하고 있으며, 샷 텍스트가 요구한 막사를 향해 걸어가는 인물을 응시하지 않음.",
    "built_space": "밤의 호숫가 야영지로 모닥불, 가마솥, 자갈 바닥이 기준 이미지와 일치하게 묘사됨.",
    "entities": "안규홍은 레퍼런스와 유사한 복장을 입고 있으나 머리를 상투로 틀어 올림. 샷 텍스트에 없는 추가 인물(전경 우측 및 배경에 서 있는 인물로 인해 총 3명)이 등장함.",
    "hard_violations": [
     "[gemini-pro] 샷 텍스트에 없는 추가 인물이 프레임에 포함됨 (extra bodies/invented people)",
     "[gpt-high] 안규홍과 떠나는 강용언 외에 오른쪽 전경의 남성을 추가해, 허용되지 않은 세 번째 인물이 화면을 크게 차지한다."
    ],
    "physics": "안규홍은 바닥에 안정적으로 앉아 있고 그릇도 땅에 놓여 있으며, 비물리적으로 떠 있는 물체나 신체는 없음."
   },
   {
    "label": "B",
    "direction": "안규홍의 시선이 화면 좌측 배경을 향해 멀어지는 강용언의 뒷모습을 뚫어져라 응시하는 목표 대상에 정확히 안착함.",
    "built_space": "밤의 호숫가 야영지에 가마솥과 모닥불, 자갈밭이 기준 이미지의 배경 요소와 잘 맞아떨어지게 배치됨.",
    "entities": "안규홍의 외형과 단단히 묶은 머리가 레퍼런스와 일치하며, 지시된 대로 안규홍과 강용언의 뒷모습만 프레임에 존재함.",
    "hard_violations": [],
    "physics": "안규홍은 바닥에 자연스럽게 앉아 무릎에 팔을 지지하고 있으며, 양손에 든 빈 그릇과 닭고기 모두 손과 신체에 의해 안정적으로 지탱됨."
   }
  ],
  "cross_model_order": {
   "policy": "cross_model_order_v2_slot0_fwd_slot1_rev_parallel2",
   "slots": [
    {
     "model": "gemini-pro",
     "order": "forward",
     "display_to_canonical": {
      "A": "A",
      "B": "B"
     },
     "raw": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 7,
        "verdict_ko": "먹던 고기를 내려놓으라는 지시와 달리 고기를 손에 들고 있어 유지 상태(Carried State)를 위반했으나, 명시된 인물만 등장하였고 시선 방향이 정확하여 A보다 우수합니다."
       },
       {
        "label": "A",
        "score": 3,
        "verdict_ko": "프롬프트에 명시되지 않은 제3의 인물이 등장하는 치명적 오류(Hard Violation)가 있으며, 안규홍의 시선 또한 멀어지는 뒷모습이 아닌 전경의 인물을 향하고 있습니다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "안규홍의 시선은 화면 우측 전경에 있는 인물의 측면 및 뒷모습을 향하고 있으며, 샷 텍스트가 요구한 막사를 향해 걸어가는 인물을 응시하지 않음.",
        "built_space": "밤의 호숫가 야영지로 모닥불, 가마솥, 자갈 바닥이 기준 이미지와 일치하게 묘사됨.",
        "entities": "안규홍은 레퍼런스와 유사한 복장을 입고 있으나 머리를 상투로 틀어 올림. 샷 텍스트에 없는 추가 인물(전경 우측 및 배경에 서 있는 인물로 인해 총 3명)이 등장함.",
        "hard_violations": [
         "샷 텍스트에 없는 추가 인물이 프레임에 포함됨 (extra bodies/invented people)"
        ],
        "physics": "안규홍은 바닥에 안정적으로 앉아 있고 그릇도 땅에 놓여 있으며, 비물리적으로 떠 있는 물체나 신체는 없음."
       },
       {
        "label": "B",
        "direction": "안규홍의 시선이 화면 좌측 배경을 향해 멀어지는 강용언의 뒷모습을 뚫어져라 응시하는 목표 대상에 정확히 안착함.",
        "built_space": "밤의 호숫가 야영지에 가마솥과 모닥불, 자갈밭이 기준 이미지의 배경 요소와 잘 맞아떨어지게 배치됨.",
        "entities": "안규홍의 외형과 단단히 묶은 머리가 레퍼런스와 일치하며, 지시된 대로 안규홍과 강용언의 뒷모습만 프레임에 존재함.",
        "hard_violations": [],
        "physics": "안규홍은 바닥에 자연스럽게 앉아 무릎에 팔을 지지하고 있으며, 양손에 든 빈 그릇과 닭고기 모두 손과 신체에 의해 안정적으로 지탱됨."
       }
      ],
      "all_candidates_fail": false
     },
     "normalized": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 7,
        "verdict_ko": "먹던 고기를 내려놓으라는 지시와 달리 고기를 손에 들고 있어 유지 상태(Carried State)를 위반했으나, 명시된 인물만 등장하였고 시선 방향이 정확하여 A보다 우수합니다."
       },
       {
        "label": "A",
        "score": 3,
        "verdict_ko": "프롬프트에 명시되지 않은 제3의 인물이 등장하는 치명적 오류(Hard Violation)가 있으며, 안규홍의 시선 또한 멀어지는 뒷모습이 아닌 전경의 인물을 향하고 있습니다."
       }
      ],
      "all_candidates_fail": false,
      "readings": [
       {
        "label": "A",
        "direction": "안규홍의 시선은 화면 우측 전경에 있는 인물의 측면 및 뒷모습을 향하고 있으며, 샷 텍스트가 요구한 막사를 향해 걸어가는 인물을 응시하지 않음.",
        "built_space": "밤의 호숫가 야영지로 모닥불, 가마솥, 자갈 바닥이 기준 이미지와 일치하게 묘사됨.",
        "entities": "안규홍은 레퍼런스와 유사한 복장을 입고 있으나 머리를 상투로 틀어 올림. 샷 텍스트에 없는 추가 인물(전경 우측 및 배경에 서 있는 인물로 인해 총 3명)이 등장함.",
        "hard_violations": [
         "샷 텍스트에 없는 추가 인물이 프레임에 포함됨 (extra bodies/invented people)"
        ],
        "physics": "안규홍은 바닥에 안정적으로 앉아 있고 그릇도 땅에 놓여 있으며, 비물리적으로 떠 있는 물체나 신체는 없음."
       },
       {
        "label": "B",
        "direction": "안규홍의 시선이 화면 좌측 배경을 향해 멀어지는 강용언의 뒷모습을 뚫어져라 응시하는 목표 대상에 정확히 안착함.",
        "built_space": "밤의 호숫가 야영지에 가마솥과 모닥불, 자갈밭이 기준 이미지의 배경 요소와 잘 맞아떨어지게 배치됨.",
        "entities": "안규홍의 외형과 단단히 묶은 머리가 레퍼런스와 일치하며, 지시된 대로 안규홍과 강용언의 뒷모습만 프레임에 존재함.",
        "hard_violations": [],
        "physics": "안규홍은 바닥에 자연스럽게 앉아 무릎에 팔을 지지하고 있으며, 양손에 든 빈 그릇과 닭고기 모두 손과 신체에 의해 안정적으로 지탱됨."
       }
      ]
     },
     "ok": true
    },
    {
     "model": "gpt-high",
     "order": "reverse",
     "display_to_canonical": {
      "A": "B",
      "B": "A"
     },
     "raw": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 4,
        "verdict_ko": "안규홍이 떠나는 남자의 등을 주시하는 관계는 살렸지만, 눈빛 중심의 클로즈업이 아니라 넓은 중경이며 남자가 막사보다 호수 쪽으로 향한다."
       },
       {
        "label": "B",
        "score": 1,
        "verdict_ko": "불필요한 전경 인물을 추가했고 안규홍의 시선도 그에게 향해, 강용언의 뒷모습을 응시하는 클로즈업이라는 핵심을 놓쳤다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "오른쪽 안규홍의 얼굴과 눈은 화면 왼쪽의 등을 보인 남자 쪽으로 향한다. 응시 대상과 시선은 대체로 연결되지만, 그 남자의 몸과 발은 오른쪽 막사가 아니라 왼쪽 뒤 호숫가 방향으로 향한다.",
        "built_space": "왼쪽 아래 돌 받침 위에 철솥 하나, 그 뒤 장작불 하나, 오른쪽 뒤에 나무 기둥과 천으로 된 막사 일부가 보인다. 토기 그릇은 손에 든 것 하나와 바닥의 큰 것 하나가 보인다. 자갈과 호수, 대나무는 장소 참고와 이어지지만, 인물의 무릎과 큰 솥까지 포함한 넓은 구도가 요구된 클로즈업을 대체했다.",
        "entities": "두 남자가 보이며, 전경의 젊은 동아시아계 남성은 검은 머리를 묶고 때 탄 밝은 전통 상의를 입어 안규홍 참고의 주요 특징을 따른다. 다만 얼굴은 참고보다 길고 날카롭게 보인다. 배경 남자는 강용언의 뒷모습으로 읽힌다. 철솥, 국물, 토기, 내려 든 닭고기는 식별되지만 닭고기는 그릇 안이 아니라 손에 있으며 손의 그릇에서 남은 고기는 확인되지 않는다. 옥반지와 읽을 수 있는 글자는 보이지 않는다.",
        "hard_violations": [],
        "physics": "안규홍은 다리를 접어 낮게 앉아 있고, 한 손은 그릇을 받치며 다른 손은 닭고기를 실제로 쥔다. 배경 남자는 자갈에 발을 디딘 보행 자세다. 솥은 돌에, 큰 토기는 지면에 놓여 있다. 공중에 지지 없이 떠 있는 몸이나 물체는 없다."
       },
       {
        "label": "B",
        "direction": "왼쪽 안규홍은 오른쪽 바로 앞에 앉은 남자의 얼굴 쪽을 바라본다. 뒤에서 멀어지는 남자는 그 시선보다 안쪽 배경에 있어 응시 대상으로 연결되지 않는다. 배경 남자는 대나무와 막사 쪽으로 등을 보이고 이동한다.",
        "built_space": "중앙 오른쪽에 돌로 받친 철솥 하나, 아래 가장자리에 토기 그릇 하나가 보인다. 솥 주변에는 왼쪽·중앙·오른쪽으로 나뉜 세 불꽃 군집이 보이고, 오른쪽 뒤에는 막사 천과 기둥 일부가 있다. 자갈 호숫가의 재료는 이어지지만, 전경 남자의 어깨와 안규홍의 무릎까지 담은 대화형 구도여서 지정된 눈빛 클로즈업이 아니다.",
        "entities": "안규홍으로 보이는 젊은 동아시아계 남성, 배경에서 등을 보인 남성, 오른쪽 전경의 머리띠를 한 남성까지 세 사람이 보인다. 전경 남성은 장면에 추가된 인물이다. 안규홍의 검은 묶음머리와 얼굴은 참고의 특징을 일부 따르지만 상의는 훨씬 어둡게 보인다. 토기 안의 닭고기와 솥 속 국물은 보이며 감자는 명확하게 구별되지 않는다. 옥반지나 읽을 수 있는 글자는 없다.",
        "hard_violations": [
         "안규홍과 떠나는 강용언 외에 오른쪽 전경의 남성을 추가해, 허용되지 않은 세 번째 인물이 화면을 크게 차지한다."
        ],
        "physics": "안규홍은 다리를 접고 앉아 무릎 부근에 손을 둔다. 배경 남자의 하체는 연기에 일부 가려져 있지만 지면을 따라 걷는 자세이며 부유한 모습은 아니다. 솥은 돌 받침이 지지한다. 아래 토기는 바닥 높이에 놓인 것으로 보이나 바닥 접촉부는 프레임 밖이다. 전경 남자의 좌면과 발도 잘려 있어 지지 상태를 직접 확인할 수는 없지만 불가능한 자세는 보이지 않는다."
       }
      ],
      "all_candidates_fail": true
     },
     "normalized": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 4,
        "verdict_ko": "안규홍이 떠나는 남자의 등을 주시하는 관계는 살렸지만, 눈빛 중심의 클로즈업이 아니라 넓은 중경이며 남자가 막사보다 호수 쪽으로 향한다."
       },
       {
        "label": "A",
        "score": 1,
        "verdict_ko": "불필요한 전경 인물을 추가했고 안규홍의 시선도 그에게 향해, 강용언의 뒷모습을 응시하는 클로즈업이라는 핵심을 놓쳤다."
       }
      ],
      "all_candidates_fail": true,
      "readings": [
       {
        "label": "B",
        "direction": "오른쪽 안규홍의 얼굴과 눈은 화면 왼쪽의 등을 보인 남자 쪽으로 향한다. 응시 대상과 시선은 대체로 연결되지만, 그 남자의 몸과 발은 오른쪽 막사가 아니라 왼쪽 뒤 호숫가 방향으로 향한다.",
        "built_space": "왼쪽 아래 돌 받침 위에 철솥 하나, 그 뒤 장작불 하나, 오른쪽 뒤에 나무 기둥과 천으로 된 막사 일부가 보인다. 토기 그릇은 손에 든 것 하나와 바닥의 큰 것 하나가 보인다. 자갈과 호수, 대나무는 장소 참고와 이어지지만, 인물의 무릎과 큰 솥까지 포함한 넓은 구도가 요구된 클로즈업을 대체했다.",
        "entities": "두 남자가 보이며, 전경의 젊은 동아시아계 남성은 검은 머리를 묶고 때 탄 밝은 전통 상의를 입어 안규홍 참고의 주요 특징을 따른다. 다만 얼굴은 참고보다 길고 날카롭게 보인다. 배경 남자는 강용언의 뒷모습으로 읽힌다. 철솥, 국물, 토기, 내려 든 닭고기는 식별되지만 닭고기는 그릇 안이 아니라 손에 있으며 손의 그릇에서 남은 고기는 확인되지 않는다. 옥반지와 읽을 수 있는 글자는 보이지 않는다.",
        "hard_violations": [],
        "physics": "안규홍은 다리를 접어 낮게 앉아 있고, 한 손은 그릇을 받치며 다른 손은 닭고기를 실제로 쥔다. 배경 남자는 자갈에 발을 디딘 보행 자세다. 솥은 돌에, 큰 토기는 지면에 놓여 있다. 공중에 지지 없이 떠 있는 몸이나 물체는 없다."
       },
       {
        "label": "A",
        "direction": "왼쪽 안규홍은 오른쪽 바로 앞에 앉은 남자의 얼굴 쪽을 바라본다. 뒤에서 멀어지는 남자는 그 시선보다 안쪽 배경에 있어 응시 대상으로 연결되지 않는다. 배경 남자는 대나무와 막사 쪽으로 등을 보이고 이동한다.",
        "built_space": "중앙 오른쪽에 돌로 받친 철솥 하나, 아래 가장자리에 토기 그릇 하나가 보인다. 솥 주변에는 왼쪽·중앙·오른쪽으로 나뉜 세 불꽃 군집이 보이고, 오른쪽 뒤에는 막사 천과 기둥 일부가 있다. 자갈 호숫가의 재료는 이어지지만, 전경 남자의 어깨와 안규홍의 무릎까지 담은 대화형 구도여서 지정된 눈빛 클로즈업이 아니다.",
        "entities": "안규홍으로 보이는 젊은 동아시아계 남성, 배경에서 등을 보인 남성, 오른쪽 전경의 머리띠를 한 남성까지 세 사람이 보인다. 전경 남성은 장면에 추가된 인물이다. 안규홍의 검은 묶음머리와 얼굴은 참고의 특징을 일부 따르지만 상의는 훨씬 어둡게 보인다. 토기 안의 닭고기와 솥 속 국물은 보이며 감자는 명확하게 구별되지 않는다. 옥반지나 읽을 수 있는 글자는 없다.",
        "hard_violations": [
         "안규홍과 떠나는 강용언 외에 오른쪽 전경의 남성을 추가해, 허용되지 않은 세 번째 인물이 화면을 크게 차지한다."
        ],
        "physics": "안규홍은 다리를 접고 앉아 무릎 부근에 손을 둔다. 배경 남자의 하체는 연기에 일부 가려져 있지만 지면을 따라 걷는 자세이며 부유한 모습은 아니다. 솥은 돌 받침이 지지한다. 아래 토기는 바닥 높이에 놓인 것으로 보이나 바닥 접촉부는 프레임 밖이다. 전경 남자의 좌면과 발도 잘려 있어 지지 상태를 직접 확인할 수는 없지만 불가능한 자세는 보이지 않는다."
       }
      ]
     },
     "ok": true
    }
   ],
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "slot_winner_match": true,
   "slot_winner": {
    "gemini-pro": "B",
    "gpt-high": "B"
   },
   "route": "cross_slot_agree"
  },
  "dual": {
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "normalized": {
    "A": 0.679,
    "B": 2.0
   },
   "adjusted": {
    "A": 0.429,
    "B": 2.0
   },
   "violations": {
    "A": [
     "[gemini-pro] 샷 텍스트에 없는 추가 인물이 프레임에 포함됨 (extra bodies/invented people)",
     "[gpt-high] 안규홍과 떠나는 강용언 외에 오른쪽 전경의 남성을 추가해, 허용되지 않은 세 번째 인물이 화면을 크게 차지한다."
    ]
   },
   "per_model_winner": {
    "gemini-pro": "B",
    "gpt-high": "B"
   },
   "agreed": true
  },
  "totals": {
   "B": 2000,
   "A": 429
  },
  "selected": "B",
  "ranking": [
   "B",
   "A"
  ],
  "verdicts": [
   {
    "label": "B",
    "score": 2000,
    "verdict_ko": "먹던 고기를 내려놓으라는 지시와 달리 고기를 손에 들고 있어 유지 상태(Carried State)를 위반했으나, 명시된 인물만 등장하였고 시선 방향이 정확하여 A보다 우수합니다."
   },
   {
    "label": "A",
    "score": 429,
    "verdict_ko": "프롬프트에 명시되지 않은 제3의 인물이 등장하는 치명적 오류(Hard Violation)가 있으며, 안규홍의 시선 또한 멀어지는 뒷모습이 아닌 전경의 인물을 향하고 있습니다.  ★위반: [gemini-pro] 샷 텍스트에 없는 추가 인물이 프레임에 포함됨 (extra bodies/invented people) / [gpt-high] 안규홍과 떠나는 강용언 외에 오른쪽 전경의 남성을 추가해, 허용되지 않은 세 번째 인물이 화면을 크게 차지한다."
   }
  ],
  "refs": [
   {
    "label": "PREVIOUS SHOT STILL — a visually related earlier shot of this same place: the location's look, materials, fixed features and lighting mood are LOCKED to this photo; never copy its camera framing. The people visible in that photo are NOT in this shot: never carry their faces, bodies or clothing onto anyone here. Each person in THIS shot is defined solely by the PEOPLE list, their own CHARACTER REFERENCE images and their wardrobe notes. If this photo shows a character who cannot move (dead or unconscious), that character's exact pose, position and orientation are ALSO LOCKED — treat their body as a fixed prop of the set that only the camera moves around.",
    "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/db00517b-165b-4005-a287-f28936b59b95/images/272bd7af-9294-4490-98c6-88a40e9deacc/scene/recipe/S15sh1_sel.png",
    "asset_id": "b59e0f47-3da6-40bd-bc25-fe3308f3c7d5",
    "role": "prev_still"
   },
   {
    "label": "CHARACTER REFERENCE — 안규홍: the exact person appearing in this shot; match face, hair and build exactly.",
    "path": "<bytes:1028384>",
    "asset_id": "3b5894db-ce55-4f34-b966-5e725c3aa952",
    "role": "character_ref"
   }
  ],
  "critique_skipped": true,
  "shot_run_produced": false,
  "shot_run_uid": "06aa140f-2d96-7db0-af15-8510cccfba70",
  "ref_mode": "prev+엔티티",
  "share_plan": {
   "ref_plan": "prev",
   "prev_anchor_tag": "S15sh1"
  }
 },
 "S15sh8::cine": {
  "applied": true,
  "attempted_at": "2026-09-09T02:27:18.140956+00:00",
  "fingerprint": "ea3b4dcd73839f9f6b4ec340b25bac48e979a131a76227dabc8b666f38ad5ad1",
  "fingerprint_version": 2,
  "provider": "mai",
  "endpoint": "openrouter/chat-completions",
  "model": "microsoft/mai-image-2.6",
  "pack": "24.202608252115",
  "source_file": "S15sh8_sel.png",
  "source_sha256": "490bad75491396a9ea0c6db7b88252093ac1bf045bc88336327c48bf091a1aeb",
  "file": "S15sh8_cine.png",
  "staged_sha256": "4bba07dff096f874b2ecbaa5a152c519669e233380be61ca6da8668b96760df2",
  "latency_ms": 30705
 },
 "S15sh12::signage": {
  "fp": "5ecf8fe489bcf3b4",
  "inscriptions": [],
  "cues": [],
  "dropped": []
 },
 "S15sh12": {
  "input_fingerprint": "2021ab2572732242",
  "prompt": "Create ONE FINAL photorealistic live-action film still of the moment below — early 20th-century Korea, Jeolla region; all people are Korean unless explicitly stated otherwise. TIME OF DAY (lock): night.\n\nSHOT TEXT (authoritative, Korean): 어이가 없다는 듯 한쪽 입꼬리를 올린 채 피식 웃는 안규홍의 정면 클로즈업.\n\nLOCATION (lock): The outdoor supper circle beside the iron cooking pot in the firelit lakeside camp. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: close-up\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Unchanged campfire warmth and subdued contrast allow the brief smile to register as ordinary human affection rather than a sentimental lighting cue.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The clothing, hair and overall look of 안규홍 — who appear both in that photo and in this shot — are LOCKED to that photo. Anyone else visible in that photo is NOT in this shot: never carry their face, body or clothing onto anyone here. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nPREVIOUS STILL USAGE (follow exactly — what to take from the attached still and what to exclude; it governs place and objects only, never who is in this shot or how the camera sees them): Take the iron pot, earthenware dishes, gravel, and nighttime firelight. Exclude the leader and his companion from the eating group, since they have left for the command tent.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): Campfires continue burning beside the chicken-broth pot and food bowls. 안규홍: He remains seated at the meal, with the jade ring concealed in his clothing.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 안규홍 (한국인 남성, 청년 나이대, 검은 머리, 단단히 묶은 머리) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.",
  "roll_prompts": {
   "A": "Create ONE FINAL photorealistic live-action film still of the moment below — early 20th-century Korea, Jeolla region; all people are Korean unless explicitly stated otherwise. TIME OF DAY (lock): night.\n\nSHOT TEXT (authoritative, Korean): 어이가 없다는 듯 한쪽 입꼬리를 올린 채 피식 웃는 안규홍의 정면 클로즈업.\n\nLOCATION (lock): The outdoor supper circle beside the iron cooking pot in the firelit lakeside camp. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: close-up\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Unchanged campfire warmth and subdued contrast allow the brief smile to register as ordinary human affection rather than a sentimental lighting cue.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The clothing, hair and overall look of 안규홍 — who appear both in that photo and in this shot — are LOCKED to that photo. Anyone else visible in that photo is NOT in this shot: never carry their face, body or clothing onto anyone here. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nPREVIOUS STILL USAGE (follow exactly — what to take from the attached still and what to exclude; it governs place and objects only, never who is in this shot or how the camera sees them): Take the iron pot, earthenware dishes, gravel, and nighttime firelight. Exclude the leader and his companion from the eating group, since they have left for the command tent.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): Campfires continue burning beside the chicken-broth pot and food bowls. 안규홍: He remains seated at the meal, with the jade ring concealed in his clothing.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 안규홍 (한국인 남성, 청년 나이대, 검은 머리, 단단히 묶은 머리) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.",
   "B": "Create ONE FINAL photorealistic live-action film still of the moment below — early 20th-century Korea, Jeolla region; all people are Korean unless explicitly stated otherwise. TIME OF DAY (lock): night.\n\nSHOT TEXT (authoritative, Korean): 어이가 없다는 듯 한쪽 입꼬리를 올린 채 피식 웃는 안규홍의 정면 클로즈업.\n\nLOCATION (lock): The outdoor supper circle beside the iron cooking pot in the firelit lakeside camp. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: close-up\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Unchanged campfire warmth and subdued contrast allow the brief smile to register as ordinary human affection rather than a sentimental lighting cue.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The clothing, hair and overall look of 안규홍 — who appear both in that photo and in this shot — are LOCKED to that photo. Anyone else visible in that photo is NOT in this shot: never carry their face, body or clothing onto anyone here. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nPREVIOUS STILL USAGE (follow exactly — what to take from the attached still and what to exclude; it governs place and objects only, never who is in this shot or how the camera sees them): Take the iron pot, earthenware dishes, gravel, and nighttime firelight. Exclude the leader and his companion from the eating group, since they have left for the command tent.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): Campfires continue burning beside the chicken-broth pot and food bowls. 안규홍: He remains seated at the meal, with the jade ring concealed in his clothing.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 안규홍 (한국인 남성, 청년 나이대, 검은 머리, 단단히 묶은 머리) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.\n\nCOMPOSITION VARIATION (this candidate only): keep the place, the people and the moment exactly as specified, and keep from the CAMERA & FRAME contract its subject scale, screen placement and key background elements exactly. For this candidate only, its camera angle and height are deliberately left open: choose a camera position distinctly different from the obvious one, as a film director picking a second setup on the same blocking. Stay true to any explicit framing the SHOT TEXT itself dictates."
  },
  "shot_run_spend_attempt_count": 1,
  "readings": [
   {
    "label": "A",
    "direction": "안규홍의 시선이 카메라(정면)를 향하고 있습니다.",
    "built_space": "밤의 호숫가 야영지로, 자갈밭 위에 무쇠솥과 도자기 그릇들이 배치되어 있으며 뒤쪽에 모닥불이 보입니다.",
    "entities": "안규홍의 외모와 복장이 레퍼런스와 일치하며, 텍스트의 요구대로 한쪽 입꼬리만 명확하게 올라간 썩소를 짓고 있습니다. 다만 바닥에 날것으로 보이는 닭고기 조각들이 놓여 있습니다.",
    "hard_violations": [],
    "physics": "자갈밭 위에 안정적으로 앉아 있으며 자세에 무리가 없습니다."
   },
   {
    "label": "B",
    "direction": "안규홍의 시선이 카메라(정면)를 향하고 있습니다.",
    "built_space": "밤의 호숫가 야영지로, 자갈밭과 무쇠솥, 도자기 그릇, 모닥불이 텍스트와 레퍼런스에 맞게 배치되어 있습니다.",
    "entities": "안규홍의 외모와 복장은 레퍼런스와 일치하지만, 한쪽 입꼬리를 올린 비웃음이 아닌 양쪽 입꼬리가 올라간 부드러운 미소를 짓고 있습니다.",
    "hard_violations": [],
    "physics": "자갈밭 위에 자연스럽게 앉아 체중을 지탱하고 있습니다."
   }
  ],
  "cross_model_order": {
   "policy": "cross_model_order_v2_slot0_fwd_slot1_rev_parallel2",
   "slots": [
    {
     "model": "gemini-pro",
     "order": "forward",
     "display_to_canonical": {
      "A": "A",
      "B": "B"
     },
     "raw": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 8,
        "verdict_ko": "지시된 '어이가 없다는 듯 한쪽 입꼬리를 올린' 표정을 매우 정확하게 묘사했으나, 바닥에 놓인 닭고기 조각들이 다소 어색합니다."
       },
       {
        "label": "B",
        "score": 6,
        "verdict_ko": "요구된 '어이가 없다는 듯한 피식 웃음'보다는 대칭적이고 온화한 미소에 가까워 샷 텍스트의 핵심 감정 표현을 살리지 못했습니다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "안규홍의 시선이 카메라(정면)를 향하고 있습니다.",
        "built_space": "밤의 호숫가 야영지로, 자갈밭 위에 무쇠솥과 도자기 그릇들이 배치되어 있으며 뒤쪽에 모닥불이 보입니다.",
        "entities": "안규홍의 외모와 복장이 레퍼런스와 일치하며, 텍스트의 요구대로 한쪽 입꼬리만 명확하게 올라간 썩소를 짓고 있습니다. 다만 바닥에 날것으로 보이는 닭고기 조각들이 놓여 있습니다.",
        "hard_violations": [],
        "physics": "자갈밭 위에 안정적으로 앉아 있으며 자세에 무리가 없습니다."
       },
       {
        "label": "B",
        "direction": "안규홍의 시선이 카메라(정면)를 향하고 있습니다.",
        "built_space": "밤의 호숫가 야영지로, 자갈밭과 무쇠솥, 도자기 그릇, 모닥불이 텍스트와 레퍼런스에 맞게 배치되어 있습니다.",
        "entities": "안규홍의 외모와 복장은 레퍼런스와 일치하지만, 한쪽 입꼬리를 올린 비웃음이 아닌 양쪽 입꼬리가 올라간 부드러운 미소를 짓고 있습니다.",
        "hard_violations": [],
        "physics": "자갈밭 위에 자연스럽게 앉아 체중을 지탱하고 있습니다."
       }
      ],
      "all_candidates_fail": false
     },
     "normalized": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 8,
        "verdict_ko": "지시된 '어이가 없다는 듯 한쪽 입꼬리를 올린' 표정을 매우 정확하게 묘사했으나, 바닥에 놓인 닭고기 조각들이 다소 어색합니다."
       },
       {
        "label": "B",
        "score": 6,
        "verdict_ko": "요구된 '어이가 없다는 듯한 피식 웃음'보다는 대칭적이고 온화한 미소에 가까워 샷 텍스트의 핵심 감정 표현을 살리지 못했습니다."
       }
      ],
      "all_candidates_fail": false,
      "readings": [
       {
        "label": "A",
        "direction": "안규홍의 시선이 카메라(정면)를 향하고 있습니다.",
        "built_space": "밤의 호숫가 야영지로, 자갈밭 위에 무쇠솥과 도자기 그릇들이 배치되어 있으며 뒤쪽에 모닥불이 보입니다.",
        "entities": "안규홍의 외모와 복장이 레퍼런스와 일치하며, 텍스트의 요구대로 한쪽 입꼬리만 명확하게 올라간 썩소를 짓고 있습니다. 다만 바닥에 날것으로 보이는 닭고기 조각들이 놓여 있습니다.",
        "hard_violations": [],
        "physics": "자갈밭 위에 안정적으로 앉아 있으며 자세에 무리가 없습니다."
       },
       {
        "label": "B",
        "direction": "안규홍의 시선이 카메라(정면)를 향하고 있습니다.",
        "built_space": "밤의 호숫가 야영지로, 자갈밭과 무쇠솥, 도자기 그릇, 모닥불이 텍스트와 레퍼런스에 맞게 배치되어 있습니다.",
        "entities": "안규홍의 외모와 복장은 레퍼런스와 일치하지만, 한쪽 입꼬리를 올린 비웃음이 아닌 양쪽 입꼬리가 올라간 부드러운 미소를 짓고 있습니다.",
        "hard_violations": [],
        "physics": "자갈밭 위에 자연스럽게 앉아 체중을 지탱하고 있습니다."
       }
      ]
     },
     "ok": true
    },
    {
     "model": "gpt-high",
     "order": "reverse",
     "display_to_canonical": {
      "A": "B",
      "B": "A"
     },
     "raw": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 7,
        "verdict_ko": "두 후보 모두 정면 얼굴 클로즈업보다 넓지만, A가 안규홍의 수염 없는 얼굴과 은근한 비대칭 미소, 절제된 모닥불 조명을 더 충실하게 유지한다."
       },
       {
        "label": "B",
        "score": 5,
        "verdict_ko": "한쪽 입꼬리를 올린 표정은 보이나, 넓은 식사 장면 구도와 참조에 없는 콧수염·턱수염이 핵심 구도 및 인물 연속성을 떨어뜨린다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "얼굴은 거의 정면이고 몸통은 비스듬하다. 두 눈은 카메라 가까이를 바라보며, 화면 오른쪽 입꼬리가 조금 더 올라간 미소가 보인다. 어이없다는 반응보다는 다정한 미소에 가깝다. 방향을 따질 무기나 손에 든 물체는 보이지 않는다.",
        "built_space": "왼쪽 전경에 화면 밖으로 잘린 무쇠솥 1개와 그 아래 받침돌, 왼쪽 가장자리에 장작불 1곳이 보인다. 인물 왼쪽 뒤 자갈 위에는 큰 토기 그릇 1개와 작은 그릇 1개가 보인다. 호수는 왼쪽 배경, 대나무와 천막 일부는 오른쪽 배경에 있어 이전 장소의 재료와 배치를 대체로 유지한다. 다만 가슴과 넓은 주변 공간까지 담아 얼굴 중심의 정면 클로즈업보다 넓다.",
        "entities": "한국인 청년 남성으로 설정된 인물 1명만 보인다. 검은 머리를 뒤로 팽팽하게 넘겼으며 묶음 자체는 뚜렷하지 않다. 수염 없는 얼굴과 눈·코·턱의 인상은 참조에 비교적 가깝고, 때 탄 회베이지색 저고리와 앞매듭도 이어진다. 솥 안의 국물과 닭고기, 토기, 자갈, 밤의 호수와 불빛이 보인다. 옥반지는 노출되지 않으며 읽을 수 있는 글자는 없다.",
        "hard_violations": [],
        "physics": "머리와 목, 어깨 및 상체가 자연스럽게 연결되어 있다. 낮게 앉은 듯한 자세지만 엉덩이와 발, 실제 지지면은 프레임 밖이므로 접촉 상태를 확인할 수 없다. 솥은 돌 위에, 그릇과 장작은 자갈 위에 놓여 있다. 떠 있는 신체나 지지 없는 물체는 보이지 않는다."
       },
       {
        "label": "B",
        "direction": "몸통은 사선이고 얼굴은 카메라 쪽으로 돌아와 거의 정면을 이룬다. 시선은 렌즈 가까이를 향한다. 화면 오른쪽 입꼬리를 올리고 눈을 조금 좁힌 표정은 피식 웃는 반응으로 읽힌다. 손에 든 물체나 조준 대상은 보이지 않는다.",
        "built_space": "왼쪽에 무쇠솥 1개와 받침돌이 있고 솥 아래 작은 불빛, 뒤쪽 중앙에 별도의 장작불 1곳이 보인다. 토기는 앞쪽 2개, 중앙 1개, 오른쪽 뒤 2개로 총 5개가 식별된다. 호수와 자갈, 오른쪽 대나무는 장소 설정에 맞는다. 솥 전체와 여러 그릇을 넓게 포함한 가슴 위 구도로, 요구된 얼굴 클로즈업보다 환경 설명의 비중이 크다.",
        "entities": "동아시아계 남성 1명이며 검은 머리를 단단한 상투로 묶고 낡은 회베이지색 저고리를 입었다. 그러나 뚜렷한 콧수염과 턱수염 때문에 수염 없는 참조 인물보다 나이 들어 보이고 얼굴 동일성이 약해진다. 무쇠솥, 국물, 토기와 모닥불은 존재한다. 닭고기 조각 여러 개가 그릇이 아니라 자갈 위에 놓여 있어 이전 식사의 소품 상태와 차이가 난다. 옥반지와 읽을 수 있는 글자는 보이지 않는다.",
        "hard_violations": [],
        "physics": "상체를 앞으로 약간 기울인 자세는 앉아서 고개를 돌리는 동작으로 가능하다. 하체와 좌면은 잘려 있어 실제 착석 접촉은 확인되지 않는다. 솥은 돌들이 받치고, 그릇과 닭고기는 자갈에 닿아 있다. 솥의 김과 장작불의 연기는 위로 올라간다. 지지 없이 떠 있는 물체나 불가능한 관절은 보이지 않는다."
       }
      ],
      "all_candidates_fail": true
     },
     "normalized": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 7,
        "verdict_ko": "두 후보 모두 정면 얼굴 클로즈업보다 넓지만, A가 안규홍의 수염 없는 얼굴과 은근한 비대칭 미소, 절제된 모닥불 조명을 더 충실하게 유지한다."
       },
       {
        "label": "A",
        "score": 5,
        "verdict_ko": "한쪽 입꼬리를 올린 표정은 보이나, 넓은 식사 장면 구도와 참조에 없는 콧수염·턱수염이 핵심 구도 및 인물 연속성을 떨어뜨린다."
       }
      ],
      "all_candidates_fail": true,
      "readings": [
       {
        "label": "B",
        "direction": "얼굴은 거의 정면이고 몸통은 비스듬하다. 두 눈은 카메라 가까이를 바라보며, 화면 오른쪽 입꼬리가 조금 더 올라간 미소가 보인다. 어이없다는 반응보다는 다정한 미소에 가깝다. 방향을 따질 무기나 손에 든 물체는 보이지 않는다.",
        "built_space": "왼쪽 전경에 화면 밖으로 잘린 무쇠솥 1개와 그 아래 받침돌, 왼쪽 가장자리에 장작불 1곳이 보인다. 인물 왼쪽 뒤 자갈 위에는 큰 토기 그릇 1개와 작은 그릇 1개가 보인다. 호수는 왼쪽 배경, 대나무와 천막 일부는 오른쪽 배경에 있어 이전 장소의 재료와 배치를 대체로 유지한다. 다만 가슴과 넓은 주변 공간까지 담아 얼굴 중심의 정면 클로즈업보다 넓다.",
        "entities": "한국인 청년 남성으로 설정된 인물 1명만 보인다. 검은 머리를 뒤로 팽팽하게 넘겼으며 묶음 자체는 뚜렷하지 않다. 수염 없는 얼굴과 눈·코·턱의 인상은 참조에 비교적 가깝고, 때 탄 회베이지색 저고리와 앞매듭도 이어진다. 솥 안의 국물과 닭고기, 토기, 자갈, 밤의 호수와 불빛이 보인다. 옥반지는 노출되지 않으며 읽을 수 있는 글자는 없다.",
        "hard_violations": [],
        "physics": "머리와 목, 어깨 및 상체가 자연스럽게 연결되어 있다. 낮게 앉은 듯한 자세지만 엉덩이와 발, 실제 지지면은 프레임 밖이므로 접촉 상태를 확인할 수 없다. 솥은 돌 위에, 그릇과 장작은 자갈 위에 놓여 있다. 떠 있는 신체나 지지 없는 물체는 보이지 않는다."
       },
       {
        "label": "A",
        "direction": "몸통은 사선이고 얼굴은 카메라 쪽으로 돌아와 거의 정면을 이룬다. 시선은 렌즈 가까이를 향한다. 화면 오른쪽 입꼬리를 올리고 눈을 조금 좁힌 표정은 피식 웃는 반응으로 읽힌다. 손에 든 물체나 조준 대상은 보이지 않는다.",
        "built_space": "왼쪽에 무쇠솥 1개와 받침돌이 있고 솥 아래 작은 불빛, 뒤쪽 중앙에 별도의 장작불 1곳이 보인다. 토기는 앞쪽 2개, 중앙 1개, 오른쪽 뒤 2개로 총 5개가 식별된다. 호수와 자갈, 오른쪽 대나무는 장소 설정에 맞는다. 솥 전체와 여러 그릇을 넓게 포함한 가슴 위 구도로, 요구된 얼굴 클로즈업보다 환경 설명의 비중이 크다.",
        "entities": "동아시아계 남성 1명이며 검은 머리를 단단한 상투로 묶고 낡은 회베이지색 저고리를 입었다. 그러나 뚜렷한 콧수염과 턱수염 때문에 수염 없는 참조 인물보다 나이 들어 보이고 얼굴 동일성이 약해진다. 무쇠솥, 국물, 토기와 모닥불은 존재한다. 닭고기 조각 여러 개가 그릇이 아니라 자갈 위에 놓여 있어 이전 식사의 소품 상태와 차이가 난다. 옥반지와 읽을 수 있는 글자는 보이지 않는다.",
        "hard_violations": [],
        "physics": "상체를 앞으로 약간 기울인 자세는 앉아서 고개를 돌리는 동작으로 가능하다. 하체와 좌면은 잘려 있어 실제 착석 접촉은 확인되지 않는다. 솥은 돌들이 받치고, 그릇과 닭고기는 자갈에 닿아 있다. 솥의 김과 장작불의 연기는 위로 올라간다. 지지 없이 떠 있는 물체나 불가능한 관절은 보이지 않는다."
       }
      ]
     },
     "ok": true
    }
   ],
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "slot_winner_match": false,
   "slot_winner": {
    "gemini-pro": "A",
    "gpt-high": "B"
   },
   "route": "cross_slot_combined"
  },
  "dual": {
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "normalized": {
    "A": 1.714,
    "B": 1.75
   },
   "adjusted": {
    "A": 1.714,
    "B": 1.75
   },
   "violations": {},
   "per_model_winner": {
    "gemini-pro": "A",
    "gpt-high": "B"
   },
   "agreed": false
  },
  "totals": {
   "A": 1714,
   "B": 1750
  },
  "selected": "B",
  "ranking": [
   "B",
   "A"
  ],
  "verdicts": [
   {
    "label": "A",
    "score": 1714,
    "verdict_ko": "지시된 '어이가 없다는 듯 한쪽 입꼬리를 올린' 표정을 매우 정확하게 묘사했으나, 바닥에 놓인 닭고기 조각들이 다소 어색합니다."
   },
   {
    "label": "B",
    "score": 1750,
    "verdict_ko": "요구된 '어이가 없다는 듯한 피식 웃음'보다는 대칭적이고 온화한 미소에 가까워 샷 텍스트의 핵심 감정 표현을 살리지 못했습니다."
   }
  ],
  "refs": [
   {
    "label": "PREVIOUS SHOT STILL — a visually related earlier shot of this same place: the location's look, materials, fixed features and lighting mood are LOCKED to this photo; never copy its camera framing. The clothing, hair and overall look of 안규홍 — who appear both in that photo and in this shot — are LOCKED to that photo. Anyone else visible in that photo is NOT in this shot: never carry their face, body or clothing onto anyone here. If this photo shows a character who cannot move (dead or unconscious), that character's exact pose, position and orientation are ALSO LOCKED — treat their body as a fixed prop of the set that only the camera moves around.",
    "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/db00517b-165b-4005-a287-f28936b59b95/images/272bd7af-9294-4490-98c6-88a40e9deacc/scene/recipe/S15sh8_sel.png",
    "asset_id": "46a1b789-f37b-4ac8-8665-ee1a7b580dfe",
    "role": "prev_still"
   },
   {
    "label": "CHARACTER REFERENCE — 안규홍: the exact person appearing in this shot; match face, hair and build exactly.",
    "path": "<bytes:1028384>",
    "asset_id": "3b5894db-ce55-4f34-b966-5e725c3aa952",
    "role": "character_ref"
   }
  ],
  "critique_skipped": true,
  "shot_run_produced": false,
  "shot_run_uid": "06aa140f-2f30-7867-8a7c-d3b949f366c4",
  "ref_mode": "prev+엔티티",
  "share_plan": {
   "ref_plan": "prev",
   "prev_anchor_tag": "S15sh8"
  }
 },
 "S15sh12::cine": {
  "applied": true,
  "attempted_at": "2026-09-09T02:28:38.350113+00:00",
  "fingerprint": "9c402a80115460da6e57528b0a08b12a3abe9a02c6d4b67407babf08241a2d5b",
  "fingerprint_version": 2,
  "provider": "mai",
  "endpoint": "openrouter/chat-completions",
  "model": "microsoft/mai-image-2.6",
  "pack": "24.202608252115",
  "source_file": "S15sh12_sel.png",
  "source_sha256": "0040fbfbb836a12db103d7c42791c6c2a6b0b94a7343dde9e86a5f348d9d1d5a",
  "file": "S15sh12_cine.png",
  "staged_sha256": "28a7f84eecb0ac9ea7a521574bdd02f8b4e0ab02d201ed0961efede9a48851e2",
  "latency_ms": 26506
 },
 "S16sh4::signage": {
  "fp": "3464185f41a9c97a",
  "inscriptions": [],
  "cues": [],
  "dropped": []
 },
 "era_assess::b917082a59167e18": {
  "subjects": [
   {
    "subject_native": "20세기 초 전라도 전통 민가의 광(곡간)",
    "search_terms_native": [
     "전통 민가 광",
     "초가집 곡간",
     "조선 후기 농가 헛간 광채",
     "전라도 전통 가옥 고방"
    ],
    "language_lock_native": "반드시 한국어로만 검색해야 하며 다른 언어로 번역하거나 다른 언어의 단어를 추가하지 마십시오.",
    "reason_ko": "전통 초가 농가의 흙벽과 판자문으로 된 광 구조 대신 서양식 헛간이나 현대식 창고 형태로 생성되기 쉽습니다."
   }
  ],
  "subject_text": "민가 마당·광 입구\n닭장과 곡식 광의 출입문이 있는 농촌 흙마당. 광 안쪽으로 곡식 자루가 놓이며, 건물과 마당이 바로 맞닿아 있다.",
  "identity": "canonical",
  "scope_id": "L17",
  "scope_role": "location_exterior",
  "scope_sha": "f00615cc78bcc23b"
 },
 "era_fail::8ae00cb52d141ecd": {
  "stage": "research",
  "subject": "20세기 초 전라도 전통 민가의 광(곡간)",
  "terms": [
   "전통 민가 광",
   "초가집 곡간",
   "조선 후기 농가 헛간 광채",
   "전라도 전통 가옥 고방"
  ],
  "status": "no_candidates",
  "queries": [
   [
    "전라도 전통 민가 고방 광 곡간",
    "전라도 초가 광채 헛간 민가 곡간"
   ],
   [
    "\"곳간\" \"한국민족문화대백과사전\" \"고방\"",
    "\"남부지방 전통민가 곳간의 건축특성\"",
    "\"낙안\" \"광채\"",
    "\"전라도\" \"고방\" \"민가\""
   ]
  ],
  "attempts": 5
 },
 "groupbg::farm_granary_yard": {
  "input_fingerprint": "4fa337bb87b3ef0b",
  "meta": {
   "model": "gpt-image-2",
   "size": "1536x864",
   "pack": "11.202607220237",
   "contract": "bgfirst_full_v3",
   "group_sig": {
    "key": "farm_granary_yard",
    "tags": [
     "S16sh16",
     "S16sh4",
     "S16sh7"
    ]
   },
   "context_sig": "fcf031db037884a1"
  },
  "prompt": "Create ONE empty live-action location background photograph — NO PEOPLE, no figures, no body parts, no silhouettes, no shadows or reflections of people anywhere.\n\"Empty\" means no people only: KEEP the location's inherent occupants and stock that define the place — animals in an animal shelter, pen or farm, goods and produce in a market, moored boats in a harbour — unless the shot text explicitly removes them.\nThe FIRST attached image is a thin-line storyboard sketch of a shot that happens at this location: use it ONLY as spatial evidence — what this place contains, how its ground, structures and landmarks are arranged and proportioned. Ignore the sketched people and arrows entirely, and do NOT copy its line style: render a fully photographic, physically plausible real place that fits THE LOCATION text below.\nHUMAN-SCALE CALIBRATION: derive every structure's true size from human-scale elements — a door ≈ 2m, a window ≈ 1–1.5m wide, one storey ≈ 2.5–3m; never inflate a small structure or shrink a large one.\n\nTHE LOCATION — early 20th-century Korea, Jeolla region; all people are Korean unless explicitly stated otherwise: The dirt yard immediately outside a rural home's storehouse entrance, where a seed-grain sack has been dragged out.\n\nLOCATION DETAIL (descriptive evidence about this place from the production's location records — evidence, not a staging order): use it to understand what kind of place this is and what it permanently contains. Carry over ONLY its enduring physical features — layout, structures, ground and materials, fixed equipment and installations. Do NOT treat any lighting, weather or time-of-day wording, momentary object states, depicted events or subjective impressions in it as fixed facts of the place; the TIME OF DAY lock below is the sole authority for time and lighting.\n민가 마당·광 입구: 일반 농가의 흙바닥 마당. 한편에 나무로 얽어 만든 닭장이 있고, 안쪽으로 곡식을 보관하는 어두운 광(창고) 입구가 열려 있다. (특징: 농가 흙바닥 마당; 목재 구조의 닭장; 문이 열린 나무 광(창고) 입구; 단단히 묶인 작은 씨나락 자루들)\n\nSCENE EVIDENCE (verbatim quotes from the screenplay about this place — treat them as evidence of what the location physically contains and looks like; stage the PLACE those moments happen in, but do NOT depict the momentary actions, people or staged props themselves):\n- 의병 둘이 광 안에서 작고 단단히 묶인 씨나락 자루를 끌어내고 있다.\n- 늙은 노파가 흙바닥에 주저앉아 그 자루를 붙잡으며 통곡한다.\n- 무릎이 꺾이며 흙바닥에 쓰러진다.\n흐려지는 시야 너머로 씨나락 자루가 멀어진다.\n\nTIME OF DAY (lock): day.\n\nRender ONE photorealistic empty location photograph, 16:9, neutral enough that every shot of this place can be staged from it later. No readable writing anywhere: surfaces that would carry writing may be present, but stage any wording out of legibility — an oblique angle, distance, shallow focus. No captions, watermarks or overlay text.",
  "effective_prompt": "Create ONE empty live-action location background photograph — NO PEOPLE, no figures, no body parts, no silhouettes, no shadows or reflections of people anywhere.\n\"Empty\" means no people only: KEEP the location's inherent occupants and stock that define the place — animals in an animal shelter, pen or farm, goods and produce in a market, moored boats in a harbour — unless the shot text explicitly removes them.\nThe FIRST attached image is a thin-line storyboard sketch of a shot that happens at this location: use it ONLY as spatial evidence — what this place contains, how its ground, structures and landmarks are arranged and proportioned. Ignore the sketched people and arrows entirely, and do NOT copy its line style: render a fully photographic, physically plausible real place that fits THE LOCATION text below.\nHUMAN-SCALE CALIBRATION: derive every structure's true size from human-scale elements — a door ≈ 2m, a window ≈ 1–1.5m wide, one storey ≈ 2.5–3m; never inflate a small structure or shrink a large one.\n\nTHE LOCATION — early 20th-century Korea, Jeolla region; all people are Korean unless explicitly stated otherwise: The dirt yard immediately outside a rural home's storehouse entrance, where a seed-grain sack has been dragged out.\n\nLOCATION DETAIL (descriptive evidence about this place from the production's location records — evidence, not a staging order): use it to understand what kind of place this is and what it permanently contains. Carry over ONLY its enduring physical features — layout, structures, ground and materials, fixed equipment and installations. Do NOT treat any lighting, weather or time-of-day wording, momentary object states, depicted events or subjective impressions in it as fixed facts of the place; the TIME OF DAY lock below is the sole authority for time and lighting.\n민가 마당·광 입구: 일반 농가의 흙바닥 마당. 한편에 나무로 얽어 만든 닭장이 있고, 안쪽으로 곡식을 보관하는 어두운 광(창고) 입구가 열려 있다. (특징: 농가 흙바닥 마당; 목재 구조의 닭장; 문이 열린 나무 광(창고) 입구; 단단히 묶인 작은 씨나락 자루들)\n\nSCENE EVIDENCE (verbatim quotes from the screenplay about this place — treat them as evidence of what the location physically contains and looks like; stage the PLACE those moments happen in, but do NOT depict the momentary actions, people or staged props themselves):\n- 의병 둘이 광 안에서 작고 단단히 묶인 씨나락 자루를 끌어내고 있다.\n- 늙은 노파가 흙바닥에 주저앉아 그 자루를 붙잡으며 통곡한다.\n- 무릎이 꺾이며 흙바닥에 쓰러진다.\n흐려지는 시야 너머로 씨나락 자루가 멀어진다.\n\nTIME OF DAY (lock): day.\n\nRender ONE photorealistic empty location photograph, 16:9, neutral enough that every shot of this place can be staged from it later. No readable writing anywhere: surfaces that would carry writing may be present, but stage any wording out of legibility — an oblique angle, distance, shallow focus. No captions, watermarks or overlay text.",
  "bg_path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/db00517b-165b-4005-a287-f28936b59b95/images/272bd7af-9294-4490-98c6-88a40e9deacc/scene/recipe/groupbg_farm_granary_yard_841469.png",
  "asset_id": "478efbd2-93b2-4c89-a843-9c8c194e1849",
  "input_asset_ids": [
   "480fe76d-33b7-4385-8934-4a59175b9569"
  ],
  "origin_tag": "S16sh4",
  "place_text": "The dirt yard immediately outside a rural home's storehouse entrance, where a seed-grain sack has been dragged out.",
  "origin_inputs": {
   "place_text": "The dirt yard immediately outside a rural home's storehouse entrance, where a seed-grain sack has been dragged out.",
   "time_of_day_en": "day",
   "conti_asset_id": "480fe76d-33b7-4385-8934-4a59175b9569"
  }
 },
 "S16sh4::bgfirst_bg": {
  "input_fingerprint": "e989b3dded351281",
  "prompt": "Create the EMPTY BACKGROUND PLATE for one film shot — NO PEOPLE, no figures, no body parts, no sketch lines, no arrows anywhere.\n\"Empty\" means no people only: KEEP the location's inherent occupants and stock that define the place — animals in an animal shelter, pen or farm, goods in a market, moored boats in a harbour — unless the shot text explicitly removes them.\nThe FIRST attached image is a thin-line storyboard sketch: use ONLY its camera angle, horizon, perspective and the placement/size of buildings and set masses — ignore the sketched people and arrows entirely. The SECOND attached image (LOCATION PHOTOGRAPH) is the real place: take its architecture, materials, signage and fixed features, and RE-PROJECT them into the sketch's camera. If the photograph's camera differs from the sketch's, the sketch's camera wins.\nHUMAN-SCALE CALIBRATION: derive every structure's true size from human-scale elements — a door ≈ 2m, a window ≈ 1–1.5m wide, one storey ≈ 2.5–3m; never inflate a small structure or shrink a large one.\n\nSHOT TEXT this background must serve (Korean): 흙바닥에 주저앉은 채 씨나락 자루를 양팔로 필사적으로 끌어안고 오열하는 조선인 노파의 애처로운 웅크림.\n\nLOCATION (lock): The dirt yard immediately outside a rural home's storehouse entrance, where a seed-grain sack has been dragged out.\n\nTIME OF DAY (lock): day.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: medium shot\n- KEY BACKGROUND ELEMENTS: 씨나락 자루 (Small, tightly tied, and held with both arms by the elderly woman) — Its side and tied upper portion are visible beneath her embrace; used as Connects her curled body to the threatened seed grain while retaining realistic scale; 흙바닥 (The woman is seated directly on it); used as Surrounds the body with unprotected ground; 광 입구 (The seed-grain sack has been dragged out from inside) — A partial oblique view of the entrance remains behind the woman; used as Locates the seizure without introducing additional action into the frame.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Daylight with sober naturalistic contrast and muted earth tones exposes the distress plainly, without beautifying the ground or the woman's suffering.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nRender ONE photorealistic empty location photograph, 16:9, that this shot can be staged inside later. No readable writing anywhere: surfaces that would carry writing may be present, but stage any wording out of legibility — an oblique angle, distance, shallow focus. No captions, watermarks or overlay text.",
  "effective_prompt": "Create the EMPTY BACKGROUND PLATE for one film shot — NO PEOPLE, no figures, no body parts, no sketch lines, no arrows anywhere.\n\"Empty\" means no people only: KEEP the location's inherent occupants and stock that define the place — animals in an animal shelter, pen or farm, goods in a market, moored boats in a harbour — unless the shot text explicitly removes them.\nThe FIRST attached image is a thin-line storyboard sketch: use ONLY its camera angle, horizon, perspective and the placement/size of buildings and set masses — ignore the sketched people and arrows entirely. The SECOND attached image (LOCATION PHOTOGRAPH) is the real place: take its architecture, materials, signage and fixed features, and RE-PROJECT them into the sketch's camera. If the photograph's camera differs from the sketch's, the sketch's camera wins.\nHUMAN-SCALE CALIBRATION: derive every structure's true size from human-scale elements — a door ≈ 2m, a window ≈ 1–1.5m wide, one storey ≈ 2.5–3m; never inflate a small structure or shrink a large one.\n\nSHOT TEXT this background must serve (Korean): 흙바닥에 주저앉은 채 씨나락 자루를 양팔로 필사적으로 끌어안고 오열하는 조선인 노파의 애처로운 웅크림.\n\nLOCATION (lock): The dirt yard immediately outside a rural home's storehouse entrance, where a seed-grain sack has been dragged out.\n\nTIME OF DAY (lock): day.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: medium shot\n- KEY BACKGROUND ELEMENTS: 씨나락 자루 (Small, tightly tied, and held with both arms by the elderly woman) — Its side and tied upper portion are visible beneath her embrace; used as Connects her curled body to the threatened seed grain while retaining realistic scale; 흙바닥 (The woman is seated directly on it); used as Surrounds the body with unprotected ground; 광 입구 (The seed-grain sack has been dragged out from inside) — A partial oblique view of the entrance remains behind the woman; used as Locates the seizure without introducing additional action into the frame.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Daylight with sober naturalistic contrast and muted earth tones exposes the distress plainly, without beautifying the ground or the woman's suffering.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nRender ONE photorealistic empty location photograph, 16:9, that this shot can be staged inside later. No readable writing anywhere: surfaces that would carry writing may be present, but stage any wording out of legibility — an oblique angle, distance, shallow focus. No captions, watermarks or overlay text.",
  "bg_path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/db00517b-165b-4005-a287-f28936b59b95/images/272bd7af-9294-4490-98c6-88a40e9deacc/scene/recipe/S16sh4__bgfirst_bg.png",
  "asset_id": "18c31ecd-1fb4-4b09-bc3b-91a5d20f0ffb",
  "input_asset_ids": [
   "480fe76d-33b7-4385-8934-4a59175b9569",
   "478efbd2-93b2-4c89-a843-9c8c194e1849"
  ]
 },
 "S16sh4": {
  "input_fingerprint": "57206d4dd2df3938",
  "prompt": "Create ONE FINAL photorealistic live-action film still of the moment below — early 20th-century Korea, Jeolla region; all people are Korean unless explicitly stated otherwise. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 흙바닥에 주저앉은 채 씨나락 자루를 양팔로 필사적으로 끌어안고 오열하는 조선인 노파의 애처로운 웅크림.\n\nLOCATION (lock): The dirt yard immediately outside a rural home's storehouse entrance, where a seed-grain sack has been dragged out. The shot takes place here — the attached LOCATION PHOTOGRAPH shows the exact spot.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: medium shot\n- KEY BACKGROUND ELEMENTS: 씨나락 자루 (Small, tightly tied, and held with both arms by the elderly woman) — Its side and tied upper portion are visible beneath her embrace; used as Connects her curled body to the threatened seed grain while retaining realistic scale; 흙바닥 (The woman is seated directly on it); used as Surrounds the body with unprotected ground; 광 입구 (The seed-grain sack has been dragged out from inside) — A partial oblique view of the entrance remains behind the woman; used as Locates the seizure without introducing additional action into the frame.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Daylight with sober naturalistic contrast and muted earth tones exposes the distress plainly, without beautifying the ground or the woman's suffering.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): A small, tightly tied seed-rice sack has been dragged out of the storehouse onto the dirt yard. 조선인 노파(씨나락 주인): She sits on the dirt clutching the tied seed-rice sack.\n\nTHE HAND THAT IS DOING THIS: the object at the centre of this shot is being held, operated, read, aimed or handed over by 조선인 노파 right now, so 조선인 노파's hand — and as much of the wrist and forearm as the framing reaches — is in the frame, gripping or resting on that object exactly the way its real use requires. Match that hand to 조선인 노파: its size, build, skin, age, grooming, sleeve and anything worn on it belong to that person and to no one else.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 조선인 노파(씨나락 주인) (한국인 여성, 노년 나이대, 자연스러운 노년의 얼굴) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.",
  "roll_prompts": {
   "A": "Stage the shot. The FIRST attached image (SHOT BACKGROUND) is the finished empty background of this shot — keep it EXACTLY: its camera, perspective, architecture, lighting and every fixed feature stay untouched. The SECOND attached image (LAYOUT SKETCH) tells you ONLY where the people go: each sketched person's position, screen size, pose and the gaze/motion arrows. Ignore the sketch's background lines. The CHARACTER REFERENCE photographs show the real people.\nPlace the real people into the background at exactly the sketched positions, sizes and poses, following the arrow directions. No sketch lines or arrows may remain.\n\nCreate ONE FINAL photorealistic live-action film still of the moment below — early 20th-century Korea, Jeolla region; all people are Korean unless explicitly stated otherwise. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 흙바닥에 주저앉은 채 씨나락 자루를 양팔로 필사적으로 끌어안고 오열하는 조선인 노파의 애처로운 웅크림.\n\nLOCATION (lock): The dirt yard immediately outside a rural home's storehouse entrance, where a seed-grain sack has been dragged out. The shot takes place here — the FIRST attached image (SHOT BACKGROUND) is this exact place, already built: its ground, structures, horizon, materials and lighting are the finished truth of this location and must not be redesigned or replaced. No location photograph is attached — read the place from that image alone, and add no scenery, structure, vehicle or fixture that it does not already show. This lock governs the place only; the figures in the shot follow the staging and pose instructions.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: medium shot\n- KEY BACKGROUND ELEMENTS: 씨나락 자루 (Small, tightly tied, and held with both arms by the elderly woman) — Its side and tied upper portion are visible beneath her embrace; used as Connects her curled body to the threatened seed grain while retaining realistic scale; 흙바닥 (The woman is seated directly on it); used as Surrounds the body with unprotected ground; 광 입구 (The seed-grain sack has been dragged out from inside) — A partial oblique view of the entrance remains behind the woman; used as Locates the seizure without introducing additional action into the frame.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Daylight with sober naturalistic contrast and muted earth tones exposes the distress plainly, without beautifying the ground or the woman's suffering.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): A small, tightly tied seed-rice sack has been dragged out of the storehouse onto the dirt yard. 조선인 노파(씨나락 주인): She sits on the dirt clutching the tied seed-rice sack.\n\nTHE HAND THAT IS DOING THIS: the object at the centre of this shot is being held, operated, read, aimed or handed over by 조선인 노파 right now, so 조선인 노파's hand — and as much of the wrist and forearm as the framing reaches — is in the frame, gripping or resting on that object exactly the way its real use requires. Match that hand to 조선인 노파: its size, build, skin, age, grooming, sleeve and anything worn on it belong to that person and to no one else.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 조선인 노파(씨나락 주인) (한국인 여성, 노년 나이대, 자연스러운 노년의 얼굴) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.",
   "B": "Create ONE FINAL photorealistic live-action film still of the moment below — early 20th-century Korea, Jeolla region; all people are Korean unless explicitly stated otherwise. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 흙바닥에 주저앉은 채 씨나락 자루를 양팔로 필사적으로 끌어안고 오열하는 조선인 노파의 애처로운 웅크림.\n\nLOCATION (lock): The dirt yard immediately outside a rural home's storehouse entrance, where a seed-grain sack has been dragged out. The shot takes place here — the attached LOCATION PHOTOGRAPH shows the exact spot.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: medium shot\n- KEY BACKGROUND ELEMENTS: 씨나락 자루 (Small, tightly tied, and held with both arms by the elderly woman) — Its side and tied upper portion are visible beneath her embrace; used as Connects her curled body to the threatened seed grain while retaining realistic scale; 흙바닥 (The woman is seated directly on it); used as Surrounds the body with unprotected ground; 광 입구 (The seed-grain sack has been dragged out from inside) — A partial oblique view of the entrance remains behind the woman; used as Locates the seizure without introducing additional action into the frame.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Daylight with sober naturalistic contrast and muted earth tones exposes the distress plainly, without beautifying the ground or the woman's suffering.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): A small, tightly tied seed-rice sack has been dragged out of the storehouse onto the dirt yard. 조선인 노파(씨나락 주인): She sits on the dirt clutching the tied seed-rice sack.\n\nTHE HAND THAT IS DOING THIS: the object at the centre of this shot is being held, operated, read, aimed or handed over by 조선인 노파 right now, so 조선인 노파's hand — and as much of the wrist and forearm as the framing reaches — is in the frame, gripping or resting on that object exactly the way its real use requires. Match that hand to 조선인 노파: its size, build, skin, age, grooming, sleeve and anything worn on it belong to that person and to no one else.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 조선인 노파(씨나락 주인) (한국인 여성, 노년 나이대, 자연스러운 노년의 얼굴) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.\n\nCOMPOSITION VARIATION (this candidate only): keep the place, the people and the moment exactly as specified, and keep from the CAMERA & FRAME contract its subject scale, screen placement and key background elements exactly. For this candidate only, its camera angle and height are deliberately left open: choose a camera position distinctly different from the obvious one, as a film director picking a second setup on the same blocking. Stay true to any explicit framing the SHOT TEXT itself dictates."
  },
  "roll_refs": {
   "A": [
    {
     "label": "SHOT BACKGROUND",
     "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/db00517b-165b-4005-a287-f28936b59b95/images/272bd7af-9294-4490-98c6-88a40e9deacc/scene/recipe/S16sh4__bgfirst_bg.png",
     "asset_id": "18c31ecd-1fb4-4b09-bc3b-91a5d20f0ffb",
     "role": "bgfirst_bg"
    },
    {
     "label": "LAYOUT SKETCH (people placement only)",
     "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/db00517b-165b-4005-a287-f28936b59b95/images/272bd7af-9294-4490-98c6-88a40e9deacc/conti/conti_S16sh4.png",
     "asset_id": "480fe76d-33b7-4385-8934-4a59175b9569",
     "role": "conti_light"
    },
    {
     "label": "CHARACTER REFERENCE — 조선인 노파(씨나락 주인): the exact person appearing in this shot; match face, hair and build exactly.",
     "path": "<bytes:982368>",
     "asset_id": "aa993fab-d06c-423d-8316-10d12c20f3c0",
     "role": "character_ref"
    },
    {
     "label": "PROP REFERENCE — 곡식 자루·가마니: the exact object appearing in this shot; match its look, material and wear exactly.",
     "path": "<bytes:1807398>",
     "asset_id": "e30c319e-fdc7-4c01-a301-290ddd45d27a",
     "role": "prop_ref"
    }
   ],
   "B": [
    {
     "label": "LOCATION PHOTOGRAPH — the exact place of this shot: its architecture, materials, fixed features and lighting mood are spatial truth; stage the moment inside this place. Never copy its camera framing.",
     "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/db00517b-165b-4005-a287-f28936b59b95/images/272bd7af-9294-4490-98c6-88a40e9deacc/scene/recipe/groupbg_farm_granary_yard_841469.png",
     "asset_id": "478efbd2-93b2-4c89-a843-9c8c194e1849",
     "role": "bgfirst_group_bg"
    },
    {
     "label": "CHARACTER REFERENCE — 조선인 노파(씨나락 주인): the exact person appearing in this shot; match face, hair and build exactly.",
     "path": "<bytes:982368>",
     "asset_id": "aa993fab-d06c-423d-8316-10d12c20f3c0",
     "role": "character_ref"
    },
    {
     "label": "PROP REFERENCE — 곡식 자루·가마니: the exact object appearing in this shot; match its look, material and wear exactly.",
     "path": "<bytes:1807398>",
     "asset_id": "e30c319e-fdc7-4c01-a301-290ddd45d27a",
     "role": "prop_ref"
    }
   ]
  },
  "shot_run_spend_attempt_count": 1,
  "judge_flip_skipped": "judge_fn_owns_order",
  "readings": [
   {
    "label": "A",
    "direction": "노파의 시선과 고개는 품에 안은 씨나락 자루를 향해 아래로 숙여져 있음.",
    "built_space": "흙바닥 위에 노파가 위치하며, 좌측에 닭장, 우측 후경에 열린 광 입구가 올바르게 배치됨.",
    "entities": "레퍼런스와 일치하는 의상과 외모의 노파(비녀 제외), 노끈으로 묶인 씨나락 자루가 확인됨.",
    "hard_violations": [],
    "physics": "노파는 흙바닥에 안정적으로 앉아 있으며, 양팔과 손으로 자루를 자연스럽게 지탱하고 있음."
   },
   {
    "label": "B",
    "direction": "노파의 시선은 정면을 향하며 입을 벌려 오열하고 있음.",
    "built_space": "흙바닥 위, 좌측 닭장, 우측 후경의 광 입구 등 지정된 공간 요소가 모두 존재함.",
    "entities": "레퍼런스와 일치하는 노파(비녀 포함)와 씨나락 자루가 보이나, 자루 아래에 불필요한 미니 자루가 있음.",
    "hard_violations": [
     "[gemini-pro] invented or extra bodies: 자루를 안고 있는 노파의 손이 3개로 생성됨(물리적으로 불가능한 해부학 구조).",
     "[gpt-high] 노파가 안은 자루 외에 화면 하단 전경에 두 번째 묶인 자루를 추가하여, 마당으로 끌려 나온 작은 씨나락 자루 하나라는 소품 구성을 위반했습니다."
    ],
    "physics": "노파는 바닥에 앉아 있으나, 세 개의 손이 자루를 잡고 있는 불가능한 자세를 취하고 있음."
   }
  ],
  "cross_model_order": {
   "policy": "cross_model_order_v2_slot0_fwd_slot1_rev_parallel2",
   "slots": [
    {
     "model": "gemini-pro",
     "order": "forward",
     "display_to_canonical": {
      "A": "A",
      "B": "B"
     },
     "raw": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 7,
        "verdict_ko": "프롬프트의 연출과 공간, 인물을 무난하게 반영하였으며 물리적 오류 없이 애처로운 분위기를 잘 표현함."
       },
       {
        "label": "B",
        "score": 3,
        "verdict_ko": "인물과 공간의 묘사는 우수하나, 신체 부위가 중복 생성되는 치명적인 오류가 발생함."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "노파의 시선과 고개는 품에 안은 씨나락 자루를 향해 아래로 숙여져 있음.",
        "built_space": "흙바닥 위에 노파가 위치하며, 좌측에 닭장, 우측 후경에 열린 광 입구가 올바르게 배치됨.",
        "entities": "레퍼런스와 일치하는 의상과 외모의 노파(비녀 제외), 노끈으로 묶인 씨나락 자루가 확인됨.",
        "hard_violations": [],
        "physics": "노파는 흙바닥에 안정적으로 앉아 있으며, 양팔과 손으로 자루를 자연스럽게 지탱하고 있음."
       },
       {
        "label": "B",
        "direction": "노파의 시선은 정면을 향하며 입을 벌려 오열하고 있음.",
        "built_space": "흙바닥 위, 좌측 닭장, 우측 후경의 광 입구 등 지정된 공간 요소가 모두 존재함.",
        "entities": "레퍼런스와 일치하는 노파(비녀 포함)와 씨나락 자루가 보이나, 자루 아래에 불필요한 미니 자루가 있음.",
        "hard_violations": [
         "invented or extra bodies: 자루를 안고 있는 노파의 손이 3개로 생성됨(물리적으로 불가능한 해부학 구조)."
        ],
        "physics": "노파는 바닥에 앉아 있으나, 세 개의 손이 자루를 잡고 있는 불가능한 자세를 취하고 있음."
       }
      ],
      "all_candidates_fail": false
     },
     "normalized": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 7,
        "verdict_ko": "프롬프트의 연출과 공간, 인물을 무난하게 반영하였으며 물리적 오류 없이 애처로운 분위기를 잘 표현함."
       },
       {
        "label": "B",
        "score": 3,
        "verdict_ko": "인물과 공간의 묘사는 우수하나, 신체 부위가 중복 생성되는 치명적인 오류가 발생함."
       }
      ],
      "all_candidates_fail": false,
      "readings": [
       {
        "label": "A",
        "direction": "노파의 시선과 고개는 품에 안은 씨나락 자루를 향해 아래로 숙여져 있음.",
        "built_space": "흙바닥 위에 노파가 위치하며, 좌측에 닭장, 우측 후경에 열린 광 입구가 올바르게 배치됨.",
        "entities": "레퍼런스와 일치하는 의상과 외모의 노파(비녀 제외), 노끈으로 묶인 씨나락 자루가 확인됨.",
        "hard_violations": [],
        "physics": "노파는 흙바닥에 안정적으로 앉아 있으며, 양팔과 손으로 자루를 자연스럽게 지탱하고 있음."
       },
       {
        "label": "B",
        "direction": "노파의 시선은 정면을 향하며 입을 벌려 오열하고 있음.",
        "built_space": "흙바닥 위, 좌측 닭장, 우측 후경의 광 입구 등 지정된 공간 요소가 모두 존재함.",
        "entities": "레퍼런스와 일치하는 노파(비녀 포함)와 씨나락 자루가 보이나, 자루 아래에 불필요한 미니 자루가 있음.",
        "hard_violations": [
         "invented or extra bodies: 자루를 안고 있는 노파의 손이 3개로 생성됨(물리적으로 불가능한 해부학 구조)."
        ],
        "physics": "노파는 바닥에 앉아 있으나, 세 개의 손이 자루를 잡고 있는 불가능한 자세를 취하고 있음."
       }
      ]
     },
     "ok": true
    },
    {
     "model": "gpt-high",
     "order": "reverse",
     "display_to_canonical": {
      "A": "B",
      "B": "A"
     },
     "raw": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 3,
        "verdict_ko": "미디엄 숏과 오열 표정은 잘 구현했지만, 화면 하단에 두 번째 묶인 자루를 추가한 중대 위반이 있으며 몸의 웅크림도 약합니다."
       },
       {
        "label": "B",
        "score": 7,
        "verdict_ko": "요구한 미디엄 숏보다 넓지만, 흙바닥에서 작은 자루 하나를 양팔로 감싸고 몸을 접어 오열하는 핵심 행동과 장소를 충실히 구현했습니다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "노파는 눈을 감고 얼굴을 정면 쪽으로 든 채 입을 크게 벌려 울고 있습니다. 시선의 특정 대상은 보이지 않습니다. 양손과 팔은 가슴 앞 자루를 향해 모여 있으나, 고개와 상체가 비교적 세워져 있어 자루 위로 웅크린 동작은 약합니다.",
        "built_space": "오른쪽 뒤에 광 입구 하나, 열린 나무 문짝 하나, 문턱과 낮은 돌 디딤단이 보입니다. 왼쪽에는 닭장 하나와 그 뒤 돌담이 있어 참조 장소의 주요 구조와 배치가 일치합니다. 노파는 문턱 밖 흙마당에 앉아 있으며 시설과 충돌하지 않습니다. 인물이 크게 잡히고 하체가 아래에서 잘려 미디엄 숏 요구에 비교적 가깝습니다.",
        "entities": "노년의 한국인 여성으로 설정에 부합하는 인물 한 명이며, 회색 머리와 비녀, 짙은 깃의 낡은 베이지색 한복이 참조와 가깝습니다. 주름진 두 손도 노파의 팔과 자연스럽게 이어집니다. 품에 안은 자루는 거친 직물과 끈으로 묶인 윗부분이 보이지만 작은 자루치고는 다소 큽니다. 화면 맨 아래에는 별도의 묶인 자루가 하나 더 보입니다. 낮의 흙마당과 자연광이 표현되어 있고 읽을 수 있는 글자는 없습니다.",
        "hard_violations": [
         "노파가 안은 자루 외에 화면 하단 전경에 두 번째 묶인 자루를 추가하여, 마당으로 끌려 나온 작은 씨나락 자루 하나라는 소품 구성을 위반했습니다."
        ],
        "physics": "노파의 하체와 펼쳐진 치마가 흙바닥에 닿아 체중을 받으며, 들어 올린 무릎도 앉은 자세에서 가능합니다. 주 자루는 무릎과 몸통에 기대고 양팔과 손이 눌러 잡아 지지합니다. 추가 자루는 하단이 프레임 밖이라 바닥 접촉을 직접 확인할 수 없지만, 공중에 떠 있다고 볼 근거는 없습니다."
       },
       {
        "label": "B",
        "direction": "노파는 눈을 감고 얼굴을 자루의 묶인 윗부분 쪽으로 깊이 숙여 울고 있습니다. 양팔이 자루 양옆과 아래를 감싸 몸 안쪽으로 끌어당기며, 머리와 몸통의 방향도 보호하려는 자루에 집중됩니다.",
        "built_space": "오른쪽 뒤에 광 입구 하나와 열린 나무 문짝 하나, 문턱과 낮은 돌 디딤단이 보이고, 왼쪽에 닭장 하나와 뒤쪽 돌담이 있습니다. 참조의 목재·흙벽·돌 기초와 공간 배치를 유지합니다. 노파는 광 입구 바로 밖 흙바닥에 앉아 있습니다. 다만 앉은 몸 전체와 주변 마당을 넉넉히 보여 주어, 요구한 미디엄 숏보다는 넓은 구도입니다.",
        "entities": "노년의 한국인 여성으로 설정에 부합하는 인물 한 명이 보입니다. 회색 머리와 자연스러운 노년의 얼굴, 짙은 깃과 베이지색 한복은 참조와 대체로 일치하며, 숙인 머리에서는 참조의 비녀가 뚜렷하지 않습니다. 작은 직물 자루 하나의 옆면과 끈으로 단단히 묶인 상단이 팔 사이로 보입니다. 두 손과 소매는 해당 노파의 신체와 복장에 일관됩니다. 낮의 자연광과 흙빛 색조이며 읽을 수 있는 글자는 없습니다.",
        "hard_violations": [],
        "physics": "치마로 덮인 엉덩이와 접힌 다리가 땅에 놓이고, 왼쪽으로 나온 신발도 지면에 닿아 앉은 몸을 지지합니다. 자루는 무릎 위에서 몸통에 기대며 양손과 팔에 붙잡혀 있습니다. 등을 굽혀 자루 위로 고개를 떨어뜨린 자세는 실제로 가능한 웅크림이며, 지지 없이 떠 있는 신체나 물체는 보이지 않습니다."
       }
      ],
      "all_candidates_fail": false
     },
     "normalized": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 3,
        "verdict_ko": "미디엄 숏과 오열 표정은 잘 구현했지만, 화면 하단에 두 번째 묶인 자루를 추가한 중대 위반이 있으며 몸의 웅크림도 약합니다."
       },
       {
        "label": "A",
        "score": 7,
        "verdict_ko": "요구한 미디엄 숏보다 넓지만, 흙바닥에서 작은 자루 하나를 양팔로 감싸고 몸을 접어 오열하는 핵심 행동과 장소를 충실히 구현했습니다."
       }
      ],
      "all_candidates_fail": false,
      "readings": [
       {
        "label": "B",
        "direction": "노파는 눈을 감고 얼굴을 정면 쪽으로 든 채 입을 크게 벌려 울고 있습니다. 시선의 특정 대상은 보이지 않습니다. 양손과 팔은 가슴 앞 자루를 향해 모여 있으나, 고개와 상체가 비교적 세워져 있어 자루 위로 웅크린 동작은 약합니다.",
        "built_space": "오른쪽 뒤에 광 입구 하나, 열린 나무 문짝 하나, 문턱과 낮은 돌 디딤단이 보입니다. 왼쪽에는 닭장 하나와 그 뒤 돌담이 있어 참조 장소의 주요 구조와 배치가 일치합니다. 노파는 문턱 밖 흙마당에 앉아 있으며 시설과 충돌하지 않습니다. 인물이 크게 잡히고 하체가 아래에서 잘려 미디엄 숏 요구에 비교적 가깝습니다.",
        "entities": "노년의 한국인 여성으로 설정에 부합하는 인물 한 명이며, 회색 머리와 비녀, 짙은 깃의 낡은 베이지색 한복이 참조와 가깝습니다. 주름진 두 손도 노파의 팔과 자연스럽게 이어집니다. 품에 안은 자루는 거친 직물과 끈으로 묶인 윗부분이 보이지만 작은 자루치고는 다소 큽니다. 화면 맨 아래에는 별도의 묶인 자루가 하나 더 보입니다. 낮의 흙마당과 자연광이 표현되어 있고 읽을 수 있는 글자는 없습니다.",
        "hard_violations": [
         "노파가 안은 자루 외에 화면 하단 전경에 두 번째 묶인 자루를 추가하여, 마당으로 끌려 나온 작은 씨나락 자루 하나라는 소품 구성을 위반했습니다."
        ],
        "physics": "노파의 하체와 펼쳐진 치마가 흙바닥에 닿아 체중을 받으며, 들어 올린 무릎도 앉은 자세에서 가능합니다. 주 자루는 무릎과 몸통에 기대고 양팔과 손이 눌러 잡아 지지합니다. 추가 자루는 하단이 프레임 밖이라 바닥 접촉을 직접 확인할 수 없지만, 공중에 떠 있다고 볼 근거는 없습니다."
       },
       {
        "label": "A",
        "direction": "노파는 눈을 감고 얼굴을 자루의 묶인 윗부분 쪽으로 깊이 숙여 울고 있습니다. 양팔이 자루 양옆과 아래를 감싸 몸 안쪽으로 끌어당기며, 머리와 몸통의 방향도 보호하려는 자루에 집중됩니다.",
        "built_space": "오른쪽 뒤에 광 입구 하나와 열린 나무 문짝 하나, 문턱과 낮은 돌 디딤단이 보이고, 왼쪽에 닭장 하나와 뒤쪽 돌담이 있습니다. 참조의 목재·흙벽·돌 기초와 공간 배치를 유지합니다. 노파는 광 입구 바로 밖 흙바닥에 앉아 있습니다. 다만 앉은 몸 전체와 주변 마당을 넉넉히 보여 주어, 요구한 미디엄 숏보다는 넓은 구도입니다.",
        "entities": "노년의 한국인 여성으로 설정에 부합하는 인물 한 명이 보입니다. 회색 머리와 자연스러운 노년의 얼굴, 짙은 깃과 베이지색 한복은 참조와 대체로 일치하며, 숙인 머리에서는 참조의 비녀가 뚜렷하지 않습니다. 작은 직물 자루 하나의 옆면과 끈으로 단단히 묶인 상단이 팔 사이로 보입니다. 두 손과 소매는 해당 노파의 신체와 복장에 일관됩니다. 낮의 자연광과 흙빛 색조이며 읽을 수 있는 글자는 없습니다.",
        "hard_violations": [],
        "physics": "치마로 덮인 엉덩이와 접힌 다리가 땅에 놓이고, 왼쪽으로 나온 신발도 지면에 닿아 앉은 몸을 지지합니다. 자루는 무릎 위에서 몸통에 기대며 양손과 팔에 붙잡혀 있습니다. 등을 굽혀 자루 위로 고개를 떨어뜨린 자세는 실제로 가능한 웅크림이며, 지지 없이 떠 있는 신체나 물체는 보이지 않습니다."
       }
      ]
     },
     "ok": true
    }
   ],
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "slot_winner_match": true,
   "slot_winner": {
    "gemini-pro": "A",
    "gpt-high": "A"
   },
   "route": "cross_slot_agree"
  },
  "dual": {
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "normalized": {
    "A": 2.0,
    "B": 0.857
   },
   "adjusted": {
    "A": 2.0,
    "B": 0.607
   },
   "violations": {
    "B": [
     "[gemini-pro] invented or extra bodies: 자루를 안고 있는 노파의 손이 3개로 생성됨(물리적으로 불가능한 해부학 구조).",
     "[gpt-high] 노파가 안은 자루 외에 화면 하단 전경에 두 번째 묶인 자루를 추가하여, 마당으로 끌려 나온 작은 씨나락 자루 하나라는 소품 구성을 위반했습니다."
    ]
   },
   "per_model_winner": {
    "gemini-pro": "A",
    "gpt-high": "A"
   },
   "agreed": true
  },
  "totals": {
   "A": 2000,
   "B": 607
  },
  "selected": "A",
  "ranking": [
   "A",
   "B"
  ],
  "verdicts": [
   {
    "label": "A",
    "score": 2000,
    "verdict_ko": "프롬프트의 연출과 공간, 인물을 무난하게 반영하였으며 물리적 오류 없이 애처로운 분위기를 잘 표현함."
   },
   {
    "label": "B",
    "score": 607,
    "verdict_ko": "인물과 공간의 묘사는 우수하나, 신체 부위가 중복 생성되는 치명적인 오류가 발생함.  ★위반: [gemini-pro] invented or extra bodies: 자루를 안고 있는 노파의 손이 3개로 생성됨(물리적으로 불가능한 해부학 구조). / [gpt-high] 노파가 안은 자루 외에 화면 하단 전경에 두 번째 묶인 자루를 추가하여, 마당으로 끌려 나온 작은 씨나락 자루 하나라는 소품 구성을 위반했습니다."
   }
  ],
  "refs": [
   {
    "label": "LOCATION PHOTOGRAPH — the exact place of this shot: its architecture, materials, fixed features and lighting mood are spatial truth; stage the moment inside this place. Never copy its camera framing.",
    "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/db00517b-165b-4005-a287-f28936b59b95/images/272bd7af-9294-4490-98c6-88a40e9deacc/scene/recipe/groupbg_farm_granary_yard_841469.png",
    "asset_id": "478efbd2-93b2-4c89-a843-9c8c194e1849",
    "role": "bgfirst_group_bg"
   },
   {
    "label": "CHARACTER REFERENCE — 조선인 노파(씨나락 주인): the exact person appearing in this shot; match face, hair and build exactly.",
    "path": "<bytes:982368>",
    "asset_id": "aa993fab-d06c-423d-8316-10d12c20f3c0",
    "role": "character_ref"
   },
   {
    "label": "PROP REFERENCE — 곡식 자루·가마니: the exact object appearing in this shot; match its look, material and wear exactly.",
    "path": "<bytes:1807398>",
    "asset_id": "e30c319e-fdc7-4c01-a301-290ddd45d27a",
    "role": "prop_ref"
   }
  ],
  "critique_skipped": true,
  "shot_run_produced": false,
  "shot_run_uid": "06aa140f-30b9-79f5-823e-3fd4dead9fc2",
  "bgfirst": {
   "bg_path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/db00517b-165b-4005-a287-f28936b59b95/images/272bd7af-9294-4490-98c6-88a40e9deacc/scene/recipe/S16sh4__bgfirst_bg.png",
   "bg_asset_id": "18c31ecd-1fb4-4b09-bc3b-91a5d20f0ffb",
   "bg_record_key": "S16sh4::bgfirst_bg",
   "chain_winner": true,
   "authority": "groupbg",
   "group_key": "farm_granary_yard",
   "groupbg_asset_id": "478efbd2-93b2-4c89-a843-9c8c194e1849"
  },
  "ref_mode": "재투영 배경+콘티+엔티티 (2택1: 체인 승)",
  "share_plan": {
   "ref_plan": "background"
  }
 },
 "S16sh4::cine": {
  "applied": true,
  "attempted_at": "2026-09-09T02:34:00.735411+00:00",
  "fingerprint": "eb26b606bc65d4ce420779042cc5b46f2b1a68d35e0d5d18469f30d0a04b3231",
  "fingerprint_version": 2,
  "provider": "mai",
  "endpoint": "openrouter/chat-completions",
  "model": "microsoft/mai-image-2.6",
  "pack": "24.202608252115",
  "source_file": "S16sh4_sel.png",
  "source_sha256": "d1cbaf1d9ab0c23f2c31ce234420605f5dbe259d2ae4620c7d72a6bce1d38535",
  "file": "S16sh4_cine.png",
  "staged_sha256": "2d76a530f524572745d010f5f15497722ec94bd2dd1cb020ea447159e3a4523e",
  "latency_ms": 27495
 },
 "S16sh7::signage": {
  "fp": "2171c0bef0a78426",
  "inscriptions": [],
  "cues": [],
  "dropped": []
 },
 "S16sh7": {
  "input_fingerprint": "5d888de2f84b4541",
  "prompt": "Create ONE FINAL photorealistic live-action film still of the moment below — early 20th-century Korea, Jeolla region; all people are Korean unless explicitly stated otherwise. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 쓰러진 조선인 노파 앞에서 의병1의 앞을 떡 하니 가로막고 선 안규홍의 넓은 뒷모습.\n\nLOCATION (lock): The outdoor space before the rural storehouse doorway, beside the fallen woman and the seed-grain sack. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: wide shot\n- KEY BACKGROUND ELEMENTS: 씨나락 자루 (Held in place beneath 안규홍's straw-shod foot) — The tied sack lies across the lower center, its near side visible beneath his foot; used as Makes the physical obstruction and its moral purpose visible together; 흙바닥 (Supports the fallen woman and the sack); used as Maintains continuous ground between the three figures; 광 입구 (The sack has been removed from the granary) — Seen obliquely behind 의병1; used as Establishes the opposing side of the intervention.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Maintain the courtyard's unembellished daylight and muted palette, with natural contrast keeping the protective body and fallen woman readable.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The clothing, hair and overall look of 조선인 노파(씨나락 주인) — who appear both in that photo and in this shot — are LOCKED to that photo. Anyone else visible in that photo is NOT in this shot: never carry their face, body or clothing onto anyone here. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nPREVIOUS STILL USAGE (follow exactly — what to take from the attached still and what to exclude; it governs place and objects only, never who is in this shot or how the camera sees them): Take the dirt ground, granary entrance materials, and daylight visible in the reference. Exclude the earlier undisturbed placement of the seed-grain sack and any dust cloud unique to that instant.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The tied seed-rice sack remains on the dirt outside the storehouse. 안규홍: He stands blocking the sack, pressing it down with a straw-shod foot; the jade ring remains concealed. 조선인 노파(씨나락 주인): She lies on the dirt after being thrown aside.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 안규홍 (한국인 남성, 청년 나이대, 검은 머리, 단단히 묶은 머리); 조선인 노파(씨나락 주인) (한국인 여성, 노년 나이대, 자연스러운 노년의 얼굴) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.",
  "roll_prompts": {
   "A": "Create ONE FINAL photorealistic live-action film still of the moment below — early 20th-century Korea, Jeolla region; all people are Korean unless explicitly stated otherwise. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 쓰러진 조선인 노파 앞에서 의병1의 앞을 떡 하니 가로막고 선 안규홍의 넓은 뒷모습.\n\nLOCATION (lock): The outdoor space before the rural storehouse doorway, beside the fallen woman and the seed-grain sack. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: wide shot\n- KEY BACKGROUND ELEMENTS: 씨나락 자루 (Held in place beneath 안규홍's straw-shod foot) — The tied sack lies across the lower center, its near side visible beneath his foot; used as Makes the physical obstruction and its moral purpose visible together; 흙바닥 (Supports the fallen woman and the sack); used as Maintains continuous ground between the three figures; 광 입구 (The sack has been removed from the granary) — Seen obliquely behind 의병1; used as Establishes the opposing side of the intervention.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Maintain the courtyard's unembellished daylight and muted palette, with natural contrast keeping the protective body and fallen woman readable.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The clothing, hair and overall look of 조선인 노파(씨나락 주인) — who appear both in that photo and in this shot — are LOCKED to that photo. Anyone else visible in that photo is NOT in this shot: never carry their face, body or clothing onto anyone here. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nPREVIOUS STILL USAGE (follow exactly — what to take from the attached still and what to exclude; it governs place and objects only, never who is in this shot or how the camera sees them): Take the dirt ground, granary entrance materials, and daylight visible in the reference. Exclude the earlier undisturbed placement of the seed-grain sack and any dust cloud unique to that instant.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The tied seed-rice sack remains on the dirt outside the storehouse. 안규홍: He stands blocking the sack, pressing it down with a straw-shod foot; the jade ring remains concealed. 조선인 노파(씨나락 주인): She lies on the dirt after being thrown aside.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 안규홍 (한국인 남성, 청년 나이대, 검은 머리, 단단히 묶은 머리); 조선인 노파(씨나락 주인) (한국인 여성, 노년 나이대, 자연스러운 노년의 얼굴) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.",
   "B": "Create ONE FINAL photorealistic live-action film still of the moment below — early 20th-century Korea, Jeolla region; all people are Korean unless explicitly stated otherwise. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 쓰러진 조선인 노파 앞에서 의병1의 앞을 떡 하니 가로막고 선 안규홍의 넓은 뒷모습.\n\nLOCATION (lock): The outdoor space before the rural storehouse doorway, beside the fallen woman and the seed-grain sack. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: wide shot\n- KEY BACKGROUND ELEMENTS: 씨나락 자루 (Held in place beneath 안규홍's straw-shod foot) — The tied sack lies across the lower center, its near side visible beneath his foot; used as Makes the physical obstruction and its moral purpose visible together; 흙바닥 (Supports the fallen woman and the sack); used as Maintains continuous ground between the three figures; 광 입구 (The sack has been removed from the granary) — Seen obliquely behind 의병1; used as Establishes the opposing side of the intervention.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Maintain the courtyard's unembellished daylight and muted palette, with natural contrast keeping the protective body and fallen woman readable.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The clothing, hair and overall look of 조선인 노파(씨나락 주인) — who appear both in that photo and in this shot — are LOCKED to that photo. Anyone else visible in that photo is NOT in this shot: never carry their face, body or clothing onto anyone here. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nPREVIOUS STILL USAGE (follow exactly — what to take from the attached still and what to exclude; it governs place and objects only, never who is in this shot or how the camera sees them): Take the dirt ground, granary entrance materials, and daylight visible in the reference. Exclude the earlier undisturbed placement of the seed-grain sack and any dust cloud unique to that instant.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The tied seed-rice sack remains on the dirt outside the storehouse. 안규홍: He stands blocking the sack, pressing it down with a straw-shod foot; the jade ring remains concealed. 조선인 노파(씨나락 주인): She lies on the dirt after being thrown aside.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 안규홍 (한국인 남성, 청년 나이대, 검은 머리, 단단히 묶은 머리); 조선인 노파(씨나락 주인) (한국인 여성, 노년 나이대, 자연스러운 노년의 얼굴) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.\n\nCOMPOSITION VARIATION (this candidate only): keep the place, the people and the moment exactly as specified, and keep from the CAMERA & FRAME contract its subject scale, screen placement and key background elements exactly. For this candidate only, its camera angle and height are deliberately left open: choose a camera position distinctly different from the obvious one, as a film director picking a second setup on the same blocking. Stay true to any explicit framing the SHOT TEXT itself dictates."
  },
  "shot_run_spend_attempt_count": 1,
  "readings": [
   {
    "label": "A",
    "direction": "가운데 남성은 앞을 향해 서 있고, 오른쪽 남성과 넘어진 노파는 가운데 남성을 바라보고 있음.",
    "built_space": "흙바닥 마당에 곡식창고와 돌담이 있으며, 인물들이 자연스럽게 배치됨.",
    "entities": "가운데 남성은 안규홍의 복장과 묶은 머리가 일치함. 노파는 참조 이미지와 일치하나 자루를 안고 있음. 오른쪽 남성은 의병1. 씨나락 자루가 노파의 품과 바닥에 여러 개로 중복되어 나타남.",
    "hard_violations": [
     "[gemini-pro] 씨나락 자루 중복 생성 (노파가 안고 있는 것 외에 바닥에 추가 자루 존재)",
     "[gemini-pro] 안규홍이 자루를 짚신 신은 발로 밟고 있지 않음",
     "[gpt-high] 단일 씨나락 자루가 노파의 품과 안규홍 발 사이에 각각 하나씩, 총 두 개로 중복되어 있다."
    ],
    "physics": "모든 인물과 사물이 흙바닥에 올바르게 지탱되어 있음."
   },
   {
    "label": "B",
    "direction": "등을 보인 남성이 정면을 보는 남성을 마주보고 있으며, 정면의 남성은 등을 보인 남성을 응시함. 노파는 바닥에 엎어져 있음.",
    "built_space": "열린 곡식창고 문과 흙바닥이 보이며 참조 이미지의 장소와 일치함.",
    "entities": "등을 보인 남성은 상투를 틀고 있어 안규홍의 머리 모양과 다름. 정면을 보는 남성에게 안규홍의 얼굴과 복장이 잘못 적용됨. 바닥에 씨나락 자루 1개가 놓여 있음.",
    "hard_violations": [
     "[gemini-pro] 안규홍이 뒷모습이 아닌 정면을 향해 잘못 배치됨 (의병1과 안규홍의 인물 역할이 완전히 뒤바뀜)"
    ],
    "physics": "등을 보인 남성의 왼발이 자루 위에 안정적으로 지탱되어 있으며, 나머지 인물도 바닥에 올바르게 위치함."
   }
  ],
  "cross_model_order": {
   "policy": "cross_model_order_v2_slot0_fwd_slot1_rev_parallel2",
   "slots": [
    {
     "model": "gemini-pro",
     "order": "forward",
     "display_to_canonical": {
      "A": "A",
      "B": "B"
     },
     "raw": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 3,
        "verdict_ko": "안규홍의 뒷모습과 복장은 올바르게 연출되었으나, 노파가 여전히 자루를 안고 있고 바닥에 자루가 중복 생성되어 구체적인 행동 지시를 위반했습니다."
       },
       {
        "label": "B",
        "score": 2,
        "verdict_ko": "자루를 밟고 가로막는 행동 연출은 정확하나, 안규홍의 얼굴과 신원이 정면을 향한 의병에게 잘못 적용되어 가장 중요한 인물 배치가 완전히 실패했습니다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "가운데 남성은 앞을 향해 서 있고, 오른쪽 남성과 넘어진 노파는 가운데 남성을 바라보고 있음.",
        "built_space": "흙바닥 마당에 곡식창고와 돌담이 있으며, 인물들이 자연스럽게 배치됨.",
        "entities": "가운데 남성은 안규홍의 복장과 묶은 머리가 일치함. 노파는 참조 이미지와 일치하나 자루를 안고 있음. 오른쪽 남성은 의병1. 씨나락 자루가 노파의 품과 바닥에 여러 개로 중복되어 나타남.",
        "hard_violations": [
         "씨나락 자루 중복 생성 (노파가 안고 있는 것 외에 바닥에 추가 자루 존재)",
         "안규홍이 자루를 짚신 신은 발로 밟고 있지 않음"
        ],
        "physics": "모든 인물과 사물이 흙바닥에 올바르게 지탱되어 있음."
       },
       {
        "label": "B",
        "direction": "등을 보인 남성이 정면을 보는 남성을 마주보고 있으며, 정면의 남성은 등을 보인 남성을 응시함. 노파는 바닥에 엎어져 있음.",
        "built_space": "열린 곡식창고 문과 흙바닥이 보이며 참조 이미지의 장소와 일치함.",
        "entities": "등을 보인 남성은 상투를 틀고 있어 안규홍의 머리 모양과 다름. 정면을 보는 남성에게 안규홍의 얼굴과 복장이 잘못 적용됨. 바닥에 씨나락 자루 1개가 놓여 있음.",
        "hard_violations": [
         "안규홍이 뒷모습이 아닌 정면을 향해 잘못 배치됨 (의병1과 안규홍의 인물 역할이 완전히 뒤바뀜)"
        ],
        "physics": "등을 보인 남성의 왼발이 자루 위에 안정적으로 지탱되어 있으며, 나머지 인물도 바닥에 올바르게 위치함."
       }
      ],
      "all_candidates_fail": true
     },
     "normalized": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 3,
        "verdict_ko": "안규홍의 뒷모습과 복장은 올바르게 연출되었으나, 노파가 여전히 자루를 안고 있고 바닥에 자루가 중복 생성되어 구체적인 행동 지시를 위반했습니다."
       },
       {
        "label": "B",
        "score": 2,
        "verdict_ko": "자루를 밟고 가로막는 행동 연출은 정확하나, 안규홍의 얼굴과 신원이 정면을 향한 의병에게 잘못 적용되어 가장 중요한 인물 배치가 완전히 실패했습니다."
       }
      ],
      "all_candidates_fail": true,
      "readings": [
       {
        "label": "A",
        "direction": "가운데 남성은 앞을 향해 서 있고, 오른쪽 남성과 넘어진 노파는 가운데 남성을 바라보고 있음.",
        "built_space": "흙바닥 마당에 곡식창고와 돌담이 있으며, 인물들이 자연스럽게 배치됨.",
        "entities": "가운데 남성은 안규홍의 복장과 묶은 머리가 일치함. 노파는 참조 이미지와 일치하나 자루를 안고 있음. 오른쪽 남성은 의병1. 씨나락 자루가 노파의 품과 바닥에 여러 개로 중복되어 나타남.",
        "hard_violations": [
         "씨나락 자루 중복 생성 (노파가 안고 있는 것 외에 바닥에 추가 자루 존재)",
         "안규홍이 자루를 짚신 신은 발로 밟고 있지 않음"
        ],
        "physics": "모든 인물과 사물이 흙바닥에 올바르게 지탱되어 있음."
       },
       {
        "label": "B",
        "direction": "등을 보인 남성이 정면을 보는 남성을 마주보고 있으며, 정면의 남성은 등을 보인 남성을 응시함. 노파는 바닥에 엎어져 있음.",
        "built_space": "열린 곡식창고 문과 흙바닥이 보이며 참조 이미지의 장소와 일치함.",
        "entities": "등을 보인 남성은 상투를 틀고 있어 안규홍의 머리 모양과 다름. 정면을 보는 남성에게 안규홍의 얼굴과 복장이 잘못 적용됨. 바닥에 씨나락 자루 1개가 놓여 있음.",
        "hard_violations": [
         "안규홍이 뒷모습이 아닌 정면을 향해 잘못 배치됨 (의병1과 안규홍의 인물 역할이 완전히 뒤바뀜)"
        ],
        "physics": "등을 보인 남성의 왼발이 자루 위에 안정적으로 지탱되어 있으며, 나머지 인물도 바닥에 올바르게 위치함."
       }
      ]
     },
     "ok": true
    },
    {
     "model": "gpt-high",
     "order": "reverse",
     "display_to_canonical": {
      "A": "B",
      "B": "A"
     },
     "raw": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 8,
        "verdict_ko": "넓은 뒷모습으로 의병1을 정면에서 가로막고 짚신으로 자루를 누르는 핵심 연출이 맞지만, 왼쪽 건물 등 장소의 세부 연속성은 덜 확실하다."
       },
       {
        "label": "B",
        "score": 3,
        "verdict_ko": "씨나락 자루가 두 개로 중복되며, 안규홍은 자루를 밟지 않고 의병1 대신 광 입구를 향해 서 있어 핵심 대치 연출을 놓쳤다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "안규홍은 카메라에 넓은 등을 보인 채 오른쪽 앞의 의병1을 향하고, 의병1도 안규홍의 얼굴 쪽을 본다. 두 남자의 대치 방향이 맞으며 노파는 안규홍 뒤편 왼쪽 땅에 얼굴을 숙이고 쓰러져 있다. 겨누는 무기나 이동 중인 물체는 없다.",
        "built_space": "오른쪽에 열린 광 출입구 하나와 목재 문짝 두 장, 낮은 문턱, 흙벽과 돌 기단이 보인다. 의병1은 그 입구 앞 마당에 있고 광은 그의 뒤로 비스듬히 보인다. 중앙 뒤에는 낮은 돌담과 초가지붕 시설이 있으며 왼쪽에는 별도 건물의 격자문이 보이는데, 이 건물은 이전 사진에서 확인되지 않는다. 세 사람과 자루 사이에는 연속된 흙바닥이 있다. 반사면은 없다.",
        "entities": "인물은 안규홍, 의병1, 노파에 해당하는 세 명이다. 안규홍은 검은 머리를 단단히 묶은 청년 남성으로, 참조의 밝은 삼베옷과 허리띠 및 짚신을 갖췄으며 얼굴은 뒷모습 때문에 확인할 수 없다. 노파는 회색 머리와 짙은 깃의 누런 옷을 입은 노년 여성으로 이전 사진의 외양과 대체로 맞는다. 의병1은 밝은 옷의 청년 남성으로 안규홍과 비슷한 외양이다. 묶인 씨나락 자루는 하단 중앙에서 오른쪽에 걸쳐 하나만 놓여 있고 옥가락지는 보이지 않는다. 읽을 수 있는 글자는 없다.",
        "hard_violations": [],
        "physics": "안규홍은 왼쪽 다리를 지지축으로 세우고 오른쪽 짚신을 자루 윗면에 대어 누른다. 자루 밑면은 흙바닥에 닿아 있어 압박 동작이 성립한다. 노파는 몸통과 접힌 다리, 머리 옆 손을 땅에 대고 누워 있다. 의병1의 두 발도 땅에 닿는다. 지지 없이 떠 있는 인물이나 물체는 없다."
       },
       {
        "label": "B",
        "direction": "안규홍의 등과 머리는 정면의 어두운 광 입구를 향하며, 오른쪽에서 접근하는 의병1을 향해 돌아서 있지 않다. 의병1은 왼쪽 아래의 안규홍과 자루 쪽을 보고 노파는 안규홍 쪽을 올려다본다. 안규홍이 상대의 정면을 막아선 대치 방향은 분명하지 않다.",
        "built_space": "중앙 뒤에 광 출입구 하나와 낮은 문턱, 목재 기둥 및 흙벽이 있고 왼쪽에는 닭장 하나와 돌담이 보여 이전 장소의 특징은 잘 이어진다. 오른쪽 위에는 격자창 일부가 보인다. 그러나 광 입구는 의병1 뒤가 아니라 안규홍 바로 앞에 놓여 있다. 세 사람은 같은 흙마당에 있으며 반사면은 없다.",
        "entities": "안규홍은 묶은 검은 머리, 밝은 삼베옷, 허리띠와 짚신을 갖춘 청년 남성의 뒷모습으로 참조의 보이는 특징에 부합한다. 노파의 회색 머리와 누런 옷, 짙은 깃도 참조와 맞는다. 오른쪽 의병1은 머리띠와 수염이 있는 성인 남성이다. 그런데 노파가 안고 있는 묶인 자루 하나와 안규홍 발 사이에 놓인 묶인 자루 하나가 동시에 있어, 이어져야 할 씨나락 자루가 중복된다. 옥가락지와 읽을 수 있는 글자는 보이지 않는다.",
        "hard_violations": [
         "단일 씨나락 자루가 노파의 품과 안규홍 발 사이에 각각 하나씩, 총 두 개로 중복되어 있다."
        ],
        "physics": "안규홍의 두 짚신은 모두 흙바닥에 닿고, 어느 발도 자루 윗면을 누르지 않는다. 노파는 옆구리와 팔꿈치를 땅에 대어 상체를 지탱하면서 양손으로 별도의 자루를 안고 있다. 오른쪽 의병1은 굽힌 다리로 접근하는 자세이며 지지 없이 떠 있다고 볼 근거는 없다. 두 자루도 몸이나 땅에 받쳐져 있으나, 요구한 발의 압박 동작은 없다."
       }
      ],
      "all_candidates_fail": false
     },
     "normalized": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 8,
        "verdict_ko": "넓은 뒷모습으로 의병1을 정면에서 가로막고 짚신으로 자루를 누르는 핵심 연출이 맞지만, 왼쪽 건물 등 장소의 세부 연속성은 덜 확실하다."
       },
       {
        "label": "A",
        "score": 3,
        "verdict_ko": "씨나락 자루가 두 개로 중복되며, 안규홍은 자루를 밟지 않고 의병1 대신 광 입구를 향해 서 있어 핵심 대치 연출을 놓쳤다."
       }
      ],
      "all_candidates_fail": false,
      "readings": [
       {
        "label": "B",
        "direction": "안규홍은 카메라에 넓은 등을 보인 채 오른쪽 앞의 의병1을 향하고, 의병1도 안규홍의 얼굴 쪽을 본다. 두 남자의 대치 방향이 맞으며 노파는 안규홍 뒤편 왼쪽 땅에 얼굴을 숙이고 쓰러져 있다. 겨누는 무기나 이동 중인 물체는 없다.",
        "built_space": "오른쪽에 열린 광 출입구 하나와 목재 문짝 두 장, 낮은 문턱, 흙벽과 돌 기단이 보인다. 의병1은 그 입구 앞 마당에 있고 광은 그의 뒤로 비스듬히 보인다. 중앙 뒤에는 낮은 돌담과 초가지붕 시설이 있으며 왼쪽에는 별도 건물의 격자문이 보이는데, 이 건물은 이전 사진에서 확인되지 않는다. 세 사람과 자루 사이에는 연속된 흙바닥이 있다. 반사면은 없다.",
        "entities": "인물은 안규홍, 의병1, 노파에 해당하는 세 명이다. 안규홍은 검은 머리를 단단히 묶은 청년 남성으로, 참조의 밝은 삼베옷과 허리띠 및 짚신을 갖췄으며 얼굴은 뒷모습 때문에 확인할 수 없다. 노파는 회색 머리와 짙은 깃의 누런 옷을 입은 노년 여성으로 이전 사진의 외양과 대체로 맞는다. 의병1은 밝은 옷의 청년 남성으로 안규홍과 비슷한 외양이다. 묶인 씨나락 자루는 하단 중앙에서 오른쪽에 걸쳐 하나만 놓여 있고 옥가락지는 보이지 않는다. 읽을 수 있는 글자는 없다.",
        "hard_violations": [],
        "physics": "안규홍은 왼쪽 다리를 지지축으로 세우고 오른쪽 짚신을 자루 윗면에 대어 누른다. 자루 밑면은 흙바닥에 닿아 있어 압박 동작이 성립한다. 노파는 몸통과 접힌 다리, 머리 옆 손을 땅에 대고 누워 있다. 의병1의 두 발도 땅에 닿는다. 지지 없이 떠 있는 인물이나 물체는 없다."
       },
       {
        "label": "A",
        "direction": "안규홍의 등과 머리는 정면의 어두운 광 입구를 향하며, 오른쪽에서 접근하는 의병1을 향해 돌아서 있지 않다. 의병1은 왼쪽 아래의 안규홍과 자루 쪽을 보고 노파는 안규홍 쪽을 올려다본다. 안규홍이 상대의 정면을 막아선 대치 방향은 분명하지 않다.",
        "built_space": "중앙 뒤에 광 출입구 하나와 낮은 문턱, 목재 기둥 및 흙벽이 있고 왼쪽에는 닭장 하나와 돌담이 보여 이전 장소의 특징은 잘 이어진다. 오른쪽 위에는 격자창 일부가 보인다. 그러나 광 입구는 의병1 뒤가 아니라 안규홍 바로 앞에 놓여 있다. 세 사람은 같은 흙마당에 있으며 반사면은 없다.",
        "entities": "안규홍은 묶은 검은 머리, 밝은 삼베옷, 허리띠와 짚신을 갖춘 청년 남성의 뒷모습으로 참조의 보이는 특징에 부합한다. 노파의 회색 머리와 누런 옷, 짙은 깃도 참조와 맞는다. 오른쪽 의병1은 머리띠와 수염이 있는 성인 남성이다. 그런데 노파가 안고 있는 묶인 자루 하나와 안규홍 발 사이에 놓인 묶인 자루 하나가 동시에 있어, 이어져야 할 씨나락 자루가 중복된다. 옥가락지와 읽을 수 있는 글자는 보이지 않는다.",
        "hard_violations": [
         "단일 씨나락 자루가 노파의 품과 안규홍 발 사이에 각각 하나씩, 총 두 개로 중복되어 있다."
        ],
        "physics": "안규홍의 두 짚신은 모두 흙바닥에 닿고, 어느 발도 자루 윗면을 누르지 않는다. 노파는 옆구리와 팔꿈치를 땅에 대어 상체를 지탱하면서 양손으로 별도의 자루를 안고 있다. 오른쪽 의병1은 굽힌 다리로 접근하는 자세이며 지지 없이 떠 있다고 볼 근거는 없다. 두 자루도 몸이나 땅에 받쳐져 있으나, 요구한 발의 압박 동작은 없다."
       }
      ]
     },
     "ok": true
    }
   ],
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "slot_winner_match": false,
   "slot_winner": {
    "gemini-pro": "A",
    "gpt-high": "B"
   },
   "route": "cross_slot_combined"
  },
  "dual": {
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "normalized": {
    "A": 1.375,
    "B": 1.667
   },
   "adjusted": {
    "A": 1.125,
    "B": 1.417
   },
   "violations": {
    "A": [
     "[gemini-pro] 씨나락 자루 중복 생성 (노파가 안고 있는 것 외에 바닥에 추가 자루 존재)",
     "[gemini-pro] 안규홍이 자루를 짚신 신은 발로 밟고 있지 않음",
     "[gpt-high] 단일 씨나락 자루가 노파의 품과 안규홍 발 사이에 각각 하나씩, 총 두 개로 중복되어 있다."
    ],
    "B": [
     "[gemini-pro] 안규홍이 뒷모습이 아닌 정면을 향해 잘못 배치됨 (의병1과 안규홍의 인물 역할이 완전히 뒤바뀜)"
    ]
   },
   "per_model_winner": {
    "gemini-pro": "A",
    "gpt-high": "B"
   },
   "agreed": false
  },
  "totals": {
   "A": 1125,
   "B": 1417
  },
  "selected": "B",
  "ranking": [
   "B",
   "A"
  ],
  "verdicts": [
   {
    "label": "A",
    "score": 1125,
    "verdict_ko": "안규홍의 뒷모습과 복장은 올바르게 연출되었으나, 노파가 여전히 자루를 안고 있고 바닥에 자루가 중복 생성되어 구체적인 행동 지시를 위반했습니다.  ★위반: [gemini-pro] 씨나락 자루 중복 생성 (노파가 안고 있는 것 외에 바닥에 추가 자루 존재) / [gemini-pro] 안규홍이 자루를 짚신 신은 발로 밟고 있지 않음 / [gpt-high] 단일 씨나락 자루가 노파의 품과 안규홍 발 사이에 각각 하나씩, 총 두 개로 중복되어 있다."
   },
   {
    "label": "B",
    "score": 1417,
    "verdict_ko": "자루를 밟고 가로막는 행동 연출은 정확하나, 안규홍의 얼굴과 신원이 정면을 향한 의병에게 잘못 적용되어 가장 중요한 인물 배치가 완전히 실패했습니다.  ★위반: [gemini-pro] 안규홍이 뒷모습이 아닌 정면을 향해 잘못 배치됨 (의병1과 안규홍의 인물 역할이 완전히 뒤바뀜)"
   }
  ],
  "refs": [
   {
    "label": "PREVIOUS SHOT STILL — a visually related earlier shot of this same place: the location's look, materials, fixed features and lighting mood are LOCKED to this photo; never copy its camera framing. The clothing, hair and overall look of 조선인 노파(씨나락 주인) — who appear both in that photo and in this shot — are LOCKED to that photo. Anyone else visible in that photo is NOT in this shot: never carry their face, body or clothing onto anyone here. If this photo shows a character who cannot move (dead or unconscious), that character's exact pose, position and orientation are ALSO LOCKED — treat their body as a fixed prop of the set that only the camera moves around.",
    "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/db00517b-165b-4005-a287-f28936b59b95/images/272bd7af-9294-4490-98c6-88a40e9deacc/scene/recipe/S16sh4_sel.png",
    "asset_id": "26ab3528-eec5-4c2e-a30f-719c9aa615e0",
    "role": "prev_still"
   },
   {
    "label": "CHARACTER REFERENCE — 안규홍: the exact person appearing in this shot; match face, hair and build exactly.",
    "path": "<bytes:1028384>",
    "asset_id": "3b5894db-ce55-4f34-b966-5e725c3aa952",
    "role": "character_ref"
   },
   {
    "label": "CHARACTER REFERENCE — 조선인 노파(씨나락 주인): the exact person appearing in this shot; match face, hair and build exactly.",
    "path": "<bytes:982368>",
    "asset_id": "aa993fab-d06c-423d-8316-10d12c20f3c0",
    "role": "character_ref"
   }
  ],
  "critique_skipped": true,
  "shot_run_produced": false,
  "shot_run_uid": "06aa1411-4c31-7476-a0c9-20856b7a7c6a",
  "ref_mode": "prev+엔티티",
  "share_plan": {
   "ref_plan": "prev",
   "prev_anchor_tag": "S16sh4"
  },
  "lane_policy": "ab_select_bypass:prev"
 },
 "S16sh7::cine": {
  "applied": true,
  "attempted_at": "2026-09-09T02:36:11.250278+00:00",
  "fingerprint": "704dd765bb6369958451b31c8ec1ebcd75586d844a477061af951dd4be6561b1",
  "fingerprint_version": 2,
  "provider": "mai",
  "endpoint": "openrouter/chat-completions",
  "model": "microsoft/mai-image-2.6",
  "pack": "24.202608252115",
  "source_file": "S16sh7_sel.png",
  "source_sha256": "9d3b8bf6b986114d83e0600303a6d26fc4a0e52c2865cf613c8160e6535dc26d",
  "file": "S16sh7_cine.png",
  "staged_sha256": "2299b98be9e342e461fcdabf6cfc7c7c0d31054f68ba5e7fc0429efaf422b27e",
  "latency_ms": 22732
 },
 "S16sh16::signage": {
  "fp": "2acd964ffc0f83ed",
  "inscriptions": [],
  "cues": [],
  "dropped": []
 },
 "S16sh16": {
  "input_fingerprint": "2836572be0bc144a",
  "prompt": "Create ONE FINAL photorealistic live-action film still of the moment below — early 20th-century Korea, Jeolla region; all people are Korean unless explicitly stated otherwise. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 바닥에 뺨을 댄 안규홍의 흐려진 시야 너머로 흙바닥을 질질 끌려가는 씨나락 자루의 바닥 시점.\n\nLOCATION (lock): Ground level in the rural home's dirt yard outside the storehouse, along the path of the departing seed-grain sack. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: medium shot\n- FRAME LAYOUT: seed-grain sack receding along the dirt ground in the middle-right of the frame, midground, moves toward upper-right departure path.\n- KEY BACKGROUND ELEMENTS: 씨나락 자루 (Being dragged away across the dirt ground) — Its lower side and underside face the ground-level viewpoint as it recedes; used as The retreating object is the final readable focus of 안규홍's impaired vision; 흙바닥 (Visible along the sack's departure path); used as Provides a continuous low plane that establishes the fallen viewpoint.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Courtyard daylight remains physically unchanged while injury-related visual softness reduces perceived definition, without hallucinated colors or added atmospheric effects.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The people visible in that photo are NOT in this shot: never carry their faces, bodies or clothing onto anyone here. Each person in THIS shot is defined solely by the PEOPLE list, their own CHARACTER REFERENCE images and their wardrobe notes. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The tied seed-rice sack is being dragged away across the dirt.\n\nNO PEOPLE IN THIS SHOT: the shot text shows only the place and its state — no living person appears in frame, and no face, and no figure standing, sitting or passing through. Do not invent someone to occupy the space.\n\nOne thing this rule does NOT forbid: the hand or arm of whoever is doing what the shot text describes. If the moment is an object being held, carried, operated, read, aimed, worn or handed over, the hand doing it belongs in frame, resting on or gripping that object exactly as its real use requires. An object photographed floating unsupported, with the hand that must be holding it erased, is a worse failure than the person you were avoiding — render the hand.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.",
  "roll_prompts": {
   "A": "Create ONE FINAL photorealistic live-action film still of the moment below — early 20th-century Korea, Jeolla region; all people are Korean unless explicitly stated otherwise. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 바닥에 뺨을 댄 안규홍의 흐려진 시야 너머로 흙바닥을 질질 끌려가는 씨나락 자루의 바닥 시점.\n\nLOCATION (lock): Ground level in the rural home's dirt yard outside the storehouse, along the path of the departing seed-grain sack. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: medium shot\n- FRAME LAYOUT: seed-grain sack receding along the dirt ground in the middle-right of the frame, midground, moves toward upper-right departure path.\n- KEY BACKGROUND ELEMENTS: 씨나락 자루 (Being dragged away across the dirt ground) — Its lower side and underside face the ground-level viewpoint as it recedes; used as The retreating object is the final readable focus of 안규홍's impaired vision; 흙바닥 (Visible along the sack's departure path); used as Provides a continuous low plane that establishes the fallen viewpoint.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Courtyard daylight remains physically unchanged while injury-related visual softness reduces perceived definition, without hallucinated colors or added atmospheric effects.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The people visible in that photo are NOT in this shot: never carry their faces, bodies or clothing onto anyone here. Each person in THIS shot is defined solely by the PEOPLE list, their own CHARACTER REFERENCE images and their wardrobe notes. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The tied seed-rice sack is being dragged away across the dirt.\n\nNO PEOPLE IN THIS SHOT: the shot text shows only the place and its state — no living person appears in frame, and no face, and no figure standing, sitting or passing through. Do not invent someone to occupy the space.\n\nOne thing this rule does NOT forbid: the hand or arm of whoever is doing what the shot text describes. If the moment is an object being held, carried, operated, read, aimed, worn or handed over, the hand doing it belongs in frame, resting on or gripping that object exactly as its real use requires. An object photographed floating unsupported, with the hand that must be holding it erased, is a worse failure than the person you were avoiding — render the hand.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.",
   "B": "Create ONE FINAL photorealistic live-action film still of the moment below — early 20th-century Korea, Jeolla region; all people are Korean unless explicitly stated otherwise. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 바닥에 뺨을 댄 안규홍의 흐려진 시야 너머로 흙바닥을 질질 끌려가는 씨나락 자루의 바닥 시점.\n\nLOCATION (lock): Ground level in the rural home's dirt yard outside the storehouse, along the path of the departing seed-grain sack. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: medium shot\n- FRAME LAYOUT: seed-grain sack receding along the dirt ground in the middle-right of the frame, midground, moves toward upper-right departure path.\n- KEY BACKGROUND ELEMENTS: 씨나락 자루 (Being dragged away across the dirt ground) — Its lower side and underside face the ground-level viewpoint as it recedes; used as The retreating object is the final readable focus of 안규홍's impaired vision; 흙바닥 (Visible along the sack's departure path); used as Provides a continuous low plane that establishes the fallen viewpoint.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Courtyard daylight remains physically unchanged while injury-related visual softness reduces perceived definition, without hallucinated colors or added atmospheric effects.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The people visible in that photo are NOT in this shot: never carry their faces, bodies or clothing onto anyone here. Each person in THIS shot is defined solely by the PEOPLE list, their own CHARACTER REFERENCE images and their wardrobe notes. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The tied seed-rice sack is being dragged away across the dirt.\n\nNO PEOPLE IN THIS SHOT: the shot text shows only the place and its state — no living person appears in frame, and no face, and no figure standing, sitting or passing through. Do not invent someone to occupy the space.\n\nOne thing this rule does NOT forbid: the hand or arm of whoever is doing what the shot text describes. If the moment is an object being held, carried, operated, read, aimed, worn or handed over, the hand doing it belongs in frame, resting on or gripping that object exactly as its real use requires. An object photographed floating unsupported, with the hand that must be holding it erased, is a worse failure than the person you were avoiding — render the hand.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.\n\nCOMPOSITION VARIATION (this candidate only): keep the place, the people and the moment exactly as specified, and keep from the CAMERA & FRAME contract its subject scale, screen placement and key background elements exactly. For this candidate only, its camera angle and height are deliberately left open: choose a camera position distinctly different from the obvious one, as a film director picking a second setup on the same blocking. Stay true to any explicit framing the SHOT TEXT itself dictates."
  },
  "shot_run_spend_attempt_count": 1,
  "initial_roll_all_fail": true,
  "readings": [
   {
    "label": "A",
    "direction": "카메라는 바닥 높이에 위치하며, 화면 왼쪽의 얼굴은 바닥을 향하고 있고 정체불명의 팔이 자루를 쥐고 화면 오른쪽으로 당기고 있습니다.",
    "built_space": "흙바닥이 보이며, 배경에는 흐릿하게 전통 가옥의 기둥과 목조 구조물이 보입니다.",
    "entities": "씨나락 자루가 묘사되어 있으나, 지시문에 어긋나게 인물의 큰 얼굴과 짚신을 신은 발이 화면에 분명하게 나타납니다.",
    "hard_violations": [
     "[gemini-pro] 물리적으로 불가능한 신체 구조 (얼굴 바로 옆에 발이 위치하고, 몸통 없이 팔만 튀어나오는 등 신체 부위가 기괴하게 얽혀 있음).",
     "[gemini-pro] '인물 없음(NO PEOPLE)' 및 '얼굴 없음' 규칙을 명백히 위반하여 프레임 안에 큰 얼굴을 포함함.",
     "[gpt-high] 사람과 얼굴을 배제하라는 조건을 어기고 노인의 얼굴과 상체를 화면 전경에 크게 등장시켰다.",
     "[gpt-high] 참조에서 옆으로 웅크리고 손을 얼굴 앞 바닥에 둔 쓰러진 인물을, 상체를 들고 팔을 자루까지 뻗어 잡는 자세로 바꾸어 고정 자세 조건을 위반했다."
    ],
    "physics": "얼굴은 바닥에 닿아 지탱되고 손이 자루를 잡고 있으나, 발이나 팔이 이어지는 신체 중심(몸통)이 전혀 논리적으로 설명되지 않아 물리적 실체가 붕괴되어 있습니다."
   },
   {
    "label": "B",
    "direction": "카메라는 바닥 높이에서 앞을 향하고 있으며, 두 손이 자루의 윗부분을 잡고 끌고 있어 바닥에 끌린 자국이 뚜렷하게 보입니다.",
    "built_space": "흙바닥 마당이며, 배경에 돌기단과 나무 문짝이 열린 전통 가옥이 위치해 레퍼런스의 공간 구조와 일치합니다.",
    "entities": "화면 우측 전면에 씨나락 자루와 이를 잡은 손이 있고, 배경 좌측에는 쓰러진 인물(안규홍)이 있습니다.",
    "hard_violations": [
     "[gemini-pro] 스테이징 위반: 쓰러진 안규홍의 시점(POV)이어야 하는 카메라 위치임에도 배경에 안규홍 본인을 중복해서 배치함.",
     "[gemini-pro] '인물 없음(NO PEOPLE)' 규칙을 위반하여 쓰러진 인물의 전신과 자루를 끄는 인물의 실루엣을 프레임 안에 포함함.",
     "[gpt-high] 인물 없는 장면이라는 명시적 조건을 어기고, 쓰러진 사람의 얼굴과 몸 및 자루를 당기는 사람의 몸통을 노출했다. 허용된 손·팔만의 등장을 넘는다."
    ],
    "physics": "자루는 바닥에 끌리며 손에 의해 물리적으로 올바르게 지탱되고 이동 중이며, 쓰러진 인물은 땅에 엎드려 있어 신체적 지지 상태는 논리적입니다."
   }
  ],
  "cross_model_order": {
   "policy": "cross_model_order_v2_slot0_fwd_slot1_rev_parallel2",
   "slots": [
    {
     "model": "gemini-pro",
     "order": "forward",
     "display_to_canonical": {
      "A": "A",
      "B": "B"
     },
     "raw": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 1,
        "verdict_ko": "카메라가 쓰러진 안규홍의 시야(시점)여야 함에도 배경에 안규홍을 배치하여 스테이징 논리를 근본적으로 위반했으며, '인물 없음' 규칙을 무시하고 화면에 사람을 포함시켜 치명적인 오류를 범했습니다."
       },
       {
        "label": "A",
        "score": 0,
        "verdict_ko": "프롬프트의 '인물 및 얼굴 없음' 지시를 전면 위반하여 얼굴을 크게 노출시켰으며, 바닥에 닿은 얼굴 바로 옆에 발이 위치하는 등 인체 구조가 물리적으로 불가능하게 기괴하게 얽혀 있어 완전히 실패한 이미지입니다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "카메라는 바닥 높이에 위치하며, 화면 왼쪽의 얼굴은 바닥을 향하고 있고 정체불명의 팔이 자루를 쥐고 화면 오른쪽으로 당기고 있습니다.",
        "built_space": "흙바닥이 보이며, 배경에는 흐릿하게 전통 가옥의 기둥과 목조 구조물이 보입니다.",
        "entities": "씨나락 자루가 묘사되어 있으나, 지시문에 어긋나게 인물의 큰 얼굴과 짚신을 신은 발이 화면에 분명하게 나타납니다.",
        "hard_violations": [
         "물리적으로 불가능한 신체 구조 (얼굴 바로 옆에 발이 위치하고, 몸통 없이 팔만 튀어나오는 등 신체 부위가 기괴하게 얽혀 있음).",
         "'인물 없음(NO PEOPLE)' 및 '얼굴 없음' 규칙을 명백히 위반하여 프레임 안에 큰 얼굴을 포함함."
        ],
        "physics": "얼굴은 바닥에 닿아 지탱되고 손이 자루를 잡고 있으나, 발이나 팔이 이어지는 신체 중심(몸통)이 전혀 논리적으로 설명되지 않아 물리적 실체가 붕괴되어 있습니다."
       },
       {
        "label": "B",
        "direction": "카메라는 바닥 높이에서 앞을 향하고 있으며, 두 손이 자루의 윗부분을 잡고 끌고 있어 바닥에 끌린 자국이 뚜렷하게 보입니다.",
        "built_space": "흙바닥 마당이며, 배경에 돌기단과 나무 문짝이 열린 전통 가옥이 위치해 레퍼런스의 공간 구조와 일치합니다.",
        "entities": "화면 우측 전면에 씨나락 자루와 이를 잡은 손이 있고, 배경 좌측에는 쓰러진 인물(안규홍)이 있습니다.",
        "hard_violations": [
         "스테이징 위반: 쓰러진 안규홍의 시점(POV)이어야 하는 카메라 위치임에도 배경에 안규홍 본인을 중복해서 배치함.",
         "'인물 없음(NO PEOPLE)' 규칙을 위반하여 쓰러진 인물의 전신과 자루를 끄는 인물의 실루엣을 프레임 안에 포함함."
        ],
        "physics": "자루는 바닥에 끌리며 손에 의해 물리적으로 올바르게 지탱되고 이동 중이며, 쓰러진 인물은 땅에 엎드려 있어 신체적 지지 상태는 논리적입니다."
       }
      ],
      "all_candidates_fail": true
     },
     "normalized": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 1,
        "verdict_ko": "카메라가 쓰러진 안규홍의 시야(시점)여야 함에도 배경에 안규홍을 배치하여 스테이징 논리를 근본적으로 위반했으며, '인물 없음' 규칙을 무시하고 화면에 사람을 포함시켜 치명적인 오류를 범했습니다."
       },
       {
        "label": "A",
        "score": 0,
        "verdict_ko": "프롬프트의 '인물 및 얼굴 없음' 지시를 전면 위반하여 얼굴을 크게 노출시켰으며, 바닥에 닿은 얼굴 바로 옆에 발이 위치하는 등 인체 구조가 물리적으로 불가능하게 기괴하게 얽혀 있어 완전히 실패한 이미지입니다."
       }
      ],
      "all_candidates_fail": true,
      "readings": [
       {
        "label": "A",
        "direction": "카메라는 바닥 높이에 위치하며, 화면 왼쪽의 얼굴은 바닥을 향하고 있고 정체불명의 팔이 자루를 쥐고 화면 오른쪽으로 당기고 있습니다.",
        "built_space": "흙바닥이 보이며, 배경에는 흐릿하게 전통 가옥의 기둥과 목조 구조물이 보입니다.",
        "entities": "씨나락 자루가 묘사되어 있으나, 지시문에 어긋나게 인물의 큰 얼굴과 짚신을 신은 발이 화면에 분명하게 나타납니다.",
        "hard_violations": [
         "물리적으로 불가능한 신체 구조 (얼굴 바로 옆에 발이 위치하고, 몸통 없이 팔만 튀어나오는 등 신체 부위가 기괴하게 얽혀 있음).",
         "'인물 없음(NO PEOPLE)' 및 '얼굴 없음' 규칙을 명백히 위반하여 프레임 안에 큰 얼굴을 포함함."
        ],
        "physics": "얼굴은 바닥에 닿아 지탱되고 손이 자루를 잡고 있으나, 발이나 팔이 이어지는 신체 중심(몸통)이 전혀 논리적으로 설명되지 않아 물리적 실체가 붕괴되어 있습니다."
       },
       {
        "label": "B",
        "direction": "카메라는 바닥 높이에서 앞을 향하고 있으며, 두 손이 자루의 윗부분을 잡고 끌고 있어 바닥에 끌린 자국이 뚜렷하게 보입니다.",
        "built_space": "흙바닥 마당이며, 배경에 돌기단과 나무 문짝이 열린 전통 가옥이 위치해 레퍼런스의 공간 구조와 일치합니다.",
        "entities": "화면 우측 전면에 씨나락 자루와 이를 잡은 손이 있고, 배경 좌측에는 쓰러진 인물(안규홍)이 있습니다.",
        "hard_violations": [
         "스테이징 위반: 쓰러진 안규홍의 시점(POV)이어야 하는 카메라 위치임에도 배경에 안규홍 본인을 중복해서 배치함.",
         "'인물 없음(NO PEOPLE)' 규칙을 위반하여 쓰러진 인물의 전신과 자루를 끄는 인물의 실루엣을 프레임 안에 포함함."
        ],
        "physics": "자루는 바닥에 끌리며 손에 의해 물리적으로 올바르게 지탱되고 이동 중이며, 쓰러진 인물은 땅에 엎드려 있어 신체적 지지 상태는 논리적입니다."
       }
      ]
     },
     "ok": true
    },
    {
     "model": "gpt-high",
     "order": "reverse",
     "display_to_canonical": {
      "A": "B",
      "B": "A"
     },
     "raw": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 2,
        "verdict_ko": "금지된 인물들이 등장해 탈락이지만, B보다 자루 중심의 바닥 시점과 끌리는 동작을 더 가깝게 구현했다."
       },
       {
        "label": "B",
        "score": 1,
        "verdict_ko": "금지된 얼굴을 전경에 크게 넣고 쓰러진 인물의 고정 자세까지 바꾸어, 인물 없는 중경 자루 시점을 크게 벗어났다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "자루의 묶인 윗부분을 화면 위쪽의 두 손이 잡고 있으며, 아래쪽 흙의 끌림 자국은 카메라 쪽으로 이어진다. 멀어지는 동작은 읽히지만 지정된 우상단 출구보다는 화면 위쪽으로 당겨지는 방향이다. 왼쪽 인물의 얼굴은 땅을 향하고, 자루를 보는 눈은 확인되지 않는다.",
        "built_space": "왼쪽 배경에 격자 문짝 두 면, 흙벽과 목재 기둥, 낮은 돌 기단이 있고 중앙 뒤에는 낮은 초가 구조물이 보인다. 참조의 농가 재료와 낮 조명은 대체로 이어지지만, 오른쪽 창고 출입구는 거의 프레임 밖이다. 카메라는 지면에 가깝지만 자루가 중경보다 크고 가깝게 차지한다.",
        "entities": "거친 갈색 직물의 곡물 자루 하나와 연속된 흙바닥이 보이며, 자루 입구는 모아져 손에 잡혀 있다. 내용물이 씨나락인지는 외관만으로 확인할 수 없다. 왼쪽에는 참조의 노년 한국인 남성처럼 보이는 회색 머리와 베이지색 옷의 쓰러진 인물이 있고, 위쪽에는 자루를 잡은 다른 인물의 손뿐 아니라 몸통과 옷도 보인다. 읽을 수 있는 글자는 없다.",
        "hard_violations": [
         "인물 없는 장면이라는 명시적 조건을 어기고, 쓰러진 사람의 얼굴과 몸 및 자루를 당기는 사람의 몸통을 노출했다. 허용된 손·팔만의 등장을 넘는다."
        ],
        "physics": "자루 아랫면은 흙바닥에 눌려 있고 윗부분은 두 손이 잡아 지지한다. 무게를 바닥에 둔 채 당기는 동작과 흙의 긁힌 자국은 물리적으로 가능하다. 누운 인물도 머리와 몸이 지면에 닿아 있으며, 지지 없이 떠 있는 물체는 보이지 않는다."
       },
       {
        "label": "B",
        "direction": "자루는 오른쪽에 있지만, 왼쪽 인물 쪽에서 뻗은 손이 윗부분을 잡고 있어 우상단으로 멀어지는 견인 방향이 명확하지 않다. 노인의 눈은 감기거나 아래로 향해 있어 자루를 바라보는 시선은 보이지 않는다. 화면은 흐려진 주관 시점보다 얼굴 옆에서 찍은 외부 시점으로 읽힌다.",
        "built_space": "오른쪽 배경에 어두운 창고 출입구 하나, 열린 목재 문짝, 흙벽, 목재 기둥과 돌 기단이 보이고 뒤에는 낮은 초가 구조물이 있다. 참조의 창고 쪽 재료는 비교적 잘 이어진다. 그러나 거대한 얼굴과 상체가 왼쪽 전경을 채워 지정된 중간 크기의 자루 중심 구도 대신 인물 근접 구도가 되었다.",
        "entities": "오른쪽에 갈색 직물 자루 하나와 흙바닥이 보인다. 자루의 입구 쪽은 손에 가려져 묶인 상태가 확실하지 않고, 씨나락 내용물도 보이지 않는다. 왼쪽에는 참조 노인과 유사한 회색 머리의 한국인 노년 남성 얼굴과 베이지색 상체가 크게 보이며 중앙에는 짚신 신은 발이 보인다. 읽을 수 있는 글자는 없다.",
        "hard_violations": [
         "사람과 얼굴을 배제하라는 조건을 어기고 노인의 얼굴과 상체를 화면 전경에 크게 등장시켰다.",
         "참조에서 옆으로 웅크리고 손을 얼굴 앞 바닥에 둔 쓰러진 인물을, 상체를 들고 팔을 자루까지 뻗어 잡는 자세로 바꾸어 고정 자세 조건을 위반했다."
        ],
        "physics": "자루의 무게는 지면이 받치고 윗부분에는 손이 닿아 잡고 있다. 인물의 뺨과 짚신 신은 발도 땅에 닿아 있어 완전히 지지 없이 뜬 몸은 아니다. 다만 얼굴을 바닥에 댄 채 상체를 높이고 자루를 잡은 자세는 참조의 움직이지 않는 쓰러진 상태와 다른 행동이다."
       }
      ],
      "all_candidates_fail": true
     },
     "normalized": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 2,
        "verdict_ko": "금지된 인물들이 등장해 탈락이지만, B보다 자루 중심의 바닥 시점과 끌리는 동작을 더 가깝게 구현했다."
       },
       {
        "label": "A",
        "score": 1,
        "verdict_ko": "금지된 얼굴을 전경에 크게 넣고 쓰러진 인물의 고정 자세까지 바꾸어, 인물 없는 중경 자루 시점을 크게 벗어났다."
       }
      ],
      "all_candidates_fail": true,
      "readings": [
       {
        "label": "B",
        "direction": "자루의 묶인 윗부분을 화면 위쪽의 두 손이 잡고 있으며, 아래쪽 흙의 끌림 자국은 카메라 쪽으로 이어진다. 멀어지는 동작은 읽히지만 지정된 우상단 출구보다는 화면 위쪽으로 당겨지는 방향이다. 왼쪽 인물의 얼굴은 땅을 향하고, 자루를 보는 눈은 확인되지 않는다.",
        "built_space": "왼쪽 배경에 격자 문짝 두 면, 흙벽과 목재 기둥, 낮은 돌 기단이 있고 중앙 뒤에는 낮은 초가 구조물이 보인다. 참조의 농가 재료와 낮 조명은 대체로 이어지지만, 오른쪽 창고 출입구는 거의 프레임 밖이다. 카메라는 지면에 가깝지만 자루가 중경보다 크고 가깝게 차지한다.",
        "entities": "거친 갈색 직물의 곡물 자루 하나와 연속된 흙바닥이 보이며, 자루 입구는 모아져 손에 잡혀 있다. 내용물이 씨나락인지는 외관만으로 확인할 수 없다. 왼쪽에는 참조의 노년 한국인 남성처럼 보이는 회색 머리와 베이지색 옷의 쓰러진 인물이 있고, 위쪽에는 자루를 잡은 다른 인물의 손뿐 아니라 몸통과 옷도 보인다. 읽을 수 있는 글자는 없다.",
        "hard_violations": [
         "인물 없는 장면이라는 명시적 조건을 어기고, 쓰러진 사람의 얼굴과 몸 및 자루를 당기는 사람의 몸통을 노출했다. 허용된 손·팔만의 등장을 넘는다."
        ],
        "physics": "자루 아랫면은 흙바닥에 눌려 있고 윗부분은 두 손이 잡아 지지한다. 무게를 바닥에 둔 채 당기는 동작과 흙의 긁힌 자국은 물리적으로 가능하다. 누운 인물도 머리와 몸이 지면에 닿아 있으며, 지지 없이 떠 있는 물체는 보이지 않는다."
       },
       {
        "label": "A",
        "direction": "자루는 오른쪽에 있지만, 왼쪽 인물 쪽에서 뻗은 손이 윗부분을 잡고 있어 우상단으로 멀어지는 견인 방향이 명확하지 않다. 노인의 눈은 감기거나 아래로 향해 있어 자루를 바라보는 시선은 보이지 않는다. 화면은 흐려진 주관 시점보다 얼굴 옆에서 찍은 외부 시점으로 읽힌다.",
        "built_space": "오른쪽 배경에 어두운 창고 출입구 하나, 열린 목재 문짝, 흙벽, 목재 기둥과 돌 기단이 보이고 뒤에는 낮은 초가 구조물이 있다. 참조의 창고 쪽 재료는 비교적 잘 이어진다. 그러나 거대한 얼굴과 상체가 왼쪽 전경을 채워 지정된 중간 크기의 자루 중심 구도 대신 인물 근접 구도가 되었다.",
        "entities": "오른쪽에 갈색 직물 자루 하나와 흙바닥이 보인다. 자루의 입구 쪽은 손에 가려져 묶인 상태가 확실하지 않고, 씨나락 내용물도 보이지 않는다. 왼쪽에는 참조 노인과 유사한 회색 머리의 한국인 노년 남성 얼굴과 베이지색 상체가 크게 보이며 중앙에는 짚신 신은 발이 보인다. 읽을 수 있는 글자는 없다.",
        "hard_violations": [
         "사람과 얼굴을 배제하라는 조건을 어기고 노인의 얼굴과 상체를 화면 전경에 크게 등장시켰다.",
         "참조에서 옆으로 웅크리고 손을 얼굴 앞 바닥에 둔 쓰러진 인물을, 상체를 들고 팔을 자루까지 뻗어 잡는 자세로 바꾸어 고정 자세 조건을 위반했다."
        ],
        "physics": "자루의 무게는 지면이 받치고 윗부분에는 손이 닿아 잡고 있다. 인물의 뺨과 짚신 신은 발도 땅에 닿아 있어 완전히 지지 없이 뜬 몸은 아니다. 다만 얼굴을 바닥에 댄 채 상체를 높이고 자루를 잡은 자세는 참조의 움직이지 않는 쓰러진 상태와 다른 행동이다."
       }
      ]
     },
     "ok": true
    }
   ],
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "slot_winner_match": true,
   "slot_winner": {
    "gemini-pro": "B",
    "gpt-high": "B"
   },
   "route": "cross_slot_agree"
  },
  "dual": {
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "normalized": {
    "A": 0.5,
    "B": 2.0
   },
   "adjusted": {
    "A": 0.25,
    "B": 1.75
   },
   "violations": {
    "A": [
     "[gemini-pro] 물리적으로 불가능한 신체 구조 (얼굴 바로 옆에 발이 위치하고, 몸통 없이 팔만 튀어나오는 등 신체 부위가 기괴하게 얽혀 있음).",
     "[gemini-pro] '인물 없음(NO PEOPLE)' 및 '얼굴 없음' 규칙을 명백히 위반하여 프레임 안에 큰 얼굴을 포함함.",
     "[gpt-high] 사람과 얼굴을 배제하라는 조건을 어기고 노인의 얼굴과 상체를 화면 전경에 크게 등장시켰다.",
     "[gpt-high] 참조에서 옆으로 웅크리고 손을 얼굴 앞 바닥에 둔 쓰러진 인물을, 상체를 들고 팔을 자루까지 뻗어 잡는 자세로 바꾸어 고정 자세 조건을 위반했다."
    ],
    "B": [
     "[gemini-pro] 스테이징 위반: 쓰러진 안규홍의 시점(POV)이어야 하는 카메라 위치임에도 배경에 안규홍 본인을 중복해서 배치함.",
     "[gemini-pro] '인물 없음(NO PEOPLE)' 규칙을 위반하여 쓰러진 인물의 전신과 자루를 끄는 인물의 실루엣을 프레임 안에 포함함.",
     "[gpt-high] 인물 없는 장면이라는 명시적 조건을 어기고, 쓰러진 사람의 얼굴과 몸 및 자루를 당기는 사람의 몸통을 노출했다. 허용된 손·팔만의 등장을 넘는다."
    ]
   },
   "per_model_winner": {
    "gemini-pro": "B",
    "gpt-high": "B"
   },
   "agreed": true
  },
  "totals": {
   "B": 1750,
   "A": 250
  },
  "selected": "B",
  "ranking": [
   "B",
   "A"
  ],
  "verdicts": [
   {
    "label": "B",
    "score": 1750,
    "verdict_ko": "카메라가 쓰러진 안규홍의 시야(시점)여야 함에도 배경에 안규홍을 배치하여 스테이징 논리를 근본적으로 위반했으며, '인물 없음' 규칙을 무시하고 화면에 사람을 포함시켜 치명적인 오류를 범했습니다.  ★위반: [gemini-pro] 스테이징 위반: 쓰러진 안규홍의 시점(POV)이어야 하는 카메라 위치임에도 배경에 안규홍 본인을 중복해서 배치함. / [gemini-pro] '인물 없음(NO PEOPLE)' 규칙을 위반하여 쓰러진 인물의 전신과 자루를 끄는 인물의 실루엣을 프레임 안에 포함함. / [gpt-high] 인물 없는 장면이라는 명시적 조건을 어기고, 쓰러진 사람의 얼굴과 몸 및 자루를 당기는 사람의 몸통을 노출했다. 허용된 손·팔만의 등장을 넘는다."
   },
   {
    "label": "A",
    "score": 250,
    "verdict_ko": "프롬프트의 '인물 및 얼굴 없음' 지시를 전면 위반하여 얼굴을 크게 노출시켰으며, 바닥에 닿은 얼굴 바로 옆에 발이 위치하는 등 인체 구조가 물리적으로 불가능하게 기괴하게 얽혀 있어 완전히 실패한 이미지입니다.  ★위반: [gemini-pro] 물리적으로 불가능한 신체 구조 (얼굴 바로 옆에 발이 위치하고, 몸통 없이 팔만 튀어나오는 등 신체 부위가 기괴하게 얽혀 있음). / [gemini-pro] '인물 없음(NO PEOPLE)' 및 '얼굴 없음' 규칙을 명백히 위반하여 프레임 안에 큰 얼굴을 포함함. / [gpt-high] 사람과 얼굴을 배제하라는 조건을 어기고 노인의 얼굴과 상체를 화면 전경에 크게 등장시켰다. / [gpt-high] 참조에서 옆으로 웅크리고 손을 얼굴 앞 바닥에 둔 쓰러진 인물을, 상체를 들고 팔을 자루까지 뻗어 잡는 자세로 바꾸어 고정 자세 조건을 위반했다."
   }
  ],
  "refs": [
   {
    "label": "PREVIOUS SHOT STILL — a visually related earlier shot of this same place: the location's look, materials, fixed features and lighting mood are LOCKED to this photo; never copy its camera framing. The people visible in that photo are NOT in this shot: never carry their faces, bodies or clothing onto anyone here. Each person in THIS shot is defined solely by the PEOPLE list, their own CHARACTER REFERENCE images and their wardrobe notes. If this photo shows a character who cannot move (dead or unconscious), that character's exact pose, position and orientation are ALSO LOCKED — treat their body as a fixed prop of the set that only the camera moves around.",
    "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/db00517b-165b-4005-a287-f28936b59b95/images/272bd7af-9294-4490-98c6-88a40e9deacc/scene/recipe/S16sh7_sel.png",
    "asset_id": "f7b6f939-cb67-44e8-a4f0-c09d6863d594",
    "role": "prev_still"
   }
  ],
  "critique_skipped": true,
  "needs_reshoot": true,
  "shot_run_produced": false,
  "shot_run_uid": "06aa1411-4dcb-7fc2-bb3d-2707e1a84179",
  "ref_mode": "prev만 (배경 전용·공유 계획)",
  "share_plan": {
   "ref_plan": "prev",
   "prev_anchor_tag": "S16sh7"
  },
  "lane_policy": "share_plan_prev_bgonly"
 },
 "S16sh16::cine": {
  "applied": true,
  "attempted_at": "2026-09-09T02:37:52.995733+00:00",
  "fingerprint": "780b23156240342ec0affec5075ea9d4c0758bea0e751684a678b13c8c71ac70",
  "fingerprint_version": 2,
  "provider": "mai",
  "endpoint": "openrouter/chat-completions",
  "model": "microsoft/mai-image-2.6",
  "pack": "24.202608252115",
  "source_file": "S16sh16_sel.png",
  "source_sha256": "b488a022f0b6995e20ca874dde758a066f9b45a29b2ae47613c112af22871910",
  "file": "S16sh16_cine.png",
  "staged_sha256": "6b1bd803474dfdbcedb2ac1d419583ba2f434b7e9494f8c091d021f2b6bbba0f",
  "latency_ms": 26068
 },
 "S17sh8::signage": {
  "fp": "f586a817131edc42",
  "inscriptions": [],
  "cues": [],
  "dropped": []
 },
 "groupbg::lake_open_ground": {
  "input_fingerprint": "3c7a1c536272ac6c",
  "meta": {
   "model": "gpt-image-2",
   "size": "1536x864",
   "pack": "11.202607220237",
   "contract": "bgfirst_full_v3",
   "group_sig": {
    "key": "lake_open_ground",
    "tags": [
     "S17sh18",
     "S17sh21",
     "S17sh8"
    ]
   },
   "context_sig": "345f04284682b840"
  },
  "prompt": "Create ONE empty live-action location background photograph — NO PEOPLE, no figures, no body parts, no silhouettes, no shadows or reflections of people anywhere.\n\"Empty\" means no people only: KEEP the location's inherent occupants and stock that define the place — animals in an animal shelter, pen or farm, goods and produce in a market, moored boats in a harbour — unless the shot text explicitly removes them.\nThe FIRST attached image is a thin-line storyboard sketch of a shot that happens at this location: use it ONLY as spatial evidence — what this place contains, how its ground, structures and landmarks are arranged and proportioned. Ignore the sketched people and arrows entirely, and do NOT copy its line style: render a fully photographic, physically plausible real place that fits THE LOCATION text below.\nHUMAN-SCALE CALIBRATION: derive every structure's true size from human-scale elements — a door ≈ 2m, a window ≈ 1–1.5m wide, one storey ≈ 2.5–3m; never inflate a small structure or shrink a large one.\n\nTHE LOCATION — early 20th-century Korea, Jeolla region; all people are Korean unless explicitly stated otherwise: The open assembly ground beside the lake in a temporary camp, ringed by onlookers under slanting afternoon sunlight.\n\nLOCATION DETAIL (descriptive evidence about this place from the production's location records — evidence, not a staging order): use it to understand what kind of place this is and what it permanently contains. Carry over ONLY its enduring physical features — layout, structures, ground and materials, fixed equipment and installations. Do NOT treat any lighting, weather or time-of-day wording, momentary object states, depicted events or subjective impressions in it as fixed facts of the place; the TIME OF DAY lock below is the sole authority for time and lighting.\n보성 호숫가 강용언 임시거점 자갈밭·공터·막사 앞: 잿빛 호숫가에 펼쳐진 자갈밭과 흙 공터. 대나무 숲을 등지고 있으며, 낡은 천과 거적을 엮어 만든 조잡한 임시 막사 여러 채, 큰 무쇠솥이 얹힌 돌 화덕이 배치되어 있다. (특징: 넓은 잿빛 호수면; 자갈과 흙이 섞인 공터; 낡은 천, 짚, 거적으로 거칠게 엮은 임시 천막/막사; 불이 타오르는 돌 화덕; 끓고 있는 큰 무쇠 가마솥; 쌓여 있는 짐 꾸러미와 곡식 자루)\n\nSCENE EVIDENCE (verbatim quotes from the screenplay about this place — treat them as evidence of what the location physically contains and looks like; stage the PLACE those moments happen in, but do NOT depict the momentary actions, people or staged props themselves):\n- 호숫가옆 공터. 의병들과 규홍 일행이 숨을 죽인 채 둘러서 있다.\n그 한가운데, 민가에서 가져온 씨나락 자루 두 개가 놓여 있다.\n\nTIME OF DAY (lock): late afternoon, low reddish sunlight.\n\nRender ONE photorealistic empty location photograph, 16:9, neutral enough that every shot of this place can be staged from it later. No readable writing anywhere: surfaces that would carry writing may be present, but stage any wording out of legibility — an oblique angle, distance, shallow focus. No captions, watermarks or overlay text.",
  "effective_prompt": "Create ONE empty live-action location background photograph — NO PEOPLE, no figures, no body parts, no silhouettes, no shadows or reflections of people anywhere.\n\"Empty\" means no people only: KEEP the location's inherent occupants and stock that define the place — animals in an animal shelter, pen or farm, goods and produce in a market, moored boats in a harbour — unless the shot text explicitly removes them.\nThe FIRST attached image is a thin-line storyboard sketch of a shot that happens at this location: use it ONLY as spatial evidence — what this place contains, how its ground, structures and landmarks are arranged and proportioned. Ignore the sketched people and arrows entirely, and do NOT copy its line style: render a fully photographic, physically plausible real place that fits THE LOCATION text below.\nHUMAN-SCALE CALIBRATION: derive every structure's true size from human-scale elements — a door ≈ 2m, a window ≈ 1–1.5m wide, one storey ≈ 2.5–3m; never inflate a small structure or shrink a large one.\n\nTHE LOCATION — early 20th-century Korea, Jeolla region; all people are Korean unless explicitly stated otherwise: The open assembly ground beside the lake in a temporary camp, ringed by onlookers under slanting afternoon sunlight.\n\nLOCATION DETAIL (descriptive evidence about this place from the production's location records — evidence, not a staging order): use it to understand what kind of place this is and what it permanently contains. Carry over ONLY its enduring physical features — layout, structures, ground and materials, fixed equipment and installations. Do NOT treat any lighting, weather or time-of-day wording, momentary object states, depicted events or subjective impressions in it as fixed facts of the place; the TIME OF DAY lock below is the sole authority for time and lighting.\n보성 호숫가 강용언 임시거점 자갈밭·공터·막사 앞: 잿빛 호숫가에 펼쳐진 자갈밭과 흙 공터. 대나무 숲을 등지고 있으며, 낡은 천과 거적을 엮어 만든 조잡한 임시 막사 여러 채, 큰 무쇠솥이 얹힌 돌 화덕이 배치되어 있다. (특징: 넓은 잿빛 호수면; 자갈과 흙이 섞인 공터; 낡은 천, 짚, 거적으로 거칠게 엮은 임시 천막/막사; 불이 타오르는 돌 화덕; 끓고 있는 큰 무쇠 가마솥; 쌓여 있는 짐 꾸러미와 곡식 자루)\n\nSCENE EVIDENCE (verbatim quotes from the screenplay about this place — treat them as evidence of what the location physically contains and looks like; stage the PLACE those moments happen in, but do NOT depict the momentary actions, people or staged props themselves):\n- 호숫가옆 공터. 의병들과 규홍 일행이 숨을 죽인 채 둘러서 있다.\n그 한가운데, 민가에서 가져온 씨나락 자루 두 개가 놓여 있다.\n\nTIME OF DAY (lock): late afternoon, low reddish sunlight.\n\nRender ONE photorealistic empty location photograph, 16:9, neutral enough that every shot of this place can be staged from it later. No readable writing anywhere: surfaces that would carry writing may be present, but stage any wording out of legibility — an oblique angle, distance, shallow focus. No captions, watermarks or overlay text.",
  "bg_path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/db00517b-165b-4005-a287-f28936b59b95/images/272bd7af-9294-4490-98c6-88a40e9deacc/scene/recipe/groupbg_lake_open_ground_bdb4de.png",
  "asset_id": "3e7f42e9-dcbe-4340-9422-44dd23901bc6",
  "input_asset_ids": [
   "1fe19e6c-d8b6-47dc-859d-698afc7389dc"
  ],
  "origin_tag": "S17sh8",
  "place_text": "The open assembly ground beside the lake in a temporary camp, ringed by onlookers under slanting afternoon sunlight.",
  "origin_inputs": {
   "place_text": "The open assembly ground beside the lake in a temporary camp, ringed by onlookers under slanting afternoon sunlight.",
   "time_of_day_en": "late afternoon, low reddish sunlight",
   "conti_asset_id": "1fe19e6c-d8b6-47dc-859d-698afc7389dc"
  }
 },
 "S17sh8::bgfirst_bg": {
  "input_fingerprint": "45d8277ff2d49f9f",
  "prompt": "Create the EMPTY BACKGROUND PLATE for one film shot — NO PEOPLE, no figures, no body parts, no sketch lines, no arrows anywhere.\n\"Empty\" means no people only: KEEP the location's inherent occupants and stock that define the place — animals in an animal shelter, pen or farm, goods in a market, moored boats in a harbour — unless the shot text explicitly removes them.\nThe FIRST attached image is a thin-line storyboard sketch: use ONLY its camera angle, horizon, perspective and the placement/size of buildings and set masses — ignore the sketched people and arrows entirely. The SECOND attached image (LOCATION PHOTOGRAPH) is the real place: take its architecture, materials, signage and fixed features, and RE-PROJECT them into the sketch's camera. If the photograph's camera differs from the sketch's, the sketch's camera wins.\nHUMAN-SCALE CALIBRATION: derive every structure's true size from human-scale elements — a door ≈ 2m, a window ≈ 1–1.5m wide, one storey ≈ 2.5–3m; never inflate a small structure or shrink a large one.\n\nSHOT TEXT this background must serve (Korean): 안규홍의 코앞까지 성큼 바짝 다가선 강용언의 위협적인 상체.\n\nLOCATION (lock): The open assembly ground beside the lake in a temporary camp, ringed by onlookers under slanting afternoon sunlight.\n\nTIME OF DAY (lock): late afternoon, low reddish sunlight.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: medium shot\n- KEY BACKGROUND ELEMENTS: 호수 (Gray beyond the lakeside clearing); used as A narrow, soft background interval preserves the established lakeside location.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Slanting afternoon sunlight carries the scene's stated reddish cast with restrained saturation and firm natural contrast, leaving the confrontation exposed rather than theatrically spotlit.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nRender ONE photorealistic empty location photograph, 16:9, that this shot can be staged inside later. No readable writing anywhere: surfaces that would carry writing may be present, but stage any wording out of legibility — an oblique angle, distance, shallow focus. No captions, watermarks or overlay text.",
  "effective_prompt": "Create the EMPTY BACKGROUND PLATE for one film shot — NO PEOPLE, no figures, no body parts, no sketch lines, no arrows anywhere.\n\"Empty\" means no people only: KEEP the location's inherent occupants and stock that define the place — animals in an animal shelter, pen or farm, goods in a market, moored boats in a harbour — unless the shot text explicitly removes them.\nThe FIRST attached image is a thin-line storyboard sketch: use ONLY its camera angle, horizon, perspective and the placement/size of buildings and set masses — ignore the sketched people and arrows entirely. The SECOND attached image (LOCATION PHOTOGRAPH) is the real place: take its architecture, materials, signage and fixed features, and RE-PROJECT them into the sketch's camera. If the photograph's camera differs from the sketch's, the sketch's camera wins.\nHUMAN-SCALE CALIBRATION: derive every structure's true size from human-scale elements — a door ≈ 2m, a window ≈ 1–1.5m wide, one storey ≈ 2.5–3m; never inflate a small structure or shrink a large one.\n\nSHOT TEXT this background must serve (Korean): 안규홍의 코앞까지 성큼 바짝 다가선 강용언의 위협적인 상체.\n\nLOCATION (lock): The open assembly ground beside the lake in a temporary camp, ringed by onlookers under slanting afternoon sunlight.\n\nTIME OF DAY (lock): late afternoon, low reddish sunlight.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: medium shot\n- KEY BACKGROUND ELEMENTS: 호수 (Gray beyond the lakeside clearing); used as A narrow, soft background interval preserves the established lakeside location.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Slanting afternoon sunlight carries the scene's stated reddish cast with restrained saturation and firm natural contrast, leaving the confrontation exposed rather than theatrically spotlit.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nRender ONE photorealistic empty location photograph, 16:9, that this shot can be staged inside later. No readable writing anywhere: surfaces that would carry writing may be present, but stage any wording out of legibility — an oblique angle, distance, shallow focus. No captions, watermarks or overlay text.",
  "bg_path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/db00517b-165b-4005-a287-f28936b59b95/images/272bd7af-9294-4490-98c6-88a40e9deacc/scene/recipe/S17sh8__bgfirst_bg.png",
  "asset_id": "a456262d-cccd-4851-9155-e55c8779db35",
  "input_asset_ids": [
   "1fe19e6c-d8b6-47dc-859d-698afc7389dc",
   "3e7f42e9-dcbe-4340-9422-44dd23901bc6"
  ]
 },
 "S17sh8": {
  "input_fingerprint": "6b9665404d2cd41f",
  "prompt": "Create ONE FINAL photorealistic live-action film still of the moment below — early 20th-century Korea, Jeolla region; all people are Korean unless explicitly stated otherwise. TIME OF DAY (lock): late afternoon, low reddish sunlight.\n\nSHOT TEXT (authoritative, Korean): 안규홍의 코앞까지 성큼 바짝 다가선 강용언의 위협적인 상체.\n\nLOCATION (lock): The open assembly ground beside the lake in a temporary camp, ringed by onlookers under slanting afternoon sunlight. The shot takes place here — the attached LOCATION PHOTOGRAPH shows the exact spot.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: medium shot\n- KEY BACKGROUND ELEMENTS: 호수 (Gray beyond the lakeside clearing); used as A narrow, soft background interval preserves the established lakeside location.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Slanting afternoon sunlight carries the scene's stated reddish cast with restrained saturation and firm natural contrast, leaving the confrontation exposed rather than theatrically spotlit.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): Two tied seed-rice sacks lie in the camp clearing under slanting red afternoon light. 안규홍: He stands with the blood wiped from his head wound, wearing a carrying frame and concealing the jade ring. 강용언: He stands close in confrontation, with his sword still at his waist.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 안규홍 (한국인 남성, 청년 나이대, 검은 머리, 단단히 묶은 머리); 강용언 (한국인 남성, 성인 나이대, 햇볕에 그을린 얼굴, 자연스러운 검은 머리) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.",
  "roll_prompts": {
   "A": "Stage the shot. The FIRST attached image (SHOT BACKGROUND) is the finished empty background of this shot — keep it EXACTLY: its camera, perspective, architecture, lighting and every fixed feature stay untouched. The SECOND attached image (LAYOUT SKETCH) tells you ONLY where the people go: each sketched person's position, screen size, pose and the gaze/motion arrows. Ignore the sketch's background lines. The CHARACTER REFERENCE photographs show the real people.\nPlace the real people into the background at exactly the sketched positions, sizes and poses, following the arrow directions. No sketch lines or arrows may remain.\n\nCreate ONE FINAL photorealistic live-action film still of the moment below — early 20th-century Korea, Jeolla region; all people are Korean unless explicitly stated otherwise. TIME OF DAY (lock): late afternoon, low reddish sunlight.\n\nSHOT TEXT (authoritative, Korean): 안규홍의 코앞까지 성큼 바짝 다가선 강용언의 위협적인 상체.\n\nLOCATION (lock): The open assembly ground beside the lake in a temporary camp, ringed by onlookers under slanting afternoon sunlight. The shot takes place here — the FIRST attached image (SHOT BACKGROUND) is this exact place, already built: its ground, structures, horizon, materials and lighting are the finished truth of this location and must not be redesigned or replaced. No location photograph is attached — read the place from that image alone, and add no scenery, structure, vehicle or fixture that it does not already show. This lock governs the place only; the figures in the shot follow the staging and pose instructions.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: medium shot\n- KEY BACKGROUND ELEMENTS: 호수 (Gray beyond the lakeside clearing); used as A narrow, soft background interval preserves the established lakeside location.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Slanting afternoon sunlight carries the scene's stated reddish cast with restrained saturation and firm natural contrast, leaving the confrontation exposed rather than theatrically spotlit.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): Two tied seed-rice sacks lie in the camp clearing under slanting red afternoon light. 안규홍: He stands with the blood wiped from his head wound, wearing a carrying frame and concealing the jade ring. 강용언: He stands close in confrontation, with his sword still at his waist.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 안규홍 (한국인 남성, 청년 나이대, 검은 머리, 단단히 묶은 머리); 강용언 (한국인 남성, 성인 나이대, 햇볕에 그을린 얼굴, 자연스러운 검은 머리) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.",
   "B": "Create ONE FINAL photorealistic live-action film still of the moment below — early 20th-century Korea, Jeolla region; all people are Korean unless explicitly stated otherwise. TIME OF DAY (lock): late afternoon, low reddish sunlight.\n\nSHOT TEXT (authoritative, Korean): 안규홍의 코앞까지 성큼 바짝 다가선 강용언의 위협적인 상체.\n\nLOCATION (lock): The open assembly ground beside the lake in a temporary camp, ringed by onlookers under slanting afternoon sunlight. The shot takes place here — the attached LOCATION PHOTOGRAPH shows the exact spot.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: medium shot\n- KEY BACKGROUND ELEMENTS: 호수 (Gray beyond the lakeside clearing); used as A narrow, soft background interval preserves the established lakeside location.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Slanting afternoon sunlight carries the scene's stated reddish cast with restrained saturation and firm natural contrast, leaving the confrontation exposed rather than theatrically spotlit.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): Two tied seed-rice sacks lie in the camp clearing under slanting red afternoon light. 안규홍: He stands with the blood wiped from his head wound, wearing a carrying frame and concealing the jade ring. 강용언: He stands close in confrontation, with his sword still at his waist.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 안규홍 (한국인 남성, 청년 나이대, 검은 머리, 단단히 묶은 머리); 강용언 (한국인 남성, 성인 나이대, 햇볕에 그을린 얼굴, 자연스러운 검은 머리) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.\n\nCOMPOSITION VARIATION (this candidate only): keep the place, the people and the moment exactly as specified, and keep from the CAMERA & FRAME contract its subject scale, screen placement and key background elements exactly. For this candidate only, its camera angle and height are deliberately left open: choose a camera position distinctly different from the obvious one, as a film director picking a second setup on the same blocking. Stay true to any explicit framing the SHOT TEXT itself dictates."
  },
  "roll_refs": {
   "A": [
    {
     "label": "SHOT BACKGROUND",
     "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/db00517b-165b-4005-a287-f28936b59b95/images/272bd7af-9294-4490-98c6-88a40e9deacc/scene/recipe/S17sh8__bgfirst_bg.png",
     "asset_id": "a456262d-cccd-4851-9155-e55c8779db35",
     "role": "bgfirst_bg"
    },
    {
     "label": "LAYOUT SKETCH (people placement only)",
     "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/db00517b-165b-4005-a287-f28936b59b95/images/272bd7af-9294-4490-98c6-88a40e9deacc/conti/conti_S17sh8.png",
     "asset_id": "1fe19e6c-d8b6-47dc-859d-698afc7389dc",
     "role": "conti_light"
    },
    {
     "label": "CHARACTER REFERENCE — 안규홍: the exact person appearing in this shot; match face, hair and build exactly.",
     "path": "<bytes:1028384>",
     "asset_id": "3b5894db-ce55-4f34-b966-5e725c3aa952",
     "role": "character_ref"
    },
    {
     "label": "CHARACTER REFERENCE — 강용언: the exact person appearing in this shot; match face, hair and build exactly.",
     "path": "<bytes:822588>",
     "asset_id": "1a75090f-1eb9-4224-a132-de7a4b454b73",
     "role": "character_ref"
    }
   ],
   "B": [
    {
     "label": "LOCATION PHOTOGRAPH — the exact place of this shot: its architecture, materials, fixed features and lighting mood are spatial truth; stage the moment inside this place. Never copy its camera framing.",
     "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/db00517b-165b-4005-a287-f28936b59b95/images/272bd7af-9294-4490-98c6-88a40e9deacc/scene/recipe/groupbg_lake_open_ground_bdb4de.png",
     "asset_id": "3e7f42e9-dcbe-4340-9422-44dd23901bc6",
     "role": "bgfirst_group_bg"
    },
    {
     "label": "CHARACTER REFERENCE — 안규홍: the exact person appearing in this shot; match face, hair and build exactly.",
     "path": "<bytes:1028384>",
     "asset_id": "3b5894db-ce55-4f34-b966-5e725c3aa952",
     "role": "character_ref"
    },
    {
     "label": "CHARACTER REFERENCE — 강용언: the exact person appearing in this shot; match face, hair and build exactly.",
     "path": "<bytes:822588>",
     "asset_id": "1a75090f-1eb9-4224-a132-de7a4b454b73",
     "role": "character_ref"
    }
   ]
  },
  "shot_run_spend_attempt_count": 1,
  "judge_flip_skipped": "judge_fn_owns_order",
  "initial_roll_all_fail": true,
  "readings": [
   {
    "label": "A",
    "direction": "강용언은 비스듬히 몸을 숙여 안규홍을 내려다보고 있으며, 안규홍은 등진 채로 그를 마주보고 있음.",
    "built_space": "호숫가 옆 임시 숙영지 공터. 왼편에 대나무 숲과 초가 막사들이 장소 레퍼런스와 일치하게 배치되어 있으며, 배경에 구경꾼들이 서 있음.",
    "entities": "안규홍(뒷모습, 단단히 묶은 머리, 지게 착용, 뒷모습이라 상처는 보이지 않음), 강용언(푸른색 의상과 검 착용, 햇볕에 그을린 얼굴, 그러나 레퍼런스의 '갓'이 누락됨), 배경에 여러 개의 자루가 흩어져 있음(프롬프트가 요구한 '두 개'가 아님).",
    "hard_violations": [
     "[gemini-pro] 지게의 오른쪽 가로 대나무 기둥이 안규홍의 어깨 의복을 뚫고 융합되어 물리적으로 불가능한 형태를 띰.",
     "[gemini-pro] 캐릭터 레퍼런스에 명시된 강용언의 필수 의상인 갓(모자)이 누락됨.",
     "[gpt-high] 등장인물을 안규홍과 강용언으로 제한한 명시적 조건과 달리 공터 양옆과 중앙에 다수의 구경꾼을 추가했다.",
     "[gpt-high] 공터의 볏자루를 두 개로 고정한 조건과 달리 최소 세 개를 배치했다."
    ],
    "physics": "두 인물 모두 땅에 서서 체중을 지탱하고 있으나, 안규홍의 지게 구조물이 신체와 물리적으로 잘못 융합되어 지지 기반이 비현실적임."
   },
   {
    "label": "B",
    "direction": "강용언이 안규홍의 코앞까지 다가가 서로의 눈을 수평선상에서 똑바로 응시하고 있음.",
    "built_space": "호숫가 공터, 배경에 회색빛 호수와 구경꾼들이 위치하여 장소 레퍼런스의 공간적 특징을 정확히 반영함.",
    "entities": "안규홍(측면, 묶은 머리, 이마의 피 묻은 상처, 지게 착용), 강용언(푸른 의상, 갓 착용, 검 착용), 바닥에 정확히 두 개의 묶인 볏짚 자루가 놓여 있음.",
    "hard_violations": [
     "[gemini-pro] 강용언의 턱 피부에서 붉은색 장식술이 자라나듯 융합되어 해부학적으로 불가능한 형태가 생성됨.",
     "[gemini-pro] 안규홍 가슴 쪽의 흰색 어깨 끈이 등 뒤의 나무 지게 틀과 전혀 연결되지 않은 채 끊어져 있어 물리적으로 성립할 수 없음.",
     "[gpt-high] 등장인물을 안규홍과 강용언으로 제한한 명시적 조건과 달리 구경꾼 네 명이 추가되어 있다."
    ],
    "physics": "두 인물 모두 땅에 안정적으로 서 있으나, 안규홍의 지게는 어깨 끈의 연결 지지가 전혀 없이 등 뒤에 마법처럼 떠 있으며, 강용언의 붉은 장식은 모자가 아닌 피부에 고정되어 있음."
   }
  ],
  "cross_model_order": {
   "policy": "cross_model_order_v2_slot0_fwd_slot1_rev_parallel2",
   "slots": [
    {
     "model": "gemini-pro",
     "order": "forward",
     "display_to_canonical": {
      "A": "A",
      "B": "B"
     },
     "raw": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 5,
        "verdict_ko": "텍스트의 '코앞까지' 다가선 위협적인 구도와 요구된 디테일(상처, 두 개의 자루, 갓 착용)을 완벽히 구현했으나, 턱에 융합된 장식과 끈이 연결되지 않은 지게 등 치명적인 물리적/해부학적 오류로 인해 최종 사용은 불가합니다."
       },
       {
        "label": "A",
        "score": 3,
        "verdict_ko": "샷 텍스트가 요구한 '코앞까지 바짝 다가선' 거리감을 살리지 못했으며, 강용언의 필수 의상인 갓이 누락되고 지게가 몸에 융합되는 물리적 오류가 발생했습니다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "강용언은 비스듬히 몸을 숙여 안규홍을 내려다보고 있으며, 안규홍은 등진 채로 그를 마주보고 있음.",
        "built_space": "호숫가 옆 임시 숙영지 공터. 왼편에 대나무 숲과 초가 막사들이 장소 레퍼런스와 일치하게 배치되어 있으며, 배경에 구경꾼들이 서 있음.",
        "entities": "안규홍(뒷모습, 단단히 묶은 머리, 지게 착용, 뒷모습이라 상처는 보이지 않음), 강용언(푸른색 의상과 검 착용, 햇볕에 그을린 얼굴, 그러나 레퍼런스의 '갓'이 누락됨), 배경에 여러 개의 자루가 흩어져 있음(프롬프트가 요구한 '두 개'가 아님).",
        "hard_violations": [
         "지게의 오른쪽 가로 대나무 기둥이 안규홍의 어깨 의복을 뚫고 융합되어 물리적으로 불가능한 형태를 띰.",
         "캐릭터 레퍼런스에 명시된 강용언의 필수 의상인 갓(모자)이 누락됨."
        ],
        "physics": "두 인물 모두 땅에 서서 체중을 지탱하고 있으나, 안규홍의 지게 구조물이 신체와 물리적으로 잘못 융합되어 지지 기반이 비현실적임."
       },
       {
        "label": "B",
        "direction": "강용언이 안규홍의 코앞까지 다가가 서로의 눈을 수평선상에서 똑바로 응시하고 있음.",
        "built_space": "호숫가 공터, 배경에 회색빛 호수와 구경꾼들이 위치하여 장소 레퍼런스의 공간적 특징을 정확히 반영함.",
        "entities": "안규홍(측면, 묶은 머리, 이마의 피 묻은 상처, 지게 착용), 강용언(푸른 의상, 갓 착용, 검 착용), 바닥에 정확히 두 개의 묶인 볏짚 자루가 놓여 있음.",
        "hard_violations": [
         "강용언의 턱 피부에서 붉은색 장식술이 자라나듯 융합되어 해부학적으로 불가능한 형태가 생성됨.",
         "안규홍 가슴 쪽의 흰색 어깨 끈이 등 뒤의 나무 지게 틀과 전혀 연결되지 않은 채 끊어져 있어 물리적으로 성립할 수 없음."
        ],
        "physics": "두 인물 모두 땅에 안정적으로 서 있으나, 안규홍의 지게는 어깨 끈의 연결 지지가 전혀 없이 등 뒤에 마법처럼 떠 있으며, 강용언의 붉은 장식은 모자가 아닌 피부에 고정되어 있음."
       }
      ],
      "all_candidates_fail": true
     },
     "normalized": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 5,
        "verdict_ko": "텍스트의 '코앞까지' 다가선 위협적인 구도와 요구된 디테일(상처, 두 개의 자루, 갓 착용)을 완벽히 구현했으나, 턱에 융합된 장식과 끈이 연결되지 않은 지게 등 치명적인 물리적/해부학적 오류로 인해 최종 사용은 불가합니다."
       },
       {
        "label": "A",
        "score": 3,
        "verdict_ko": "샷 텍스트가 요구한 '코앞까지 바짝 다가선' 거리감을 살리지 못했으며, 강용언의 필수 의상인 갓이 누락되고 지게가 몸에 융합되는 물리적 오류가 발생했습니다."
       }
      ],
      "all_candidates_fail": true,
      "readings": [
       {
        "label": "A",
        "direction": "강용언은 비스듬히 몸을 숙여 안규홍을 내려다보고 있으며, 안규홍은 등진 채로 그를 마주보고 있음.",
        "built_space": "호숫가 옆 임시 숙영지 공터. 왼편에 대나무 숲과 초가 막사들이 장소 레퍼런스와 일치하게 배치되어 있으며, 배경에 구경꾼들이 서 있음.",
        "entities": "안규홍(뒷모습, 단단히 묶은 머리, 지게 착용, 뒷모습이라 상처는 보이지 않음), 강용언(푸른색 의상과 검 착용, 햇볕에 그을린 얼굴, 그러나 레퍼런스의 '갓'이 누락됨), 배경에 여러 개의 자루가 흩어져 있음(프롬프트가 요구한 '두 개'가 아님).",
        "hard_violations": [
         "지게의 오른쪽 가로 대나무 기둥이 안규홍의 어깨 의복을 뚫고 융합되어 물리적으로 불가능한 형태를 띰.",
         "캐릭터 레퍼런스에 명시된 강용언의 필수 의상인 갓(모자)이 누락됨."
        ],
        "physics": "두 인물 모두 땅에 서서 체중을 지탱하고 있으나, 안규홍의 지게 구조물이 신체와 물리적으로 잘못 융합되어 지지 기반이 비현실적임."
       },
       {
        "label": "B",
        "direction": "강용언이 안규홍의 코앞까지 다가가 서로의 눈을 수평선상에서 똑바로 응시하고 있음.",
        "built_space": "호숫가 공터, 배경에 회색빛 호수와 구경꾼들이 위치하여 장소 레퍼런스의 공간적 특징을 정확히 반영함.",
        "entities": "안규홍(측면, 묶은 머리, 이마의 피 묻은 상처, 지게 착용), 강용언(푸른 의상, 갓 착용, 검 착용), 바닥에 정확히 두 개의 묶인 볏짚 자루가 놓여 있음.",
        "hard_violations": [
         "강용언의 턱 피부에서 붉은색 장식술이 자라나듯 융합되어 해부학적으로 불가능한 형태가 생성됨.",
         "안규홍 가슴 쪽의 흰색 어깨 끈이 등 뒤의 나무 지게 틀과 전혀 연결되지 않은 채 끊어져 있어 물리적으로 성립할 수 없음."
        ],
        "physics": "두 인물 모두 땅에 안정적으로 서 있으나, 안규홍의 지게는 어깨 끈의 연결 지지가 전혀 없이 등 뒤에 마법처럼 떠 있으며, 강용언의 붉은 장식은 모자가 아닌 피부에 고정되어 있음."
       }
      ]
     },
     "ok": true
    },
    {
     "model": "gpt-high",
     "order": "reverse",
     "display_to_canonical": {
      "A": "B",
      "B": "A"
     },
     "raw": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 4,
        "verdict_ko": "가까이 맞선 강용언의 위협적인 상체와 참조 복식을 더 충실히 구현했지만, 금지된 구경꾼 네 명과 안규홍 이마에 남은 선명한 피가 요구에 어긋난다."
       },
       {
        "label": "B",
        "score": 2,
        "verdict_ko": "강용언의 전진 자세와 붉은 오후빛은 적절하지만, 금지된 다수 인물과 추가 볏자루가 등장하며 두 사람의 간격과 모자 누락도 충실도를 낮춘다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "오른쪽 강용언은 고개와 시선을 왼쪽 아래 안규홍의 얼굴로 향하고, 안규홍은 턱을 들어 강용언의 눈을 바라본다. 서로를 겨냥한 대치가 분명하다. 칼은 뽑지 않았고 허리에서 손잡이가 왼쪽 위로 향한다.",
        "built_space": "왼쪽에 천과 갈대로 만든 막사 일부, 뒤에 대나무숲과 산, 두 사람 사이에 좁은 호수 띠가 보인다. 중앙 아래에는 바닥에 놓인 볏자루 두 개가 있다. 참조의 호숫가 야영지 재료와 배치에 부합하며, 오른쪽 막사와 취사시설은 화면 밖이다. 두 주인공 외에 왼쪽 두 명, 오른쪽 두 명의 구경꾼이 보인다.",
        "entities": "안규홍은 검은 머리를 뒤로 묶은 한국인 청년 남성으로 보이며, 때 탄 흰옷과 어깨끈으로 멘 목제 지게가 보인다. 얼굴은 옆모습이라 참조와의 정확한 일치는 제한적으로 확인되며, 이마에는 닦였다고 보기 어려운 붉은 피가 남아 있다. 강용언은 그을린 얼굴의 한국인 성인 남성으로 보이고, 모자·청회색 군복·붉은 허리띠·허리의 칼이 참조와 대체로 맞는다. 옥가락지는 노출되지 않는다. 읽을 수 있는 글자는 없다.",
        "hard_violations": [
         "등장인물을 안규홍과 강용언으로 제한한 명시적 조건과 달리 구경꾼 네 명이 추가되어 있다."
        ],
        "physics": "강용언의 상체는 안규홍 쪽으로 약간 기울어 있고, 지게는 안규홍의 어깨끈과 등에 의해 지지된다. 지게 뒤 바구니는 목제 틀에 붙어 있다. 칼집은 허리띠에 연결되어 있고 볏자루 두 개는 땅에 놓여 있다. 두 사람의 발은 잘려 있어 보폭이나 정확한 지면 접촉은 확인할 수 없지만, 공중부양이나 불가능한 관절은 보이지 않는다."
       },
       {
        "label": "B",
        "direction": "강용언은 왼쪽 전경 안규홍의 얼굴을 똑바로 응시하며 상체를 앞으로 숙인다. 안규홍도 오른쪽 강용언을 향해 머리를 돌리고 있다. 대치 대상은 정확하지만 얼굴 사이에는 상당한 간격이 남아 '코앞까지 바짝'의 압박은 약하다. 칼은 칼집 안에 있고 손잡이는 왼쪽 위로 향한다.",
        "built_space": "왼쪽의 연속된 낮은 막사와 대나무숲, 오른쪽 천막 한 채, 중앙 뒤의 좁은 호수와 산이 참조 장소를 잘 드러낸다. 안규홍은 왼쪽 전경에 등을 보이고 강용언은 오른쪽에서 허벅지 윗부분까지 드러난다. 두 사람 사이로 넓은 공터가 보이며, 앞쪽 볏자루 두 개와 더 뒤쪽 볏자루 최소 한 개가 확인된다. 공터 양옆과 중앙에는 다수의 구경꾼이 서 있다.",
        "entities": "안규홍은 단단히 올려 묶은 검은 머리, 흰옷, 지게를 갖춘 한국인 청년 남성으로 보인다. 얼굴과 이마가 대부분 가려져 신원과 상처 상태를 충분히 확인할 수 없다. 강용언은 그을린 얼굴의 한국인 성인 남성으로 보이며 청회색 옷, 붉은 허리띠, 허리의 칼은 맞지만 참조의 모자가 없다. 옥가락지는 노출되지 않는다. 볏자루는 지정된 두 개보다 많고, 읽을 수 있는 글자는 없다.",
        "hard_violations": [
         "등장인물을 안규홍과 강용언으로 제한한 명시적 조건과 달리 공터 양옆과 중앙에 다수의 구경꾼을 추가했다.",
         "공터의 볏자루를 두 개로 고정한 조건과 달리 최소 세 개를 배치했다."
        ],
        "physics": "강용언의 전방 기울기는 허리와 하체로 이어지는 자연스러운 자세이며, 잘린 발 때문에 실제 전진 보폭은 확인할 수 없다. 안규홍의 지게는 어깨끈과 등에 지지되고, 칼집은 허리띠에 매달려 있다. 볏자루들은 지면에 놓이고 배경 인물들의 보이는 발도 땅에 닿아 있다. 지지 없이 떠 있는 물체나 명백히 불가능한 신체 자세는 없다."
       }
      ],
      "all_candidates_fail": true
     },
     "normalized": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 4,
        "verdict_ko": "가까이 맞선 강용언의 위협적인 상체와 참조 복식을 더 충실히 구현했지만, 금지된 구경꾼 네 명과 안규홍 이마에 남은 선명한 피가 요구에 어긋난다."
       },
       {
        "label": "A",
        "score": 2,
        "verdict_ko": "강용언의 전진 자세와 붉은 오후빛은 적절하지만, 금지된 다수 인물과 추가 볏자루가 등장하며 두 사람의 간격과 모자 누락도 충실도를 낮춘다."
       }
      ],
      "all_candidates_fail": true,
      "readings": [
       {
        "label": "B",
        "direction": "오른쪽 강용언은 고개와 시선을 왼쪽 아래 안규홍의 얼굴로 향하고, 안규홍은 턱을 들어 강용언의 눈을 바라본다. 서로를 겨냥한 대치가 분명하다. 칼은 뽑지 않았고 허리에서 손잡이가 왼쪽 위로 향한다.",
        "built_space": "왼쪽에 천과 갈대로 만든 막사 일부, 뒤에 대나무숲과 산, 두 사람 사이에 좁은 호수 띠가 보인다. 중앙 아래에는 바닥에 놓인 볏자루 두 개가 있다. 참조의 호숫가 야영지 재료와 배치에 부합하며, 오른쪽 막사와 취사시설은 화면 밖이다. 두 주인공 외에 왼쪽 두 명, 오른쪽 두 명의 구경꾼이 보인다.",
        "entities": "안규홍은 검은 머리를 뒤로 묶은 한국인 청년 남성으로 보이며, 때 탄 흰옷과 어깨끈으로 멘 목제 지게가 보인다. 얼굴은 옆모습이라 참조와의 정확한 일치는 제한적으로 확인되며, 이마에는 닦였다고 보기 어려운 붉은 피가 남아 있다. 강용언은 그을린 얼굴의 한국인 성인 남성으로 보이고, 모자·청회색 군복·붉은 허리띠·허리의 칼이 참조와 대체로 맞는다. 옥가락지는 노출되지 않는다. 읽을 수 있는 글자는 없다.",
        "hard_violations": [
         "등장인물을 안규홍과 강용언으로 제한한 명시적 조건과 달리 구경꾼 네 명이 추가되어 있다."
        ],
        "physics": "강용언의 상체는 안규홍 쪽으로 약간 기울어 있고, 지게는 안규홍의 어깨끈과 등에 의해 지지된다. 지게 뒤 바구니는 목제 틀에 붙어 있다. 칼집은 허리띠에 연결되어 있고 볏자루 두 개는 땅에 놓여 있다. 두 사람의 발은 잘려 있어 보폭이나 정확한 지면 접촉은 확인할 수 없지만, 공중부양이나 불가능한 관절은 보이지 않는다."
       },
       {
        "label": "A",
        "direction": "강용언은 왼쪽 전경 안규홍의 얼굴을 똑바로 응시하며 상체를 앞으로 숙인다. 안규홍도 오른쪽 강용언을 향해 머리를 돌리고 있다. 대치 대상은 정확하지만 얼굴 사이에는 상당한 간격이 남아 '코앞까지 바짝'의 압박은 약하다. 칼은 칼집 안에 있고 손잡이는 왼쪽 위로 향한다.",
        "built_space": "왼쪽의 연속된 낮은 막사와 대나무숲, 오른쪽 천막 한 채, 중앙 뒤의 좁은 호수와 산이 참조 장소를 잘 드러낸다. 안규홍은 왼쪽 전경에 등을 보이고 강용언은 오른쪽에서 허벅지 윗부분까지 드러난다. 두 사람 사이로 넓은 공터가 보이며, 앞쪽 볏자루 두 개와 더 뒤쪽 볏자루 최소 한 개가 확인된다. 공터 양옆과 중앙에는 다수의 구경꾼이 서 있다.",
        "entities": "안규홍은 단단히 올려 묶은 검은 머리, 흰옷, 지게를 갖춘 한국인 청년 남성으로 보인다. 얼굴과 이마가 대부분 가려져 신원과 상처 상태를 충분히 확인할 수 없다. 강용언은 그을린 얼굴의 한국인 성인 남성으로 보이며 청회색 옷, 붉은 허리띠, 허리의 칼은 맞지만 참조의 모자가 없다. 옥가락지는 노출되지 않는다. 볏자루는 지정된 두 개보다 많고, 읽을 수 있는 글자는 없다.",
        "hard_violations": [
         "등장인물을 안규홍과 강용언으로 제한한 명시적 조건과 달리 공터 양옆과 중앙에 다수의 구경꾼을 추가했다.",
         "공터의 볏자루를 두 개로 고정한 조건과 달리 최소 세 개를 배치했다."
        ],
        "physics": "강용언의 전방 기울기는 허리와 하체로 이어지는 자연스러운 자세이며, 잘린 발 때문에 실제 전진 보폭은 확인할 수 없다. 안규홍의 지게는 어깨끈과 등에 지지되고, 칼집은 허리띠에 매달려 있다. 볏자루들은 지면에 놓이고 배경 인물들의 보이는 발도 땅에 닿아 있다. 지지 없이 떠 있는 물체나 명백히 불가능한 신체 자세는 없다."
       }
      ]
     },
     "ok": true
    }
   ],
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "slot_winner_match": true,
   "slot_winner": {
    "gemini-pro": "B",
    "gpt-high": "B"
   },
   "route": "cross_slot_agree"
  },
  "dual": {
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "normalized": {
    "A": 1.1,
    "B": 2.0
   },
   "adjusted": {
    "A": 0.85,
    "B": 1.75
   },
   "violations": {
    "A": [
     "[gemini-pro] 지게의 오른쪽 가로 대나무 기둥이 안규홍의 어깨 의복을 뚫고 융합되어 물리적으로 불가능한 형태를 띰.",
     "[gemini-pro] 캐릭터 레퍼런스에 명시된 강용언의 필수 의상인 갓(모자)이 누락됨.",
     "[gpt-high] 등장인물을 안규홍과 강용언으로 제한한 명시적 조건과 달리 공터 양옆과 중앙에 다수의 구경꾼을 추가했다.",
     "[gpt-high] 공터의 볏자루를 두 개로 고정한 조건과 달리 최소 세 개를 배치했다."
    ],
    "B": [
     "[gemini-pro] 강용언의 턱 피부에서 붉은색 장식술이 자라나듯 융합되어 해부학적으로 불가능한 형태가 생성됨.",
     "[gemini-pro] 안규홍 가슴 쪽의 흰색 어깨 끈이 등 뒤의 나무 지게 틀과 전혀 연결되지 않은 채 끊어져 있어 물리적으로 성립할 수 없음.",
     "[gpt-high] 등장인물을 안규홍과 강용언으로 제한한 명시적 조건과 달리 구경꾼 네 명이 추가되어 있다."
    ]
   },
   "per_model_winner": {
    "gemini-pro": "B",
    "gpt-high": "B"
   },
   "agreed": true
  },
  "totals": {
   "B": 1750,
   "A": 850
  },
  "selected": "B",
  "ranking": [
   "B",
   "A"
  ],
  "verdicts": [
   {
    "label": "B",
    "score": 1750,
    "verdict_ko": "텍스트의 '코앞까지' 다가선 위협적인 구도와 요구된 디테일(상처, 두 개의 자루, 갓 착용)을 완벽히 구현했으나, 턱에 융합된 장식과 끈이 연결되지 않은 지게 등 치명적인 물리적/해부학적 오류로 인해 최종 사용은 불가합니다.  ★위반: [gemini-pro] 강용언의 턱 피부에서 붉은색 장식술이 자라나듯 융합되어 해부학적으로 불가능한 형태가 생성됨. / [gemini-pro] 안규홍 가슴 쪽의 흰색 어깨 끈이 등 뒤의 나무 지게 틀과 전혀 연결되지 않은 채 끊어져 있어 물리적으로 성립할 수 없음. / [gpt-high] 등장인물을 안규홍과 강용언으로 제한한 명시적 조건과 달리 구경꾼 네 명이 추가되어 있다."
   },
   {
    "label": "A",
    "score": 850,
    "verdict_ko": "샷 텍스트가 요구한 '코앞까지 바짝 다가선' 거리감을 살리지 못했으며, 강용언의 필수 의상인 갓이 누락되고 지게가 몸에 융합되는 물리적 오류가 발생했습니다.  ★위반: [gemini-pro] 지게의 오른쪽 가로 대나무 기둥이 안규홍의 어깨 의복을 뚫고 융합되어 물리적으로 불가능한 형태를 띰. / [gemini-pro] 캐릭터 레퍼런스에 명시된 강용언의 필수 의상인 갓(모자)이 누락됨. / [gpt-high] 등장인물을 안규홍과 강용언으로 제한한 명시적 조건과 달리 공터 양옆과 중앙에 다수의 구경꾼을 추가했다. / [gpt-high] 공터의 볏자루를 두 개로 고정한 조건과 달리 최소 세 개를 배치했다."
   }
  ],
  "refs": [
   {
    "label": "LOCATION PHOTOGRAPH — the exact place of this shot: its architecture, materials, fixed features and lighting mood are spatial truth; stage the moment inside this place. Never copy its camera framing.",
    "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/db00517b-165b-4005-a287-f28936b59b95/images/272bd7af-9294-4490-98c6-88a40e9deacc/scene/recipe/groupbg_lake_open_ground_bdb4de.png",
    "asset_id": "3e7f42e9-dcbe-4340-9422-44dd23901bc6",
    "role": "bgfirst_group_bg"
   },
   {
    "label": "CHARACTER REFERENCE — 안규홍: the exact person appearing in this shot; match face, hair and build exactly.",
    "path": "<bytes:1028384>",
    "asset_id": "3b5894db-ce55-4f34-b966-5e725c3aa952",
    "role": "character_ref"
   },
   {
    "label": "CHARACTER REFERENCE — 강용언: the exact person appearing in this shot; match face, hair and build exactly.",
    "path": "<bytes:822588>",
    "asset_id": "1a75090f-1eb9-4224-a132-de7a4b454b73",
    "role": "character_ref"
   }
  ],
  "critique_skipped": true,
  "needs_reshoot": true,
  "shot_run_produced": false,
  "shot_run_uid": "06aa1411-4f5c-767d-ab61-0590cadc7dc2",
  "bgfirst": {
   "bg_path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/db00517b-165b-4005-a287-f28936b59b95/images/272bd7af-9294-4490-98c6-88a40e9deacc/scene/recipe/S17sh8__bgfirst_bg.png",
   "bg_asset_id": "a456262d-cccd-4851-9155-e55c8779db35",
   "bg_record_key": "S17sh8::bgfirst_bg",
   "chain_winner": false,
   "authority": "groupbg",
   "group_key": "lake_open_ground",
   "groupbg_asset_id": "3e7f42e9-dcbe-4340-9422-44dd23901bc6"
  },
  "ref_mode": "그룹 배경+엔티티 (2택1: 무콘티 승)",
  "share_plan": {
   "ref_plan": "background"
  }
 },
 "S17sh8::cine": {
  "applied": true,
  "attempted_at": "2026-09-09T02:43:11.842659+00:00",
  "fingerprint": "d04d6f6dd365da70683b5c59fc6f779e72c18a63fcde00347c47d132dcbf32dd",
  "fingerprint_version": 2,
  "provider": "mai",
  "endpoint": "openrouter/chat-completions",
  "model": "microsoft/mai-image-2.6",
  "pack": "24.202608252115",
  "source_file": "S17sh8_sel.png",
  "source_sha256": "efb9df8232f73cca7bf35b8d7dc151ea02d411e8eb16dc1763140f64b4b4558d",
  "file": "S17sh8_cine.png",
  "staged_sha256": "29686de6d0d294ac63359e24df37c20786a73172c11bc93acaa36a167a2aac8b",
  "latency_ms": 28959
 },
 "S17sh18::signage": {
  "fp": "31eb30eafabc00b0",
  "inscriptions": [],
  "cues": [],
  "dropped": []
 },
 "S17sh18": {
  "input_fingerprint": "c7c0a03b3b92af9e",
  "prompt": "Create ONE FINAL photorealistic live-action film still of the moment below — early 20th-century Korea, Jeolla region; all people are Korean unless explicitly stated otherwise. TIME OF DAY (lock): late afternoon, low reddish sunlight.\n\nSHOT TEXT (authoritative, Korean): 허공을 향해 은비녀를 쥔 주먹을 번쩍 치켜든 안규홍의 당당한 상체.\n\nLOCATION (lock): The outdoor gathering circle in the lakeside camp's clearing, beside the men surrounding two seed-grain sacks. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: medium shot\n- KEY BACKGROUND ELEMENTS: 은비녀 (Recovered from the militia member and held aloft in 안규홍's fist) — A short exposed length projects obliquely from his grip; used as Makes the accusation of theft visible without enlarging the object; 호수 (Gray beyond the clearing); used as Forms a subdued background margin beside the raised arm.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: The same slanting reddish afternoon light is held to muted saturation and sober contrast, preserving resolve without turning the gesture into a heroic glow.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The clothing, hair and overall look of 안규홍 — who appear both in that photo and in this shot — are LOCKED to that photo. Anyone else visible in that photo is NOT in this shot: never carry their face, body or clothing onto anyone here. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The two tied seed-rice sacks remain in the red-lit clearing. 안규홍: He now holds up the recovered silver hairpin; his head wound has been wiped clean, and he retains his carrying frame and concealed jade ring.\n\nTHE HAND THAT IS DOING THIS: the object at the centre of this shot is being held, operated, read, aimed or handed over by 안규홍 right now, so 안규홍's hand — and as much of the wrist and forearm as the framing reaches — is in the frame, gripping or resting on that object exactly the way its real use requires. Match that hand to 안규홍: its size, build, skin, age, grooming, sleeve and anything worn on it belong to that person and to no one else.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 안규홍 (한국인 남성, 청년 나이대, 검은 머리, 단단히 묶은 머리) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.",
  "roll_prompts": {
   "A": "Create ONE FINAL photorealistic live-action film still of the moment below — early 20th-century Korea, Jeolla region; all people are Korean unless explicitly stated otherwise. TIME OF DAY (lock): late afternoon, low reddish sunlight.\n\nSHOT TEXT (authoritative, Korean): 허공을 향해 은비녀를 쥔 주먹을 번쩍 치켜든 안규홍의 당당한 상체.\n\nLOCATION (lock): The outdoor gathering circle in the lakeside camp's clearing, beside the men surrounding two seed-grain sacks. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: medium shot\n- KEY BACKGROUND ELEMENTS: 은비녀 (Recovered from the militia member and held aloft in 안규홍's fist) — A short exposed length projects obliquely from his grip; used as Makes the accusation of theft visible without enlarging the object; 호수 (Gray beyond the clearing); used as Forms a subdued background margin beside the raised arm.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: The same slanting reddish afternoon light is held to muted saturation and sober contrast, preserving resolve without turning the gesture into a heroic glow.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The clothing, hair and overall look of 안규홍 — who appear both in that photo and in this shot — are LOCKED to that photo. Anyone else visible in that photo is NOT in this shot: never carry their face, body or clothing onto anyone here. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The two tied seed-rice sacks remain in the red-lit clearing. 안규홍: He now holds up the recovered silver hairpin; his head wound has been wiped clean, and he retains his carrying frame and concealed jade ring.\n\nTHE HAND THAT IS DOING THIS: the object at the centre of this shot is being held, operated, read, aimed or handed over by 안규홍 right now, so 안규홍's hand — and as much of the wrist and forearm as the framing reaches — is in the frame, gripping or resting on that object exactly the way its real use requires. Match that hand to 안규홍: its size, build, skin, age, grooming, sleeve and anything worn on it belong to that person and to no one else.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 안규홍 (한국인 남성, 청년 나이대, 검은 머리, 단단히 묶은 머리) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.",
   "B": "Create ONE FINAL photorealistic live-action film still of the moment below — early 20th-century Korea, Jeolla region; all people are Korean unless explicitly stated otherwise. TIME OF DAY (lock): late afternoon, low reddish sunlight.\n\nSHOT TEXT (authoritative, Korean): 허공을 향해 은비녀를 쥔 주먹을 번쩍 치켜든 안규홍의 당당한 상체.\n\nLOCATION (lock): The outdoor gathering circle in the lakeside camp's clearing, beside the men surrounding two seed-grain sacks. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: medium shot\n- KEY BACKGROUND ELEMENTS: 은비녀 (Recovered from the militia member and held aloft in 안규홍's fist) — A short exposed length projects obliquely from his grip; used as Makes the accusation of theft visible without enlarging the object; 호수 (Gray beyond the clearing); used as Forms a subdued background margin beside the raised arm.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: The same slanting reddish afternoon light is held to muted saturation and sober contrast, preserving resolve without turning the gesture into a heroic glow.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The clothing, hair and overall look of 안규홍 — who appear both in that photo and in this shot — are LOCKED to that photo. Anyone else visible in that photo is NOT in this shot: never carry their face, body or clothing onto anyone here. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The two tied seed-rice sacks remain in the red-lit clearing. 안규홍: He now holds up the recovered silver hairpin; his head wound has been wiped clean, and he retains his carrying frame and concealed jade ring.\n\nTHE HAND THAT IS DOING THIS: the object at the centre of this shot is being held, operated, read, aimed or handed over by 안규홍 right now, so 안규홍's hand — and as much of the wrist and forearm as the framing reaches — is in the frame, gripping or resting on that object exactly the way its real use requires. Match that hand to 안규홍: its size, build, skin, age, grooming, sleeve and anything worn on it belong to that person and to no one else.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 안규홍 (한국인 남성, 청년 나이대, 검은 머리, 단단히 묶은 머리) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.\n\nCOMPOSITION VARIATION (this candidate only): keep the place, the people and the moment exactly as specified, and keep from the CAMERA & FRAME contract its subject scale, screen placement and key background elements exactly. For this candidate only, its camera angle and height are deliberately left open: choose a camera position distinctly different from the obvious one, as a film director picking a second setup on the same blocking. Stay true to any explicit framing the SHOT TEXT itself dictates."
  },
  "shot_run_spend_attempt_count": 1,
  "initial_roll_all_fail": true,
  "readings": [
   {
    "label": "A",
    "direction": "안규홍은 은비녀를 쥔 오른 주먹을 대각선 위로 들어 올리고 있으며, 시선은 정면을 향하고 있습니다.",
    "built_space": "호숫가 공터이며, 바닥에 곡식 자루 2개가 놓여 있고 뒤로는 호수가 보입니다. 전경 양옆으로 사람들의 뒷모습과 어깨가 보입니다.",
    "entities": "안규홍(레퍼런스와 일치, 지게 착용, 흰색 한복, 얼굴의 핏자국 없음), 은비녀(주먹에 쥐어진 짧은 형태). 프롬프트에 명시되지 않은 추가 인물들이 전경에 존재합니다.",
    "hard_violations": [
     "[gemini-pro] 프롬프트 샷 텍스트에 명시되지 않은 추가 인물(전경의 남녀)이 화면에 포함됨 (엄격한 인물 제한 및 추가 배제 규칙 위반)",
     "[gpt-high] 안규홍만 등장해야 하는데 화면 좌우 전경에 다른 인물 두 명의 머리와 어깨를 추가했다."
    ],
    "physics": "안규홍은 땅에 서 있고, 은비녀는 오른손 주먹에 단단히 쥐어져 있습니다. 곡식 자루는 땅에 안정적으로 놓여 있습니다."
   },
   {
    "label": "B",
    "direction": "안규홍은 긴 은비녀를 쥔 오른손을 위로 들고 있으며, 시선은 정면을 향합니다.",
    "built_space": "호숫가 공터이며, 곡식 자루 2개가 바닥에 있습니다. 배경과 전경에 여러 명의 남성들이 서 있습니다.",
    "entities": "안규홍(레퍼런스와 일치, 지게 착용), 은비녀(지시와 달리 너무 길고 꼬치처럼 묘사됨). 프롬프트에서 명시적으로 금지한 이전 샷의 인물들(파란 옷의 남성 등)이 배경에 그대로 등장합니다.",
    "hard_violations": [
     "[gemini-pro] 프롬프트 샷 텍스트에 명시되지 않은 추가 인물들이 포함됨",
     "[gemini-pro] 이전 샷 레퍼런스에 있던 인물(파란 옷을 입은 남성 등)을 절대 포함하지 말라는 명시적인 지시를 위반하고 배경에 그대로 묘사함",
     "[gpt-high] 안규홍만 등장해야 하는데 군복 인물과 뒤쪽 남성들, 전경의 다른 인물 및 신체 일부를 추가했다."
    ],
    "physics": "안규홍은 땅에 서 있고, 오른손으로 은비녀를 쥐고 있습니다. 곡식 자루는 바닥에 놓여 있습니다."
   }
  ],
  "cross_model_order": {
   "policy": "cross_model_order_v2_slot0_fwd_slot1_rev_parallel2",
   "slots": [
    {
     "model": "gemini-pro",
     "order": "forward",
     "display_to_canonical": {
      "A": "A",
      "B": "B"
     },
     "raw": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 4,
        "verdict_ko": "전경에 명시되지 않은 인물이 등장하는 치명적인 위반이 있으나, 주먹 밖으로 짧게 튀어나온 은비녀의 형태와 동작을 정확히 구현했으며 레퍼런스 인물 배제 지시를 상대적으로 잘 따랐습니다."
       },
       {
        "label": "B",
        "score": 1,
        "verdict_ko": "명시적으로 배제해야 할 이전 샷의 인물(파란 옷의 남성 등)을 배경에 그대로 포함하는 심각한 지시 위반을 저질렀으며, 은비녀도 지시와 달리 너무 길게 묘사되었습니다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "안규홍은 은비녀를 쥔 오른 주먹을 대각선 위로 들어 올리고 있으며, 시선은 정면을 향하고 있습니다.",
        "built_space": "호숫가 공터이며, 바닥에 곡식 자루 2개가 놓여 있고 뒤로는 호수가 보입니다. 전경 양옆으로 사람들의 뒷모습과 어깨가 보입니다.",
        "entities": "안규홍(레퍼런스와 일치, 지게 착용, 흰색 한복, 얼굴의 핏자국 없음), 은비녀(주먹에 쥐어진 짧은 형태). 프롬프트에 명시되지 않은 추가 인물들이 전경에 존재합니다.",
        "hard_violations": [
         "프롬프트 샷 텍스트에 명시되지 않은 추가 인물(전경의 남녀)이 화면에 포함됨 (엄격한 인물 제한 및 추가 배제 규칙 위반)"
        ],
        "physics": "안규홍은 땅에 서 있고, 은비녀는 오른손 주먹에 단단히 쥐어져 있습니다. 곡식 자루는 땅에 안정적으로 놓여 있습니다."
       },
       {
        "label": "B",
        "direction": "안규홍은 긴 은비녀를 쥔 오른손을 위로 들고 있으며, 시선은 정면을 향합니다.",
        "built_space": "호숫가 공터이며, 곡식 자루 2개가 바닥에 있습니다. 배경과 전경에 여러 명의 남성들이 서 있습니다.",
        "entities": "안규홍(레퍼런스와 일치, 지게 착용), 은비녀(지시와 달리 너무 길고 꼬치처럼 묘사됨). 프롬프트에서 명시적으로 금지한 이전 샷의 인물들(파란 옷의 남성 등)이 배경에 그대로 등장합니다.",
        "hard_violations": [
         "프롬프트 샷 텍스트에 명시되지 않은 추가 인물들이 포함됨",
         "이전 샷 레퍼런스에 있던 인물(파란 옷을 입은 남성 등)을 절대 포함하지 말라는 명시적인 지시를 위반하고 배경에 그대로 묘사함"
        ],
        "physics": "안규홍은 땅에 서 있고, 오른손으로 은비녀를 쥐고 있습니다. 곡식 자루는 바닥에 놓여 있습니다."
       }
      ],
      "all_candidates_fail": true
     },
     "normalized": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 4,
        "verdict_ko": "전경에 명시되지 않은 인물이 등장하는 치명적인 위반이 있으나, 주먹 밖으로 짧게 튀어나온 은비녀의 형태와 동작을 정확히 구현했으며 레퍼런스 인물 배제 지시를 상대적으로 잘 따랐습니다."
       },
       {
        "label": "B",
        "score": 1,
        "verdict_ko": "명시적으로 배제해야 할 이전 샷의 인물(파란 옷의 남성 등)을 배경에 그대로 포함하는 심각한 지시 위반을 저질렀으며, 은비녀도 지시와 달리 너무 길게 묘사되었습니다."
       }
      ],
      "all_candidates_fail": true,
      "readings": [
       {
        "label": "A",
        "direction": "안규홍은 은비녀를 쥔 오른 주먹을 대각선 위로 들어 올리고 있으며, 시선은 정면을 향하고 있습니다.",
        "built_space": "호숫가 공터이며, 바닥에 곡식 자루 2개가 놓여 있고 뒤로는 호수가 보입니다. 전경 양옆으로 사람들의 뒷모습과 어깨가 보입니다.",
        "entities": "안규홍(레퍼런스와 일치, 지게 착용, 흰색 한복, 얼굴의 핏자국 없음), 은비녀(주먹에 쥐어진 짧은 형태). 프롬프트에 명시되지 않은 추가 인물들이 전경에 존재합니다.",
        "hard_violations": [
         "프롬프트 샷 텍스트에 명시되지 않은 추가 인물(전경의 남녀)이 화면에 포함됨 (엄격한 인물 제한 및 추가 배제 규칙 위반)"
        ],
        "physics": "안규홍은 땅에 서 있고, 은비녀는 오른손 주먹에 단단히 쥐어져 있습니다. 곡식 자루는 땅에 안정적으로 놓여 있습니다."
       },
       {
        "label": "B",
        "direction": "안규홍은 긴 은비녀를 쥔 오른손을 위로 들고 있으며, 시선은 정면을 향합니다.",
        "built_space": "호숫가 공터이며, 곡식 자루 2개가 바닥에 있습니다. 배경과 전경에 여러 명의 남성들이 서 있습니다.",
        "entities": "안규홍(레퍼런스와 일치, 지게 착용), 은비녀(지시와 달리 너무 길고 꼬치처럼 묘사됨). 프롬프트에서 명시적으로 금지한 이전 샷의 인물들(파란 옷의 남성 등)이 배경에 그대로 등장합니다.",
        "hard_violations": [
         "프롬프트 샷 텍스트에 명시되지 않은 추가 인물들이 포함됨",
         "이전 샷 레퍼런스에 있던 인물(파란 옷을 입은 남성 등)을 절대 포함하지 말라는 명시적인 지시를 위반하고 배경에 그대로 묘사함"
        ],
        "physics": "안규홍은 땅에 서 있고, 오른손으로 은비녀를 쥐고 있습니다. 곡식 자루는 바닥에 놓여 있습니다."
       }
      ]
     },
     "ok": true
    },
    {
     "model": "gpt-high",
     "order": "reverse",
     "display_to_canonical": {
      "A": "B",
      "B": "A"
     },
     "raw": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 2,
        "verdict_ko": "상체 미디엄 숏과 치켜든 주먹은 맞지만, 금지된 주변 인물을 다수 등장시켰고 은비녀의 노출 길이도 지나치게 길다."
       },
       {
        "label": "B",
        "score": 4,
        "verdict_ko": "치켜든 주먹과 짧게 비스듬히 돌출된 은비녀는 더 충실하지만, 전경의 추가 인물 두 명 때문에 이 후보도 실격이다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "안규홍은 오른주먹을 머리 위 허공으로 치켜들고 화면 왼쪽 앞을 바라본다. 시선의 정확한 대상은 확정하기 어렵다. 은비녀의 긴 뾰족한 끝은 주먹에서 왼쪽 아래 빈 공간을 향하고, 장식 끝은 오른쪽 위로 나온다. 특정 사람을 겨누는 동작은 아니며 주먹을 허공으로 드는 지시는 맞는다. 주변 인물들은 대체로 안규홍 쪽을 향한다.",
        "built_space": "흙 공터 뒤로 호수와 먼 산이 보이고, 묶인 곡식 자루 두 개가 화면 왼쪽 아래 지면에 놓여 있다. 안규홍 등 뒤에는 세로 기둥 두 개와 가로 연결부가 있는 지게 하나가 보인다. 참조의 호숫가 공터와 재질은 대체로 이어지지만, 안규홍만 보여야 할 프레임에 왼쪽 가장자리의 군복 인물, 뒤쪽 남성 세 명, 오른쪽 전경 인물 및 왼쪽 아래 어깨 일부가 들어왔다. 반사나 고정 시설 중복 문제는 보이지 않는다.",
        "entities": "중앙 인물은 참조와 유사한 얼굴의 한국인 청년 남성으로 보이며 검은 머리를 뒤로 묶었다. 낡은 미색 저고리와 천 허리띠, 지게와 어깨끈을 착용했고 이마에 흐르는 피는 없다. 손에 든 금속 물체는 은색 비녀로 읽히지만 주먹 아래로 드러난 침 부분이 길어 짧은 노출이라는 지시와 다르다. 자루는 두 개이며 내용물이 볍씨인지는 외관만으로 확인할 수 없다. 숨긴 옥가락지는 보이지 않아 불일치로 판단하지 않는다. 안규홍 외 인물들은 명시적인 출연 제한을 어긴다. 읽을 수 있는 글자는 없다.",
        "hard_violations": [
         "안규홍만 등장해야 하는데 군복 인물과 뒤쪽 남성들, 전경의 다른 인물 및 신체 일부를 추가했다."
        ],
        "physics": "은비녀는 안규홍의 접힌 손가락과 엄지가 감싼 주먹에 지지되어 있다. 들어 올린 팔은 어깨와 팔꿈치가 자연스럽게 연결되어 실제로 가능한 자세다. 지게는 양쪽 어깨끈으로 몸에 매여 있고 자루 두 개는 지면에 닿아 있다. 발은 프레임 밖이지만 상체가 공중에 떠 있다는 징후는 없다. 보이는 주변 인물에도 명백한 부유나 불가능한 접촉은 없다."
       },
       {
        "label": "B",
        "direction": "안규홍은 오른주먹을 머리 위 허공으로 들고 화면 왼쪽 앞을 응시한다. 왼쪽 전경 인물 방향과 가깝지만 정확한 눈맞춤은 확인하기 어렵다. 은비녀의 뾰족한 끝은 왼쪽 아래로, 장식 끝은 오른쪽 위로 짧게 돌출되어 있다. 주먹은 허공을 향하며 비녀로 사람을 찌르거나 겨누는 자세가 아니다. 전경의 두 인물은 안규홍 쪽을 향한다.",
        "built_space": "호수 바로 앞 흙 공터에 안규홍이 서 있고, 화면 왼쪽 아래에 묶인 자루 두 개가 서로 기대어 있다. 등 뒤에는 세로 기둥 두 개와 가로 연결부를 갖춘 지게 하나가 있다. 호수와 흙의 재질은 참조 장소에 부합하지만 호수가 배경 대부분을 차지해 팔 옆의 절제된 배경 여백이라는 요구보다 두드러진다. 양쪽 전경에 다른 사람의 머리와 어깨를 배치해 단독 상체 숏에 불필요한 어깨너머 구도를 만들었다. 불가능한 반사나 중복 시설은 없다.",
        "entities": "안규홍은 참조와 유사한 한국인 청년 남성의 얼굴과 체격을 보이며 검은 머리를 뒤로 묶었다. 미색의 때 묻은 저고리, 천 허리띠, 지게와 어깨끈이 이어지고 이마의 피는 닦인 상태로 보인다. 은색 금속 비녀는 주먹 양옆으로 비교적 짧게 드러나 요구에 더 가깝다. 묶인 자루는 정확히 두 개이며 내용물은 확인할 수 없다. 옥가락지는 숨겨져 있어 보이지 않아도 문제없다. 양쪽 전경의 인물 두 명은 허용되지 않은 추가 인물이다. 읽을 수 있는 글자는 없다.",
        "hard_violations": [
         "안규홍만 등장해야 하는데 화면 좌우 전경에 다른 인물 두 명의 머리와 어깨를 추가했다."
        ],
        "physics": "비녀는 주먹에 실제로 쥐어져 있고 손목과 팔은 자연스럽게 연결된다. 팔을 위로 들어 유지하는 자세는 신체적으로 가능하다. 지게는 어깨끈으로 몸에 지지되며 두 자루는 지면과 서로의 표면에 기대어 있다. 하체가 잘렸다는 이유로 지지 결여를 판단할 수 없으며, 떠 있는 신체나 물체는 보이지 않는다."
       }
      ],
      "all_candidates_fail": true
     },
     "normalized": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 2,
        "verdict_ko": "상체 미디엄 숏과 치켜든 주먹은 맞지만, 금지된 주변 인물을 다수 등장시켰고 은비녀의 노출 길이도 지나치게 길다."
       },
       {
        "label": "A",
        "score": 4,
        "verdict_ko": "치켜든 주먹과 짧게 비스듬히 돌출된 은비녀는 더 충실하지만, 전경의 추가 인물 두 명 때문에 이 후보도 실격이다."
       }
      ],
      "all_candidates_fail": true,
      "readings": [
       {
        "label": "B",
        "direction": "안규홍은 오른주먹을 머리 위 허공으로 치켜들고 화면 왼쪽 앞을 바라본다. 시선의 정확한 대상은 확정하기 어렵다. 은비녀의 긴 뾰족한 끝은 주먹에서 왼쪽 아래 빈 공간을 향하고, 장식 끝은 오른쪽 위로 나온다. 특정 사람을 겨누는 동작은 아니며 주먹을 허공으로 드는 지시는 맞는다. 주변 인물들은 대체로 안규홍 쪽을 향한다.",
        "built_space": "흙 공터 뒤로 호수와 먼 산이 보이고, 묶인 곡식 자루 두 개가 화면 왼쪽 아래 지면에 놓여 있다. 안규홍 등 뒤에는 세로 기둥 두 개와 가로 연결부가 있는 지게 하나가 보인다. 참조의 호숫가 공터와 재질은 대체로 이어지지만, 안규홍만 보여야 할 프레임에 왼쪽 가장자리의 군복 인물, 뒤쪽 남성 세 명, 오른쪽 전경 인물 및 왼쪽 아래 어깨 일부가 들어왔다. 반사나 고정 시설 중복 문제는 보이지 않는다.",
        "entities": "중앙 인물은 참조와 유사한 얼굴의 한국인 청년 남성으로 보이며 검은 머리를 뒤로 묶었다. 낡은 미색 저고리와 천 허리띠, 지게와 어깨끈을 착용했고 이마에 흐르는 피는 없다. 손에 든 금속 물체는 은색 비녀로 읽히지만 주먹 아래로 드러난 침 부분이 길어 짧은 노출이라는 지시와 다르다. 자루는 두 개이며 내용물이 볍씨인지는 외관만으로 확인할 수 없다. 숨긴 옥가락지는 보이지 않아 불일치로 판단하지 않는다. 안규홍 외 인물들은 명시적인 출연 제한을 어긴다. 읽을 수 있는 글자는 없다.",
        "hard_violations": [
         "안규홍만 등장해야 하는데 군복 인물과 뒤쪽 남성들, 전경의 다른 인물 및 신체 일부를 추가했다."
        ],
        "physics": "은비녀는 안규홍의 접힌 손가락과 엄지가 감싼 주먹에 지지되어 있다. 들어 올린 팔은 어깨와 팔꿈치가 자연스럽게 연결되어 실제로 가능한 자세다. 지게는 양쪽 어깨끈으로 몸에 매여 있고 자루 두 개는 지면에 닿아 있다. 발은 프레임 밖이지만 상체가 공중에 떠 있다는 징후는 없다. 보이는 주변 인물에도 명백한 부유나 불가능한 접촉은 없다."
       },
       {
        "label": "A",
        "direction": "안규홍은 오른주먹을 머리 위 허공으로 들고 화면 왼쪽 앞을 응시한다. 왼쪽 전경 인물 방향과 가깝지만 정확한 눈맞춤은 확인하기 어렵다. 은비녀의 뾰족한 끝은 왼쪽 아래로, 장식 끝은 오른쪽 위로 짧게 돌출되어 있다. 주먹은 허공을 향하며 비녀로 사람을 찌르거나 겨누는 자세가 아니다. 전경의 두 인물은 안규홍 쪽을 향한다.",
        "built_space": "호수 바로 앞 흙 공터에 안규홍이 서 있고, 화면 왼쪽 아래에 묶인 자루 두 개가 서로 기대어 있다. 등 뒤에는 세로 기둥 두 개와 가로 연결부를 갖춘 지게 하나가 있다. 호수와 흙의 재질은 참조 장소에 부합하지만 호수가 배경 대부분을 차지해 팔 옆의 절제된 배경 여백이라는 요구보다 두드러진다. 양쪽 전경에 다른 사람의 머리와 어깨를 배치해 단독 상체 숏에 불필요한 어깨너머 구도를 만들었다. 불가능한 반사나 중복 시설은 없다.",
        "entities": "안규홍은 참조와 유사한 한국인 청년 남성의 얼굴과 체격을 보이며 검은 머리를 뒤로 묶었다. 미색의 때 묻은 저고리, 천 허리띠, 지게와 어깨끈이 이어지고 이마의 피는 닦인 상태로 보인다. 은색 금속 비녀는 주먹 양옆으로 비교적 짧게 드러나 요구에 더 가깝다. 묶인 자루는 정확히 두 개이며 내용물은 확인할 수 없다. 옥가락지는 숨겨져 있어 보이지 않아도 문제없다. 양쪽 전경의 인물 두 명은 허용되지 않은 추가 인물이다. 읽을 수 있는 글자는 없다.",
        "hard_violations": [
         "안규홍만 등장해야 하는데 화면 좌우 전경에 다른 인물 두 명의 머리와 어깨를 추가했다."
        ],
        "physics": "비녀는 주먹에 실제로 쥐어져 있고 손목과 팔은 자연스럽게 연결된다. 팔을 위로 들어 유지하는 자세는 신체적으로 가능하다. 지게는 어깨끈으로 몸에 지지되며 두 자루는 지면과 서로의 표면에 기대어 있다. 하체가 잘렸다는 이유로 지지 결여를 판단할 수 없으며, 떠 있는 신체나 물체는 보이지 않는다."
       }
      ]
     },
     "ok": true
    }
   ],
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "slot_winner_match": true,
   "slot_winner": {
    "gemini-pro": "A",
    "gpt-high": "A"
   },
   "route": "cross_slot_agree"
  },
  "dual": {
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "normalized": {
    "A": 2.0,
    "B": 0.75
   },
   "adjusted": {
    "A": 1.75,
    "B": 0.5
   },
   "violations": {
    "A": [
     "[gemini-pro] 프롬프트 샷 텍스트에 명시되지 않은 추가 인물(전경의 남녀)이 화면에 포함됨 (엄격한 인물 제한 및 추가 배제 규칙 위반)",
     "[gpt-high] 안규홍만 등장해야 하는데 화면 좌우 전경에 다른 인물 두 명의 머리와 어깨를 추가했다."
    ],
    "B": [
     "[gemini-pro] 프롬프트 샷 텍스트에 명시되지 않은 추가 인물들이 포함됨",
     "[gemini-pro] 이전 샷 레퍼런스에 있던 인물(파란 옷을 입은 남성 등)을 절대 포함하지 말라는 명시적인 지시를 위반하고 배경에 그대로 묘사함",
     "[gpt-high] 안규홍만 등장해야 하는데 군복 인물과 뒤쪽 남성들, 전경의 다른 인물 및 신체 일부를 추가했다."
    ]
   },
   "per_model_winner": {
    "gemini-pro": "A",
    "gpt-high": "A"
   },
   "agreed": true
  },
  "totals": {
   "A": 1750,
   "B": 500
  },
  "selected": "A",
  "ranking": [
   "A",
   "B"
  ],
  "verdicts": [
   {
    "label": "A",
    "score": 1750,
    "verdict_ko": "전경에 명시되지 않은 인물이 등장하는 치명적인 위반이 있으나, 주먹 밖으로 짧게 튀어나온 은비녀의 형태와 동작을 정확히 구현했으며 레퍼런스 인물 배제 지시를 상대적으로 잘 따랐습니다.  ★위반: [gemini-pro] 프롬프트 샷 텍스트에 명시되지 않은 추가 인물(전경의 남녀)이 화면에 포함됨 (엄격한 인물 제한 및 추가 배제 규칙 위반) / [gpt-high] 안규홍만 등장해야 하는데 화면 좌우 전경에 다른 인물 두 명의 머리와 어깨를 추가했다."
   },
   {
    "label": "B",
    "score": 500,
    "verdict_ko": "명시적으로 배제해야 할 이전 샷의 인물(파란 옷의 남성 등)을 배경에 그대로 포함하는 심각한 지시 위반을 저질렀으며, 은비녀도 지시와 달리 너무 길게 묘사되었습니다.  ★위반: [gemini-pro] 프롬프트 샷 텍스트에 명시되지 않은 추가 인물들이 포함됨 / [gemini-pro] 이전 샷 레퍼런스에 있던 인물(파란 옷을 입은 남성 등)을 절대 포함하지 말라는 명시적인 지시를 위반하고 배경에 그대로 묘사함 / [gpt-high] 안규홍만 등장해야 하는데 군복 인물과 뒤쪽 남성들, 전경의 다른 인물 및 신체 일부를 추가했다."
   }
  ],
  "refs": [
   {
    "label": "PREVIOUS SHOT STILL — a visually related earlier shot of this same place: the location's look, materials, fixed features and lighting mood are LOCKED to this photo; never copy its camera framing. The clothing, hair and overall look of 안규홍 — who appear both in that photo and in this shot — are LOCKED to that photo. Anyone else visible in that photo is NOT in this shot: never carry their face, body or clothing onto anyone here. If this photo shows a character who cannot move (dead or unconscious), that character's exact pose, position and orientation are ALSO LOCKED — treat their body as a fixed prop of the set that only the camera moves around.",
    "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/db00517b-165b-4005-a287-f28936b59b95/images/272bd7af-9294-4490-98c6-88a40e9deacc/scene/recipe/S17sh8_sel.png",
    "asset_id": "75c9150e-7e4a-4a53-8bec-2481d8b961e0",
    "role": "prev_still"
   },
   {
    "label": "CHARACTER REFERENCE — 안규홍: the exact person appearing in this shot; match face, hair and build exactly.",
    "path": "<bytes:1028384>",
    "asset_id": "3b5894db-ce55-4f34-b966-5e725c3aa952",
    "role": "character_ref"
   }
  ],
  "critique_skipped": true,
  "needs_reshoot": true,
  "shot_run_produced": false,
  "shot_run_uid": "06aa1411-5370-7df9-a04c-7f40ab0482ad",
  "ref_mode": "prev+엔티티",
  "share_plan": {
   "ref_plan": "prev",
   "prev_anchor_tag": "S17sh8"
  }
 },
 "S17sh18::cine": {
  "applied": true,
  "attempted_at": "2026-09-09T02:44:49.808854+00:00",
  "fingerprint": "72864b039a6340684ce25bf65e254e2423a9014c639f76f0e9189634f59a2ae2",
  "fingerprint_version": 2,
  "provider": "mai",
  "endpoint": "openrouter/chat-completions",
  "model": "microsoft/mai-image-2.6",
  "pack": "24.202608252115",
  "source_file": "S17sh18_sel.png",
  "source_sha256": "cbc8f02674cedb82403c3053c1d91fa7c601e343cc54275b2a34e925d80581d4",
  "file": "S17sh18_cine.png",
  "staged_sha256": "4971b548e39d8598e0fd0349d941576bf6e3a20ef9170cdc63439a843520c765",
  "latency_ms": 30758
 },
 "S17sh21::signage": {
  "fp": "2ea1717f97dad5d3",
  "inscriptions": [],
  "cues": [],
  "dropped": []
 },
 "S17sh21": {
  "input_fingerprint": "7e9bf871629f8299",
  "prompt": "Create ONE FINAL photorealistic live-action film still of the moment below — early 20th-century Korea, Jeolla region; all people are Korean unless explicitly stated otherwise. TIME OF DAY (lock): late afternoon, low reddish sunlight.\n\nSHOT TEXT (authoritative, Korean): 강용언의 발끝이 흙바닥의 씨나락 자루 표면에 강하게 맞닿아 자루가 움푹 눌려 있는 mid-impact 찰나.\n\nLOCATION (lock): Ground level in the lakeside camp's dirt clearing, at the two tied seed-grain sacks in the center of the gathering. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: close-up\n- FRAME LAYOUT: struck seed-grain sack with visible toe-contact indentation in the lower-center of the frame, foreground.\n- KEY BACKGROUND ELEMENTS: 발에 맞는 씨나락 자루 (Locally indented at the instant 강용언's toe makes contact) — The struck side is shown obliquely so the toe and depression remain separately readable; used as Provides the contact surface without overwhelming the lower-body frame; 두 번째 씨나락 자루 (Remains on the ground nearby) — Partially visible farther back and to the right of the struck sack; used as Preserves the established pair of sacks and supplies scale; 흙바닥 (Visible beneath the sacks and 강용언's supporting foot); used as Keeps the impact grounded in the clearing rather than isolating it against an abstract background.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Preserve the established slanting afternoon light and restrained reddish cast, using natural contrast to reveal the contact without a stylized impact flash.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The people visible in that photo are NOT in this shot: never carry their faces, bodies or clothing onto anyone here. Each person in THIS shot is defined solely by the PEOPLE list, their own CHARACTER REFERENCE images and their wardrobe notes. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): Both tied seed-rice sacks remain in the clearing; one is indented by the kick but has not opened. 강용언: He stands with one foot extended in a kick, still wearing his sword.\n\nTHE HAND THAT IS DOING THIS: the object at the centre of this shot is being held, operated, read, aimed or handed over by 강용언 right now, so 강용언's hand — and as much of the wrist and forearm as the framing reaches — is in the frame, gripping or resting on that object exactly the way its real use requires. Match that hand to 강용언: its size, build, skin, age, grooming, sleeve and anything worn on it belong to that person and to no one else.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 강용언 (한국인 남성, 성인 나이대, 햇볕에 그을린 얼굴, 자연스러운 검은 머리) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.",
  "roll_prompts": {
   "A": "Create ONE FINAL photorealistic live-action film still of the moment below — early 20th-century Korea, Jeolla region; all people are Korean unless explicitly stated otherwise. TIME OF DAY (lock): late afternoon, low reddish sunlight.\n\nSHOT TEXT (authoritative, Korean): 강용언의 발끝이 흙바닥의 씨나락 자루 표면에 강하게 맞닿아 자루가 움푹 눌려 있는 mid-impact 찰나.\n\nLOCATION (lock): Ground level in the lakeside camp's dirt clearing, at the two tied seed-grain sacks in the center of the gathering. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: close-up\n- FRAME LAYOUT: struck seed-grain sack with visible toe-contact indentation in the lower-center of the frame, foreground.\n- KEY BACKGROUND ELEMENTS: 발에 맞는 씨나락 자루 (Locally indented at the instant 강용언's toe makes contact) — The struck side is shown obliquely so the toe and depression remain separately readable; used as Provides the contact surface without overwhelming the lower-body frame; 두 번째 씨나락 자루 (Remains on the ground nearby) — Partially visible farther back and to the right of the struck sack; used as Preserves the established pair of sacks and supplies scale; 흙바닥 (Visible beneath the sacks and 강용언's supporting foot); used as Keeps the impact grounded in the clearing rather than isolating it against an abstract background.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Preserve the established slanting afternoon light and restrained reddish cast, using natural contrast to reveal the contact without a stylized impact flash.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The people visible in that photo are NOT in this shot: never carry their faces, bodies or clothing onto anyone here. Each person in THIS shot is defined solely by the PEOPLE list, their own CHARACTER REFERENCE images and their wardrobe notes. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): Both tied seed-rice sacks remain in the clearing; one is indented by the kick but has not opened. 강용언: He stands with one foot extended in a kick, still wearing his sword.\n\nTHE HAND THAT IS DOING THIS: the object at the centre of this shot is being held, operated, read, aimed or handed over by 강용언 right now, so 강용언's hand — and as much of the wrist and forearm as the framing reaches — is in the frame, gripping or resting on that object exactly the way its real use requires. Match that hand to 강용언: its size, build, skin, age, grooming, sleeve and anything worn on it belong to that person and to no one else.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 강용언 (한국인 남성, 성인 나이대, 햇볕에 그을린 얼굴, 자연스러운 검은 머리) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.",
   "B": "Create ONE FINAL photorealistic live-action film still of the moment below — early 20th-century Korea, Jeolla region; all people are Korean unless explicitly stated otherwise. TIME OF DAY (lock): late afternoon, low reddish sunlight.\n\nSHOT TEXT (authoritative, Korean): 강용언의 발끝이 흙바닥의 씨나락 자루 표면에 강하게 맞닿아 자루가 움푹 눌려 있는 mid-impact 찰나.\n\nLOCATION (lock): Ground level in the lakeside camp's dirt clearing, at the two tied seed-grain sacks in the center of the gathering. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: close-up\n- FRAME LAYOUT: struck seed-grain sack with visible toe-contact indentation in the lower-center of the frame, foreground.\n- KEY BACKGROUND ELEMENTS: 발에 맞는 씨나락 자루 (Locally indented at the instant 강용언's toe makes contact) — The struck side is shown obliquely so the toe and depression remain separately readable; used as Provides the contact surface without overwhelming the lower-body frame; 두 번째 씨나락 자루 (Remains on the ground nearby) — Partially visible farther back and to the right of the struck sack; used as Preserves the established pair of sacks and supplies scale; 흙바닥 (Visible beneath the sacks and 강용언's supporting foot); used as Keeps the impact grounded in the clearing rather than isolating it against an abstract background.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Preserve the established slanting afternoon light and restrained reddish cast, using natural contrast to reveal the contact without a stylized impact flash.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The people visible in that photo are NOT in this shot: never carry their faces, bodies or clothing onto anyone here. Each person in THIS shot is defined solely by the PEOPLE list, their own CHARACTER REFERENCE images and their wardrobe notes. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): Both tied seed-rice sacks remain in the clearing; one is indented by the kick but has not opened. 강용언: He stands with one foot extended in a kick, still wearing his sword.\n\nTHE HAND THAT IS DOING THIS: the object at the centre of this shot is being held, operated, read, aimed or handed over by 강용언 right now, so 강용언's hand — and as much of the wrist and forearm as the framing reaches — is in the frame, gripping or resting on that object exactly the way its real use requires. Match that hand to 강용언: its size, build, skin, age, grooming, sleeve and anything worn on it belong to that person and to no one else.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 강용언 (한국인 남성, 성인 나이대, 햇볕에 그을린 얼굴, 자연스러운 검은 머리) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.\n\nCOMPOSITION VARIATION (this candidate only): keep the place, the people and the moment exactly as specified, and keep from the CAMERA & FRAME contract its subject scale, screen placement and key background elements exactly. For this candidate only, its camera angle and height are deliberately left open: choose a camera position distinctly different from the obvious one, as a film director picking a second setup on the same blocking. Stay true to any explicit framing the SHOT TEXT itself dictates."
  },
  "shot_run_spend_attempt_count": 1,
  "readings": [
   {
    "label": "A",
    "direction": "맨발이 앞쪽 씨나락 자루의 윗부분을 아래로 밟고 있음.",
    "built_space": "호숫가 흙바닥 공터. 화면 뒤쪽으로 다수의 사람 다리가 보임.",
    "entities": "강용언의 하반신(파란 바지, 맨발), 타격받는 자루, 프롬프트에 없는 다수의 엑스트라들.",
    "hard_violations": [
     "[gemini-pro] 프롬프트에 명시되지 않은 다수의 인물(엑스트라)이 배경에 추가됨.",
     "[gpt-high] 강용언 외의 배경 인물들이 최소 다섯 명 추가되어, 다른 사람을 넣지 말라는 명시적 제한을 위반한다.",
     "[gpt-high] 씨나락 자루가 전경 하나와 배경 두 개로 총 세 개여서, 두 자루로 고정된 소품 수를 위반한다."
    ],
    "physics": "발이 자루 위에 올려져 체중으로 누르고 있으며, 타격보다는 밟고 서 있는 형태에 가까움."
   },
   {
    "label": "B",
    "direction": "짚신을 신은 발끝이 씨나락 자루의 전면을 강하게 차서 누르고 있음.",
    "built_space": "호숫가 흙바닥 공터. 지정된 대로 중앙 앞쪽에 타격받는 자루가 있고 우측 뒤편에 두 번째 자루가 놓여 있음.",
    "entities": "강용언의 하반신(흰색 바지, 짚신, 칼 착용), 앞쪽의 움푹 패인 자루, 뒤쪽의 온전한 자루. 바지 색상이 캐릭터 레퍼런스(파란색)와 일치하지 않음.",
    "hard_violations": [],
    "physics": "한쪽 다리로 몸을 지탱하고 뻗은 발끝이 자루 표면을 타격하여 자루가 움푹 들어가고 곡식 알갱이가 바닥에 떨어지는 등 타격의 물리력이 자연스러움."
   }
  ],
  "cross_model_order": {
   "policy": "cross_model_order_v2_slot0_fwd_slot1_rev_parallel2",
   "slots": [
    {
     "model": "gemini-pro",
     "order": "forward",
     "display_to_canonical": {
      "A": "A",
      "B": "B"
     },
     "raw": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 2,
        "verdict_ko": "지정되지 않은 다수의 인물을 배경에 추가하여 엄격한 인물 제한 지시를 위반한 치명적 오류가 있습니다."
       },
       {
        "label": "B",
        "score": 8,
        "verdict_ko": "의상 색상이 레퍼런스와 다르지만, 지시된 클로즈업 구도 내에서 발끝이 자루에 맞닿아 움푹 패인 타격의 순간을 매우 정확하고 사실적으로 구현했습니다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "맨발이 앞쪽 씨나락 자루의 윗부분을 아래로 밟고 있음.",
        "built_space": "호숫가 흙바닥 공터. 화면 뒤쪽으로 다수의 사람 다리가 보임.",
        "entities": "강용언의 하반신(파란 바지, 맨발), 타격받는 자루, 프롬프트에 없는 다수의 엑스트라들.",
        "hard_violations": [
         "프롬프트에 명시되지 않은 다수의 인물(엑스트라)이 배경에 추가됨."
        ],
        "physics": "발이 자루 위에 올려져 체중으로 누르고 있으며, 타격보다는 밟고 서 있는 형태에 가까움."
       },
       {
        "label": "B",
        "direction": "짚신을 신은 발끝이 씨나락 자루의 전면을 강하게 차서 누르고 있음.",
        "built_space": "호숫가 흙바닥 공터. 지정된 대로 중앙 앞쪽에 타격받는 자루가 있고 우측 뒤편에 두 번째 자루가 놓여 있음.",
        "entities": "강용언의 하반신(흰색 바지, 짚신, 칼 착용), 앞쪽의 움푹 패인 자루, 뒤쪽의 온전한 자루. 바지 색상이 캐릭터 레퍼런스(파란색)와 일치하지 않음.",
        "hard_violations": [],
        "physics": "한쪽 다리로 몸을 지탱하고 뻗은 발끝이 자루 표면을 타격하여 자루가 움푹 들어가고 곡식 알갱이가 바닥에 떨어지는 등 타격의 물리력이 자연스러움."
       }
      ],
      "all_candidates_fail": false
     },
     "normalized": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 2,
        "verdict_ko": "지정되지 않은 다수의 인물을 배경에 추가하여 엄격한 인물 제한 지시를 위반한 치명적 오류가 있습니다."
       },
       {
        "label": "B",
        "score": 8,
        "verdict_ko": "의상 색상이 레퍼런스와 다르지만, 지시된 클로즈업 구도 내에서 발끝이 자루에 맞닿아 움푹 패인 타격의 순간을 매우 정확하고 사실적으로 구현했습니다."
       }
      ],
      "all_candidates_fail": false,
      "readings": [
       {
        "label": "A",
        "direction": "맨발이 앞쪽 씨나락 자루의 윗부분을 아래로 밟고 있음.",
        "built_space": "호숫가 흙바닥 공터. 화면 뒤쪽으로 다수의 사람 다리가 보임.",
        "entities": "강용언의 하반신(파란 바지, 맨발), 타격받는 자루, 프롬프트에 없는 다수의 엑스트라들.",
        "hard_violations": [
         "프롬프트에 명시되지 않은 다수의 인물(엑스트라)이 배경에 추가됨."
        ],
        "physics": "발이 자루 위에 올려져 체중으로 누르고 있으며, 타격보다는 밟고 서 있는 형태에 가까움."
       },
       {
        "label": "B",
        "direction": "짚신을 신은 발끝이 씨나락 자루의 전면을 강하게 차서 누르고 있음.",
        "built_space": "호숫가 흙바닥 공터. 지정된 대로 중앙 앞쪽에 타격받는 자루가 있고 우측 뒤편에 두 번째 자루가 놓여 있음.",
        "entities": "강용언의 하반신(흰색 바지, 짚신, 칼 착용), 앞쪽의 움푹 패인 자루, 뒤쪽의 온전한 자루. 바지 색상이 캐릭터 레퍼런스(파란색)와 일치하지 않음.",
        "hard_violations": [],
        "physics": "한쪽 다리로 몸을 지탱하고 뻗은 발끝이 자루 표면을 타격하여 자루가 움푹 들어가고 곡식 알갱이가 바닥에 떨어지는 등 타격의 물리력이 자연스러움."
       }
      ]
     },
     "ok": true
    },
    {
     "model": "gpt-high",
     "order": "reverse",
     "display_to_canonical": {
      "A": "B",
      "B": "A"
     },
     "raw": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 7,
        "verdict_ko": "발끝이 자루 옆면을 누르는 접촉과 지지발, 뒤쪽 오른편의 두 번째 자루를 근접 구도로 구현했지만, 밝은색 바지는 강용언의 청회색 복식과 다르다."
       },
       {
        "label": "B",
        "score": 2,
        "verdict_ko": "금지된 배경 인물들과 세 번째 자루가 추가되었으며, 발가락이 위로 들려 있어 핵심인 발끝 충돌보다 발뒤꿈치로 누르는 모습에 가깝다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "왼쪽에서 뻗은 짚신의 앞끝이 화면 아래 중앙 자루의 비스듬한 옆면을 향하고 실제로 맞닿아 있다. 접촉점 주변 천이 안으로 눌려 발끝과 함몰부를 따로 읽을 수 있다. 얼굴과 시선은 화면 밖이며, 왼쪽 칼집은 아래로 늘어져 있다.",
        "built_space": "흙바닥과 상단의 좁은 호수 띠가 보이며 건축물이나 고정 설비는 없다. 주요 자루는 두 개로, 맞는 자루가 전경 중앙에 있고 다른 자루는 그 뒤 오른쪽에서 일부 가려진다. 지지발은 왼쪽 흙바닥에 놓여 있다.",
        "entities": "한 사람의 하반신만 보인다. 얼굴이 없어 한국인 성인 남성의 정확한 신원은 확인할 수 없다. 짚신과 착용한 칼집은 참조와 부합하지만, 밝은 회백색 바지와 발목 감개는 강용언의 청회색 바지 및 어두운 행전과 다르다. 두 자루는 거친 갈색 직물로 되어 있고 입구가 묶여 있으며 뚜렷한 파열은 없다. 낮은 따뜻한 햇빛과 흙의 질감은 장소 참조에 가깝고 읽을 수 있는 글자는 없다.",
        "hard_violations": [],
        "physics": "왼쪽 지지발이 땅에 닿아 체중을 받치고, 반대쪽 다리는 무릎을 굽힌 채 자루로 뻗어 있다. 앞 자루의 아래쪽은 흙에 닿고 뒤쪽은 두 번째 자루에 기대어 있어 기울어진 상태에도 지지가 있다. 짚신 앞끝에서 천이 눌리는 변형은 발차기 중간 순간으로 성립하며, 지지 없이 떠 있는 물체는 보이지 않는다."
       },
       {
        "label": "B",
        "direction": "맨발의 발바닥은 카메라를 향하고 발가락은 위로 들려 있다. 자루의 깊은 함몰은 발뒤꿈치 아래쪽에 있어, 발끝이 자루 표면을 강하게 치는 지정 순간으로 읽히지 않는다. 얼굴과 시선은 잘려 있으며 오른쪽 칼집은 아래로 향한다.",
        "built_space": "호숫가 흙 공터가 보인다. 전경에 맞는 자루 하나, 뒤 오른쪽에 낮게 누운 자루 하나와 세워진 자루 하나가 있어 총 세 개다. 배경에는 왼쪽과 중앙, 오른쪽에 여러 사람의 몸이 들어와 지정된 단독 하반신 장면을 벗어난다. 건축 설비나 반사는 없다.",
        "entities": "중앙 인물의 청회색 겉옷과 바지, 어두운 행전, 붉은 띠와 칼집은 강용언 참조에 가깝지만 차는 발은 참조의 짚신 없이 맨발이다. 얼굴은 보이지 않아 정확한 신원은 확인할 수 없다. 배경에는 허용되지 않은 인물들이 최소 다섯 명 보인다. 거친 직물 자루는 지정된 두 개보다 하나 많고, 전경 자루의 묶인 입구는 확인되지 않는다. 호수와 따뜻한 오후 햇빛은 보이며 읽을 수 있는 글자는 없다.",
        "hard_violations": [
         "강용언 외의 배경 인물들이 최소 다섯 명 추가되어, 다른 사람을 넣지 말라는 명시적 제한을 위반한다.",
         "씨나락 자루가 전경 하나와 배경 두 개로 총 세 개여서, 두 자루로 고정된 소품 수를 위반한다."
        ],
        "physics": "세 자루는 모두 흙바닥에 닿아 있다. 차는 다리는 몸에 연결되어 있고 반대 다리는 아래로 내려오지만, 지지발과 지면의 접촉은 전경 자루에 가려 확인할 수 없다. 자루가 발 아래에서 눌리는 현상 자체는 가능하나, 보이는 접촉 부위는 발끝보다 뒤꿈치 쪽이다. 떠오른 미세한 먼지는 충격으로 설명할 수 있으며, 몸이 지지 없이 공중에 떠 있다고 단정할 근거는 없다."
       }
      ],
      "all_candidates_fail": false
     },
     "normalized": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 7,
        "verdict_ko": "발끝이 자루 옆면을 누르는 접촉과 지지발, 뒤쪽 오른편의 두 번째 자루를 근접 구도로 구현했지만, 밝은색 바지는 강용언의 청회색 복식과 다르다."
       },
       {
        "label": "A",
        "score": 2,
        "verdict_ko": "금지된 배경 인물들과 세 번째 자루가 추가되었으며, 발가락이 위로 들려 있어 핵심인 발끝 충돌보다 발뒤꿈치로 누르는 모습에 가깝다."
       }
      ],
      "all_candidates_fail": false,
      "readings": [
       {
        "label": "B",
        "direction": "왼쪽에서 뻗은 짚신의 앞끝이 화면 아래 중앙 자루의 비스듬한 옆면을 향하고 실제로 맞닿아 있다. 접촉점 주변 천이 안으로 눌려 발끝과 함몰부를 따로 읽을 수 있다. 얼굴과 시선은 화면 밖이며, 왼쪽 칼집은 아래로 늘어져 있다.",
        "built_space": "흙바닥과 상단의 좁은 호수 띠가 보이며 건축물이나 고정 설비는 없다. 주요 자루는 두 개로, 맞는 자루가 전경 중앙에 있고 다른 자루는 그 뒤 오른쪽에서 일부 가려진다. 지지발은 왼쪽 흙바닥에 놓여 있다.",
        "entities": "한 사람의 하반신만 보인다. 얼굴이 없어 한국인 성인 남성의 정확한 신원은 확인할 수 없다. 짚신과 착용한 칼집은 참조와 부합하지만, 밝은 회백색 바지와 발목 감개는 강용언의 청회색 바지 및 어두운 행전과 다르다. 두 자루는 거친 갈색 직물로 되어 있고 입구가 묶여 있으며 뚜렷한 파열은 없다. 낮은 따뜻한 햇빛과 흙의 질감은 장소 참조에 가깝고 읽을 수 있는 글자는 없다.",
        "hard_violations": [],
        "physics": "왼쪽 지지발이 땅에 닿아 체중을 받치고, 반대쪽 다리는 무릎을 굽힌 채 자루로 뻗어 있다. 앞 자루의 아래쪽은 흙에 닿고 뒤쪽은 두 번째 자루에 기대어 있어 기울어진 상태에도 지지가 있다. 짚신 앞끝에서 천이 눌리는 변형은 발차기 중간 순간으로 성립하며, 지지 없이 떠 있는 물체는 보이지 않는다."
       },
       {
        "label": "A",
        "direction": "맨발의 발바닥은 카메라를 향하고 발가락은 위로 들려 있다. 자루의 깊은 함몰은 발뒤꿈치 아래쪽에 있어, 발끝이 자루 표면을 강하게 치는 지정 순간으로 읽히지 않는다. 얼굴과 시선은 잘려 있으며 오른쪽 칼집은 아래로 향한다.",
        "built_space": "호숫가 흙 공터가 보인다. 전경에 맞는 자루 하나, 뒤 오른쪽에 낮게 누운 자루 하나와 세워진 자루 하나가 있어 총 세 개다. 배경에는 왼쪽과 중앙, 오른쪽에 여러 사람의 몸이 들어와 지정된 단독 하반신 장면을 벗어난다. 건축 설비나 반사는 없다.",
        "entities": "중앙 인물의 청회색 겉옷과 바지, 어두운 행전, 붉은 띠와 칼집은 강용언 참조에 가깝지만 차는 발은 참조의 짚신 없이 맨발이다. 얼굴은 보이지 않아 정확한 신원은 확인할 수 없다. 배경에는 허용되지 않은 인물들이 최소 다섯 명 보인다. 거친 직물 자루는 지정된 두 개보다 하나 많고, 전경 자루의 묶인 입구는 확인되지 않는다. 호수와 따뜻한 오후 햇빛은 보이며 읽을 수 있는 글자는 없다.",
        "hard_violations": [
         "강용언 외의 배경 인물들이 최소 다섯 명 추가되어, 다른 사람을 넣지 말라는 명시적 제한을 위반한다.",
         "씨나락 자루가 전경 하나와 배경 두 개로 총 세 개여서, 두 자루로 고정된 소품 수를 위반한다."
        ],
        "physics": "세 자루는 모두 흙바닥에 닿아 있다. 차는 다리는 몸에 연결되어 있고 반대 다리는 아래로 내려오지만, 지지발과 지면의 접촉은 전경 자루에 가려 확인할 수 없다. 자루가 발 아래에서 눌리는 현상 자체는 가능하나, 보이는 접촉 부위는 발끝보다 뒤꿈치 쪽이다. 떠오른 미세한 먼지는 충격으로 설명할 수 있으며, 몸이 지지 없이 공중에 떠 있다고 단정할 근거는 없다."
       }
      ]
     },
     "ok": true
    }
   ],
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "slot_winner_match": true,
   "slot_winner": {
    "gemini-pro": "B",
    "gpt-high": "B"
   },
   "route": "cross_slot_agree"
  },
  "dual": {
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "normalized": {
    "A": 0.536,
    "B": 2.0
   },
   "adjusted": {
    "A": 0.286,
    "B": 2.0
   },
   "violations": {
    "A": [
     "[gemini-pro] 프롬프트에 명시되지 않은 다수의 인물(엑스트라)이 배경에 추가됨.",
     "[gpt-high] 강용언 외의 배경 인물들이 최소 다섯 명 추가되어, 다른 사람을 넣지 말라는 명시적 제한을 위반한다.",
     "[gpt-high] 씨나락 자루가 전경 하나와 배경 두 개로 총 세 개여서, 두 자루로 고정된 소품 수를 위반한다."
    ]
   },
   "per_model_winner": {
    "gemini-pro": "B",
    "gpt-high": "B"
   },
   "agreed": true
  },
  "totals": {
   "A": 286,
   "B": 2000
  },
  "selected": "B",
  "ranking": [
   "B",
   "A"
  ],
  "verdicts": [
   {
    "label": "A",
    "score": 286,
    "verdict_ko": "지정되지 않은 다수의 인물을 배경에 추가하여 엄격한 인물 제한 지시를 위반한 치명적 오류가 있습니다.  ★위반: [gemini-pro] 프롬프트에 명시되지 않은 다수의 인물(엑스트라)이 배경에 추가됨. / [gpt-high] 강용언 외의 배경 인물들이 최소 다섯 명 추가되어, 다른 사람을 넣지 말라는 명시적 제한을 위반한다. / [gpt-high] 씨나락 자루가 전경 하나와 배경 두 개로 총 세 개여서, 두 자루로 고정된 소품 수를 위반한다."
   },
   {
    "label": "B",
    "score": 2000,
    "verdict_ko": "의상 색상이 레퍼런스와 다르지만, 지시된 클로즈업 구도 내에서 발끝이 자루에 맞닿아 움푹 패인 타격의 순간을 매우 정확하고 사실적으로 구현했습니다."
   }
  ],
  "refs": [
   {
    "label": "PREVIOUS SHOT STILL — a visually related earlier shot of this same place: the location's look, materials, fixed features and lighting mood are LOCKED to this photo; never copy its camera framing. The people visible in that photo are NOT in this shot: never carry their faces, bodies or clothing onto anyone here. Each person in THIS shot is defined solely by the PEOPLE list, their own CHARACTER REFERENCE images and their wardrobe notes. If this photo shows a character who cannot move (dead or unconscious), that character's exact pose, position and orientation are ALSO LOCKED — treat their body as a fixed prop of the set that only the camera moves around.",
    "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/db00517b-165b-4005-a287-f28936b59b95/images/272bd7af-9294-4490-98c6-88a40e9deacc/scene/recipe/S17sh18_sel.png",
    "asset_id": "75dc390f-509c-4a28-a21f-384f7270e810",
    "role": "prev_still"
   },
   {
    "label": "CHARACTER REFERENCE — 강용언: the exact person appearing in this shot; match face, hair and build exactly.",
    "path": "<bytes:822588>",
    "asset_id": "1a75090f-1eb9-4224-a132-de7a4b454b73",
    "role": "character_ref"
   },
   {
    "label": "PROP REFERENCE — 곡식 자루·가마니: the exact object appearing in this shot; match its look, material and wear exactly.",
    "path": "<bytes:1807398>",
    "asset_id": "e30c319e-fdc7-4c01-a301-290ddd45d27a",
    "role": "prop_ref"
   }
  ],
  "critique_skipped": true,
  "shot_run_produced": false,
  "shot_run_uid": "06aa1411-54f0-7ac3-9cea-89ec03801d20",
  "ref_mode": "prev+엔티티",
  "share_plan": {
   "ref_plan": "prev",
   "prev_anchor_tag": "S17sh18"
  }
 },
 "S17sh21::cine": {
  "applied": true,
  "attempted_at": "2026-09-09T02:46:27.049820+00:00",
  "fingerprint": "6652f1da1354af4b27ff4ecbeee0cc312e1a57b17135c4e77f53a5b6a0f5c507",
  "fingerprint_version": 2,
  "provider": "mai",
  "endpoint": "openrouter/chat-completions",
  "model": "microsoft/mai-image-2.6",
  "pack": "24.202608252115",
  "source_file": "S17sh21_sel.png",
  "source_sha256": "4e632eec20913fd7dcc81f8e532defa264857b47c89b32d72db6a5576e2efa0d",
  "file": "S17sh21_cine.png",
  "staged_sha256": "9da160685d08238bb12e78ec15a6585f62b83fb719d6593fa5da70ba426f205a",
  "latency_ms": 27564
 },
 "S18sh1::signage": {
  "fp": "e93f8f3ce7b7b295",
  "inscriptions": [],
  "cues": [],
  "dropped": []
 },
 "S18sh1::bgfirst_bg": {
  "input_fingerprint": "be69fe219505c6a3",
  "prompt": "Create the EMPTY BACKGROUND PLATE for one film shot — NO PEOPLE, no figures, no body parts, no sketch lines, no arrows anywhere.\n\"Empty\" means no people only: KEEP the location's inherent occupants and stock that define the place — animals in an animal shelter, pen or farm, goods in a market, moored boats in a harbour — unless the shot text explicitly removes them.\nThe FIRST attached image is a thin-line storyboard sketch: use ONLY its camera angle, horizon, perspective and the placement/size of buildings and set masses — ignore the sketched people and arrows entirely. The SECOND attached image (LOCATION PHOTOGRAPH) is the real place: take its architecture, materials, signage and fixed features, and RE-PROJECT them into the sketch's camera. If the photograph's camera differs from the sketch's, the sketch's camera wins.\nHUMAN-SCALE CALIBRATION: derive every structure's true size from human-scale elements — a door ≈ 2m, a window ≈ 1–1.5m wide, one storey ≈ 2.5–3m; never inflate a small structure or shrink a large one.\n\nSHOT TEXT this background must serve (Korean): 빈 무쇠솥 주변 자갈밭에 축 늘어져 웅크린 조선인 의병들의 퀭한 전신.\n\nLOCATION (lock): The gravel eating area around an empty iron pot in the outdoor lakeside camp, with two untouched seed-grain sacks nearby.\n\nTIME OF DAY (lock): late afternoon, reddish natural light.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: wide shot\n- KEY BACKGROUND ELEMENTS: 빈 무쇠솥 (Empty, with exhausted militia members gathered around it) — The oblique elevated view exposes the empty interior; used as A modest central reference that makes the men's hunger legible without overpowering their bodies; 씨나락 자루 두 개 (Both tied and untouched beside the group) — Both sacks and their tied closures remain visible beside the men; used as A secondary visual counterpoint to the empty pot; 자갈밭 (Beneath the resting men and cooking area); used as Connects the separate resting poses within a continuous ground plane.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Late-afternoon ambient light preserves the scene's subdued warmth with muted saturation and naturalistic contrast, without anticipating the later cooking fire.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nRender ONE photorealistic empty location photograph, 16:9, that this shot can be staged inside later. No readable writing anywhere: surfaces that would carry writing may be present, but stage any wording out of legibility — an oblique angle, distance, shallow focus. No captions, watermarks or overlay text.",
  "effective_prompt": "Create the EMPTY BACKGROUND PLATE for one film shot — NO PEOPLE, no figures, no body parts, no sketch lines, no arrows anywhere.\n\"Empty\" means no people only: KEEP the location's inherent occupants and stock that define the place — animals in an animal shelter, pen or farm, goods in a market, moored boats in a harbour — unless the shot text explicitly removes them.\nThe FIRST attached image is a thin-line storyboard sketch: use ONLY its camera angle, horizon, perspective and the placement/size of buildings and set masses — ignore the sketched people and arrows entirely. The SECOND attached image (LOCATION PHOTOGRAPH) is the real place: take its architecture, materials, signage and fixed features, and RE-PROJECT them into the sketch's camera. If the photograph's camera differs from the sketch's, the sketch's camera wins.\nHUMAN-SCALE CALIBRATION: derive every structure's true size from human-scale elements — a door ≈ 2m, a window ≈ 1–1.5m wide, one storey ≈ 2.5–3m; never inflate a small structure or shrink a large one.\n\nSHOT TEXT this background must serve (Korean): 빈 무쇠솥 주변 자갈밭에 축 늘어져 웅크린 조선인 의병들의 퀭한 전신.\n\nLOCATION (lock): The gravel eating area around an empty iron pot in the outdoor lakeside camp, with two untouched seed-grain sacks nearby.\n\nTIME OF DAY (lock): late afternoon, reddish natural light.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: wide shot\n- KEY BACKGROUND ELEMENTS: 빈 무쇠솥 (Empty, with exhausted militia members gathered around it) — The oblique elevated view exposes the empty interior; used as A modest central reference that makes the men's hunger legible without overpowering their bodies; 씨나락 자루 두 개 (Both tied and untouched beside the group) — Both sacks and their tied closures remain visible beside the men; used as A secondary visual counterpoint to the empty pot; 자갈밭 (Beneath the resting men and cooking area); used as Connects the separate resting poses within a continuous ground plane.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Late-afternoon ambient light preserves the scene's subdued warmth with muted saturation and naturalistic contrast, without anticipating the later cooking fire.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nRender ONE photorealistic empty location photograph, 16:9, that this shot can be staged inside later. No readable writing anywhere: surfaces that would carry writing may be present, but stage any wording out of legibility — an oblique angle, distance, shallow focus. No captions, watermarks or overlay text.",
  "bg_path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/db00517b-165b-4005-a287-f28936b59b95/images/272bd7af-9294-4490-98c6-88a40e9deacc/scene/recipe/S18sh1__bgfirst_bg.png",
  "asset_id": "c96df81e-fc8e-4d6f-ba80-2177c04ce045",
  "input_asset_ids": [
   "0ead871c-d53e-42a7-a406-a8613400bae4",
   "f7b5ab04-6778-439a-b1ab-10ec63e47377"
  ]
 },
 "S18sh1": {
  "input_fingerprint": "ebfd930bfb68009c",
  "prompt": "Create ONE FINAL photorealistic live-action film still of the moment below — early 20th-century Korea, Jeolla region; all people are Korean unless explicitly stated otherwise. TIME OF DAY (lock): late afternoon, reddish natural light.\n\nSHOT TEXT (authoritative, Korean): 빈 무쇠솥 주변 자갈밭에 축 늘어져 웅크린 조선인 의병들의 퀭한 전신.\n\nLOCATION (lock): The gravel eating area around an empty iron pot in the outdoor lakeside camp, with two untouched seed-grain sacks nearby. The shot takes place here — the attached LOCATION PHOTOGRAPH shows the exact spot.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: wide shot\n- KEY BACKGROUND ELEMENTS: 빈 무쇠솥 (Empty, with exhausted militia members gathered around it) — The oblique elevated view exposes the empty interior; used as A modest central reference that makes the men's hunger legible without overpowering their bodies; 씨나락 자루 두 개 (Both tied and untouched beside the group) — Both sacks and their tied closures remain visible beside the men; used as A secondary visual counterpoint to the empty pot; 자갈밭 (Beneath the resting men and cooking area); used as Connects the separate resting poses within a continuous ground plane.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Late-afternoon ambient light preserves the scene's subdued warmth with muted saturation and naturalistic contrast, without anticipating the later cooking fire.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The iron pot is empty; two untouched, tightly tied seed-rice sacks remain beside it in the red evening light.\n\nNo people appear unless the shot text itself says so.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.",
  "roll_prompts": {
   "A": "Stage the shot. The FIRST attached image (SHOT BACKGROUND) is the finished empty background of this shot — keep it EXACTLY: its camera, perspective, architecture, lighting and every fixed feature stay untouched. The SECOND attached image (LAYOUT SKETCH) tells you ONLY where the people go: each sketched person's position, screen size, pose and the gaze/motion arrows. Ignore the sketch's background lines. The CHARACTER REFERENCE photographs show the real people.\nPlace the real people into the background at exactly the sketched positions, sizes and poses, following the arrow directions. No sketch lines or arrows may remain.\n\nCreate ONE FINAL photorealistic live-action film still of the moment below — early 20th-century Korea, Jeolla region; all people are Korean unless explicitly stated otherwise. TIME OF DAY (lock): late afternoon, reddish natural light.\n\nSHOT TEXT (authoritative, Korean): 빈 무쇠솥 주변 자갈밭에 축 늘어져 웅크린 조선인 의병들의 퀭한 전신.\n\nLOCATION (lock): The gravel eating area around an empty iron pot in the outdoor lakeside camp, with two untouched seed-grain sacks nearby. The shot takes place here — the FIRST attached image (SHOT BACKGROUND) is this exact place, already built: its ground, structures, horizon, materials and lighting are the finished truth of this location and must not be redesigned or replaced. No location photograph is attached — read the place from that image alone, and add no scenery, structure, vehicle or fixture that it does not already show. This lock governs the place only; the figures in the shot follow the staging and pose instructions.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: wide shot\n- KEY BACKGROUND ELEMENTS: 빈 무쇠솥 (Empty, with exhausted militia members gathered around it) — The oblique elevated view exposes the empty interior; used as A modest central reference that makes the men's hunger legible without overpowering their bodies; 씨나락 자루 두 개 (Both tied and untouched beside the group) — Both sacks and their tied closures remain visible beside the men; used as A secondary visual counterpoint to the empty pot; 자갈밭 (Beneath the resting men and cooking area); used as Connects the separate resting poses within a continuous ground plane.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Late-afternoon ambient light preserves the scene's subdued warmth with muted saturation and naturalistic contrast, without anticipating the later cooking fire.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The iron pot is empty; two untouched, tightly tied seed-rice sacks remain beside it in the red evening light.\n\nNo people appear unless the shot text itself says so.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.",
   "B": "Create ONE FINAL photorealistic live-action film still of the moment below — early 20th-century Korea, Jeolla region; all people are Korean unless explicitly stated otherwise. TIME OF DAY (lock): late afternoon, reddish natural light.\n\nSHOT TEXT (authoritative, Korean): 빈 무쇠솥 주변 자갈밭에 축 늘어져 웅크린 조선인 의병들의 퀭한 전신.\n\nLOCATION (lock): The gravel eating area around an empty iron pot in the outdoor lakeside camp, with two untouched seed-grain sacks nearby. The shot takes place here — the attached LOCATION PHOTOGRAPH shows the exact spot.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: wide shot\n- KEY BACKGROUND ELEMENTS: 빈 무쇠솥 (Empty, with exhausted militia members gathered around it) — The oblique elevated view exposes the empty interior; used as A modest central reference that makes the men's hunger legible without overpowering their bodies; 씨나락 자루 두 개 (Both tied and untouched beside the group) — Both sacks and their tied closures remain visible beside the men; used as A secondary visual counterpoint to the empty pot; 자갈밭 (Beneath the resting men and cooking area); used as Connects the separate resting poses within a continuous ground plane.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Late-afternoon ambient light preserves the scene's subdued warmth with muted saturation and naturalistic contrast, without anticipating the later cooking fire.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The iron pot is empty; two untouched, tightly tied seed-rice sacks remain beside it in the red evening light.\n\nNo people appear unless the shot text itself says so.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.\n\nCOMPOSITION VARIATION (this candidate only): keep the place, the people and the moment exactly as specified, and keep from the CAMERA & FRAME contract its subject scale, screen placement and key background elements exactly. For this candidate only, its camera angle and height are deliberately left open: choose a camera position distinctly different from the obvious one, as a film director picking a second setup on the same blocking. Stay true to any explicit framing the SHOT TEXT itself dictates."
  },
  "roll_refs": {
   "A": [
    {
     "label": "SHOT BACKGROUND",
     "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/db00517b-165b-4005-a287-f28936b59b95/images/272bd7af-9294-4490-98c6-88a40e9deacc/scene/recipe/S18sh1__bgfirst_bg.png",
     "asset_id": "c96df81e-fc8e-4d6f-ba80-2177c04ce045",
     "role": "bgfirst_bg"
    },
    {
     "label": "LAYOUT SKETCH (people placement only)",
     "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/db00517b-165b-4005-a287-f28936b59b95/images/272bd7af-9294-4490-98c6-88a40e9deacc/conti/conti_S18sh1.png",
     "asset_id": "0ead871c-d53e-42a7-a406-a8613400bae4",
     "role": "conti_light"
    },
    {
     "label": "PROP REFERENCE — 무쇠솥: the exact object appearing in this shot; match its look, material and wear exactly.",
     "path": "<bytes:1131910>",
     "asset_id": "8510d902-c12d-45d6-9f5c-8a90c9a902b3",
     "role": "prop_ref"
    }
   ],
   "B": [
    {
     "label": "LOCATION PHOTOGRAPH — the exact place of this shot: its architecture, materials, fixed features and lighting mood are spatial truth; stage the moment inside this place. Never copy its camera framing.",
     "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/db00517b-165b-4005-a287-f28936b59b95/images/272bd7af-9294-4490-98c6-88a40e9deacc/scene/recipe/groupbg_lake_cooking_circle_fb9493.png",
     "asset_id": "f7b5ab04-6778-439a-b1ab-10ec63e47377",
     "role": "bgfirst_group_bg"
    },
    {
     "label": "PROP REFERENCE — 무쇠솥: the exact object appearing in this shot; match its look, material and wear exactly.",
     "path": "<bytes:1131910>",
     "asset_id": "8510d902-c12d-45d6-9f5c-8a90c9a902b3",
     "role": "prop_ref"
    }
   ]
  },
  "shot_run_spend_attempt_count": 1,
  "judge_flip_skipped": "judge_fn_owns_order",
  "readings": [
   {
    "label": "A",
    "direction": "인물들은 고개를 숙이고 무릎에 얼굴을 묻거나 바닥을 향해 시선을 두고 있음.",
    "built_space": "호숫가 야영지의 자갈밭. 위치 레퍼런스에 있는 돌 화덕과 멍석이 묘사되지 않고 맨바닥으로 처리됨.",
    "entities": "지친 모습의 조선인 남성 3명, 뚜껑이 열려 내부가 비어 있는 무쇠솥, 솥 옆에 놓인 묶인 자루 2개.",
    "hard_violations": [],
    "physics": "세 명의 인물과 무쇠솥, 자루 모두 자갈밭 바닥에 무게를 싣고 안정적으로 놓여 있음."
   },
   {
    "label": "B",
    "direction": "인물들의 시선이 중앙에 타오르는 불과 무쇠솥을 향하고 있음.",
    "built_space": "호숫가 야영지의 자갈밭. 돌 화덕과 멍석 등 위치 레퍼런스의 공간 구조와 소품 배치를 충실히 반영함.",
    "entities": "조선인 남성 4명, 물이 끓고 있는 무쇠솥, 불이 피워진 화덕. (프롬프트가 요구한 인물 옆의 자루 2개는 생략됨).",
    "hard_violations": [],
    "physics": "인물들이 멍석 주변에 쭈그려 앉아 바닥에 무게를 지탱하고 있으며 구조적인 무리 없음."
   }
  ],
  "cross_model_order": {
   "policy": "cross_model_order_v2_slot0_fwd_slot1_rev_parallel2",
   "slots": [
    {
     "model": "gemini-pro",
     "order": "forward",
     "display_to_canonical": {
      "A": "A",
      "B": "B"
     },
     "raw": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 7,
        "verdict_ko": "위치 레퍼런스의 돌 화덕은 생략되었으나, 내부가 노출된 빈 무쇠솥, 그룹 옆에 놓인 자루 두 개, 불이 피워지지 않은 설정 등 프롬프트의 핵심 연출을 정확히 구현함."
       },
       {
        "label": "B",
        "score": 3,
        "verdict_ko": "위치 레퍼런스의 형태는 잘 유지했으나, '빈 솥'과 '요리용 불이 없는 상태'라는 명시적 지시를 무시하고 솥에 물을 채우고 불을 피웠으며 인물 옆의 자루도 누락하여 크게 감점됨."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "인물들은 고개를 숙이고 무릎에 얼굴을 묻거나 바닥을 향해 시선을 두고 있음.",
        "built_space": "호숫가 야영지의 자갈밭. 위치 레퍼런스에 있는 돌 화덕과 멍석이 묘사되지 않고 맨바닥으로 처리됨.",
        "entities": "지친 모습의 조선인 남성 3명, 뚜껑이 열려 내부가 비어 있는 무쇠솥, 솥 옆에 놓인 묶인 자루 2개.",
        "hard_violations": [],
        "physics": "세 명의 인물과 무쇠솥, 자루 모두 자갈밭 바닥에 무게를 싣고 안정적으로 놓여 있음."
       },
       {
        "label": "B",
        "direction": "인물들의 시선이 중앙에 타오르는 불과 무쇠솥을 향하고 있음.",
        "built_space": "호숫가 야영지의 자갈밭. 돌 화덕과 멍석 등 위치 레퍼런스의 공간 구조와 소품 배치를 충실히 반영함.",
        "entities": "조선인 남성 4명, 물이 끓고 있는 무쇠솥, 불이 피워진 화덕. (프롬프트가 요구한 인물 옆의 자루 2개는 생략됨).",
        "hard_violations": [],
        "physics": "인물들이 멍석 주변에 쭈그려 앉아 바닥에 무게를 지탱하고 있으며 구조적인 무리 없음."
       }
      ],
      "all_candidates_fail": false
     },
     "normalized": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 7,
        "verdict_ko": "위치 레퍼런스의 돌 화덕은 생략되었으나, 내부가 노출된 빈 무쇠솥, 그룹 옆에 놓인 자루 두 개, 불이 피워지지 않은 설정 등 프롬프트의 핵심 연출을 정확히 구현함."
       },
       {
        "label": "B",
        "score": 3,
        "verdict_ko": "위치 레퍼런스의 형태는 잘 유지했으나, '빈 솥'과 '요리용 불이 없는 상태'라는 명시적 지시를 무시하고 솥에 물을 채우고 불을 피웠으며 인물 옆의 자루도 누락하여 크게 감점됨."
       }
      ],
      "all_candidates_fail": false,
      "readings": [
       {
        "label": "A",
        "direction": "인물들은 고개를 숙이고 무릎에 얼굴을 묻거나 바닥을 향해 시선을 두고 있음.",
        "built_space": "호숫가 야영지의 자갈밭. 위치 레퍼런스에 있는 돌 화덕과 멍석이 묘사되지 않고 맨바닥으로 처리됨.",
        "entities": "지친 모습의 조선인 남성 3명, 뚜껑이 열려 내부가 비어 있는 무쇠솥, 솥 옆에 놓인 묶인 자루 2개.",
        "hard_violations": [],
        "physics": "세 명의 인물과 무쇠솥, 자루 모두 자갈밭 바닥에 무게를 싣고 안정적으로 놓여 있음."
       },
       {
        "label": "B",
        "direction": "인물들의 시선이 중앙에 타오르는 불과 무쇠솥을 향하고 있음.",
        "built_space": "호숫가 야영지의 자갈밭. 돌 화덕과 멍석 등 위치 레퍼런스의 공간 구조와 소품 배치를 충실히 반영함.",
        "entities": "조선인 남성 4명, 물이 끓고 있는 무쇠솥, 불이 피워진 화덕. (프롬프트가 요구한 인물 옆의 자루 2개는 생략됨).",
        "hard_violations": [],
        "physics": "인물들이 멍석 주변에 쭈그려 앉아 바닥에 무게를 지탱하고 있으며 구조적인 무리 없음."
       }
      ]
     },
     "ok": true
    },
    {
     "model": "gpt-high",
     "order": "reverse",
     "display_to_canonical": {
      "A": "B",
      "B": "A"
     },
     "raw": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 5,
        "verdict_ko": "전신과 빈 솥은 보이지만, 인물들이 돗자리에 밀집해 있고 솥 아래 불이 이미 타며 두 자루의 묶인 입구도 명확하지 않아 요구한 순간과 어긋납니다."
       },
       {
        "label": "B",
        "score": 8,
        "verdict_ko": "자갈 위에 축 늘어진 의병들의 전신, 중앙의 빈 솥, 입구가 묶인 두 자루를 정확히 배치했으나 솥의 형태와 화덕 주변 지형은 참조와 다릅니다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "네 남자 모두 고개와 시선을 무릎이나 앞쪽 바닥으로 내리고 있다. 가운데 남자는 자기 앞쪽을 멍하니 내려다본다. 카메라를 응시하는 사람은 없으며, 솥 입구는 위를 향해 높은 사선 시점에서 빈 내부가 보인다.",
        "built_space": "왼쪽 호수와 바위 해안, 뒤쪽 대나무숲, 오른쪽 천막과 발, 묶음 짐이 참조 장소와 대응한다. 돗자리 한 장 왼쪽에 네 사람이 밀집하고 오른쪽에는 큰 돌들에 걸친 솥 하나가 있다. 가까운 자루들은 화덕 돌 뒤에 일부 가려져 두 입구의 결박을 모두 확인하기 어렵다. 솥과 돌 화덕이 화면에서 상당한 비중을 차지한다.",
        "entities": "조선인 의병으로 설정된 성인 남성 네 명이 동아시아계 외관과 낡은 회백색 전통 복식, 상투 및 머리띠로 표현됐다. 얼굴은 지쳐 보이며 비정상적인 눈 표현은 없다. 빈 검은 무쇠솥의 둥근 몸통은 참조와 비슷하지만 참조 소품의 양쪽 고리 손잡이는 뚜렷하지 않다. 자루의 내용물은 확인할 수 없다. 읽을 수 있는 글자는 없다. 솥 아래 불꽃은 나중의 취사 불을 앞당기지 말라는 지시와 다르다.",
        "hard_violations": [],
        "physics": "앞쪽 남자는 발로 지면을 딛고 쪼그려 앉았고, 가운데 두 남자는 엉덩이와 접은 다리를 돗자리에 댄다. 오른쪽 남자도 접은 다리와 발로 지지된다. 솥은 큰 돌들 위에 걸쳐 있고 자루들은 지면에 놓여 있다. 지지 없이 떠 있는 신체나 물체는 보이지 않는다."
       },
       {
        "label": "B",
        "direction": "왼쪽 남자는 무릎 위 팔에 얼굴을 묻고, 뒤쪽 남자는 자기 무릎과 자갈을 내려다보며, 오른쪽 남자는 고개를 옆으로 떨어뜨린 채 앞쪽 바닥을 향한다. 모두 탈진한 정지 상태로 읽힌다. 솥 입구는 위를 향하고 사선 시점에서 빈 내부가 드러난다.",
        "built_space": "왼쪽 호수와 바위 해안, 뒤쪽 대나무숲, 오른쪽 천막과 발이 참조 장소의 배치를 유지한다. 세 남자는 연속된 자갈 지면 위에서 솥의 왼쪽 앞, 뒤, 오른쪽에 나뉘어 앉아 있다. 솥 하나가 중앙에 작게 놓이고 그 오른쪽 앞에는 입구가 각각 묶인 자루 두 개가 온전히 보인다. 다만 참조의 큰 돌 화덕과 돗자리는 이 취사 자리에서 보이지 않아 장소의 세부 배치는 달라졌다.",
        "entities": "조선인 의병으로 설정된 성인 남성 세 명이 동아시아계 외관과 낡은 회백색 전통 복식으로 표현됐다. 짧은 머리이며 지친 표정과 늘어진 자세를 보인다. 빈 검은 금속 솥 하나와 닫힌 천 자루 두 개가 명확하지만, 솥은 참조의 불룩한 몸통과 돌출 테두리 대신 단순한 원통형에 가깝다. 자루 속 씨나락 자체는 보이지 않지만 개봉 흔적은 없다. 불꽃이나 읽을 수 있는 글자는 없다.",
        "hard_violations": [],
        "physics": "세 남자 모두 엉덩이와 다리를 자갈 바닥에 대고 있다. 왼쪽 남자의 머리는 무릎에 얹은 팔이 받치며, 오른쪽 남자는 옆으로 짚은 손으로 상체를 지지한다. 솥과 두 자루는 지면에 안정적으로 놓여 있다. 지지 없는 부유나 불가능한 자세는 보이지 않는다."
       }
      ],
      "all_candidates_fail": false
     },
     "normalized": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 5,
        "verdict_ko": "전신과 빈 솥은 보이지만, 인물들이 돗자리에 밀집해 있고 솥 아래 불이 이미 타며 두 자루의 묶인 입구도 명확하지 않아 요구한 순간과 어긋납니다."
       },
       {
        "label": "A",
        "score": 8,
        "verdict_ko": "자갈 위에 축 늘어진 의병들의 전신, 중앙의 빈 솥, 입구가 묶인 두 자루를 정확히 배치했으나 솥의 형태와 화덕 주변 지형은 참조와 다릅니다."
       }
      ],
      "all_candidates_fail": false,
      "readings": [
       {
        "label": "B",
        "direction": "네 남자 모두 고개와 시선을 무릎이나 앞쪽 바닥으로 내리고 있다. 가운데 남자는 자기 앞쪽을 멍하니 내려다본다. 카메라를 응시하는 사람은 없으며, 솥 입구는 위를 향해 높은 사선 시점에서 빈 내부가 보인다.",
        "built_space": "왼쪽 호수와 바위 해안, 뒤쪽 대나무숲, 오른쪽 천막과 발, 묶음 짐이 참조 장소와 대응한다. 돗자리 한 장 왼쪽에 네 사람이 밀집하고 오른쪽에는 큰 돌들에 걸친 솥 하나가 있다. 가까운 자루들은 화덕 돌 뒤에 일부 가려져 두 입구의 결박을 모두 확인하기 어렵다. 솥과 돌 화덕이 화면에서 상당한 비중을 차지한다.",
        "entities": "조선인 의병으로 설정된 성인 남성 네 명이 동아시아계 외관과 낡은 회백색 전통 복식, 상투 및 머리띠로 표현됐다. 얼굴은 지쳐 보이며 비정상적인 눈 표현은 없다. 빈 검은 무쇠솥의 둥근 몸통은 참조와 비슷하지만 참조 소품의 양쪽 고리 손잡이는 뚜렷하지 않다. 자루의 내용물은 확인할 수 없다. 읽을 수 있는 글자는 없다. 솥 아래 불꽃은 나중의 취사 불을 앞당기지 말라는 지시와 다르다.",
        "hard_violations": [],
        "physics": "앞쪽 남자는 발로 지면을 딛고 쪼그려 앉았고, 가운데 두 남자는 엉덩이와 접은 다리를 돗자리에 댄다. 오른쪽 남자도 접은 다리와 발로 지지된다. 솥은 큰 돌들 위에 걸쳐 있고 자루들은 지면에 놓여 있다. 지지 없이 떠 있는 신체나 물체는 보이지 않는다."
       },
       {
        "label": "A",
        "direction": "왼쪽 남자는 무릎 위 팔에 얼굴을 묻고, 뒤쪽 남자는 자기 무릎과 자갈을 내려다보며, 오른쪽 남자는 고개를 옆으로 떨어뜨린 채 앞쪽 바닥을 향한다. 모두 탈진한 정지 상태로 읽힌다. 솥 입구는 위를 향하고 사선 시점에서 빈 내부가 드러난다.",
        "built_space": "왼쪽 호수와 바위 해안, 뒤쪽 대나무숲, 오른쪽 천막과 발이 참조 장소의 배치를 유지한다. 세 남자는 연속된 자갈 지면 위에서 솥의 왼쪽 앞, 뒤, 오른쪽에 나뉘어 앉아 있다. 솥 하나가 중앙에 작게 놓이고 그 오른쪽 앞에는 입구가 각각 묶인 자루 두 개가 온전히 보인다. 다만 참조의 큰 돌 화덕과 돗자리는 이 취사 자리에서 보이지 않아 장소의 세부 배치는 달라졌다.",
        "entities": "조선인 의병으로 설정된 성인 남성 세 명이 동아시아계 외관과 낡은 회백색 전통 복식으로 표현됐다. 짧은 머리이며 지친 표정과 늘어진 자세를 보인다. 빈 검은 금속 솥 하나와 닫힌 천 자루 두 개가 명확하지만, 솥은 참조의 불룩한 몸통과 돌출 테두리 대신 단순한 원통형에 가깝다. 자루 속 씨나락 자체는 보이지 않지만 개봉 흔적은 없다. 불꽃이나 읽을 수 있는 글자는 없다.",
        "hard_violations": [],
        "physics": "세 남자 모두 엉덩이와 다리를 자갈 바닥에 대고 있다. 왼쪽 남자의 머리는 무릎에 얹은 팔이 받치며, 오른쪽 남자는 옆으로 짚은 손으로 상체를 지지한다. 솥과 두 자루는 지면에 안정적으로 놓여 있다. 지지 없는 부유나 불가능한 자세는 보이지 않는다."
       }
      ]
     },
     "ok": true
    }
   ],
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "slot_winner_match": true,
   "slot_winner": {
    "gemini-pro": "A",
    "gpt-high": "A"
   },
   "route": "cross_slot_agree"
  },
  "dual": {
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "normalized": {
    "A": 2.0,
    "B": 1.054
   },
   "adjusted": {
    "A": 2.0,
    "B": 1.054
   },
   "violations": {},
   "per_model_winner": {
    "gemini-pro": "A",
    "gpt-high": "A"
   },
   "agreed": true
  },
  "totals": {
   "A": 2000,
   "B": 1054
  },
  "selected": "A",
  "ranking": [
   "A",
   "B"
  ],
  "verdicts": [
   {
    "label": "A",
    "score": 2000,
    "verdict_ko": "위치 레퍼런스의 돌 화덕은 생략되었으나, 내부가 노출된 빈 무쇠솥, 그룹 옆에 놓인 자루 두 개, 불이 피워지지 않은 설정 등 프롬프트의 핵심 연출을 정확히 구현함."
   },
   {
    "label": "B",
    "score": 1054,
    "verdict_ko": "위치 레퍼런스의 형태는 잘 유지했으나, '빈 솥'과 '요리용 불이 없는 상태'라는 명시적 지시를 무시하고 솥에 물을 채우고 불을 피웠으며 인물 옆의 자루도 누락하여 크게 감점됨."
   }
  ],
  "refs": [
   {
    "label": "LOCATION PHOTOGRAPH — the exact place of this shot: its architecture, materials, fixed features and lighting mood are spatial truth; stage the moment inside this place. Never copy its camera framing.",
    "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/db00517b-165b-4005-a287-f28936b59b95/images/272bd7af-9294-4490-98c6-88a40e9deacc/scene/recipe/groupbg_lake_cooking_circle_fb9493.png",
    "asset_id": "f7b5ab04-6778-439a-b1ab-10ec63e47377",
    "role": "bgfirst_group_bg"
   },
   {
    "label": "PROP REFERENCE — 무쇠솥: the exact object appearing in this shot; match its look, material and wear exactly.",
    "path": "<bytes:1131910>",
    "asset_id": "8510d902-c12d-45d6-9f5c-8a90c9a902b3",
    "role": "prop_ref"
   }
  ],
  "critique_skipped": true,
  "shot_run_produced": false,
  "shot_run_uid": "06aa1411-5692-7e0e-a923-0e79e386632b",
  "bgfirst": {
   "bg_path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/db00517b-165b-4005-a287-f28936b59b95/images/272bd7af-9294-4490-98c6-88a40e9deacc/scene/recipe/S18sh1__bgfirst_bg.png",
   "bg_asset_id": "c96df81e-fc8e-4d6f-ba80-2177c04ce045",
   "bg_record_key": "S18sh1::bgfirst_bg",
   "chain_winner": true,
   "authority": "groupbg",
   "group_key": "lake_cooking_circle",
   "groupbg_asset_id": "f7b5ab04-6778-439a-b1ab-10ec63e47377"
  },
  "ref_mode": "재투영 배경+콘티+엔티티 (2택1: 체인 승)",
  "share_plan": {
   "ref_plan": "background"
  }
 },
 "S18sh1::cine": {
  "applied": false,
  "attempted_at": "2026-09-09T08:57:26.746271+00:00",
  "fingerprint": "54518c45cbfe3f36040821d0af5648c68f2caccbc0d9cf01a4f8f5e8894fd9dd",
  "fingerprint_version": 2,
  "provider": "mai",
  "endpoint": "openrouter/chat-completions",
  "model": "microsoft/mai-image-2.6",
  "pack": "24.202608252115",
  "source_file": "S18sh1_sel.png",
  "source_sha256": "f1ec102f72664b861b034717d815e6ab87e0b11fb44fca2f21f19711d7d286be",
  "error": "RuntimeError: moderation blocked: {\"message\": \"Response content blocked by label 'MultiSeverity_ViolenceScore'.\", \"code\": 400}",
  "moderation_refusals": 2,
  "declined": true,
  "declined_reason": "moderation"
 },
 "S18sh4::signage": {
  "fp": "c93830bfb8ce3481",
  "inscriptions": [],
  "cues": [],
  "dropped": []
 },
 "groupbg::lake_bamboo_entry": {
  "input_fingerprint": "1634f2b24d2d9b16",
  "meta": {
   "model": "gpt-image-2",
   "size": "1536x864",
   "pack": "11.202607220237",
   "contract": "bgfirst_full_v3",
   "group_sig": {
    "key": "lake_bamboo_entry",
    "tags": [
     "S18sh4"
    ]
   },
   "context_sig": "c9e2ea936fd0bed2"
  },
  "prompt": "Create ONE empty live-action location background photograph — NO PEOPLE, no figures, no body parts, no silhouettes, no shadows or reflections of people anywhere.\n\"Empty\" means no people only: KEEP the location's inherent occupants and stock that define the place — animals in an animal shelter, pen or farm, goods and produce in a market, moored boats in a harbour — unless the shot text explicitly removes them.\nThe FIRST attached image is a thin-line storyboard sketch of a shot that happens at this location: use it ONLY as spatial evidence — what this place contains, how its ground, structures and landmarks are arranged and proportioned. Ignore the sketched people and arrows entirely, and do NOT copy its line style: render a fully photographic, physically plausible real place that fits THE LOCATION text below.\nHUMAN-SCALE CALIBRATION: derive every structure's true size from human-scale elements — a door ≈ 2m, a window ≈ 1–1.5m wide, one storey ≈ 2.5–3m; never inflate a small structure or shrink a large one.\n\nTHE LOCATION — early 20th-century Korea, Jeolla region; all people are Korean unless explicitly stated otherwise: The narrow bamboo approach opening into the temporary lakeside camp in reddish late-afternoon light.\n\nLOCATION DETAIL (descriptive evidence about this place from the production's location records — evidence, not a staging order): use it to understand what kind of place this is and what it permanently contains. Carry over ONLY its enduring physical features — layout, structures, ground and materials, fixed equipment and installations. Do NOT treat any lighting, weather or time-of-day wording, momentary object states, depicted events or subjective impressions in it as fixed facts of the place; the TIME OF DAY lock below is the sole authority for time and lighting.\n보성 호숫가 강용언 임시거점 자갈밭·공터·막사 앞: 잿빛 호숫가에 펼쳐진 자갈밭과 흙 공터. 대나무 숲을 등지고 있으며, 낡은 천과 거적을 엮어 만든 조잡한 임시 막사 여러 채, 큰 무쇠솥이 얹힌 돌 화덕이 배치되어 있다. (특징: 넓은 잿빛 호수면; 자갈과 흙이 섞인 공터; 낡은 천, 짚, 거적으로 거칠게 엮은 임시 천막/막사; 불이 타오르는 돌 화덕; 끓고 있는 큰 무쇠 가마솥; 쌓여 있는 짐 꾸러미와 곡식 자루)\n\nSCENE EVIDENCE (verbatim quotes from the screenplay about this place — treat them as evidence of what the location physically contains and looks like; stage the PLACE those moments happen in, but do NOT depict the momentary actions, people or staged props themselves):\n- 그때, 대나무 숲 사이로 지게를 진 규홍과 일행이 걸어 들어온다.\n- 보리쌀 자루와 감자 망태가 가마솥 앞 자갈밭에 묵직하게 내려앉는다.\n\nTIME OF DAY (lock): late afternoon, reddish natural light.\n\nRender ONE photorealistic empty location photograph, 16:9, neutral enough that every shot of this place can be staged from it later. No readable writing anywhere: surfaces that would carry writing may be present, but stage any wording out of legibility — an oblique angle, distance, shallow focus. No captions, watermarks or overlay text.",
  "effective_prompt": "Create ONE empty live-action location background photograph — NO PEOPLE, no figures, no body parts, no silhouettes, no shadows or reflections of people anywhere.\n\"Empty\" means no people only: KEEP the location's inherent occupants and stock that define the place — animals in an animal shelter, pen or farm, goods and produce in a market, moored boats in a harbour — unless the shot text explicitly removes them.\nThe FIRST attached image is a thin-line storyboard sketch of a shot that happens at this location: use it ONLY as spatial evidence — what this place contains, how its ground, structures and landmarks are arranged and proportioned. Ignore the sketched people and arrows entirely, and do NOT copy its line style: render a fully photographic, physically plausible real place that fits THE LOCATION text below.\nHUMAN-SCALE CALIBRATION: derive every structure's true size from human-scale elements — a door ≈ 2m, a window ≈ 1–1.5m wide, one storey ≈ 2.5–3m; never inflate a small structure or shrink a large one.\n\nTHE LOCATION — early 20th-century Korea, Jeolla region; all people are Korean unless explicitly stated otherwise: The narrow bamboo approach opening into the temporary lakeside camp in reddish late-afternoon light.\n\nLOCATION DETAIL (descriptive evidence about this place from the production's location records — evidence, not a staging order): use it to understand what kind of place this is and what it permanently contains. Carry over ONLY its enduring physical features — layout, structures, ground and materials, fixed equipment and installations. Do NOT treat any lighting, weather or time-of-day wording, momentary object states, depicted events or subjective impressions in it as fixed facts of the place; the TIME OF DAY lock below is the sole authority for time and lighting.\n보성 호숫가 강용언 임시거점 자갈밭·공터·막사 앞: 잿빛 호숫가에 펼쳐진 자갈밭과 흙 공터. 대나무 숲을 등지고 있으며, 낡은 천과 거적을 엮어 만든 조잡한 임시 막사 여러 채, 큰 무쇠솥이 얹힌 돌 화덕이 배치되어 있다. (특징: 넓은 잿빛 호수면; 자갈과 흙이 섞인 공터; 낡은 천, 짚, 거적으로 거칠게 엮은 임시 천막/막사; 불이 타오르는 돌 화덕; 끓고 있는 큰 무쇠 가마솥; 쌓여 있는 짐 꾸러미와 곡식 자루)\n\nSCENE EVIDENCE (verbatim quotes from the screenplay about this place — treat them as evidence of what the location physically contains and looks like; stage the PLACE those moments happen in, but do NOT depict the momentary actions, people or staged props themselves):\n- 그때, 대나무 숲 사이로 지게를 진 규홍과 일행이 걸어 들어온다.\n- 보리쌀 자루와 감자 망태가 가마솥 앞 자갈밭에 묵직하게 내려앉는다.\n\nTIME OF DAY (lock): late afternoon, reddish natural light.\n\nRender ONE photorealistic empty location photograph, 16:9, neutral enough that every shot of this place can be staged from it later. No readable writing anywhere: surfaces that would carry writing may be present, but stage any wording out of legibility — an oblique angle, distance, shallow focus. No captions, watermarks or overlay text.",
  "bg_path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/db00517b-165b-4005-a287-f28936b59b95/images/272bd7af-9294-4490-98c6-88a40e9deacc/scene/recipe/groupbg_lake_bamboo_entry_11c8f7.png",
  "asset_id": "cc3149c9-bdcd-4a60-b603-ee58e6473a40",
  "input_asset_ids": [
   "7f934b4a-3c6a-4635-bdbb-61ec165cbeb3"
  ],
  "origin_tag": "S18sh4",
  "place_text": "The narrow bamboo approach opening into the temporary lakeside camp in reddish late-afternoon light.",
  "origin_inputs": {
   "place_text": "The narrow bamboo approach opening into the temporary lakeside camp in reddish late-afternoon light.",
   "time_of_day_en": "late afternoon, reddish natural light",
   "conti_asset_id": "7f934b4a-3c6a-4635-bdbb-61ec165cbeb3"
  }
 },
 "S18sh4::bgfirst_bg": {
  "input_fingerprint": "c8df2809c8427ed0",
  "prompt": "Create the EMPTY BACKGROUND PLATE for one film shot — NO PEOPLE, no figures, no body parts, no sketch lines, no arrows anywhere.\n\"Empty\" means no people only: KEEP the location's inherent occupants and stock that define the place — animals in an animal shelter, pen or farm, goods in a market, moored boats in a harbour — unless the shot text explicitly removes them.\nThe FIRST attached image is a thin-line storyboard sketch: use ONLY its camera angle, horizon, perspective and the placement/size of buildings and set masses — ignore the sketched people and arrows entirely. The SECOND attached image (LOCATION PHOTOGRAPH) is the real place: take its architecture, materials, signage and fixed features, and RE-PROJECT them into the sketch's camera. If the photograph's camera differs from the sketch's, the sketch's camera wins.\nHUMAN-SCALE CALIBRATION: derive every structure's true size from human-scale elements — a door ≈ 2m, a window ≈ 1–1.5m wide, one storey ≈ 2.5–3m; never inflate a small structure or shrink a large one.\n\nSHOT TEXT this background must serve (Korean): 빽빽한 대나무 숲길 사이로 무거운 짐을 진 안규홍과 이관회가 한쪽 발을 앞으로 내디딘 미드액션 전신.\n\nLOCATION (lock): The narrow bamboo approach opening into the temporary lakeside camp in reddish late-afternoon light.\n\nTIME OF DAY (lock): late afternoon, reddish natural light.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: wide shot\n- KEY BACKGROUND ELEMENTS: 대나무 숲길 입구 (The men are emerging through the dense bamboo into the clearing) — The entrance recedes obliquely behind the men; used as Frames the arrival along the sides without concealing either man's feet or load; 지게와 운반 중인 식량 짐 (Carried toward the cooking area before unloading) — The carrying frames and loads are seen obliquely behind the men's shoulders; used as Makes the burden visible while retaining the carriers as the primary subjects.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Maintain the preceding late-afternoon ambient exposure and restrained saturation, allowing physical effort rather than a lighting change to carry the arrival.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nRender ONE photorealistic empty location photograph, 16:9, that this shot can be staged inside later. No readable writing anywhere: surfaces that would carry writing may be present, but stage any wording out of legibility — an oblique angle, distance, shallow focus. No captions, watermarks or overlay text.",
  "effective_prompt": "Create the EMPTY BACKGROUND PLATE for one film shot — NO PEOPLE, no figures, no body parts, no sketch lines, no arrows anywhere.\n\"Empty\" means no people only: KEEP the location's inherent occupants and stock that define the place — animals in an animal shelter, pen or farm, goods in a market, moored boats in a harbour — unless the shot text explicitly removes them.\nThe FIRST attached image is a thin-line storyboard sketch: use ONLY its camera angle, horizon, perspective and the placement/size of buildings and set masses — ignore the sketched people and arrows entirely. The SECOND attached image (LOCATION PHOTOGRAPH) is the real place: take its architecture, materials, signage and fixed features, and RE-PROJECT them into the sketch's camera. If the photograph's camera differs from the sketch's, the sketch's camera wins.\nHUMAN-SCALE CALIBRATION: derive every structure's true size from human-scale elements — a door ≈ 2m, a window ≈ 1–1.5m wide, one storey ≈ 2.5–3m; never inflate a small structure or shrink a large one.\n\nSHOT TEXT this background must serve (Korean): 빽빽한 대나무 숲길 사이로 무거운 짐을 진 안규홍과 이관회가 한쪽 발을 앞으로 내디딘 미드액션 전신.\n\nLOCATION (lock): The narrow bamboo approach opening into the temporary lakeside camp in reddish late-afternoon light.\n\nTIME OF DAY (lock): late afternoon, reddish natural light.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: wide shot\n- KEY BACKGROUND ELEMENTS: 대나무 숲길 입구 (The men are emerging through the dense bamboo into the clearing) — The entrance recedes obliquely behind the men; used as Frames the arrival along the sides without concealing either man's feet or load; 지게와 운반 중인 식량 짐 (Carried toward the cooking area before unloading) — The carrying frames and loads are seen obliquely behind the men's shoulders; used as Makes the burden visible while retaining the carriers as the primary subjects.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Maintain the preceding late-afternoon ambient exposure and restrained saturation, allowing physical effort rather than a lighting change to carry the arrival.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nRender ONE photorealistic empty location photograph, 16:9, that this shot can be staged inside later. No readable writing anywhere: surfaces that would carry writing may be present, but stage any wording out of legibility — an oblique angle, distance, shallow focus. No captions, watermarks or overlay text.",
  "bg_path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/db00517b-165b-4005-a287-f28936b59b95/images/272bd7af-9294-4490-98c6-88a40e9deacc/scene/recipe/S18sh4__bgfirst_bg.png",
  "asset_id": "a9dee6e0-0fb8-4884-975e-987652aa7464",
  "input_asset_ids": [
   "7f934b4a-3c6a-4635-bdbb-61ec165cbeb3",
   "cc3149c9-bdcd-4a60-b603-ee58e6473a40"
  ]
 },
 "S18sh4": {
  "input_fingerprint": "ae777aecdd3a6db9",
  "prompt": "Create ONE FINAL photorealistic live-action film still of the moment below — early 20th-century Korea, Jeolla region; all people are Korean unless explicitly stated otherwise. TIME OF DAY (lock): late afternoon, reddish natural light.\n\nSHOT TEXT (authoritative, Korean): 빽빽한 대나무 숲길 사이로 무거운 짐을 진 안규홍과 이관회가 한쪽 발을 앞으로 내디딘 미드액션 전신.\n\nLOCATION (lock): The narrow bamboo approach opening into the temporary lakeside camp in reddish late-afternoon light. The shot takes place here — the attached LOCATION PHOTOGRAPH shows the exact spot.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: wide shot\n- KEY BACKGROUND ELEMENTS: 대나무 숲길 입구 (The men are emerging through the dense bamboo into the clearing) — The entrance recedes obliquely behind the men; used as Frames the arrival along the sides without concealing either man's feet or load; 지게와 운반 중인 식량 짐 (Carried toward the cooking area before unloading) — The carrying frames and loads are seen obliquely behind the men's shoulders; used as Makes the burden visible while retaining the carriers as the primary subjects.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Maintain the preceding late-afternoon ambient exposure and restrained saturation, allowing physical effort rather than a lighting change to carry the arrival.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The empty pot and two unopened seed-rice sacks remain on the gravel in red late-afternoon light. 안규홍: He enters with a loaded carrying frame, dirt-covered hands and the jade ring concealed in his clothing. 이관회: He arrives sweaty and tired from building a wall and chopping wood.\n\nTHE HAND THAT IS DOING THIS: the object at the centre of this shot is being held, operated, read, aimed or handed over by 안규홍과 이관회 right now, so 안규홍과 이관회's hand — and as much of the wrist and forearm as the framing reaches — is in the frame, gripping or resting on that object exactly the way its real use requires. Match that hand to 안규홍과 이관회: its size, build, skin, age, grooming, sleeve and anything worn on it belong to that person and to no one else.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 안규홍 (한국인 남성, 청년 나이대, 검은 머리, 단단히 묶은 머리); 이관회 (한국인 남성, 청년 나이대, 자연스러운 검은 머리) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.",
  "roll_prompts": {
   "A": "Stage the shot. The FIRST attached image (SHOT BACKGROUND) is the finished empty background of this shot — keep it EXACTLY: its camera, perspective, architecture, lighting and every fixed feature stay untouched. The SECOND attached image (LAYOUT SKETCH) tells you ONLY where the people go: each sketched person's position, screen size, pose and the gaze/motion arrows. Ignore the sketch's background lines. The CHARACTER REFERENCE photographs show the real people.\nPlace the real people into the background at exactly the sketched positions, sizes and poses, following the arrow directions. No sketch lines or arrows may remain.\n\nCreate ONE FINAL photorealistic live-action film still of the moment below — early 20th-century Korea, Jeolla region; all people are Korean unless explicitly stated otherwise. TIME OF DAY (lock): late afternoon, reddish natural light.\n\nSHOT TEXT (authoritative, Korean): 빽빽한 대나무 숲길 사이로 무거운 짐을 진 안규홍과 이관회가 한쪽 발을 앞으로 내디딘 미드액션 전신.\n\nLOCATION (lock): The narrow bamboo approach opening into the temporary lakeside camp in reddish late-afternoon light. The shot takes place here — the FIRST attached image (SHOT BACKGROUND) is this exact place, already built: its ground, structures, horizon, materials and lighting are the finished truth of this location and must not be redesigned or replaced. No location photograph is attached — read the place from that image alone, and add no scenery, structure, vehicle or fixture that it does not already show. This lock governs the place only; the figures in the shot follow the staging and pose instructions.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: wide shot\n- KEY BACKGROUND ELEMENTS: 대나무 숲길 입구 (The men are emerging through the dense bamboo into the clearing) — The entrance recedes obliquely behind the men; used as Frames the arrival along the sides without concealing either man's feet or load; 지게와 운반 중인 식량 짐 (Carried toward the cooking area before unloading) — The carrying frames and loads are seen obliquely behind the men's shoulders; used as Makes the burden visible while retaining the carriers as the primary subjects.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Maintain the preceding late-afternoon ambient exposure and restrained saturation, allowing physical effort rather than a lighting change to carry the arrival.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The empty pot and two unopened seed-rice sacks remain on the gravel in red late-afternoon light. 안규홍: He enters with a loaded carrying frame, dirt-covered hands and the jade ring concealed in his clothing. 이관회: He arrives sweaty and tired from building a wall and chopping wood.\n\nTHE HAND THAT IS DOING THIS: the object at the centre of this shot is being held, operated, read, aimed or handed over by 안규홍과 이관회 right now, so 안규홍과 이관회's hand — and as much of the wrist and forearm as the framing reaches — is in the frame, gripping or resting on that object exactly the way its real use requires. Match that hand to 안규홍과 이관회: its size, build, skin, age, grooming, sleeve and anything worn on it belong to that person and to no one else.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 안규홍 (한국인 남성, 청년 나이대, 검은 머리, 단단히 묶은 머리); 이관회 (한국인 남성, 청년 나이대, 자연스러운 검은 머리) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.",
   "B": "Create ONE FINAL photorealistic live-action film still of the moment below — early 20th-century Korea, Jeolla region; all people are Korean unless explicitly stated otherwise. TIME OF DAY (lock): late afternoon, reddish natural light.\n\nSHOT TEXT (authoritative, Korean): 빽빽한 대나무 숲길 사이로 무거운 짐을 진 안규홍과 이관회가 한쪽 발을 앞으로 내디딘 미드액션 전신.\n\nLOCATION (lock): The narrow bamboo approach opening into the temporary lakeside camp in reddish late-afternoon light. The shot takes place here — the attached LOCATION PHOTOGRAPH shows the exact spot.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: wide shot\n- KEY BACKGROUND ELEMENTS: 대나무 숲길 입구 (The men are emerging through the dense bamboo into the clearing) — The entrance recedes obliquely behind the men; used as Frames the arrival along the sides without concealing either man's feet or load; 지게와 운반 중인 식량 짐 (Carried toward the cooking area before unloading) — The carrying frames and loads are seen obliquely behind the men's shoulders; used as Makes the burden visible while retaining the carriers as the primary subjects.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Maintain the preceding late-afternoon ambient exposure and restrained saturation, allowing physical effort rather than a lighting change to carry the arrival.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The empty pot and two unopened seed-rice sacks remain on the gravel in red late-afternoon light. 안규홍: He enters with a loaded carrying frame, dirt-covered hands and the jade ring concealed in his clothing. 이관회: He arrives sweaty and tired from building a wall and chopping wood.\n\nTHE HAND THAT IS DOING THIS: the object at the centre of this shot is being held, operated, read, aimed or handed over by 안규홍과 이관회 right now, so 안규홍과 이관회's hand — and as much of the wrist and forearm as the framing reaches — is in the frame, gripping or resting on that object exactly the way its real use requires. Match that hand to 안규홍과 이관회: its size, build, skin, age, grooming, sleeve and anything worn on it belong to that person and to no one else.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 안규홍 (한국인 남성, 청년 나이대, 검은 머리, 단단히 묶은 머리); 이관회 (한국인 남성, 청년 나이대, 자연스러운 검은 머리) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.\n\nCOMPOSITION VARIATION (this candidate only): keep the place, the people and the moment exactly as specified, and keep from the CAMERA & FRAME contract its subject scale, screen placement and key background elements exactly. For this candidate only, its camera angle and height are deliberately left open: choose a camera position distinctly different from the obvious one, as a film director picking a second setup on the same blocking. Stay true to any explicit framing the SHOT TEXT itself dictates."
  },
  "roll_refs": {
   "A": [
    {
     "label": "SHOT BACKGROUND",
     "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/db00517b-165b-4005-a287-f28936b59b95/images/272bd7af-9294-4490-98c6-88a40e9deacc/scene/recipe/S18sh4__bgfirst_bg.png",
     "asset_id": "a9dee6e0-0fb8-4884-975e-987652aa7464",
     "role": "bgfirst_bg"
    },
    {
     "label": "LAYOUT SKETCH (people placement only)",
     "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/db00517b-165b-4005-a287-f28936b59b95/images/272bd7af-9294-4490-98c6-88a40e9deacc/conti/conti_S18sh4.png",
     "asset_id": "7f934b4a-3c6a-4635-bdbb-61ec165cbeb3",
     "role": "conti_light"
    },
    {
     "label": "CHARACTER REFERENCE — 안규홍: the exact person appearing in this shot; match face, hair and build exactly.",
     "path": "<bytes:1028384>",
     "asset_id": "3b5894db-ce55-4f34-b966-5e725c3aa952",
     "role": "character_ref"
    },
    {
     "label": "CHARACTER REFERENCE — 이관회: the exact person appearing in this shot; match face, hair and build exactly.",
     "path": "<bytes:933540>",
     "asset_id": "851b450f-ac84-47fe-ac5a-03a56f63fcae",
     "role": "character_ref"
    },
    {
     "label": "PROP REFERENCE — 지게: the exact object appearing in this shot; match its look, material and wear exactly.",
     "path": "<bytes:1157225>",
     "asset_id": "bdfa122b-f62f-4ee2-807f-3e7997ed8125",
     "role": "prop_ref"
    },
    {
     "label": "PROP REFERENCE — 곡식 자루·가마니: the exact object appearing in this shot; match its look, material and wear exactly.",
     "path": "<bytes:1807398>",
     "asset_id": "e30c319e-fdc7-4c01-a301-290ddd45d27a",
     "role": "prop_ref"
    }
   ],
   "B": [
    {
     "label": "LOCATION PHOTOGRAPH — the exact place of this shot: its architecture, materials, fixed features and lighting mood are spatial truth; stage the moment inside this place. Never copy its camera framing.",
     "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/db00517b-165b-4005-a287-f28936b59b95/images/272bd7af-9294-4490-98c6-88a40e9deacc/scene/recipe/groupbg_lake_bamboo_entry_11c8f7.png",
     "asset_id": "cc3149c9-bdcd-4a60-b603-ee58e6473a40",
     "role": "bgfirst_group_bg"
    },
    {
     "label": "CHARACTER REFERENCE — 안규홍: the exact person appearing in this shot; match face, hair and build exactly.",
     "path": "<bytes:1028384>",
     "asset_id": "3b5894db-ce55-4f34-b966-5e725c3aa952",
     "role": "character_ref"
    },
    {
     "label": "CHARACTER REFERENCE — 이관회: the exact person appearing in this shot; match face, hair and build exactly.",
     "path": "<bytes:933540>",
     "asset_id": "851b450f-ac84-47fe-ac5a-03a56f63fcae",
     "role": "character_ref"
    },
    {
     "label": "PROP REFERENCE — 지게: the exact object appearing in this shot; match its look, material and wear exactly.",
     "path": "<bytes:1157225>",
     "asset_id": "bdfa122b-f62f-4ee2-807f-3e7997ed8125",
     "role": "prop_ref"
    },
    {
     "label": "PROP REFERENCE — 곡식 자루·가마니: the exact object appearing in this shot; match its look, material and wear exactly.",
     "path": "<bytes:1807398>",
     "asset_id": "e30c319e-fdc7-4c01-a301-290ddd45d27a",
     "role": "prop_ref"
    }
   ]
  },
  "shot_run_spend_attempt_count": 1,
  "judge_flip_skipped": "judge_fn_owns_order",
  "initial_roll_all_fail": true,
  "readings": [
   {
    "label": "A",
    "direction": "두 인물 모두 대나무 숲을 빠져나와 앞으로 걸어오고 있으며, 시선은 전방 및 약간 우측을 향함.",
    "built_space": "좌측의 대나무 숲길 입구와 우측의 호숫가 캠프(천막, 바닥의 자루 등)가 위치 레퍼런스와 일치하게 배치되어 있음.",
    "entities": "안규홍(묶은 머리, 밝은 옷)과 이관회(짧은 머리, 갈색 상의)의 인상착의는 레퍼런스와 부합하나, 등에 있는 지게의 형태가 무너져 있음.",
    "hard_violations": [
     "[gemini-pro] 두 인물 모두 등에 무거운 지게와 짐을 짊어지고 있으나, 이를 인물의 몸에 고정하는 어깨 끈(멜빵)이 전혀 묘사되지 않아 짐이 등에 아무런 지지 없이 공중에 떠 있음."
    ],
    "physics": "걷는 동작에 따른 발의 위치는 바닥에 닿아 있으나, 등에 짊어진 무거운 지게를 지탱하는 끈이나 손의 파지가 전혀 없어 지게가 허공에 떠 있는 상태임."
   },
   {
    "label": "B",
    "direction": "두 인물 모두 전방을 향해 걷고 있으며 시선 역시 앞으로 향함.",
    "built_space": "대나무 숲길에서 빠져나오는 동선과 우측 호숫가 임시 캠프의 구조물 배치가 레퍼런스와 잘 맞아떨어짐.",
    "entities": "안규홍과 이관회의 복장, 헤어스타일 등은 레퍼런스와 일치하며, 지게의 사다리 같은 구조도 레퍼런스에 가깝게 표현됨.",
    "hard_violations": [
     "[gemini-pro] 두 인물의 등에 지게가 위치해 있으나, 이를 어깨에 걸쳐 지탱하는 끈이 아예 그려지지 않아 지게와 짐이 공중에 떠 있음 (물리적 지지 기반 부재)."
    ],
    "physics": "인물의 걷는 자세와 바닥의 접지는 구현되었으나, 무거운 짐을 실은 지게를 지지하는 끈이 생략되어 물리적으로 불가능한 공중 부양 상태가 됨."
   }
  ],
  "cross_model_order": {
   "policy": "cross_model_order_v2_slot0_fwd_slot1_rev_parallel2",
   "slots": [
    {
     "model": "gemini-pro",
     "order": "forward",
     "display_to_canonical": {
      "A": "A",
      "B": "B"
     },
     "raw": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 1,
        "verdict_ko": "지게의 형태는 레퍼런스와 가깝게 구현되었으나, 지게를 지탱하는 어깨 끈이 전혀 없어 무거운 짐이 등에 지지 없이 떠 있는 심각한 물리적 오류가 발생했습니다."
       },
       {
        "label": "A",
        "score": 1,
        "verdict_ko": "지게를 지탱하는 어깨 끈이 없어 짐이 허공에 떠 있으며, 지게의 기본 형태조차 레퍼런스와 크게 다르게 왜곡되었습니다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "두 인물 모두 대나무 숲을 빠져나와 앞으로 걸어오고 있으며, 시선은 전방 및 약간 우측을 향함.",
        "built_space": "좌측의 대나무 숲길 입구와 우측의 호숫가 캠프(천막, 바닥의 자루 등)가 위치 레퍼런스와 일치하게 배치되어 있음.",
        "entities": "안규홍(묶은 머리, 밝은 옷)과 이관회(짧은 머리, 갈색 상의)의 인상착의는 레퍼런스와 부합하나, 등에 있는 지게의 형태가 무너져 있음.",
        "hard_violations": [
         "두 인물 모두 등에 무거운 지게와 짐을 짊어지고 있으나, 이를 인물의 몸에 고정하는 어깨 끈(멜빵)이 전혀 묘사되지 않아 짐이 등에 아무런 지지 없이 공중에 떠 있음."
        ],
        "physics": "걷는 동작에 따른 발의 위치는 바닥에 닿아 있으나, 등에 짊어진 무거운 지게를 지탱하는 끈이나 손의 파지가 전혀 없어 지게가 허공에 떠 있는 상태임."
       },
       {
        "label": "B",
        "direction": "두 인물 모두 전방을 향해 걷고 있으며 시선 역시 앞으로 향함.",
        "built_space": "대나무 숲길에서 빠져나오는 동선과 우측 호숫가 임시 캠프의 구조물 배치가 레퍼런스와 잘 맞아떨어짐.",
        "entities": "안규홍과 이관회의 복장, 헤어스타일 등은 레퍼런스와 일치하며, 지게의 사다리 같은 구조도 레퍼런스에 가깝게 표현됨.",
        "hard_violations": [
         "두 인물의 등에 지게가 위치해 있으나, 이를 어깨에 걸쳐 지탱하는 끈이 아예 그려지지 않아 지게와 짐이 공중에 떠 있음 (물리적 지지 기반 부재)."
        ],
        "physics": "인물의 걷는 자세와 바닥의 접지는 구현되었으나, 무거운 짐을 실은 지게를 지지하는 끈이 생략되어 물리적으로 불가능한 공중 부양 상태가 됨."
       }
      ],
      "all_candidates_fail": true
     },
     "normalized": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 1,
        "verdict_ko": "지게의 형태는 레퍼런스와 가깝게 구현되었으나, 지게를 지탱하는 어깨 끈이 전혀 없어 무거운 짐이 등에 지지 없이 떠 있는 심각한 물리적 오류가 발생했습니다."
       },
       {
        "label": "A",
        "score": 1,
        "verdict_ko": "지게를 지탱하는 어깨 끈이 없어 짐이 허공에 떠 있으며, 지게의 기본 형태조차 레퍼런스와 크게 다르게 왜곡되었습니다."
       }
      ],
      "all_candidates_fail": true,
      "readings": [
       {
        "label": "A",
        "direction": "두 인물 모두 대나무 숲을 빠져나와 앞으로 걸어오고 있으며, 시선은 전방 및 약간 우측을 향함.",
        "built_space": "좌측의 대나무 숲길 입구와 우측의 호숫가 캠프(천막, 바닥의 자루 등)가 위치 레퍼런스와 일치하게 배치되어 있음.",
        "entities": "안규홍(묶은 머리, 밝은 옷)과 이관회(짧은 머리, 갈색 상의)의 인상착의는 레퍼런스와 부합하나, 등에 있는 지게의 형태가 무너져 있음.",
        "hard_violations": [
         "두 인물 모두 등에 무거운 지게와 짐을 짊어지고 있으나, 이를 인물의 몸에 고정하는 어깨 끈(멜빵)이 전혀 묘사되지 않아 짐이 등에 아무런 지지 없이 공중에 떠 있음."
        ],
        "physics": "걷는 동작에 따른 발의 위치는 바닥에 닿아 있으나, 등에 짊어진 무거운 지게를 지탱하는 끈이나 손의 파지가 전혀 없어 지게가 허공에 떠 있는 상태임."
       },
       {
        "label": "B",
        "direction": "두 인물 모두 전방을 향해 걷고 있으며 시선 역시 앞으로 향함.",
        "built_space": "대나무 숲길에서 빠져나오는 동선과 우측 호숫가 임시 캠프의 구조물 배치가 레퍼런스와 잘 맞아떨어짐.",
        "entities": "안규홍과 이관회의 복장, 헤어스타일 등은 레퍼런스와 일치하며, 지게의 사다리 같은 구조도 레퍼런스에 가깝게 표현됨.",
        "hard_violations": [
         "두 인물의 등에 지게가 위치해 있으나, 이를 어깨에 걸쳐 지탱하는 끈이 아예 그려지지 않아 지게와 짐이 공중에 떠 있음 (물리적 지지 기반 부재)."
        ],
        "physics": "인물의 걷는 자세와 바닥의 접지는 구현되었으나, 무거운 짐을 실은 지게를 지지하는 끈이 생략되어 물리적으로 불가능한 공중 부양 상태가 됨."
       }
      ]
     },
     "ok": true
    },
    {
     "model": "gpt-high",
     "order": "reverse",
     "display_to_canonical": {
      "A": "B",
      "B": "A"
     },
     "raw": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 6,
        "verdict_ko": "전신 보행과 뒤로 드러난 지게는 맞지만, 숲길 너머 호수가 수목으로 바뀌고 두 사람 모두 짐을 잡지 않아 장소와 손 동작의 충실도가 떨어진다."
       },
       {
        "label": "B",
        "score": 8,
        "verdict_ko": "전신 미드액션, 대각선으로 물러나는 입구, 호숫가 야영지와 이관회의 피로 표현이 더 충실하지만, 두 사람의 손이 지게나 멜끈을 잡지 않는 조건 위반은 남는다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "두 사람은 화면 오른쪽 앞쪽으로 걸으며 시선도 진행 방향의 낮은 지면을 향한다. 뒤쪽 왼편 대나무 입구에서 야영지로 나온 방향은 맞으며, 카메라를 응시하지 않는다. 짐은 각각 어깨 뒤에 놓여 있다.",
        "built_space": "왼쪽에 대나무 통로 하나, 중앙에 천막 오두막 하나, 오른쪽에 연결된 취사 천막 구역이 보인다. 돌 화덕 위 큰 솥 하나와 오른쪽의 작은 뚜껑 있는 용기 하나, 전경의 큰 자루 세 개가 보인다. 두 사람은 시설 앞 자갈 공터에 있으며 발과 짐이 가려지지 않는다. 다만 참고에서 호수가 보이던 숲길 끝이 수목으로 바뀌고 호수와 산의 관계도 대부분 사라져 정확한 장소 일치가 약하다.",
        "entities": "한국인 청년 남성으로 설정된 두 사람만 보인다. 왼쪽의 묶은 검은 머리, 밝은 저고리와 발목을 묶은 바지, 오른쪽의 자연스러운 검은 머리, 갈색 저고리와 짧은 밝은 바지는 인물 참고와 대체로 맞는다. 나무 지게 두 개에 묶은 자루가 있으나 참고의 가마니 대신 장작 더미가 두드러진다. 안규홍의 손이 흙투성이인지는 뚜렷하지 않고 이관회의 땀도 약하다. 옥가락지는 노출되지 않는다. 바닥 자루 중 정확히 어느 두 개가 씨벼인지는 판별할 수 없고 솥 내부도 확인하기 어렵다. 읽을 수 있는 글자는 없다.",
        "hard_violations": [],
        "physics": "각자 앞발을 자갈에 딛고 뒤쪽 다리를 굽혀 옮기는 보행 순간으로, 몸이 공중에 뜬 모습은 아니다. 자루와 장작은 지게의 받침과 묶음에 지지된다. 지게는 등 뒤에 붙어 있으나 몸과 연결되는 멜끈은 명확하게 식별되지 않는다. 두 사람 모두 손을 아래로 늘어뜨려 지게나 끈을 잡지 않으므로 명시된 손 접촉 조건을 충족하지 않는다. 바닥의 자루와 취사 도구는 지면이나 화덕에 놓여 있다."
       },
       {
        "label": "B",
        "direction": "두 사람의 몸과 발은 화면 오른쪽 앞의 공터를 향하고, 시선도 그 진행 방향을 따른다. 특히 이관회는 고개와 상체를 앞으로 내밀어 도착하며 주변을 보는 모습이다. 대나무 입구는 인물 뒤 왼쪽으로 비스듬히 물러나고, 운반 짐은 어깨 뒤에서 사선으로 드러난다.",
        "built_space": "왼쪽 대나무 입구 하나, 중앙 천막 오두막 하나, 오른쪽 연결 취사 천막 구역이 보인다. 큰 솥 하나는 인물 뒤 돌 화덕에 부분적으로 가려져 있고, 오른쪽에는 작은 뚜껑 있는 용기 하나와 큰 자루 세 개가 보인다. 숲길 끝과 천막 뒤로 호수 및 산이 이어져 참고 장소의 공간 관계를 잘 유지한다. 두 사람 모두 자갈 공터에 서며 전신, 발, 적재물까지 프레임 안에 들어온다.",
        "entities": "두 명 모두 참고에 부합하는 청년 남성 외형이며, 안규홍의 묶은 검은 머리와 밝은 작업복, 이관회의 흐트러진 검은 머리와 갈색 저고리·짧은 밝은 바지가 구별된다. 이관회의 젖고 헝클어진 머리, 벌어진 입과 피곤한 표정은 노동 직후 상태에 더 가깝다. 목재와 밧줄로 된 지게 두 개에 묶인 곡식 자루가 보이지만 참고의 짚 가마니는 뚜렷하지 않다. 안규홍 지게 아래에는 둥근 그릇 같은 추가 적재물이 보인다. 손의 흙은 강하게 드러나지 않으며 옥가락지는 보이지 않는다. 씨벼 자루의 내용과 솥이 비었는지는 확인할 수 없다. 읽을 수 있는 글자는 없다.",
        "hard_violations": [],
        "physics": "두 사람 모두 앞으로 내디딘 발로 지면을 지지하고 반대쪽 무릎을 굽힌 보행 순간이다. 이관회의 앞으로 기운 상체가 짐의 무게에 대응한다. 지게의 목재 받침과 결박이 자루를 받치고, 안규홍의 어깨에는 멜끈이 분명히 보인다. 지면의 자루와 화덕의 솥도 지지면이 있다. 다만 두 사람 모두 손이 몸 옆에 내려와 있어 지게나 멜끈을 잡으라는 명시 조건에는 실패한다."
       }
      ],
      "all_candidates_fail": true
     },
     "normalized": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 6,
        "verdict_ko": "전신 보행과 뒤로 드러난 지게는 맞지만, 숲길 너머 호수가 수목으로 바뀌고 두 사람 모두 짐을 잡지 않아 장소와 손 동작의 충실도가 떨어진다."
       },
       {
        "label": "A",
        "score": 8,
        "verdict_ko": "전신 미드액션, 대각선으로 물러나는 입구, 호숫가 야영지와 이관회의 피로 표현이 더 충실하지만, 두 사람의 손이 지게나 멜끈을 잡지 않는 조건 위반은 남는다."
       }
      ],
      "all_candidates_fail": true,
      "readings": [
       {
        "label": "B",
        "direction": "두 사람은 화면 오른쪽 앞쪽으로 걸으며 시선도 진행 방향의 낮은 지면을 향한다. 뒤쪽 왼편 대나무 입구에서 야영지로 나온 방향은 맞으며, 카메라를 응시하지 않는다. 짐은 각각 어깨 뒤에 놓여 있다.",
        "built_space": "왼쪽에 대나무 통로 하나, 중앙에 천막 오두막 하나, 오른쪽에 연결된 취사 천막 구역이 보인다. 돌 화덕 위 큰 솥 하나와 오른쪽의 작은 뚜껑 있는 용기 하나, 전경의 큰 자루 세 개가 보인다. 두 사람은 시설 앞 자갈 공터에 있으며 발과 짐이 가려지지 않는다. 다만 참고에서 호수가 보이던 숲길 끝이 수목으로 바뀌고 호수와 산의 관계도 대부분 사라져 정확한 장소 일치가 약하다.",
        "entities": "한국인 청년 남성으로 설정된 두 사람만 보인다. 왼쪽의 묶은 검은 머리, 밝은 저고리와 발목을 묶은 바지, 오른쪽의 자연스러운 검은 머리, 갈색 저고리와 짧은 밝은 바지는 인물 참고와 대체로 맞는다. 나무 지게 두 개에 묶은 자루가 있으나 참고의 가마니 대신 장작 더미가 두드러진다. 안규홍의 손이 흙투성이인지는 뚜렷하지 않고 이관회의 땀도 약하다. 옥가락지는 노출되지 않는다. 바닥 자루 중 정확히 어느 두 개가 씨벼인지는 판별할 수 없고 솥 내부도 확인하기 어렵다. 읽을 수 있는 글자는 없다.",
        "hard_violations": [],
        "physics": "각자 앞발을 자갈에 딛고 뒤쪽 다리를 굽혀 옮기는 보행 순간으로, 몸이 공중에 뜬 모습은 아니다. 자루와 장작은 지게의 받침과 묶음에 지지된다. 지게는 등 뒤에 붙어 있으나 몸과 연결되는 멜끈은 명확하게 식별되지 않는다. 두 사람 모두 손을 아래로 늘어뜨려 지게나 끈을 잡지 않으므로 명시된 손 접촉 조건을 충족하지 않는다. 바닥의 자루와 취사 도구는 지면이나 화덕에 놓여 있다."
       },
       {
        "label": "A",
        "direction": "두 사람의 몸과 발은 화면 오른쪽 앞의 공터를 향하고, 시선도 그 진행 방향을 따른다. 특히 이관회는 고개와 상체를 앞으로 내밀어 도착하며 주변을 보는 모습이다. 대나무 입구는 인물 뒤 왼쪽으로 비스듬히 물러나고, 운반 짐은 어깨 뒤에서 사선으로 드러난다.",
        "built_space": "왼쪽 대나무 입구 하나, 중앙 천막 오두막 하나, 오른쪽 연결 취사 천막 구역이 보인다. 큰 솥 하나는 인물 뒤 돌 화덕에 부분적으로 가려져 있고, 오른쪽에는 작은 뚜껑 있는 용기 하나와 큰 자루 세 개가 보인다. 숲길 끝과 천막 뒤로 호수 및 산이 이어져 참고 장소의 공간 관계를 잘 유지한다. 두 사람 모두 자갈 공터에 서며 전신, 발, 적재물까지 프레임 안에 들어온다.",
        "entities": "두 명 모두 참고에 부합하는 청년 남성 외형이며, 안규홍의 묶은 검은 머리와 밝은 작업복, 이관회의 흐트러진 검은 머리와 갈색 저고리·짧은 밝은 바지가 구별된다. 이관회의 젖고 헝클어진 머리, 벌어진 입과 피곤한 표정은 노동 직후 상태에 더 가깝다. 목재와 밧줄로 된 지게 두 개에 묶인 곡식 자루가 보이지만 참고의 짚 가마니는 뚜렷하지 않다. 안규홍 지게 아래에는 둥근 그릇 같은 추가 적재물이 보인다. 손의 흙은 강하게 드러나지 않으며 옥가락지는 보이지 않는다. 씨벼 자루의 내용과 솥이 비었는지는 확인할 수 없다. 읽을 수 있는 글자는 없다.",
        "hard_violations": [],
        "physics": "두 사람 모두 앞으로 내디딘 발로 지면을 지지하고 반대쪽 무릎을 굽힌 보행 순간이다. 이관회의 앞으로 기운 상체가 짐의 무게에 대응한다. 지게의 목재 받침과 결박이 자루를 받치고, 안규홍의 어깨에는 멜끈이 분명히 보인다. 지면의 자루와 화덕의 솥도 지지면이 있다. 다만 두 사람 모두 손이 몸 옆에 내려와 있어 지게나 멜끈을 잡으라는 명시 조건에는 실패한다."
       }
      ]
     },
     "ok": true
    }
   ],
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "slot_winner_match": false,
   "slot_winner": {
    "gemini-pro": "B",
    "gpt-high": "A"
   },
   "route": "cross_slot_combined"
  },
  "dual": {
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "normalized": {
    "A": 2.0,
    "B": 1.75
   },
   "adjusted": {
    "A": 1.75,
    "B": 1.5
   },
   "violations": {
    "A": [
     "[gemini-pro] 두 인물 모두 등에 무거운 지게와 짐을 짊어지고 있으나, 이를 인물의 몸에 고정하는 어깨 끈(멜빵)이 전혀 묘사되지 않아 짐이 등에 아무런 지지 없이 공중에 떠 있음."
    ],
    "B": [
     "[gemini-pro] 두 인물의 등에 지게가 위치해 있으나, 이를 어깨에 걸쳐 지탱하는 끈이 아예 그려지지 않아 지게와 짐이 공중에 떠 있음 (물리적 지지 기반 부재)."
    ]
   },
   "per_model_winner": {
    "gemini-pro": "B",
    "gpt-high": "A"
   },
   "agreed": false
  },
  "totals": {
   "B": 1500,
   "A": 1750
  },
  "selected": "A",
  "ranking": [
   "A",
   "B"
  ],
  "verdicts": [
   {
    "label": "B",
    "score": 1500,
    "verdict_ko": "지게의 형태는 레퍼런스와 가깝게 구현되었으나, 지게를 지탱하는 어깨 끈이 전혀 없어 무거운 짐이 등에 지지 없이 떠 있는 심각한 물리적 오류가 발생했습니다.  ★위반: [gemini-pro] 두 인물의 등에 지게가 위치해 있으나, 이를 어깨에 걸쳐 지탱하는 끈이 아예 그려지지 않아 지게와 짐이 공중에 떠 있음 (물리적 지지 기반 부재)."
   },
   {
    "label": "A",
    "score": 1750,
    "verdict_ko": "지게를 지탱하는 어깨 끈이 없어 짐이 허공에 떠 있으며, 지게의 기본 형태조차 레퍼런스와 크게 다르게 왜곡되었습니다.  ★위반: [gemini-pro] 두 인물 모두 등에 무거운 지게와 짐을 짊어지고 있으나, 이를 인물의 몸에 고정하는 어깨 끈(멜빵)이 전혀 묘사되지 않아 짐이 등에 아무런 지지 없이 공중에 떠 있음."
   }
  ],
  "refs": [
   {
    "label": "LOCATION PHOTOGRAPH — the exact place of this shot: its architecture, materials, fixed features and lighting mood are spatial truth; stage the moment inside this place. Never copy its camera framing.",
    "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/db00517b-165b-4005-a287-f28936b59b95/images/272bd7af-9294-4490-98c6-88a40e9deacc/scene/recipe/groupbg_lake_bamboo_entry_11c8f7.png",
    "asset_id": "cc3149c9-bdcd-4a60-b603-ee58e6473a40",
    "role": "bgfirst_group_bg"
   },
   {
    "label": "CHARACTER REFERENCE — 안규홍: the exact person appearing in this shot; match face, hair and build exactly.",
    "path": "<bytes:1028384>",
    "asset_id": "3b5894db-ce55-4f34-b966-5e725c3aa952",
    "role": "character_ref"
   },
   {
    "label": "CHARACTER REFERENCE — 이관회: the exact person appearing in this shot; match face, hair and build exactly.",
    "path": "<bytes:933540>",
    "asset_id": "851b450f-ac84-47fe-ac5a-03a56f63fcae",
    "role": "character_ref"
   },
   {
    "label": "PROP REFERENCE — 지게: the exact object appearing in this shot; match its look, material and wear exactly.",
    "path": "<bytes:1157225>",
    "asset_id": "bdfa122b-f62f-4ee2-807f-3e7997ed8125",
    "role": "prop_ref"
   },
   {
    "label": "PROP REFERENCE — 곡식 자루·가마니: the exact object appearing in this shot; match its look, material and wear exactly.",
    "path": "<bytes:1807398>",
    "asset_id": "e30c319e-fdc7-4c01-a301-290ddd45d27a",
    "role": "prop_ref"
   }
  ],
  "critique_skipped": true,
  "needs_reshoot": true,
  "shot_run_produced": false,
  "shot_run_uid": "06aa1412-526f-7240-b951-b14e3d586fef",
  "bgfirst": {
   "bg_path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/db00517b-165b-4005-a287-f28936b59b95/images/272bd7af-9294-4490-98c6-88a40e9deacc/scene/recipe/S18sh4__bgfirst_bg.png",
   "bg_asset_id": "a9dee6e0-0fb8-4884-975e-987652aa7464",
   "bg_record_key": "S18sh4::bgfirst_bg",
   "chain_winner": true,
   "authority": "groupbg",
   "group_key": "lake_bamboo_entry",
   "groupbg_asset_id": "cc3149c9-bdcd-4a60-b603-ee58e6473a40"
  },
  "ref_mode": "재투영 배경+콘티+엔티티 (2택1: 체인 승)",
  "share_plan": {
   "ref_plan": "background"
  }
 },
 "S18sh4::cine": {
  "applied": true,
  "attempted_at": "2026-09-09T02:54:54.989478+00:00",
  "fingerprint": "c345e584ee80e8091cd7f774709f08a8cd81eec8e9c16a91628d13f113d164cc",
  "fingerprint_version": 2,
  "provider": "mai",
  "endpoint": "openrouter/chat-completions",
  "model": "microsoft/mai-image-2.6",
  "pack": "24.202608252115",
  "source_file": "S18sh4_sel.png",
  "source_sha256": "257314b8d92263d3b3e5d5c253e2661c66f7ecff92e5541d27e76b02d6889b5d",
  "file": "S18sh4_cine.png",
  "staged_sha256": "fdbbfd26d62b374dd47c56d81b3d026866113ff247c12a6f77529c32a49a8acf",
  "latency_ms": 29008
 },
 "S18sh14::signage": {
  "fp": "bc43ed7e322036e1",
  "inscriptions": [],
  "cues": [],
  "dropped": []
 },
 "S18sh14": {
  "input_fingerprint": "956b91a62b586c5f",
  "prompt": "Create ONE FINAL photorealistic live-action film still of the moment below — early 20th-century Korea, Jeolla region; all people are Korean unless explicitly stated otherwise. TIME OF DAY (lock): late afternoon, reddish natural light.\n\nSHOT TEXT (authoritative, Korean): 왁자지껄하게 웃음을 터뜨리는 조선인 의병들 너머로 팔짱을 낀 채 우두커니 선 강용언의 전신.\n\nLOCATION (lock): The exterior frontage of the commander's tent, across the active cooking clearing of the lakeside camp. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: wide shot\n- FRAME LAYOUT: Tent frontage visible through the gap between foreground militia members in the middle-center of the frame, background.\n- KEY BACKGROUND ELEMENTS: 대장 막사 (Behind 강용언 as he remains outside the gathering) — The front of the tent is seen obliquely behind his full figure; used as Anchors his separation from the foreground cooking group; 취사 공간과 막사 사이 공터 (Separates the lively group from 강용언); used as A visible interval through the foreground gap prevents the leader from appearing part of the gathering.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Keep the established late-afternoon ambient light and muted contrast, withholding the explicit red firelight that belongs to the subsequent beat.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The clothing, hair and overall look of 강용언 — who appear both in that photo and in this shot — are LOCKED to that photo. Anyone else visible in that photo is NOT in this shot: never carry their face, body or clothing onto anyone here. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): Barley sacks and a potato basket now rest on the gravel beside the pot; the two seed-rice sacks remain tied and untouched while the cooking fire burns strongly. 강용언: He stands rigidly with folded arms before the tent, his face lit red by the cooking fire.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 강용언 (한국인 남성, 성인 나이대, 햇볕에 그을린 얼굴, 자연스러운 검은 머리) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.",
  "roll_prompts": {
   "A": "Create ONE FINAL photorealistic live-action film still of the moment below — early 20th-century Korea, Jeolla region; all people are Korean unless explicitly stated otherwise. TIME OF DAY (lock): late afternoon, reddish natural light.\n\nSHOT TEXT (authoritative, Korean): 왁자지껄하게 웃음을 터뜨리는 조선인 의병들 너머로 팔짱을 낀 채 우두커니 선 강용언의 전신.\n\nLOCATION (lock): The exterior frontage of the commander's tent, across the active cooking clearing of the lakeside camp. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: wide shot\n- FRAME LAYOUT: Tent frontage visible through the gap between foreground militia members in the middle-center of the frame, background.\n- KEY BACKGROUND ELEMENTS: 대장 막사 (Behind 강용언 as he remains outside the gathering) — The front of the tent is seen obliquely behind his full figure; used as Anchors his separation from the foreground cooking group; 취사 공간과 막사 사이 공터 (Separates the lively group from 강용언); used as A visible interval through the foreground gap prevents the leader from appearing part of the gathering.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Keep the established late-afternoon ambient light and muted contrast, withholding the explicit red firelight that belongs to the subsequent beat.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The clothing, hair and overall look of 강용언 — who appear both in that photo and in this shot — are LOCKED to that photo. Anyone else visible in that photo is NOT in this shot: never carry their face, body or clothing onto anyone here. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): Barley sacks and a potato basket now rest on the gravel beside the pot; the two seed-rice sacks remain tied and untouched while the cooking fire burns strongly. 강용언: He stands rigidly with folded arms before the tent, his face lit red by the cooking fire.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 강용언 (한국인 남성, 성인 나이대, 햇볕에 그을린 얼굴, 자연스러운 검은 머리) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.",
   "B": "Create ONE FINAL photorealistic live-action film still of the moment below — early 20th-century Korea, Jeolla region; all people are Korean unless explicitly stated otherwise. TIME OF DAY (lock): late afternoon, reddish natural light.\n\nSHOT TEXT (authoritative, Korean): 왁자지껄하게 웃음을 터뜨리는 조선인 의병들 너머로 팔짱을 낀 채 우두커니 선 강용언의 전신.\n\nLOCATION (lock): The exterior frontage of the commander's tent, across the active cooking clearing of the lakeside camp. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: wide shot\n- FRAME LAYOUT: Tent frontage visible through the gap between foreground militia members in the middle-center of the frame, background.\n- KEY BACKGROUND ELEMENTS: 대장 막사 (Behind 강용언 as he remains outside the gathering) — The front of the tent is seen obliquely behind his full figure; used as Anchors his separation from the foreground cooking group; 취사 공간과 막사 사이 공터 (Separates the lively group from 강용언); used as A visible interval through the foreground gap prevents the leader from appearing part of the gathering.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Keep the established late-afternoon ambient light and muted contrast, withholding the explicit red firelight that belongs to the subsequent beat.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The clothing, hair and overall look of 강용언 — who appear both in that photo and in this shot — are LOCKED to that photo. Anyone else visible in that photo is NOT in this shot: never carry their face, body or clothing onto anyone here. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): Barley sacks and a potato basket now rest on the gravel beside the pot; the two seed-rice sacks remain tied and untouched while the cooking fire burns strongly. 강용언: He stands rigidly with folded arms before the tent, his face lit red by the cooking fire.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 강용언 (한국인 남성, 성인 나이대, 햇볕에 그을린 얼굴, 자연스러운 검은 머리) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.\n\nCOMPOSITION VARIATION (this candidate only): keep the place, the people and the moment exactly as specified, and keep from the CAMERA & FRAME contract its subject scale, screen placement and key background elements exactly. For this candidate only, its camera angle and height are deliberately left open: choose a camera position distinctly different from the obvious one, as a film director picking a second setup on the same blocking. Stay true to any explicit framing the SHOT TEXT itself dictates."
  },
  "shot_run_spend_attempt_count": 1,
  "readings": [
   {
    "label": "A",
    "direction": "전경의 의병들은 서로를 마주보며 활짝 웃고 있으며, 배경의 강용언은 팔짱을 낀 채 정면 방향을 묵묵히 응시하고 있습니다.",
    "built_space": "우측에 강, 좌측과 배경에 대나무 숲이 있으며, 중앙에 대장 막사가 위치합니다. 전경의 요리 공간과 막사 사이에 지시된 빈 공터가 적절하게 확보되어 공간감이 잘 느껴집니다.",
    "entities": "강용언은 제공된 레퍼런스와 일치하는 얼굴, 수염, 깃털 달린 모자와 푸른색 복장을 정확히 착용하고 있습니다. 전경에는 가마솥과 모닥불, 감자가 담긴 바구니와 자루 등 요구된 소품들이 글씨 없이 자연스럽게 배치되어 있습니다.",
    "hard_violations": [],
    "physics": "모든 인물이 땅에 서 있거나 안정적인 자세로 웅크려 앉아 있으며, 팔짱을 낀 자세나 땅에 놓인 솥과 소품들 모두 중력에 맞게 자연스럽게 지지되고 있습니다."
   },
   {
    "label": "B",
    "direction": "의병들은 불가를 둘러싸고 앉아 서로를 바라보며 웃고 있고, 막사 앞의 강용언은 팔짱을 낀 채 앞을 바라보고 있습니다.",
    "built_space": "막사, 대나무 숲, 강의 위치가 레퍼런스 구도에 맞게 배치되어 있으며 전경의 솥과 막사 사이의 거리감도 적절히 표현되었습니다.",
    "entities": "강용언의 외모와 의상은 레퍼런스와 부합하며, 전경에 의병들과 가마솥, 감자 바구니가 배치되어 있습니다. 그러나 씨앗 자루에 영문으로 된 글씨가 뚜렷하게 적혀 있습니다.",
    "hard_violations": [
     "[gemini-pro] 이미지 내에 읽을 수 있는 텍스트 금지('No readable writing anywhere') 지시를 완전히 위반하여, 전경의 자루들에 'SEED-RICE'라는 영어 텍스트가 선명하게 새겨져 있음.",
     "[gpt-high] 종자 자루에 읽을 수 있는 영문 표기가 있어 모든 가독성 있는 글자를 금지한 지시를 위반한다.",
     "[gpt-high] 중앙 전경과 오른쪽 후경에 가마솥·돌 화덕이 각각 있어 기준 장소의 취사 설비를 중복 생성했다."
    ],
    "physics": "모닥불을 받치는 돌과 솥, 인물들의 서 있거나 앉은 자세 모두 물리적으로 어긋남 없이 바닥에 잘 지지되어 있습니다."
   }
  ],
  "cross_model_order": {
   "policy": "cross_model_order_v2_slot0_fwd_slot1_rev_parallel2",
   "slots": [
    {
     "model": "gemini-pro",
     "order": "forward",
     "display_to_canonical": {
      "A": "A",
      "B": "B"
     },
     "raw": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 9,
        "verdict_ko": "지시된 화면 구도와 인물 배치를 훌륭하게 구현했으며, 텍스트가 나타나지 않아야 한다는 제약 조건을 완벽히 준수했습니다."
       },
       {
        "label": "B",
        "score": 3,
        "verdict_ko": "이미지 내에 어떠한 글씨도 있어서는 안 된다는 강력한 지시사항이 무색하게, 전경의 자루에 'SEED-RICE'라는 영어 텍스트가 명확하게 생성되어 심각한 오류가 발생했습니다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "전경의 의병들은 서로를 마주보며 활짝 웃고 있으며, 배경의 강용언은 팔짱을 낀 채 정면 방향을 묵묵히 응시하고 있습니다.",
        "built_space": "우측에 강, 좌측과 배경에 대나무 숲이 있으며, 중앙에 대장 막사가 위치합니다. 전경의 요리 공간과 막사 사이에 지시된 빈 공터가 적절하게 확보되어 공간감이 잘 느껴집니다.",
        "entities": "강용언은 제공된 레퍼런스와 일치하는 얼굴, 수염, 깃털 달린 모자와 푸른색 복장을 정확히 착용하고 있습니다. 전경에는 가마솥과 모닥불, 감자가 담긴 바구니와 자루 등 요구된 소품들이 글씨 없이 자연스럽게 배치되어 있습니다.",
        "hard_violations": [],
        "physics": "모든 인물이 땅에 서 있거나 안정적인 자세로 웅크려 앉아 있으며, 팔짱을 낀 자세나 땅에 놓인 솥과 소품들 모두 중력에 맞게 자연스럽게 지지되고 있습니다."
       },
       {
        "label": "B",
        "direction": "의병들은 불가를 둘러싸고 앉아 서로를 바라보며 웃고 있고, 막사 앞의 강용언은 팔짱을 낀 채 앞을 바라보고 있습니다.",
        "built_space": "막사, 대나무 숲, 강의 위치가 레퍼런스 구도에 맞게 배치되어 있으며 전경의 솥과 막사 사이의 거리감도 적절히 표현되었습니다.",
        "entities": "강용언의 외모와 의상은 레퍼런스와 부합하며, 전경에 의병들과 가마솥, 감자 바구니가 배치되어 있습니다. 그러나 씨앗 자루에 영문으로 된 글씨가 뚜렷하게 적혀 있습니다.",
        "hard_violations": [
         "이미지 내에 읽을 수 있는 텍스트 금지('No readable writing anywhere') 지시를 완전히 위반하여, 전경의 자루들에 'SEED-RICE'라는 영어 텍스트가 선명하게 새겨져 있음."
        ],
        "physics": "모닥불을 받치는 돌과 솥, 인물들의 서 있거나 앉은 자세 모두 물리적으로 어긋남 없이 바닥에 잘 지지되어 있습니다."
       }
      ],
      "all_candidates_fail": false
     },
     "normalized": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 9,
        "verdict_ko": "지시된 화면 구도와 인물 배치를 훌륭하게 구현했으며, 텍스트가 나타나지 않아야 한다는 제약 조건을 완벽히 준수했습니다."
       },
       {
        "label": "B",
        "score": 3,
        "verdict_ko": "이미지 내에 어떠한 글씨도 있어서는 안 된다는 강력한 지시사항이 무색하게, 전경의 자루에 'SEED-RICE'라는 영어 텍스트가 명확하게 생성되어 심각한 오류가 발생했습니다."
       }
      ],
      "all_candidates_fail": false,
      "readings": [
       {
        "label": "A",
        "direction": "전경의 의병들은 서로를 마주보며 활짝 웃고 있으며, 배경의 강용언은 팔짱을 낀 채 정면 방향을 묵묵히 응시하고 있습니다.",
        "built_space": "우측에 강, 좌측과 배경에 대나무 숲이 있으며, 중앙에 대장 막사가 위치합니다. 전경의 요리 공간과 막사 사이에 지시된 빈 공터가 적절하게 확보되어 공간감이 잘 느껴집니다.",
        "entities": "강용언은 제공된 레퍼런스와 일치하는 얼굴, 수염, 깃털 달린 모자와 푸른색 복장을 정확히 착용하고 있습니다. 전경에는 가마솥과 모닥불, 감자가 담긴 바구니와 자루 등 요구된 소품들이 글씨 없이 자연스럽게 배치되어 있습니다.",
        "hard_violations": [],
        "physics": "모든 인물이 땅에 서 있거나 안정적인 자세로 웅크려 앉아 있으며, 팔짱을 낀 자세나 땅에 놓인 솥과 소품들 모두 중력에 맞게 자연스럽게 지지되고 있습니다."
       },
       {
        "label": "B",
        "direction": "의병들은 불가를 둘러싸고 앉아 서로를 바라보며 웃고 있고, 막사 앞의 강용언은 팔짱을 낀 채 앞을 바라보고 있습니다.",
        "built_space": "막사, 대나무 숲, 강의 위치가 레퍼런스 구도에 맞게 배치되어 있으며 전경의 솥과 막사 사이의 거리감도 적절히 표현되었습니다.",
        "entities": "강용언의 외모와 의상은 레퍼런스와 부합하며, 전경에 의병들과 가마솥, 감자 바구니가 배치되어 있습니다. 그러나 씨앗 자루에 영문으로 된 글씨가 뚜렷하게 적혀 있습니다.",
        "hard_violations": [
         "이미지 내에 읽을 수 있는 텍스트 금지('No readable writing anywhere') 지시를 완전히 위반하여, 전경의 자루들에 'SEED-RICE'라는 영어 텍스트가 선명하게 새겨져 있음."
        ],
        "physics": "모닥불을 받치는 돌과 솥, 인물들의 서 있거나 앉은 자세 모두 물리적으로 어긋남 없이 바닥에 잘 지지되어 있습니다."
       }
      ]
     },
     "ok": true
    },
    {
     "model": "gpt-high",
     "order": "reverse",
     "display_to_canonical": {
      "A": "B",
      "B": "A"
     },
     "raw": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 2,
        "verdict_ko": "전경의 웃음과 후경의 팔짱 낀 전신은 구현했지만, 자루의 읽을 수 있는 영문과 중복된 가마솥·화덕이 명시적 금지 및 장소 고정을 위반한다."
       },
       {
        "label": "B",
        "score": 8,
        "verdict_ko": "전경 의병 사이 공터 너머 강용언의 전신과 정적인 팔짱 자세를 충실히 구현했으며, 막사 지붕 형태와 묶인 종자 자루의 연속성은 부족하다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "전경 의병들은 좌우 맞은편 동료를 보며 입을 벌려 웃고, 일부는 동료 쪽으로 손을 내민다. 중앙 후경의 강용언은 전경 모임 방향을 바라본다. 강용언의 칼집은 허리에서 화면 오른쪽 아래로 내려가며 누구를 겨누지 않는다.",
        "built_space": "중앙 뒤에 출입구 하나를 가진 큰 막사, 왼쪽에 작은 막사, 오른쪽에 호수, 뒤쪽에 대숲이 있다. 의병 사이로 강용언의 발끝까지 보이고 취사 모임과의 간격도 확보된다. 다만 중앙 전경에 가마솥과 돌 화덕 한 벌이 있고, 오른쪽 후경에도 별도의 가마솥과 돌 화덕이 보여 이전 장면의 단일 취사 설비가 중복된다. 주 막사는 기준 사진보다 정면에 가깝다.",
        "entities": "강용언 한 명과 웃는 의병 열 명이 보인다. 모두 성인 한국인 남성이라는 설정에 어긋나는 뚜렷한 외형은 없다. 후경 강용언의 검은 모자, 붉은 장식과 허리띠, 청회색 군복, 칼집은 기준과 대체로 맞는다. 왼쪽 의병 두 명도 강용언과 매우 비슷한 군복을 입어 인물 구별이 약하다. 감자 바구니 하나, 묶인 자루 두 개, 열린 곡물 자루 두 개가 보이지만 곡물 종류는 확정하기 어렵다. 묶인 자루에는 읽을 수 있는 영문 'SEED-RICE'가 있다. 자연광은 따뜻하지만 붉은 늦오후 느낌은 약하다.",
        "hard_violations": [
         "종자 자루에 읽을 수 있는 영문 표기가 있어 모든 가독성 있는 글자를 금지한 지시를 위반한다.",
         "중앙 전경과 오른쪽 후경에 가마솥·돌 화덕이 각각 있어 기준 장소의 취사 설비를 중복 생성했다."
        ],
        "physics": "강용언과 서 있는 의병들은 발로 지면을 딛고 있다. 낮게 앉은 의병들의 좌면은 몸과 소품에 가려져 있지만 발은 땅에 닿고 무릎이 굽혀져 있어 공중에 떠 있다고 볼 근거는 없다. 중앙 솥은 돌 화덕 위에 걸쳐 있고 장작은 지면에 놓여 타고 있다. 자루와 바구니는 땅에 놓이며 칼집은 허리띠에 매달려 있다."
       },
       {
        "label": "B",
        "direction": "좌우 전경 의병들은 서로를 향해 몸과 얼굴을 돌리고 웃는다. 등을 보인 두 의병도 모임 안쪽을 향한다. 강용언은 그들 너머에서 모임 쪽을 바라보며 팔짱을 끼고 있다. 허리의 칼집은 화면 오른쪽 아래를 향해 늘어져 있고 겨누는 동작은 없다.",
        "built_space": "왼쪽 후경에 출입구 하나인 막사가 비스듬히 보이고, 그 앞 중앙에 강용언이 서 있다. 전경 의병들 사이에 넓은 빈 땅이 드러나 모임과 강용언의 분리가 명확하다. 중앙 전경에는 가마솥과 돌 화덕 한 벌, 왼쪽에는 그릇을 둔 멍석, 오른쪽에는 자루와 바구니가 있다. 호수와 대숲의 관계는 유지되지만 막사는 기준의 거친 맞배지붕보다 둥글고 정돈된 초가지붕으로 바뀌었고 출입구의 천 덮개도 보이지 않는다.",
        "entities": "강용언 한 명과 전경 의병 여덟 명이 보인다. 성인 한국인 남성 설정과 대체로 부합하며, 강용언의 얼굴 인상과 체격, 장식 달린 검은 모자, 청회색 군복, 붉은 허리띠, 칼집과 짚신이 기준에 가깝다. 의병들의 웃음과 강용언의 무표정한 정지가 대비된다. 솥 하나, 감자 바구니 하나와 열린 자루 두 개는 보이지만 묶인 종자 자루 두 개는 확인되지 않는다. 읽을 수 있는 글자는 없다. 불빛이 얼굴을 과장되게 붉히지 않는 점은 조명 우선 지시와 맞지만 늦오후의 붉은 자연광은 약하다.",
        "hard_violations": [],
        "physics": "강용언은 두 발로 땅을 딛고 팔을 가슴 앞에 접어 서 있다. 전경 인물들의 하체는 대부분 프레임 밖이지만 보이는 몸통의 기울기와 팔 자세는 웃거나 앞으로 숙이는 동작으로 가능하며 부유 징후는 없다. 솥은 돌 화덕에 받쳐져 있고 불과 장작은 그 아래에 있다. 자루·바구니·멍석·그릇은 지면에 놓여 있으며 칼집은 허리띠에서 지지된다."
       }
      ],
      "all_candidates_fail": false
     },
     "normalized": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 2,
        "verdict_ko": "전경의 웃음과 후경의 팔짱 낀 전신은 구현했지만, 자루의 읽을 수 있는 영문과 중복된 가마솥·화덕이 명시적 금지 및 장소 고정을 위반한다."
       },
       {
        "label": "A",
        "score": 8,
        "verdict_ko": "전경 의병 사이 공터 너머 강용언의 전신과 정적인 팔짱 자세를 충실히 구현했으며, 막사 지붕 형태와 묶인 종자 자루의 연속성은 부족하다."
       }
      ],
      "all_candidates_fail": false,
      "readings": [
       {
        "label": "B",
        "direction": "전경 의병들은 좌우 맞은편 동료를 보며 입을 벌려 웃고, 일부는 동료 쪽으로 손을 내민다. 중앙 후경의 강용언은 전경 모임 방향을 바라본다. 강용언의 칼집은 허리에서 화면 오른쪽 아래로 내려가며 누구를 겨누지 않는다.",
        "built_space": "중앙 뒤에 출입구 하나를 가진 큰 막사, 왼쪽에 작은 막사, 오른쪽에 호수, 뒤쪽에 대숲이 있다. 의병 사이로 강용언의 발끝까지 보이고 취사 모임과의 간격도 확보된다. 다만 중앙 전경에 가마솥과 돌 화덕 한 벌이 있고, 오른쪽 후경에도 별도의 가마솥과 돌 화덕이 보여 이전 장면의 단일 취사 설비가 중복된다. 주 막사는 기준 사진보다 정면에 가깝다.",
        "entities": "강용언 한 명과 웃는 의병 열 명이 보인다. 모두 성인 한국인 남성이라는 설정에 어긋나는 뚜렷한 외형은 없다. 후경 강용언의 검은 모자, 붉은 장식과 허리띠, 청회색 군복, 칼집은 기준과 대체로 맞는다. 왼쪽 의병 두 명도 강용언과 매우 비슷한 군복을 입어 인물 구별이 약하다. 감자 바구니 하나, 묶인 자루 두 개, 열린 곡물 자루 두 개가 보이지만 곡물 종류는 확정하기 어렵다. 묶인 자루에는 읽을 수 있는 영문 'SEED-RICE'가 있다. 자연광은 따뜻하지만 붉은 늦오후 느낌은 약하다.",
        "hard_violations": [
         "종자 자루에 읽을 수 있는 영문 표기가 있어 모든 가독성 있는 글자를 금지한 지시를 위반한다.",
         "중앙 전경과 오른쪽 후경에 가마솥·돌 화덕이 각각 있어 기준 장소의 취사 설비를 중복 생성했다."
        ],
        "physics": "강용언과 서 있는 의병들은 발로 지면을 딛고 있다. 낮게 앉은 의병들의 좌면은 몸과 소품에 가려져 있지만 발은 땅에 닿고 무릎이 굽혀져 있어 공중에 떠 있다고 볼 근거는 없다. 중앙 솥은 돌 화덕 위에 걸쳐 있고 장작은 지면에 놓여 타고 있다. 자루와 바구니는 땅에 놓이며 칼집은 허리띠에 매달려 있다."
       },
       {
        "label": "A",
        "direction": "좌우 전경 의병들은 서로를 향해 몸과 얼굴을 돌리고 웃는다. 등을 보인 두 의병도 모임 안쪽을 향한다. 강용언은 그들 너머에서 모임 쪽을 바라보며 팔짱을 끼고 있다. 허리의 칼집은 화면 오른쪽 아래를 향해 늘어져 있고 겨누는 동작은 없다.",
        "built_space": "왼쪽 후경에 출입구 하나인 막사가 비스듬히 보이고, 그 앞 중앙에 강용언이 서 있다. 전경 의병들 사이에 넓은 빈 땅이 드러나 모임과 강용언의 분리가 명확하다. 중앙 전경에는 가마솥과 돌 화덕 한 벌, 왼쪽에는 그릇을 둔 멍석, 오른쪽에는 자루와 바구니가 있다. 호수와 대숲의 관계는 유지되지만 막사는 기준의 거친 맞배지붕보다 둥글고 정돈된 초가지붕으로 바뀌었고 출입구의 천 덮개도 보이지 않는다.",
        "entities": "강용언 한 명과 전경 의병 여덟 명이 보인다. 성인 한국인 남성 설정과 대체로 부합하며, 강용언의 얼굴 인상과 체격, 장식 달린 검은 모자, 청회색 군복, 붉은 허리띠, 칼집과 짚신이 기준에 가깝다. 의병들의 웃음과 강용언의 무표정한 정지가 대비된다. 솥 하나, 감자 바구니 하나와 열린 자루 두 개는 보이지만 묶인 종자 자루 두 개는 확인되지 않는다. 읽을 수 있는 글자는 없다. 불빛이 얼굴을 과장되게 붉히지 않는 점은 조명 우선 지시와 맞지만 늦오후의 붉은 자연광은 약하다.",
        "hard_violations": [],
        "physics": "강용언은 두 발로 땅을 딛고 팔을 가슴 앞에 접어 서 있다. 전경 인물들의 하체는 대부분 프레임 밖이지만 보이는 몸통의 기울기와 팔 자세는 웃거나 앞으로 숙이는 동작으로 가능하며 부유 징후는 없다. 솥은 돌 화덕에 받쳐져 있고 불과 장작은 그 아래에 있다. 자루·바구니·멍석·그릇은 지면에 놓여 있으며 칼집은 허리띠에서 지지된다."
       }
      ]
     },
     "ok": true
    }
   ],
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "slot_winner_match": true,
   "slot_winner": {
    "gemini-pro": "A",
    "gpt-high": "A"
   },
   "route": "cross_slot_agree"
  },
  "dual": {
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "normalized": {
    "A": 2.0,
    "B": 0.583
   },
   "adjusted": {
    "A": 2.0,
    "B": 0.333
   },
   "violations": {
    "B": [
     "[gemini-pro] 이미지 내에 읽을 수 있는 텍스트 금지('No readable writing anywhere') 지시를 완전히 위반하여, 전경의 자루들에 'SEED-RICE'라는 영어 텍스트가 선명하게 새겨져 있음.",
     "[gpt-high] 종자 자루에 읽을 수 있는 영문 표기가 있어 모든 가독성 있는 글자를 금지한 지시를 위반한다.",
     "[gpt-high] 중앙 전경과 오른쪽 후경에 가마솥·돌 화덕이 각각 있어 기준 장소의 취사 설비를 중복 생성했다."
    ]
   },
   "per_model_winner": {
    "gemini-pro": "A",
    "gpt-high": "A"
   },
   "agreed": true
  },
  "totals": {
   "A": 2000,
   "B": 333
  },
  "selected": "A",
  "ranking": [
   "A",
   "B"
  ],
  "verdicts": [
   {
    "label": "A",
    "score": 2000,
    "verdict_ko": "지시된 화면 구도와 인물 배치를 훌륭하게 구현했으며, 텍스트가 나타나지 않아야 한다는 제약 조건을 완벽히 준수했습니다."
   },
   {
    "label": "B",
    "score": 333,
    "verdict_ko": "이미지 내에 어떠한 글씨도 있어서는 안 된다는 강력한 지시사항이 무색하게, 전경의 자루에 'SEED-RICE'라는 영어 텍스트가 명확하게 생성되어 심각한 오류가 발생했습니다.  ★위반: [gemini-pro] 이미지 내에 읽을 수 있는 텍스트 금지('No readable writing anywhere') 지시를 완전히 위반하여, 전경의 자루들에 'SEED-RICE'라는 영어 텍스트가 선명하게 새겨져 있음. / [gpt-high] 종자 자루에 읽을 수 있는 영문 표기가 있어 모든 가독성 있는 글자를 금지한 지시를 위반한다. / [gpt-high] 중앙 전경과 오른쪽 후경에 가마솥·돌 화덕이 각각 있어 기준 장소의 취사 설비를 중복 생성했다."
   }
  ],
  "refs": [
   {
    "label": "PREVIOUS SHOT STILL — a visually related earlier shot of this same place: the location's look, materials, fixed features and lighting mood are LOCKED to this photo; never copy its camera framing. The clothing, hair and overall look of 강용언 — who appear both in that photo and in this shot — are LOCKED to that photo. Anyone else visible in that photo is NOT in this shot: never carry their face, body or clothing onto anyone here. If this photo shows a character who cannot move (dead or unconscious), that character's exact pose, position and orientation are ALSO LOCKED — treat their body as a fixed prop of the set that only the camera moves around.",
    "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/db00517b-165b-4005-a287-f28936b59b95/images/272bd7af-9294-4490-98c6-88a40e9deacc/scene/recipe/S14sh9_sel.png",
    "asset_id": "1e7bf128-78d9-4835-b119-cb0a12b6268a",
    "role": "prev_still"
   },
   {
    "label": "CHARACTER REFERENCE — 강용언: the exact person appearing in this shot; match face, hair and build exactly.",
    "path": "<bytes:822588>",
    "asset_id": "1a75090f-1eb9-4224-a132-de7a4b454b73",
    "role": "character_ref"
   }
  ],
  "critique_skipped": true,
  "shot_run_produced": false,
  "shot_run_uid": "06aa1412-56bc-774f-89b4-654fbb00948d",
  "ref_mode": "prev+엔티티",
  "share_plan": {
   "ref_plan": "prev",
   "prev_anchor_tag": "S14sh9"
  },
  "lane_policy": "ab_select_bypass:prev"
 },
 "S18sh14::cine": {
  "applied": true,
  "attempted_at": "2026-09-09T02:56:35.251004+00:00",
  "fingerprint": "319bcdae5fceab60aa1bb156a0d3e13e1a9c69d5fe8d5a67fdf6c6ea777cabd5",
  "fingerprint_version": 2,
  "provider": "mai",
  "endpoint": "openrouter/chat-completions",
  "model": "microsoft/mai-image-2.6",
  "pack": "24.202608252115",
  "source_file": "S18sh14_sel.png",
  "source_sha256": "b9c5c3047db7578b80dcc310ca80bce556ca956c725d8c9be9373d57a0196f55",
  "file": "S18sh14_cine.png",
  "staged_sha256": "552d705f17f3f0bc71e42040c5bbd8b25457d0ee5dc8408d92ea593e1d191ebd",
  "latency_ms": 26167
 },
 "S19sh17::signage": {
  "fp": "8d2526fe3600278e",
  "inscriptions": [],
  "cues": [],
  "dropped": []
 },
 "groupbg::lake_fire_circle": {
  "input_fingerprint": "d756fef7aa3ebdfa",
  "meta": {
   "model": "gpt-image-2",
   "size": "1536x864",
   "pack": "11.202607220237",
   "contract": "bgfirst_full_v3",
   "group_sig": {
    "key": "lake_fire_circle",
    "tags": [
     "S19sh17",
     "S19sh25",
     "S19sh31"
    ]
   },
   "context_sig": "d77ba094f12420a5"
  },
  "prompt": "Create ONE empty live-action location background photograph — NO PEOPLE, no figures, no body parts, no silhouettes, no shadows or reflections of people anywhere.\n\"Empty\" means no people only: KEEP the location's inherent occupants and stock that define the place — animals in an animal shelter, pen or farm, goods and produce in a market, moored boats in a harbour — unless the shot text explicitly removes them.\nThe FIRST attached image is a thin-line storyboard sketch of a shot that happens at this location: use it ONLY as spatial evidence — what this place contains, how its ground, structures and landmarks are arranged and proportioned. Ignore the sketched people and arrows entirely, and do NOT copy its line style: render a fully photographic, physically plausible real place that fits THE LOCATION text below.\nHUMAN-SCALE CALIBRATION: derive every structure's true size from human-scale elements — a door ≈ 2m, a window ≈ 1–1.5m wide, one storey ≈ 2.5–3m; never inflate a small structure or shrink a large one.\n\nTHE LOCATION — early 20th-century Korea, Jeolla region; all people are Korean unless explicitly stated otherwise: The dirt space beside the communal campfire in the lakeside camp, where a wooden carrying frame lies within the fighting area.\n\nLOCATION DETAIL (descriptive evidence about this place from the production's location records — evidence, not a staging order): use it to understand what kind of place this is and what it permanently contains. Carry over ONLY its enduring physical features — layout, structures, ground and materials, fixed equipment and installations. Do NOT treat any lighting, weather or time-of-day wording, momentary object states, depicted events or subjective impressions in it as fixed facts of the place; the TIME OF DAY lock below is the sole authority for time and lighting.\n보성 호숫가 강용언 임시거점 자갈밭·공터·막사 앞: 잿빛 호숫가에 펼쳐진 자갈밭과 흙 공터. 대나무 숲을 등지고 있으며, 낡은 천과 거적을 엮어 만든 조잡한 임시 막사 여러 채, 큰 무쇠솥이 얹힌 돌 화덕이 배치되어 있다. (특징: 넓은 잿빛 호수면; 자갈과 흙이 섞인 공터; 낡은 천, 짚, 거적으로 거칠게 엮은 임시 천막/막사; 불이 타오르는 돌 화덕; 끓고 있는 큰 무쇠 가마솥; 쌓여 있는 짐 꾸러미와 곡식 자루)\n\nSCENE EVIDENCE (verbatim quotes from the screenplay about this place — treat them as evidence of what the location physically contains and looks like; stage the PLACE those moments happen in, but do NOT depict the momentary actions, people or staged props themselves):\n- 사내들이 뜨거운 닭고기를 뜯어 허겁지겁 먹는다. 저만치 모닥불가.\n- 모닥불 주변.\n배를 채운 의병들이 규홍옆에 모여 웃고 떠들고 있다\n- 강용언이 무릎을 꿇은 채 그 모습을 본다. 복만만 남아 있다.\n\nTIME OF DAY (lock): night.\n\nRender ONE photorealistic empty location photograph, 16:9, neutral enough that every shot of this place can be staged from it later. No readable writing anywhere: surfaces that would carry writing may be present, but stage any wording out of legibility — an oblique angle, distance, shallow focus. No captions, watermarks or overlay text.",
  "effective_prompt": "Create ONE empty live-action location background photograph — NO PEOPLE, no figures, no body parts, no silhouettes, no shadows or reflections of people anywhere.\n\"Empty\" means no people only: KEEP the location's inherent occupants and stock that define the place — animals in an animal shelter, pen or farm, goods and produce in a market, moored boats in a harbour — unless the shot text explicitly removes them.\nThe FIRST attached image is a thin-line storyboard sketch of a shot that happens at this location: use it ONLY as spatial evidence — what this place contains, how its ground, structures and landmarks are arranged and proportioned. Ignore the sketched people and arrows entirely, and do NOT copy its line style: render a fully photographic, physically plausible real place that fits THE LOCATION text below.\nHUMAN-SCALE CALIBRATION: derive every structure's true size from human-scale elements — a door ≈ 2m, a window ≈ 1–1.5m wide, one storey ≈ 2.5–3m; never inflate a small structure or shrink a large one.\n\nTHE LOCATION — early 20th-century Korea, Jeolla region; all people are Korean unless explicitly stated otherwise: The dirt space beside the communal campfire in the lakeside camp, where a wooden carrying frame lies within the fighting area.\n\nLOCATION DETAIL (descriptive evidence about this place from the production's location records — evidence, not a staging order): use it to understand what kind of place this is and what it permanently contains. Carry over ONLY its enduring physical features — layout, structures, ground and materials, fixed equipment and installations. Do NOT treat any lighting, weather or time-of-day wording, momentary object states, depicted events or subjective impressions in it as fixed facts of the place; the TIME OF DAY lock below is the sole authority for time and lighting.\n보성 호숫가 강용언 임시거점 자갈밭·공터·막사 앞: 잿빛 호숫가에 펼쳐진 자갈밭과 흙 공터. 대나무 숲을 등지고 있으며, 낡은 천과 거적을 엮어 만든 조잡한 임시 막사 여러 채, 큰 무쇠솥이 얹힌 돌 화덕이 배치되어 있다. (특징: 넓은 잿빛 호수면; 자갈과 흙이 섞인 공터; 낡은 천, 짚, 거적으로 거칠게 엮은 임시 천막/막사; 불이 타오르는 돌 화덕; 끓고 있는 큰 무쇠 가마솥; 쌓여 있는 짐 꾸러미와 곡식 자루)\n\nSCENE EVIDENCE (verbatim quotes from the screenplay about this place — treat them as evidence of what the location physically contains and looks like; stage the PLACE those moments happen in, but do NOT depict the momentary actions, people or staged props themselves):\n- 사내들이 뜨거운 닭고기를 뜯어 허겁지겁 먹는다. 저만치 모닥불가.\n- 모닥불 주변.\n배를 채운 의병들이 규홍옆에 모여 웃고 떠들고 있다\n- 강용언이 무릎을 꿇은 채 그 모습을 본다. 복만만 남아 있다.\n\nTIME OF DAY (lock): night.\n\nRender ONE photorealistic empty location photograph, 16:9, neutral enough that every shot of this place can be staged from it later. No readable writing anywhere: surfaces that would carry writing may be present, but stage any wording out of legibility — an oblique angle, distance, shallow focus. No captions, watermarks or overlay text.",
  "bg_path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/db00517b-165b-4005-a287-f28936b59b95/images/272bd7af-9294-4490-98c6-88a40e9deacc/scene/recipe/groupbg_lake_fire_circle_7fdda1.png",
  "asset_id": "56a98b90-775e-401a-9419-f0314b337734",
  "input_asset_ids": [
   "5aac126f-7b05-4f78-b00f-cfa59765f394"
  ],
  "origin_tag": "S19sh17",
  "place_text": "The dirt space beside the communal campfire in the lakeside camp, where a wooden carrying frame lies within the fighting area.",
  "origin_inputs": {
   "place_text": "The dirt space beside the communal campfire in the lakeside camp, where a wooden carrying frame lies within the fighting area.",
   "time_of_day_en": "night",
   "conti_asset_id": "5aac126f-7b05-4f78-b00f-cfa59765f394"
  }
 },
 "S19sh17::bgfirst_bg": {
  "input_fingerprint": "2d08b1940a19924d",
  "prompt": "Create the EMPTY BACKGROUND PLATE for one film shot — NO PEOPLE, no figures, no body parts, no sketch lines, no arrows anywhere.\n\"Empty\" means no people only: KEEP the location's inherent occupants and stock that define the place — animals in an animal shelter, pen or farm, goods in a market, moored boats in a harbour — unless the shot text explicitly removes them.\nThe FIRST attached image is a thin-line storyboard sketch: use ONLY its camera angle, horizon, perspective and the placement/size of buildings and set masses — ignore the sketched people and arrows entirely. The SECOND attached image (LOCATION PHOTOGRAPH) is the real place: take its architecture, materials, signage and fixed features, and RE-PROJECT them into the sketch's camera. If the photograph's camera differs from the sketch's, the sketch's camera wins.\nHUMAN-SCALE CALIBRATION: derive every structure's true size from human-scale elements — a door ≈ 2m, a window ≈ 1–1.5m wide, one storey ≈ 2.5–3m; never inflate a small structure or shrink a large one.\n\nSHOT TEXT this background must serve (Korean): 안규홍의 굵은 주먹이 강용언의 턱을 무자비하게 강타한 찰나.\n\nLOCATION (lock): The dirt space beside the communal campfire in the lakeside camp, where a wooden carrying frame lies within the fighting area.\n\nTIME OF DAY (lock): night.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: close-up\n- KEY BACKGROUND ELEMENTS: 임시거점 공터 (The confrontation remains within the camp clearing); used as A narrow, unfocused spatial margin keeps the bodily impact grounded in the existing location.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: The established nighttime campfire illumination gives the impact naturalistic moving contrast and restrained warmth without stylized flashes or added sources.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nRender ONE photorealistic empty location photograph, 16:9, that this shot can be staged inside later. No readable writing anywhere: surfaces that would carry writing may be present, but stage any wording out of legibility — an oblique angle, distance, shallow focus. No captions, watermarks or overlay text.",
  "effective_prompt": "Create the EMPTY BACKGROUND PLATE for one film shot — NO PEOPLE, no figures, no body parts, no sketch lines, no arrows anywhere.\n\"Empty\" means no people only: KEEP the location's inherent occupants and stock that define the place — animals in an animal shelter, pen or farm, goods in a market, moored boats in a harbour — unless the shot text explicitly removes them.\nThe FIRST attached image is a thin-line storyboard sketch: use ONLY its camera angle, horizon, perspective and the placement/size of buildings and set masses — ignore the sketched people and arrows entirely. The SECOND attached image (LOCATION PHOTOGRAPH) is the real place: take its architecture, materials, signage and fixed features, and RE-PROJECT them into the sketch's camera. If the photograph's camera differs from the sketch's, the sketch's camera wins.\nHUMAN-SCALE CALIBRATION: derive every structure's true size from human-scale elements — a door ≈ 2m, a window ≈ 1–1.5m wide, one storey ≈ 2.5–3m; never inflate a small structure or shrink a large one.\n\nSHOT TEXT this background must serve (Korean): 안규홍의 굵은 주먹이 강용언의 턱을 무자비하게 강타한 찰나.\n\nLOCATION (lock): The dirt space beside the communal campfire in the lakeside camp, where a wooden carrying frame lies within the fighting area.\n\nTIME OF DAY (lock): night.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: close-up\n- KEY BACKGROUND ELEMENTS: 임시거점 공터 (The confrontation remains within the camp clearing); used as A narrow, unfocused spatial margin keeps the bodily impact grounded in the existing location.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: The established nighttime campfire illumination gives the impact naturalistic moving contrast and restrained warmth without stylized flashes or added sources.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nRender ONE photorealistic empty location photograph, 16:9, that this shot can be staged inside later. No readable writing anywhere: surfaces that would carry writing may be present, but stage any wording out of legibility — an oblique angle, distance, shallow focus. No captions, watermarks or overlay text.",
  "bg_path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/db00517b-165b-4005-a287-f28936b59b95/images/272bd7af-9294-4490-98c6-88a40e9deacc/scene/recipe/S19sh17__bgfirst_bg.png",
  "asset_id": "90a91ee0-fd9f-4f7f-a4e0-edf4084124f1",
  "input_asset_ids": [
   "5aac126f-7b05-4f78-b00f-cfa59765f394",
   "56a98b90-775e-401a-9419-f0314b337734"
  ]
 },
 "S19sh17": {
  "input_fingerprint": "bb72a0f5fe66ea54",
  "prompt": "Create ONE FINAL photorealistic live-action film still of the moment below — early 20th-century Korea, Jeolla region; all people are Korean unless explicitly stated otherwise. TIME OF DAY (lock): night.\n\nSHOT TEXT (authoritative, Korean): 안규홍의 굵은 주먹이 강용언의 턱을 무자비하게 강타한 찰나.\n\nLOCATION (lock): The dirt space beside the communal campfire in the lakeside camp, where a wooden carrying frame lies within the fighting area. The shot takes place here — the attached LOCATION PHOTOGRAPH shows the exact spot.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: close-up\n- KEY BACKGROUND ELEMENTS: 임시거점 공터 (The confrontation remains within the camp clearing); used as A narrow, unfocused spatial margin keeps the bodily impact grounded in the existing location.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: The established nighttime campfire illumination gives the impact naturalistic moving contrast and restrained warmth without stylized flashes or added sources.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): A sword strike has damaged the carrying frame beside the campfire; the two seed-rice sacks remain unopened. 안규홍: His lip is freshly split and bleeding; the jade ring remains concealed in his clothing. 강용언: He is unsteady and red-eyed from drink, still in the armed struggle.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 안규홍 (한국인 남성, 청년 나이대, 검은 머리, 단단히 묶은 머리); 강용언 (한국인 남성, 성인 나이대, 햇볕에 그을린 얼굴, 자연스러운 검은 머리) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.",
  "roll_prompts": {
   "A": "Stage the shot. The FIRST attached image (SHOT BACKGROUND) is the finished empty background of this shot — keep it EXACTLY: its camera, perspective, architecture, lighting and every fixed feature stay untouched. The SECOND attached image (LAYOUT SKETCH) tells you ONLY where the people go: each sketched person's position, screen size, pose and the gaze/motion arrows. Ignore the sketch's background lines. The CHARACTER REFERENCE photographs show the real people.\nPlace the real people into the background at exactly the sketched positions, sizes and poses, following the arrow directions. No sketch lines or arrows may remain.\n\nCreate ONE FINAL photorealistic live-action film still of the moment below — early 20th-century Korea, Jeolla region; all people are Korean unless explicitly stated otherwise. TIME OF DAY (lock): night.\n\nSHOT TEXT (authoritative, Korean): 안규홍의 굵은 주먹이 강용언의 턱을 무자비하게 강타한 찰나.\n\nLOCATION (lock): The dirt space beside the communal campfire in the lakeside camp, where a wooden carrying frame lies within the fighting area. The shot takes place here — the FIRST attached image (SHOT BACKGROUND) is this exact place, already built: its ground, structures, horizon, materials and lighting are the finished truth of this location and must not be redesigned or replaced. No location photograph is attached — read the place from that image alone, and add no scenery, structure, vehicle or fixture that it does not already show. This lock governs the place only; the figures in the shot follow the staging and pose instructions.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: close-up\n- KEY BACKGROUND ELEMENTS: 임시거점 공터 (The confrontation remains within the camp clearing); used as A narrow, unfocused spatial margin keeps the bodily impact grounded in the existing location.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: The established nighttime campfire illumination gives the impact naturalistic moving contrast and restrained warmth without stylized flashes or added sources.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): A sword strike has damaged the carrying frame beside the campfire; the two seed-rice sacks remain unopened. 안규홍: His lip is freshly split and bleeding; the jade ring remains concealed in his clothing. 강용언: He is unsteady and red-eyed from drink, still in the armed struggle.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 안규홍 (한국인 남성, 청년 나이대, 검은 머리, 단단히 묶은 머리); 강용언 (한국인 남성, 성인 나이대, 햇볕에 그을린 얼굴, 자연스러운 검은 머리) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.",
   "B": "Create ONE FINAL photorealistic live-action film still of the moment below — early 20th-century Korea, Jeolla region; all people are Korean unless explicitly stated otherwise. TIME OF DAY (lock): night.\n\nSHOT TEXT (authoritative, Korean): 안규홍의 굵은 주먹이 강용언의 턱을 무자비하게 강타한 찰나.\n\nLOCATION (lock): The dirt space beside the communal campfire in the lakeside camp, where a wooden carrying frame lies within the fighting area. The shot takes place here — the attached LOCATION PHOTOGRAPH shows the exact spot.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: close-up\n- KEY BACKGROUND ELEMENTS: 임시거점 공터 (The confrontation remains within the camp clearing); used as A narrow, unfocused spatial margin keeps the bodily impact grounded in the existing location.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: The established nighttime campfire illumination gives the impact naturalistic moving contrast and restrained warmth without stylized flashes or added sources.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): A sword strike has damaged the carrying frame beside the campfire; the two seed-rice sacks remain unopened. 안규홍: His lip is freshly split and bleeding; the jade ring remains concealed in his clothing. 강용언: He is unsteady and red-eyed from drink, still in the armed struggle.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 안규홍 (한국인 남성, 청년 나이대, 검은 머리, 단단히 묶은 머리); 강용언 (한국인 남성, 성인 나이대, 햇볕에 그을린 얼굴, 자연스러운 검은 머리) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.\n\nCOMPOSITION VARIATION (this candidate only): keep the place, the people and the moment exactly as specified, and keep from the CAMERA & FRAME contract its subject scale, screen placement and key background elements exactly. For this candidate only, its camera angle and height are deliberately left open: choose a camera position distinctly different from the obvious one, as a film director picking a second setup on the same blocking. Stay true to any explicit framing the SHOT TEXT itself dictates."
  },
  "roll_refs": {
   "A": [
    {
     "label": "SHOT BACKGROUND",
     "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/db00517b-165b-4005-a287-f28936b59b95/images/272bd7af-9294-4490-98c6-88a40e9deacc/scene/recipe/S19sh17__bgfirst_bg.png",
     "asset_id": "90a91ee0-fd9f-4f7f-a4e0-edf4084124f1",
     "role": "bgfirst_bg"
    },
    {
     "label": "LAYOUT SKETCH (people placement only)",
     "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/db00517b-165b-4005-a287-f28936b59b95/images/272bd7af-9294-4490-98c6-88a40e9deacc/conti/conti_S19sh17.png",
     "asset_id": "5aac126f-7b05-4f78-b00f-cfa59765f394",
     "role": "conti_light"
    },
    {
     "label": "CHARACTER REFERENCE — 안규홍: the exact person appearing in this shot; match face, hair and build exactly.",
     "path": "<bytes:1028384>",
     "asset_id": "3b5894db-ce55-4f34-b966-5e725c3aa952",
     "role": "character_ref"
    },
    {
     "label": "CHARACTER REFERENCE — 강용언: the exact person appearing in this shot; match face, hair and build exactly.",
     "path": "<bytes:822588>",
     "asset_id": "1a75090f-1eb9-4224-a132-de7a4b454b73",
     "role": "character_ref"
    }
   ],
   "B": [
    {
     "label": "LOCATION PHOTOGRAPH — the exact place of this shot: its architecture, materials, fixed features and lighting mood are spatial truth; stage the moment inside this place. Never copy its camera framing.",
     "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/db00517b-165b-4005-a287-f28936b59b95/images/272bd7af-9294-4490-98c6-88a40e9deacc/scene/recipe/groupbg_lake_fire_circle_7fdda1.png",
     "asset_id": "56a98b90-775e-401a-9419-f0314b337734",
     "role": "bgfirst_group_bg"
    },
    {
     "label": "CHARACTER REFERENCE — 안규홍: the exact person appearing in this shot; match face, hair and build exactly.",
     "path": "<bytes:1028384>",
     "asset_id": "3b5894db-ce55-4f34-b966-5e725c3aa952",
     "role": "character_ref"
    },
    {
     "label": "CHARACTER REFERENCE — 강용언: the exact person appearing in this shot; match face, hair and build exactly.",
     "path": "<bytes:822588>",
     "asset_id": "1a75090f-1eb9-4224-a132-de7a4b454b73",
     "role": "character_ref"
    }
   ]
  },
  "shot_run_spend_attempt_count": 1,
  "judge_flip_skipped": "judge_fn_owns_order",
  "readings": [
   {
    "label": "A",
    "direction": "안규홍의 오른쪽 주먹이 강용언의 우측 턱과 뺨 부위를 정확하게 강타하고 있으며, 시선과 타격의 방향이 일치함.",
    "built_space": "배경에 호수, 대나무 천막, 모닥불 등 야영지의 구조물이 레퍼런스와 일치하게 배치되어 있으며, 화면 우측 바닥에 볏짚 포대가 얹힌 지게가 놓여 있음.",
    "entities": "안규홍(흰옷)과 강용언(푸른 관복) 단 두 명만 화면에 등장함. 안규홍이 뒷모습으로 잡혀 터진 입술은 확인할 수 없으나, 두 인물의 복장과 헤어스타일, 강용언이 검을 쥔 모습은 지시된 바와 일치함.",
    "hard_violations": [],
    "physics": "강한 주먹 타격으로 인해 침과 땀이 허공으로 튀는 물리적 반동이 사실적이며, 두 인물 모두 지면에 두 발을 단단히 딛고 체중을 실은 채 서 있음."
   },
   {
    "label": "B",
    "direction": "안규홍의 오른쪽 주먹이 강용언의 좌측 턱에 닿아 있으며, 두 인물은 서로의 얼굴을 향해 시선을 두고 있음.",
    "built_space": "아웃포커싱된 배경 중앙에 모닥불이 켜진 공터가 있고 좌측에 나무 거치 구조물이 보이나 지게의 형태나 파손 여부는 식별하기 어려움.",
    "entities": "안규홍의 피가 나는 터진 입술과 강용언의 붉게 충혈된 눈이 명확하게 표현됨. 그러나 배경의 모닥불 근처에 프롬프트에 명시되지 않은 정체불명의 인물 두 명이 흐릿하게 등장함.",
    "hard_violations": [
     "[gemini-pro] 프롬프트에 명시되지 않은 정체불명의 추가 인물 2명이 배경에 등장함 (invented people).",
     "[gpt-high] 등장이 허용된 안규홍과 강용언 외에 배경 인물 두 명을 추가했다."
    ],
    "physics": "두 인물 모두 지면에 서 있으나, 강타하는 찰나의 역동적인 운동감이라기보다는 주먹을 턱에 댄 상태로 멈춰 서서 포즈를 취한 듯 물리적 충격의 묘사가 부족함."
   }
  ],
  "cross_model_order": {
   "policy": "cross_model_order_v2_slot0_fwd_slot1_rev_parallel2",
   "slots": [
    {
     "model": "gemini-pro",
     "order": "forward",
     "display_to_canonical": {
      "A": "A",
      "B": "B"
     },
     "raw": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 7,
        "verdict_ko": "지정된 인물 외에 다른 사람을 추가하지 않고 주먹이 강타하는 찰나의 물리적 역동성과 충격감을 훌륭하게 구현했습니다. 다만 요구된 클로즈업 샷보다 다소 넓게 프레이밍되어 안규홍의 터진 입술 묘사가 화면에 드러나지 않은 점이 아쉽습니다."
       },
       {
        "label": "B",
        "score": 3,
        "verdict_ko": "안규홍의 터진 입술과 강용언의 충혈된 눈 등 세부적인 인물 묘사를 훌륭히 구현하고 클로즈업 구도도 잘 따랐으나, 프롬프트에서 엄격히 금지한 정체불명의 추가 인물이 배경에 둘이나 등장하여 치명적인 규칙 위반이 발생했습니다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "안규홍의 오른쪽 주먹이 강용언의 우측 턱과 뺨 부위를 정확하게 강타하고 있으며, 시선과 타격의 방향이 일치함.",
        "built_space": "배경에 호수, 대나무 천막, 모닥불 등 야영지의 구조물이 레퍼런스와 일치하게 배치되어 있으며, 화면 우측 바닥에 볏짚 포대가 얹힌 지게가 놓여 있음.",
        "entities": "안규홍(흰옷)과 강용언(푸른 관복) 단 두 명만 화면에 등장함. 안규홍이 뒷모습으로 잡혀 터진 입술은 확인할 수 없으나, 두 인물의 복장과 헤어스타일, 강용언이 검을 쥔 모습은 지시된 바와 일치함.",
        "hard_violations": [],
        "physics": "강한 주먹 타격으로 인해 침과 땀이 허공으로 튀는 물리적 반동이 사실적이며, 두 인물 모두 지면에 두 발을 단단히 딛고 체중을 실은 채 서 있음."
       },
       {
        "label": "B",
        "direction": "안규홍의 오른쪽 주먹이 강용언의 좌측 턱에 닿아 있으며, 두 인물은 서로의 얼굴을 향해 시선을 두고 있음.",
        "built_space": "아웃포커싱된 배경 중앙에 모닥불이 켜진 공터가 있고 좌측에 나무 거치 구조물이 보이나 지게의 형태나 파손 여부는 식별하기 어려움.",
        "entities": "안규홍의 피가 나는 터진 입술과 강용언의 붉게 충혈된 눈이 명확하게 표현됨. 그러나 배경의 모닥불 근처에 프롬프트에 명시되지 않은 정체불명의 인물 두 명이 흐릿하게 등장함.",
        "hard_violations": [
         "프롬프트에 명시되지 않은 정체불명의 추가 인물 2명이 배경에 등장함 (invented people)."
        ],
        "physics": "두 인물 모두 지면에 서 있으나, 강타하는 찰나의 역동적인 운동감이라기보다는 주먹을 턱에 댄 상태로 멈춰 서서 포즈를 취한 듯 물리적 충격의 묘사가 부족함."
       }
      ],
      "all_candidates_fail": false
     },
     "normalized": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 7,
        "verdict_ko": "지정된 인물 외에 다른 사람을 추가하지 않고 주먹이 강타하는 찰나의 물리적 역동성과 충격감을 훌륭하게 구현했습니다. 다만 요구된 클로즈업 샷보다 다소 넓게 프레이밍되어 안규홍의 터진 입술 묘사가 화면에 드러나지 않은 점이 아쉽습니다."
       },
       {
        "label": "B",
        "score": 3,
        "verdict_ko": "안규홍의 터진 입술과 강용언의 충혈된 눈 등 세부적인 인물 묘사를 훌륭히 구현하고 클로즈업 구도도 잘 따랐으나, 프롬프트에서 엄격히 금지한 정체불명의 추가 인물이 배경에 둘이나 등장하여 치명적인 규칙 위반이 발생했습니다."
       }
      ],
      "all_candidates_fail": false,
      "readings": [
       {
        "label": "A",
        "direction": "안규홍의 오른쪽 주먹이 강용언의 우측 턱과 뺨 부위를 정확하게 강타하고 있으며, 시선과 타격의 방향이 일치함.",
        "built_space": "배경에 호수, 대나무 천막, 모닥불 등 야영지의 구조물이 레퍼런스와 일치하게 배치되어 있으며, 화면 우측 바닥에 볏짚 포대가 얹힌 지게가 놓여 있음.",
        "entities": "안규홍(흰옷)과 강용언(푸른 관복) 단 두 명만 화면에 등장함. 안규홍이 뒷모습으로 잡혀 터진 입술은 확인할 수 없으나, 두 인물의 복장과 헤어스타일, 강용언이 검을 쥔 모습은 지시된 바와 일치함.",
        "hard_violations": [],
        "physics": "강한 주먹 타격으로 인해 침과 땀이 허공으로 튀는 물리적 반동이 사실적이며, 두 인물 모두 지면에 두 발을 단단히 딛고 체중을 실은 채 서 있음."
       },
       {
        "label": "B",
        "direction": "안규홍의 오른쪽 주먹이 강용언의 좌측 턱에 닿아 있으며, 두 인물은 서로의 얼굴을 향해 시선을 두고 있음.",
        "built_space": "아웃포커싱된 배경 중앙에 모닥불이 켜진 공터가 있고 좌측에 나무 거치 구조물이 보이나 지게의 형태나 파손 여부는 식별하기 어려움.",
        "entities": "안규홍의 피가 나는 터진 입술과 강용언의 붉게 충혈된 눈이 명확하게 표현됨. 그러나 배경의 모닥불 근처에 프롬프트에 명시되지 않은 정체불명의 인물 두 명이 흐릿하게 등장함.",
        "hard_violations": [
         "프롬프트에 명시되지 않은 정체불명의 추가 인물 2명이 배경에 등장함 (invented people)."
        ],
        "physics": "두 인물 모두 지면에 서 있으나, 강타하는 찰나의 역동적인 운동감이라기보다는 주먹을 턱에 댄 상태로 멈춰 서서 포즈를 취한 듯 물리적 충격의 묘사가 부족함."
       }
      ]
     },
     "ok": true
    },
    {
     "model": "gpt-high",
     "order": "reverse",
     "display_to_canonical": {
      "A": "B",
      "B": "A"
     },
     "raw": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 2,
        "verdict_ko": "밀착 구도와 안규홍의 입술 출혈은 맞지만, 배경에 허용되지 않은 인물 두 명을 추가해 탈락하며 주먹도 턱보다 볼에 닿는다."
       },
       {
        "label": "B",
        "score": 6,
        "verdict_ko": "추가 인물 없이 야영지에서의 타격과 무장 상태를 구현했지만, 클로즈업을 허리까지 보이는 넓은 구도로 바꾸었고 타격점도 턱보다 볼에 가깝다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "오른쪽 안규홍은 왼쪽 강용언의 얼굴을 바라보며 주먹을 왼쪽으로 뻗는다. 주먹은 강용언의 입 옆 볼에 닿아 있으며 턱끝이나 아래턱을 직접 강타하는 모습은 아니다. 강용언의 시선은 오른쪽 위의 안규홍 쪽으로 향하고, 머리의 반동은 작다. 무기는 보이지 않는다.",
        "built_space": "두 사람의 얼굴과 상체가 화면을 크게 채운다. 뒤쪽 아래에 돌로 둘러싼 모닥불 한 곳, 왼쪽에 천이 걸린 목재 지지대와 낮은 목재 틀, 뒤쪽에 천막형 거처가 보인다. 참조 야영지의 재료와 밤 분위기는 이어지지만, 배경 중앙에 사람이 두 명 더 서 있다. 손상된 운반틀과 종자 벼 자루는 이 크롭에서 확인되지 않는다.",
        "entities": "전경의 두 사람은 한국인 남성이라는 설정에 부합하는 외형이다. 안규홍은 젊은 얼굴, 묶은 검은 머리, 거친 밝은색 저고리와 피가 흐르는 입술을 갖췄다. 강용언은 성인 남성의 그을린 얼굴, 검은 머리, 충혈된 눈과 붉은 테두리의 짙은 군복을 보이지만 참조의 모자는 없다. 배경의 성인 남성 두 명은 허용된 인물이 아니다. 옥가락지는 노출되지 않으며 읽을 수 있는 글자도 없다.",
        "hard_violations": [
         "등장이 허용된 안규홍과 강용언 외에 배경 인물 두 명을 추가했다."
        ],
        "physics": "주먹과 손목은 안규홍의 소매 속 팔로 이어지고, 주먹이 강용언의 볼에 실제 접촉한다. 두 사람의 몸통은 화면 아래로 이어지며 발이 잘렸을 뿐 공중에 떠 있다는 증거는 없다. 다만 강타에 따른 얼굴 압박과 목의 반동이 약해 주먹을 볼에 댄 정지 자세처럼 읽힌다. 모닥불 장작과 목재 시설은 지면에 놓여 있다."
       },
       {
        "label": "B",
        "direction": "왼쪽 안규홍은 오른쪽 강용언을 향해 팔을 뻗고 얼굴도 상대 쪽으로 돌리고 있다. 주먹은 강용언의 볼과 턱선 위쪽에 닿으며 머리는 화면 오른쪽으로 기울어진다. 정확한 턱 강타보다는 옆볼 타격에 가깝다. 강용언의 반쯤 감긴 눈은 왼쪽 상대 쪽을 향한다. 손에 든 칼날은 오른쪽 아래로 뻗어 안규홍을 직접 겨누지는 않지만, 문구가 칼의 특정 조준점을 요구하지는 않는다.",
        "built_space": "두 사람을 허리 부근까지 보여 주는 넓은 구도여서, 요구된 클로즈업과 좁고 흐린 배경 여백을 벗어난다. 뒤에는 모닥불 한 곳, 호수와 산, 천막형 거처 열, 왼쪽 목재 지지대가 보인다. 오른쪽 땅에는 운반틀 한 개와 포대가 놓여 있다. 장소의 배치와 재료는 참조와 대체로 맞으며 추가 인물이나 불가능한 반사는 없다.",
        "entities": "두 인물만 보인다. 안규홍의 묶은 검은 머리와 밝은색 낡은 옷은 참조에 부합하지만, 얼굴이 돌아가 있어 얼굴 동일성과 입술 출혈은 판정하기 어렵다. 강용언은 성인 한국인 남성 설정에 부합하는 외형과 짙은 군복, 붉은 허리띠, 칼을 갖췄다. 다만 참조의 장식 달린 모자 대신 검은 머리띠와 상투를 보이며, 충혈은 뚜렷하지 않다. 운반틀 위 포대는 닫힌 상태로 보이나 정확히 두 자루인지와 칼에 의한 틀 손상은 명확하지 않다. 옥가락지와 읽을 수 있는 글자는 보이지 않는다.",
        "hard_violations": [],
        "physics": "안규홍의 어깨에서 팔과 주먹이 자연스럽게 이어져 강용언의 얼굴에 접촉하고, 강용언의 머리와 상체가 타격 방향으로 기운다. 얼굴 주변의 작은 비말은 충격으로 튄 것으로 해석 가능하다. 강용언의 한 손이 칼자루를 잡고 있어 칼의 지지가 확인된다. 하체 일부는 프레임 밖이지만 몸이 떠 있는 모습은 아니며, 운반틀과 포대는 땅에 놓여 있다."
       }
      ],
      "all_candidates_fail": true
     },
     "normalized": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 2,
        "verdict_ko": "밀착 구도와 안규홍의 입술 출혈은 맞지만, 배경에 허용되지 않은 인물 두 명을 추가해 탈락하며 주먹도 턱보다 볼에 닿는다."
       },
       {
        "label": "A",
        "score": 6,
        "verdict_ko": "추가 인물 없이 야영지에서의 타격과 무장 상태를 구현했지만, 클로즈업을 허리까지 보이는 넓은 구도로 바꾸었고 타격점도 턱보다 볼에 가깝다."
       }
      ],
      "all_candidates_fail": true,
      "readings": [
       {
        "label": "B",
        "direction": "오른쪽 안규홍은 왼쪽 강용언의 얼굴을 바라보며 주먹을 왼쪽으로 뻗는다. 주먹은 강용언의 입 옆 볼에 닿아 있으며 턱끝이나 아래턱을 직접 강타하는 모습은 아니다. 강용언의 시선은 오른쪽 위의 안규홍 쪽으로 향하고, 머리의 반동은 작다. 무기는 보이지 않는다.",
        "built_space": "두 사람의 얼굴과 상체가 화면을 크게 채운다. 뒤쪽 아래에 돌로 둘러싼 모닥불 한 곳, 왼쪽에 천이 걸린 목재 지지대와 낮은 목재 틀, 뒤쪽에 천막형 거처가 보인다. 참조 야영지의 재료와 밤 분위기는 이어지지만, 배경 중앙에 사람이 두 명 더 서 있다. 손상된 운반틀과 종자 벼 자루는 이 크롭에서 확인되지 않는다.",
        "entities": "전경의 두 사람은 한국인 남성이라는 설정에 부합하는 외형이다. 안규홍은 젊은 얼굴, 묶은 검은 머리, 거친 밝은색 저고리와 피가 흐르는 입술을 갖췄다. 강용언은 성인 남성의 그을린 얼굴, 검은 머리, 충혈된 눈과 붉은 테두리의 짙은 군복을 보이지만 참조의 모자는 없다. 배경의 성인 남성 두 명은 허용된 인물이 아니다. 옥가락지는 노출되지 않으며 읽을 수 있는 글자도 없다.",
        "hard_violations": [
         "등장이 허용된 안규홍과 강용언 외에 배경 인물 두 명을 추가했다."
        ],
        "physics": "주먹과 손목은 안규홍의 소매 속 팔로 이어지고, 주먹이 강용언의 볼에 실제 접촉한다. 두 사람의 몸통은 화면 아래로 이어지며 발이 잘렸을 뿐 공중에 떠 있다는 증거는 없다. 다만 강타에 따른 얼굴 압박과 목의 반동이 약해 주먹을 볼에 댄 정지 자세처럼 읽힌다. 모닥불 장작과 목재 시설은 지면에 놓여 있다."
       },
       {
        "label": "A",
        "direction": "왼쪽 안규홍은 오른쪽 강용언을 향해 팔을 뻗고 얼굴도 상대 쪽으로 돌리고 있다. 주먹은 강용언의 볼과 턱선 위쪽에 닿으며 머리는 화면 오른쪽으로 기울어진다. 정확한 턱 강타보다는 옆볼 타격에 가깝다. 강용언의 반쯤 감긴 눈은 왼쪽 상대 쪽을 향한다. 손에 든 칼날은 오른쪽 아래로 뻗어 안규홍을 직접 겨누지는 않지만, 문구가 칼의 특정 조준점을 요구하지는 않는다.",
        "built_space": "두 사람을 허리 부근까지 보여 주는 넓은 구도여서, 요구된 클로즈업과 좁고 흐린 배경 여백을 벗어난다. 뒤에는 모닥불 한 곳, 호수와 산, 천막형 거처 열, 왼쪽 목재 지지대가 보인다. 오른쪽 땅에는 운반틀 한 개와 포대가 놓여 있다. 장소의 배치와 재료는 참조와 대체로 맞으며 추가 인물이나 불가능한 반사는 없다.",
        "entities": "두 인물만 보인다. 안규홍의 묶은 검은 머리와 밝은색 낡은 옷은 참조에 부합하지만, 얼굴이 돌아가 있어 얼굴 동일성과 입술 출혈은 판정하기 어렵다. 강용언은 성인 한국인 남성 설정에 부합하는 외형과 짙은 군복, 붉은 허리띠, 칼을 갖췄다. 다만 참조의 장식 달린 모자 대신 검은 머리띠와 상투를 보이며, 충혈은 뚜렷하지 않다. 운반틀 위 포대는 닫힌 상태로 보이나 정확히 두 자루인지와 칼에 의한 틀 손상은 명확하지 않다. 옥가락지와 읽을 수 있는 글자는 보이지 않는다.",
        "hard_violations": [],
        "physics": "안규홍의 어깨에서 팔과 주먹이 자연스럽게 이어져 강용언의 얼굴에 접촉하고, 강용언의 머리와 상체가 타격 방향으로 기운다. 얼굴 주변의 작은 비말은 충격으로 튄 것으로 해석 가능하다. 강용언의 한 손이 칼자루를 잡고 있어 칼의 지지가 확인된다. 하체 일부는 프레임 밖이지만 몸이 떠 있는 모습은 아니며, 운반틀과 포대는 땅에 놓여 있다."
       }
      ]
     },
     "ok": true
    }
   ],
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "slot_winner_match": true,
   "slot_winner": {
    "gemini-pro": "A",
    "gpt-high": "A"
   },
   "route": "cross_slot_agree"
  },
  "dual": {
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "normalized": {
    "A": 2.0,
    "B": 0.762
   },
   "adjusted": {
    "A": 2.0,
    "B": 0.512
   },
   "violations": {
    "B": [
     "[gemini-pro] 프롬프트에 명시되지 않은 정체불명의 추가 인물 2명이 배경에 등장함 (invented people).",
     "[gpt-high] 등장이 허용된 안규홍과 강용언 외에 배경 인물 두 명을 추가했다."
    ]
   },
   "per_model_winner": {
    "gemini-pro": "A",
    "gpt-high": "A"
   },
   "agreed": true
  },
  "totals": {
   "A": 2000,
   "B": 512
  },
  "selected": "A",
  "ranking": [
   "A",
   "B"
  ],
  "verdicts": [
   {
    "label": "A",
    "score": 2000,
    "verdict_ko": "지정된 인물 외에 다른 사람을 추가하지 않고 주먹이 강타하는 찰나의 물리적 역동성과 충격감을 훌륭하게 구현했습니다. 다만 요구된 클로즈업 샷보다 다소 넓게 프레이밍되어 안규홍의 터진 입술 묘사가 화면에 드러나지 않은 점이 아쉽습니다."
   },
   {
    "label": "B",
    "score": 512,
    "verdict_ko": "안규홍의 터진 입술과 강용언의 충혈된 눈 등 세부적인 인물 묘사를 훌륭히 구현하고 클로즈업 구도도 잘 따랐으나, 프롬프트에서 엄격히 금지한 정체불명의 추가 인물이 배경에 둘이나 등장하여 치명적인 규칙 위반이 발생했습니다.  ★위반: [gemini-pro] 프롬프트에 명시되지 않은 정체불명의 추가 인물 2명이 배경에 등장함 (invented people). / [gpt-high] 등장이 허용된 안규홍과 강용언 외에 배경 인물 두 명을 추가했다."
   }
  ],
  "refs": [
   {
    "label": "LOCATION PHOTOGRAPH — the exact place of this shot: its architecture, materials, fixed features and lighting mood are spatial truth; stage the moment inside this place. Never copy its camera framing.",
    "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/db00517b-165b-4005-a287-f28936b59b95/images/272bd7af-9294-4490-98c6-88a40e9deacc/scene/recipe/groupbg_lake_fire_circle_7fdda1.png",
    "asset_id": "56a98b90-775e-401a-9419-f0314b337734",
    "role": "bgfirst_group_bg"
   },
   {
    "label": "CHARACTER REFERENCE — 안규홍: the exact person appearing in this shot; match face, hair and build exactly.",
    "path": "<bytes:1028384>",
    "asset_id": "3b5894db-ce55-4f34-b966-5e725c3aa952",
    "role": "character_ref"
   },
   {
    "label": "CHARACTER REFERENCE — 강용언: the exact person appearing in this shot; match face, hair and build exactly.",
    "path": "<bytes:822588>",
    "asset_id": "1a75090f-1eb9-4224-a132-de7a4b454b73",
    "role": "character_ref"
   }
  ],
  "critique_skipped": true,
  "shot_run_produced": false,
  "shot_run_uid": "06aa1412-584b-7dec-b73a-1eab3e4e9fe7",
  "bgfirst": {
   "bg_path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/db00517b-165b-4005-a287-f28936b59b95/images/272bd7af-9294-4490-98c6-88a40e9deacc/scene/recipe/S19sh17__bgfirst_bg.png",
   "bg_asset_id": "90a91ee0-fd9f-4f7f-a4e0-edf4084124f1",
   "bg_record_key": "S19sh17::bgfirst_bg",
   "chain_winner": true,
   "authority": "groupbg",
   "group_key": "lake_fire_circle",
   "groupbg_asset_id": "56a98b90-775e-401a-9419-f0314b337734"
  },
  "ref_mode": "재투영 배경+콘티+엔티티 (2택1: 체인 승)",
  "share_plan": {
   "ref_plan": "background"
  }
 },
 "S19sh17::cine": {
  "applied": true,
  "attempted_at": "2026-09-09T03:01:55.392866+00:00",
  "fingerprint": "ed12e6f2d94bd4f7a22e70e1e92b581c1938e21847789e36b76974d58f49e9dc",
  "fingerprint_version": 2,
  "provider": "mai",
  "endpoint": "openrouter/chat-completions",
  "model": "microsoft/mai-image-2.6",
  "pack": "24.202608252115",
  "source_file": "S19sh17_sel.png",
  "source_sha256": "e6bb38489e981e3abdb38d0fda697fa2daa0719615a1446c835f998509049548",
  "file": "S19sh17_cine.png",
  "staged_sha256": "a778ccbbb0a03e1c6118d6022bdb739a36d90a9d894f78d1d31f592da3a1d656",
  "latency_ms": 32636
 },
 "S19sh25::signage": {
  "fp": "355eb66c615e0f7b",
  "inscriptions": [],
  "cues": [],
  "dropped": []
 },
 "S19sh25": {
  "input_fingerprint": "1c0418f21df9a77e",
  "prompt": "Create ONE FINAL photorealistic live-action film still of the moment below — early 20th-century Korea, Jeolla region; all people are Korean unless explicitly stated otherwise. TIME OF DAY (lock): night.\n\nSHOT TEXT (authoritative, Korean): 거리가 벌어진 안규홍의 등 뒤로 일제히 한쪽 발을 내딛은 자세인 조선인 의병들의 무리.\n\nLOCATION (lock): The outdoor departure route away from the lakeside camp's communal fire, between the gathering place and the camp edge. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: wide shot\n- FRAME LAYOUT: Outgoing route continuing beyond the upper-right edge in the upper-right of the frame, background.\n- KEY BACKGROUND ELEMENTS: 공터의 흙바닥 (Crossed by the departing group); used as Keeps advancing feet and the growing intervals between men readable; 안규홍의 봇짐 (Lifted and carried on his shoulder) — Seen from behind beside his shoulder without obscuring his receding outline; used as A small narrative cue that this is a departure rather than a temporary retreat.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Retain the campfire-established nighttime contrast and muted palette without adding illumination along the departure route.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The clothing, hair and overall look of 안규홍 — who appear both in that photo and in this shot — are LOCKED to that photo. Anyone else visible in that photo is NOT in this shot: never carry their face, body or clothing onto anyone here. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The dropped sword lies on the dirt beside the sword-damaged carrying frame and campfire; the seed-rice sacks remain tied. 안규홍: He leaves carrying his travel bundle, with a split lip and the jade ring concealed in his clothing.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 안규홍 (한국인 남성, 청년 나이대, 검은 머리, 단단히 묶은 머리) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.",
  "roll_prompts": {
   "A": "Create ONE FINAL photorealistic live-action film still of the moment below — early 20th-century Korea, Jeolla region; all people are Korean unless explicitly stated otherwise. TIME OF DAY (lock): night.\n\nSHOT TEXT (authoritative, Korean): 거리가 벌어진 안규홍의 등 뒤로 일제히 한쪽 발을 내딛은 자세인 조선인 의병들의 무리.\n\nLOCATION (lock): The outdoor departure route away from the lakeside camp's communal fire, between the gathering place and the camp edge. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: wide shot\n- FRAME LAYOUT: Outgoing route continuing beyond the upper-right edge in the upper-right of the frame, background.\n- KEY BACKGROUND ELEMENTS: 공터의 흙바닥 (Crossed by the departing group); used as Keeps advancing feet and the growing intervals between men readable; 안규홍의 봇짐 (Lifted and carried on his shoulder) — Seen from behind beside his shoulder without obscuring his receding outline; used as A small narrative cue that this is a departure rather than a temporary retreat.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Retain the campfire-established nighttime contrast and muted palette without adding illumination along the departure route.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The clothing, hair and overall look of 안규홍 — who appear both in that photo and in this shot — are LOCKED to that photo. Anyone else visible in that photo is NOT in this shot: never carry their face, body or clothing onto anyone here. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The dropped sword lies on the dirt beside the sword-damaged carrying frame and campfire; the seed-rice sacks remain tied. 안규홍: He leaves carrying his travel bundle, with a split lip and the jade ring concealed in his clothing.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 안규홍 (한국인 남성, 청년 나이대, 검은 머리, 단단히 묶은 머리) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.",
   "B": "Create ONE FINAL photorealistic live-action film still of the moment below — early 20th-century Korea, Jeolla region; all people are Korean unless explicitly stated otherwise. TIME OF DAY (lock): night.\n\nSHOT TEXT (authoritative, Korean): 거리가 벌어진 안규홍의 등 뒤로 일제히 한쪽 발을 내딛은 자세인 조선인 의병들의 무리.\n\nLOCATION (lock): The outdoor departure route away from the lakeside camp's communal fire, between the gathering place and the camp edge. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: wide shot\n- FRAME LAYOUT: Outgoing route continuing beyond the upper-right edge in the upper-right of the frame, background.\n- KEY BACKGROUND ELEMENTS: 공터의 흙바닥 (Crossed by the departing group); used as Keeps advancing feet and the growing intervals between men readable; 안규홍의 봇짐 (Lifted and carried on his shoulder) — Seen from behind beside his shoulder without obscuring his receding outline; used as A small narrative cue that this is a departure rather than a temporary retreat.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Retain the campfire-established nighttime contrast and muted palette without adding illumination along the departure route.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The clothing, hair and overall look of 안규홍 — who appear both in that photo and in this shot — are LOCKED to that photo. Anyone else visible in that photo is NOT in this shot: never carry their face, body or clothing onto anyone here. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The dropped sword lies on the dirt beside the sword-damaged carrying frame and campfire; the seed-rice sacks remain tied. 안규홍: He leaves carrying his travel bundle, with a split lip and the jade ring concealed in his clothing.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 안규홍 (한국인 남성, 청년 나이대, 검은 머리, 단단히 묶은 머리) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.\n\nCOMPOSITION VARIATION (this candidate only): keep the place, the people and the moment exactly as specified, and keep from the CAMERA & FRAME contract its subject scale, screen placement and key background elements exactly. For this candidate only, its camera angle and height are deliberately left open: choose a camera position distinctly different from the obvious one, as a film director picking a second setup on the same blocking. Stay true to any explicit framing the SHOT TEXT itself dictates."
  },
  "shot_run_spend_attempt_count": 1,
  "readings": [
   {
    "label": "A",
    "direction": "인물들 모두 카메라를 등지고 화면 우측 상단의 배경 방향으로 이동하고 있음.",
    "built_space": "호숫가 야영지 형태가 유지됨. 좌측 중경에 모닥불, 우측 배경에 묶인 포대들이 위치하며 전경에 떨어진 칼과 부서진 지게가 있음.",
    "entities": "화면 중앙에 봇짐을 멘 안규홍이 있으며, 주위에 비슷한 복장의 의병 무리가 함께 뒷모습으로 보임.",
    "hard_violations": [
     "[gpt-high] 왼쪽의 눕혀진 운반틀 외에 오른쪽 전경에 별도의 세워진 운반틀을 추가하여, 유지해야 할 단일 손상 운반틀을 중복시켰다."
    ],
    "physics": "땅을 딛고 걷는 정상적인 체중 이동과 자세가 나타남."
   },
   {
    "label": "B",
    "direction": "전경의 안규홍이 카메라를 등진 채 서 있고, 의병 무리는 안규홍(및 카메라)을 마주보는 방향으로 늘어서 있음.",
    "built_space": "호숫가 배경 및 모닥불, 포대 등 기존 구조물들이 배치되어 있음.",
    "entities": "봇짐을 든 안규홍과 다수의 의병들. 그러나 이전 샷에 등장했던 파란 옷의 인물이 무리 중앙에 섞여 있음.",
    "hard_violations": [
     "[gemini-pro] 이전 샷에 등장한 인물(파란 옷) 재등장 금지 지시 위반",
     "[gemini-pro] 물리적으로 불가능하고 기이한 인체 포즈 (다수의 인물이 체중 지지나 모멘텀 없이 다리를 허공에 뻗고 있음)",
     "[gpt-high] 안규홍 뒤에서 출발해야 할 의병들을 그의 앞에 마주 보는 대열로 배치했다.",
     "[gpt-high] 이전 장면의 다른 인물은 재등장시키지 말라는 지시에도 청색 도포와 붉은 허리띠의 남성을 다시 등장시켰다."
    ],
    "physics": "다수의 인물이 한쪽 다리를 부자연스럽게 앞으로 들고 있는데, 근육의 긴장이나 체중 이동 없이 다리가 허공에 떠 있어 물리적 지지 및 자세 묘사가 실패함."
   }
  ],
  "cross_model_order": {
   "policy": "cross_model_order_v2_slot0_fwd_slot1_rev_parallel2",
   "slots": [
    {
     "model": "gemini-pro",
     "order": "forward",
     "display_to_canonical": {
      "A": "A",
      "B": "B"
     },
     "raw": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 8,
        "verdict_ko": "요구된 배경 요소(떨어진 칼, 부서진 지게, 모닥불)와 봇짐을 멘 안규홍의 뒷모습을 안정적으로 구현했으나, 무리가 일제히 발을 내딛는 역동성은 다소 평범한 걷기로 표현되었습니다."
       },
       {
        "label": "B",
        "score": 2,
        "verdict_ko": "이전 샷의 파란 옷 인물을 재등장시켜 등장 금지 지시를 위반했으며, 의병들이 체중 지지 없이 비정상적으로 다리를 들고 있는 물리적 오류가 발생했습니다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "인물들 모두 카메라를 등지고 화면 우측 상단의 배경 방향으로 이동하고 있음.",
        "built_space": "호숫가 야영지 형태가 유지됨. 좌측 중경에 모닥불, 우측 배경에 묶인 포대들이 위치하며 전경에 떨어진 칼과 부서진 지게가 있음.",
        "entities": "화면 중앙에 봇짐을 멘 안규홍이 있으며, 주위에 비슷한 복장의 의병 무리가 함께 뒷모습으로 보임.",
        "hard_violations": [],
        "physics": "땅을 딛고 걷는 정상적인 체중 이동과 자세가 나타남."
       },
       {
        "label": "B",
        "direction": "전경의 안규홍이 카메라를 등진 채 서 있고, 의병 무리는 안규홍(및 카메라)을 마주보는 방향으로 늘어서 있음.",
        "built_space": "호숫가 배경 및 모닥불, 포대 등 기존 구조물들이 배치되어 있음.",
        "entities": "봇짐을 든 안규홍과 다수의 의병들. 그러나 이전 샷에 등장했던 파란 옷의 인물이 무리 중앙에 섞여 있음.",
        "hard_violations": [
         "이전 샷에 등장한 인물(파란 옷) 재등장 금지 지시 위반",
         "물리적으로 불가능하고 기이한 인체 포즈 (다수의 인물이 체중 지지나 모멘텀 없이 다리를 허공에 뻗고 있음)"
        ],
        "physics": "다수의 인물이 한쪽 다리를 부자연스럽게 앞으로 들고 있는데, 근육의 긴장이나 체중 이동 없이 다리가 허공에 떠 있어 물리적 지지 및 자세 묘사가 실패함."
       }
      ],
      "all_candidates_fail": false
     },
     "normalized": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 8,
        "verdict_ko": "요구된 배경 요소(떨어진 칼, 부서진 지게, 모닥불)와 봇짐을 멘 안규홍의 뒷모습을 안정적으로 구현했으나, 무리가 일제히 발을 내딛는 역동성은 다소 평범한 걷기로 표현되었습니다."
       },
       {
        "label": "B",
        "score": 2,
        "verdict_ko": "이전 샷의 파란 옷 인물을 재등장시켜 등장 금지 지시를 위반했으며, 의병들이 체중 지지 없이 비정상적으로 다리를 들고 있는 물리적 오류가 발생했습니다."
       }
      ],
      "all_candidates_fail": false,
      "readings": [
       {
        "label": "A",
        "direction": "인물들 모두 카메라를 등지고 화면 우측 상단의 배경 방향으로 이동하고 있음.",
        "built_space": "호숫가 야영지 형태가 유지됨. 좌측 중경에 모닥불, 우측 배경에 묶인 포대들이 위치하며 전경에 떨어진 칼과 부서진 지게가 있음.",
        "entities": "화면 중앙에 봇짐을 멘 안규홍이 있으며, 주위에 비슷한 복장의 의병 무리가 함께 뒷모습으로 보임.",
        "hard_violations": [],
        "physics": "땅을 딛고 걷는 정상적인 체중 이동과 자세가 나타남."
       },
       {
        "label": "B",
        "direction": "전경의 안규홍이 카메라를 등진 채 서 있고, 의병 무리는 안규홍(및 카메라)을 마주보는 방향으로 늘어서 있음.",
        "built_space": "호숫가 배경 및 모닥불, 포대 등 기존 구조물들이 배치되어 있음.",
        "entities": "봇짐을 든 안규홍과 다수의 의병들. 그러나 이전 샷에 등장했던 파란 옷의 인물이 무리 중앙에 섞여 있음.",
        "hard_violations": [
         "이전 샷에 등장한 인물(파란 옷) 재등장 금지 지시 위반",
         "물리적으로 불가능하고 기이한 인체 포즈 (다수의 인물이 체중 지지나 모멘텀 없이 다리를 허공에 뻗고 있음)"
        ],
        "physics": "다수의 인물이 한쪽 다리를 부자연스럽게 앞으로 들고 있는데, 근육의 긴장이나 체중 이동 없이 다리가 허공에 떠 있어 물리적 지지 및 자세 묘사가 실패함."
       }
      ]
     },
     "ok": true
    },
    {
     "model": "gpt-high",
     "order": "reverse",
     "display_to_canonical": {
      "A": "B",
      "B": "A"
     },
     "raw": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 2,
        "verdict_ko": "의병들이 떠나는 안규홍의 뒤를 따르지 않고 정면에서 마주하며, 이전 장면의 제외 대상인 청색 도포 남성까지 재등장한다."
       },
       {
        "label": "B",
        "score": 5,
        "verdict_ko": "안규홍과 간격을 두고 뒤따르는 의병들의 야간 와이드 숏은 맞지만, 운반틀이 추가되고 봇짐도 어깨 옆이 아닌 등 중앙에 놓였다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "안규홍은 화면 중앙 오른쪽에서 등을 보인 채 배경의 의병들을 향한다. 약 14명의 의병은 카메라 쪽 또는 안규홍을 바라보며 그와 마주 서 있다. 따라서 안규홍의 진행 방향 뒤에서 함께 출발하는 관계가 아니라 서로 접근하는 방향이다. 오른쪽 위 바깥으로 이어지는 출발 동선도 명확하지 않다.",
        "built_space": "왼쪽에 호수와 돌 해안, 뒤쪽과 오른쪽에 목재·천막 숙영시설, 중앙 배경에 모닥불 하나와 처마의 등불 하나가 보인다. 흙 공터와 야간 재질은 참조와 대체로 연결된다. 그러나 의병 무리가 안규홍의 앞을 가로막는 위치에 배치되어 지정된 출발 대열과 맞지 않는다. 떨어진 칼과 손상된 운반틀은 식별되지 않는다.",
        "entities": "안규홍의 묶은 검은 머리, 때 묻은 담색 옷과 허리띠, 짚신은 참조와 부합하며 얼굴과 입술 상처는 뒷모습이라 확인할 수 없다. 오른쪽 어깨 옆에는 천 봇짐이 있다. 의병들은 성인 한국인 남성으로 설정된 시대극 인물들로 보인다. 다만 중앙의 검은 망건·청색 도포·붉은 허리띠 남성은 이전 장면에서 안규홍에게 맞던 인물의 외형과 의상을 그대로 재사용했다. 읽을 수 있는 글자는 없다.",
        "hard_violations": [
         "안규홍 뒤에서 출발해야 할 의병들을 그의 앞에 마주 보는 대열로 배치했다.",
         "이전 장면의 다른 인물은 재등장시키지 말라는 지시에도 청색 도포와 붉은 허리띠의 남성을 다시 등장시켰다."
        ],
        "physics": "의병들은 한 발을 땅에 둔 채 다른 발을 앞으로 들고 있어 체중 지지는 성립한다. 동작이 다소 발차기나 제식처럼 보이지만 공중에 뜬 몸은 없다. 안규홍의 봇짐은 올라간 오른손이 쥔 매듭과 어깨로 지지되며, 몸 역시 지면에 닿은 발로 지탱된다."
       },
       {
        "label": "B",
        "direction": "안규홍은 중앙에서 등을 보이며 멀어지고, 좌우의 의병 여섯 명도 그의 진행 방향을 바라보며 뒤따른다. 선두와 후속 인물 사이의 간격이 읽힌다. 이동축은 주로 화면 위쪽 중앙으로 향해, 오른쪽 위 가장자리로 빠지는 경로라는 지정은 다소 약하다. 바닥의 칼은 오른쪽 아래로 향해 놓여 있으며 누구를 겨누는 상태가 아니다.",
        "built_space": "왼쪽의 호수와 돌 해안, 오른쪽의 목재·천막 숙영시설, 왼쪽 중경의 모닥불 하나, 숙영시설의 등불 하나가 참조의 공간과 조명을 이어간다. 안규홍은 흙길 중앙 앞쪽, 의병 여섯 명은 그보다 카메라 가까운 좌우에 있다. 왼쪽 아래에는 눕혀진 사각 목제 운반틀이 보이는데, 별도로 오른쪽 아래 전경에 세워진 삼각형 운반틀도 있어 지정된 단일 손상 운반틀 외에 하나가 추가된다. 칼은 이 전경 운반틀 옆 바닥에 있다.",
        "entities": "안규홍은 참조에 맞는 묶은 검은 머리와 담색 저고리·바지, 허리띠와 짚신을 착용한다. 얼굴이 가려져 정확한 얼굴 동일성과 입술 상처는 확인할 수 없고, 옥가락지는 드러나지 않는다. 봇짐은 존재하지만 어깨 옆에 들어 메기보다 등 중앙에 매단 형태다. 의병 여섯 명은 성인 한국인 남성으로 설정된 인물들로 보이며, 이전 장면의 청색 도포 남성은 보이지 않는다. 오른쪽에는 묶인 자루들이 있고 바닥에는 칼이 있다. 읽을 수 있는 글자는 없다.",
        "hard_violations": [
         "왼쪽의 눕혀진 운반틀 외에 오른쪽 전경에 별도의 세워진 운반틀을 추가하여, 유지해야 할 단일 손상 운반틀을 중복시켰다."
        ],
        "physics": "안규홍과 의병들은 지면에 닿은 발로 체중을 지탱하며 일부는 뒤꿈치나 반대 발을 들어 보행 중간 자세를 보인다. 모든 사람이 똑같은 순간에 발을 내딛었다는 동시성은 약하지만 불가능한 자세는 아니다. 봇짐은 목 아래의 매듭과 어깨를 지나는 천으로 등에 지지된다. 칼과 자루, 두 운반틀은 지면이나 아래 물체에 닿아 있어 지지 없이 떠 있는 물체는 없다."
       }
      ],
      "all_candidates_fail": true
     },
     "normalized": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 2,
        "verdict_ko": "의병들이 떠나는 안규홍의 뒤를 따르지 않고 정면에서 마주하며, 이전 장면의 제외 대상인 청색 도포 남성까지 재등장한다."
       },
       {
        "label": "A",
        "score": 5,
        "verdict_ko": "안규홍과 간격을 두고 뒤따르는 의병들의 야간 와이드 숏은 맞지만, 운반틀이 추가되고 봇짐도 어깨 옆이 아닌 등 중앙에 놓였다."
       }
      ],
      "all_candidates_fail": true,
      "readings": [
       {
        "label": "B",
        "direction": "안규홍은 화면 중앙 오른쪽에서 등을 보인 채 배경의 의병들을 향한다. 약 14명의 의병은 카메라 쪽 또는 안규홍을 바라보며 그와 마주 서 있다. 따라서 안규홍의 진행 방향 뒤에서 함께 출발하는 관계가 아니라 서로 접근하는 방향이다. 오른쪽 위 바깥으로 이어지는 출발 동선도 명확하지 않다.",
        "built_space": "왼쪽에 호수와 돌 해안, 뒤쪽과 오른쪽에 목재·천막 숙영시설, 중앙 배경에 모닥불 하나와 처마의 등불 하나가 보인다. 흙 공터와 야간 재질은 참조와 대체로 연결된다. 그러나 의병 무리가 안규홍의 앞을 가로막는 위치에 배치되어 지정된 출발 대열과 맞지 않는다. 떨어진 칼과 손상된 운반틀은 식별되지 않는다.",
        "entities": "안규홍의 묶은 검은 머리, 때 묻은 담색 옷과 허리띠, 짚신은 참조와 부합하며 얼굴과 입술 상처는 뒷모습이라 확인할 수 없다. 오른쪽 어깨 옆에는 천 봇짐이 있다. 의병들은 성인 한국인 남성으로 설정된 시대극 인물들로 보인다. 다만 중앙의 검은 망건·청색 도포·붉은 허리띠 남성은 이전 장면에서 안규홍에게 맞던 인물의 외형과 의상을 그대로 재사용했다. 읽을 수 있는 글자는 없다.",
        "hard_violations": [
         "안규홍 뒤에서 출발해야 할 의병들을 그의 앞에 마주 보는 대열로 배치했다.",
         "이전 장면의 다른 인물은 재등장시키지 말라는 지시에도 청색 도포와 붉은 허리띠의 남성을 다시 등장시켰다."
        ],
        "physics": "의병들은 한 발을 땅에 둔 채 다른 발을 앞으로 들고 있어 체중 지지는 성립한다. 동작이 다소 발차기나 제식처럼 보이지만 공중에 뜬 몸은 없다. 안규홍의 봇짐은 올라간 오른손이 쥔 매듭과 어깨로 지지되며, 몸 역시 지면에 닿은 발로 지탱된다."
       },
       {
        "label": "A",
        "direction": "안규홍은 중앙에서 등을 보이며 멀어지고, 좌우의 의병 여섯 명도 그의 진행 방향을 바라보며 뒤따른다. 선두와 후속 인물 사이의 간격이 읽힌다. 이동축은 주로 화면 위쪽 중앙으로 향해, 오른쪽 위 가장자리로 빠지는 경로라는 지정은 다소 약하다. 바닥의 칼은 오른쪽 아래로 향해 놓여 있으며 누구를 겨누는 상태가 아니다.",
        "built_space": "왼쪽의 호수와 돌 해안, 오른쪽의 목재·천막 숙영시설, 왼쪽 중경의 모닥불 하나, 숙영시설의 등불 하나가 참조의 공간과 조명을 이어간다. 안규홍은 흙길 중앙 앞쪽, 의병 여섯 명은 그보다 카메라 가까운 좌우에 있다. 왼쪽 아래에는 눕혀진 사각 목제 운반틀이 보이는데, 별도로 오른쪽 아래 전경에 세워진 삼각형 운반틀도 있어 지정된 단일 손상 운반틀 외에 하나가 추가된다. 칼은 이 전경 운반틀 옆 바닥에 있다.",
        "entities": "안규홍은 참조에 맞는 묶은 검은 머리와 담색 저고리·바지, 허리띠와 짚신을 착용한다. 얼굴이 가려져 정확한 얼굴 동일성과 입술 상처는 확인할 수 없고, 옥가락지는 드러나지 않는다. 봇짐은 존재하지만 어깨 옆에 들어 메기보다 등 중앙에 매단 형태다. 의병 여섯 명은 성인 한국인 남성으로 설정된 인물들로 보이며, 이전 장면의 청색 도포 남성은 보이지 않는다. 오른쪽에는 묶인 자루들이 있고 바닥에는 칼이 있다. 읽을 수 있는 글자는 없다.",
        "hard_violations": [
         "왼쪽의 눕혀진 운반틀 외에 오른쪽 전경에 별도의 세워진 운반틀을 추가하여, 유지해야 할 단일 손상 운반틀을 중복시켰다."
        ],
        "physics": "안규홍과 의병들은 지면에 닿은 발로 체중을 지탱하며 일부는 뒤꿈치나 반대 발을 들어 보행 중간 자세를 보인다. 모든 사람이 똑같은 순간에 발을 내딛었다는 동시성은 약하지만 불가능한 자세는 아니다. 봇짐은 목 아래의 매듭과 어깨를 지나는 천으로 등에 지지된다. 칼과 자루, 두 운반틀은 지면이나 아래 물체에 닿아 있어 지지 없이 떠 있는 물체는 없다."
       }
      ]
     },
     "ok": true
    }
   ],
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "slot_winner_match": true,
   "slot_winner": {
    "gemini-pro": "A",
    "gpt-high": "A"
   },
   "route": "cross_slot_agree"
  },
  "dual": {
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "normalized": {
    "A": 2.0,
    "B": 0.65
   },
   "adjusted": {
    "A": 1.75,
    "B": 0.4
   },
   "violations": {
    "B": [
     "[gemini-pro] 이전 샷에 등장한 인물(파란 옷) 재등장 금지 지시 위반",
     "[gemini-pro] 물리적으로 불가능하고 기이한 인체 포즈 (다수의 인물이 체중 지지나 모멘텀 없이 다리를 허공에 뻗고 있음)",
     "[gpt-high] 안규홍 뒤에서 출발해야 할 의병들을 그의 앞에 마주 보는 대열로 배치했다.",
     "[gpt-high] 이전 장면의 다른 인물은 재등장시키지 말라는 지시에도 청색 도포와 붉은 허리띠의 남성을 다시 등장시켰다."
    ],
    "A": [
     "[gpt-high] 왼쪽의 눕혀진 운반틀 외에 오른쪽 전경에 별도의 세워진 운반틀을 추가하여, 유지해야 할 단일 손상 운반틀을 중복시켰다."
    ]
   },
   "per_model_winner": {
    "gemini-pro": "A",
    "gpt-high": "A"
   },
   "agreed": true
  },
  "totals": {
   "A": 1750,
   "B": 400
  },
  "selected": "A",
  "ranking": [
   "A",
   "B"
  ],
  "verdicts": [
   {
    "label": "A",
    "score": 1750,
    "verdict_ko": "요구된 배경 요소(떨어진 칼, 부서진 지게, 모닥불)와 봇짐을 멘 안규홍의 뒷모습을 안정적으로 구현했으나, 무리가 일제히 발을 내딛는 역동성은 다소 평범한 걷기로 표현되었습니다.  ★위반: [gpt-high] 왼쪽의 눕혀진 운반틀 외에 오른쪽 전경에 별도의 세워진 운반틀을 추가하여, 유지해야 할 단일 손상 운반틀을 중복시켰다."
   },
   {
    "label": "B",
    "score": 400,
    "verdict_ko": "이전 샷의 파란 옷 인물을 재등장시켜 등장 금지 지시를 위반했으며, 의병들이 체중 지지 없이 비정상적으로 다리를 들고 있는 물리적 오류가 발생했습니다.  ★위반: [gemini-pro] 이전 샷에 등장한 인물(파란 옷) 재등장 금지 지시 위반 / [gemini-pro] 물리적으로 불가능하고 기이한 인체 포즈 (다수의 인물이 체중 지지나 모멘텀 없이 다리를 허공에 뻗고 있음) / [gpt-high] 안규홍 뒤에서 출발해야 할 의병들을 그의 앞에 마주 보는 대열로 배치했다. / [gpt-high] 이전 장면의 다른 인물은 재등장시키지 말라는 지시에도 청색 도포와 붉은 허리띠의 남성을 다시 등장시켰다."
   }
  ],
  "refs": [
   {
    "label": "PREVIOUS SHOT STILL — a visually related earlier shot of this same place: the location's look, materials, fixed features and lighting mood are LOCKED to this photo; never copy its camera framing. The clothing, hair and overall look of 안규홍 — who appear both in that photo and in this shot — are LOCKED to that photo. Anyone else visible in that photo is NOT in this shot: never carry their face, body or clothing onto anyone here. If this photo shows a character who cannot move (dead or unconscious), that character's exact pose, position and orientation are ALSO LOCKED — treat their body as a fixed prop of the set that only the camera moves around.",
    "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/db00517b-165b-4005-a287-f28936b59b95/images/272bd7af-9294-4490-98c6-88a40e9deacc/scene/recipe/S19sh17_sel.png",
    "asset_id": "242d6a68-f9c0-4c42-b31a-bf04272e7146",
    "role": "prev_still"
   },
   {
    "label": "CHARACTER REFERENCE — 안규홍: the exact person appearing in this shot; match face, hair and build exactly.",
    "path": "<bytes:1028384>",
    "asset_id": "3b5894db-ce55-4f34-b966-5e725c3aa952",
    "role": "character_ref"
   }
  ],
  "critique_skipped": true,
  "shot_run_produced": false,
  "shot_run_uid": "06aa1412-5c79-7d0a-8eff-2e7296ef3b8b",
  "ref_mode": "prev+엔티티",
  "share_plan": {
   "ref_plan": "prev",
   "prev_anchor_tag": "S19sh17"
  }
 },
 "S19sh25::cine": {
  "applied": true,
  "attempted_at": "2026-09-09T03:03:42.117218+00:00",
  "fingerprint": "5524f5f4b162fadcb16c8aa390c70917890875fc7f64b6f198fda965fe215583",
  "fingerprint_version": 2,
  "provider": "mai",
  "endpoint": "openrouter/chat-completions",
  "model": "microsoft/mai-image-2.6",
  "pack": "24.202608252115",
  "source_file": "S19sh25_sel.png",
  "source_sha256": "cf254dec5c1904c5192896a0db75514dcd7aa081c2574666e751857512d5160e",
  "file": "S19sh25_cine.png",
  "staged_sha256": "42e07ba2cf4694f0a9c7ebdab6070b336b9290905d04d2f864e664f3feefc299",
  "latency_ms": 29449
 },
 "S19sh31::signage": {
  "fp": "df4ae0aba52182f9",
  "inscriptions": [],
  "cues": [],
  "dropped": []
 },
 "S19sh31": {
  "input_fingerprint": "83432e22253abc87",
  "prompt": "Create ONE FINAL photorealistic live-action film still of the moment below — early 20th-century Korea, Jeolla region; all people are Korean unless explicitly stated otherwise. TIME OF DAY (lock): night.\n\nSHOT TEXT (authoritative, Korean): 텅 빈 모닥불 가에 홀로 무릎 꿇은 강용언의 고립된 전신.\n\nLOCATION (lock): The deserted dirt space beside the lakeside camp's fire at night. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: wide shot\n- KEY BACKGROUND ELEMENTS: 모닥불 (Burning beside the now solitary 강용언); used as A modest spatial counterweight to his kneeling figure, not an oversized foreground feature; 비워진 공터 (The other men have departed); used as Negative space surrounding the entire kneeling figure establishes the loss of his following; 강용언의 환도 (Lying on the dirt where it fell, unclaimed) — Seen obliquely along the ground near the former struggle; used as A small secondary trace of the confrontation within the wider abandonment.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: The campfire remains the established nighttime illumination, preserving sober contrast around 강용언 without introducing a new spotlight or changing the scene's palette.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The people visible in that photo are NOT in this shot: never carry their faces, bodies or clothing onto anyone here. Each person in THIS shot is defined solely by the PEOPLE list, their own CHARACTER REFERENCE images and their wardrobe notes. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The sword remains on the dirt beside the damaged carrying frame; campfire light falls over the abandoned meal area and unopened seed-rice sacks. 강용언: He remains kneeling alone, red-eyed and unsteady, without his sword in hand.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 강용언 (한국인 남성, 성인 나이대, 햇볕에 그을린 얼굴, 자연스러운 검은 머리) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.",
  "roll_prompts": {
   "A": "Create ONE FINAL photorealistic live-action film still of the moment below — early 20th-century Korea, Jeolla region; all people are Korean unless explicitly stated otherwise. TIME OF DAY (lock): night.\n\nSHOT TEXT (authoritative, Korean): 텅 빈 모닥불 가에 홀로 무릎 꿇은 강용언의 고립된 전신.\n\nLOCATION (lock): The deserted dirt space beside the lakeside camp's fire at night. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: wide shot\n- KEY BACKGROUND ELEMENTS: 모닥불 (Burning beside the now solitary 강용언); used as A modest spatial counterweight to his kneeling figure, not an oversized foreground feature; 비워진 공터 (The other men have departed); used as Negative space surrounding the entire kneeling figure establishes the loss of his following; 강용언의 환도 (Lying on the dirt where it fell, unclaimed) — Seen obliquely along the ground near the former struggle; used as A small secondary trace of the confrontation within the wider abandonment.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: The campfire remains the established nighttime illumination, preserving sober contrast around 강용언 without introducing a new spotlight or changing the scene's palette.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The people visible in that photo are NOT in this shot: never carry their faces, bodies or clothing onto anyone here. Each person in THIS shot is defined solely by the PEOPLE list, their own CHARACTER REFERENCE images and their wardrobe notes. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The sword remains on the dirt beside the damaged carrying frame; campfire light falls over the abandoned meal area and unopened seed-rice sacks. 강용언: He remains kneeling alone, red-eyed and unsteady, without his sword in hand.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 강용언 (한국인 남성, 성인 나이대, 햇볕에 그을린 얼굴, 자연스러운 검은 머리) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.",
   "B": "Create ONE FINAL photorealistic live-action film still of the moment below — early 20th-century Korea, Jeolla region; all people are Korean unless explicitly stated otherwise. TIME OF DAY (lock): night.\n\nSHOT TEXT (authoritative, Korean): 텅 빈 모닥불 가에 홀로 무릎 꿇은 강용언의 고립된 전신.\n\nLOCATION (lock): The deserted dirt space beside the lakeside camp's fire at night. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: wide shot\n- KEY BACKGROUND ELEMENTS: 모닥불 (Burning beside the now solitary 강용언); used as A modest spatial counterweight to his kneeling figure, not an oversized foreground feature; 비워진 공터 (The other men have departed); used as Negative space surrounding the entire kneeling figure establishes the loss of his following; 강용언의 환도 (Lying on the dirt where it fell, unclaimed) — Seen obliquely along the ground near the former struggle; used as A small secondary trace of the confrontation within the wider abandonment.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: The campfire remains the established nighttime illumination, preserving sober contrast around 강용언 without introducing a new spotlight or changing the scene's palette.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The people visible in that photo are NOT in this shot: never carry their faces, bodies or clothing onto anyone here. Each person in THIS shot is defined solely by the PEOPLE list, their own CHARACTER REFERENCE images and their wardrobe notes. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The sword remains on the dirt beside the damaged carrying frame; campfire light falls over the abandoned meal area and unopened seed-rice sacks. 강용언: He remains kneeling alone, red-eyed and unsteady, without his sword in hand.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 강용언 (한국인 남성, 성인 나이대, 햇볕에 그을린 얼굴, 자연스러운 검은 머리) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.\n\nCOMPOSITION VARIATION (this candidate only): keep the place, the people and the moment exactly as specified, and keep from the CAMERA & FRAME contract its subject scale, screen placement and key background elements exactly. For this candidate only, its camera angle and height are deliberately left open: choose a camera position distinctly different from the obvious one, as a film director picking a second setup on the same blocking. Stay true to any explicit framing the SHOT TEXT itself dictates."
  },
  "shot_run_spend_attempt_count": 1,
  "readings": [
   {
    "label": "A",
    "direction": "강용언은 오른쪽을 향해 무릎을 꿇고 있으며, 시선은 모닥불을 향하고 있습니다.",
    "built_space": "기준 이미지와 비교해 야영지의 공간 배치가 완전히 달라졌습니다. 오른쪽의 오두막과 짐들이 아주 멀리 밀려났고, 해안선 모양이 바뀌었으며, 지게의 위치도 오른쪽에서 왼쪽 전경으로 이동했습니다.",
    "entities": "강용언은 기준 이미지의 복장을 입고 있으나 삿갓을 쓰고 있어 지시된 머리카락이 가려져 있습니다. 환도는 전경 오른쪽에 놓여 있으나 지게와 떨어져 있습니다. 캐릭터 뒤에 지시되지 않은 정체불명의 삼각대 구조물이 생성되었습니다.",
    "hard_violations": [
     "[gemini-pro] 캐릭터 뒤에 붉은 장식이 달린 정체불명의 삼각대 구조물이 창조됨 (발명된 사물).",
     "[gemini-pro] 이전 샷에서 하나만 보이던 지게가 위치와 형태를 달리하여 나타남 (복제 및 변형된 사물)."
    ],
    "physics": "인물이 바닥에 안정적으로 무릎을 꿇고 있으며, 사물들도 땅 위에 자연스럽게 놓여 있습니다."
   },
   {
    "label": "B",
    "direction": "강용언은 오른쪽을 향해 무릎을 꿇고 고개를 숙인 채, 바닥에 놓인 검들을 내려다보고 있습니다.",
    "built_space": "이전 샷의 장소 연속성을 완벽하게 유지하고 있습니다. 모닥불, 오른쪽의 오두막들, 포대 자루와 통, 오른쪽 전경의 지게 등 모든 고정 구조물과 사물이 기준 이미지와 정확히 같은 위치에 배치되어 있습니다.",
    "entities": "강용언은 기준 의상을 입고 상투를 튼 검은 머리를 보여줍니다. 모닥불과 빈 공터가 잘 묘사되었습니다. 그러나 강용언의 환도가 여러 개로 복제되어 허리에 1개, 앞 바닥에 2개, 오른쪽 지게 옆에 1개가 존재합니다.",
    "hard_violations": [
     "[gemini-pro] 캐릭터의 환도가 4개로 복제됨 (허리에 1개, 캐릭터 앞 바닥에 2개, 오른쪽 지게 옆에 1개)."
    ],
    "physics": "인물이 바닥에 무겁게 주저앉은 자세를 안정적으로 지탱하고 있으며, 모든 사물이 땅에 닿아 있습니다."
   }
  ],
  "cross_model_order": {
   "policy": "cross_model_order_v2_slot0_fwd_slot1_rev_parallel2",
   "slots": [
    {
     "model": "gemini-pro",
     "order": "forward",
     "display_to_canonical": {
      "A": "A",
      "B": "B"
     },
     "raw": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 5,
        "verdict_ko": "이전 샷의 장소 연속성(오두막, 지게의 위치 등)을 완벽하게 유지하고 인물의 고립된 분위기를 잘 살렸으나, 환도가 4개로 복제되는 치명적인 오류가 발생했습니다."
       },
       {
        "label": "A",
        "score": 3,
        "verdict_ko": "인물과 모닥불의 배치는 무난하나, 이전 샷의 공간 배치를 완전히 왜곡하고 정체불명의 구조물을 창조하는 등 록(lock) 지시사항을 크게 위반했습니다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "강용언은 오른쪽을 향해 무릎을 꿇고 있으며, 시선은 모닥불을 향하고 있습니다.",
        "built_space": "기준 이미지와 비교해 야영지의 공간 배치가 완전히 달라졌습니다. 오른쪽의 오두막과 짐들이 아주 멀리 밀려났고, 해안선 모양이 바뀌었으며, 지게의 위치도 오른쪽에서 왼쪽 전경으로 이동했습니다.",
        "entities": "강용언은 기준 이미지의 복장을 입고 있으나 삿갓을 쓰고 있어 지시된 머리카락이 가려져 있습니다. 환도는 전경 오른쪽에 놓여 있으나 지게와 떨어져 있습니다. 캐릭터 뒤에 지시되지 않은 정체불명의 삼각대 구조물이 생성되었습니다.",
        "hard_violations": [
         "캐릭터 뒤에 붉은 장식이 달린 정체불명의 삼각대 구조물이 창조됨 (발명된 사물).",
         "이전 샷에서 하나만 보이던 지게가 위치와 형태를 달리하여 나타남 (복제 및 변형된 사물)."
        ],
        "physics": "인물이 바닥에 안정적으로 무릎을 꿇고 있으며, 사물들도 땅 위에 자연스럽게 놓여 있습니다."
       },
       {
        "label": "B",
        "direction": "강용언은 오른쪽을 향해 무릎을 꿇고 고개를 숙인 채, 바닥에 놓인 검들을 내려다보고 있습니다.",
        "built_space": "이전 샷의 장소 연속성을 완벽하게 유지하고 있습니다. 모닥불, 오른쪽의 오두막들, 포대 자루와 통, 오른쪽 전경의 지게 등 모든 고정 구조물과 사물이 기준 이미지와 정확히 같은 위치에 배치되어 있습니다.",
        "entities": "강용언은 기준 의상을 입고 상투를 튼 검은 머리를 보여줍니다. 모닥불과 빈 공터가 잘 묘사되었습니다. 그러나 강용언의 환도가 여러 개로 복제되어 허리에 1개, 앞 바닥에 2개, 오른쪽 지게 옆에 1개가 존재합니다.",
        "hard_violations": [
         "캐릭터의 환도가 4개로 복제됨 (허리에 1개, 캐릭터 앞 바닥에 2개, 오른쪽 지게 옆에 1개)."
        ],
        "physics": "인물이 바닥에 무겁게 주저앉은 자세를 안정적으로 지탱하고 있으며, 모든 사물이 땅에 닿아 있습니다."
       }
      ],
      "all_candidates_fail": true
     },
     "normalized": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 5,
        "verdict_ko": "이전 샷의 장소 연속성(오두막, 지게의 위치 등)을 완벽하게 유지하고 인물의 고립된 분위기를 잘 살렸으나, 환도가 4개로 복제되는 치명적인 오류가 발생했습니다."
       },
       {
        "label": "A",
        "score": 3,
        "verdict_ko": "인물과 모닥불의 배치는 무난하나, 이전 샷의 공간 배치를 완전히 왜곡하고 정체불명의 구조물을 창조하는 등 록(lock) 지시사항을 크게 위반했습니다."
       }
      ],
      "all_candidates_fail": true,
      "readings": [
       {
        "label": "A",
        "direction": "강용언은 오른쪽을 향해 무릎을 꿇고 있으며, 시선은 모닥불을 향하고 있습니다.",
        "built_space": "기준 이미지와 비교해 야영지의 공간 배치가 완전히 달라졌습니다. 오른쪽의 오두막과 짐들이 아주 멀리 밀려났고, 해안선 모양이 바뀌었으며, 지게의 위치도 오른쪽에서 왼쪽 전경으로 이동했습니다.",
        "entities": "강용언은 기준 이미지의 복장을 입고 있으나 삿갓을 쓰고 있어 지시된 머리카락이 가려져 있습니다. 환도는 전경 오른쪽에 놓여 있으나 지게와 떨어져 있습니다. 캐릭터 뒤에 지시되지 않은 정체불명의 삼각대 구조물이 생성되었습니다.",
        "hard_violations": [
         "캐릭터 뒤에 붉은 장식이 달린 정체불명의 삼각대 구조물이 창조됨 (발명된 사물).",
         "이전 샷에서 하나만 보이던 지게가 위치와 형태를 달리하여 나타남 (복제 및 변형된 사물)."
        ],
        "physics": "인물이 바닥에 안정적으로 무릎을 꿇고 있으며, 사물들도 땅 위에 자연스럽게 놓여 있습니다."
       },
       {
        "label": "B",
        "direction": "강용언은 오른쪽을 향해 무릎을 꿇고 고개를 숙인 채, 바닥에 놓인 검들을 내려다보고 있습니다.",
        "built_space": "이전 샷의 장소 연속성을 완벽하게 유지하고 있습니다. 모닥불, 오른쪽의 오두막들, 포대 자루와 통, 오른쪽 전경의 지게 등 모든 고정 구조물과 사물이 기준 이미지와 정확히 같은 위치에 배치되어 있습니다.",
        "entities": "강용언은 기준 의상을 입고 상투를 튼 검은 머리를 보여줍니다. 모닥불과 빈 공터가 잘 묘사되었습니다. 그러나 강용언의 환도가 여러 개로 복제되어 허리에 1개, 앞 바닥에 2개, 오른쪽 지게 옆에 1개가 존재합니다.",
        "hard_violations": [
         "캐릭터의 환도가 4개로 복제됨 (허리에 1개, 캐릭터 앞 바닥에 2개, 오른쪽 지게 옆에 1개)."
        ],
        "physics": "인물이 바닥에 무겁게 주저앉은 자세를 안정적으로 지탱하고 있으며, 모든 사물이 땅에 닿아 있습니다."
       }
      ]
     },
     "ok": true
    },
    {
     "model": "gpt-high",
     "order": "reverse",
     "display_to_canonical": {
      "A": "B",
      "B": "A"
     },
     "raw": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 7,
        "verdict_ko": "빈 공터 속 작은 무릎 꿇은 전신과 야간 장소의 연속성은 충실하지만, 기준의 모자가 사라지고 허리 뒤 칼집 같은 물체가 남아 인물·휴대 상태의 일치도가 떨어진다."
       },
       {
        "label": "B",
        "score": 8,
        "verdict_ko": "고립된 전신을 담은 와이드 구도와 적절한 크기의 모닥불을 유지하면서 기준의 모자·복장을 더 충실히 재현하지만, 불안정한 자세와 지게 바로 곁에 떨어진 환도의 관계는 약하다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "강용언은 고개를 숙이고 무릎 앞 흙바닥을 내려다본다. 바닥의 환도는 화면 왼쪽 위에서 오른쪽 아래로 비스듬히 놓였으며 누구를 겨누지 않고, 두 손도 칼을 잡지 않는다.",
        "built_space": "왼쪽에는 호수와 돌 많은 물가, 오른쪽에는 목재 기둥과 거적 지붕의 연속된 야영막이 보인다. 모닥불은 한 곳이고, 강용언은 그 오른쪽 공터에 홀로 무릎 꿇었으며 지게 한 개가 전경에 있다. 오른쪽 막사 앞에는 통들과 묶인 포대들이 모여 있어 이전 장소의 배치를 잘 이어가지만, 환도는 전경 지게보다 인물 쪽에 가까워졌다.",
        "entities": "보이는 사람은 검은 상투를 한 성인 동아시아계 남성 한 명으로, 청회색 옷과 붉은 허리띠는 기준과 대체로 맞는다. 다만 기준의 깃털 달린 모자가 없고 얼굴의 정확한 동일성과 충혈은 이 크기에서 확인하기 어렵다. 모닥불, 흙 공터, 지게, 바닥의 환도와 칼집 모양 물체, 닫힌 포대가 보이며 허리 뒤에도 긴 칼집 같은 물체가 남아 있다. 식사 흔적과 지게의 파손 상태는 뚜렷하지 않고 읽을 수 있는 글자는 없다.",
        "hard_violations": [],
        "physics": "양 무릎과 접힌 다리가 흙바닥에 닿아 몸을 지탱하고, 앞으로 기운 상체와 허벅지에 댄 손이 자연스럽게 연결된다. 환도와 지게는 지면에 놓였고 포대와 통도 바닥이나 아래 물건의 지지를 받는다. 허리 뒤 긴 물체는 허리 부근에서 이어져 내려가므로 무지지 부유물로 보이지 않는다."
       },
       {
        "label": "B",
        "direction": "강용언은 화면 오른쪽의 모닥불 쪽으로 얼굴을 돌린 채 시선을 약간 아래에 둔다. 환도는 오른쪽 전경의 흙 위에서 칼끝이 왼쪽 아래를 향하도록 놓여 있으며, 강용언의 두 손은 허벅지 위에 있어 칼을 들지 않는다.",
        "built_space": "호수와 돌 많은 물가가 배경을 이루고, 오른쪽 가장자리에 목재·거적 야영막 일부와 통들이 보인다. 모닥불은 한 곳이며 인물 오른쪽에 놓여 전신보다 작게 보인다. 전경 왼쪽의 세워진 지게 한 개와 그 뒤의 넘어진 운반틀 한 개가 보이고, 묶인 포대들이 그 주변에 놓였다. 카메라가 호수 쪽을 더 담는 구도는 허용되지만 환도와 전경 지게 사이의 거리는 이전 장면보다 벌어져 있다.",
        "entities": "성인 동아시아계 남성 한 명만 등장하며, 검은 머리와 기준에 가까운 얼굴 윤곽, 깃털 달린 검은 모자, 청회색 군복과 붉은 허리띠가 확인된다. 작은 얼굴에서 충혈까지 판별하기는 어렵다. 바닥의 환도, 모닥불, 빈 흙 공터, 운반틀과 닫힌 포대는 보이지만 식사 흔적과 지게의 구체적인 파손은 명확하지 않다. 읽을 수 있는 글자는 없다.",
        "hard_violations": [],
        "physics": "접힌 다리와 무릎이 지면에 닿고 엉덩이가 뒤꿈치 쪽에 내려앉아 정상적으로 체중을 지탱한다. 손은 허벅지에 놓였으며, 환도는 흙과 작은 돌 위에 걸쳐 있고 지게와 포대도 지면의 지지를 받는다. 불가능한 부유나 해부학적 이상은 없지만 곧게 세운 상체는 요구된 비틀거림보다 안정된 정지 자세로 보인다."
       }
      ],
      "all_candidates_fail": false
     },
     "normalized": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 7,
        "verdict_ko": "빈 공터 속 작은 무릎 꿇은 전신과 야간 장소의 연속성은 충실하지만, 기준의 모자가 사라지고 허리 뒤 칼집 같은 물체가 남아 인물·휴대 상태의 일치도가 떨어진다."
       },
       {
        "label": "A",
        "score": 8,
        "verdict_ko": "고립된 전신을 담은 와이드 구도와 적절한 크기의 모닥불을 유지하면서 기준의 모자·복장을 더 충실히 재현하지만, 불안정한 자세와 지게 바로 곁에 떨어진 환도의 관계는 약하다."
       }
      ],
      "all_candidates_fail": false,
      "readings": [
       {
        "label": "B",
        "direction": "강용언은 고개를 숙이고 무릎 앞 흙바닥을 내려다본다. 바닥의 환도는 화면 왼쪽 위에서 오른쪽 아래로 비스듬히 놓였으며 누구를 겨누지 않고, 두 손도 칼을 잡지 않는다.",
        "built_space": "왼쪽에는 호수와 돌 많은 물가, 오른쪽에는 목재 기둥과 거적 지붕의 연속된 야영막이 보인다. 모닥불은 한 곳이고, 강용언은 그 오른쪽 공터에 홀로 무릎 꿇었으며 지게 한 개가 전경에 있다. 오른쪽 막사 앞에는 통들과 묶인 포대들이 모여 있어 이전 장소의 배치를 잘 이어가지만, 환도는 전경 지게보다 인물 쪽에 가까워졌다.",
        "entities": "보이는 사람은 검은 상투를 한 성인 동아시아계 남성 한 명으로, 청회색 옷과 붉은 허리띠는 기준과 대체로 맞는다. 다만 기준의 깃털 달린 모자가 없고 얼굴의 정확한 동일성과 충혈은 이 크기에서 확인하기 어렵다. 모닥불, 흙 공터, 지게, 바닥의 환도와 칼집 모양 물체, 닫힌 포대가 보이며 허리 뒤에도 긴 칼집 같은 물체가 남아 있다. 식사 흔적과 지게의 파손 상태는 뚜렷하지 않고 읽을 수 있는 글자는 없다.",
        "hard_violations": [],
        "physics": "양 무릎과 접힌 다리가 흙바닥에 닿아 몸을 지탱하고, 앞으로 기운 상체와 허벅지에 댄 손이 자연스럽게 연결된다. 환도와 지게는 지면에 놓였고 포대와 통도 바닥이나 아래 물건의 지지를 받는다. 허리 뒤 긴 물체는 허리 부근에서 이어져 내려가므로 무지지 부유물로 보이지 않는다."
       },
       {
        "label": "A",
        "direction": "강용언은 화면 오른쪽의 모닥불 쪽으로 얼굴을 돌린 채 시선을 약간 아래에 둔다. 환도는 오른쪽 전경의 흙 위에서 칼끝이 왼쪽 아래를 향하도록 놓여 있으며, 강용언의 두 손은 허벅지 위에 있어 칼을 들지 않는다.",
        "built_space": "호수와 돌 많은 물가가 배경을 이루고, 오른쪽 가장자리에 목재·거적 야영막 일부와 통들이 보인다. 모닥불은 한 곳이며 인물 오른쪽에 놓여 전신보다 작게 보인다. 전경 왼쪽의 세워진 지게 한 개와 그 뒤의 넘어진 운반틀 한 개가 보이고, 묶인 포대들이 그 주변에 놓였다. 카메라가 호수 쪽을 더 담는 구도는 허용되지만 환도와 전경 지게 사이의 거리는 이전 장면보다 벌어져 있다.",
        "entities": "성인 동아시아계 남성 한 명만 등장하며, 검은 머리와 기준에 가까운 얼굴 윤곽, 깃털 달린 검은 모자, 청회색 군복과 붉은 허리띠가 확인된다. 작은 얼굴에서 충혈까지 판별하기는 어렵다. 바닥의 환도, 모닥불, 빈 흙 공터, 운반틀과 닫힌 포대는 보이지만 식사 흔적과 지게의 구체적인 파손은 명확하지 않다. 읽을 수 있는 글자는 없다.",
        "hard_violations": [],
        "physics": "접힌 다리와 무릎이 지면에 닿고 엉덩이가 뒤꿈치 쪽에 내려앉아 정상적으로 체중을 지탱한다. 손은 허벅지에 놓였으며, 환도는 흙과 작은 돌 위에 걸쳐 있고 지게와 포대도 지면의 지지를 받는다. 불가능한 부유나 해부학적 이상은 없지만 곧게 세운 상체는 요구된 비틀거림보다 안정된 정지 자세로 보인다."
       }
      ]
     },
     "ok": true
    }
   ],
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "slot_winner_match": false,
   "slot_winner": {
    "gemini-pro": "B",
    "gpt-high": "A"
   },
   "route": "cross_slot_combined"
  },
  "dual": {
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "normalized": {
    "A": 1.6,
    "B": 1.875
   },
   "adjusted": {
    "A": 1.35,
    "B": 1.625
   },
   "violations": {
    "A": [
     "[gemini-pro] 캐릭터 뒤에 붉은 장식이 달린 정체불명의 삼각대 구조물이 창조됨 (발명된 사물).",
     "[gemini-pro] 이전 샷에서 하나만 보이던 지게가 위치와 형태를 달리하여 나타남 (복제 및 변형된 사물)."
    ],
    "B": [
     "[gemini-pro] 캐릭터의 환도가 4개로 복제됨 (허리에 1개, 캐릭터 앞 바닥에 2개, 오른쪽 지게 옆에 1개)."
    ]
   },
   "per_model_winner": {
    "gemini-pro": "B",
    "gpt-high": "A"
   },
   "agreed": false
  },
  "totals": {
   "B": 1625,
   "A": 1350
  },
  "selected": "B",
  "ranking": [
   "B",
   "A"
  ],
  "verdicts": [
   {
    "label": "B",
    "score": 1625,
    "verdict_ko": "이전 샷의 장소 연속성(오두막, 지게의 위치 등)을 완벽하게 유지하고 인물의 고립된 분위기를 잘 살렸으나, 환도가 4개로 복제되는 치명적인 오류가 발생했습니다.  ★위반: [gemini-pro] 캐릭터의 환도가 4개로 복제됨 (허리에 1개, 캐릭터 앞 바닥에 2개, 오른쪽 지게 옆에 1개)."
   },
   {
    "label": "A",
    "score": 1350,
    "verdict_ko": "인물과 모닥불의 배치는 무난하나, 이전 샷의 공간 배치를 완전히 왜곡하고 정체불명의 구조물을 창조하는 등 록(lock) 지시사항을 크게 위반했습니다.  ★위반: [gemini-pro] 캐릭터 뒤에 붉은 장식이 달린 정체불명의 삼각대 구조물이 창조됨 (발명된 사물). / [gemini-pro] 이전 샷에서 하나만 보이던 지게가 위치와 형태를 달리하여 나타남 (복제 및 변형된 사물)."
   }
  ],
  "refs": [
   {
    "label": "PREVIOUS SHOT STILL — a visually related earlier shot of this same place: the location's look, materials, fixed features and lighting mood are LOCKED to this photo; never copy its camera framing. The people visible in that photo are NOT in this shot: never carry their faces, bodies or clothing onto anyone here. Each person in THIS shot is defined solely by the PEOPLE list, their own CHARACTER REFERENCE images and their wardrobe notes. If this photo shows a character who cannot move (dead or unconscious), that character's exact pose, position and orientation are ALSO LOCKED — treat their body as a fixed prop of the set that only the camera moves around.",
    "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/db00517b-165b-4005-a287-f28936b59b95/images/272bd7af-9294-4490-98c6-88a40e9deacc/scene/recipe/S19sh25_sel.png",
    "asset_id": "b2ae4b72-259e-42e8-a0b6-30287403ba01",
    "role": "prev_still"
   },
   {
    "label": "CHARACTER REFERENCE — 강용언: the exact person appearing in this shot; match face, hair and build exactly.",
    "path": "<bytes:822588>",
    "asset_id": "1a75090f-1eb9-4224-a132-de7a4b454b73",
    "role": "character_ref"
   }
  ],
  "critique_skipped": true,
  "shot_run_produced": false,
  "shot_run_uid": "06aa1412-5e1b-720b-913c-a507eb6f281f",
  "ref_mode": "prev+엔티티",
  "share_plan": {
   "ref_plan": "prev",
   "prev_anchor_tag": "S19sh25"
  }
 },
 "S19sh31::cine": {
  "applied": true,
  "attempted_at": "2026-09-09T03:06:02.194493+00:00",
  "fingerprint": "bb7b59189323d0c2eba8d18f6191cd612de18b9b14702bb825034e8ccfa932b7",
  "fingerprint_version": 2,
  "provider": "mai",
  "endpoint": "openrouter/chat-completions",
  "model": "microsoft/mai-image-2.6",
  "pack": "24.202608252115",
  "source_file": "S19sh31_sel.png",
  "source_sha256": "ccb726ed392e3f6a8726fa811912a531df5e31e8af3bbdfb742135c33a69f76c",
  "file": "S19sh31_cine.png",
  "staged_sha256": "a598bdf6ba7ec089561d0bd3c1c913ad62cf0e8d453ee4e74864392b694c723c",
  "latency_ms": 29389
 },
 "S20sh2::signage": {
  "fp": "11e5bb4095cda906",
  "inscriptions": [],
  "cues": [
   {
    "text_native": "",
    "source": "scene_text_implied",
    "source_quote": "붉은 압류 봉인지"
   }
  ],
  "dropped": []
 },
 "S20sh2": {
  "input_fingerprint": "900af1f4644c0f7d",
  "prompt": "Create ONE FINAL photorealistic live-action film still of the moment below — early 20th-century Korea, Jeolla region; all people are Korean unless explicitly stated otherwise. TIME OF DAY (lock): night.\n\nSHOT TEXT (authoritative, Korean): 마당 너머, 곡간 문 한가운데 붉은 압류 봉인지가 선명하게 붙어 있는 클로즈업.\n\nLOCATION (lock): The exterior face of a traditional estate's granary door across the nighttime courtyard, bearing a red confiscation seal. The shot takes place here — the attached LOCATION PHOTOGRAPH shows the exact spot.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: close-up\n- FRAME LAYOUT: Red seizure seal attached to the center of the storehouse door in the middle-center of the frame, midground.\n- KEY BACKGROUND ELEMENTS: 붉은 압류 봉인지 (Clearly attached at the center of the storehouse door) — Its content-bearing front faces the camera obliquely and identifies it as a seizure seal; no additional text is specified; used as The focal evidence of confiscation, surrounded by enough setting to establish where it is attached; 곡간 문 (Bears the seizure seal at its center) — The yard-facing side is visible at a slight oblique angle; used as Provides the physical attachment surface and contextual scale for the seal.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Use subdued nighttime ambient illumination appropriate to the setting, preserving the explicitly red seal's visibility without inventing a lamp or spotlight.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): A red seizure seal is attached to the grain-store door; the dark sarangchae has no drinking table, and the usual carrying-frame corner is empty.\n\nNO PEOPLE IN THIS SHOT: the shot text shows only the place and its state — no living person appears in frame, and no face, and no figure standing, sitting or passing through. Do not invent someone to occupy the space.\n\nOne thing this rule does NOT forbid: the hand or arm of whoever is doing what the shot text describes. If the moment is an object being held, carried, operated, read, aimed, worn or handed over, the hand doing it belongs in frame, resting on or gripping that object exactly as its real use requires. An object photographed floating unsupported, with the hand that must be holding it erased, is a worse failure than the person you were avoiding — render the hand.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.",
  "roll_prompts": {
   "A": "Create ONE FINAL photorealistic live-action film still of the moment below — early 20th-century Korea, Jeolla region; all people are Korean unless explicitly stated otherwise. TIME OF DAY (lock): night.\n\nSHOT TEXT (authoritative, Korean): 마당 너머, 곡간 문 한가운데 붉은 압류 봉인지가 선명하게 붙어 있는 클로즈업.\n\nLOCATION (lock): The exterior face of a traditional estate's granary door across the nighttime courtyard, bearing a red confiscation seal. The shot takes place here — the attached LOCATION PHOTOGRAPH shows the exact spot.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: close-up\n- FRAME LAYOUT: Red seizure seal attached to the center of the storehouse door in the middle-center of the frame, midground.\n- KEY BACKGROUND ELEMENTS: 붉은 압류 봉인지 (Clearly attached at the center of the storehouse door) — Its content-bearing front faces the camera obliquely and identifies it as a seizure seal; no additional text is specified; used as The focal evidence of confiscation, surrounded by enough setting to establish where it is attached; 곡간 문 (Bears the seizure seal at its center) — The yard-facing side is visible at a slight oblique angle; used as Provides the physical attachment surface and contextual scale for the seal.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Use subdued nighttime ambient illumination appropriate to the setting, preserving the explicitly red seal's visibility without inventing a lamp or spotlight.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): A red seizure seal is attached to the grain-store door; the dark sarangchae has no drinking table, and the usual carrying-frame corner is empty.\n\nNO PEOPLE IN THIS SHOT: the shot text shows only the place and its state — no living person appears in frame, and no face, and no figure standing, sitting or passing through. Do not invent someone to occupy the space.\n\nOne thing this rule does NOT forbid: the hand or arm of whoever is doing what the shot text describes. If the moment is an object being held, carried, operated, read, aimed, worn or handed over, the hand doing it belongs in frame, resting on or gripping that object exactly as its real use requires. An object photographed floating unsupported, with the hand that must be holding it erased, is a worse failure than the person you were avoiding — render the hand.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.",
   "B": "Create ONE FINAL photorealistic live-action film still of the moment below — early 20th-century Korea, Jeolla region; all people are Korean unless explicitly stated otherwise. TIME OF DAY (lock): night.\n\nSHOT TEXT (authoritative, Korean): 마당 너머, 곡간 문 한가운데 붉은 압류 봉인지가 선명하게 붙어 있는 클로즈업.\n\nLOCATION (lock): The exterior face of a traditional estate's granary door across the nighttime courtyard, bearing a red confiscation seal. The shot takes place here — the attached LOCATION PHOTOGRAPH shows the exact spot.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: close-up\n- FRAME LAYOUT: Red seizure seal attached to the center of the storehouse door in the middle-center of the frame, midground.\n- KEY BACKGROUND ELEMENTS: 붉은 압류 봉인지 (Clearly attached at the center of the storehouse door) — Its content-bearing front faces the camera obliquely and identifies it as a seizure seal; no additional text is specified; used as The focal evidence of confiscation, surrounded by enough setting to establish where it is attached; 곡간 문 (Bears the seizure seal at its center) — The yard-facing side is visible at a slight oblique angle; used as Provides the physical attachment surface and contextual scale for the seal.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Use subdued nighttime ambient illumination appropriate to the setting, preserving the explicitly red seal's visibility without inventing a lamp or spotlight.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): A red seizure seal is attached to the grain-store door; the dark sarangchae has no drinking table, and the usual carrying-frame corner is empty.\n\nNO PEOPLE IN THIS SHOT: the shot text shows only the place and its state — no living person appears in frame, and no face, and no figure standing, sitting or passing through. Do not invent someone to occupy the space.\n\nOne thing this rule does NOT forbid: the hand or arm of whoever is doing what the shot text describes. If the moment is an object being held, carried, operated, read, aimed, worn or handed over, the hand doing it belongs in frame, resting on or gripping that object exactly as its real use requires. An object photographed floating unsupported, with the hand that must be holding it erased, is a worse failure than the person you were avoiding — render the hand.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.\n\nCOMPOSITION VARIATION (this candidate only): keep the place, the people and the moment exactly as specified, and keep from the CAMERA & FRAME contract its subject scale, screen placement and key background elements exactly. For this candidate only, its camera angle and height are deliberately left open: choose a camera position distinctly different from the obvious one, as a film director picking a second setup on the same blocking. Stay true to any explicit framing the SHOT TEXT itself dictates."
  },
  "shot_run_spend_attempt_count": 1,
  "readings": [
   {
    "label": "A",
    "direction": "카메라가 정면에서 곡간 문을 똑바로 바라보고 있습니다.",
    "built_space": "레퍼런스 이미지와 동일하게 전체 문, 지붕, 담장, 돌구유가 넓은 화각에 정면으로 담겨 있습니다.",
    "entities": "곡간 문 중앙에 붉은 봉인지가 있으나, 문짝 하나의 너비와 맞먹을 정도로 비정상적으로 거대하게 왜곡되어 있습니다.",
    "hard_violations": [
     "[gemini-pro] 금지된 레퍼런스 카메라 구도 그대로 복사 (Copied reference framing)",
     "[gemini-pro] 사물의 물리적 스케일 왜곡 (Enlarged object physical size instead of using close-up framing)"
    ],
    "physics": "거대한 봉인지가 문 표면에 평면적으로 붙어 있습니다."
   },
   {
    "label": "B",
    "direction": "카메라가 왼쪽 아래에서 곡간 문을 향해 비스듬한 각도(oblique)로 향하고 있습니다.",
    "built_space": "곡간의 목재 문과 기둥, 뒷편의 담장이 사선 구도로 배치되어 있으며, 레퍼런스의 돌구유가 전경에 아웃포커싱되어 위치합니다.",
    "entities": "문 중앙에 현실적인 크기의 붉은 압류 봉인지가 붙어 있으며, 야간 시간대의 조명 분위기가 정확히 반영되었습니다.",
    "hard_violations": [],
    "physics": "봉인지가 문 틈을 가로질러 표면에 단단히 붙어 있습니다."
   }
  ],
  "cross_model_order": {
   "policy": "cross_model_order_v2_slot0_fwd_slot1_rev_parallel2",
   "slots": [
    {
     "model": "gemini-pro",
     "order": "forward",
     "display_to_canonical": {
      "A": "A",
      "B": "B"
     },
     "raw": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 8,
        "verdict_ko": "지시된 비스듬한 앵글(oblique angle)을 잘 적용했으며, 사물의 사실적인 크기를 유지하면서 야간의 영화적인 분위기를 훌륭하게 연출했습니다."
       },
       {
        "label": "A",
        "score": 2,
        "verdict_ko": "레퍼런스의 프레이밍을 그대로 복사하지 말라는 지시를 어겼으며, 카메라를 피사체에 다가가게 하는 대신 봉인지 자체의 크기를 비현실적으로 거대하게 키워 물리적 스케일 규정을 심각하게 위반했습니다."
       }
      ],
      "readings": [
       {
        "label": "B",
        "direction": "카메라가 왼쪽 아래에서 곡간 문을 향해 비스듬한 각도(oblique)로 향하고 있습니다.",
        "built_space": "곡간의 목재 문과 기둥, 뒷편의 담장이 사선 구도로 배치되어 있으며, 레퍼런스의 돌구유가 전경에 아웃포커싱되어 위치합니다.",
        "entities": "문 중앙에 현실적인 크기의 붉은 압류 봉인지가 붙어 있으며, 야간 시간대의 조명 분위기가 정확히 반영되었습니다.",
        "hard_violations": [],
        "physics": "봉인지가 문 틈을 가로질러 표면에 단단히 붙어 있습니다."
       },
       {
        "label": "A",
        "direction": "카메라가 정면에서 곡간 문을 똑바로 바라보고 있습니다.",
        "built_space": "레퍼런스 이미지와 동일하게 전체 문, 지붕, 담장, 돌구유가 넓은 화각에 정면으로 담겨 있습니다.",
        "entities": "곡간 문 중앙에 붉은 봉인지가 있으나, 문짝 하나의 너비와 맞먹을 정도로 비정상적으로 거대하게 왜곡되어 있습니다.",
        "hard_violations": [
         "금지된 레퍼런스 카메라 구도 그대로 복사 (Copied reference framing)",
         "사물의 물리적 스케일 왜곡 (Enlarged object physical size instead of using close-up framing)"
        ],
        "physics": "거대한 봉인지가 문 표면에 평면적으로 붙어 있습니다."
       }
      ],
      "all_candidates_fail": false
     },
     "normalized": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 8,
        "verdict_ko": "지시된 비스듬한 앵글(oblique angle)을 잘 적용했으며, 사물의 사실적인 크기를 유지하면서 야간의 영화적인 분위기를 훌륭하게 연출했습니다."
       },
       {
        "label": "A",
        "score": 2,
        "verdict_ko": "레퍼런스의 프레이밍을 그대로 복사하지 말라는 지시를 어겼으며, 카메라를 피사체에 다가가게 하는 대신 봉인지 자체의 크기를 비현실적으로 거대하게 키워 물리적 스케일 규정을 심각하게 위반했습니다."
       }
      ],
      "all_candidates_fail": false,
      "readings": [
       {
        "label": "B",
        "direction": "카메라가 왼쪽 아래에서 곡간 문을 향해 비스듬한 각도(oblique)로 향하고 있습니다.",
        "built_space": "곡간의 목재 문과 기둥, 뒷편의 담장이 사선 구도로 배치되어 있으며, 레퍼런스의 돌구유가 전경에 아웃포커싱되어 위치합니다.",
        "entities": "문 중앙에 현실적인 크기의 붉은 압류 봉인지가 붙어 있으며, 야간 시간대의 조명 분위기가 정확히 반영되었습니다.",
        "hard_violations": [],
        "physics": "봉인지가 문 틈을 가로질러 표면에 단단히 붙어 있습니다."
       },
       {
        "label": "A",
        "direction": "카메라가 정면에서 곡간 문을 똑바로 바라보고 있습니다.",
        "built_space": "레퍼런스 이미지와 동일하게 전체 문, 지붕, 담장, 돌구유가 넓은 화각에 정면으로 담겨 있습니다.",
        "entities": "곡간 문 중앙에 붉은 봉인지가 있으나, 문짝 하나의 너비와 맞먹을 정도로 비정상적으로 거대하게 왜곡되어 있습니다.",
        "hard_violations": [
         "금지된 레퍼런스 카메라 구도 그대로 복사 (Copied reference framing)",
         "사물의 물리적 스케일 왜곡 (Enlarged object physical size instead of using close-up framing)"
        ],
        "physics": "거대한 봉인지가 문 표면에 평면적으로 붙어 있습니다."
       }
      ]
     },
     "ok": true
    },
    {
     "model": "gpt-high",
     "order": "reverse",
     "display_to_canonical": {
      "A": "B",
      "B": "A"
     },
     "raw": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 4,
        "verdict_ko": "문의 재질·보강대와 봉인지의 실제 크기, 비스듬한 관찰 방향은 더 충실하지만, 봉인지가 작고 오른쪽으로 치우쳐 중앙 클로즈업을 실현하지 못했습니다."
       },
       {
        "label": "B",
        "score": 3,
        "verdict_ko": "봉인지는 화면 중앙에 있지만, 건물 전면을 넓게 보여주면서 종이 자체를 과대하게 키워 클로즈업과 실제 크기 유지 지시를 모두 놓쳤습니다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "봉인지의 인쇄된 앞면과 마당 쪽 문 표면이 카메라를 향해 비스듬히 보입니다. 화면 전체도 크게 기울어져 있습니다. 사람의 시선이나 겨누는 물체는 없습니다.",
        "built_space": "맞닫이 목문 두 짝, 각 문짝의 가로 보강대 두 개씩 총 네 개, 양옆 기둥 두 개와 아래 문턱이 보입니다. 왼쪽에는 기와를 얹은 돌흙담과 돌확 하나가 있습니다. 참조의 주요 구조와 배치는 대체로 유지되지만, 문 대부분과 전경 돌확까지 포함해 봉인지 클로즈업보다 넓으며 봉인지는 화면 중앙 오른쪽에 놓입니다. 반사는 없습니다.",
        "entities": "사람은 없습니다. 붉은 가로형 종이 한 장이 문짝 사이를 가로질러 붙어 있고, 두 원형 인장 같은 흔적이 있어 참조 봉인지와 형태 및 상대 크기가 가깝습니다. 확실히 읽을 수 있는 문구는 식별되지 않습니다. 낡은 목재와 돌흙담, 어두운 야간 분위기는 부합합니다. 사랑채와 지게 보관 구석은 화면 밖이라 확인할 수 없습니다.",
        "hard_violations": [],
        "physics": "봉인지는 목문 표면에 밀착되어 문이 지지하고 있습니다. 문은 문틀 안에 있으며 기둥은 주춧돌 위에 놓여 있고, 돌확은 지면에 놓여 있습니다. 공중에 떠 있거나 지지 없는 물체는 보이지 않습니다."
       },
       {
        "label": "B",
        "direction": "봉인지 앞면과 문의 마당 쪽 면이 거의 정면으로 카메라를 향합니다. 요구한 약한 사선 방향은 거의 드러나지 않습니다. 사람의 시선이나 이동 방향은 없습니다.",
        "built_space": "맞닫이 목문 두 짝, 기둥 두 개, 기와지붕 하나, 양쪽 돌흙담과 왼쪽 돌확 하나가 보입니다. 문의 위쪽 가로 보강대는 종이 양옆으로 일부 보이지만 참조의 아래쪽 보강대는 재현되지 않았습니다. 지붕부터 기단까지 건물 전면을 담아 장소 소개에 가까운 넓은 구도이며 클로즈업이 아닙니다. 반사는 없습니다.",
        "entities": "사람은 없습니다. 붉은 종이 한 장이 문 중앙과 화면 중앙에 붙어 있으나, 참조의 작은 가로형 봉인지와 달리 문 폭의 절반을 훨씬 넘는 큰 종이입니다. 인쇄 흔적도 두 원형 인장 대신 큰 사각 인장 형태이며 읽을 수 있는 문구는 식별되지 않습니다. 목재·기와·돌흙담과 야간 상태는 대체로 부합합니다. 사랑채와 지게 보관 구석은 보이지 않습니다.",
        "hard_violations": [],
        "physics": "주름진 종이는 중앙 문틈을 가로질러 목문에 붙어 있으며 지지 표면이 분명합니다. 지붕은 보와 기둥이 받치고, 기둥은 주춧돌 위에 있으며 돌확도 땅에 놓여 있습니다. 봉인지 크기는 부정확하지만 물리적으로 떠 있는 상태는 아닙니다."
       }
      ],
      "all_candidates_fail": true
     },
     "normalized": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 4,
        "verdict_ko": "문의 재질·보강대와 봉인지의 실제 크기, 비스듬한 관찰 방향은 더 충실하지만, 봉인지가 작고 오른쪽으로 치우쳐 중앙 클로즈업을 실현하지 못했습니다."
       },
       {
        "label": "A",
        "score": 3,
        "verdict_ko": "봉인지는 화면 중앙에 있지만, 건물 전면을 넓게 보여주면서 종이 자체를 과대하게 키워 클로즈업과 실제 크기 유지 지시를 모두 놓쳤습니다."
       }
      ],
      "all_candidates_fail": true,
      "readings": [
       {
        "label": "B",
        "direction": "봉인지의 인쇄된 앞면과 마당 쪽 문 표면이 카메라를 향해 비스듬히 보입니다. 화면 전체도 크게 기울어져 있습니다. 사람의 시선이나 겨누는 물체는 없습니다.",
        "built_space": "맞닫이 목문 두 짝, 각 문짝의 가로 보강대 두 개씩 총 네 개, 양옆 기둥 두 개와 아래 문턱이 보입니다. 왼쪽에는 기와를 얹은 돌흙담과 돌확 하나가 있습니다. 참조의 주요 구조와 배치는 대체로 유지되지만, 문 대부분과 전경 돌확까지 포함해 봉인지 클로즈업보다 넓으며 봉인지는 화면 중앙 오른쪽에 놓입니다. 반사는 없습니다.",
        "entities": "사람은 없습니다. 붉은 가로형 종이 한 장이 문짝 사이를 가로질러 붙어 있고, 두 원형 인장 같은 흔적이 있어 참조 봉인지와 형태 및 상대 크기가 가깝습니다. 확실히 읽을 수 있는 문구는 식별되지 않습니다. 낡은 목재와 돌흙담, 어두운 야간 분위기는 부합합니다. 사랑채와 지게 보관 구석은 화면 밖이라 확인할 수 없습니다.",
        "hard_violations": [],
        "physics": "봉인지는 목문 표면에 밀착되어 문이 지지하고 있습니다. 문은 문틀 안에 있으며 기둥은 주춧돌 위에 놓여 있고, 돌확은 지면에 놓여 있습니다. 공중에 떠 있거나 지지 없는 물체는 보이지 않습니다."
       },
       {
        "label": "A",
        "direction": "봉인지 앞면과 문의 마당 쪽 면이 거의 정면으로 카메라를 향합니다. 요구한 약한 사선 방향은 거의 드러나지 않습니다. 사람의 시선이나 이동 방향은 없습니다.",
        "built_space": "맞닫이 목문 두 짝, 기둥 두 개, 기와지붕 하나, 양쪽 돌흙담과 왼쪽 돌확 하나가 보입니다. 문의 위쪽 가로 보강대는 종이 양옆으로 일부 보이지만 참조의 아래쪽 보강대는 재현되지 않았습니다. 지붕부터 기단까지 건물 전면을 담아 장소 소개에 가까운 넓은 구도이며 클로즈업이 아닙니다. 반사는 없습니다.",
        "entities": "사람은 없습니다. 붉은 종이 한 장이 문 중앙과 화면 중앙에 붙어 있으나, 참조의 작은 가로형 봉인지와 달리 문 폭의 절반을 훨씬 넘는 큰 종이입니다. 인쇄 흔적도 두 원형 인장 대신 큰 사각 인장 형태이며 읽을 수 있는 문구는 식별되지 않습니다. 목재·기와·돌흙담과 야간 상태는 대체로 부합합니다. 사랑채와 지게 보관 구석은 보이지 않습니다.",
        "hard_violations": [],
        "physics": "주름진 종이는 중앙 문틈을 가로질러 목문에 붙어 있으며 지지 표면이 분명합니다. 지붕은 보와 기둥이 받치고, 기둥은 주춧돌 위에 있으며 돌확도 땅에 놓여 있습니다. 봉인지 크기는 부정확하지만 물리적으로 떠 있는 상태는 아닙니다."
       }
      ]
     },
     "ok": true
    }
   ],
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "slot_winner_match": true,
   "slot_winner": {
    "gemini-pro": "B",
    "gpt-high": "B"
   },
   "route": "cross_slot_agree"
  },
  "dual": {
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "normalized": {
    "A": 1.0,
    "B": 2.0
   },
   "adjusted": {
    "A": 0.75,
    "B": 2.0
   },
   "violations": {
    "A": [
     "[gemini-pro] 금지된 레퍼런스 카메라 구도 그대로 복사 (Copied reference framing)",
     "[gemini-pro] 사물의 물리적 스케일 왜곡 (Enlarged object physical size instead of using close-up framing)"
    ]
   },
   "per_model_winner": {
    "gemini-pro": "B",
    "gpt-high": "B"
   },
   "agreed": true
  },
  "totals": {
   "B": 2000,
   "A": 750
  },
  "selected": "B",
  "ranking": [
   "B",
   "A"
  ],
  "verdicts": [
   {
    "label": "B",
    "score": 2000,
    "verdict_ko": "지시된 비스듬한 앵글(oblique angle)을 잘 적용했으며, 사물의 사실적인 크기를 유지하면서 야간의 영화적인 분위기를 훌륭하게 연출했습니다."
   },
   {
    "label": "A",
    "score": 750,
    "verdict_ko": "레퍼런스의 프레이밍을 그대로 복사하지 말라는 지시를 어겼으며, 카메라를 피사체에 다가가게 하는 대신 봉인지 자체의 크기를 비현실적으로 거대하게 키워 물리적 스케일 규정을 심각하게 위반했습니다.  ★위반: [gemini-pro] 금지된 레퍼런스 카메라 구도 그대로 복사 (Copied reference framing) / [gemini-pro] 사물의 물리적 스케일 왜곡 (Enlarged object physical size instead of using close-up framing)"
   }
  ],
  "refs": [
   {
    "label": "LOCATION PHOTOGRAPH — the exact place of this shot: its architecture, materials, fixed features and lighting mood are spatial truth; stage the moment inside this place. Never copy its camera framing.",
    "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/db00517b-165b-4005-a287-f28936b59b95/episodes/272bd7af-9294-4490-98c6-88a40e9deacc/images/background_chain/L09B01.png",
    "asset_id": "d34a5459-08b9-4763-b66e-2894dac771d6",
    "role": "location_plate"
   }
  ],
  "critique_skipped": true,
  "shot_run_produced": false,
  "shot_run_uid": "06aa1412-5fb8-790d-a620-f494a857e654",
  "ref_mode": "플레이트만 (배경 전용)",
  "share_plan": {
   "ref_plan": "background"
  }
 },
 "S20sh2::cine": {
  "applied": true,
  "attempted_at": "2026-09-09T03:07:21.176244+00:00",
  "fingerprint": "b009daf0900ac5165dd14ac5691c54e008d90ab43dfad727fed6b2cadab17c93",
  "fingerprint_version": 2,
  "provider": "mai",
  "endpoint": "openrouter/chat-completions",
  "model": "microsoft/mai-image-2.6",
  "pack": "24.202608252115",
  "source_file": "S20sh2_sel.png",
  "source_sha256": "50768ac5306279beca4a68d98721dae4af1a8229d6165553fe34d364c0ba8589",
  "file": "S20sh2_cine.png",
  "staged_sha256": "98cd43cd057d89648d46da1b1dea640f7ad501dd5250ddbdaa184dbf88a2b92b",
  "latency_ms": 26218
 },
 "S20sh6::signage": {
  "fp": "5c828b24a7f4dbbc",
  "inscriptions": [],
  "cues": [],
  "dropped": []
 },
 "era_assess::e5f5f373e7b60323": {
  "subjects": [
   {
    "subject_native": "20세기 초 전라도 전통 한옥 사랑채 방 내부",
    "search_terms_native": [
     "근대 한옥 사랑채 방",
     "전라도 고택 사랑방 내부",
     "전통 한옥 온돌방 장판 문창살"
    ],
    "language_lock_native": "반드시 한국어로만 검색어를 작성하고 다른 언어로 번역하거나 추가하지 마십시오.",
    "reason_ko": "AI 생성 모델은 한국 전통 한옥 방을 일본식 다다미방이나 중국식 격자창 구조와 자주 혼동합니다."
   }
  ],
  "subject_text": "박제현 집 사랑채 방\n목조 한옥 사랑채의 좌식 방. 낮은 상을 놓을 수 있는 바닥과 목재 문틀이 있으며, 열린 방문 너머 마당이 보인다.",
  "identity": "canonical",
  "scope_id": "L09",
  "scope_role": "location_interior",
  "scope_sha": "bfafc49acfc3e77b"
 },
 "era_fail::b9f6ea2defafdb26": {
  "stage": "research",
  "subject": "20세기 초 전라도 전통 한옥 사랑채 방 내부",
  "terms": [
   "근대 한옥 사랑채 방",
   "전라도 고택 사랑방 내부",
   "전통 한옥 온돌방 장판 문창살"
  ],
  "status": "no_usable",
  "queries": [
   [
    "전라도 고택 사랑채 사랑방 내부 온돌방 장판 문창살",
    "전남 전통 한옥 사랑방 내부 장판 창호"
   ],
   [
    "전라도 고택 사랑방 내부 지장 장판",
    "나주 남파고택 사랑채 방 내부"
   ]
  ],
  "candidate_urls": [
   "https://img1.daumcdn.net/thumb/R1280x0.fjpg/?fname=http%3A%2F%2Ft1.daumcdn.net%2Fbrunch%2Fservice%2Fuser%2FwlQ%2Fimage%2FvPPUWlW1jmJmg8worstPcGH1Xc8.jpg",
   "https://stat.ameba.jp/user_images/20161210/23/jnkorea/a1/7d/j/o1024068213818456338.jpg",
   "https://i.pinimg.com/736x/7f/70/ce/7f70ce54eead120c1e6716fbbc166210--korean-traditional-korean-style.jpg",
   "https://yaimg.yanolja.com/v5/2024/08/01/09/1280/66ab5928db02b3.11353592.jpg"
  ],
  "coarse": {
   "eligible": [],
   "chosen_index": 0,
   "reason": "종류·보임·기준을 다 만족하는 후보가 없다 — 이 라운드에선 안 고른다",
   "single_judge": true,
   "rejected_judges": {}
  },
  "verdicts": [
   {
    "index": 1,
    "object_type_match": "yes",
    "visible": true,
    "criteria_match": "unsure",
    "similarity": 70
   },
   {
    "index": 2,
    "object_type_match": "yes",
    "visible": true,
    "criteria_match": "unsure",
    "similarity": 75
   },
   {
    "index": 3,
    "object_type_match": "yes",
    "visible": true,
    "criteria_match": "unsure",
    "similarity": 95
   },
   {
    "index": 4,
    "object_type_match": "yes",
    "visible": true,
    "criteria_match": "unsure",
    "similarity": 70
   }
  ],
  "chosen_reason_ko": "",
  "attempts": 5
 },
 "S20sh6::bgfirst_bg": {
  "input_fingerprint": "ba6d1a2cbf78c0c6",
  "prompt": "Create the EMPTY BACKGROUND PLATE for one film shot — NO PEOPLE, no figures, no body parts, no sketch lines, no arrows anywhere.\n\"Empty\" means no people only: KEEP the location's inherent occupants and stock that define the place — animals in an animal shelter, pen or farm, goods in a market, moored boats in a harbour — unless the shot text explicitly removes them.\nThe FIRST attached image is a thin-line storyboard sketch: use ONLY its camera angle, horizon, perspective and the placement/size of buildings and set masses — ignore the sketched people and arrows entirely. The SECOND attached image (LOCATION PHOTOGRAPH) is the real place: take its architecture, materials, signage and fixed features, and RE-PROJECT them into the sketch's camera. If the photograph's camera differs from the sketch's, the sketch's camera wins.\nHUMAN-SCALE CALIBRATION: derive every structure's true size from human-scale elements — a door ≈ 2m, a window ≈ 1–1.5m wide, one storey ≈ 2.5–3m; never inflate a small structure or shrink a large one.\n\nSHOT TEXT this background must serve (Korean): 박제현과 시선을 마주한 채 눈빛이 파르르 흔들리는 순간의 박상현의 얼굴 클로즈업.\n\nLOCATION (lock): Inside the traditional estate's dim guest-wing meeting room at night, in the floor-seating area with no drinking table.\n\nTIME OF DAY (lock): night.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: close-up\n- KEY BACKGROUND ELEMENTS: 사랑채 실내 (The seated conversation continues without the customary drinking table); used as A minimal, defocused interior margin keeps the reaction in the shared room without adding furnishings.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Preserve the explicitly dark interior with restrained ambient modeling and readable eyes, without assigning an unsupported source or color cast.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nRender ONE photorealistic empty location photograph, 16:9, that this shot can be staged inside later. No readable writing anywhere: surfaces that would carry writing may be present, but stage any wording out of legibility — an oblique angle, distance, shallow focus. No captions, watermarks or overlay text.",
  "effective_prompt": "Create the EMPTY BACKGROUND PLATE for one film shot — NO PEOPLE, no figures, no body parts, no sketch lines, no arrows anywhere.\n\"Empty\" means no people only: KEEP the location's inherent occupants and stock that define the place — animals in an animal shelter, pen or farm, goods in a market, moored boats in a harbour — unless the shot text explicitly removes them.\nThe FIRST attached image is a thin-line storyboard sketch: use ONLY its camera angle, horizon, perspective and the placement/size of buildings and set masses — ignore the sketched people and arrows entirely. The SECOND attached image (LOCATION PHOTOGRAPH) is the real place: take its architecture, materials, signage and fixed features, and RE-PROJECT them into the sketch's camera. If the photograph's camera differs from the sketch's, the sketch's camera wins.\nHUMAN-SCALE CALIBRATION: derive every structure's true size from human-scale elements — a door ≈ 2m, a window ≈ 1–1.5m wide, one storey ≈ 2.5–3m; never inflate a small structure or shrink a large one.\n\nSHOT TEXT this background must serve (Korean): 박제현과 시선을 마주한 채 눈빛이 파르르 흔들리는 순간의 박상현의 얼굴 클로즈업.\n\nLOCATION (lock): Inside the traditional estate's dim guest-wing meeting room at night, in the floor-seating area with no drinking table.\n\nTIME OF DAY (lock): night.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: close-up\n- KEY BACKGROUND ELEMENTS: 사랑채 실내 (The seated conversation continues without the customary drinking table); used as A minimal, defocused interior margin keeps the reaction in the shared room without adding furnishings.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Preserve the explicitly dark interior with restrained ambient modeling and readable eyes, without assigning an unsupported source or color cast.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nRender ONE photorealistic empty location photograph, 16:9, that this shot can be staged inside later. No readable writing anywhere: surfaces that would carry writing may be present, but stage any wording out of legibility — an oblique angle, distance, shallow focus. No captions, watermarks or overlay text.",
  "bg_path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/db00517b-165b-4005-a287-f28936b59b95/images/272bd7af-9294-4490-98c6-88a40e9deacc/scene/recipe/S20sh6__bgfirst_bg.png",
  "asset_id": "94bf919b-6f27-43e6-80df-67920439a9a2",
  "input_asset_ids": [
   "278fc18d-344d-447a-8a88-b10740614a49",
   "2554259b-2ac4-4116-a9a5-ab2266901e39"
  ]
 },
 "S20sh6": {
  "input_fingerprint": "f6764d03f22ddf6f",
  "prompt": "Create ONE FINAL photorealistic live-action film still of the moment below — early 20th-century Korea, Jeolla region; all people are Korean unless explicitly stated otherwise. TIME OF DAY (lock): night.\n\nSHOT TEXT (authoritative, Korean): 박제현과 시선을 마주한 채 눈빛이 파르르 흔들리는 순간의 박상현의 얼굴 클로즈업.\n\nLOCATION (lock): Inside the traditional estate's dim guest-wing meeting room at night, in the floor-seating area with no drinking table. The shot takes place here — the attached LOCATION PHOTOGRAPH shows the exact spot.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: close-up\n- KEY BACKGROUND ELEMENTS: 사랑채 실내 (The seated conversation continues without the customary drinking table); used as A minimal, defocused interior margin keeps the reaction in the shared room without adding furnishings.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Preserve the explicitly dark interior with restrained ambient modeling and readable eyes, without assigning an unsupported source or color cast.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The red seizure seal remains on the grain-store door; no drinking table is set, and the frame's usual yard corner is empty. 박상현: He remains seated, his expression losing its composure.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 박상현 (한국인 남성, 성인 나이대, 자연스러운 검은 머리) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.",
  "roll_prompts": {
   "A": "Stage the shot. The FIRST attached image (SHOT BACKGROUND) is the finished empty background of this shot — keep it EXACTLY: its camera, perspective, architecture, lighting and every fixed feature stay untouched. The SECOND attached image (LAYOUT SKETCH) tells you ONLY where the people go: each sketched person's position, screen size, pose and the gaze/motion arrows. Ignore the sketch's background lines. The CHARACTER REFERENCE photographs show the real people.\nPlace the real people into the background at exactly the sketched positions, sizes and poses, following the arrow directions. No sketch lines or arrows may remain.\n\nCreate ONE FINAL photorealistic live-action film still of the moment below — early 20th-century Korea, Jeolla region; all people are Korean unless explicitly stated otherwise. TIME OF DAY (lock): night.\n\nSHOT TEXT (authoritative, Korean): 박제현과 시선을 마주한 채 눈빛이 파르르 흔들리는 순간의 박상현의 얼굴 클로즈업.\n\nLOCATION (lock): Inside the traditional estate's dim guest-wing meeting room at night, in the floor-seating area with no drinking table. The shot takes place here — the FIRST attached image (SHOT BACKGROUND) is this exact place, already built: its ground, structures, horizon, materials and lighting are the finished truth of this location and must not be redesigned or replaced. No location photograph is attached — read the place from that image alone, and add no scenery, structure, vehicle or fixture that it does not already show. This lock governs the place only; the figures in the shot follow the staging and pose instructions.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: close-up\n- KEY BACKGROUND ELEMENTS: 사랑채 실내 (The seated conversation continues without the customary drinking table); used as A minimal, defocused interior margin keeps the reaction in the shared room without adding furnishings.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Preserve the explicitly dark interior with restrained ambient modeling and readable eyes, without assigning an unsupported source or color cast.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The red seizure seal remains on the grain-store door; no drinking table is set, and the frame's usual yard corner is empty. 박상현: He remains seated, his expression losing its composure.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 박상현 (한국인 남성, 성인 나이대, 자연스러운 검은 머리) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.",
   "B": "Create ONE FINAL photorealistic live-action film still of the moment below — early 20th-century Korea, Jeolla region; all people are Korean unless explicitly stated otherwise. TIME OF DAY (lock): night.\n\nSHOT TEXT (authoritative, Korean): 박제현과 시선을 마주한 채 눈빛이 파르르 흔들리는 순간의 박상현의 얼굴 클로즈업.\n\nLOCATION (lock): Inside the traditional estate's dim guest-wing meeting room at night, in the floor-seating area with no drinking table. The shot takes place here — the attached LOCATION PHOTOGRAPH shows the exact spot.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: close-up\n- KEY BACKGROUND ELEMENTS: 사랑채 실내 (The seated conversation continues without the customary drinking table); used as A minimal, defocused interior margin keeps the reaction in the shared room without adding furnishings.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Preserve the explicitly dark interior with restrained ambient modeling and readable eyes, without assigning an unsupported source or color cast.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The red seizure seal remains on the grain-store door; no drinking table is set, and the frame's usual yard corner is empty. 박상현: He remains seated, his expression losing its composure.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 박상현 (한국인 남성, 성인 나이대, 자연스러운 검은 머리) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.\n\nCOMPOSITION VARIATION (this candidate only): keep the place, the people and the moment exactly as specified, and keep from the CAMERA & FRAME contract its subject scale, screen placement and key background elements exactly. For this candidate only, its camera angle and height are deliberately left open: choose a camera position distinctly different from the obvious one, as a film director picking a second setup on the same blocking. Stay true to any explicit framing the SHOT TEXT itself dictates."
  },
  "roll_refs": {
   "A": [
    {
     "label": "SHOT BACKGROUND",
     "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/db00517b-165b-4005-a287-f28936b59b95/images/272bd7af-9294-4490-98c6-88a40e9deacc/scene/recipe/S20sh6__bgfirst_bg.png",
     "asset_id": "94bf919b-6f27-43e6-80df-67920439a9a2",
     "role": "bgfirst_bg"
    },
    {
     "label": "LAYOUT SKETCH (people placement only)",
     "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/db00517b-165b-4005-a287-f28936b59b95/images/272bd7af-9294-4490-98c6-88a40e9deacc/conti/conti_S20sh6.png",
     "asset_id": "278fc18d-344d-447a-8a88-b10740614a49",
     "role": "conti_light"
    },
    {
     "label": "CHARACTER REFERENCE — 박상현: the exact person appearing in this shot; match face, hair and build exactly.",
     "path": "<bytes:840781>",
     "asset_id": "6187751d-5df3-4eb7-9d30-d6fc98a005e6",
     "role": "character_ref"
    }
   ],
   "B": [
    {
     "label": "LOCATION PHOTOGRAPH — the exact place of this shot: its architecture, materials, fixed features and lighting mood are spatial truth; stage the moment inside this place. Never copy its camera framing.",
     "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/db00517b-165b-4005-a287-f28936b59b95/episodes/272bd7af-9294-4490-98c6-88a40e9deacc/images/background_chain/L09B02.png",
     "asset_id": "2554259b-2ac4-4116-a9a5-ab2266901e39",
     "role": "location_plate"
    },
    {
     "label": "CHARACTER REFERENCE — 박상현: the exact person appearing in this shot; match face, hair and build exactly.",
     "path": "<bytes:840781>",
     "asset_id": "6187751d-5df3-4eb7-9d30-d6fc98a005e6",
     "role": "character_ref"
    }
   ]
  },
  "shot_run_spend_attempt_count": 1,
  "judge_flip_skipped": "judge_fn_owns_order",
  "readings": [
   {
    "label": "A",
    "direction": "박상현의 시선이 프레임 전경에 있는 인물(박제현)의 얼굴 쪽을 정확히 향하고 있어 '시선을 마주한 채'라는 텍스트를 충실히 따름.",
    "built_space": "배경은 로케이션 레퍼런스와 완벽히 일치함. 열린 문 너머의 마당, 건너편 건물의 붉은 부적, 바구니 등 공간적 디테일이 정확히 묘사되었으며 술상 없이 구성됨.",
    "entities": "박상현은 레퍼런스와 일치하는 얼굴과 회색 한복을 착용함. 프롬프트가 명시한 '자연스러운 검은 머리'가 잘 드러나며(갓 미착용), 전경의 인물은 텍스트에 등장하는 박제현으로 인식됨.",
    "hard_violations": [
     "[gpt-high] 가시 인물을 박상현으로 제한한 조건과 달리, 왼쪽 전경에 두 번째 남성의 머리와 상체를 추가했다."
    ],
    "physics": "어색함 없이 바닥에 안정적으로 앉아있는 자세가 물리적으로 타당함."
   },
   {
    "label": "B",
    "direction": "박상현의 시선이 화면 밖 아래쪽을 향하고 있어, 누군가와 시선을 마주치는 상황으로 보이지 않음.",
    "built_space": "로케이션 레퍼런스 특유의 마당과 건물 구조가 보이지 않으며, 정체불명의 기둥과 밋밋한 벽면만 묘사되어 장소 일치도가 매우 낮음.",
    "entities": "박상현의 얼굴과 회색 한복, 갓은 캐릭터 레퍼런스와 잘 일치함.",
    "hard_violations": [],
    "physics": "자연스럽게 앉아있는 자세가 안정적으로 묘사됨."
   }
  ],
  "cross_model_order": {
   "policy": "cross_model_order_v2_slot0_fwd_slot1_rev_parallel2",
   "slots": [
    {
     "model": "gemini-pro",
     "order": "forward",
     "display_to_canonical": {
      "A": "A",
      "B": "B"
     },
     "raw": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 9,
        "verdict_ko": "텍스트가 요구한 '시선을 마주하는' 연출과 지정된 로케이션 배경을 정확하게 구현한 훌륭한 샷입니다."
       },
       {
        "label": "B",
        "score": 4,
        "verdict_ko": "시선이 아래를 향해 핵심 연출(시선 맞춤)을 놓쳤으며, 지정된 배경 로케이션의 특징도 전혀 반영되지 않았습니다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "박상현의 시선이 프레임 전경에 있는 인물(박제현)의 얼굴 쪽을 정확히 향하고 있어 '시선을 마주한 채'라는 텍스트를 충실히 따름.",
        "built_space": "배경은 로케이션 레퍼런스와 완벽히 일치함. 열린 문 너머의 마당, 건너편 건물의 붉은 부적, 바구니 등 공간적 디테일이 정확히 묘사되었으며 술상 없이 구성됨.",
        "entities": "박상현은 레퍼런스와 일치하는 얼굴과 회색 한복을 착용함. 프롬프트가 명시한 '자연스러운 검은 머리'가 잘 드러나며(갓 미착용), 전경의 인물은 텍스트에 등장하는 박제현으로 인식됨.",
        "hard_violations": [],
        "physics": "어색함 없이 바닥에 안정적으로 앉아있는 자세가 물리적으로 타당함."
       },
       {
        "label": "B",
        "direction": "박상현의 시선이 화면 밖 아래쪽을 향하고 있어, 누군가와 시선을 마주치는 상황으로 보이지 않음.",
        "built_space": "로케이션 레퍼런스 특유의 마당과 건물 구조가 보이지 않으며, 정체불명의 기둥과 밋밋한 벽면만 묘사되어 장소 일치도가 매우 낮음.",
        "entities": "박상현의 얼굴과 회색 한복, 갓은 캐릭터 레퍼런스와 잘 일치함.",
        "hard_violations": [],
        "physics": "자연스럽게 앉아있는 자세가 안정적으로 묘사됨."
       }
      ],
      "all_candidates_fail": false
     },
     "normalized": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 9,
        "verdict_ko": "텍스트가 요구한 '시선을 마주하는' 연출과 지정된 로케이션 배경을 정확하게 구현한 훌륭한 샷입니다."
       },
       {
        "label": "B",
        "score": 4,
        "verdict_ko": "시선이 아래를 향해 핵심 연출(시선 맞춤)을 놓쳤으며, 지정된 배경 로케이션의 특징도 전혀 반영되지 않았습니다."
       }
      ],
      "all_candidates_fail": false,
      "readings": [
       {
        "label": "A",
        "direction": "박상현의 시선이 프레임 전경에 있는 인물(박제현)의 얼굴 쪽을 정확히 향하고 있어 '시선을 마주한 채'라는 텍스트를 충실히 따름.",
        "built_space": "배경은 로케이션 레퍼런스와 완벽히 일치함. 열린 문 너머의 마당, 건너편 건물의 붉은 부적, 바구니 등 공간적 디테일이 정확히 묘사되었으며 술상 없이 구성됨.",
        "entities": "박상현은 레퍼런스와 일치하는 얼굴과 회색 한복을 착용함. 프롬프트가 명시한 '자연스러운 검은 머리'가 잘 드러나며(갓 미착용), 전경의 인물은 텍스트에 등장하는 박제현으로 인식됨.",
        "hard_violations": [],
        "physics": "어색함 없이 바닥에 안정적으로 앉아있는 자세가 물리적으로 타당함."
       },
       {
        "label": "B",
        "direction": "박상현의 시선이 화면 밖 아래쪽을 향하고 있어, 누군가와 시선을 마주치는 상황으로 보이지 않음.",
        "built_space": "로케이션 레퍼런스 특유의 마당과 건물 구조가 보이지 않으며, 정체불명의 기둥과 밋밋한 벽면만 묘사되어 장소 일치도가 매우 낮음.",
        "entities": "박상현의 얼굴과 회색 한복, 갓은 캐릭터 레퍼런스와 잘 일치함.",
        "hard_violations": [],
        "physics": "자연스럽게 앉아있는 자세가 안정적으로 묘사됨."
       }
      ]
     },
     "ok": true
    },
    {
     "model": "gpt-high",
     "order": "reverse",
     "display_to_canonical": {
      "A": "B",
      "B": "A"
     },
     "raw": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 8,
        "verdict_ko": "화면 밖 박제현을 향한 불안한 시선과 박상현 단독 얼굴 클로즈업, 갓과 회색 의복을 충실히 구현했으나 배경 여백은 다소 넓다."
       },
       {
        "label": "B",
        "score": 3,
        "verdict_ko": "상대와 눈을 맞추는 불안한 표정은 명확하지만, 허용되지 않은 두 번째 인물을 전경에 넣었고 갓을 없애며 배경까지 넓혀 단독 반응 클로즈업에서 벗어났다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "박상현의 얼굴과 두 눈은 화면 왼쪽의 카메라 밖 상대를 향한다. 박제현을 바라보는 대화 시선으로 자연스럽게 읽히며 렌즈를 정면 응시하지 않는다. 상대가 보이지 않아 실제 상호 눈맞춤까지 확인할 수는 없다. 무기나 방향성 소품은 없다.",
        "built_space": "왼쪽에 흐릿한 흙벽과 세로 목재 구획, 아래쪽 가로 목재가 보이고, 오른쪽 뒤에는 문설주 하나와 어두운 개구부 일부가 보인다. 참조의 목재·흙벽 실내와 양립한다. 술상이나 새 가구, 중복 설비, 반사는 없다. 얼굴과 어깨 중심의 구도라 바닥 좌정 위치는 화면 밖이다.",
        "entities": "성인 한국인 남성으로 설정된 박상현 한 명만 보인다. 얼굴 윤곽, 짧은 콧수염과 턱수염, 검은 갓과 망건, 회색 전통 의복이 인물 참조에 대체로 부합한다. 머리카락은 머리쓰개에 가려져 있다. 눈은 정상적인 홍채와 동공을 유지하고, 긴장한 눈썹과 살짝 벌어진 입으로 평정이 흔들리는 반응을 표현한다. 곡간 봉인과 마당 구석은 구도 밖이며 읽을 수 있는 글자는 없다.",
        "hard_violations": [],
        "physics": "머리는 목과 어깨에 자연스럽게 이어지고 갓은 머리에 얹혀 턱끈으로 고정되어 있다. 옷은 어깨에 걸려 자연스럽게 주름진다. 엉덩이와 바닥 접촉은 클로즈업 밖이므로 좌정 자체는 확인할 수 없지만, 보이는 상체에 부유나 불가능한 자세는 없다."
       },
       {
        "label": "B",
        "direction": "박상현의 눈은 왼쪽 전경 남성의 얼굴 쪽을 향하고, 전경 남성도 박상현 쪽으로 고개를 돌리고 있다. 눈맞춤 관계는 A보다 직접적으로 드러나지만 상대의 눈은 보이지 않는다. 무기나 방향성 소품은 없다.",
        "built_space": "왼쪽에 격자창 일부 하나, 오른쪽에 큰 출입구 하나와 그 너머 마루 기둥 하나, 맞은편 기와지붕과 곡간 문 하나가 보인다. 참조 장소의 구조와 밤 풍경은 상당히 가깝다. 술상과 반사는 없지만, 최소한의 흐린 실내 여백이어야 할 배경에 마당과 외부 건물이 넓고 비교적 또렷하게 노출된다. 두 사람의 좌면과 바닥 접촉은 화면 밖이다.",
        "entities": "중앙에 성인 한국인 남성으로 보이는 박상현이 있고, 왼쪽 전경에 별도의 성인 남성의 뒷머리·귀·어깨가 크게 들어온다. 박상현의 회색 의복과 수염은 참조와 유사하지만 참조의 갓과 망건 대신 짧게 빗은 검은 머리가 드러난다. 눈과 표정은 해부학적으로 자연스럽고 불안감이 읽힌다. 뒤쪽 곡간 문에는 작은 붉은 봉인이 보이며 글자는 판독되지 않는다.",
        "hard_violations": [
         "가시 인물을 박상현으로 제한한 조건과 달리, 왼쪽 전경에 두 번째 남성의 머리와 상체를 추가했다."
        ],
        "physics": "박상현의 머리와 상체, 전경 남성의 머리와 어깨는 각각 자연스럽게 연결된다. 의복도 몸에 지지되어 늘어진다. 하체는 잘려 있어 좌정 지지점은 확인할 수 없으나, 떠 있는 신체나 무지지 물체, 불가능한 관절 자세는 보이지 않는다."
       }
      ],
      "all_candidates_fail": false
     },
     "normalized": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 8,
        "verdict_ko": "화면 밖 박제현을 향한 불안한 시선과 박상현 단독 얼굴 클로즈업, 갓과 회색 의복을 충실히 구현했으나 배경 여백은 다소 넓다."
       },
       {
        "label": "A",
        "score": 3,
        "verdict_ko": "상대와 눈을 맞추는 불안한 표정은 명확하지만, 허용되지 않은 두 번째 인물을 전경에 넣었고 갓을 없애며 배경까지 넓혀 단독 반응 클로즈업에서 벗어났다."
       }
      ],
      "all_candidates_fail": false,
      "readings": [
       {
        "label": "B",
        "direction": "박상현의 얼굴과 두 눈은 화면 왼쪽의 카메라 밖 상대를 향한다. 박제현을 바라보는 대화 시선으로 자연스럽게 읽히며 렌즈를 정면 응시하지 않는다. 상대가 보이지 않아 실제 상호 눈맞춤까지 확인할 수는 없다. 무기나 방향성 소품은 없다.",
        "built_space": "왼쪽에 흐릿한 흙벽과 세로 목재 구획, 아래쪽 가로 목재가 보이고, 오른쪽 뒤에는 문설주 하나와 어두운 개구부 일부가 보인다. 참조의 목재·흙벽 실내와 양립한다. 술상이나 새 가구, 중복 설비, 반사는 없다. 얼굴과 어깨 중심의 구도라 바닥 좌정 위치는 화면 밖이다.",
        "entities": "성인 한국인 남성으로 설정된 박상현 한 명만 보인다. 얼굴 윤곽, 짧은 콧수염과 턱수염, 검은 갓과 망건, 회색 전통 의복이 인물 참조에 대체로 부합한다. 머리카락은 머리쓰개에 가려져 있다. 눈은 정상적인 홍채와 동공을 유지하고, 긴장한 눈썹과 살짝 벌어진 입으로 평정이 흔들리는 반응을 표현한다. 곡간 봉인과 마당 구석은 구도 밖이며 읽을 수 있는 글자는 없다.",
        "hard_violations": [],
        "physics": "머리는 목과 어깨에 자연스럽게 이어지고 갓은 머리에 얹혀 턱끈으로 고정되어 있다. 옷은 어깨에 걸려 자연스럽게 주름진다. 엉덩이와 바닥 접촉은 클로즈업 밖이므로 좌정 자체는 확인할 수 없지만, 보이는 상체에 부유나 불가능한 자세는 없다."
       },
       {
        "label": "A",
        "direction": "박상현의 눈은 왼쪽 전경 남성의 얼굴 쪽을 향하고, 전경 남성도 박상현 쪽으로 고개를 돌리고 있다. 눈맞춤 관계는 A보다 직접적으로 드러나지만 상대의 눈은 보이지 않는다. 무기나 방향성 소품은 없다.",
        "built_space": "왼쪽에 격자창 일부 하나, 오른쪽에 큰 출입구 하나와 그 너머 마루 기둥 하나, 맞은편 기와지붕과 곡간 문 하나가 보인다. 참조 장소의 구조와 밤 풍경은 상당히 가깝다. 술상과 반사는 없지만, 최소한의 흐린 실내 여백이어야 할 배경에 마당과 외부 건물이 넓고 비교적 또렷하게 노출된다. 두 사람의 좌면과 바닥 접촉은 화면 밖이다.",
        "entities": "중앙에 성인 한국인 남성으로 보이는 박상현이 있고, 왼쪽 전경에 별도의 성인 남성의 뒷머리·귀·어깨가 크게 들어온다. 박상현의 회색 의복과 수염은 참조와 유사하지만 참조의 갓과 망건 대신 짧게 빗은 검은 머리가 드러난다. 눈과 표정은 해부학적으로 자연스럽고 불안감이 읽힌다. 뒤쪽 곡간 문에는 작은 붉은 봉인이 보이며 글자는 판독되지 않는다.",
        "hard_violations": [
         "가시 인물을 박상현으로 제한한 조건과 달리, 왼쪽 전경에 두 번째 남성의 머리와 상체를 추가했다."
        ],
        "physics": "박상현의 머리와 상체, 전경 남성의 머리와 어깨는 각각 자연스럽게 연결된다. 의복도 몸에 지지되어 늘어진다. 하체는 잘려 있어 좌정 지지점은 확인할 수 없으나, 떠 있는 신체나 무지지 물체, 불가능한 관절 자세는 보이지 않는다."
       }
      ]
     },
     "ok": true
    }
   ],
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "slot_winner_match": false,
   "slot_winner": {
    "gemini-pro": "A",
    "gpt-high": "B"
   },
   "route": "cross_slot_combined"
  },
  "dual": {
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "normalized": {
    "A": 1.375,
    "B": 1.444
   },
   "adjusted": {
    "A": 1.125,
    "B": 1.444
   },
   "violations": {
    "A": [
     "[gpt-high] 가시 인물을 박상현으로 제한한 조건과 달리, 왼쪽 전경에 두 번째 남성의 머리와 상체를 추가했다."
    ]
   },
   "per_model_winner": {
    "gemini-pro": "A",
    "gpt-high": "B"
   },
   "agreed": false
  },
  "totals": {
   "A": 1125,
   "B": 1444
  },
  "selected": "B",
  "ranking": [
   "B",
   "A"
  ],
  "verdicts": [
   {
    "label": "A",
    "score": 1125,
    "verdict_ko": "텍스트가 요구한 '시선을 마주하는' 연출과 지정된 로케이션 배경을 정확하게 구현한 훌륭한 샷입니다.  ★위반: [gpt-high] 가시 인물을 박상현으로 제한한 조건과 달리, 왼쪽 전경에 두 번째 남성의 머리와 상체를 추가했다."
   },
   {
    "label": "B",
    "score": 1444,
    "verdict_ko": "시선이 아래를 향해 핵심 연출(시선 맞춤)을 놓쳤으며, 지정된 배경 로케이션의 특징도 전혀 반영되지 않았습니다."
   }
  ],
  "refs": [
   {
    "label": "LOCATION PHOTOGRAPH — the exact place of this shot: its architecture, materials, fixed features and lighting mood are spatial truth; stage the moment inside this place. Never copy its camera framing.",
    "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/db00517b-165b-4005-a287-f28936b59b95/episodes/272bd7af-9294-4490-98c6-88a40e9deacc/images/background_chain/L09B02.png",
    "asset_id": "2554259b-2ac4-4116-a9a5-ab2266901e39",
    "role": "location_plate"
   },
   {
    "label": "CHARACTER REFERENCE — 박상현: the exact person appearing in this shot; match face, hair and build exactly.",
    "path": "<bytes:840781>",
    "asset_id": "6187751d-5df3-4eb7-9d30-d6fc98a005e6",
    "role": "character_ref"
   }
  ],
  "critique_skipped": true,
  "shot_run_produced": false,
  "shot_run_uid": "06aa1412-615c-70b9-9f89-75fd03300491",
  "bgfirst": {
   "bg_path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/db00517b-165b-4005-a287-f28936b59b95/images/272bd7af-9294-4490-98c6-88a40e9deacc/scene/recipe/S20sh6__bgfirst_bg.png",
   "bg_asset_id": "94bf919b-6f27-43e6-80df-67920439a9a2",
   "bg_record_key": "S20sh6::bgfirst_bg",
   "chain_winner": false,
   "authority": "plate"
  },
  "ref_mode": "플레이트+엔티티 (2택1: 무콘티 승)",
  "share_plan": {
   "ref_plan": "background"
  }
 },
 "S20sh6::cine": {
  "applied": true,
  "attempted_at": "2026-09-09T08:58:25.221854+00:00",
  "fingerprint": "5c775db44619f558aed33a52a81bca798a5be4b6dbec3cfbca47d2f0d8a61828",
  "fingerprint_version": 2,
  "provider": "mai",
  "endpoint": "openrouter/chat-completions",
  "model": "microsoft/mai-image-2.6",
  "pack": "24.202608252115",
  "source_file": "S20sh6_sel.png",
  "source_sha256": "f305fd912283d6cb8b92179c752ae08f2c44727f4121be479b0b9800c5d5b959",
  "file": "S20sh6_cine.png",
  "staged_sha256": "056353162c16191e11e969f7b70a465bc1e17966cb9d83e29113bbbcdd048d6e",
  "latency_ms": 32488
 },
 "S20sh8::signage": {
  "fp": "c7e7ba08bbba15e6",
  "inscriptions": [],
  "cues": [],
  "dropped": []
 },
 "S20sh8": {
  "input_fingerprint": "e29233a2c3105ffd",
  "prompt": "Create ONE FINAL photorealistic live-action film still of the moment below — early 20th-century Korea, Jeolla region; all people are Korean unless explicitly stated otherwise. TIME OF DAY (lock): night.\n\nSHOT TEXT (authoritative, Korean): 마당 구석, 나무 지게가 사라진 텅 빈 흙바닥 곁의 어두운 공간.\n\nLOCATION (lock): The dark dirt corner of the estate's outdoor courtyard where a wooden carrying frame normally stood. The shot takes place here — the attached LOCATION PHOTOGRAPH shows the exact spot.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: wide shot\n- KEY BACKGROUND ELEMENTS: 마당 구석의 빈자리 (The carrying frame that normally stood here is absent); used as The vacant space itself is the subject, held long enough for absence to register; 마당 흙바닥 (Visible beneath and around the empty place); used as Provides continuous spatial context without invented footprints, stains, or object-shaped marks.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Keep the yard corner explicitly dark with subdued nighttime ambient contrast, without extending any unestablished interior light source into the exterior.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The carrying frame is absent from its usual yard corner; the sarangchae door is open, the drinking table absent, and the grain-store door bears its red seal.\n\nNO PEOPLE IN THIS SHOT: the shot text shows only the place and its state — no living person appears in frame, and no face, and no figure standing, sitting or passing through. Do not invent someone to occupy the space.\n\nOne thing this rule does NOT forbid: the hand or arm of whoever is doing what the shot text describes. If the moment is an object being held, carried, operated, read, aimed, worn or handed over, the hand doing it belongs in frame, resting on or gripping that object exactly as its real use requires. An object photographed floating unsupported, with the hand that must be holding it erased, is a worse failure than the person you were avoiding — render the hand.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.",
  "roll_prompts": {
   "A": "Create ONE FINAL photorealistic live-action film still of the moment below — early 20th-century Korea, Jeolla region; all people are Korean unless explicitly stated otherwise. TIME OF DAY (lock): night.\n\nSHOT TEXT (authoritative, Korean): 마당 구석, 나무 지게가 사라진 텅 빈 흙바닥 곁의 어두운 공간.\n\nLOCATION (lock): The dark dirt corner of the estate's outdoor courtyard where a wooden carrying frame normally stood. The shot takes place here — the attached LOCATION PHOTOGRAPH shows the exact spot.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: wide shot\n- KEY BACKGROUND ELEMENTS: 마당 구석의 빈자리 (The carrying frame that normally stood here is absent); used as The vacant space itself is the subject, held long enough for absence to register; 마당 흙바닥 (Visible beneath and around the empty place); used as Provides continuous spatial context without invented footprints, stains, or object-shaped marks.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Keep the yard corner explicitly dark with subdued nighttime ambient contrast, without extending any unestablished interior light source into the exterior.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The carrying frame is absent from its usual yard corner; the sarangchae door is open, the drinking table absent, and the grain-store door bears its red seal.\n\nNO PEOPLE IN THIS SHOT: the shot text shows only the place and its state — no living person appears in frame, and no face, and no figure standing, sitting or passing through. Do not invent someone to occupy the space.\n\nOne thing this rule does NOT forbid: the hand or arm of whoever is doing what the shot text describes. If the moment is an object being held, carried, operated, read, aimed, worn or handed over, the hand doing it belongs in frame, resting on or gripping that object exactly as its real use requires. An object photographed floating unsupported, with the hand that must be holding it erased, is a worse failure than the person you were avoiding — render the hand.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.",
   "B": "Create ONE FINAL photorealistic live-action film still of the moment below — early 20th-century Korea, Jeolla region; all people are Korean unless explicitly stated otherwise. TIME OF DAY (lock): night.\n\nSHOT TEXT (authoritative, Korean): 마당 구석, 나무 지게가 사라진 텅 빈 흙바닥 곁의 어두운 공간.\n\nLOCATION (lock): The dark dirt corner of the estate's outdoor courtyard where a wooden carrying frame normally stood. The shot takes place here — the attached LOCATION PHOTOGRAPH shows the exact spot.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: wide shot\n- KEY BACKGROUND ELEMENTS: 마당 구석의 빈자리 (The carrying frame that normally stood here is absent); used as The vacant space itself is the subject, held long enough for absence to register; 마당 흙바닥 (Visible beneath and around the empty place); used as Provides continuous spatial context without invented footprints, stains, or object-shaped marks.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Keep the yard corner explicitly dark with subdued nighttime ambient contrast, without extending any unestablished interior light source into the exterior.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The carrying frame is absent from its usual yard corner; the sarangchae door is open, the drinking table absent, and the grain-store door bears its red seal.\n\nNO PEOPLE IN THIS SHOT: the shot text shows only the place and its state — no living person appears in frame, and no face, and no figure standing, sitting or passing through. Do not invent someone to occupy the space.\n\nOne thing this rule does NOT forbid: the hand or arm of whoever is doing what the shot text describes. If the moment is an object being held, carried, operated, read, aimed, worn or handed over, the hand doing it belongs in frame, resting on or gripping that object exactly as its real use requires. An object photographed floating unsupported, with the hand that must be holding it erased, is a worse failure than the person you were avoiding — render the hand.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.\n\nCOMPOSITION VARIATION (this candidate only): keep the place, the people and the moment exactly as specified, and keep from the CAMERA & FRAME contract its subject scale, screen placement and key background elements exactly. For this candidate only, its camera angle and height are deliberately left open: choose a camera position distinctly different from the obvious one, as a film director picking a second setup on the same blocking. Stay true to any explicit framing the SHOT TEXT itself dictates."
  },
  "shot_run_spend_attempt_count": 1,
  "readings": [
   {
    "label": "A",
    "direction": "특정 방향성을 가지는 인물이나 무기, 이동하는 물체가 화면에 존재하지 않음.",
    "built_space": "참조 이미지에 기반한 공간 구조(왼쪽 사랑채 기둥, 담장, 오른쪽 곡간)를 잘 재현함. 왼쪽 측면에 열린 사랑채 문이 보임.",
    "entities": "프롬프트에서 '사라졌다(absent)'고 명시한 나무 지게가 담장 곁에 그려져 있음. 곡간 문에 붉은색 자국(봉인)이 있음. 인물은 존재하지 않음.",
    "hard_violations": [
     "[gemini-pro] 지게가 없는 빈 공간이 피사체가 되어야 한다는 지시(The carrying frame is absent)를 어기고 나무 지게를 화면 중앙에 배치함",
     "[gpt-high] 없어야 한다고 명시한 나무 지게를 마당 구석에 배치하여, 빈자리만 보여야 하는 장면에 금지된 소품을 추가했다."
    ],
    "physics": "담장에 기대어 있는 지게를 포함하여, 물리 법칙을 거스르거나 지지대 없이 떠 있는 사물은 발견되지 않음."
   },
   {
    "label": "B",
    "direction": "특정 방향성을 가지는 인물이나 무기, 이동하는 물체가 화면에 존재하지 않음.",
    "built_space": "참조 이미지와 동일한 구조의 마당을 보여줌. 왼쪽에는 사랑채의 기둥과 열려 있는 격자문이 배치되어 있으며, 중앙의 돌담과 오른쪽의 곡간 건물이 올바른 위치에 자연스럽게 묘사됨.",
    "entities": "프롬프트의 요구대로 마당 구석에 나무 지게가 존재하지 않음. 곡간 문 중앙에 붉은색 종이(봉인)가 제대로 붙어 있으며, 화면 내에 인물이나 주안상은 보이지 않음.",
    "hard_violations": [],
    "physics": "모든 건축물과 지형이 물리적으로 안정감 있게 바닥에 위치해 있으며, 떠 있거나 불가능한 형태의 사물은 없음."
   }
  ],
  "cross_model_order": {
   "policy": "cross_model_order_v2_slot0_fwd_slot1_rev_parallel2",
   "slots": [
    {
     "model": "gemini-pro",
     "order": "forward",
     "display_to_canonical": {
      "A": "A",
      "B": "B"
     },
     "raw": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 10,
        "verdict_ko": "지게가 사라진 텅 빈 마당의 모습, 곡간 문의 붉은 봉인, 열린 사랑채 문 등 프롬프트의 지시사항과 부재(absence)의 상태를 완벽하게 구현한 훌륭한 결과물입니다."
       },
       {
        "label": "A",
        "score": 3,
        "verdict_ko": "프롬프트의 핵심인 '나무 지게가 사라진 텅 빈 흙바닥'이라는 명시적인 지시에도 불구하고 담장에 지게를 그려 넣어 가장 중요한 요구사항을 위반했습니다."
       }
      ],
      "readings": [
       {
        "label": "B",
        "direction": "특정 방향성을 가지는 인물이나 무기, 이동하는 물체가 화면에 존재하지 않음.",
        "built_space": "참조 이미지와 동일한 구조의 마당을 보여줌. 왼쪽에는 사랑채의 기둥과 열려 있는 격자문이 배치되어 있으며, 중앙의 돌담과 오른쪽의 곡간 건물이 올바른 위치에 자연스럽게 묘사됨.",
        "entities": "프롬프트의 요구대로 마당 구석에 나무 지게가 존재하지 않음. 곡간 문 중앙에 붉은색 종이(봉인)가 제대로 붙어 있으며, 화면 내에 인물이나 주안상은 보이지 않음.",
        "hard_violations": [],
        "physics": "모든 건축물과 지형이 물리적으로 안정감 있게 바닥에 위치해 있으며, 떠 있거나 불가능한 형태의 사물은 없음."
       },
       {
        "label": "A",
        "direction": "특정 방향성을 가지는 인물이나 무기, 이동하는 물체가 화면에 존재하지 않음.",
        "built_space": "참조 이미지에 기반한 공간 구조(왼쪽 사랑채 기둥, 담장, 오른쪽 곡간)를 잘 재현함. 왼쪽 측면에 열린 사랑채 문이 보임.",
        "entities": "프롬프트에서 '사라졌다(absent)'고 명시한 나무 지게가 담장 곁에 그려져 있음. 곡간 문에 붉은색 자국(봉인)이 있음. 인물은 존재하지 않음.",
        "hard_violations": [
         "지게가 없는 빈 공간이 피사체가 되어야 한다는 지시(The carrying frame is absent)를 어기고 나무 지게를 화면 중앙에 배치함"
        ],
        "physics": "담장에 기대어 있는 지게를 포함하여, 물리 법칙을 거스르거나 지지대 없이 떠 있는 사물은 발견되지 않음."
       }
      ],
      "all_candidates_fail": false
     },
     "normalized": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 10,
        "verdict_ko": "지게가 사라진 텅 빈 마당의 모습, 곡간 문의 붉은 봉인, 열린 사랑채 문 등 프롬프트의 지시사항과 부재(absence)의 상태를 완벽하게 구현한 훌륭한 결과물입니다."
       },
       {
        "label": "A",
        "score": 3,
        "verdict_ko": "프롬프트의 핵심인 '나무 지게가 사라진 텅 빈 흙바닥'이라는 명시적인 지시에도 불구하고 담장에 지게를 그려 넣어 가장 중요한 요구사항을 위반했습니다."
       }
      ],
      "all_candidates_fail": false,
      "readings": [
       {
        "label": "B",
        "direction": "특정 방향성을 가지는 인물이나 무기, 이동하는 물체가 화면에 존재하지 않음.",
        "built_space": "참조 이미지와 동일한 구조의 마당을 보여줌. 왼쪽에는 사랑채의 기둥과 열려 있는 격자문이 배치되어 있으며, 중앙의 돌담과 오른쪽의 곡간 건물이 올바른 위치에 자연스럽게 묘사됨.",
        "entities": "프롬프트의 요구대로 마당 구석에 나무 지게가 존재하지 않음. 곡간 문 중앙에 붉은색 종이(봉인)가 제대로 붙어 있으며, 화면 내에 인물이나 주안상은 보이지 않음.",
        "hard_violations": [],
        "physics": "모든 건축물과 지형이 물리적으로 안정감 있게 바닥에 위치해 있으며, 떠 있거나 불가능한 형태의 사물은 없음."
       },
       {
        "label": "A",
        "direction": "특정 방향성을 가지는 인물이나 무기, 이동하는 물체가 화면에 존재하지 않음.",
        "built_space": "참조 이미지에 기반한 공간 구조(왼쪽 사랑채 기둥, 담장, 오른쪽 곡간)를 잘 재현함. 왼쪽 측면에 열린 사랑채 문이 보임.",
        "entities": "프롬프트에서 '사라졌다(absent)'고 명시한 나무 지게가 담장 곁에 그려져 있음. 곡간 문에 붉은색 자국(봉인)이 있음. 인물은 존재하지 않음.",
        "hard_violations": [
         "지게가 없는 빈 공간이 피사체가 되어야 한다는 지시(The carrying frame is absent)를 어기고 나무 지게를 화면 중앙에 배치함"
        ],
        "physics": "담장에 기대어 있는 지게를 포함하여, 물리 법칙을 거스르거나 지지대 없이 떠 있는 사물은 발견되지 않음."
       }
      ]
     },
     "ok": true
    },
    {
     "model": "gpt-high",
     "order": "reverse",
     "display_to_canonical": {
      "A": "B",
      "B": "A"
     },
     "raw": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 9,
        "verdict_ko": "지게 없는 흙바닥을 와이드 숏의 중심으로 삼고 어두운 밤과 참조 장소를 유지해, 빈자리 자체가 주제라는 핵심 지시를 충실히 구현한다."
       },
       {
        "label": "B",
        "score": 2,
        "verdict_ko": "밤의 마당과 건축은 맞지만, 사라져야 할 나무 지게를 담장에 세워 두어 장면의 핵심인 부재를 정면으로 위반한다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "사람이나 시선, 무기, 이동 물체는 없다. 카메라는 왼쪽 툇마루 곁에서 빈 흙마당과 맞은편 담장을 향하며, 곡간 문 정면이 보인다.",
        "built_space": "왼쪽에는 전경의 굵은 기둥과 뒤쪽 기둥, 처마 및 마루 가장자리가 보인다. 뒤에는 기와를 얹은 돌·흙 담장, 지붕이 있는 양짝문 한 곳, 그 앞 기단과 왼쪽 돌덩이 하나가 있다. 참조의 주요 구조와 배치가 유지되며, 빈 흙바닥이 화면 중앙과 아래를 넓게 차지한다. 사랑채 문의 개방 여부는 왼쪽의 가림과 어둠 때문에 확정하기 어렵다.",
        "entities": "사람, 얼굴, 지게, 술상은 없다. 흙마당과 목재 기둥, 기와, 돌담은 참조 장소에 부합한다. 닫힌 곡간 문 중앙에는 붉은 봉인처럼 보이는 직사각형 표식이 있으며 읽을 수 있는 글자는 없다. 바닥에는 자연스러운 요철만 보이고 뚜렷한 발자국이나 지게 윤곽의 흔적은 없다. 밤의 어두운 주변광이며 외부로 쏟아지는 실내 조명은 없다.",
        "hard_violations": [],
        "physics": "기둥은 돌 주초 위에, 마루는 건물 구조 위에 놓여 있다. 문과 지붕은 기둥 및 기단으로 지지되고 돌덩이는 흙바닥에 놓여 있다. 붉은 봉인은 문 표면에 붙어 있으며, 떠 있거나 지지 없는 물체는 없다."
       },
       {
        "label": "B",
        "direction": "사람이나 시선, 무기, 이동 물체는 없다. 카메라는 마루 곁에서 맞은편 문과 담장을 향한다. 나무 지게는 세로로 서서 윗부분이 담장 쪽으로 기울어 있다.",
        "built_space": "왼쪽에 목재 기둥 두 개와 처마, 마루 가장자리, 열린 것으로 보이는 격자문 한 짝이 있다. 뒤에는 기와 덮인 돌·흙 담장과 지붕 달린 양짝문 한 곳, 문 앞 기단, 왼쪽 돌덩이 하나가 보인다. 참조의 공간 구성은 대체로 유지되지만, 비어 있어야 할 담장 앞 구석에 지게가 놓여 있다.",
        "entities": "사람과 얼굴, 술상은 없다. 그러나 곡간 왼쪽에 긴 세로대와 가로대를 갖춘 나무 지게가 명확히 보인다. 곡간 문에는 붉은 봉인 같은 불규칙한 표식이 있고 읽을 수 있는 글자는 없다. 흙바닥과 한옥 재료, 어두운 밤의 분위기는 요구와 맞는다.",
        "hard_violations": [
         "없어야 한다고 명시한 나무 지게를 마당 구석에 배치하여, 빈자리만 보여야 하는 장면에 금지된 소품을 추가했다."
        ],
        "physics": "지게의 두 하단은 흙바닥에 닿고 윗부분은 담장에 기대어 있어 지지는 물리적으로 가능하다. 건물 기둥과 기단은 지면에 놓이고, 지붕은 문 구조 위에 얹혀 있다. 돌덩이도 바닥에 놓여 있으며 지지 없이 떠 있는 물체는 없다."
       }
      ],
      "all_candidates_fail": false
     },
     "normalized": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 9,
        "verdict_ko": "지게 없는 흙바닥을 와이드 숏의 중심으로 삼고 어두운 밤과 참조 장소를 유지해, 빈자리 자체가 주제라는 핵심 지시를 충실히 구현한다."
       },
       {
        "label": "A",
        "score": 2,
        "verdict_ko": "밤의 마당과 건축은 맞지만, 사라져야 할 나무 지게를 담장에 세워 두어 장면의 핵심인 부재를 정면으로 위반한다."
       }
      ],
      "all_candidates_fail": false,
      "readings": [
       {
        "label": "B",
        "direction": "사람이나 시선, 무기, 이동 물체는 없다. 카메라는 왼쪽 툇마루 곁에서 빈 흙마당과 맞은편 담장을 향하며, 곡간 문 정면이 보인다.",
        "built_space": "왼쪽에는 전경의 굵은 기둥과 뒤쪽 기둥, 처마 및 마루 가장자리가 보인다. 뒤에는 기와를 얹은 돌·흙 담장, 지붕이 있는 양짝문 한 곳, 그 앞 기단과 왼쪽 돌덩이 하나가 있다. 참조의 주요 구조와 배치가 유지되며, 빈 흙바닥이 화면 중앙과 아래를 넓게 차지한다. 사랑채 문의 개방 여부는 왼쪽의 가림과 어둠 때문에 확정하기 어렵다.",
        "entities": "사람, 얼굴, 지게, 술상은 없다. 흙마당과 목재 기둥, 기와, 돌담은 참조 장소에 부합한다. 닫힌 곡간 문 중앙에는 붉은 봉인처럼 보이는 직사각형 표식이 있으며 읽을 수 있는 글자는 없다. 바닥에는 자연스러운 요철만 보이고 뚜렷한 발자국이나 지게 윤곽의 흔적은 없다. 밤의 어두운 주변광이며 외부로 쏟아지는 실내 조명은 없다.",
        "hard_violations": [],
        "physics": "기둥은 돌 주초 위에, 마루는 건물 구조 위에 놓여 있다. 문과 지붕은 기둥 및 기단으로 지지되고 돌덩이는 흙바닥에 놓여 있다. 붉은 봉인은 문 표면에 붙어 있으며, 떠 있거나 지지 없는 물체는 없다."
       },
       {
        "label": "A",
        "direction": "사람이나 시선, 무기, 이동 물체는 없다. 카메라는 마루 곁에서 맞은편 문과 담장을 향한다. 나무 지게는 세로로 서서 윗부분이 담장 쪽으로 기울어 있다.",
        "built_space": "왼쪽에 목재 기둥 두 개와 처마, 마루 가장자리, 열린 것으로 보이는 격자문 한 짝이 있다. 뒤에는 기와 덮인 돌·흙 담장과 지붕 달린 양짝문 한 곳, 문 앞 기단, 왼쪽 돌덩이 하나가 보인다. 참조의 공간 구성은 대체로 유지되지만, 비어 있어야 할 담장 앞 구석에 지게가 놓여 있다.",
        "entities": "사람과 얼굴, 술상은 없다. 그러나 곡간 왼쪽에 긴 세로대와 가로대를 갖춘 나무 지게가 명확히 보인다. 곡간 문에는 붉은 봉인 같은 불규칙한 표식이 있고 읽을 수 있는 글자는 없다. 흙바닥과 한옥 재료, 어두운 밤의 분위기는 요구와 맞는다.",
        "hard_violations": [
         "없어야 한다고 명시한 나무 지게를 마당 구석에 배치하여, 빈자리만 보여야 하는 장면에 금지된 소품을 추가했다."
        ],
        "physics": "지게의 두 하단은 흙바닥에 닿고 윗부분은 담장에 기대어 있어 지지는 물리적으로 가능하다. 건물 기둥과 기단은 지면에 놓이고, 지붕은 문 구조 위에 얹혀 있다. 돌덩이도 바닥에 놓여 있으며 지지 없이 떠 있는 물체는 없다."
       }
      ]
     },
     "ok": true
    }
   ],
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "slot_winner_match": true,
   "slot_winner": {
    "gemini-pro": "B",
    "gpt-high": "B"
   },
   "route": "cross_slot_agree"
  },
  "dual": {
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "normalized": {
    "A": 0.522,
    "B": 2.0
   },
   "adjusted": {
    "A": 0.272,
    "B": 2.0
   },
   "violations": {
    "A": [
     "[gemini-pro] 지게가 없는 빈 공간이 피사체가 되어야 한다는 지시(The carrying frame is absent)를 어기고 나무 지게를 화면 중앙에 배치함",
     "[gpt-high] 없어야 한다고 명시한 나무 지게를 마당 구석에 배치하여, 빈자리만 보여야 하는 장면에 금지된 소품을 추가했다."
    ]
   },
   "per_model_winner": {
    "gemini-pro": "B",
    "gpt-high": "B"
   },
   "agreed": true
  },
  "totals": {
   "B": 2000,
   "A": 272
  },
  "selected": "B",
  "ranking": [
   "B",
   "A"
  ],
  "verdicts": [
   {
    "label": "B",
    "score": 2000,
    "verdict_ko": "지게가 사라진 텅 빈 마당의 모습, 곡간 문의 붉은 봉인, 열린 사랑채 문 등 프롬프트의 지시사항과 부재(absence)의 상태를 완벽하게 구현한 훌륭한 결과물입니다."
   },
   {
    "label": "A",
    "score": 272,
    "verdict_ko": "프롬프트의 핵심인 '나무 지게가 사라진 텅 빈 흙바닥'이라는 명시적인 지시에도 불구하고 담장에 지게를 그려 넣어 가장 중요한 요구사항을 위반했습니다.  ★위반: [gemini-pro] 지게가 없는 빈 공간이 피사체가 되어야 한다는 지시(The carrying frame is absent)를 어기고 나무 지게를 화면 중앙에 배치함 / [gpt-high] 없어야 한다고 명시한 나무 지게를 마당 구석에 배치하여, 빈자리만 보여야 하는 장면에 금지된 소품을 추가했다."
   }
  ],
  "refs": [
   {
    "label": "LOCATION PHOTOGRAPH — the exact place of this shot: its architecture, materials, fixed features and lighting mood are spatial truth; stage the moment inside this place. Never copy its camera framing.",
    "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/db00517b-165b-4005-a287-f28936b59b95/episodes/272bd7af-9294-4490-98c6-88a40e9deacc/images/background_chain/L09B03.png",
    "asset_id": "9e7a40ea-733e-4a97-a54e-668135b02b9c",
    "role": "location_plate"
   }
  ],
  "critique_skipped": true,
  "shot_run_produced": false,
  "shot_run_uid": "06aa1414-f35a-78ef-8de5-a0fa90651a2c",
  "ref_mode": "플레이트만 (배경 전용)",
  "share_plan": {
   "ref_plan": "background"
  }
 },
 "S20sh8::cine": {
  "applied": true,
  "attempted_at": "2026-09-09T08:58:58.003299+00:00",
  "fingerprint": "baf2ad57cc254c55610112b34e4c9911e56574d3676aa961b72d0b71a6cc0fda",
  "fingerprint_version": 2,
  "provider": "mai",
  "endpoint": "openrouter/chat-completions",
  "model": "microsoft/mai-image-2.6",
  "pack": "24.202608252115",
  "source_file": "S20sh8_sel.png",
  "source_sha256": "c8747884eda268d1cefe6aae2854e3cc6a081728b10b04983430027bcd348b9e",
  "file": "S20sh8_cine.png",
  "staged_sha256": "a3d5f2a20be14c319fde246270268ec02520e9614c7ca643ea5099b052c33fa1",
  "latency_ms": 22525
 },
 "S21sh1::signage": {
  "fp": "14627ba04692f577",
  "inscriptions": [],
  "cues": [],
  "dropped": []
 },
 "era_assess::d34e7ae3491d3cf1": {
  "subjects": [
   {
    "subject_native": "20세기 초 전라도 전통 농가 헛간 내부",
    "search_terms_native": [
     "전통 초가 헛간 내부",
     "일제강점기 농가 헛간",
     "근대 헛간 목재 골조"
    ],
    "language_lock_native": "모든 검색어는 반드시 한국어로만 작성해야 하며 다른 언어로 번역하거나 추가하지 마십시오.",
    "reason_ko": "일반적인 서구식 헛간이나 현대 창고 형태로 왜곡되기 쉬워 전통 가구조 목재 골조와 흙벽 구조의 고증이 필요합니다."
   }
  ],
  "subject_text": "산자락 외딴 헛간 내부\n산자락에 떨어져 있는 어두운 헛간 내부. 거친 목재 골조와 단순한 바닥, 바깥빛이 들어오는 출입문이 보인다.",
  "identity": "canonical",
  "scope_id": "L18",
  "scope_role": "location_interior",
  "scope_sha": "937b7bfb54a5ddd9"
 },
 "era_fail::7fd57ef13cbba43d": {
  "stage": "research",
  "subject": "20세기 초 전라도 전통 농가 헛간 내부",
  "terms": [
   "전통 초가 헛간 내부",
   "일제강점기 농가 헛간",
   "근대 헛간 목재 골조"
  ],
  "status": "no_usable",
  "queries": [
   [
    "전라도 전통 농가 헛간 목재 골조 초가",
    "국립민속박물관 헛간 농기구 흙바닥"
   ],
   [
    "\"헛간\" \"흙바닥\"",
    "\"헛간\" \"낙안\"",
    "\"부안 김상만 가옥\" \"헛간\""
   ],
   [
    "곳간 한국민족문화대백과사전 헛간 구조",
    "순천 낙안읍성 서문성벽집 헛간 토담 국가유산"
   ]
  ],
  "candidate_urls": [
   "https://minio.nculture.org/amsweb-opt/multimedia_assets/228/72628/66822/c/%EC%A0%84%EB%B6%81-%EB%B6%80%EC%95%88%EA%B5%B0-%EB%B6%80%EC%95%88-%EA%B9%80%EC%83%81%EB%A7%8C%EA%B0%80%EC%98%A5-%EC%95%88%EC%B1%84-%EB%82%B4%EB%B6%80-%281%29-medium-size.jpg",
   "https://photos.smugmug.com/USA/Georgia/Macon/i-KtfkmhV/0/L/Ocmulgee-8-L.jpg",
   "https://colonialkorea.com/wp-content/uploads/2015/02/img_7269.jpg?w=950",
   "https://assets.st-note.com/img/1737518124-rpBLK69CGxVI4oXW7PARNcJw.jpg?width=1200"
  ],
  "coarse": {
   "eligible": [],
   "chosen_index": 0,
   "reason": "종류·보임·기준을 다 만족하는 후보가 없다 — 이 라운드에선 안 고른다",
   "single_judge": true,
   "rejected_judges": {}
  },
  "verdicts": [
   {
    "index": 1,
    "object_type_match": "unsure",
    "visible": true,
    "criteria_match": "unsure",
    "similarity": 70
   },
   {
    "index": 2,
    "object_type_match": "no",
    "visible": true,
    "criteria_match": "unsure",
    "similarity": 20
   },
   {
    "index": 3,
    "object_type_match": "no",
    "visible": true,
    "criteria_match": "unsure",
    "similarity": 40
   },
   {
    "index": 4,
    "object_type_match": "no",
    "visible": true,
    "criteria_match": "unsure",
    "similarity": 20
   }
  ],
  "chosen_reason_ko": "",
  "attempts": 5
 },
 "S21sh1::bgfirst_bg": {
  "input_fingerprint": "7ec9fc10dadacae9",
  "prompt": "Create the EMPTY BACKGROUND PLATE for one film shot — NO PEOPLE, no figures, no body parts, no sketch lines, no arrows anywhere.\n\"Empty\" means no people only: KEEP the location's inherent occupants and stock that define the place — animals in an animal shelter, pen or farm, goods in a market, moored boats in a harbour — unless the shot text explicitly removes them.\nThe FIRST attached image is a thin-line storyboard sketch: use ONLY its camera angle, horizon, perspective and the placement/size of buildings and set masses — ignore the sketched people and arrows entirely. The SECOND attached image (LOCATION PHOTOGRAPH) is the real place: take its architecture, materials, signage and fixed features, and RE-PROJECT them into the sketch's camera. If the photograph's camera differs from the sketch's, the sketch's camera wins.\nHUMAN-SCALE CALIBRATION: derive every structure's true size from human-scale elements — a door ≈ 2m, a window ≈ 1–1.5m wide, one storey ≈ 2.5–3m; never inflate a small structure or shrink a large one.\n\nSHOT TEXT this background must serve (Korean): 어두운 헛간 바닥에 지친 기색으로 널브러진 안규홍, 이관회, 염재보, 복만과 세 명의 조선인 남성들의 전경.\n\nLOCATION (lock): The open floor inside an isolated hillside barn, dim despite the daytime outside.\n\nTIME OF DAY (lock): day.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: wide shot\n- KEY BACKGROUND ELEMENTS: 헛간 바닥 (Occupied by seven exhausted men resting in irregular positions); used as A continuous ground plane reveals different resting postures and unequal intervals between bodies; 헛간 문 (Not yet opened for 달래's entrance) — Seen from inside at an oblique angle along the right edge; used as Establishes the entrance for the subsequent movement without prematurely showing 달래.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Daytime ambient illumination remains subdued inside the explicitly dark barn, with naturalistic contrast and restrained saturation that keep the exhausted bodies readable.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nRender ONE photorealistic empty location photograph, 16:9, that this shot can be staged inside later. No readable writing anywhere: surfaces that would carry writing may be present, but stage any wording out of legibility — an oblique angle, distance, shallow focus. No captions, watermarks or overlay text.",
  "effective_prompt": "Create the EMPTY BACKGROUND PLATE for one film shot — NO PEOPLE, no figures, no body parts, no sketch lines, no arrows anywhere.\n\"Empty\" means no people only: KEEP the location's inherent occupants and stock that define the place — animals in an animal shelter, pen or farm, goods in a market, moored boats in a harbour — unless the shot text explicitly removes them.\nThe FIRST attached image is a thin-line storyboard sketch: use ONLY its camera angle, horizon, perspective and the placement/size of buildings and set masses — ignore the sketched people and arrows entirely. The SECOND attached image (LOCATION PHOTOGRAPH) is the real place: take its architecture, materials, signage and fixed features, and RE-PROJECT them into the sketch's camera. If the photograph's camera differs from the sketch's, the sketch's camera wins.\nHUMAN-SCALE CALIBRATION: derive every structure's true size from human-scale elements — a door ≈ 2m, a window ≈ 1–1.5m wide, one storey ≈ 2.5–3m; never inflate a small structure or shrink a large one.\n\nSHOT TEXT this background must serve (Korean): 어두운 헛간 바닥에 지친 기색으로 널브러진 안규홍, 이관회, 염재보, 복만과 세 명의 조선인 남성들의 전경.\n\nLOCATION (lock): The open floor inside an isolated hillside barn, dim despite the daytime outside.\n\nTIME OF DAY (lock): day.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: wide shot\n- KEY BACKGROUND ELEMENTS: 헛간 바닥 (Occupied by seven exhausted men resting in irregular positions); used as A continuous ground plane reveals different resting postures and unequal intervals between bodies; 헛간 문 (Not yet opened for 달래's entrance) — Seen from inside at an oblique angle along the right edge; used as Establishes the entrance for the subsequent movement without prematurely showing 달래.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Daytime ambient illumination remains subdued inside the explicitly dark barn, with naturalistic contrast and restrained saturation that keep the exhausted bodies readable.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nRender ONE photorealistic empty location photograph, 16:9, that this shot can be staged inside later. No readable writing anywhere: surfaces that would carry writing may be present, but stage any wording out of legibility — an oblique angle, distance, shallow focus. No captions, watermarks or overlay text.",
  "bg_path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/db00517b-165b-4005-a287-f28936b59b95/images/272bd7af-9294-4490-98c6-88a40e9deacc/scene/recipe/S21sh1__bgfirst_bg.png",
  "asset_id": "3375f1f5-42f2-49ad-8af2-799fff54a331",
  "input_asset_ids": [
   "f7c5032e-6432-4695-86f3-7827a0cdf745",
   "b0df7402-b140-4d36-b1f5-a319a54c9d77"
  ]
 },
 "S21sh1": {
  "input_fingerprint": "b3429a10528b3af9",
  "prompt": "Create ONE FINAL photorealistic live-action film still of the moment below — early 20th-century Korea, Jeolla region; all people are Korean unless explicitly stated otherwise. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 어두운 헛간 바닥에 지친 기색으로 널브러진 안규홍, 이관회, 염재보, 복만과 세 명의 조선인 남성들의 전경.\n\nLOCATION (lock): The open floor inside an isolated hillside barn, dim despite the daytime outside. The shot takes place here — the attached LOCATION PHOTOGRAPH shows the exact spot.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: wide shot\n- KEY BACKGROUND ELEMENTS: 헛간 바닥 (Occupied by seven exhausted men resting in irregular positions); used as A continuous ground plane reveals different resting postures and unequal intervals between bodies; 헛간 문 (Not yet opened for 달래's entrance) — Seen from inside at an oblique angle along the right edge; used as Establishes the entrance for the subsequent movement without prematurely showing 달래.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Daytime ambient illumination remains subdued inside the explicitly dark barn, with naturalistic contrast and restrained saturation that keep the exhausted bodies readable.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The isolated barn is dim. 안규홍: He sprawls exhausted with one eye badly swollen and his lip split; the jade ring remains concealed in his clothing. 이관회: He sprawls exhausted on the barn floor. 염재보: He sprawls exhausted on the barn floor. 복만: He sprawls exhausted in his worn durumagi, retaining his matchlock.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 안규홍 (한국인 남성, 청년 나이대, 검은 머리, 단단히 묶은 머리); 이관회 (한국인 남성, 청년 나이대, 자연스러운 검은 머리); 염재보 (한국인 남성, 성인 나이대, 자연스러운 검은 머리); 복만 (한국인 남성, 성인 나이대, 거칠게 자란 수염, 자연스러운 검은 머리) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.",
  "roll_prompts": {
   "A": "Stage the shot. The FIRST attached image (SHOT BACKGROUND) is the finished empty background of this shot — keep it EXACTLY: its camera, perspective, architecture, lighting and every fixed feature stay untouched. The SECOND attached image (LAYOUT SKETCH) tells you ONLY where the people go: each sketched person's position, screen size, pose and the gaze/motion arrows. Ignore the sketch's background lines. The CHARACTER REFERENCE photographs show the real people.\nPlace the real people into the background at exactly the sketched positions, sizes and poses, following the arrow directions. No sketch lines or arrows may remain.\n\nCreate ONE FINAL photorealistic live-action film still of the moment below — early 20th-century Korea, Jeolla region; all people are Korean unless explicitly stated otherwise. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 어두운 헛간 바닥에 지친 기색으로 널브러진 안규홍, 이관회, 염재보, 복만과 세 명의 조선인 남성들의 전경.\n\nLOCATION (lock): The open floor inside an isolated hillside barn, dim despite the daytime outside. The shot takes place here — the FIRST attached image (SHOT BACKGROUND) is this exact place, already built: its ground, structures, horizon, materials and lighting are the finished truth of this location and must not be redesigned or replaced. No location photograph is attached — read the place from that image alone, and add no scenery, structure, vehicle or fixture that it does not already show. This lock governs the place only; the figures in the shot follow the staging and pose instructions.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: wide shot\n- KEY BACKGROUND ELEMENTS: 헛간 바닥 (Occupied by seven exhausted men resting in irregular positions); used as A continuous ground plane reveals different resting postures and unequal intervals between bodies; 헛간 문 (Not yet opened for 달래's entrance) — Seen from inside at an oblique angle along the right edge; used as Establishes the entrance for the subsequent movement without prematurely showing 달래.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Daytime ambient illumination remains subdued inside the explicitly dark barn, with naturalistic contrast and restrained saturation that keep the exhausted bodies readable.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The isolated barn is dim. 안규홍: He sprawls exhausted with one eye badly swollen and his lip split; the jade ring remains concealed in his clothing. 이관회: He sprawls exhausted on the barn floor. 염재보: He sprawls exhausted on the barn floor. 복만: He sprawls exhausted in his worn durumagi, retaining his matchlock.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 안규홍 (한국인 남성, 청년 나이대, 검은 머리, 단단히 묶은 머리); 이관회 (한국인 남성, 청년 나이대, 자연스러운 검은 머리); 염재보 (한국인 남성, 성인 나이대, 자연스러운 검은 머리); 복만 (한국인 남성, 성인 나이대, 거칠게 자란 수염, 자연스러운 검은 머리) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.",
   "B": "Create ONE FINAL photorealistic live-action film still of the moment below — early 20th-century Korea, Jeolla region; all people are Korean unless explicitly stated otherwise. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 어두운 헛간 바닥에 지친 기색으로 널브러진 안규홍, 이관회, 염재보, 복만과 세 명의 조선인 남성들의 전경.\n\nLOCATION (lock): The open floor inside an isolated hillside barn, dim despite the daytime outside. The shot takes place here — the attached LOCATION PHOTOGRAPH shows the exact spot.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: wide shot\n- KEY BACKGROUND ELEMENTS: 헛간 바닥 (Occupied by seven exhausted men resting in irregular positions); used as A continuous ground plane reveals different resting postures and unequal intervals between bodies; 헛간 문 (Not yet opened for 달래's entrance) — Seen from inside at an oblique angle along the right edge; used as Establishes the entrance for the subsequent movement without prematurely showing 달래.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Daytime ambient illumination remains subdued inside the explicitly dark barn, with naturalistic contrast and restrained saturation that keep the exhausted bodies readable.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The isolated barn is dim. 안규홍: He sprawls exhausted with one eye badly swollen and his lip split; the jade ring remains concealed in his clothing. 이관회: He sprawls exhausted on the barn floor. 염재보: He sprawls exhausted on the barn floor. 복만: He sprawls exhausted in his worn durumagi, retaining his matchlock.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 안규홍 (한국인 남성, 청년 나이대, 검은 머리, 단단히 묶은 머리); 이관회 (한국인 남성, 청년 나이대, 자연스러운 검은 머리); 염재보 (한국인 남성, 성인 나이대, 자연스러운 검은 머리); 복만 (한국인 남성, 성인 나이대, 거칠게 자란 수염, 자연스러운 검은 머리) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.\n\nCOMPOSITION VARIATION (this candidate only): keep the place, the people and the moment exactly as specified, and keep from the CAMERA & FRAME contract its subject scale, screen placement and key background elements exactly. For this candidate only, its camera angle and height are deliberately left open: choose a camera position distinctly different from the obvious one, as a film director picking a second setup on the same blocking. Stay true to any explicit framing the SHOT TEXT itself dictates."
  },
  "roll_refs": {
   "A": [
    {
     "label": "SHOT BACKGROUND",
     "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/db00517b-165b-4005-a287-f28936b59b95/images/272bd7af-9294-4490-98c6-88a40e9deacc/scene/recipe/S21sh1__bgfirst_bg.png",
     "asset_id": "3375f1f5-42f2-49ad-8af2-799fff54a331",
     "role": "bgfirst_bg"
    },
    {
     "label": "LAYOUT SKETCH (people placement only)",
     "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/db00517b-165b-4005-a287-f28936b59b95/images/272bd7af-9294-4490-98c6-88a40e9deacc/conti/conti_S21sh1.png",
     "asset_id": "f7c5032e-6432-4695-86f3-7827a0cdf745",
     "role": "conti_light"
    },
    {
     "label": "CHARACTER REFERENCE — 안규홍: the exact person appearing in this shot; match face, hair and build exactly.",
     "path": "<bytes:1028384>",
     "asset_id": "3b5894db-ce55-4f34-b966-5e725c3aa952",
     "role": "character_ref"
    },
    {
     "label": "CHARACTER REFERENCE — 이관회: the exact person appearing in this shot; match face, hair and build exactly.",
     "path": "<bytes:933540>",
     "asset_id": "851b450f-ac84-47fe-ac5a-03a56f63fcae",
     "role": "character_ref"
    },
    {
     "label": "CHARACTER REFERENCE — 염재보: the exact person appearing in this shot; match face, hair and build exactly.",
     "path": "<bytes:964351>",
     "asset_id": "ab287289-d0b7-436d-8f70-68452f898f73",
     "role": "character_ref"
    },
    {
     "label": "CHARACTER REFERENCE — 복만: the exact person appearing in this shot; match face, hair and build exactly.",
     "path": "<bytes:1027479>",
     "asset_id": "b7e69570-d15d-4241-b5ea-4cb8d4ece1cf",
     "role": "character_ref"
    }
   ],
   "B": [
    {
     "label": "LOCATION PHOTOGRAPH — the exact place of this shot: its architecture, materials, fixed features and lighting mood are spatial truth; stage the moment inside this place. Never copy its camera framing.",
     "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/db00517b-165b-4005-a287-f28936b59b95/episodes/272bd7af-9294-4490-98c6-88a40e9deacc/images/background_chain/L18B03.png",
     "asset_id": "b0df7402-b140-4d36-b1f5-a319a54c9d77",
     "role": "location_plate"
    },
    {
     "label": "CHARACTER REFERENCE — 안규홍: the exact person appearing in this shot; match face, hair and build exactly.",
     "path": "<bytes:1028384>",
     "asset_id": "3b5894db-ce55-4f34-b966-5e725c3aa952",
     "role": "character_ref"
    },
    {
     "label": "CHARACTER REFERENCE — 이관회: the exact person appearing in this shot; match face, hair and build exactly.",
     "path": "<bytes:933540>",
     "asset_id": "851b450f-ac84-47fe-ac5a-03a56f63fcae",
     "role": "character_ref"
    },
    {
     "label": "CHARACTER REFERENCE — 염재보: the exact person appearing in this shot; match face, hair and build exactly.",
     "path": "<bytes:964351>",
     "asset_id": "ab287289-d0b7-436d-8f70-68452f898f73",
     "role": "character_ref"
    },
    {
     "label": "CHARACTER REFERENCE — 복만: the exact person appearing in this shot; match face, hair and build exactly.",
     "path": "<bytes:1027479>",
     "asset_id": "b7e69570-d15d-4241-b5ea-4cb8d4ece1cf",
     "role": "character_ref"
    }
   ]
  },
  "shot_run_spend_attempt_count": 1,
  "judge_flip_skipped": "judge_fn_owns_order",
  "readings": [
   {
    "label": "A",
    "direction": "복만(좌측 상단)의 시선은 정면을 향하고 있으며, 그가 쥔 화승총의 총구는 화면 우측을 향하고 있습니다. 나머지 인물들은 고개를 숙이거나 눈을 감고 있습니다.",
    "built_space": "레퍼런스 사진과 정확히 일치하는 흙바닥, 나무 기둥, 후면의 작은 쇠창살 창문과 우측 나무문이 구현된 헛간입니다. 7명의 남성이 흙바닥 위 곳곳에 불규칙하게 흩어져 있습니다. 화면 양 측면에 문틀 역할을 하는 나무 구조물이 보입니다.",
    "entities": "총 7명의 남성이 확인됩니다. 안규홍(중앙 상단)과 이관회(중앙)는 흰옷을 입고 있으며, 복만(좌측 상단)은 화승총을 들고 있으나 레퍼런스와 달리 검은 갓을 쓰고 있습니다. 염재보로 추정되는 인물을 포함해 지친 상태의 남성들이 바닥에 널브러져 있습니다.",
    "hard_violations": [
     "[gpt-high] 출입문을 열기 전의 장면이어야 하지만, 전경 출입구가 열린 채 문턱 너머로 실내가 보인다."
    ],
    "physics": "모든 인물들은 흙바닥이나 벽면에 체중을 싣고 자연스럽게 누워있거나 기대어 쉬고 있습니다."
   },
   {
    "label": "B",
    "direction": "인물들의 시선은 각기 다른 방향을 향하거나 감겨 있습니다. 화면 우측 전경에 있는 복만이 안은 화승총의 총구는 우측 하단 바닥을 향합니다.",
    "built_space": "레퍼런스와 달리 후면에 2개의 쇠창살 창문이 있고, 우측 벽에 거대한 여닫이 나무문이 추가되어 공간 구조가 변형되었습니다. 바닥에는 프롬프트와 레퍼런스에 없는 커다란 건초더미들이 쌓여 있습니다.",
    "entities": "총 7명이 등장해야 하나 6명의 남성만 존재하여 인원수가 누락되었습니다. 안규홍(중앙)과 복만(우측 전경)의 복장과 특징은 일부 확인되나 전체 인물 구성이 맞지 않습니다.",
    "hard_violations": [
     "[gemini-pro] 지시된 7명의 인원수 중 1명 누락 (6명만 렌더링됨)",
     "[gemini-pro] 레퍼런스에 없는 건초더미 임의 생성",
     "[gemini-pro] 레퍼런스와 구조가 다른 헛간 (창문 추가, 우측 나무문 변경)",
     "[gpt-high] 참고 장소에 하나인 창살창을 두 개로 늘렸다.",
     "[gpt-high] 아직 열리지 않아야 하는 오른쪽 출입문이 열린 상태다."
    ],
    "physics": "인물들은 바닥과 추가된 건초더미 위에 몸을 지탱하고 기대어 누워 있습니다."
   }
  ],
  "cross_model_order": {
   "policy": "cross_model_order_v2_slot0_fwd_slot1_rev_parallel2",
   "slots": [
    {
     "model": "gemini-pro",
     "order": "forward",
     "display_to_canonical": {
      "A": "A",
      "B": "B"
     },
     "raw": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 7,
        "verdict_ko": "제시된 인원수(7명)와 레퍼런스 공간의 구조를 정확히 구현했으나, 복만의 모자가 다르고 요구된 어두운 헛간의 조명 분위기보다 다소 밝게 묘사되었습니다."
       },
       {
        "label": "B",
        "score": 3,
        "verdict_ko": "등장해야 할 7명 중 6명만 그려졌으며, 레퍼런스에 없는 건초더미와 추가 창문을 임의로 생성하여 공간적 사실성을 위반했습니다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "복만(좌측 상단)의 시선은 정면을 향하고 있으며, 그가 쥔 화승총의 총구는 화면 우측을 향하고 있습니다. 나머지 인물들은 고개를 숙이거나 눈을 감고 있습니다.",
        "built_space": "레퍼런스 사진과 정확히 일치하는 흙바닥, 나무 기둥, 후면의 작은 쇠창살 창문과 우측 나무문이 구현된 헛간입니다. 7명의 남성이 흙바닥 위 곳곳에 불규칙하게 흩어져 있습니다. 화면 양 측면에 문틀 역할을 하는 나무 구조물이 보입니다.",
        "entities": "총 7명의 남성이 확인됩니다. 안규홍(중앙 상단)과 이관회(중앙)는 흰옷을 입고 있으며, 복만(좌측 상단)은 화승총을 들고 있으나 레퍼런스와 달리 검은 갓을 쓰고 있습니다. 염재보로 추정되는 인물을 포함해 지친 상태의 남성들이 바닥에 널브러져 있습니다.",
        "hard_violations": [],
        "physics": "모든 인물들은 흙바닥이나 벽면에 체중을 싣고 자연스럽게 누워있거나 기대어 쉬고 있습니다."
       },
       {
        "label": "B",
        "direction": "인물들의 시선은 각기 다른 방향을 향하거나 감겨 있습니다. 화면 우측 전경에 있는 복만이 안은 화승총의 총구는 우측 하단 바닥을 향합니다.",
        "built_space": "레퍼런스와 달리 후면에 2개의 쇠창살 창문이 있고, 우측 벽에 거대한 여닫이 나무문이 추가되어 공간 구조가 변형되었습니다. 바닥에는 프롬프트와 레퍼런스에 없는 커다란 건초더미들이 쌓여 있습니다.",
        "entities": "총 7명이 등장해야 하나 6명의 남성만 존재하여 인원수가 누락되었습니다. 안규홍(중앙)과 복만(우측 전경)의 복장과 특징은 일부 확인되나 전체 인물 구성이 맞지 않습니다.",
        "hard_violations": [
         "지시된 7명의 인원수 중 1명 누락 (6명만 렌더링됨)",
         "레퍼런스에 없는 건초더미 임의 생성",
         "레퍼런스와 구조가 다른 헛간 (창문 추가, 우측 나무문 변경)"
        ],
        "physics": "인물들은 바닥과 추가된 건초더미 위에 몸을 지탱하고 기대어 누워 있습니다."
       }
      ],
      "all_candidates_fail": false
     },
     "normalized": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 7,
        "verdict_ko": "제시된 인원수(7명)와 레퍼런스 공간의 구조를 정확히 구현했으나, 복만의 모자가 다르고 요구된 어두운 헛간의 조명 분위기보다 다소 밝게 묘사되었습니다."
       },
       {
        "label": "B",
        "score": 3,
        "verdict_ko": "등장해야 할 7명 중 6명만 그려졌으며, 레퍼런스에 없는 건초더미와 추가 창문을 임의로 생성하여 공간적 사실성을 위반했습니다."
       }
      ],
      "all_candidates_fail": false,
      "readings": [
       {
        "label": "A",
        "direction": "복만(좌측 상단)의 시선은 정면을 향하고 있으며, 그가 쥔 화승총의 총구는 화면 우측을 향하고 있습니다. 나머지 인물들은 고개를 숙이거나 눈을 감고 있습니다.",
        "built_space": "레퍼런스 사진과 정확히 일치하는 흙바닥, 나무 기둥, 후면의 작은 쇠창살 창문과 우측 나무문이 구현된 헛간입니다. 7명의 남성이 흙바닥 위 곳곳에 불규칙하게 흩어져 있습니다. 화면 양 측면에 문틀 역할을 하는 나무 구조물이 보입니다.",
        "entities": "총 7명의 남성이 확인됩니다. 안규홍(중앙 상단)과 이관회(중앙)는 흰옷을 입고 있으며, 복만(좌측 상단)은 화승총을 들고 있으나 레퍼런스와 달리 검은 갓을 쓰고 있습니다. 염재보로 추정되는 인물을 포함해 지친 상태의 남성들이 바닥에 널브러져 있습니다.",
        "hard_violations": [],
        "physics": "모든 인물들은 흙바닥이나 벽면에 체중을 싣고 자연스럽게 누워있거나 기대어 쉬고 있습니다."
       },
       {
        "label": "B",
        "direction": "인물들의 시선은 각기 다른 방향을 향하거나 감겨 있습니다. 화면 우측 전경에 있는 복만이 안은 화승총의 총구는 우측 하단 바닥을 향합니다.",
        "built_space": "레퍼런스와 달리 후면에 2개의 쇠창살 창문이 있고, 우측 벽에 거대한 여닫이 나무문이 추가되어 공간 구조가 변형되었습니다. 바닥에는 프롬프트와 레퍼런스에 없는 커다란 건초더미들이 쌓여 있습니다.",
        "entities": "총 7명이 등장해야 하나 6명의 남성만 존재하여 인원수가 누락되었습니다. 안규홍(중앙)과 복만(우측 전경)의 복장과 특징은 일부 확인되나 전체 인물 구성이 맞지 않습니다.",
        "hard_violations": [
         "지시된 7명의 인원수 중 1명 누락 (6명만 렌더링됨)",
         "레퍼런스에 없는 건초더미 임의 생성",
         "레퍼런스와 구조가 다른 헛간 (창문 추가, 우측 나무문 변경)"
        ],
        "physics": "인물들은 바닥과 추가된 건초더미 위에 몸을 지탱하고 기대어 누워 있습니다."
       }
      ]
     },
     "ok": true
    },
    {
     "model": "gpt-high",
     "order": "reverse",
     "display_to_canonical": {
      "A": "B",
      "B": "A"
     },
     "raw": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 3,
        "verdict_ko": "일곱 남성의 탈진과 복만의 조총 보유는 잘 보이지만, 창살창을 추가하고 오른쪽 출입문을 열어 장소와 입장 직전의 설정을 위반한다."
       },
       {
        "label": "B",
        "score": 5,
        "verdict_ko": "흙바닥·흙벽·목재 골조와 일곱 남성의 불규칙한 휴식 배치는 더 충실하지만, 열린 입구 너머의 구도로 닫힌 문을 내부 오른쪽 가장자리에 보여주라는 조건을 충족하지 못한다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "전경 왼쪽 남성은 팔로 눈을 가리고 얼굴을 위로 향한다. 흰옷 남성과 오른쪽 수염 난 남성은 눈을 감거나 내리며, 중앙 짙은 옷 남성은 무릎 쪽으로 고개를 숙인다. 뒤쪽 세 남성도 아래나 옆을 향해 쉬며 카메라를 뚜렷이 응시하지 않는다. 복만으로 보이는 남성의 총구는 화면 오른쪽 위를 향하며 특정 사람을 겨누지 않는다.",
        "built_space": "흙바닥 위에 전경 세 명, 중경 한 명, 후경 세 명으로 총 일곱 명이 놓여 있다. 간격과 자세가 서로 다르고 내부는 어둡다. 흙벽과 목재 골조는 유사하지만 창살창이 왼쪽 위와 오른쪽 위에 두 개 있어 참고의 한 개와 다르다. 뒤쪽 닫힌 판자문 외에 오른쪽 가장자리의 큰 출입문이 열려 있다. 참고의 빈 바닥과 달리 상당한 짚더미와 바닥의 천류도 보인다.",
        "entities": "일곱 명 모두 한국인 남성 설정에 부합하는 동아시아계 청년 또는 성인으로 보이며 달래는 없다. 흰옷과 묶은 검은 머리의 안규홍 후보, 갈색 저고리와 밝은 바지의 이관회 후보, 짙은 옷과 각반의 염재보 후보, 수염과 낡은 겉옷을 갖춘 복만 후보가 보인다. 염재보의 모자는 참고의 넓은 삿갓과 다르고 복만은 참고의 모자를 쓰지 않았다. 안규홍의 심하게 부은 한쪽 눈과 찢어진 입술은 명확히 식별되지 않는다. 옥가락지는 노출되지 않으며 복만은 긴 목제 총을 품고 있다. 읽을 수 있는 글자는 없다.",
        "hard_violations": [
         "참고 장소에 하나인 창살창을 두 개로 늘렸다.",
         "아직 열리지 않아야 하는 오른쪽 출입문이 열린 상태다."
        ],
        "physics": "누운 남성들은 몸통과 다리가 바닥 또는 짚에 닿고, 기대앉은 남성들은 엉덩이와 다리를 바닥에 두고 벽·짚더미·팔로 상체를 지탱한다. 복만의 총은 두 손과 몸, 바닥 가까이에 놓인 개머리판으로 지지된다. 공중에 뜬 인체나 무지지 물체는 보이지 않는다."
       },
       {
        "label": "B",
        "direction": "중앙에 드러누운 남성은 얼굴을 천장 쪽으로 향하고, 나머지 누운 남성들은 바닥이나 자기 팔 쪽으로 얼굴을 내린다. 중앙에 앉은 흰옷 남성도 아래를 본다. 왼쪽 벽에 앉은 복만 후보는 방 안쪽을 바라보며 총을 무릎 위에 놓았다. 총구는 화면 오른쪽 위의 빈 공간을 향하고 사람에게 조준되어 있지 않다.",
        "built_space": "한 개의 높은 창살창, 그 오른쪽의 닫힌 판자문, 흙벽과 가로·세로 목재 골조, 넓고 연속된 흙바닥이 참고 장소에 가깝다. 총 일곱 명이 왼쪽 벽가 두 명, 중앙 두 명, 오른쪽 중경 한 명, 전경 두 명으로 불균등하게 놓여 있다. 다만 카메라는 열린 입구의 문턱 너머에서 안을 보는 구도이며, 오른쪽 전경에는 닫힌 문 대신 입구의 목재 가장자리가 크게 보인다. 닫힌 문은 후경에 있다. 실내는 참고와 비슷한 자연광을 받지만 요청한 어두운 헛간치고 밝다.",
        "entities": "일곱 명 모두 한국인 남성 설정에 부합하는 동아시아계 청년 또는 성인으로 보이고 추가 여성은 없다. 중앙에 앉은 묶은 머리의 흰옷 남성이 안규홍, 갈색 저고리와 밝은 바지의 남성이 이관회, 전경 오른쪽의 짙은 옷과 각반을 착용한 남성이 염재보에 대응한다. 염재보의 삿갓은 보이지 않는다. 왼쪽 벽가 남성은 복만 참고와 비슷한 수염·모자·낡은 두루마기를 갖추고 긴 총을 보유한다. 안규홍 후보의 얼굴에 붉은 기운은 있지만 심하게 부은 한쪽 눈과 갈라진 입술까지 확실하지 않다. 옥가락지는 보이지 않고 읽을 수 있는 글자도 없다.",
        "hard_violations": [
         "출입문을 열기 전의 장면이어야 하지만, 전경 출입구가 열린 채 문턱 너머로 실내가 보인다."
        ],
        "physics": "옆으로 누운 남성들은 옆구리·골반·다리를 바닥에 대고 팔꿈치나 팔로 머리와 상체를 받친다. 중앙에 앉은 남성은 바닥에 짚은 손과 엉덩이로 체중을 지탱하고, 드러누운 남성은 등과 팔다리가 바닥에 놓인다. 복만은 엉덩이를 바닥에 둔 채 벽에 기대고 두 손과 무릎으로 총을 받친다. 부유하거나 물리적으로 불가능한 자세는 보이지 않지만 복만의 자세는 널브러진 모습보다 비교적 꼿꼿한 휴식에 가깝다."
       }
      ],
      "all_candidates_fail": true
     },
     "normalized": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 3,
        "verdict_ko": "일곱 남성의 탈진과 복만의 조총 보유는 잘 보이지만, 창살창을 추가하고 오른쪽 출입문을 열어 장소와 입장 직전의 설정을 위반한다."
       },
       {
        "label": "A",
        "score": 5,
        "verdict_ko": "흙바닥·흙벽·목재 골조와 일곱 남성의 불규칙한 휴식 배치는 더 충실하지만, 열린 입구 너머의 구도로 닫힌 문을 내부 오른쪽 가장자리에 보여주라는 조건을 충족하지 못한다."
       }
      ],
      "all_candidates_fail": true,
      "readings": [
       {
        "label": "B",
        "direction": "전경 왼쪽 남성은 팔로 눈을 가리고 얼굴을 위로 향한다. 흰옷 남성과 오른쪽 수염 난 남성은 눈을 감거나 내리며, 중앙 짙은 옷 남성은 무릎 쪽으로 고개를 숙인다. 뒤쪽 세 남성도 아래나 옆을 향해 쉬며 카메라를 뚜렷이 응시하지 않는다. 복만으로 보이는 남성의 총구는 화면 오른쪽 위를 향하며 특정 사람을 겨누지 않는다.",
        "built_space": "흙바닥 위에 전경 세 명, 중경 한 명, 후경 세 명으로 총 일곱 명이 놓여 있다. 간격과 자세가 서로 다르고 내부는 어둡다. 흙벽과 목재 골조는 유사하지만 창살창이 왼쪽 위와 오른쪽 위에 두 개 있어 참고의 한 개와 다르다. 뒤쪽 닫힌 판자문 외에 오른쪽 가장자리의 큰 출입문이 열려 있다. 참고의 빈 바닥과 달리 상당한 짚더미와 바닥의 천류도 보인다.",
        "entities": "일곱 명 모두 한국인 남성 설정에 부합하는 동아시아계 청년 또는 성인으로 보이며 달래는 없다. 흰옷과 묶은 검은 머리의 안규홍 후보, 갈색 저고리와 밝은 바지의 이관회 후보, 짙은 옷과 각반의 염재보 후보, 수염과 낡은 겉옷을 갖춘 복만 후보가 보인다. 염재보의 모자는 참고의 넓은 삿갓과 다르고 복만은 참고의 모자를 쓰지 않았다. 안규홍의 심하게 부은 한쪽 눈과 찢어진 입술은 명확히 식별되지 않는다. 옥가락지는 노출되지 않으며 복만은 긴 목제 총을 품고 있다. 읽을 수 있는 글자는 없다.",
        "hard_violations": [
         "참고 장소에 하나인 창살창을 두 개로 늘렸다.",
         "아직 열리지 않아야 하는 오른쪽 출입문이 열린 상태다."
        ],
        "physics": "누운 남성들은 몸통과 다리가 바닥 또는 짚에 닿고, 기대앉은 남성들은 엉덩이와 다리를 바닥에 두고 벽·짚더미·팔로 상체를 지탱한다. 복만의 총은 두 손과 몸, 바닥 가까이에 놓인 개머리판으로 지지된다. 공중에 뜬 인체나 무지지 물체는 보이지 않는다."
       },
       {
        "label": "A",
        "direction": "중앙에 드러누운 남성은 얼굴을 천장 쪽으로 향하고, 나머지 누운 남성들은 바닥이나 자기 팔 쪽으로 얼굴을 내린다. 중앙에 앉은 흰옷 남성도 아래를 본다. 왼쪽 벽에 앉은 복만 후보는 방 안쪽을 바라보며 총을 무릎 위에 놓았다. 총구는 화면 오른쪽 위의 빈 공간을 향하고 사람에게 조준되어 있지 않다.",
        "built_space": "한 개의 높은 창살창, 그 오른쪽의 닫힌 판자문, 흙벽과 가로·세로 목재 골조, 넓고 연속된 흙바닥이 참고 장소에 가깝다. 총 일곱 명이 왼쪽 벽가 두 명, 중앙 두 명, 오른쪽 중경 한 명, 전경 두 명으로 불균등하게 놓여 있다. 다만 카메라는 열린 입구의 문턱 너머에서 안을 보는 구도이며, 오른쪽 전경에는 닫힌 문 대신 입구의 목재 가장자리가 크게 보인다. 닫힌 문은 후경에 있다. 실내는 참고와 비슷한 자연광을 받지만 요청한 어두운 헛간치고 밝다.",
        "entities": "일곱 명 모두 한국인 남성 설정에 부합하는 동아시아계 청년 또는 성인으로 보이고 추가 여성은 없다. 중앙에 앉은 묶은 머리의 흰옷 남성이 안규홍, 갈색 저고리와 밝은 바지의 남성이 이관회, 전경 오른쪽의 짙은 옷과 각반을 착용한 남성이 염재보에 대응한다. 염재보의 삿갓은 보이지 않는다. 왼쪽 벽가 남성은 복만 참고와 비슷한 수염·모자·낡은 두루마기를 갖추고 긴 총을 보유한다. 안규홍 후보의 얼굴에 붉은 기운은 있지만 심하게 부은 한쪽 눈과 갈라진 입술까지 확실하지 않다. 옥가락지는 보이지 않고 읽을 수 있는 글자도 없다.",
        "hard_violations": [
         "출입문을 열기 전의 장면이어야 하지만, 전경 출입구가 열린 채 문턱 너머로 실내가 보인다."
        ],
        "physics": "옆으로 누운 남성들은 옆구리·골반·다리를 바닥에 대고 팔꿈치나 팔로 머리와 상체를 받친다. 중앙에 앉은 남성은 바닥에 짚은 손과 엉덩이로 체중을 지탱하고, 드러누운 남성은 등과 팔다리가 바닥에 놓인다. 복만은 엉덩이를 바닥에 둔 채 벽에 기대고 두 손과 무릎으로 총을 받친다. 부유하거나 물리적으로 불가능한 자세는 보이지 않지만 복만의 자세는 널브러진 모습보다 비교적 꼿꼿한 휴식에 가깝다."
       }
      ]
     },
     "ok": true
    }
   ],
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "slot_winner_match": true,
   "slot_winner": {
    "gemini-pro": "A",
    "gpt-high": "A"
   },
   "route": "cross_slot_agree"
  },
  "dual": {
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "normalized": {
    "A": 2.0,
    "B": 1.029
   },
   "adjusted": {
    "A": 1.75,
    "B": 0.779
   },
   "violations": {
    "B": [
     "[gemini-pro] 지시된 7명의 인원수 중 1명 누락 (6명만 렌더링됨)",
     "[gemini-pro] 레퍼런스에 없는 건초더미 임의 생성",
     "[gemini-pro] 레퍼런스와 구조가 다른 헛간 (창문 추가, 우측 나무문 변경)",
     "[gpt-high] 참고 장소에 하나인 창살창을 두 개로 늘렸다.",
     "[gpt-high] 아직 열리지 않아야 하는 오른쪽 출입문이 열린 상태다."
    ],
    "A": [
     "[gpt-high] 출입문을 열기 전의 장면이어야 하지만, 전경 출입구가 열린 채 문턱 너머로 실내가 보인다."
    ]
   },
   "per_model_winner": {
    "gemini-pro": "A",
    "gpt-high": "A"
   },
   "agreed": true
  },
  "totals": {
   "A": 1750,
   "B": 779
  },
  "selected": "A",
  "ranking": [
   "A",
   "B"
  ],
  "verdicts": [
   {
    "label": "A",
    "score": 1750,
    "verdict_ko": "제시된 인원수(7명)와 레퍼런스 공간의 구조를 정확히 구현했으나, 복만의 모자가 다르고 요구된 어두운 헛간의 조명 분위기보다 다소 밝게 묘사되었습니다.  ★위반: [gpt-high] 출입문을 열기 전의 장면이어야 하지만, 전경 출입구가 열린 채 문턱 너머로 실내가 보인다."
   },
   {
    "label": "B",
    "score": 779,
    "verdict_ko": "등장해야 할 7명 중 6명만 그려졌으며, 레퍼런스에 없는 건초더미와 추가 창문을 임의로 생성하여 공간적 사실성을 위반했습니다.  ★위반: [gemini-pro] 지시된 7명의 인원수 중 1명 누락 (6명만 렌더링됨) / [gemini-pro] 레퍼런스에 없는 건초더미 임의 생성 / [gemini-pro] 레퍼런스와 구조가 다른 헛간 (창문 추가, 우측 나무문 변경) / [gpt-high] 참고 장소에 하나인 창살창을 두 개로 늘렸다. / [gpt-high] 아직 열리지 않아야 하는 오른쪽 출입문이 열린 상태다."
   }
  ],
  "refs": [
   {
    "label": "LOCATION PHOTOGRAPH — the exact place of this shot: its architecture, materials, fixed features and lighting mood are spatial truth; stage the moment inside this place. Never copy its camera framing.",
    "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/db00517b-165b-4005-a287-f28936b59b95/episodes/272bd7af-9294-4490-98c6-88a40e9deacc/images/background_chain/L18B03.png",
    "asset_id": "b0df7402-b140-4d36-b1f5-a319a54c9d77",
    "role": "location_plate"
   },
   {
    "label": "CHARACTER REFERENCE — 안규홍: the exact person appearing in this shot; match face, hair and build exactly.",
    "path": "<bytes:1028384>",
    "asset_id": "3b5894db-ce55-4f34-b966-5e725c3aa952",
    "role": "character_ref"
   },
   {
    "label": "CHARACTER REFERENCE — 이관회: the exact person appearing in this shot; match face, hair and build exactly.",
    "path": "<bytes:933540>",
    "asset_id": "851b450f-ac84-47fe-ac5a-03a56f63fcae",
    "role": "character_ref"
   },
   {
    "label": "CHARACTER REFERENCE — 염재보: the exact person appearing in this shot; match face, hair and build exactly.",
    "path": "<bytes:964351>",
    "asset_id": "ab287289-d0b7-436d-8f70-68452f898f73",
    "role": "character_ref"
   },
   {
    "label": "CHARACTER REFERENCE — 복만: the exact person appearing in this shot; match face, hair and build exactly.",
    "path": "<bytes:1027479>",
    "asset_id": "b7e69570-d15d-4241-b5ea-4cb8d4ece1cf",
    "role": "character_ref"
   }
  ],
  "critique_skipped": true,
  "shot_run_produced": false,
  "shot_run_uid": "06aa1414-f4e4-7b12-b6c4-fe1c7bc0dd4c",
  "bgfirst": {
   "bg_path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/db00517b-165b-4005-a287-f28936b59b95/images/272bd7af-9294-4490-98c6-88a40e9deacc/scene/recipe/S21sh1__bgfirst_bg.png",
   "bg_asset_id": "3375f1f5-42f2-49ad-8af2-799fff54a331",
   "bg_record_key": "S21sh1::bgfirst_bg",
   "chain_winner": true,
   "authority": "plate"
  },
  "ref_mode": "재투영 배경+콘티+엔티티 (2택1: 체인 승)",
  "share_plan": {
   "ref_plan": "background"
  }
 },
 "S21sh1::cine": {
  "applied": true,
  "attempted_at": "2026-09-09T09:00:12.552200+00:00",
  "fingerprint": "9c1ba551c3530bfc0dd0e61cf2c2b08f5c50ca271dce08fef895abf93a17983c",
  "fingerprint_version": 2,
  "provider": "mai",
  "endpoint": "openrouter/chat-completions",
  "model": "microsoft/mai-image-2.6",
  "pack": "24.202608252115",
  "source_file": "S21sh1_sel.png",
  "source_sha256": "39c5071172a4caaeecbc296a2f35403ef042ea0f7bf0b00e397139a1f52ba356",
  "file": "S21sh1_cine.png",
  "staged_sha256": "fe2e5e05254281378e0d1f2885a640af62c650f5558a93f7fd24609495d0b6c2",
  "latency_ms": 30076
 },
 "S21sh7::signage": {
  "fp": "0f4d5deb7fc7465d",
  "inscriptions": [],
  "cues": [],
  "dropped": []
 },
 "S21sh7": {
  "input_fingerprint": "660bad1c0226b764",
  "prompt": "Create ONE FINAL photorealistic live-action film still of the moment below — early 20th-century Korea, Jeolla region; all people are Korean unless explicitly stated otherwise. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 보자기가 열려 드러난 주먹밥들을 향해 여러 손이 일제히 뻗어 나온 찰나.\n\nLOCATION (lock): The floor-level food-sharing spot inside the dim hillside barn, with daytime light available through the opened entrance. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: close-up\n- KEY BACKGROUND ELEMENTS: Opened bundle (Untied and open, revealing rice balls) — The open interior faces upward toward the oblique camera; used as A compact central anchor for the inward movement of the hands; Rice balls (Exposed and not yet taken by the converging hands); used as Shared destination of the reaching gestures; Barn floor (Visible around the opened bundle); used as Surrounding space that preserves the scale of food and hands.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Preserve the barn's dim interior with restrained saturation and sufficient tonal separation to distinguish the overlapping hands.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The people visible in that photo are NOT in this shot: never carry their faces, bodies or clothing onto anyone here. Each person in THIS shot is defined solely by the PEOPLE list, their own CHARACTER REFERENCE images and their wardrobe notes. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nPREVIOUS STILL USAGE (follow exactly — what to take from the attached still and what to exclude; it governs place and objects only, never who is in this shot or how the camera sees them): Take the barn's floor, walls, structural materials, and dim daytime light visible in the reference. Exclude loose bedding or bundles occupying the spot where the food cloth is now opened.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The barn door is open, and an unwrapped cloth bundle exposes rice balls.\n\nNo people appear unless the shot text itself says so.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.",
  "roll_prompts": {
   "A": "Create ONE FINAL photorealistic live-action film still of the moment below — early 20th-century Korea, Jeolla region; all people are Korean unless explicitly stated otherwise. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 보자기가 열려 드러난 주먹밥들을 향해 여러 손이 일제히 뻗어 나온 찰나.\n\nLOCATION (lock): The floor-level food-sharing spot inside the dim hillside barn, with daytime light available through the opened entrance. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: close-up\n- KEY BACKGROUND ELEMENTS: Opened bundle (Untied and open, revealing rice balls) — The open interior faces upward toward the oblique camera; used as A compact central anchor for the inward movement of the hands; Rice balls (Exposed and not yet taken by the converging hands); used as Shared destination of the reaching gestures; Barn floor (Visible around the opened bundle); used as Surrounding space that preserves the scale of food and hands.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Preserve the barn's dim interior with restrained saturation and sufficient tonal separation to distinguish the overlapping hands.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The people visible in that photo are NOT in this shot: never carry their faces, bodies or clothing onto anyone here. Each person in THIS shot is defined solely by the PEOPLE list, their own CHARACTER REFERENCE images and their wardrobe notes. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nPREVIOUS STILL USAGE (follow exactly — what to take from the attached still and what to exclude; it governs place and objects only, never who is in this shot or how the camera sees them): Take the barn's floor, walls, structural materials, and dim daytime light visible in the reference. Exclude loose bedding or bundles occupying the spot where the food cloth is now opened.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The barn door is open, and an unwrapped cloth bundle exposes rice balls.\n\nNo people appear unless the shot text itself says so.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.",
   "B": "Create ONE FINAL photorealistic live-action film still of the moment below — early 20th-century Korea, Jeolla region; all people are Korean unless explicitly stated otherwise. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 보자기가 열려 드러난 주먹밥들을 향해 여러 손이 일제히 뻗어 나온 찰나.\n\nLOCATION (lock): The floor-level food-sharing spot inside the dim hillside barn, with daytime light available through the opened entrance. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: close-up\n- KEY BACKGROUND ELEMENTS: Opened bundle (Untied and open, revealing rice balls) — The open interior faces upward toward the oblique camera; used as A compact central anchor for the inward movement of the hands; Rice balls (Exposed and not yet taken by the converging hands); used as Shared destination of the reaching gestures; Barn floor (Visible around the opened bundle); used as Surrounding space that preserves the scale of food and hands.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Preserve the barn's dim interior with restrained saturation and sufficient tonal separation to distinguish the overlapping hands.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The people visible in that photo are NOT in this shot: never carry their faces, bodies or clothing onto anyone here. Each person in THIS shot is defined solely by the PEOPLE list, their own CHARACTER REFERENCE images and their wardrobe notes. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nPREVIOUS STILL USAGE (follow exactly — what to take from the attached still and what to exclude; it governs place and objects only, never who is in this shot or how the camera sees them): Take the barn's floor, walls, structural materials, and dim daytime light visible in the reference. Exclude loose bedding or bundles occupying the spot where the food cloth is now opened.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The barn door is open, and an unwrapped cloth bundle exposes rice balls.\n\nNo people appear unless the shot text itself says so.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.\n\nCOMPOSITION VARIATION (this candidate only): keep the place, the people and the moment exactly as specified, and keep from the CAMERA & FRAME contract its subject scale, screen placement and key background elements exactly. For this candidate only, its camera angle and height are deliberately left open: choose a camera position distinctly different from the obvious one, as a film director picking a second setup on the same blocking. Stay true to any explicit framing the SHOT TEXT itself dictates."
  },
  "shot_run_spend_attempt_count": 1,
  "readings": [
   {
    "label": "A",
    "direction": "인물들의 팔과 손이 중심에 놓인 주먹밥을 향하고 있음.",
    "built_space": "어두운 헛간의 흙바닥이 레퍼런스와 일치하게 묘사됨.",
    "entities": "녹색 보자기와 주먹밥, 전통 복식을 한 인물들이 확인됨.",
    "hard_violations": [],
    "physics": "인물들이 바닥에 쪼그려 앉아 있으며, 신체 지지가 자연스러움."
   },
   {
    "label": "B",
    "direction": "프레임 가장자리에서 중심의 주먹밥을 향해 다수의 손이 집중적으로 뻗어 나옴.",
    "built_space": "조명과 흙바닥의 질감이 지시된 헛간 환경과 잘 들어맞음.",
    "entities": "레퍼런스와 동일한 형태의 보자기, 주먹밥, 한복 및 짚신이 확인됨.",
    "hard_violations": [],
    "physics": "프레임 주변의 신체와 연결된 팔들이 올바른 궤적으로 뻗어 나옴."
   }
  ],
  "cross_model_order": {
   "policy": "cross_model_order_v2_slot0_fwd_slot1_rev_parallel2",
   "slots": [
    {
     "model": "gemini-pro",
     "order": "forward",
     "display_to_canonical": {
      "A": "A",
      "B": "B"
     },
     "raw": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 7,
        "verdict_ko": "클로즈업 프레이밍 지시를 정확히 따랐으며, 여러 손이 주먹밥을 향해 일제히 뻗어 나오는 역동적인 찰나를 훌륭하게 구현했습니다."
       },
       {
        "label": "A",
        "score": 4,
        "verdict_ko": "프레이밍이 너무 넓어 클로즈업 지시를 어겼으며, 손이 뻗어 나오는 찰나보다는 이미 손을 얹고 멈춰 있는 듯한 인상을 줍니다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "인물들의 팔과 손이 중심에 놓인 주먹밥을 향하고 있음.",
        "built_space": "어두운 헛간의 흙바닥이 레퍼런스와 일치하게 묘사됨.",
        "entities": "녹색 보자기와 주먹밥, 전통 복식을 한 인물들이 확인됨.",
        "hard_violations": [],
        "physics": "인물들이 바닥에 쪼그려 앉아 있으며, 신체 지지가 자연스러움."
       },
       {
        "label": "B",
        "direction": "프레임 가장자리에서 중심의 주먹밥을 향해 다수의 손이 집중적으로 뻗어 나옴.",
        "built_space": "조명과 흙바닥의 질감이 지시된 헛간 환경과 잘 들어맞음.",
        "entities": "레퍼런스와 동일한 형태의 보자기, 주먹밥, 한복 및 짚신이 확인됨.",
        "hard_violations": [],
        "physics": "프레임 주변의 신체와 연결된 팔들이 올바른 궤적으로 뻗어 나옴."
       }
      ],
      "all_candidates_fail": false
     },
     "normalized": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 7,
        "verdict_ko": "클로즈업 프레이밍 지시를 정확히 따랐으며, 여러 손이 주먹밥을 향해 일제히 뻗어 나오는 역동적인 찰나를 훌륭하게 구현했습니다."
       },
       {
        "label": "A",
        "score": 4,
        "verdict_ko": "프레이밍이 너무 넓어 클로즈업 지시를 어겼으며, 손이 뻗어 나오는 찰나보다는 이미 손을 얹고 멈춰 있는 듯한 인상을 줍니다."
       }
      ],
      "all_candidates_fail": false,
      "readings": [
       {
        "label": "A",
        "direction": "인물들의 팔과 손이 중심에 놓인 주먹밥을 향하고 있음.",
        "built_space": "어두운 헛간의 흙바닥이 레퍼런스와 일치하게 묘사됨.",
        "entities": "녹색 보자기와 주먹밥, 전통 복식을 한 인물들이 확인됨.",
        "hard_violations": [],
        "physics": "인물들이 바닥에 쪼그려 앉아 있으며, 신체 지지가 자연스러움."
       },
       {
        "label": "B",
        "direction": "프레임 가장자리에서 중심의 주먹밥을 향해 다수의 손이 집중적으로 뻗어 나옴.",
        "built_space": "조명과 흙바닥의 질감이 지시된 헛간 환경과 잘 들어맞음.",
        "entities": "레퍼런스와 동일한 형태의 보자기, 주먹밥, 한복 및 짚신이 확인됨.",
        "hard_violations": [],
        "physics": "프레임 주변의 신체와 연결된 팔들이 올바른 궤적으로 뻗어 나옴."
       }
      ]
     },
     "ok": true
    },
    {
     "model": "gpt-high",
     "order": "reverse",
     "display_to_canonical": {
      "A": "B",
      "B": "A"
     },
     "raw": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 9,
        "verdict_ko": "일곱 손이 중앙의 노출된 주먹밥을 아직 집지 않은 채 향하는 클로즈업으로, 요청한 찰나와 위로 열린 보자기의 배치를 가장 충실히 구현했다."
       },
       {
        "label": "B",
        "score": 7,
        "verdict_ko": "손과 음식의 클로즈업은 맞지만 일부 손이 이미 주먹밥을 잡기 시작했고, 낮은 시점에서 보자기 내부가 가려져 요청한 동시 접근 순간은 A보다 덜 정확하다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "화면 좌우·상단·하단에서 들어오는 일곱 손의 손가락이 모두 중앙 주먹밥 무리를 향한다. 손들은 밥을 움켜쥐기 직전에 있으며, 들려 나간 주먹밥은 없다. 얼굴과 시선은 프레임 밖이다.",
        "built_space": "보자기 한 장이 다져진 흙바닥 중앙에 펼쳐져 있고, 비스듬히 내려다보는 카메라에 내부가 잘 드러난다. 주변에는 사람들의 팔과 일부 다리·신발만 들어온다. 문·창·기둥은 크롭 밖이므로 수량이나 개방 상태를 확인할 수 없다. 어두운 실내와 상단에서 유입되는 낮빛은 참고 장소의 조건에 부합한다.",
        "entities": "청록색 보자기 한 장과 여러 개의 흰 주먹밥, 거친 무채색 직물 소매를 입은 사람들의 손이 보인다. 보자기의 색과 직물 질감은 소품 참고와 대체로 맞지만 참고의 굵은 땋은 끈은 식별되지 않는다. 손은 성인의 손처럼 보이나 얼굴이 없어 한국인 여부와 정확한 성별·나이는 확인할 수 없다. 읽을 수 있는 글자나 화면 표식은 없다.",
        "hard_violations": [],
        "physics": "주먹밥은 보자기와 아래쪽 주먹밥에 받쳐져 있고 보자기는 흙바닥에 놓여 있다. 뻗은 손들은 손목과 소매 속 팔로 자연스럽게 이어진다. 보이는 신발은 바닥에 닿아 있으며, 손을 모으는 동작에 불가능한 관절이나 지지 없는 물체는 보이지 않는다."
       },
       {
        "label": "B",
        "direction": "네 손이 좌우와 뒤쪽에서 중앙 주먹밥으로 모인다. 왼쪽과 오른쪽 전경 손은 이미 주먹밥에 닿아 잡기 시작한 모양이다. 뒤쪽 인물의 다른 한 손은 무릎 위에 머문다. 얼굴은 잘려 시선은 확인할 수 없다.",
        "built_space": "보자기 한 장이 흙바닥 위에 펼쳐져 있고 사람들은 그 뒤와 양옆에 쪼그려 앉아 있다. 낮은 사선 시점 때문에 보자기 앞쪽 주름이 내부 일부를 가리며, 뒤쪽 인물의 몸통과 무릎도 들어온다. 배경에 어두운 목재 면이 보이지만 문·창·기둥의 수량이나 열린 출입구는 확인되지 않는다.",
        "entities": "청록색 직물 보자기와 여러 주먹밥, 흙 묻은 손, 낡은 소매와 신발이 보인다. 보자기의 색과 짜임은 참고 소품과 유사하지만 땋은 운반 끈은 식별되지 않는다. 주먹밥은 아직 모두 보자기 위에 있다. 얼굴이 없어 인종·정확한 나이·성별을 확정할 수 없으며 읽을 수 있는 글자는 없다.",
        "hard_violations": [],
        "physics": "보자기는 바닥에 밀착하고 주먹밥은 천 위에 놓여 있다. 음식에 닿은 손에는 실제 접촉이 있으며 공중으로 들린 밥은 없다. 팔은 소매와 몸으로 이어지고 양옆 신발은 바닥을 딛는다. 뒤쪽 인물의 굽힌 다리와 무릎에 놓인 손도 자연스러워 명백한 지지·해부학 오류는 없다."
       }
      ],
      "all_candidates_fail": false
     },
     "normalized": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 9,
        "verdict_ko": "일곱 손이 중앙의 노출된 주먹밥을 아직 집지 않은 채 향하는 클로즈업으로, 요청한 찰나와 위로 열린 보자기의 배치를 가장 충실히 구현했다."
       },
       {
        "label": "A",
        "score": 7,
        "verdict_ko": "손과 음식의 클로즈업은 맞지만 일부 손이 이미 주먹밥을 잡기 시작했고, 낮은 시점에서 보자기 내부가 가려져 요청한 동시 접근 순간은 A보다 덜 정확하다."
       }
      ],
      "all_candidates_fail": false,
      "readings": [
       {
        "label": "B",
        "direction": "화면 좌우·상단·하단에서 들어오는 일곱 손의 손가락이 모두 중앙 주먹밥 무리를 향한다. 손들은 밥을 움켜쥐기 직전에 있으며, 들려 나간 주먹밥은 없다. 얼굴과 시선은 프레임 밖이다.",
        "built_space": "보자기 한 장이 다져진 흙바닥 중앙에 펼쳐져 있고, 비스듬히 내려다보는 카메라에 내부가 잘 드러난다. 주변에는 사람들의 팔과 일부 다리·신발만 들어온다. 문·창·기둥은 크롭 밖이므로 수량이나 개방 상태를 확인할 수 없다. 어두운 실내와 상단에서 유입되는 낮빛은 참고 장소의 조건에 부합한다.",
        "entities": "청록색 보자기 한 장과 여러 개의 흰 주먹밥, 거친 무채색 직물 소매를 입은 사람들의 손이 보인다. 보자기의 색과 직물 질감은 소품 참고와 대체로 맞지만 참고의 굵은 땋은 끈은 식별되지 않는다. 손은 성인의 손처럼 보이나 얼굴이 없어 한국인 여부와 정확한 성별·나이는 확인할 수 없다. 읽을 수 있는 글자나 화면 표식은 없다.",
        "hard_violations": [],
        "physics": "주먹밥은 보자기와 아래쪽 주먹밥에 받쳐져 있고 보자기는 흙바닥에 놓여 있다. 뻗은 손들은 손목과 소매 속 팔로 자연스럽게 이어진다. 보이는 신발은 바닥에 닿아 있으며, 손을 모으는 동작에 불가능한 관절이나 지지 없는 물체는 보이지 않는다."
       },
       {
        "label": "A",
        "direction": "네 손이 좌우와 뒤쪽에서 중앙 주먹밥으로 모인다. 왼쪽과 오른쪽 전경 손은 이미 주먹밥에 닿아 잡기 시작한 모양이다. 뒤쪽 인물의 다른 한 손은 무릎 위에 머문다. 얼굴은 잘려 시선은 확인할 수 없다.",
        "built_space": "보자기 한 장이 흙바닥 위에 펼쳐져 있고 사람들은 그 뒤와 양옆에 쪼그려 앉아 있다. 낮은 사선 시점 때문에 보자기 앞쪽 주름이 내부 일부를 가리며, 뒤쪽 인물의 몸통과 무릎도 들어온다. 배경에 어두운 목재 면이 보이지만 문·창·기둥의 수량이나 열린 출입구는 확인되지 않는다.",
        "entities": "청록색 직물 보자기와 여러 주먹밥, 흙 묻은 손, 낡은 소매와 신발이 보인다. 보자기의 색과 짜임은 참고 소품과 유사하지만 땋은 운반 끈은 식별되지 않는다. 주먹밥은 아직 모두 보자기 위에 있다. 얼굴이 없어 인종·정확한 나이·성별을 확정할 수 없으며 읽을 수 있는 글자는 없다.",
        "hard_violations": [],
        "physics": "보자기는 바닥에 밀착하고 주먹밥은 천 위에 놓여 있다. 음식에 닿은 손에는 실제 접촉이 있으며 공중으로 들린 밥은 없다. 팔은 소매와 몸으로 이어지고 양옆 신발은 바닥을 딛는다. 뒤쪽 인물의 굽힌 다리와 무릎에 놓인 손도 자연스러워 명백한 지지·해부학 오류는 없다."
       }
      ]
     },
     "ok": true
    }
   ],
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "slot_winner_match": true,
   "slot_winner": {
    "gemini-pro": "B",
    "gpt-high": "B"
   },
   "route": "cross_slot_agree"
  },
  "dual": {
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "normalized": {
    "A": 1.349,
    "B": 2.0
   },
   "adjusted": {
    "A": 1.349,
    "B": 2.0
   },
   "violations": {},
   "per_model_winner": {
    "gemini-pro": "B",
    "gpt-high": "B"
   },
   "agreed": true
  },
  "totals": {
   "B": 2000,
   "A": 1349
  },
  "selected": "B",
  "ranking": [
   "B",
   "A"
  ],
  "verdicts": [
   {
    "label": "B",
    "score": 2000,
    "verdict_ko": "클로즈업 프레이밍 지시를 정확히 따랐으며, 여러 손이 주먹밥을 향해 일제히 뻗어 나오는 역동적인 찰나를 훌륭하게 구현했습니다."
   },
   {
    "label": "A",
    "score": 1349,
    "verdict_ko": "프레이밍이 너무 넓어 클로즈업 지시를 어겼으며, 손이 뻗어 나오는 찰나보다는 이미 손을 얹고 멈춰 있는 듯한 인상을 줍니다."
   }
  ],
  "refs": [
   {
    "label": "PREVIOUS SHOT STILL — a visually related earlier shot of this same place: the location's look, materials, fixed features and lighting mood are LOCKED to this photo; never copy its camera framing. The people visible in that photo are NOT in this shot: never carry their faces, bodies or clothing onto anyone here. Each person in THIS shot is defined solely by the PEOPLE list, their own CHARACTER REFERENCE images and their wardrobe notes. If this photo shows a character who cannot move (dead or unconscious), that character's exact pose, position and orientation are ALSO LOCKED — treat their body as a fixed prop of the set that only the camera moves around.",
    "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/db00517b-165b-4005-a287-f28936b59b95/images/272bd7af-9294-4490-98c6-88a40e9deacc/scene/recipe/S21sh1_sel.png",
    "asset_id": "57fb6e88-e7a9-44d5-85fb-7fbc410bdd91",
    "role": "prev_still"
   },
   {
    "label": "PROP REFERENCE — 보자기 꾸러미·봇짐: the exact object appearing in this shot; match its look, material and wear exactly.",
    "path": "<bytes:1465734>",
    "asset_id": "3ca92b24-68b4-4883-b851-cabc06c00d1a",
    "role": "prop_ref"
   }
  ],
  "critique_skipped": true,
  "shot_run_produced": false,
  "shot_run_uid": "06aa1418-9c3c-7b4f-bfac-f0312971b700",
  "ref_mode": "prev+엔티티",
  "share_plan": {
   "ref_plan": "prev",
   "prev_anchor_tag": "S21sh1"
  }
 },
 "S21sh7::cine": {
  "applied": true,
  "attempted_at": "2026-09-09T09:00:42.908711+00:00",
  "fingerprint": "4bcca20a0dc452973a06a94e284f23123b094b29cc7c9ab602eb27a51bc7a5f2",
  "fingerprint_version": 2,
  "provider": "mai",
  "endpoint": "openrouter/chat-completions",
  "model": "microsoft/mai-image-2.6",
  "pack": "24.202608252115",
  "source_file": "S21sh7_sel.png",
  "source_sha256": "08aac65c5ace97d500c0a9df395cb5d8fe83da3e7b8bf801edba788f0e8fecba",
  "file": "S21sh7_cine.png",
  "staged_sha256": "03ed53762a553fad3eb056164e5fd9f396d4ea568b90b17729ca6ed15d938851",
  "latency_ms": 26253
 },
 "S21sh16::signage": {
  "fp": "c6297a92dcbb911d",
  "inscriptions": [],
  "cues": [],
  "dropped": []
 },
 "S21sh16": {
  "input_fingerprint": "8ca812fd42184afd",
  "prompt": "Create ONE FINAL photorealistic live-action film still of the moment below — early 20th-century Korea, Jeolla region; all people are Korean unless explicitly stated otherwise. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 달래를 내려다보며 한쪽 입꼬리를 올려 피식 웃는 안규홍의 붓기 가득한 정면.\n\nLOCATION (lock): The standing space beside the opened food bundle inside the hillside barn, in subdued daylight from its entrance. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: close-up\n- KEY BACKGROUND ELEMENTS: Barn interior (Only a narrow background area remains visible behind 안규홍); used as Unobtrusive spatial continuity behind the close reaction.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Keep the barn dim, preserving gentle facial separation and subdued contrast so the warmth of the smile does not erase the injuries.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The clothing, hair and overall look of 안규홍 — who appear both in that photo and in this shot — are LOCKED to that photo. Anyone else visible in that photo is NOT in this shot: never carry their face, body or clothing onto anyone here. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nPREVIOUS STILL USAGE (follow exactly — what to take from the attached still and what to exclude; it governs place and objects only, never who is in this shot or how the camera sees them): Take the barn's structural materials, floor texture, and subdued daylight. Exclude the earlier exhausted resting tableau and do not add food that has already been eaten.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The open rice-ball cloth remains in the dim barn. 안규홍: He is standing with his mouth still full, one eye swollen and his lip split; the jade ring remains concealed.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 안규홍 (한국인 남성, 청년 나이대, 검은 머리, 단단히 묶은 머리) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.",
  "roll_prompts": {
   "A": "Create ONE FINAL photorealistic live-action film still of the moment below — early 20th-century Korea, Jeolla region; all people are Korean unless explicitly stated otherwise. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 달래를 내려다보며 한쪽 입꼬리를 올려 피식 웃는 안규홍의 붓기 가득한 정면.\n\nLOCATION (lock): The standing space beside the opened food bundle inside the hillside barn, in subdued daylight from its entrance. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: close-up\n- KEY BACKGROUND ELEMENTS: Barn interior (Only a narrow background area remains visible behind 안규홍); used as Unobtrusive spatial continuity behind the close reaction.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Keep the barn dim, preserving gentle facial separation and subdued contrast so the warmth of the smile does not erase the injuries.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The clothing, hair and overall look of 안규홍 — who appear both in that photo and in this shot — are LOCKED to that photo. Anyone else visible in that photo is NOT in this shot: never carry their face, body or clothing onto anyone here. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nPREVIOUS STILL USAGE (follow exactly — what to take from the attached still and what to exclude; it governs place and objects only, never who is in this shot or how the camera sees them): Take the barn's structural materials, floor texture, and subdued daylight. Exclude the earlier exhausted resting tableau and do not add food that has already been eaten.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The open rice-ball cloth remains in the dim barn. 안규홍: He is standing with his mouth still full, one eye swollen and his lip split; the jade ring remains concealed.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 안규홍 (한국인 남성, 청년 나이대, 검은 머리, 단단히 묶은 머리) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.",
   "B": "Create ONE FINAL photorealistic live-action film still of the moment below — early 20th-century Korea, Jeolla region; all people are Korean unless explicitly stated otherwise. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 달래를 내려다보며 한쪽 입꼬리를 올려 피식 웃는 안규홍의 붓기 가득한 정면.\n\nLOCATION (lock): The standing space beside the opened food bundle inside the hillside barn, in subdued daylight from its entrance. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: close-up\n- KEY BACKGROUND ELEMENTS: Barn interior (Only a narrow background area remains visible behind 안규홍); used as Unobtrusive spatial continuity behind the close reaction.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Keep the barn dim, preserving gentle facial separation and subdued contrast so the warmth of the smile does not erase the injuries.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The clothing, hair and overall look of 안규홍 — who appear both in that photo and in this shot — are LOCKED to that photo. Anyone else visible in that photo is NOT in this shot: never carry their face, body or clothing onto anyone here. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nPREVIOUS STILL USAGE (follow exactly — what to take from the attached still and what to exclude; it governs place and objects only, never who is in this shot or how the camera sees them): Take the barn's structural materials, floor texture, and subdued daylight. Exclude the earlier exhausted resting tableau and do not add food that has already been eaten.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The open rice-ball cloth remains in the dim barn. 안규홍: He is standing with his mouth still full, one eye swollen and his lip split; the jade ring remains concealed.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 안규홍 (한국인 남성, 청년 나이대, 검은 머리, 단단히 묶은 머리) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.\n\nCOMPOSITION VARIATION (this candidate only): keep the place, the people and the moment exactly as specified, and keep from the CAMERA & FRAME contract its subject scale, screen placement and key background elements exactly. For this candidate only, its camera angle and height are deliberately left open: choose a camera position distinctly different from the obvious one, as a film director picking a second setup on the same blocking. Stay true to any explicit framing the SHOT TEXT itself dictates."
  },
  "shot_run_spend_attempt_count": 1,
  "readings": [
   {
    "label": "A",
    "direction": "시선은 아래를 향하고 한쪽 입꼬리가 올라감.",
    "built_space": "흙벽과 기둥이 있는 헛간 내부. 지시대로 배경이 좁게 렌더링됨.",
    "entities": "안규홍(부은 한쪽 눈, 터진 입술, 음식을 물어 볼록한 볼), 내용물이 없는 보따리 천.",
    "hard_violations": [
     "[gpt-high] 이미 먹은 음식을 추가하지 말라는 명시적 지시에도 전경의 열린 보자기에 밥 덩어리를 추가했다."
    ],
    "physics": "자연스럽게 서 있는 자세이며 어색함 없음."
   },
   {
    "label": "B",
    "direction": "시선은 아래를 향하고 한쪽 입꼬리가 올라감.",
    "built_space": "헛간 내부가 깊게 렌더링되어 문 밖까지 보임(좁은 배경 지시 불일치).",
    "entities": "안규홍(부은 한쪽 눈, 터진 입술), 주먹밥이 들어있는 보따리 천.",
    "hard_violations": [
     "[gemini-pro] 명시적으로 추가하지 말라고 지시한 음식(주먹밥)을 보따리 안에 생성함 (invented objects 위반).",
     "[gpt-high] 이미 먹은 음식을 추가하지 말라는 명시적 지시와 달리, 전경 보자기에 주먹밥을 적어도 세 개 추가했다."
    ],
    "physics": "서서 두 손으로 보따리를 들고 있는 자세."
   }
  ],
  "cross_model_order": {
   "policy": "cross_model_order_v2_slot0_fwd_slot1_rev_parallel2",
   "slots": [
    {
     "model": "gemini-pro",
     "order": "forward",
     "display_to_canonical": {
      "A": "A",
      "B": "B"
     },
     "raw": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 9,
        "verdict_ko": "클로즈업과 좁은 배경 프레이밍을 정확히 따랐으며, 음식을 물고 볼이 빵빵한 상태와 부은 얼굴 등 지시사항을 완벽히 구현함."
       },
       {
        "label": "B",
        "score": 4,
        "verdict_ko": "배경이 너무 깊게 렌더링되었으며, 입에 음식이 가득 찬 디테일이 부족하고, 명시적으로 금지된 남은 주먹밥을 추가하여 감점됨."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "시선은 아래를 향하고 한쪽 입꼬리가 올라감.",
        "built_space": "흙벽과 기둥이 있는 헛간 내부. 지시대로 배경이 좁게 렌더링됨.",
        "entities": "안규홍(부은 한쪽 눈, 터진 입술, 음식을 물어 볼록한 볼), 내용물이 없는 보따리 천.",
        "hard_violations": [],
        "physics": "자연스럽게 서 있는 자세이며 어색함 없음."
       },
       {
        "label": "B",
        "direction": "시선은 아래를 향하고 한쪽 입꼬리가 올라감.",
        "built_space": "헛간 내부가 깊게 렌더링되어 문 밖까지 보임(좁은 배경 지시 불일치).",
        "entities": "안규홍(부은 한쪽 눈, 터진 입술), 주먹밥이 들어있는 보따리 천.",
        "hard_violations": [
         "명시적으로 추가하지 말라고 지시한 음식(주먹밥)을 보따리 안에 생성함 (invented objects 위반)."
        ],
        "physics": "서서 두 손으로 보따리를 들고 있는 자세."
       }
      ],
      "all_candidates_fail": false
     },
     "normalized": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 9,
        "verdict_ko": "클로즈업과 좁은 배경 프레이밍을 정확히 따랐으며, 음식을 물고 볼이 빵빵한 상태와 부은 얼굴 등 지시사항을 완벽히 구현함."
       },
       {
        "label": "B",
        "score": 4,
        "verdict_ko": "배경이 너무 깊게 렌더링되었으며, 입에 음식이 가득 찬 디테일이 부족하고, 명시적으로 금지된 남은 주먹밥을 추가하여 감점됨."
       }
      ],
      "all_candidates_fail": false,
      "readings": [
       {
        "label": "A",
        "direction": "시선은 아래를 향하고 한쪽 입꼬리가 올라감.",
        "built_space": "흙벽과 기둥이 있는 헛간 내부. 지시대로 배경이 좁게 렌더링됨.",
        "entities": "안규홍(부은 한쪽 눈, 터진 입술, 음식을 물어 볼록한 볼), 내용물이 없는 보따리 천.",
        "hard_violations": [],
        "physics": "자연스럽게 서 있는 자세이며 어색함 없음."
       },
       {
        "label": "B",
        "direction": "시선은 아래를 향하고 한쪽 입꼬리가 올라감.",
        "built_space": "헛간 내부가 깊게 렌더링되어 문 밖까지 보임(좁은 배경 지시 불일치).",
        "entities": "안규홍(부은 한쪽 눈, 터진 입술), 주먹밥이 들어있는 보따리 천.",
        "hard_violations": [
         "명시적으로 추가하지 말라고 지시한 음식(주먹밥)을 보따리 안에 생성함 (invented objects 위반)."
        ],
        "physics": "서서 두 손으로 보따리를 들고 있는 자세."
       }
      ]
     },
     "ok": true
    },
    {
     "model": "gpt-high",
     "order": "reverse",
     "display_to_canonical": {
      "A": "B",
      "B": "A"
     },
     "raw": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 2,
        "verdict_ko": "내려보는 시선과 비웃음은 맞지만, 이미 먹은 주먹밥을 여러 개 추가했고 가슴과 헛간을 넓게 보여 정면 얼굴 클로즈업에서 벗어납니다."
       },
       {
        "label": "B",
        "score": 4,
        "verdict_ko": "부은 눈과 터진 입술을 유지한 정면 클로즈업, 아래로 향한 시선과 한쪽 입꼬리는 더 충실하지만, 전경에 음식이 든 보자기를 추가한 위반은 남습니다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "얼굴은 정면이고 두 눈은 화면 아래쪽을 향한다. 화면 밖 낮은 위치의 달래를 내려다보는 방향으로 해석할 수 있지만, 바로 아래에 주먹밥 보자기가 있어 시선의 실제 대상이 음식처럼도 읽힌다. 화면 오른쪽 입꼬리가 올라가 있다. 무기나 이동하는 물체는 없다.",
        "built_space": "흙벽과 목재 골조, 흙바닥이 보인다. 인물 뒤 중앙에 밝은 출입구 하나, 왼쪽 벽에 작은 목재 창 또는 덧문 하나가 보이며 천장 부재와 양쪽 벽까지 드러난다. 재료는 참조와 대체로 이어지지만, 좁은 배경만 남기는 얼굴 클로즈업보다 공간을 훨씬 넓게 보여준다. 가슴 앞 보자기는 그 옆에 서 있다는 배치와 다르다. 반사면은 없다.",
        "entities": "젊은 한국인 남성으로 보이는 인물 한 명이며 검은 머리를 뒤로 단단히 넘겼다. 얼굴과 때 묻은 미색 옷은 안규홍 참조에 대체로 부합한다. 화면 오른쪽 눈이 붓고 입술에 갈라진 상처가 있으며 비대칭 미소가 보인다. 다만 입안이 가득 찬 볼의 팽창은 약하다. 전경의 열린 천 안에 주먹밥이 적어도 세 개 보여, 이미 먹은 음식을 추가하지 말라는 지시와 충돌한다. 달래나 다른 인물, 옥가락지, 읽을 수 있는 글자는 보이지 않는다.",
        "hard_violations": [
         "이미 먹은 음식을 추가하지 말라는 명시적 지시와 달리, 전경 보자기에 주먹밥을 적어도 세 개 추가했다."
        ],
        "physics": "머리와 목은 어깨 및 몸통에 자연스럽게 연결되어 있고 부상도 실제 피부 위에 나타난다. 하체는 프레임 밖이므로 서 있는 자세의 발 접촉은 확인할 수 없다. 주먹밥은 접힌 보자기 안에 놓여 있다. 보자기 밑면과 손이 잘려 있어 누가 어떻게 받치는지는 확인되지 않지만, 화면 밖 지지를 배제하여 공중 부양이라고 단정할 근거는 없다."
       },
       {
        "label": "B",
        "direction": "얼굴은 거의 정확한 정면이고 눈동자는 렌즈보다 아래를 향한다. 화면 밖 아래에 있는 달래를 내려다보는 시선에 부합한다. 화면 오른쪽 입꼬리를 살짝 올린 표정이 분명하다. 다만 아래 전경의 음식 보자기가 시선 대상으로 경쟁한다. 무기나 이동하는 물체는 없다.",
        "built_space": "얼굴과 어깨가 화면 대부분을 차지한다. 뒤에는 흙벽, 가로 목재 한 줄, 왼쪽 세로 골조 한 줄과 오른쪽 굵은 기둥 하나가 보이고 왼쪽 아래에 짚이 있다. 출입구나 창, 바닥은 이 크롭에서 보이지 않는다. 참조의 흙벽과 낡은 목재 재질 및 어두운 낮 분위기를 잘 유지하며 배경이 얼굴 반응을 압도하지 않는다. 보자기는 인물 옆이 아니라 화면 아래 가슴 앞 전경에 놓여 있다. 반사면은 없다.",
        "entities": "안규홍 참조와 닮은 젊은 한국인 남성 한 명이 보이며, 검은 머리를 뒤로 넘긴 모습과 오염된 미색 옷깃이 일치한다. 화면 오른쪽 눈의 부기와 멍, 갈라진 아랫입술이 보인다. 다문 입과 약간 팽팽한 볼은 입에 음식이 남은 상태를 어느 정도 전달한다. 아래쪽 열린 보자기에는 밝은 밥 덩어리가 드러나 이미 먹은 음식의 추가 금지와 충돌한다. 다른 인물, 노출된 옥가락지, 읽을 수 있는 글자는 없다.",
        "hard_violations": [
         "이미 먹은 음식을 추가하지 말라는 명시적 지시에도 전경의 열린 보자기에 밥 덩어리를 추가했다."
        ],
        "physics": "머리와 목, 어깨의 연결 및 얼굴 근육의 비대칭 미소는 자연스럽다. 몸통 아래와 발은 클로즈업 밖이므로 지면 접촉이나 기립 여부를 직접 확인할 수 없다. 음식은 보자기 주름 안에 놓여 있으며, 보자기의 아래쪽과 이를 받칠 손 또는 표면은 화면 밖이다. 지지 방식은 확인할 수 없지만, 잘린 화면만으로 물체가 떠 있다고 판단할 수는 없다."
       }
      ],
      "all_candidates_fail": true
     },
     "normalized": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 2,
        "verdict_ko": "내려보는 시선과 비웃음은 맞지만, 이미 먹은 주먹밥을 여러 개 추가했고 가슴과 헛간을 넓게 보여 정면 얼굴 클로즈업에서 벗어납니다."
       },
       {
        "label": "A",
        "score": 4,
        "verdict_ko": "부은 눈과 터진 입술을 유지한 정면 클로즈업, 아래로 향한 시선과 한쪽 입꼬리는 더 충실하지만, 전경에 음식이 든 보자기를 추가한 위반은 남습니다."
       }
      ],
      "all_candidates_fail": true,
      "readings": [
       {
        "label": "B",
        "direction": "얼굴은 정면이고 두 눈은 화면 아래쪽을 향한다. 화면 밖 낮은 위치의 달래를 내려다보는 방향으로 해석할 수 있지만, 바로 아래에 주먹밥 보자기가 있어 시선의 실제 대상이 음식처럼도 읽힌다. 화면 오른쪽 입꼬리가 올라가 있다. 무기나 이동하는 물체는 없다.",
        "built_space": "흙벽과 목재 골조, 흙바닥이 보인다. 인물 뒤 중앙에 밝은 출입구 하나, 왼쪽 벽에 작은 목재 창 또는 덧문 하나가 보이며 천장 부재와 양쪽 벽까지 드러난다. 재료는 참조와 대체로 이어지지만, 좁은 배경만 남기는 얼굴 클로즈업보다 공간을 훨씬 넓게 보여준다. 가슴 앞 보자기는 그 옆에 서 있다는 배치와 다르다. 반사면은 없다.",
        "entities": "젊은 한국인 남성으로 보이는 인물 한 명이며 검은 머리를 뒤로 단단히 넘겼다. 얼굴과 때 묻은 미색 옷은 안규홍 참조에 대체로 부합한다. 화면 오른쪽 눈이 붓고 입술에 갈라진 상처가 있으며 비대칭 미소가 보인다. 다만 입안이 가득 찬 볼의 팽창은 약하다. 전경의 열린 천 안에 주먹밥이 적어도 세 개 보여, 이미 먹은 음식을 추가하지 말라는 지시와 충돌한다. 달래나 다른 인물, 옥가락지, 읽을 수 있는 글자는 보이지 않는다.",
        "hard_violations": [
         "이미 먹은 음식을 추가하지 말라는 명시적 지시와 달리, 전경 보자기에 주먹밥을 적어도 세 개 추가했다."
        ],
        "physics": "머리와 목은 어깨 및 몸통에 자연스럽게 연결되어 있고 부상도 실제 피부 위에 나타난다. 하체는 프레임 밖이므로 서 있는 자세의 발 접촉은 확인할 수 없다. 주먹밥은 접힌 보자기 안에 놓여 있다. 보자기 밑면과 손이 잘려 있어 누가 어떻게 받치는지는 확인되지 않지만, 화면 밖 지지를 배제하여 공중 부양이라고 단정할 근거는 없다."
       },
       {
        "label": "A",
        "direction": "얼굴은 거의 정확한 정면이고 눈동자는 렌즈보다 아래를 향한다. 화면 밖 아래에 있는 달래를 내려다보는 시선에 부합한다. 화면 오른쪽 입꼬리를 살짝 올린 표정이 분명하다. 다만 아래 전경의 음식 보자기가 시선 대상으로 경쟁한다. 무기나 이동하는 물체는 없다.",
        "built_space": "얼굴과 어깨가 화면 대부분을 차지한다. 뒤에는 흙벽, 가로 목재 한 줄, 왼쪽 세로 골조 한 줄과 오른쪽 굵은 기둥 하나가 보이고 왼쪽 아래에 짚이 있다. 출입구나 창, 바닥은 이 크롭에서 보이지 않는다. 참조의 흙벽과 낡은 목재 재질 및 어두운 낮 분위기를 잘 유지하며 배경이 얼굴 반응을 압도하지 않는다. 보자기는 인물 옆이 아니라 화면 아래 가슴 앞 전경에 놓여 있다. 반사면은 없다.",
        "entities": "안규홍 참조와 닮은 젊은 한국인 남성 한 명이 보이며, 검은 머리를 뒤로 넘긴 모습과 오염된 미색 옷깃이 일치한다. 화면 오른쪽 눈의 부기와 멍, 갈라진 아랫입술이 보인다. 다문 입과 약간 팽팽한 볼은 입에 음식이 남은 상태를 어느 정도 전달한다. 아래쪽 열린 보자기에는 밝은 밥 덩어리가 드러나 이미 먹은 음식의 추가 금지와 충돌한다. 다른 인물, 노출된 옥가락지, 읽을 수 있는 글자는 없다.",
        "hard_violations": [
         "이미 먹은 음식을 추가하지 말라는 명시적 지시에도 전경의 열린 보자기에 밥 덩어리를 추가했다."
        ],
        "physics": "머리와 목, 어깨의 연결 및 얼굴 근육의 비대칭 미소는 자연스럽다. 몸통 아래와 발은 클로즈업 밖이므로 지면 접촉이나 기립 여부를 직접 확인할 수 없다. 음식은 보자기 주름 안에 놓여 있으며, 보자기의 아래쪽과 이를 받칠 손 또는 표면은 화면 밖이다. 지지 방식은 확인할 수 없지만, 잘린 화면만으로 물체가 떠 있다고 판단할 수는 없다."
       }
      ]
     },
     "ok": true
    }
   ],
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "slot_winner_match": true,
   "slot_winner": {
    "gemini-pro": "A",
    "gpt-high": "A"
   },
   "route": "cross_slot_agree"
  },
  "dual": {
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "normalized": {
    "A": 2.0,
    "B": 0.944
   },
   "adjusted": {
    "A": 1.75,
    "B": 0.694
   },
   "violations": {
    "B": [
     "[gemini-pro] 명시적으로 추가하지 말라고 지시한 음식(주먹밥)을 보따리 안에 생성함 (invented objects 위반).",
     "[gpt-high] 이미 먹은 음식을 추가하지 말라는 명시적 지시와 달리, 전경 보자기에 주먹밥을 적어도 세 개 추가했다."
    ],
    "A": [
     "[gpt-high] 이미 먹은 음식을 추가하지 말라는 명시적 지시에도 전경의 열린 보자기에 밥 덩어리를 추가했다."
    ]
   },
   "per_model_winner": {
    "gemini-pro": "A",
    "gpt-high": "A"
   },
   "agreed": true
  },
  "totals": {
   "A": 1750,
   "B": 694
  },
  "selected": "A",
  "ranking": [
   "A",
   "B"
  ],
  "verdicts": [
   {
    "label": "A",
    "score": 1750,
    "verdict_ko": "클로즈업과 좁은 배경 프레이밍을 정확히 따랐으며, 음식을 물고 볼이 빵빵한 상태와 부은 얼굴 등 지시사항을 완벽히 구현함.  ★위반: [gpt-high] 이미 먹은 음식을 추가하지 말라는 명시적 지시에도 전경의 열린 보자기에 밥 덩어리를 추가했다."
   },
   {
    "label": "B",
    "score": 694,
    "verdict_ko": "배경이 너무 깊게 렌더링되었으며, 입에 음식이 가득 찬 디테일이 부족하고, 명시적으로 금지된 남은 주먹밥을 추가하여 감점됨.  ★위반: [gemini-pro] 명시적으로 추가하지 말라고 지시한 음식(주먹밥)을 보따리 안에 생성함 (invented objects 위반). / [gpt-high] 이미 먹은 음식을 추가하지 말라는 명시적 지시와 달리, 전경 보자기에 주먹밥을 적어도 세 개 추가했다."
   }
  ],
  "refs": [
   {
    "label": "PREVIOUS SHOT STILL — a visually related earlier shot of this same place: the location's look, materials, fixed features and lighting mood are LOCKED to this photo; never copy its camera framing. The clothing, hair and overall look of 안규홍 — who appear both in that photo and in this shot — are LOCKED to that photo. Anyone else visible in that photo is NOT in this shot: never carry their face, body or clothing onto anyone here. If this photo shows a character who cannot move (dead or unconscious), that character's exact pose, position and orientation are ALSO LOCKED — treat their body as a fixed prop of the set that only the camera moves around.",
    "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/db00517b-165b-4005-a287-f28936b59b95/images/272bd7af-9294-4490-98c6-88a40e9deacc/scene/recipe/S21sh1_sel.png",
    "asset_id": "57fb6e88-e7a9-44d5-85fb-7fbc410bdd91",
    "role": "prev_still"
   },
   {
    "label": "CHARACTER REFERENCE — 안규홍: the exact person appearing in this shot; match face, hair and build exactly.",
    "path": "<bytes:1028384>",
    "asset_id": "3b5894db-ce55-4f34-b966-5e725c3aa952",
    "role": "character_ref"
   }
  ],
  "critique_skipped": true,
  "shot_run_produced": false,
  "shot_run_uid": "06aa1418-9dda-75eb-b065-9be71448c625",
  "ref_mode": "prev+엔티티",
  "share_plan": {
   "ref_plan": "prev",
   "prev_anchor_tag": "S21sh1"
  }
 },
 "S21sh16::cine": {
  "applied": true,
  "attempted_at": "2026-09-09T09:01:09.471349+00:00",
  "fingerprint": "96b54f27fa4d7ac5ede95fca27ed5cd3914b2a7c4048d2e41472c66f634a726c",
  "fingerprint_version": 2,
  "provider": "mai",
  "endpoint": "openrouter/chat-completions",
  "model": "microsoft/mai-image-2.6",
  "pack": "24.202608252115",
  "source_file": "S21sh16_sel.png",
  "source_sha256": "e52ea2cf636aeb14739ef88fa76a8ff607a97bbeeba882f24f478ec0211e9595",
  "file": "S21sh16_cine.png",
  "staged_sha256": "8f5502e8331929852916693cf40bca42a0e6608563616247a5c81bf1133ab04d",
  "latency_ms": 26033
 },
 "S22sh2::signage": {
  "fp": "16fca2ac873ac6a4",
  "inscriptions": [],
  "cues": [],
  "dropped": []
 },
 "era_assess::46f14f6fab0859e3": {
  "subjects": [
   {
    "subject_native": "20세기 초 한국의 관솔과 관솔대",
    "search_terms_native": [
     "관솔불",
     "관솔대",
     "민속 관솔",
     "전통 관솔불"
    ],
    "language_lock_native": "한국어 외의 다른 언어로 검색어를 번역하거나 추가하지 마십시오.",
    "reason_ko": "AI 모델이 서양식 횃불이나 일반 양초로 묘사하기 쉬우나, 소나무 옹이(관솔)를 태우는 한국 전통 조명 방식의 형태를 갖추어야 합니다."
   }
  ],
  "subject_text": "배상도 끄나풀 집 곳간 내부\n쌀가마니와 씨나락 자루가 높이 쌓인 흙바닥 곳간. 은비녀와 유기그릇이 함께 놓이고, 관솔불이 내부를 밝힌다.",
  "identity": "canonical",
  "scope_id": "L20",
  "scope_role": "location_interior",
  "scope_sha": "a815b1c887b6d4d8"
 },
 "era_fail::c8863aa244f0bbe3": {
  "stage": "research",
  "subject": "20세기 초 한국의 관솔과 관솔대",
  "terms": [
   "관솔불",
   "관솔대",
   "민속 관솔",
   "전통 관솔불"
  ],
  "status": "no_candidates",
  "queries": [
   [
    "관솔불 관솔대 민속",
    "관솔대 20세기 초 등잔"
   ],
   [
    "\"관솔대\" \"박물관\"",
    "\"관솔불\" \"한국민족문화대백과\"",
    "\"관솔\" \"등경\" \"민속\""
   ],
   [
    "\"관솔대\" \"정료대\"",
    "\"관솔불\" \"일제강점기\"",
    "\"관솔\" \"국립민속박물관\" \"송진\""
   ]
  ],
  "attempts": 5
 },
 "S22sh2::bgfirst_bg": {
  "input_fingerprint": "6ec65923f5c5ad3f",
  "prompt": "Create the EMPTY BACKGROUND PLATE for one film shot — NO PEOPLE, no figures, no body parts, no sketch lines, no arrows anywhere.\n\"Empty\" means no people only: KEEP the location's inherent occupants and stock that define the place — animals in an animal shelter, pen or farm, goods in a market, moored boats in a harbour — unless the shot text explicitly removes them.\nThe FIRST attached image is a thin-line storyboard sketch: use ONLY its camera angle, horizon, perspective and the placement/size of buildings and set masses — ignore the sketched people and arrows entirely. The SECOND attached image (LOCATION PHOTOGRAPH) is the real place: take its architecture, materials, signage and fixed features, and RE-PROJECT them into the sketch's camera. If the photograph's camera differs from the sketch's, the sketch's camera wins.\nHUMAN-SCALE CALIBRATION: derive every structure's true size from human-scale elements — a door ≈ 2m, a window ≈ 1–1.5m wide, one storey ≈ 2.5–3m; never inflate a small structure or shrink a large one.\n\nSHOT TEXT this background must serve (Korean): 활짝 열린 곳간 문 안으로 성큼 발을 내디딘 복면 쓴 안규홍의 전신.\n\nLOCATION (lock): Just inside the open doorway of a private home's grain storehouse at night, with pine-torch light falling over stored sacks and valuables.\n\nTIME OF DAY (lock): night.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: wide shot\n- FRAME LAYOUT: Open storehouse doorway behind the entering group in the upper-left of the frame, background.\n- KEY BACKGROUND ELEMENTS: Storehouse doorway (Fully open during the group's entry) — Seen obliquely from inside, with the entry passage visible behind 안규홍; used as Background threshold establishing movement from outside into the storehouse; Rice bales and seed sacks (Piled inside the storehouse) — Partial sides of the piles appear beside the entry path; used as Side framing and scale references without blocking the full-body action.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Use the scene's pine-knot firelight with restrained saturation and sober nighttime contrast, retaining readable bodies within the storehouse darkness.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nRender ONE photorealistic empty location photograph, 16:9, that this shot can be staged inside later. No readable writing anywhere: surfaces that would carry writing may be present, but stage any wording out of legibility — an oblique angle, distance, shallow focus. No captions, watermarks or overlay text.",
  "effective_prompt": "Create the EMPTY BACKGROUND PLATE for one film shot — NO PEOPLE, no figures, no body parts, no sketch lines, no arrows anywhere.\n\"Empty\" means no people only: KEEP the location's inherent occupants and stock that define the place — animals in an animal shelter, pen or farm, goods in a market, moored boats in a harbour — unless the shot text explicitly removes them.\nThe FIRST attached image is a thin-line storyboard sketch: use ONLY its camera angle, horizon, perspective and the placement/size of buildings and set masses — ignore the sketched people and arrows entirely. The SECOND attached image (LOCATION PHOTOGRAPH) is the real place: take its architecture, materials, signage and fixed features, and RE-PROJECT them into the sketch's camera. If the photograph's camera differs from the sketch's, the sketch's camera wins.\nHUMAN-SCALE CALIBRATION: derive every structure's true size from human-scale elements — a door ≈ 2m, a window ≈ 1–1.5m wide, one storey ≈ 2.5–3m; never inflate a small structure or shrink a large one.\n\nSHOT TEXT this background must serve (Korean): 활짝 열린 곳간 문 안으로 성큼 발을 내디딘 복면 쓴 안규홍의 전신.\n\nLOCATION (lock): Just inside the open doorway of a private home's grain storehouse at night, with pine-torch light falling over stored sacks and valuables.\n\nTIME OF DAY (lock): night.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: wide shot\n- FRAME LAYOUT: Open storehouse doorway behind the entering group in the upper-left of the frame, background.\n- KEY BACKGROUND ELEMENTS: Storehouse doorway (Fully open during the group's entry) — Seen obliquely from inside, with the entry passage visible behind 안규홍; used as Background threshold establishing movement from outside into the storehouse; Rice bales and seed sacks (Piled inside the storehouse) — Partial sides of the piles appear beside the entry path; used as Side framing and scale references without blocking the full-body action.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Use the scene's pine-knot firelight with restrained saturation and sober nighttime contrast, retaining readable bodies within the storehouse darkness.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nRender ONE photorealistic empty location photograph, 16:9, that this shot can be staged inside later. No readable writing anywhere: surfaces that would carry writing may be present, but stage any wording out of legibility — an oblique angle, distance, shallow focus. No captions, watermarks or overlay text.",
  "bg_path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/db00517b-165b-4005-a287-f28936b59b95/images/272bd7af-9294-4490-98c6-88a40e9deacc/scene/recipe/S22sh2__bgfirst_bg.png",
  "asset_id": "dfba00df-2a39-454d-8669-2847b09f54c8",
  "input_asset_ids": [
   "44af9002-9e8c-42e1-a3a9-c56745c9477f",
   "8ac26e75-11f5-4112-b737-07fc5fbda49a"
  ]
 },
 "S22sh2": {
  "input_fingerprint": "15101915e0612580",
  "prompt": "Create ONE FINAL photorealistic live-action film still of the moment below — early 20th-century Korea, Jeolla region; all people are Korean unless explicitly stated otherwise. TIME OF DAY (lock): night.\n\nSHOT TEXT (authoritative, Korean): 활짝 열린 곳간 문 안으로 성큼 발을 내디딘 복면 쓴 안규홍의 전신.\n\nLOCATION (lock): Just inside the open doorway of a private home's grain storehouse at night, with pine-torch light falling over stored sacks and valuables. The shot takes place here — the attached LOCATION PHOTOGRAPH shows the exact spot.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: wide shot\n- FRAME LAYOUT: Open storehouse doorway behind the entering group in the upper-left of the frame, background.\n- KEY BACKGROUND ELEMENTS: Storehouse doorway (Fully open during the group's entry) — Seen obliquely from inside, with the entry passage visible behind 안규홍; used as Background threshold establishing movement from outside into the storehouse; Rice bales and seed sacks (Piled inside the storehouse) — Partial sides of the piles appear beside the entry path; used as Side framing and scale references without blocking the full-body action.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Use the scene's pine-knot firelight with restrained saturation and sober nighttime contrast, retaining readable bodies within the storehouse darkness.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): A broken iron padlock lies on the floor; pine-resin light reveals rice bales, seed sacks, silver hairpins and brass bowls, including sacks marked '최 서방 종자' and '이 씨 볍씨'. 안규홍: He enters masked, with the swollen eye, split lip and concealed jade ring retained.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 안규홍 (한국인 남성, 청년 나이대, 검은 머리, 단단히 묶은 머리) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.",
  "roll_prompts": {
   "A": "Stage the shot. The FIRST attached image (SHOT BACKGROUND) is the finished empty background of this shot — keep it EXACTLY: its camera, perspective, architecture, lighting and every fixed feature stay untouched. The SECOND attached image (LAYOUT SKETCH) tells you ONLY where the people go: each sketched person's position, screen size, pose and the gaze/motion arrows. Ignore the sketch's background lines. The CHARACTER REFERENCE photographs show the real people.\nPlace the real people into the background at exactly the sketched positions, sizes and poses, following the arrow directions. No sketch lines or arrows may remain.\n\nCreate ONE FINAL photorealistic live-action film still of the moment below — early 20th-century Korea, Jeolla region; all people are Korean unless explicitly stated otherwise. TIME OF DAY (lock): night.\n\nSHOT TEXT (authoritative, Korean): 활짝 열린 곳간 문 안으로 성큼 발을 내디딘 복면 쓴 안규홍의 전신.\n\nLOCATION (lock): Just inside the open doorway of a private home's grain storehouse at night, with pine-torch light falling over stored sacks and valuables. The shot takes place here — the FIRST attached image (SHOT BACKGROUND) is this exact place, already built: its ground, structures, horizon, materials and lighting are the finished truth of this location and must not be redesigned or replaced. No location photograph is attached — read the place from that image alone, and add no scenery, structure, vehicle or fixture that it does not already show. This lock governs the place only; the figures in the shot follow the staging and pose instructions.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: wide shot\n- FRAME LAYOUT: Open storehouse doorway behind the entering group in the upper-left of the frame, background.\n- KEY BACKGROUND ELEMENTS: Storehouse doorway (Fully open during the group's entry) — Seen obliquely from inside, with the entry passage visible behind 안규홍; used as Background threshold establishing movement from outside into the storehouse; Rice bales and seed sacks (Piled inside the storehouse) — Partial sides of the piles appear beside the entry path; used as Side framing and scale references without blocking the full-body action.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Use the scene's pine-knot firelight with restrained saturation and sober nighttime contrast, retaining readable bodies within the storehouse darkness.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): A broken iron padlock lies on the floor; pine-resin light reveals rice bales, seed sacks, silver hairpins and brass bowls, including sacks marked '최 서방 종자' and '이 씨 볍씨'. 안규홍: He enters masked, with the swollen eye, split lip and concealed jade ring retained.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 안규홍 (한국인 남성, 청년 나이대, 검은 머리, 단단히 묶은 머리) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.",
   "B": "Create ONE FINAL photorealistic live-action film still of the moment below — early 20th-century Korea, Jeolla region; all people are Korean unless explicitly stated otherwise. TIME OF DAY (lock): night.\n\nSHOT TEXT (authoritative, Korean): 활짝 열린 곳간 문 안으로 성큼 발을 내디딘 복면 쓴 안규홍의 전신.\n\nLOCATION (lock): Just inside the open doorway of a private home's grain storehouse at night, with pine-torch light falling over stored sacks and valuables. The shot takes place here — the attached LOCATION PHOTOGRAPH shows the exact spot.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: wide shot\n- FRAME LAYOUT: Open storehouse doorway behind the entering group in the upper-left of the frame, background.\n- KEY BACKGROUND ELEMENTS: Storehouse doorway (Fully open during the group's entry) — Seen obliquely from inside, with the entry passage visible behind 안규홍; used as Background threshold establishing movement from outside into the storehouse; Rice bales and seed sacks (Piled inside the storehouse) — Partial sides of the piles appear beside the entry path; used as Side framing and scale references without blocking the full-body action.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Use the scene's pine-knot firelight with restrained saturation and sober nighttime contrast, retaining readable bodies within the storehouse darkness.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): A broken iron padlock lies on the floor; pine-resin light reveals rice bales, seed sacks, silver hairpins and brass bowls, including sacks marked '최 서방 종자' and '이 씨 볍씨'. 안규홍: He enters masked, with the swollen eye, split lip and concealed jade ring retained.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 안규홍 (한국인 남성, 청년 나이대, 검은 머리, 단단히 묶은 머리) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.\n\nCOMPOSITION VARIATION (this candidate only): keep the place, the people and the moment exactly as specified, and keep from the CAMERA & FRAME contract its subject scale, screen placement and key background elements exactly. For this candidate only, its camera angle and height are deliberately left open: choose a camera position distinctly different from the obvious one, as a film director picking a second setup on the same blocking. Stay true to any explicit framing the SHOT TEXT itself dictates."
  },
  "roll_refs": {
   "A": [
    {
     "label": "SHOT BACKGROUND",
     "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/db00517b-165b-4005-a287-f28936b59b95/images/272bd7af-9294-4490-98c6-88a40e9deacc/scene/recipe/S22sh2__bgfirst_bg.png",
     "asset_id": "dfba00df-2a39-454d-8669-2847b09f54c8",
     "role": "bgfirst_bg"
    },
    {
     "label": "LAYOUT SKETCH (people placement only)",
     "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/db00517b-165b-4005-a287-f28936b59b95/images/272bd7af-9294-4490-98c6-88a40e9deacc/conti/conti_S22sh2.png",
     "asset_id": "44af9002-9e8c-42e1-a3a9-c56745c9477f",
     "role": "conti_light"
    },
    {
     "label": "CHARACTER REFERENCE — 안규홍: the exact person appearing in this shot; match face, hair and build exactly.",
     "path": "<bytes:1028384>",
     "asset_id": "3b5894db-ce55-4f34-b966-5e725c3aa952",
     "role": "character_ref"
    }
   ],
   "B": [
    {
     "label": "LOCATION PHOTOGRAPH — the exact place of this shot: its architecture, materials, fixed features and lighting mood are spatial truth; stage the moment inside this place. Never copy its camera framing.",
     "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/db00517b-165b-4005-a287-f28936b59b95/episodes/272bd7af-9294-4490-98c6-88a40e9deacc/images/background_chain/L20B01.png",
     "asset_id": "8ac26e75-11f5-4112-b737-07fc5fbda49a",
     "role": "location_plate"
    },
    {
     "label": "CHARACTER REFERENCE — 안규홍: the exact person appearing in this shot; match face, hair and build exactly.",
     "path": "<bytes:1028384>",
     "asset_id": "3b5894db-ce55-4f34-b966-5e725c3aa952",
     "role": "character_ref"
    }
   ]
  },
  "shot_run_spend_attempt_count": 1,
  "judge_flip_skipped": "judge_fn_owns_order",
  "readings": [
   {
    "label": "A",
    "direction": "안규홍은 곳간 문턱을 넘어 화면 우측 하단을 향해 시선을 두고 걷고 있음.",
    "built_space": "곳간 내부 전경으로 왼쪽에 개방된 문이 있고, 참조 사진에 있는 가마니와 놋그릇 세트, 부서진 자물쇠가 바닥에 배치됨. 횃불(촛불)이 안쪽을 밝힘.",
    "entities": "복면을 쓴 안규홍 본인은 잘 묘사되었으나, 그의 뒤로 샷 텍스트에 언급되지 않은 조선인 남성 두 명이 추가로 서 있음. 자물쇠와 글씨가 적힌 가마니 등 프랍은 일치함.",
    "hard_violations": [
     "[gemini-pro] invented people (샷 텍스트에 없는 추가 인물 2명 등장)",
     "[gpt-high] 등장 허용 인물인 안규홍 외에 남성 두 명을 추가했다.",
     "[gpt-high] 실내에서 출입구를 배경으로 보아야 하는 카메라를 문 밖에 두어, 지정된 실내 시점과 진입 동선을 뒤집었다.",
     "[gpt-high] 포대의 한글 표기가 선명하게 판독되어 글씨 노출 금지 조건을 위반한다."
    ],
    "physics": "걷는 포즈로 양발이 지면에 닿아 체중을 지탱하고 있으며 중력에 어긋남이 없음."
   },
   {
    "label": "B",
    "direction": "안규홍은 문을 들어서며 고개를 들어 화면 우측 전방을 강렬하게 응시하고 있음.",
    "built_space": "참조 사진의 구조를 따른 곳간 내부. 좌측 후경에 밖으로 통하는 열린 문이 있으며, 내부에 지정된 가마니 더미와 그릇들, 부서진 자물쇠가 정확한 위치에 존재함.",
    "entities": "지정된 의상과 복면을 착용한 안규홍 단 한 명만 프레임에 존재함. 텍스트가 적힌 가마니와 바닥의 부서진 자물쇠가 요구사항과 일치함.",
    "hard_violations": [
     "[gpt-high] 포대 여러 개에 한글이 판독 가능하게 노출되어 '어디에도 읽을 수 있는 글씨가 없어야 한다'는 조건을 위반한다."
    ],
    "physics": "성큼 들어서는 동작을 포착하여 뒷발은 문지방 부근을, 앞발은 바닥을 확실히 딛고 있어 동작과 지지 기반이 사실적임."
   }
  ],
  "cross_model_order": {
   "policy": "cross_model_order_v2_slot0_fwd_slot1_rev_parallel2",
   "slots": [
    {
     "model": "gemini-pro",
     "order": "forward",
     "display_to_canonical": {
      "A": "A",
      "B": "B"
     },
     "raw": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 7,
        "verdict_ko": "샷 텍스트에 명시된 대로 복면을 쓴 안규홍 1인만을 정확히 묘사하였으며, 곳간으로 들어서는 역동적인 발걸음과 공간의 배치를 훌륭하게 구현했습니다."
       },
       {
        "label": "A",
        "score": 3,
        "verdict_ko": "곳간 배경과 주인공의 묘사는 준수하나, 샷 텍스트에 없는 인물 두 명을 추가로 배치한 치명적인 오류가 있습니다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "안규홍은 곳간 문턱을 넘어 화면 우측 하단을 향해 시선을 두고 걷고 있음.",
        "built_space": "곳간 내부 전경으로 왼쪽에 개방된 문이 있고, 참조 사진에 있는 가마니와 놋그릇 세트, 부서진 자물쇠가 바닥에 배치됨. 횃불(촛불)이 안쪽을 밝힘.",
        "entities": "복면을 쓴 안규홍 본인은 잘 묘사되었으나, 그의 뒤로 샷 텍스트에 언급되지 않은 조선인 남성 두 명이 추가로 서 있음. 자물쇠와 글씨가 적힌 가마니 등 프랍은 일치함.",
        "hard_violations": [
         "invented people (샷 텍스트에 없는 추가 인물 2명 등장)"
        ],
        "physics": "걷는 포즈로 양발이 지면에 닿아 체중을 지탱하고 있으며 중력에 어긋남이 없음."
       },
       {
        "label": "B",
        "direction": "안규홍은 문을 들어서며 고개를 들어 화면 우측 전방을 강렬하게 응시하고 있음.",
        "built_space": "참조 사진의 구조를 따른 곳간 내부. 좌측 후경에 밖으로 통하는 열린 문이 있으며, 내부에 지정된 가마니 더미와 그릇들, 부서진 자물쇠가 정확한 위치에 존재함.",
        "entities": "지정된 의상과 복면을 착용한 안규홍 단 한 명만 프레임에 존재함. 텍스트가 적힌 가마니와 바닥의 부서진 자물쇠가 요구사항과 일치함.",
        "hard_violations": [],
        "physics": "성큼 들어서는 동작을 포착하여 뒷발은 문지방 부근을, 앞발은 바닥을 확실히 딛고 있어 동작과 지지 기반이 사실적임."
       }
      ],
      "all_candidates_fail": false
     },
     "normalized": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 7,
        "verdict_ko": "샷 텍스트에 명시된 대로 복면을 쓴 안규홍 1인만을 정확히 묘사하였으며, 곳간으로 들어서는 역동적인 발걸음과 공간의 배치를 훌륭하게 구현했습니다."
       },
       {
        "label": "A",
        "score": 3,
        "verdict_ko": "곳간 배경과 주인공의 묘사는 준수하나, 샷 텍스트에 없는 인물 두 명을 추가로 배치한 치명적인 오류가 있습니다."
       }
      ],
      "all_candidates_fail": false,
      "readings": [
       {
        "label": "A",
        "direction": "안규홍은 곳간 문턱을 넘어 화면 우측 하단을 향해 시선을 두고 걷고 있음.",
        "built_space": "곳간 내부 전경으로 왼쪽에 개방된 문이 있고, 참조 사진에 있는 가마니와 놋그릇 세트, 부서진 자물쇠가 바닥에 배치됨. 횃불(촛불)이 안쪽을 밝힘.",
        "entities": "복면을 쓴 안규홍 본인은 잘 묘사되었으나, 그의 뒤로 샷 텍스트에 언급되지 않은 조선인 남성 두 명이 추가로 서 있음. 자물쇠와 글씨가 적힌 가마니 등 프랍은 일치함.",
        "hard_violations": [
         "invented people (샷 텍스트에 없는 추가 인물 2명 등장)"
        ],
        "physics": "걷는 포즈로 양발이 지면에 닿아 체중을 지탱하고 있으며 중력에 어긋남이 없음."
       },
       {
        "label": "B",
        "direction": "안규홍은 문을 들어서며 고개를 들어 화면 우측 전방을 강렬하게 응시하고 있음.",
        "built_space": "참조 사진의 구조를 따른 곳간 내부. 좌측 후경에 밖으로 통하는 열린 문이 있으며, 내부에 지정된 가마니 더미와 그릇들, 부서진 자물쇠가 정확한 위치에 존재함.",
        "entities": "지정된 의상과 복면을 착용한 안규홍 단 한 명만 프레임에 존재함. 텍스트가 적힌 가마니와 바닥의 부서진 자물쇠가 요구사항과 일치함.",
        "hard_violations": [],
        "physics": "성큼 들어서는 동작을 포착하여 뒷발은 문지방 부근을, 앞발은 바닥을 확실히 딛고 있어 동작과 지지 기반이 사실적임."
       }
      ]
     },
     "ok": true
    },
    {
     "model": "gpt-high",
     "order": "reverse",
     "display_to_canonical": {
      "A": "B",
      "B": "A"
     },
     "raw": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 4,
        "verdict_ko": "실내에서 비스듬히 본 왼쪽 위 출입구와 복면 인물의 전신 진입 동작은 맞지만, 포대의 읽을 수 있는 글씨가 명시적 금지 조건을 위반한다."
       },
       {
        "label": "B",
        "score": 1,
        "verdict_ko": "추가 인물 두 명과 읽히는 글씨가 등장하며, 실내 진입을 실외에서 바라본 퇴장처럼 뒤집어 핵심 시점과 동선을 어긴다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "안규홍의 시선과 몸통은 화면 오른쪽 곳간 내부를 향한다. 앞발은 실내 바닥에 놓이고 뒷발은 문턱을 넘어오는 중이어서 밖에서 안으로 진입하는 방향이 명확하다. 겨누는 무기나 지시하는 물체는 없다.",
        "built_space": "출입구 하나와 열린 목제 문짝 하나가 인물 뒤 왼쪽 위에 있고, 그 너머 밤의 외부 통로가 보인다. 카메라는 실내에서 출입구를 비스듬히 본다. 흙벽·목재 골조·흙바닥은 참조와 부합한다. 쌀가마니 더미는 중앙 뒤쪽, 세운 포대들은 오른쪽과 전경 가장자리에 있다. 낮은 불받침 하나가 뒤쪽 바닥에 있고 오른쪽 위에는 횃불 일부가 보인다. 귀중품 쟁반은 전경 양쪽에 나뉘어 있어 참조의 오른쪽 집중 배치와 다르지만 전신 동선을 막지는 않는다.",
        "entities": "인물은 한 명으로, 검은 머리를 뒤로 단정히 묶은 한국인 청년 남성이라는 설정에 부합하는 외양이다. 어두운 복면, 때 탄 밝은 저고리와 바지, 허리띠와 짚신이 참조 의상에 가깝다. 눈 주변 부기는 뚜렷하게 확인되지 않고 입술은 복면에 가려져 있으며 옥반지는 드러나지 않는다. 쌀가마니·종자 포대·은비녀·놋그릇과 바닥의 분리된 자물쇠 몸통 및 고리가 보인다. 포대에는 '서방', '종자', '이 씨', '볍씨' 등 읽을 수 있는 글씨가 있다.",
        "hard_violations": [
         "포대 여러 개에 한글이 판독 가능하게 노출되어 '어디에도 읽을 수 있는 글씨가 없어야 한다'는 조건을 위반한다."
        ],
        "physics": "실내에 디딘 앞발이 체중을 지탱하고, 뒤쪽 다리는 문턱을 넘도록 들어 올려져 있다. 몸의 전방 기울기와 벌어진 보폭은 성큼 들어오는 동작으로 가능하다. 포대는 바닥과 아래 포대에, 그릇과 비녀는 쟁반에, 자물쇠 조각은 바닥에 지지된다. 오른쪽 위 횃불은 대부분 화면 밖이라 손이나 고정구는 확인할 수 없지만, 공중에 독립적으로 떠 있는 물체로 보이지는 않는다."
       },
       {
        "label": "B",
        "direction": "주인공의 앞발과 몸은 뒤쪽 저장 공간에서 전경 문턱과 카메라 쪽으로 향한다. 시선은 화면 오른쪽 아래를 향한다. 출입구를 뒤에 두고 내부로 들어오기보다 내부에서 문 밖으로 나오는 방향으로 읽힌다. 뒤의 두 남성도 주인공을 따라 전경을 향한다.",
        "built_space": "출입구 하나와 왼쪽으로 열린 문짝 하나가 화면 전경을 둘러싸고, 문턱은 화면 아래를 가로지른다. 카메라는 곳간 밖에서 안을 바라보므로 요구된 실내 시점 및 왼쪽 위 배경 출입구 배치와 반대다. 왼쪽 가마니 더미, 오른쪽 포대와 귀중품 쟁반, 뒤쪽의 낮은 불받침 하나는 참조 배치와 가깝다. 추가 남성 두 명이 주인공 뒤 왼쪽 가마니 앞에 서 있다.",
        "entities": "복면을 쓴 검은 머리의 청년 남성 한 명 외에 복면 없는 남성 두 명이 더 있다. 주인공의 밝은 저고리·바지·짚신은 대체로 맞지만 참조의 두드러진 허리띠가 없고 의상이 더 깨끗하다. 눈 부기는 명확하지 않으며 입술은 가려져 있다. 화면 왼쪽 손에는 반지가 노출되어 있어 옥반지를 숨긴 상태와 맞지 않는다. 쌀가마니·종자 포대·은비녀·놋그릇·깨진 자물쇠가 보인다. 여러 포대에 '최서방종자' 등의 글씨가 선명하게 읽힌다.",
        "hard_violations": [
         "등장 허용 인물인 안규홍 외에 남성 두 명을 추가했다.",
         "실내에서 출입구를 배경으로 보아야 하는 카메라를 문 밖에 두어, 지정된 실내 시점과 진입 동선을 뒤집었다.",
         "포대의 한글 표기가 선명하게 판독되어 글씨 노출 금지 조건을 위반한다."
        ],
        "physics": "주인공은 뒤쪽 발로 바닥을 지지하면서 앞발을 전경으로 내딛는 자세이고, 추가 인물들의 발도 바닥에 닿아 있다. 보행 자세 자체가 부유하거나 해부학적으로 불가능하지는 않지만 동작 방향이 요청과 반대다. 포대와 자물쇠는 바닥에, 그릇과 비녀는 바닥에 놓인 쟁반에 지지된다."
       }
      ],
      "all_candidates_fail": true
     },
     "normalized": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 4,
        "verdict_ko": "실내에서 비스듬히 본 왼쪽 위 출입구와 복면 인물의 전신 진입 동작은 맞지만, 포대의 읽을 수 있는 글씨가 명시적 금지 조건을 위반한다."
       },
       {
        "label": "A",
        "score": 1,
        "verdict_ko": "추가 인물 두 명과 읽히는 글씨가 등장하며, 실내 진입을 실외에서 바라본 퇴장처럼 뒤집어 핵심 시점과 동선을 어긴다."
       }
      ],
      "all_candidates_fail": true,
      "readings": [
       {
        "label": "B",
        "direction": "안규홍의 시선과 몸통은 화면 오른쪽 곳간 내부를 향한다. 앞발은 실내 바닥에 놓이고 뒷발은 문턱을 넘어오는 중이어서 밖에서 안으로 진입하는 방향이 명확하다. 겨누는 무기나 지시하는 물체는 없다.",
        "built_space": "출입구 하나와 열린 목제 문짝 하나가 인물 뒤 왼쪽 위에 있고, 그 너머 밤의 외부 통로가 보인다. 카메라는 실내에서 출입구를 비스듬히 본다. 흙벽·목재 골조·흙바닥은 참조와 부합한다. 쌀가마니 더미는 중앙 뒤쪽, 세운 포대들은 오른쪽과 전경 가장자리에 있다. 낮은 불받침 하나가 뒤쪽 바닥에 있고 오른쪽 위에는 횃불 일부가 보인다. 귀중품 쟁반은 전경 양쪽에 나뉘어 있어 참조의 오른쪽 집중 배치와 다르지만 전신 동선을 막지는 않는다.",
        "entities": "인물은 한 명으로, 검은 머리를 뒤로 단정히 묶은 한국인 청년 남성이라는 설정에 부합하는 외양이다. 어두운 복면, 때 탄 밝은 저고리와 바지, 허리띠와 짚신이 참조 의상에 가깝다. 눈 주변 부기는 뚜렷하게 확인되지 않고 입술은 복면에 가려져 있으며 옥반지는 드러나지 않는다. 쌀가마니·종자 포대·은비녀·놋그릇과 바닥의 분리된 자물쇠 몸통 및 고리가 보인다. 포대에는 '서방', '종자', '이 씨', '볍씨' 등 읽을 수 있는 글씨가 있다.",
        "hard_violations": [
         "포대 여러 개에 한글이 판독 가능하게 노출되어 '어디에도 읽을 수 있는 글씨가 없어야 한다'는 조건을 위반한다."
        ],
        "physics": "실내에 디딘 앞발이 체중을 지탱하고, 뒤쪽 다리는 문턱을 넘도록 들어 올려져 있다. 몸의 전방 기울기와 벌어진 보폭은 성큼 들어오는 동작으로 가능하다. 포대는 바닥과 아래 포대에, 그릇과 비녀는 쟁반에, 자물쇠 조각은 바닥에 지지된다. 오른쪽 위 횃불은 대부분 화면 밖이라 손이나 고정구는 확인할 수 없지만, 공중에 독립적으로 떠 있는 물체로 보이지는 않는다."
       },
       {
        "label": "A",
        "direction": "주인공의 앞발과 몸은 뒤쪽 저장 공간에서 전경 문턱과 카메라 쪽으로 향한다. 시선은 화면 오른쪽 아래를 향한다. 출입구를 뒤에 두고 내부로 들어오기보다 내부에서 문 밖으로 나오는 방향으로 읽힌다. 뒤의 두 남성도 주인공을 따라 전경을 향한다.",
        "built_space": "출입구 하나와 왼쪽으로 열린 문짝 하나가 화면 전경을 둘러싸고, 문턱은 화면 아래를 가로지른다. 카메라는 곳간 밖에서 안을 바라보므로 요구된 실내 시점 및 왼쪽 위 배경 출입구 배치와 반대다. 왼쪽 가마니 더미, 오른쪽 포대와 귀중품 쟁반, 뒤쪽의 낮은 불받침 하나는 참조 배치와 가깝다. 추가 남성 두 명이 주인공 뒤 왼쪽 가마니 앞에 서 있다.",
        "entities": "복면을 쓴 검은 머리의 청년 남성 한 명 외에 복면 없는 남성 두 명이 더 있다. 주인공의 밝은 저고리·바지·짚신은 대체로 맞지만 참조의 두드러진 허리띠가 없고 의상이 더 깨끗하다. 눈 부기는 명확하지 않으며 입술은 가려져 있다. 화면 왼쪽 손에는 반지가 노출되어 있어 옥반지를 숨긴 상태와 맞지 않는다. 쌀가마니·종자 포대·은비녀·놋그릇·깨진 자물쇠가 보인다. 여러 포대에 '최서방종자' 등의 글씨가 선명하게 읽힌다.",
        "hard_violations": [
         "등장 허용 인물인 안규홍 외에 남성 두 명을 추가했다.",
         "실내에서 출입구를 배경으로 보아야 하는 카메라를 문 밖에 두어, 지정된 실내 시점과 진입 동선을 뒤집었다.",
         "포대의 한글 표기가 선명하게 판독되어 글씨 노출 금지 조건을 위반한다."
        ],
        "physics": "주인공은 뒤쪽 발로 바닥을 지지하면서 앞발을 전경으로 내딛는 자세이고, 추가 인물들의 발도 바닥에 닿아 있다. 보행 자세 자체가 부유하거나 해부학적으로 불가능하지는 않지만 동작 방향이 요청과 반대다. 포대와 자물쇠는 바닥에, 그릇과 비녀는 바닥에 놓인 쟁반에 지지된다."
       }
      ]
     },
     "ok": true
    }
   ],
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "slot_winner_match": true,
   "slot_winner": {
    "gemini-pro": "B",
    "gpt-high": "B"
   },
   "route": "cross_slot_agree"
  },
  "dual": {
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "normalized": {
    "A": 0.679,
    "B": 2.0
   },
   "adjusted": {
    "A": 0.429,
    "B": 1.75
   },
   "violations": {
    "A": [
     "[gemini-pro] invented people (샷 텍스트에 없는 추가 인물 2명 등장)",
     "[gpt-high] 등장 허용 인물인 안규홍 외에 남성 두 명을 추가했다.",
     "[gpt-high] 실내에서 출입구를 배경으로 보아야 하는 카메라를 문 밖에 두어, 지정된 실내 시점과 진입 동선을 뒤집었다.",
     "[gpt-high] 포대의 한글 표기가 선명하게 판독되어 글씨 노출 금지 조건을 위반한다."
    ],
    "B": [
     "[gpt-high] 포대 여러 개에 한글이 판독 가능하게 노출되어 '어디에도 읽을 수 있는 글씨가 없어야 한다'는 조건을 위반한다."
    ]
   },
   "per_model_winner": {
    "gemini-pro": "B",
    "gpt-high": "B"
   },
   "agreed": true
  },
  "totals": {
   "B": 1750,
   "A": 429
  },
  "selected": "B",
  "ranking": [
   "B",
   "A"
  ],
  "verdicts": [
   {
    "label": "B",
    "score": 1750,
    "verdict_ko": "샷 텍스트에 명시된 대로 복면을 쓴 안규홍 1인만을 정확히 묘사하였으며, 곳간으로 들어서는 역동적인 발걸음과 공간의 배치를 훌륭하게 구현했습니다.  ★위반: [gpt-high] 포대 여러 개에 한글이 판독 가능하게 노출되어 '어디에도 읽을 수 있는 글씨가 없어야 한다'는 조건을 위반한다."
   },
   {
    "label": "A",
    "score": 429,
    "verdict_ko": "곳간 배경과 주인공의 묘사는 준수하나, 샷 텍스트에 없는 인물 두 명을 추가로 배치한 치명적인 오류가 있습니다.  ★위반: [gemini-pro] invented people (샷 텍스트에 없는 추가 인물 2명 등장) / [gpt-high] 등장 허용 인물인 안규홍 외에 남성 두 명을 추가했다. / [gpt-high] 실내에서 출입구를 배경으로 보아야 하는 카메라를 문 밖에 두어, 지정된 실내 시점과 진입 동선을 뒤집었다. / [gpt-high] 포대의 한글 표기가 선명하게 판독되어 글씨 노출 금지 조건을 위반한다."
   }
  ],
  "refs": [
   {
    "label": "LOCATION PHOTOGRAPH — the exact place of this shot: its architecture, materials, fixed features and lighting mood are spatial truth; stage the moment inside this place. Never copy its camera framing.",
    "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/db00517b-165b-4005-a287-f28936b59b95/episodes/272bd7af-9294-4490-98c6-88a40e9deacc/images/background_chain/L20B01.png",
    "asset_id": "8ac26e75-11f5-4112-b737-07fc5fbda49a",
    "role": "location_plate"
   },
   {
    "label": "CHARACTER REFERENCE — 안규홍: the exact person appearing in this shot; match face, hair and build exactly.",
    "path": "<bytes:1028384>",
    "asset_id": "3b5894db-ce55-4f34-b966-5e725c3aa952",
    "role": "character_ref"
   }
  ],
  "critique_skipped": true,
  "shot_run_produced": false,
  "shot_run_uid": "06aa1418-9f74-780a-99dd-5086cf0a2530",
  "bgfirst": {
   "bg_path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/db00517b-165b-4005-a287-f28936b59b95/images/272bd7af-9294-4490-98c6-88a40e9deacc/scene/recipe/S22sh2__bgfirst_bg.png",
   "bg_asset_id": "dfba00df-2a39-454d-8669-2847b09f54c8",
   "bg_record_key": "S22sh2::bgfirst_bg",
   "chain_winner": false,
   "authority": "plate"
  },
  "ref_mode": "플레이트+엔티티 (2택1: 무콘티 승)",
  "share_plan": {
   "ref_plan": "background"
  }
 },
 "S22sh2::cine": {
  "applied": true,
  "attempted_at": "2026-09-09T03:22:42.474398+00:00",
  "fingerprint": "59edb5273edf3e1b6127a896fd603092e4567a3ccbcab1784aac8e4807d47f49",
  "fingerprint_version": 2,
  "provider": "mai",
  "endpoint": "openrouter/chat-completions",
  "model": "microsoft/mai-image-2.6",
  "pack": "24.202608252115",
  "source_file": "S22sh2_sel.png",
  "source_sha256": "7c47ebc518d183fc3350b62718806bdfb4942f79da30ded7f39145123faf6e1f",
  "file": "S22sh2_cine.png",
  "staged_sha256": "a484de456e8958de27da9e5865359b8a605e9438568a9be8e9e7db01a2b7c5a7",
  "latency_ms": 32601
 },
 "S22sh9::signage": {
  "fp": "598570f863a50175",
  "inscriptions": [],
  "cues": [],
  "dropped": []
 },
 "S22sh9": {
  "input_fingerprint": "f3c90f0fae3a9967",
  "prompt": "Create ONE FINAL photorealistic live-action film still of the moment below — early 20th-century Korea, Jeolla region; all people are Korean unless explicitly stated otherwise. TIME OF DAY (lock): night.\n\nSHOT TEXT (authoritative, Korean): 조선인 남성의 관자놀이에 서늘한 총구를 콱 들이대고 있는 염재보의 굳은 옆얼굴.\n\nLOCATION (lock): The passage beside stacked rice sacks inside the home's storehouse, illuminated by a pine torch at night. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: close-up\n- KEY BACKGROUND ELEMENTS: Gun (Muzzle pressed against the man's temple) — Seen from the side, with the muzzle terminating visibly at the temple; used as A narrow connecting line between the two faces, keeping the contact unambiguous; Rice-bale pile (Piled behind the confrontation) — Only a partial side of the pile is visible; used as Soft background continuity with the preceding struggle.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Maintain the established pine-knot firelight and muted nighttime palette, using firm but naturalistic separation around the faces and gun.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The people visible in that photo are NOT in this shot: never carry their faces, bodies or clothing onto anyone here. Each person in THIS shot is defined solely by the PEOPLE list, their own CHARACTER REFERENCE images and their wardrobe notes. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nPREVIOUS STILL USAGE (follow exactly — what to take from the attached still and what to exclude; it governs place and objects only, never who is in this shot or how the camera sees them): Take the granary doorway, dirt floor, stored grain sacks, and pine-torch lighting visible in the reference. Exclude an intact locked-door state, since the entrance has already been forced open.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The broken padlock remains on the floor; the storehouse door is open and the labeled seed sacks remain among the piled goods under resin light. 염재보: He is masked with his gun raised and held steady at head height. 조선인 남성(배상도 끄나풀 집 곳간): He has frozen with a club in hand and a gun muzzle against his temple.\n\nTHE HAND THAT IS DOING THIS: the object at the centre of this shot is being held, operated, read, aimed or handed over by 염재보 right now, so 염재보's hand — and as much of the wrist and forearm as the framing reaches — is in the frame, gripping or resting on that object exactly the way its real use requires. Match that hand to 염재보: its size, build, skin, age, grooming, sleeve and anything worn on it belong to that person and to no one else.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 염재보 (한국인 남성, 성인 나이대, 자연스러운 검은 머리); 조선인 남성(배상도 끄나풀 집 곳간) (한국인 남성, 성인 나이대, 자연스러운 검은 머리) — wearing: 배상도 일당의 앞잡이답게 평범한 무명옷 위에 거들먹거리는 느낌의 검은 조끼를 걸친 차림. — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.",
  "roll_prompts": {
   "A": "Create ONE FINAL photorealistic live-action film still of the moment below — early 20th-century Korea, Jeolla region; all people are Korean unless explicitly stated otherwise. TIME OF DAY (lock): night.\n\nSHOT TEXT (authoritative, Korean): 조선인 남성의 관자놀이에 서늘한 총구를 콱 들이대고 있는 염재보의 굳은 옆얼굴.\n\nLOCATION (lock): The passage beside stacked rice sacks inside the home's storehouse, illuminated by a pine torch at night. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: close-up\n- KEY BACKGROUND ELEMENTS: Gun (Muzzle pressed against the man's temple) — Seen from the side, with the muzzle terminating visibly at the temple; used as A narrow connecting line between the two faces, keeping the contact unambiguous; Rice-bale pile (Piled behind the confrontation) — Only a partial side of the pile is visible; used as Soft background continuity with the preceding struggle.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Maintain the established pine-knot firelight and muted nighttime palette, using firm but naturalistic separation around the faces and gun.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The people visible in that photo are NOT in this shot: never carry their faces, bodies or clothing onto anyone here. Each person in THIS shot is defined solely by the PEOPLE list, their own CHARACTER REFERENCE images and their wardrobe notes. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nPREVIOUS STILL USAGE (follow exactly — what to take from the attached still and what to exclude; it governs place and objects only, never who is in this shot or how the camera sees them): Take the granary doorway, dirt floor, stored grain sacks, and pine-torch lighting visible in the reference. Exclude an intact locked-door state, since the entrance has already been forced open.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The broken padlock remains on the floor; the storehouse door is open and the labeled seed sacks remain among the piled goods under resin light. 염재보: He is masked with his gun raised and held steady at head height. 조선인 남성(배상도 끄나풀 집 곳간): He has frozen with a club in hand and a gun muzzle against his temple.\n\nTHE HAND THAT IS DOING THIS: the object at the centre of this shot is being held, operated, read, aimed or handed over by 염재보 right now, so 염재보's hand — and as much of the wrist and forearm as the framing reaches — is in the frame, gripping or resting on that object exactly the way its real use requires. Match that hand to 염재보: its size, build, skin, age, grooming, sleeve and anything worn on it belong to that person and to no one else.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 염재보 (한국인 남성, 성인 나이대, 자연스러운 검은 머리); 조선인 남성(배상도 끄나풀 집 곳간) (한국인 남성, 성인 나이대, 자연스러운 검은 머리) — wearing: 배상도 일당의 앞잡이답게 평범한 무명옷 위에 거들먹거리는 느낌의 검은 조끼를 걸친 차림. — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.",
   "B": "Create ONE FINAL photorealistic live-action film still of the moment below — early 20th-century Korea, Jeolla region; all people are Korean unless explicitly stated otherwise. TIME OF DAY (lock): night.\n\nSHOT TEXT (authoritative, Korean): 조선인 남성의 관자놀이에 서늘한 총구를 콱 들이대고 있는 염재보의 굳은 옆얼굴.\n\nLOCATION (lock): The passage beside stacked rice sacks inside the home's storehouse, illuminated by a pine torch at night. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: close-up\n- KEY BACKGROUND ELEMENTS: Gun (Muzzle pressed against the man's temple) — Seen from the side, with the muzzle terminating visibly at the temple; used as A narrow connecting line between the two faces, keeping the contact unambiguous; Rice-bale pile (Piled behind the confrontation) — Only a partial side of the pile is visible; used as Soft background continuity with the preceding struggle.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Maintain the established pine-knot firelight and muted nighttime palette, using firm but naturalistic separation around the faces and gun.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The people visible in that photo are NOT in this shot: never carry their faces, bodies or clothing onto anyone here. Each person in THIS shot is defined solely by the PEOPLE list, their own CHARACTER REFERENCE images and their wardrobe notes. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nPREVIOUS STILL USAGE (follow exactly — what to take from the attached still and what to exclude; it governs place and objects only, never who is in this shot or how the camera sees them): Take the granary doorway, dirt floor, stored grain sacks, and pine-torch lighting visible in the reference. Exclude an intact locked-door state, since the entrance has already been forced open.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The broken padlock remains on the floor; the storehouse door is open and the labeled seed sacks remain among the piled goods under resin light. 염재보: He is masked with his gun raised and held steady at head height. 조선인 남성(배상도 끄나풀 집 곳간): He has frozen with a club in hand and a gun muzzle against his temple.\n\nTHE HAND THAT IS DOING THIS: the object at the centre of this shot is being held, operated, read, aimed or handed over by 염재보 right now, so 염재보's hand — and as much of the wrist and forearm as the framing reaches — is in the frame, gripping or resting on that object exactly the way its real use requires. Match that hand to 염재보: its size, build, skin, age, grooming, sleeve and anything worn on it belong to that person and to no one else.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 염재보 (한국인 남성, 성인 나이대, 자연스러운 검은 머리); 조선인 남성(배상도 끄나풀 집 곳간) (한국인 남성, 성인 나이대, 자연스러운 검은 머리) — wearing: 배상도 일당의 앞잡이답게 평범한 무명옷 위에 거들먹거리는 느낌의 검은 조끼를 걸친 차림. — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.\n\nCOMPOSITION VARIATION (this candidate only): keep the place, the people and the moment exactly as specified, and keep from the CAMERA & FRAME contract its subject scale, screen placement and key background elements exactly. For this candidate only, its camera angle and height are deliberately left open: choose a camera position distinctly different from the obvious one, as a film director picking a second setup on the same blocking. Stay true to any explicit framing the SHOT TEXT itself dictates."
  },
  "shot_run_spend_attempt_count": 1,
  "readings": [
   {
    "label": "A",
    "direction": "왼쪽 남성의 시선은 오른쪽 남성의 얼굴을 강렬하게 향하고 있으며, 권총의 총구는 오른쪽 남성의 관자놀이와 이마 부근에 정확히 맞닿아 있습니다.",
    "built_space": "어두운 곳간 내부입니다. 화면 왼쪽 뒤로 프롬프트가 요구한 열린 나무 문이 보이며, 오른쪽 뒤로는 쌀가마니가 쌓여 있고 화면 우측 밖의 횃불에서 빛이 들어옵니다.",
    "entities": "왼쪽 남성은 성인 한국인 남성이지만 지시된 복면을 착용하지 않았습니다. 오른쪽 남성은 흰옷 위에 검은 조끼를 입은 지시문과 일치하는 복장입니다. 권총과 몽둥이(나무 막대)가 명확히 보입니다.",
    "hard_violations": [],
    "physics": "왼쪽 남성의 오른손이 권총의 손잡이를 단단히 쥐어 지탱하고 있으며, 오른쪽 남성은 자신의 무기인 몽둥이를 손으로 쥐고 있습니다. 두 인물의 자세와 물체를 든 상태 모두 물리적으로 자연스럽습니다."
   },
   {
    "label": "B",
    "direction": "왼쪽 남성의 시선이 오른쪽 남성을 향하고, 총구는 오른쪽 남성의 이마에 닿아 있습니다.",
    "built_space": "곳간 내부이며, 인물들 바로 뒤편으로 쌀가마니 더미가 쌓여 있습니다.",
    "entities": "두 남성 모두 성인 한국인 남성으로 묘사되었으나 왼쪽 남성은 복면이 없습니다. 오른쪽 남성은 검은 조끼를 입고 있습니다.",
    "hard_violations": [
     "[gemini-pro] 총을 지탱하는 손이 전혀 존재하지 않으며, 총의 뒷부분이 왼쪽 남성의 코와 눈 부근에 완전히 융합되어 허공에 떠 있는 물리적 불가능 상태(Unsupported object, Physically impossible anatomy/staging)입니다.",
     "[gpt-high] 몽둥이를 하나 든 남성이라는 설정과 달리, 하단의 서로 다른 두 손에 각각 별개의 나무 손잡이가 나타나 불필요한 추가 소품이 생겼습니다."
    ],
    "physics": "오른쪽 남성은 몽둥이를 쥐고 있지만, 가장 중요한 물체인 권총을 잡고 있는 손이 없어 총이 공중에 떠서 얼굴에 붙어 있는 물리적 오류가 발생했습니다."
   }
  ],
  "cross_model_order": {
   "policy": "cross_model_order_v2_slot0_fwd_slot1_rev_parallel2",
   "slots": [
    {
     "model": "gemini-pro",
     "order": "forward",
     "display_to_canonical": {
      "A": "A",
      "B": "B"
     },
     "raw": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 7,
        "verdict_ko": "프롬프트가 요구한 클로즈업 앵글과 인물 간의 총구 겨눔(관자놀이에 밀착)을 훌륭하게 구현했으나, 염재보가 지시된 복면을 쓰지 않고 의상도 레퍼런스와 다르게 묘사된 점은 감점 요인입니다."
       },
       {
        "label": "B",
        "score": 0,
        "verdict_ko": "권총을 쥐고 있는 손이 전혀 없이 총이 인물의 얼굴에 융합되어 허공에 떠 있는 치명적인 물리적 오류(Hard Violation)가 발생하여 사용할 수 없는 이미지입니다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "왼쪽 남성의 시선은 오른쪽 남성의 얼굴을 강렬하게 향하고 있으며, 권총의 총구는 오른쪽 남성의 관자놀이와 이마 부근에 정확히 맞닿아 있습니다.",
        "built_space": "어두운 곳간 내부입니다. 화면 왼쪽 뒤로 프롬프트가 요구한 열린 나무 문이 보이며, 오른쪽 뒤로는 쌀가마니가 쌓여 있고 화면 우측 밖의 횃불에서 빛이 들어옵니다.",
        "entities": "왼쪽 남성은 성인 한국인 남성이지만 지시된 복면을 착용하지 않았습니다. 오른쪽 남성은 흰옷 위에 검은 조끼를 입은 지시문과 일치하는 복장입니다. 권총과 몽둥이(나무 막대)가 명확히 보입니다.",
        "hard_violations": [],
        "physics": "왼쪽 남성의 오른손이 권총의 손잡이를 단단히 쥐어 지탱하고 있으며, 오른쪽 남성은 자신의 무기인 몽둥이를 손으로 쥐고 있습니다. 두 인물의 자세와 물체를 든 상태 모두 물리적으로 자연스럽습니다."
       },
       {
        "label": "B",
        "direction": "왼쪽 남성의 시선이 오른쪽 남성을 향하고, 총구는 오른쪽 남성의 이마에 닿아 있습니다.",
        "built_space": "곳간 내부이며, 인물들 바로 뒤편으로 쌀가마니 더미가 쌓여 있습니다.",
        "entities": "두 남성 모두 성인 한국인 남성으로 묘사되었으나 왼쪽 남성은 복면이 없습니다. 오른쪽 남성은 검은 조끼를 입고 있습니다.",
        "hard_violations": [
         "총을 지탱하는 손이 전혀 존재하지 않으며, 총의 뒷부분이 왼쪽 남성의 코와 눈 부근에 완전히 융합되어 허공에 떠 있는 물리적 불가능 상태(Unsupported object, Physically impossible anatomy/staging)입니다."
        ],
        "physics": "오른쪽 남성은 몽둥이를 쥐고 있지만, 가장 중요한 물체인 권총을 잡고 있는 손이 없어 총이 공중에 떠서 얼굴에 붙어 있는 물리적 오류가 발생했습니다."
       }
      ],
      "all_candidates_fail": false
     },
     "normalized": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 7,
        "verdict_ko": "프롬프트가 요구한 클로즈업 앵글과 인물 간의 총구 겨눔(관자놀이에 밀착)을 훌륭하게 구현했으나, 염재보가 지시된 복면을 쓰지 않고 의상도 레퍼런스와 다르게 묘사된 점은 감점 요인입니다."
       },
       {
        "label": "B",
        "score": 0,
        "verdict_ko": "권총을 쥐고 있는 손이 전혀 없이 총이 인물의 얼굴에 융합되어 허공에 떠 있는 치명적인 물리적 오류(Hard Violation)가 발생하여 사용할 수 없는 이미지입니다."
       }
      ],
      "all_candidates_fail": false,
      "readings": [
       {
        "label": "A",
        "direction": "왼쪽 남성의 시선은 오른쪽 남성의 얼굴을 강렬하게 향하고 있으며, 권총의 총구는 오른쪽 남성의 관자놀이와 이마 부근에 정확히 맞닿아 있습니다.",
        "built_space": "어두운 곳간 내부입니다. 화면 왼쪽 뒤로 프롬프트가 요구한 열린 나무 문이 보이며, 오른쪽 뒤로는 쌀가마니가 쌓여 있고 화면 우측 밖의 횃불에서 빛이 들어옵니다.",
        "entities": "왼쪽 남성은 성인 한국인 남성이지만 지시된 복면을 착용하지 않았습니다. 오른쪽 남성은 흰옷 위에 검은 조끼를 입은 지시문과 일치하는 복장입니다. 권총과 몽둥이(나무 막대)가 명확히 보입니다.",
        "hard_violations": [],
        "physics": "왼쪽 남성의 오른손이 권총의 손잡이를 단단히 쥐어 지탱하고 있으며, 오른쪽 남성은 자신의 무기인 몽둥이를 손으로 쥐고 있습니다. 두 인물의 자세와 물체를 든 상태 모두 물리적으로 자연스럽습니다."
       },
       {
        "label": "B",
        "direction": "왼쪽 남성의 시선이 오른쪽 남성을 향하고, 총구는 오른쪽 남성의 이마에 닿아 있습니다.",
        "built_space": "곳간 내부이며, 인물들 바로 뒤편으로 쌀가마니 더미가 쌓여 있습니다.",
        "entities": "두 남성 모두 성인 한국인 남성으로 묘사되었으나 왼쪽 남성은 복면이 없습니다. 오른쪽 남성은 검은 조끼를 입고 있습니다.",
        "hard_violations": [
         "총을 지탱하는 손이 전혀 존재하지 않으며, 총의 뒷부분이 왼쪽 남성의 코와 눈 부근에 완전히 융합되어 허공에 떠 있는 물리적 불가능 상태(Unsupported object, Physically impossible anatomy/staging)입니다."
        ],
        "physics": "오른쪽 남성은 몽둥이를 쥐고 있지만, 가장 중요한 물체인 권총을 잡고 있는 손이 없어 총이 공중에 떠서 얼굴에 붙어 있는 물리적 오류가 발생했습니다."
       }
      ]
     },
     "ok": true
    },
    {
     "model": "gpt-high",
     "order": "reverse",
     "display_to_canonical": {
      "A": "B",
      "B": "A"
     },
     "raw": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 2,
        "verdict_ko": "얼굴 중심의 밀착 구도는 가깝지만, 복면과 염재보의 인물·의상이 틀리고 총구는 관자놀이보다 이마에 닿으며 불필요한 두 번째 나무 손잡이까지 보입니다."
       },
       {
        "label": "B",
        "score": 4,
        "verdict_ko": "염재보가 직접 총을 쥔 동작과 열린 곳간의 연속성은 더 명확하지만, 구도가 넓고 복면·인물·의상이 불일치하며 총구도 관자놀이가 아닌 이마를 누릅니다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "왼쪽 남성은 오른쪽 남성을 노려보고 오른쪽 남성도 그를 올려다봅니다. 리볼버는 왼쪽에서 오른쪽으로 향하지만 총구의 접촉점은 귀 앞쪽 관자놀이가 아니라 눈썹 위쪽 이마의 머리카락 경계입니다. 총열이 두 얼굴을 잇는 선으로 보인다는 조건은 충족합니다.",
        "built_space": "두 얼굴 뒤로 가로로 쌓인 곡물 자루 약 일곱 층과 오른쪽의 추가 자루 일부가 보입니다. 흙벽과 따뜻한 불빛은 이전 곳간과 어울리지만 자루 더미가 배경 대부분을 차지하고 그리 부드럽게 흐려지지 않았습니다. 문, 바닥, 자물쇠, 횃불 자체는 이 근접 구도 밖이므로 확인할 수 없습니다. 두 사람 사이에 구조물 간섭이나 불가능한 반사는 없습니다.",
        "entities": "성인 한국인 남성으로 설정 가능한 검은 머리의 남성 두 명이 보입니다. 왼쪽 총잡이는 복면 없이 얼굴 전체를 드러내며, 묶은 긴 머리와 밝은 무명옷은 염재보 참고의 얼굴·짧은 머리·짙은 의상과 다릅니다. 오른쪽 남성의 밝은 무명옷과 검은 조끼는 지정에 맞습니다. 리볼버 한 정과 오른쪽 아래의 굵은 몽둥이 외에, 아래 중앙의 다른 손에도 별도 나무 손잡이가 보입니다. 읽을 수 있는 글자는 없습니다.",
        "hard_violations": [
         "몽둥이를 하나 든 남성이라는 설정과 달리, 하단의 서로 다른 두 손에 각각 별개의 나무 손잡이가 나타나 불필요한 추가 소품이 생겼습니다."
        ],
        "physics": "총의 방아쇠울 안에 손가락은 보이지만 손잡이와 이를 감싸는 손바닥, 손목은 왼쪽 얼굴 뒤에 가려져 염재보의 실제 파지를 확인하기 어렵습니다. 가림만으로 총이 공중에 뜬다고 단정할 수는 없지만, 총을 쥔 손을 프레임에 명확히 보여 달라는 요구에는 못 미칩니다. 하단의 나무 물체들은 각각 손에 잡혀 있습니다. 몸은 화면 아래로 이어지며 도약이나 부유는 없습니다."
       },
       {
        "label": "B",
        "direction": "왼쪽 총잡이와 오른쪽 남성은 서로를 바라봅니다. 총잡이의 팔과 총열은 오른쪽 남성의 머리를 향하며 총구 접촉도 분명하지만, 접촉점은 관자놀이가 아니라 눈썹 안쪽 위의 이마입니다. 옆에서 본 총열이 두 얼굴 사이를 연결합니다.",
        "built_space": "왼쪽 뒤에 열린 나무문 한 짝과 문틀, 바깥의 어둠이 보이고 중앙 뒤에는 약 여섯 층의 곡물 자루 더미가 있습니다. 흙벽, 목재 기둥과 가로재, 흙바닥, 오른쪽 가장자리의 불꽃 하나가 이전 장소의 재료와 야간 조명을 잘 이어갑니다. 두 사람은 자루 앞 통로에 서 있습니다. 다만 상반신과 문·바닥까지 넓게 보여, 요구한 굳은 옆얼굴의 클로즈업보다 넓습니다. 불가능한 반사나 중복된 문은 없습니다.",
        "entities": "검은 머리의 성인 한국인 남성으로 보이는 두 사람이 있습니다. 왼쪽 총잡이는 복면이 없고 밝은 무명옷을 입어 염재보의 짙은 의상 참고와 맞지 않으며, 드러난 얼굴과 묶은 머리도 참고 인물과 다릅니다. 오른쪽 남성의 검은 조끼와 밝은 무명옷은 지정에 맞습니다. 리볼버 한 정과 가슴 앞의 나무 몽둥이 하나가 보입니다. 자루 표기나 다른 읽을 수 있는 글자는 없습니다.",
        "hard_violations": [],
        "physics": "총잡이의 손바닥과 손가락이 권총 손잡이를 감싸고, 검지가 방아쇠울 안에 있으며 손목과 소매 입은 팔이 연속적으로 보입니다. 팔로 머리 높이의 총을 지탱하는 자세는 가능합니다. 몽둥이는 오른쪽 남성의 몸 앞에서 화면 아래로 이어져 잡은 손은 잘렸지만, 이것만으로 무지지 부유라고 볼 근거는 없습니다. 두 상체는 직립하며 화면 밖 하체로 이어지고, 불가능한 신체 자세는 보이지 않습니다."
       }
      ],
      "all_candidates_fail": true
     },
     "normalized": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 2,
        "verdict_ko": "얼굴 중심의 밀착 구도는 가깝지만, 복면과 염재보의 인물·의상이 틀리고 총구는 관자놀이보다 이마에 닿으며 불필요한 두 번째 나무 손잡이까지 보입니다."
       },
       {
        "label": "A",
        "score": 4,
        "verdict_ko": "염재보가 직접 총을 쥔 동작과 열린 곳간의 연속성은 더 명확하지만, 구도가 넓고 복면·인물·의상이 불일치하며 총구도 관자놀이가 아닌 이마를 누릅니다."
       }
      ],
      "all_candidates_fail": true,
      "readings": [
       {
        "label": "B",
        "direction": "왼쪽 남성은 오른쪽 남성을 노려보고 오른쪽 남성도 그를 올려다봅니다. 리볼버는 왼쪽에서 오른쪽으로 향하지만 총구의 접촉점은 귀 앞쪽 관자놀이가 아니라 눈썹 위쪽 이마의 머리카락 경계입니다. 총열이 두 얼굴을 잇는 선으로 보인다는 조건은 충족합니다.",
        "built_space": "두 얼굴 뒤로 가로로 쌓인 곡물 자루 약 일곱 층과 오른쪽의 추가 자루 일부가 보입니다. 흙벽과 따뜻한 불빛은 이전 곳간과 어울리지만 자루 더미가 배경 대부분을 차지하고 그리 부드럽게 흐려지지 않았습니다. 문, 바닥, 자물쇠, 횃불 자체는 이 근접 구도 밖이므로 확인할 수 없습니다. 두 사람 사이에 구조물 간섭이나 불가능한 반사는 없습니다.",
        "entities": "성인 한국인 남성으로 설정 가능한 검은 머리의 남성 두 명이 보입니다. 왼쪽 총잡이는 복면 없이 얼굴 전체를 드러내며, 묶은 긴 머리와 밝은 무명옷은 염재보 참고의 얼굴·짧은 머리·짙은 의상과 다릅니다. 오른쪽 남성의 밝은 무명옷과 검은 조끼는 지정에 맞습니다. 리볼버 한 정과 오른쪽 아래의 굵은 몽둥이 외에, 아래 중앙의 다른 손에도 별도 나무 손잡이가 보입니다. 읽을 수 있는 글자는 없습니다.",
        "hard_violations": [
         "몽둥이를 하나 든 남성이라는 설정과 달리, 하단의 서로 다른 두 손에 각각 별개의 나무 손잡이가 나타나 불필요한 추가 소품이 생겼습니다."
        ],
        "physics": "총의 방아쇠울 안에 손가락은 보이지만 손잡이와 이를 감싸는 손바닥, 손목은 왼쪽 얼굴 뒤에 가려져 염재보의 실제 파지를 확인하기 어렵습니다. 가림만으로 총이 공중에 뜬다고 단정할 수는 없지만, 총을 쥔 손을 프레임에 명확히 보여 달라는 요구에는 못 미칩니다. 하단의 나무 물체들은 각각 손에 잡혀 있습니다. 몸은 화면 아래로 이어지며 도약이나 부유는 없습니다."
       },
       {
        "label": "A",
        "direction": "왼쪽 총잡이와 오른쪽 남성은 서로를 바라봅니다. 총잡이의 팔과 총열은 오른쪽 남성의 머리를 향하며 총구 접촉도 분명하지만, 접촉점은 관자놀이가 아니라 눈썹 안쪽 위의 이마입니다. 옆에서 본 총열이 두 얼굴 사이를 연결합니다.",
        "built_space": "왼쪽 뒤에 열린 나무문 한 짝과 문틀, 바깥의 어둠이 보이고 중앙 뒤에는 약 여섯 층의 곡물 자루 더미가 있습니다. 흙벽, 목재 기둥과 가로재, 흙바닥, 오른쪽 가장자리의 불꽃 하나가 이전 장소의 재료와 야간 조명을 잘 이어갑니다. 두 사람은 자루 앞 통로에 서 있습니다. 다만 상반신과 문·바닥까지 넓게 보여, 요구한 굳은 옆얼굴의 클로즈업보다 넓습니다. 불가능한 반사나 중복된 문은 없습니다.",
        "entities": "검은 머리의 성인 한국인 남성으로 보이는 두 사람이 있습니다. 왼쪽 총잡이는 복면이 없고 밝은 무명옷을 입어 염재보의 짙은 의상 참고와 맞지 않으며, 드러난 얼굴과 묶은 머리도 참고 인물과 다릅니다. 오른쪽 남성의 검은 조끼와 밝은 무명옷은 지정에 맞습니다. 리볼버 한 정과 가슴 앞의 나무 몽둥이 하나가 보입니다. 자루 표기나 다른 읽을 수 있는 글자는 없습니다.",
        "hard_violations": [],
        "physics": "총잡이의 손바닥과 손가락이 권총 손잡이를 감싸고, 검지가 방아쇠울 안에 있으며 손목과 소매 입은 팔이 연속적으로 보입니다. 팔로 머리 높이의 총을 지탱하는 자세는 가능합니다. 몽둥이는 오른쪽 남성의 몸 앞에서 화면 아래로 이어져 잡은 손은 잘렸지만, 이것만으로 무지지 부유라고 볼 근거는 없습니다. 두 상체는 직립하며 화면 밖 하체로 이어지고, 불가능한 신체 자세는 보이지 않습니다."
       }
      ]
     },
     "ok": true
    }
   ],
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "slot_winner_match": true,
   "slot_winner": {
    "gemini-pro": "A",
    "gpt-high": "A"
   },
   "route": "cross_slot_agree"
  },
  "dual": {
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "normalized": {
    "A": 2.0,
    "B": 0.5
   },
   "adjusted": {
    "A": 2.0,
    "B": 0.25
   },
   "violations": {
    "B": [
     "[gemini-pro] 총을 지탱하는 손이 전혀 존재하지 않으며, 총의 뒷부분이 왼쪽 남성의 코와 눈 부근에 완전히 융합되어 허공에 떠 있는 물리적 불가능 상태(Unsupported object, Physically impossible anatomy/staging)입니다.",
     "[gpt-high] 몽둥이를 하나 든 남성이라는 설정과 달리, 하단의 서로 다른 두 손에 각각 별개의 나무 손잡이가 나타나 불필요한 추가 소품이 생겼습니다."
    ]
   },
   "per_model_winner": {
    "gemini-pro": "A",
    "gpt-high": "A"
   },
   "agreed": true
  },
  "totals": {
   "A": 2000,
   "B": 250
  },
  "selected": "A",
  "ranking": [
   "A",
   "B"
  ],
  "verdicts": [
   {
    "label": "A",
    "score": 2000,
    "verdict_ko": "프롬프트가 요구한 클로즈업 앵글과 인물 간의 총구 겨눔(관자놀이에 밀착)을 훌륭하게 구현했으나, 염재보가 지시된 복면을 쓰지 않고 의상도 레퍼런스와 다르게 묘사된 점은 감점 요인입니다."
   },
   {
    "label": "B",
    "score": 250,
    "verdict_ko": "권총을 쥐고 있는 손이 전혀 없이 총이 인물의 얼굴에 융합되어 허공에 떠 있는 치명적인 물리적 오류(Hard Violation)가 발생하여 사용할 수 없는 이미지입니다.  ★위반: [gemini-pro] 총을 지탱하는 손이 전혀 존재하지 않으며, 총의 뒷부분이 왼쪽 남성의 코와 눈 부근에 완전히 융합되어 허공에 떠 있는 물리적 불가능 상태(Unsupported object, Physically impossible anatomy/staging)입니다. / [gpt-high] 몽둥이를 하나 든 남성이라는 설정과 달리, 하단의 서로 다른 두 손에 각각 별개의 나무 손잡이가 나타나 불필요한 추가 소품이 생겼습니다."
   }
  ],
  "refs": [
   {
    "label": "PREVIOUS SHOT STILL — a visually related earlier shot of this same place: the location's look, materials, fixed features and lighting mood are LOCKED to this photo; never copy its camera framing. The people visible in that photo are NOT in this shot: never carry their faces, bodies or clothing onto anyone here. Each person in THIS shot is defined solely by the PEOPLE list, their own CHARACTER REFERENCE images and their wardrobe notes. If this photo shows a character who cannot move (dead or unconscious), that character's exact pose, position and orientation are ALSO LOCKED — treat their body as a fixed prop of the set that only the camera moves around.",
    "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/db00517b-165b-4005-a287-f28936b59b95/images/272bd7af-9294-4490-98c6-88a40e9deacc/scene/recipe/S22sh2_sel.png",
    "asset_id": "4906f314-8cec-4877-b2c7-f99b3a303b84",
    "role": "prev_still"
   },
   {
    "label": "CHARACTER REFERENCE — 염재보: the exact person appearing in this shot; match face, hair and build exactly.",
    "path": "<bytes:964351>",
    "asset_id": "ab287289-d0b7-436d-8f70-68452f898f73",
    "role": "character_ref"
   }
  ],
  "critique_skipped": true,
  "shot_run_produced": false,
  "shot_run_uid": "06aa141a-812e-719b-89f5-c3d8bfb472d0",
  "ref_mode": "prev+엔티티",
  "share_plan": {
   "ref_plan": "prev",
   "prev_anchor_tag": "S22sh2"
  }
 },
 "S22sh9::cine": {
  "applied": true,
  "attempted_at": "2026-09-09T03:24:11.107668+00:00",
  "fingerprint": "0674b7dc9abd785593945d8f661850c989dfb6dc45509dd44b787906c8c0425e",
  "fingerprint_version": 2,
  "provider": "mai",
  "endpoint": "openrouter/chat-completions",
  "model": "microsoft/mai-image-2.6",
  "pack": "24.202608252115",
  "source_file": "S22sh9_sel.png",
  "source_sha256": "00ac9f7c13fffe818439c130271cb93e1b84667ebc107b2bec8772f0110af410",
  "file": "S22sh9_cine.png",
  "staged_sha256": "0dd4e8108755fce952e5d67612501f71984009e179eb37b03fe832d248395f41",
  "latency_ms": 26214
 },
 "S22sh11::signage": {
  "fp": "5cd0669948e3ec8d",
  "inscriptions": [],
  "cues": [],
  "dropped": []
 },
 "S22sh11": {
  "input_fingerprint": "5f87fef8225331a6",
  "prompt": "Create ONE FINAL photorealistic live-action film still of the moment below — early 20th-century Korea, Jeolla region; all people are Korean unless explicitly stated otherwise. TIME OF DAY (lock): night.\n\nSHOT TEXT (authoritative, Korean): 묵직한 씨나락 자루를 한쪽 어깨 위로 번쩍 둘러멘 안규홍의 거대한 상체.\n\nLOCATION (lock): The seed-grain storage area inside the torchlit granary, beside labeled sacks and piled household valuables. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: medium shot\n- KEY BACKGROUND ELEMENTS: Seed sack on 안규홍's shoulder (Lifted onto one shoulder and still supported by his arm) — The sack's side is visible above the supporting shoulder; no inscription is presented to the camera; used as Makes the burden tangible while preserving the shoulder and torso as scale references; Remaining stored sacks (Piled within the storehouse) — Partial sides remain visible low behind 안규홍; used as Maintains the connection to the supply pile without competing with the lifting action.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Continue the same pine-knot firelight and restrained nighttime contrast, retaining detail in the supporting arm and masked face.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The people visible in that photo are NOT in this shot: never carry their faces, bodies or clothing onto anyone here. Each person in THIS shot is defined solely by the PEOPLE list, their own CHARACTER REFERENCE images and their wardrobe notes. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nPREVIOUS STILL USAGE (follow exactly — what to take from the attached still and what to exclude; it governs place and objects only, never who is in this shot or how the camera sees them): Take the granary's structural materials, dirt floor, remaining grain stores, and torchlit color. Exclude a duplicate of the seed-grain sack now being carried and do not restore the broken lock.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The broken padlock and open door remain; labeled seed sacks and other stolen goods are still piled under pine-resin light. 안규홍: He is masked and carries a seed-rice sack on his shoulder; his eye and lip remain injured, and the jade ring stays concealed.\n\nTHE HAND THAT IS DOING THIS: the object at the centre of this shot is being held, operated, read, aimed or handed over by 안규홍 right now, so 안규홍's hand — and as much of the wrist and forearm as the framing reaches — is in the frame, gripping or resting on that object exactly the way its real use requires. Match that hand to 안규홍: its size, build, skin, age, grooming, sleeve and anything worn on it belong to that person and to no one else.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 안규홍 (한국인 남성, 청년 나이대, 검은 머리, 단단히 묶은 머리) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.",
  "roll_prompts": {
   "A": "Create ONE FINAL photorealistic live-action film still of the moment below — early 20th-century Korea, Jeolla region; all people are Korean unless explicitly stated otherwise. TIME OF DAY (lock): night.\n\nSHOT TEXT (authoritative, Korean): 묵직한 씨나락 자루를 한쪽 어깨 위로 번쩍 둘러멘 안규홍의 거대한 상체.\n\nLOCATION (lock): The seed-grain storage area inside the torchlit granary, beside labeled sacks and piled household valuables. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: medium shot\n- KEY BACKGROUND ELEMENTS: Seed sack on 안규홍's shoulder (Lifted onto one shoulder and still supported by his arm) — The sack's side is visible above the supporting shoulder; no inscription is presented to the camera; used as Makes the burden tangible while preserving the shoulder and torso as scale references; Remaining stored sacks (Piled within the storehouse) — Partial sides remain visible low behind 안규홍; used as Maintains the connection to the supply pile without competing with the lifting action.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Continue the same pine-knot firelight and restrained nighttime contrast, retaining detail in the supporting arm and masked face.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The people visible in that photo are NOT in this shot: never carry their faces, bodies or clothing onto anyone here. Each person in THIS shot is defined solely by the PEOPLE list, their own CHARACTER REFERENCE images and their wardrobe notes. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nPREVIOUS STILL USAGE (follow exactly — what to take from the attached still and what to exclude; it governs place and objects only, never who is in this shot or how the camera sees them): Take the granary's structural materials, dirt floor, remaining grain stores, and torchlit color. Exclude a duplicate of the seed-grain sack now being carried and do not restore the broken lock.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The broken padlock and open door remain; labeled seed sacks and other stolen goods are still piled under pine-resin light. 안규홍: He is masked and carries a seed-rice sack on his shoulder; his eye and lip remain injured, and the jade ring stays concealed.\n\nTHE HAND THAT IS DOING THIS: the object at the centre of this shot is being held, operated, read, aimed or handed over by 안규홍 right now, so 안규홍's hand — and as much of the wrist and forearm as the framing reaches — is in the frame, gripping or resting on that object exactly the way its real use requires. Match that hand to 안규홍: its size, build, skin, age, grooming, sleeve and anything worn on it belong to that person and to no one else.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 안규홍 (한국인 남성, 청년 나이대, 검은 머리, 단단히 묶은 머리) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.",
   "B": "Create ONE FINAL photorealistic live-action film still of the moment below — early 20th-century Korea, Jeolla region; all people are Korean unless explicitly stated otherwise. TIME OF DAY (lock): night.\n\nSHOT TEXT (authoritative, Korean): 묵직한 씨나락 자루를 한쪽 어깨 위로 번쩍 둘러멘 안규홍의 거대한 상체.\n\nLOCATION (lock): The seed-grain storage area inside the torchlit granary, beside labeled sacks and piled household valuables. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: medium shot\n- KEY BACKGROUND ELEMENTS: Seed sack on 안규홍's shoulder (Lifted onto one shoulder and still supported by his arm) — The sack's side is visible above the supporting shoulder; no inscription is presented to the camera; used as Makes the burden tangible while preserving the shoulder and torso as scale references; Remaining stored sacks (Piled within the storehouse) — Partial sides remain visible low behind 안규홍; used as Maintains the connection to the supply pile without competing with the lifting action.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Continue the same pine-knot firelight and restrained nighttime contrast, retaining detail in the supporting arm and masked face.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The people visible in that photo are NOT in this shot: never carry their faces, bodies or clothing onto anyone here. Each person in THIS shot is defined solely by the PEOPLE list, their own CHARACTER REFERENCE images and their wardrobe notes. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nPREVIOUS STILL USAGE (follow exactly — what to take from the attached still and what to exclude; it governs place and objects only, never who is in this shot or how the camera sees them): Take the granary's structural materials, dirt floor, remaining grain stores, and torchlit color. Exclude a duplicate of the seed-grain sack now being carried and do not restore the broken lock.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The broken padlock and open door remain; labeled seed sacks and other stolen goods are still piled under pine-resin light. 안규홍: He is masked and carries a seed-rice sack on his shoulder; his eye and lip remain injured, and the jade ring stays concealed.\n\nTHE HAND THAT IS DOING THIS: the object at the centre of this shot is being held, operated, read, aimed or handed over by 안규홍 right now, so 안규홍's hand — and as much of the wrist and forearm as the framing reaches — is in the frame, gripping or resting on that object exactly the way its real use requires. Match that hand to 안규홍: its size, build, skin, age, grooming, sleeve and anything worn on it belong to that person and to no one else.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 안규홍 (한국인 남성, 청년 나이대, 검은 머리, 단단히 묶은 머리) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.\n\nCOMPOSITION VARIATION (this candidate only): keep the place, the people and the moment exactly as specified, and keep from the CAMERA & FRAME contract its subject scale, screen placement and key background elements exactly. For this candidate only, its camera angle and height are deliberately left open: choose a camera position distinctly different from the obvious one, as a film director picking a second setup on the same blocking. Stay true to any explicit framing the SHOT TEXT itself dictates."
  },
  "shot_run_spend_attempt_count": 1,
  "readings": [
   {
    "label": "A",
    "direction": "인물의 시선은 화면 오른쪽 밖을 향하고 있으며, 오른쪽 팔로 어깨 위의 자루를 지탱하고 있습니다.",
    "built_space": "나무 기둥과 흙바닥으로 이루어진 창고 내부이며, 뒤쪽에는 곡식 자루들이 쌓여 있습니다. 화면 오른쪽 끝에 횃불이 타고 있습니다.",
    "entities": "안규홍의 외모는 레퍼런스와 비슷하나 프롬프트가 요구한 복면을 착용하지 않았습니다. 어깨에 멘 자루에 희미한 자국이 있고, 배경 왼쪽 아래의 자루에 알파벳(SUP...) 같은 읽을 수 있는 글씨가 있습니다.",
    "hard_violations": [
     "[gemini-pro] 프롬프트에 명시된 복면(마스크) 미착용",
     "[gemini-pro] 배경 자루에 읽을 수 있는 문자 포함(글씨 금지 조항 위반)",
     "[gpt-high] 어깨에 멘 자루 윗면에 한글 글자가 노출되어, 읽을 수 있는 글씨를 배제하라는 조건을 위반한다."
    ],
    "physics": "자루가 인물의 오른쪽 어깨에 얹혀 있으며, 오른쪽 팔의 근육과 자세가 무거운 하중을 지탱하는 모습을 자연스럽게 보여줍니다."
   },
   {
    "label": "B",
    "direction": "인물의 시선은 오른쪽 아래를 향하고 있으며, 두 손으로 어깨 위의 자루를 단단히 잡고 있습니다.",
    "built_space": "이전 샷의 연속성을 살려 화면 왼쪽에 부서진 자물쇠가 달린 채 열려 있는 문이 보입니다. 뒤쪽으로 자루들이 쌓여 있으며, 오른쪽에서 횃불이 타오르고 있습니다.",
    "entities": "안규홍은 요구된 대로 하관을 가리는 복면을 썼으며 눈 주변의 상처가 확인됩니다. 어깨에 멘 곡식 자루와 배경의 자루들 모두 글씨 없이 재질이 잘 표현되었습니다.",
    "hard_violations": [],
    "physics": "오른쪽 어깨에 멘 무거운 자루를 양손으로 끌어당겨 지탱하고 있으며, 무게 중심과 근육의 긴장감이 사실적으로 묘사되었습니다."
   }
  ],
  "cross_model_order": {
   "policy": "cross_model_order_v2_slot0_fwd_slot1_rev_parallel2",
   "slots": [
    {
     "model": "gemini-pro",
     "order": "forward",
     "display_to_canonical": {
      "A": "A",
      "B": "B"
     },
     "raw": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 9,
        "verdict_ko": "프롬프트가 요구한 복면(마스크), 상처 입은 눈, 그리고 배경의 열린 문과 부서진 자물쇠까지 완벽하게 구현하여 디테일과 연속성을 훌륭하게 살려냈습니다."
       },
       {
        "label": "A",
        "score": 4,
        "verdict_ko": "인물이 프롬프트에 명시된 복면을 착용하지 않았으며, 배경의 자루에 금지된 읽을 수 있는 문자가 표기되어 있어 중요한 지시 사항을 위반했습니다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "인물의 시선은 화면 오른쪽 밖을 향하고 있으며, 오른쪽 팔로 어깨 위의 자루를 지탱하고 있습니다.",
        "built_space": "나무 기둥과 흙바닥으로 이루어진 창고 내부이며, 뒤쪽에는 곡식 자루들이 쌓여 있습니다. 화면 오른쪽 끝에 횃불이 타고 있습니다.",
        "entities": "안규홍의 외모는 레퍼런스와 비슷하나 프롬프트가 요구한 복면을 착용하지 않았습니다. 어깨에 멘 자루에 희미한 자국이 있고, 배경 왼쪽 아래의 자루에 알파벳(SUP...) 같은 읽을 수 있는 글씨가 있습니다.",
        "hard_violations": [
         "프롬프트에 명시된 복면(마스크) 미착용",
         "배경 자루에 읽을 수 있는 문자 포함(글씨 금지 조항 위반)"
        ],
        "physics": "자루가 인물의 오른쪽 어깨에 얹혀 있으며, 오른쪽 팔의 근육과 자세가 무거운 하중을 지탱하는 모습을 자연스럽게 보여줍니다."
       },
       {
        "label": "B",
        "direction": "인물의 시선은 오른쪽 아래를 향하고 있으며, 두 손으로 어깨 위의 자루를 단단히 잡고 있습니다.",
        "built_space": "이전 샷의 연속성을 살려 화면 왼쪽에 부서진 자물쇠가 달린 채 열려 있는 문이 보입니다. 뒤쪽으로 자루들이 쌓여 있으며, 오른쪽에서 횃불이 타오르고 있습니다.",
        "entities": "안규홍은 요구된 대로 하관을 가리는 복면을 썼으며 눈 주변의 상처가 확인됩니다. 어깨에 멘 곡식 자루와 배경의 자루들 모두 글씨 없이 재질이 잘 표현되었습니다.",
        "hard_violations": [],
        "physics": "오른쪽 어깨에 멘 무거운 자루를 양손으로 끌어당겨 지탱하고 있으며, 무게 중심과 근육의 긴장감이 사실적으로 묘사되었습니다."
       }
      ],
      "all_candidates_fail": false
     },
     "normalized": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 9,
        "verdict_ko": "프롬프트가 요구한 복면(마스크), 상처 입은 눈, 그리고 배경의 열린 문과 부서진 자물쇠까지 완벽하게 구현하여 디테일과 연속성을 훌륭하게 살려냈습니다."
       },
       {
        "label": "A",
        "score": 4,
        "verdict_ko": "인물이 프롬프트에 명시된 복면을 착용하지 않았으며, 배경의 자루에 금지된 읽을 수 있는 문자가 표기되어 있어 중요한 지시 사항을 위반했습니다."
       }
      ],
      "all_candidates_fail": false,
      "readings": [
       {
        "label": "A",
        "direction": "인물의 시선은 화면 오른쪽 밖을 향하고 있으며, 오른쪽 팔로 어깨 위의 자루를 지탱하고 있습니다.",
        "built_space": "나무 기둥과 흙바닥으로 이루어진 창고 내부이며, 뒤쪽에는 곡식 자루들이 쌓여 있습니다. 화면 오른쪽 끝에 횃불이 타고 있습니다.",
        "entities": "안규홍의 외모는 레퍼런스와 비슷하나 프롬프트가 요구한 복면을 착용하지 않았습니다. 어깨에 멘 자루에 희미한 자국이 있고, 배경 왼쪽 아래의 자루에 알파벳(SUP...) 같은 읽을 수 있는 글씨가 있습니다.",
        "hard_violations": [
         "프롬프트에 명시된 복면(마스크) 미착용",
         "배경 자루에 읽을 수 있는 문자 포함(글씨 금지 조항 위반)"
        ],
        "physics": "자루가 인물의 오른쪽 어깨에 얹혀 있으며, 오른쪽 팔의 근육과 자세가 무거운 하중을 지탱하는 모습을 자연스럽게 보여줍니다."
       },
       {
        "label": "B",
        "direction": "인물의 시선은 오른쪽 아래를 향하고 있으며, 두 손으로 어깨 위의 자루를 단단히 잡고 있습니다.",
        "built_space": "이전 샷의 연속성을 살려 화면 왼쪽에 부서진 자물쇠가 달린 채 열려 있는 문이 보입니다. 뒤쪽으로 자루들이 쌓여 있으며, 오른쪽에서 횃불이 타오르고 있습니다.",
        "entities": "안규홍은 요구된 대로 하관을 가리는 복면을 썼으며 눈 주변의 상처가 확인됩니다. 어깨에 멘 곡식 자루와 배경의 자루들 모두 글씨 없이 재질이 잘 표현되었습니다.",
        "hard_violations": [],
        "physics": "오른쪽 어깨에 멘 무거운 자루를 양손으로 끌어당겨 지탱하고 있으며, 무게 중심과 근육의 긴장감이 사실적으로 묘사되었습니다."
       }
      ]
     },
     "ok": true
    },
    {
     "model": "gpt-high",
     "order": "reverse",
     "display_to_canonical": {
      "A": "B",
      "B": "A"
     },
     "raw": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 8,
        "verdict_ko": "복면을 쓴 안규홍의 상체와 어깨·팔로 받친 무거운 자루를 미디엄 숏으로 구현했으나, 눈의 부상과 자물쇠의 파손 상태는 불명확하다."
       },
       {
        "label": "B",
        "score": 4,
        "verdict_ko": "어깨와 손의 하중 지지는 자연스럽지만 자루의 한글 노출이 금지 조건을 어기며, 민소매 복장과 확인되지 않는 복면도 인물 설정에 미달한다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "얼굴과 시선은 화면 오른쪽 아래의 화면 밖 공간을 향하며 렌즈를 응시하지 않는다. 구체적인 시선 목표는 지정되지 않았다. 자루의 넓은 무문자 옆면이 카메라를 향하고, 들어 올린 손은 어깨 위 자루의 천을 움켜쥔다.",
        "built_space": "왼쪽에 열린 널판문 하나와 문틀, 뒤쪽에 목재 기둥과 가로대가 있는 흙벽, 아래에 흙바닥이 보인다. 왼쪽 뒤에는 약 여섯 단의 곡식 자루와 일부 가려진 추가 자루가 있고, 오른쪽에는 횃불 하나와 살림살이 더미가 있다. 인물은 문 안쪽 저장물 앞에 서 있다. 문 걸쇠 부근의 자물쇠 모양 금속물은 보이지만 파손 여부는 판별하기 어렵다. 참조의 재료와 야간 불빛은 대체로 유지된다.",
        "entities": "청년 동아시아계 남성 한 명으로 한국인 설정과 모순되지 않으며, 검은 머리를 뒤로 묶고 검은 천으로 코와 입을 가렸다. 때 묻은 밝은 저고리와 걷어 올린 소매는 인물 참조에 가깝다. 눈 주변의 명확한 상처는 확인되지 않고 입술은 복면에 가려져 있다. 옥반지는 드러나지 않는다. 어깨의 자루는 거친 황갈색 마대이며 참조와 재질은 유사하지만 묶인 입구는 뚜렷하지 않다. 배경에는 저장 자루와 생활용품이 있으며 읽을 수 있는 글자는 보이지 않는다.",
        "hard_violations": [],
        "physics": "자루 밑면이 한쪽 어깨에 눌려 얹혀 있고, 같은 쪽의 굽힌 팔과 천을 쥔 손이 하중과 미끄러짐을 제어한다. 상체가 앞으로 기울어 무게에 대응하며 다른 팔은 몸 앞쪽으로 접혀 있다. 발은 프레임 밖이지만 몸이 떠 있다는 징후는 없다. 배경 자루는 바닥과 아래 자루에, 살림살이는 받침 면에 놓여 있고 횃불에는 지지대가 보인다."
       },
       {
        "label": "B",
        "direction": "남성은 오른쪽을 향한 옆모습으로 화면 오른쪽 아래를 본다. 자루는 한쪽 어깨에서 등 뒤로 길게 걸쳐지고 손은 앞쪽으로 내려온 입구를 잡는다. 카메라를 향한 자루 윗면에 한글 글자가 노출되어, 글씨 없는 옆면을 보여 달라는 지시와 어긋난다.",
        "built_space": "목재 세로 기둥과 가로대가 있는 흙벽 앞에 인물이 서 있다. 왼쪽 아래에는 약 다섯 단의 자루와 그 옆의 가려진 자루 더미, 오른쪽 아래에는 천 보따리와 상자류가 놓여 있다. 오른쪽 가장자리에는 횃불 하나가 크게 잘려 보인다. 문과 자물쇠, 흙바닥은 프레임 밖이라 상태를 확인할 수 없다. 창고 재료는 참조와 유사하지만 가까운 횃불과 큰 자루가 상체와 경쟁한다.",
        "entities": "검은 머리를 묶은 청년 동아시아계 남성 한 명이며, 드러난 팔은 인물 참조보다 훨씬 근육질이다. 밝고 때 묻은 옷이지만 소매가 거의 없는 형태여서 참조의 긴소매 저고리와 다르다. 얼굴 아래쪽은 자루에 가려져 검은 복면이 확인되지 않으며 눈의 상처도 뚜렷하지 않다. 입술과 옥반지는 보이지 않는다. 마대 자루의 거친 직조와 묶인 입구는 소품 참조에 가깝지만, 카메라 쪽 한글 표기는 명시적인 금지 사항과 충돌한다.",
        "hard_violations": [
         "어깨에 멘 자루 윗면에 한글 글자가 노출되어, 읽을 수 있는 글씨를 배제하라는 조건을 위반한다."
        ],
        "physics": "자루는 어깨와 등 위에 직접 얹혀 있고, 굽힌 팔의 손이 앞쪽 자루 입구를 잡아 지지한다. 충전물이 어깨 양쪽으로 처지는 모양과 앞으로 기운 몸통은 무거운 짐을 멘 동작으로 가능하다. 발은 프레임 밖이며 부유의 증거는 없다. 뒤쪽 자루와 보따리는 아래 물건에 받쳐져 있고, 오른쪽 횃불은 화면 아래로 이어지는 자루에 연결되어 있다."
       }
      ],
      "all_candidates_fail": false
     },
     "normalized": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 8,
        "verdict_ko": "복면을 쓴 안규홍의 상체와 어깨·팔로 받친 무거운 자루를 미디엄 숏으로 구현했으나, 눈의 부상과 자물쇠의 파손 상태는 불명확하다."
       },
       {
        "label": "A",
        "score": 4,
        "verdict_ko": "어깨와 손의 하중 지지는 자연스럽지만 자루의 한글 노출이 금지 조건을 어기며, 민소매 복장과 확인되지 않는 복면도 인물 설정에 미달한다."
       }
      ],
      "all_candidates_fail": false,
      "readings": [
       {
        "label": "B",
        "direction": "얼굴과 시선은 화면 오른쪽 아래의 화면 밖 공간을 향하며 렌즈를 응시하지 않는다. 구체적인 시선 목표는 지정되지 않았다. 자루의 넓은 무문자 옆면이 카메라를 향하고, 들어 올린 손은 어깨 위 자루의 천을 움켜쥔다.",
        "built_space": "왼쪽에 열린 널판문 하나와 문틀, 뒤쪽에 목재 기둥과 가로대가 있는 흙벽, 아래에 흙바닥이 보인다. 왼쪽 뒤에는 약 여섯 단의 곡식 자루와 일부 가려진 추가 자루가 있고, 오른쪽에는 횃불 하나와 살림살이 더미가 있다. 인물은 문 안쪽 저장물 앞에 서 있다. 문 걸쇠 부근의 자물쇠 모양 금속물은 보이지만 파손 여부는 판별하기 어렵다. 참조의 재료와 야간 불빛은 대체로 유지된다.",
        "entities": "청년 동아시아계 남성 한 명으로 한국인 설정과 모순되지 않으며, 검은 머리를 뒤로 묶고 검은 천으로 코와 입을 가렸다. 때 묻은 밝은 저고리와 걷어 올린 소매는 인물 참조에 가깝다. 눈 주변의 명확한 상처는 확인되지 않고 입술은 복면에 가려져 있다. 옥반지는 드러나지 않는다. 어깨의 자루는 거친 황갈색 마대이며 참조와 재질은 유사하지만 묶인 입구는 뚜렷하지 않다. 배경에는 저장 자루와 생활용품이 있으며 읽을 수 있는 글자는 보이지 않는다.",
        "hard_violations": [],
        "physics": "자루 밑면이 한쪽 어깨에 눌려 얹혀 있고, 같은 쪽의 굽힌 팔과 천을 쥔 손이 하중과 미끄러짐을 제어한다. 상체가 앞으로 기울어 무게에 대응하며 다른 팔은 몸 앞쪽으로 접혀 있다. 발은 프레임 밖이지만 몸이 떠 있다는 징후는 없다. 배경 자루는 바닥과 아래 자루에, 살림살이는 받침 면에 놓여 있고 횃불에는 지지대가 보인다."
       },
       {
        "label": "A",
        "direction": "남성은 오른쪽을 향한 옆모습으로 화면 오른쪽 아래를 본다. 자루는 한쪽 어깨에서 등 뒤로 길게 걸쳐지고 손은 앞쪽으로 내려온 입구를 잡는다. 카메라를 향한 자루 윗면에 한글 글자가 노출되어, 글씨 없는 옆면을 보여 달라는 지시와 어긋난다.",
        "built_space": "목재 세로 기둥과 가로대가 있는 흙벽 앞에 인물이 서 있다. 왼쪽 아래에는 약 다섯 단의 자루와 그 옆의 가려진 자루 더미, 오른쪽 아래에는 천 보따리와 상자류가 놓여 있다. 오른쪽 가장자리에는 횃불 하나가 크게 잘려 보인다. 문과 자물쇠, 흙바닥은 프레임 밖이라 상태를 확인할 수 없다. 창고 재료는 참조와 유사하지만 가까운 횃불과 큰 자루가 상체와 경쟁한다.",
        "entities": "검은 머리를 묶은 청년 동아시아계 남성 한 명이며, 드러난 팔은 인물 참조보다 훨씬 근육질이다. 밝고 때 묻은 옷이지만 소매가 거의 없는 형태여서 참조의 긴소매 저고리와 다르다. 얼굴 아래쪽은 자루에 가려져 검은 복면이 확인되지 않으며 눈의 상처도 뚜렷하지 않다. 입술과 옥반지는 보이지 않는다. 마대 자루의 거친 직조와 묶인 입구는 소품 참조에 가깝지만, 카메라 쪽 한글 표기는 명시적인 금지 사항과 충돌한다.",
        "hard_violations": [
         "어깨에 멘 자루 윗면에 한글 글자가 노출되어, 읽을 수 있는 글씨를 배제하라는 조건을 위반한다."
        ],
        "physics": "자루는 어깨와 등 위에 직접 얹혀 있고, 굽힌 팔의 손이 앞쪽 자루 입구를 잡아 지지한다. 충전물이 어깨 양쪽으로 처지는 모양과 앞으로 기운 몸통은 무거운 짐을 멘 동작으로 가능하다. 발은 프레임 밖이며 부유의 증거는 없다. 뒤쪽 자루와 보따리는 아래 물건에 받쳐져 있고, 오른쪽 횃불은 화면 아래로 이어지는 자루에 연결되어 있다."
       }
      ]
     },
     "ok": true
    }
   ],
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "slot_winner_match": true,
   "slot_winner": {
    "gemini-pro": "B",
    "gpt-high": "B"
   },
   "route": "cross_slot_agree"
  },
  "dual": {
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "normalized": {
    "A": 0.944,
    "B": 2.0
   },
   "adjusted": {
    "A": 0.694,
    "B": 2.0
   },
   "violations": {
    "A": [
     "[gemini-pro] 프롬프트에 명시된 복면(마스크) 미착용",
     "[gemini-pro] 배경 자루에 읽을 수 있는 문자 포함(글씨 금지 조항 위반)",
     "[gpt-high] 어깨에 멘 자루 윗면에 한글 글자가 노출되어, 읽을 수 있는 글씨를 배제하라는 조건을 위반한다."
    ]
   },
   "per_model_winner": {
    "gemini-pro": "B",
    "gpt-high": "B"
   },
   "agreed": true
  },
  "totals": {
   "B": 2000,
   "A": 694
  },
  "selected": "B",
  "ranking": [
   "B",
   "A"
  ],
  "verdicts": [
   {
    "label": "B",
    "score": 2000,
    "verdict_ko": "프롬프트가 요구한 복면(마스크), 상처 입은 눈, 그리고 배경의 열린 문과 부서진 자물쇠까지 완벽하게 구현하여 디테일과 연속성을 훌륭하게 살려냈습니다."
   },
   {
    "label": "A",
    "score": 694,
    "verdict_ko": "인물이 프롬프트에 명시된 복면을 착용하지 않았으며, 배경의 자루에 금지된 읽을 수 있는 문자가 표기되어 있어 중요한 지시 사항을 위반했습니다.  ★위반: [gemini-pro] 프롬프트에 명시된 복면(마스크) 미착용 / [gemini-pro] 배경 자루에 읽을 수 있는 문자 포함(글씨 금지 조항 위반) / [gpt-high] 어깨에 멘 자루 윗면에 한글 글자가 노출되어, 읽을 수 있는 글씨를 배제하라는 조건을 위반한다."
   }
  ],
  "refs": [
   {
    "label": "PREVIOUS SHOT STILL — a visually related earlier shot of this same place: the location's look, materials, fixed features and lighting mood are LOCKED to this photo; never copy its camera framing. The people visible in that photo are NOT in this shot: never carry their faces, bodies or clothing onto anyone here. Each person in THIS shot is defined solely by the PEOPLE list, their own CHARACTER REFERENCE images and their wardrobe notes. If this photo shows a character who cannot move (dead or unconscious), that character's exact pose, position and orientation are ALSO LOCKED — treat their body as a fixed prop of the set that only the camera moves around.",
    "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/db00517b-165b-4005-a287-f28936b59b95/images/272bd7af-9294-4490-98c6-88a40e9deacc/scene/recipe/S22sh9_sel.png",
    "asset_id": "b995bd83-c146-41f1-881c-1c9863173cb6",
    "role": "prev_still"
   },
   {
    "label": "CHARACTER REFERENCE — 안규홍: the exact person appearing in this shot; match face, hair and build exactly.",
    "path": "<bytes:1028384>",
    "asset_id": "3b5894db-ce55-4f34-b966-5e725c3aa952",
    "role": "character_ref"
   },
   {
    "label": "PROP REFERENCE — 곡식 자루·가마니: the exact object appearing in this shot; match its look, material and wear exactly.",
    "path": "<bytes:1807398>",
    "asset_id": "e30c319e-fdc7-4c01-a301-290ddd45d27a",
    "role": "prop_ref"
   }
  ],
  "critique_skipped": true,
  "shot_run_produced": false,
  "shot_run_uid": "06aa141a-82cf-781e-9b01-06af1726f197",
  "ref_mode": "prev+엔티티",
  "share_plan": {
   "ref_plan": "prev",
   "prev_anchor_tag": "S22sh9"
  }
 },
 "S22sh11::cine": {
  "applied": true,
  "attempted_at": "2026-09-09T03:25:42.460333+00:00",
  "fingerprint": "3c546cbc1016b4cf57226a5e1bb47eed1408f5d16437c8ed2b9f02e381772dca",
  "fingerprint_version": 2,
  "provider": "mai",
  "endpoint": "openrouter/chat-completions",
  "model": "microsoft/mai-image-2.6",
  "pack": "24.202608252115",
  "source_file": "S22sh11_sel.png",
  "source_sha256": "30faad511216e972ff5cb9a4acc7a2c8756577a78c5f3ceb69ce7b95e2008a7f",
  "file": "S22sh11_cine.png",
  "staged_sha256": "3920d4f5ab87c154fde80ca66e4e4fb94b3ee2dd660134f9b5b3c837f41af5a0",
  "latency_ms": 27818
 },
 "S23sh1::signage": {
  "fp": "653a18ec639db422",
  "inscriptions": [],
  "cues": [],
  "dropped": []
 },
 "era_assess::20aa11718aad459e": {
  "subjects": [
   {
    "subject_native": "20세기 초 전라도 전통 초가집과 사립문",
    "search_terms_native": [
     "전라도 전통 초가집 사립문",
     "일제강점기 농촌 초가집 골목",
     "근대 한국 초가 툇마루"
    ],
    "language_lock_native": "검색 시 한국어 외에 다른 언어로 번역하거나 외래어 검색어를 덧붙이지 마십시오.",
    "reason_ko": "AI가 한국 전통 초가와 사립문 구조를 일본식 초가(민가)나 중국 전통 흙집 양식과 혼동하여 잘못 그리기 쉽습니다."
   }
  ],
  "subject_text": "마을 흙길·초가집 골목·사립문 앞·툇마루\n초가집 사이로 굽어지는 농촌 흙길. 낮은 사립문이 이어지고, 길 가까이에 집마다 작은 목조 툇마루가 붙어 있다.",
  "identity": "canonical",
  "scope_id": "L21",
  "scope_role": "location_exterior",
  "scope_sha": "96515568c6e0bcd5"
 },
 "era_fail::2fd5df32bcf92080": {
  "stage": "research",
  "subject": "20세기 초 전라도 전통 초가집과 사립문",
  "terms": [
   "전라도 전통 초가집 사립문",
   "일제강점기 농촌 초가집 골목",
   "근대 한국 초가 툇마루"
  ],
  "status": "no_usable",
  "queries": [
   [
    "전라도 초가집 사립문 일제강점기 사진",
    "초가집 툇마루 옛사진 전라도"
   ]
  ],
  "candidate_urls": [
   "https://t1.daumcdn.net/thumb/R659x0/?fname=http%3A%2F%2Ft1.daumcdn.net%2Fencyclop%2Fm54%2FoiN9X85xUAkzEDd2foYtM4O6ZJD00McJnC4Duaro%3Ft%3D1522806856000",
   "https://i.pinimg.com/736x/18/89/0c/18890c3559044e9aa6ff2ce32fc9a60d.jpg",
   "https://farm3.static.flickr.com/2099/2419972834_db3969cae0_z.jpg?zz=1",
   "https://blog.kakaocdn.net/dna/bEB8yt/btqR6uVHGWN/AAAAAAAAAAAAAAAAAAAAAAQg48WPTB2gjyajm_HTiIbDLj8fZer1y1chvdtI7nED/img.jpg?allow_ip=&allow_referer=&credential=yqXZFxpELC7KVnFOS48ylbz2pIh7yKj8&expires=1777561199&signature=OeEeIZ5rVgCRRfeMwrfTfYeBvYY%3D"
  ],
  "coarse": {
   "eligible": [],
   "chosen_index": 0,
   "reason": "종류·보임·기준을 다 만족하는 후보가 없다 — 이 라운드에선 안 고른다",
   "single_judge": true,
   "rejected_judges": {}
  },
  "verdicts": [
   {
    "index": 1,
    "object_type_match": "yes",
    "visible": true,
    "criteria_match": "unsure",
    "similarity": 95
   },
   {
    "index": 2,
    "object_type_match": "no",
    "visible": true,
    "criteria_match": "unsure",
    "similarity": 40
   },
   {
    "index": 3,
    "object_type_match": "no",
    "visible": true,
    "criteria_match": "unsure",
    "similarity": 40
   },
   {
    "index": 4,
    "object_type_match": "no",
    "visible": true,
    "criteria_match": "unsure",
    "similarity": 60
   }
  ],
  "chosen_reason_ko": "",
  "attempts": 5
 },
 "S23sh1::bgfirst_bg": {
  "input_fingerprint": "3890ae5160367893",
  "prompt": "Create the EMPTY BACKGROUND PLATE for one film shot — NO PEOPLE, no figures, no body parts, no sketch lines, no arrows anywhere.\n\"Empty\" means no people only: KEEP the location's inherent occupants and stock that define the place — animals in an animal shelter, pen or farm, goods in a market, moored boats in a harbour — unless the shot text explicitly removes them.\nThe FIRST attached image is a thin-line storyboard sketch: use ONLY its camera angle, horizon, perspective and the placement/size of buildings and set masses — ignore the sketched people and arrows entirely. The SECOND attached image (LOCATION PHOTOGRAPH) is the real place: take its architecture, materials, signage and fixed features, and RE-PROJECT them into the sketch's camera. If the photograph's camera differs from the sketch's, the sketch's camera wins.\nHUMAN-SCALE CALIBRATION: derive every structure's true size from human-scale elements — a door ≈ 2m, a window ≈ 1–1.5m wide, one storey ≈ 2.5–3m; never inflate a small structure or shrink a large one.\n\nSHOT TEXT this background must serve (Korean): 어두운 초가집 골목, 무거운 자루를 멘 채 은밀한 자세로 걸음을 내딛어 한쪽 발이 허공에 뜬 복면 쓴 안규홍, 이관회, 염재보, 복만의 전경.\n\nLOCATION (lock): A dark outdoor dirt alley between thatched village homes.\n\nTIME OF DAY (lock): night.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: wide shot\n- FRAME LAYOUT: Alley continuation ahead of the group in the upper-center of the frame, background.\n- KEY BACKGROUND ELEMENTS: Thatched-house alley (The group is moving quietly through it) — The alley recedes diagonally from the lower right toward the upper center; used as Defines the shared direction of travel and the rear-oblique tracking route; Carried sacks (Borne by the masked group) — Seen mainly from behind and from the camera-side edges; used as Distinct load shapes reinforce individual silhouettes without obscuring the walkers.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Preserve the explicitly dark nighttime alley with subdued saturation and just enough tonal separation to read the burdened moving figures.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nRender ONE photorealistic empty location photograph, 16:9, that this shot can be staged inside later. No readable writing anywhere: surfaces that would carry writing may be present, but stage any wording out of legibility — an oblique angle, distance, shallow focus. No captions, watermarks or overlay text.",
  "effective_prompt": "Create the EMPTY BACKGROUND PLATE for one film shot — NO PEOPLE, no figures, no body parts, no sketch lines, no arrows anywhere.\n\"Empty\" means no people only: KEEP the location's inherent occupants and stock that define the place — animals in an animal shelter, pen or farm, goods in a market, moored boats in a harbour — unless the shot text explicitly removes them.\nThe FIRST attached image is a thin-line storyboard sketch: use ONLY its camera angle, horizon, perspective and the placement/size of buildings and set masses — ignore the sketched people and arrows entirely. The SECOND attached image (LOCATION PHOTOGRAPH) is the real place: take its architecture, materials, signage and fixed features, and RE-PROJECT them into the sketch's camera. If the photograph's camera differs from the sketch's, the sketch's camera wins.\nHUMAN-SCALE CALIBRATION: derive every structure's true size from human-scale elements — a door ≈ 2m, a window ≈ 1–1.5m wide, one storey ≈ 2.5–3m; never inflate a small structure or shrink a large one.\n\nSHOT TEXT this background must serve (Korean): 어두운 초가집 골목, 무거운 자루를 멘 채 은밀한 자세로 걸음을 내딛어 한쪽 발이 허공에 뜬 복면 쓴 안규홍, 이관회, 염재보, 복만의 전경.\n\nLOCATION (lock): A dark outdoor dirt alley between thatched village homes.\n\nTIME OF DAY (lock): night.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: wide shot\n- FRAME LAYOUT: Alley continuation ahead of the group in the upper-center of the frame, background.\n- KEY BACKGROUND ELEMENTS: Thatched-house alley (The group is moving quietly through it) — The alley recedes diagonally from the lower right toward the upper center; used as Defines the shared direction of travel and the rear-oblique tracking route; Carried sacks (Borne by the masked group) — Seen mainly from behind and from the camera-side edges; used as Distinct load shapes reinforce individual silhouettes without obscuring the walkers.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Preserve the explicitly dark nighttime alley with subdued saturation and just enough tonal separation to read the burdened moving figures.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nRender ONE photorealistic empty location photograph, 16:9, that this shot can be staged inside later. No readable writing anywhere: surfaces that would carry writing may be present, but stage any wording out of legibility — an oblique angle, distance, shallow focus. No captions, watermarks or overlay text.",
  "bg_path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/db00517b-165b-4005-a287-f28936b59b95/images/272bd7af-9294-4490-98c6-88a40e9deacc/scene/recipe/S23sh1__bgfirst_bg.png",
  "asset_id": "ab7d96c5-adea-4469-901a-914aeaa6129e",
  "input_asset_ids": [
   "37dd0a26-f584-4c3d-b385-c565ed3d6c52",
   "aeaa845f-49c4-45b0-8a1f-1c8f50c6daed"
  ]
 },
 "S23sh1": {
  "input_fingerprint": "ef54902fc1d08b94",
  "prompt": "Create ONE FINAL photorealistic live-action film still of the moment below — early 20th-century Korea, Jeolla region; all people are Korean unless explicitly stated otherwise. TIME OF DAY (lock): night.\n\nSHOT TEXT (authoritative, Korean): 어두운 초가집 골목, 무거운 자루를 멘 채 은밀한 자세로 걸음을 내딛어 한쪽 발이 허공에 뜬 복면 쓴 안규홍, 이관회, 염재보, 복만의 전경.\n\nLOCATION (lock): A dark outdoor dirt alley between thatched village homes. The shot takes place here — the attached LOCATION STRUCTURE PHOTOGRAPH is the single authority for this exact place — its fixed structure and permanent site details are LOCKED to it. No separate location photograph exists for this place. Build everything else strictly from the location text above and the shot text; the layout sketch (when attached) governs framing and placement only, and the shot text governs time of day, lighting and action.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: wide shot\n- FRAME LAYOUT: Alley continuation ahead of the group in the upper-center of the frame, background.\n- KEY BACKGROUND ELEMENTS: Thatched-house alley (The group is moving quietly through it) — The alley recedes diagonally from the lower right toward the upper center; used as Defines the shared direction of travel and the rear-oblique tracking route; Carried sacks (Borne by the masked group) — Seen mainly from behind and from the camera-side edges; used as Distinct load shapes reinforce individual silhouettes without obscuring the walkers.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Preserve the explicitly dark nighttime alley with subdued saturation and just enough tonal separation to read the burdened moving figures.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The thatched-house lane is dark. 안규홍: He walks masked with a seed-rice sack shouldered, retaining his facial injuries and concealed jade ring. 이관회: He walks masked with a travel bundle. 염재보: He walks masked carrying a rice sack and retaining his gun. 복만: He walks masked, retaining his matchlock.\n\nTHE HAND THAT IS DOING THIS: the object at the centre of this shot is being held, operated, read, aimed or handed over by 안규홍, 이관회, 염재보, 복만 right now, so 안규홍, 이관회, 염재보, 복만's hand — and as much of the wrist and forearm as the framing reaches — is in the frame, gripping or resting on that object exactly the way its real use requires. Match that hand to 안규홍, 이관회, 염재보, 복만: its size, build, skin, age, grooming, sleeve and anything worn on it belong to that person and to no one else.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 안규홍 (한국인 남성, 청년 나이대, 검은 머리, 단단히 묶은 머리); 이관회 (한국인 남성, 청년 나이대, 자연스러운 검은 머리); 염재보 (한국인 남성, 성인 나이대, 자연스러운 검은 머리); 복만 (한국인 남성, 성인 나이대, 거칠게 자란 수염, 자연스러운 검은 머리) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.",
  "roll_prompts": {
   "A": "Stage the shot. The FIRST attached image (SHOT BACKGROUND) is the finished empty background of this shot — keep it EXACTLY: its camera, perspective, architecture, lighting and every fixed feature stay untouched. The SECOND attached image (LAYOUT SKETCH) tells you ONLY where the people go: each sketched person's position, screen size, pose and the gaze/motion arrows. Ignore the sketch's background lines. The CHARACTER REFERENCE photographs show the real people.\nPlace the real people into the background at exactly the sketched positions, sizes and poses, following the arrow directions. No sketch lines or arrows may remain.\n\nCreate ONE FINAL photorealistic live-action film still of the moment below — early 20th-century Korea, Jeolla region; all people are Korean unless explicitly stated otherwise. TIME OF DAY (lock): night.\n\nSHOT TEXT (authoritative, Korean): 어두운 초가집 골목, 무거운 자루를 멘 채 은밀한 자세로 걸음을 내딛어 한쪽 발이 허공에 뜬 복면 쓴 안규홍, 이관회, 염재보, 복만의 전경.\n\nLOCATION (lock): A dark outdoor dirt alley between thatched village homes. The shot takes place here — the FIRST attached image (SHOT BACKGROUND) is this exact place, already built: its ground, structures, horizon, materials and lighting are the finished truth of this location and must not be redesigned or replaced. No location photograph is attached — read the place from that image alone, and add no scenery, structure, vehicle or fixture that it does not already show. This lock governs the place only; the figures in the shot follow the staging and pose instructions.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: wide shot\n- FRAME LAYOUT: Alley continuation ahead of the group in the upper-center of the frame, background.\n- KEY BACKGROUND ELEMENTS: Thatched-house alley (The group is moving quietly through it) — The alley recedes diagonally from the lower right toward the upper center; used as Defines the shared direction of travel and the rear-oblique tracking route; Carried sacks (Borne by the masked group) — Seen mainly from behind and from the camera-side edges; used as Distinct load shapes reinforce individual silhouettes without obscuring the walkers.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Preserve the explicitly dark nighttime alley with subdued saturation and just enough tonal separation to read the burdened moving figures.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The thatched-house lane is dark. 안규홍: He walks masked with a seed-rice sack shouldered, retaining his facial injuries and concealed jade ring. 이관회: He walks masked with a travel bundle. 염재보: He walks masked carrying a rice sack and retaining his gun. 복만: He walks masked, retaining his matchlock.\n\nTHE HAND THAT IS DOING THIS: the object at the centre of this shot is being held, operated, read, aimed or handed over by 안규홍, 이관회, 염재보, 복만 right now, so 안규홍, 이관회, 염재보, 복만's hand — and as much of the wrist and forearm as the framing reaches — is in the frame, gripping or resting on that object exactly the way its real use requires. Match that hand to 안규홍, 이관회, 염재보, 복만: its size, build, skin, age, grooming, sleeve and anything worn on it belong to that person and to no one else.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 안규홍 (한국인 남성, 청년 나이대, 검은 머리, 단단히 묶은 머리); 이관회 (한국인 남성, 청년 나이대, 자연스러운 검은 머리); 염재보 (한국인 남성, 성인 나이대, 자연스러운 검은 머리); 복만 (한국인 남성, 성인 나이대, 거칠게 자란 수염, 자연스러운 검은 머리) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.",
   "B": "Create ONE FINAL photorealistic live-action film still of the moment below — early 20th-century Korea, Jeolla region; all people are Korean unless explicitly stated otherwise. TIME OF DAY (lock): night.\n\nSHOT TEXT (authoritative, Korean): 어두운 초가집 골목, 무거운 자루를 멘 채 은밀한 자세로 걸음을 내딛어 한쪽 발이 허공에 뜬 복면 쓴 안규홍, 이관회, 염재보, 복만의 전경.\n\nLOCATION (lock): A dark outdoor dirt alley between thatched village homes. The shot takes place here — the attached LOCATION STRUCTURE PHOTOGRAPH is the single authority for this exact place — its fixed structure and permanent site details are LOCKED to it. No separate location photograph exists for this place. Build everything else strictly from the location text above and the shot text; the layout sketch (when attached) governs framing and placement only, and the shot text governs time of day, lighting and action.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: wide shot\n- FRAME LAYOUT: Alley continuation ahead of the group in the upper-center of the frame, background.\n- KEY BACKGROUND ELEMENTS: Thatched-house alley (The group is moving quietly through it) — The alley recedes diagonally from the lower right toward the upper center; used as Defines the shared direction of travel and the rear-oblique tracking route; Carried sacks (Borne by the masked group) — Seen mainly from behind and from the camera-side edges; used as Distinct load shapes reinforce individual silhouettes without obscuring the walkers.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Preserve the explicitly dark nighttime alley with subdued saturation and just enough tonal separation to read the burdened moving figures.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The thatched-house lane is dark. 안규홍: He walks masked with a seed-rice sack shouldered, retaining his facial injuries and concealed jade ring. 이관회: He walks masked with a travel bundle. 염재보: He walks masked carrying a rice sack and retaining his gun. 복만: He walks masked, retaining his matchlock.\n\nTHE HAND THAT IS DOING THIS: the object at the centre of this shot is being held, operated, read, aimed or handed over by 안규홍, 이관회, 염재보, 복만 right now, so 안규홍, 이관회, 염재보, 복만's hand — and as much of the wrist and forearm as the framing reaches — is in the frame, gripping or resting on that object exactly the way its real use requires. Match that hand to 안규홍, 이관회, 염재보, 복만: its size, build, skin, age, grooming, sleeve and anything worn on it belong to that person and to no one else.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 안규홍 (한국인 남성, 청년 나이대, 검은 머리, 단단히 묶은 머리); 이관회 (한국인 남성, 청년 나이대, 자연스러운 검은 머리); 염재보 (한국인 남성, 성인 나이대, 자연스러운 검은 머리); 복만 (한국인 남성, 성인 나이대, 거칠게 자란 수염, 자연스러운 검은 머리) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.\n\nCOMPOSITION VARIATION (this candidate only): keep the place, the people and the moment exactly as specified, and keep from the CAMERA & FRAME contract its subject scale, screen placement and key background elements exactly. For this candidate only, its camera angle and height are deliberately left open: choose a camera position distinctly different from the obvious one, as a film director picking a second setup on the same blocking. Stay true to any explicit framing the SHOT TEXT itself dictates."
  },
  "roll_refs": {
   "A": [
    {
     "label": "SHOT BACKGROUND",
     "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/db00517b-165b-4005-a287-f28936b59b95/images/272bd7af-9294-4490-98c6-88a40e9deacc/scene/recipe/S23sh1__bgfirst_bg.png",
     "asset_id": "ab7d96c5-adea-4469-901a-914aeaa6129e",
     "role": "bgfirst_bg"
    },
    {
     "label": "LAYOUT SKETCH (people placement only)",
     "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/db00517b-165b-4005-a287-f28936b59b95/images/272bd7af-9294-4490-98c6-88a40e9deacc/conti/conti_S23sh1.png",
     "asset_id": "37dd0a26-f584-4c3d-b385-c565ed3d6c52",
     "role": "conti_light"
    },
    {
     "label": "CHARACTER REFERENCE — 안규홍: the exact person appearing in this shot; match face, hair and build exactly.",
     "path": "<bytes:1028384>",
     "asset_id": "3b5894db-ce55-4f34-b966-5e725c3aa952",
     "role": "character_ref"
    },
    {
     "label": "CHARACTER REFERENCE — 이관회: the exact person appearing in this shot; match face, hair and build exactly.",
     "path": "<bytes:933540>",
     "asset_id": "851b450f-ac84-47fe-ac5a-03a56f63fcae",
     "role": "character_ref"
    },
    {
     "label": "CHARACTER REFERENCE — 염재보: the exact person appearing in this shot; match face, hair and build exactly.",
     "path": "<bytes:964351>",
     "asset_id": "ab287289-d0b7-436d-8f70-68452f898f73",
     "role": "character_ref"
    },
    {
     "label": "CHARACTER REFERENCE — 복만: the exact person appearing in this shot; match face, hair and build exactly.",
     "path": "<bytes:1027479>",
     "asset_id": "b7e69570-d15d-4241-b5ea-4cb8d4ece1cf",
     "role": "character_ref"
    },
    {
     "label": "PROP REFERENCE — 곡식 자루·가마니: the exact object appearing in this shot; match its look, material and wear exactly.",
     "path": "<bytes:1807398>",
     "asset_id": "e30c319e-fdc7-4c01-a301-290ddd45d27a",
     "role": "prop_ref"
    }
   ],
   "B": [
    {
     "label": "LOCATION STRUCTURE PHOTOGRAPH — the confirmed photograph of this exact place and its fixed structure: it is the SINGLE authority for the location, the structure's shape, proportions, materials, colors, openings and every permanent site detail. Never copy its camera framing, time of day or lighting — the shot text is the authority for those.",
     "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/db00517b-165b-4005-a287-f28936b59b95/images/background_chain/seed_bg_village_lanes_sel.png",
     "asset_id": "aeaa845f-49c4-45b0-8a1f-1c8f50c6daed",
     "role": "location_seed_bg"
    },
    {
     "label": "CHARACTER REFERENCE — 안규홍: the exact person appearing in this shot; match face, hair and build exactly.",
     "path": "<bytes:1028384>",
     "asset_id": "3b5894db-ce55-4f34-b966-5e725c3aa952",
     "role": "character_ref"
    },
    {
     "label": "CHARACTER REFERENCE — 이관회: the exact person appearing in this shot; match face, hair and build exactly.",
     "path": "<bytes:933540>",
     "asset_id": "851b450f-ac84-47fe-ac5a-03a56f63fcae",
     "role": "character_ref"
    },
    {
     "label": "CHARACTER REFERENCE — 염재보: the exact person appearing in this shot; match face, hair and build exactly.",
     "path": "<bytes:964351>",
     "asset_id": "ab287289-d0b7-436d-8f70-68452f898f73",
     "role": "character_ref"
    },
    {
     "label": "CHARACTER REFERENCE — 복만: the exact person appearing in this shot; match face, hair and build exactly.",
     "path": "<bytes:1027479>",
     "asset_id": "b7e69570-d15d-4241-b5ea-4cb8d4ece1cf",
     "role": "character_ref"
    },
    {
     "label": "PROP REFERENCE — 곡식 자루·가마니: the exact object appearing in this shot; match its look, material and wear exactly.",
     "path": "<bytes:1807398>",
     "asset_id": "e30c319e-fdc7-4c01-a301-290ddd45d27a",
     "role": "prop_ref"
    }
   ]
  },
  "shot_run_spend_attempt_count": 1,
  "judge_flip_skipped": "judge_fn_owns_order",
  "readings": [
   {
    "label": "A",
    "direction": "인물들은 골목길을 따라 카메라에서 멀어지는 방향으로 걷고 있으며, 우측의 복만은 몸을 틀어 뒤쪽을 바라보고 있음.",
    "built_space": "참조 이미지의 초가집, 돌담, 흙길 등 골목길의 구조와 일치함.",
    "entities": "4명의 인물이 등장함. 맨 앞의 안규홍(흰 옷)은 복면과 자루가 없음. 두 번째 이관회는 복면을 쓰고 자루를 메고 있으며 소총을 소지함. 세 번째 염재보는 복면을 쓰고 큰 자루를 멤. 네 번째 복만은 복면을 쓰고 소총을 들고 있음.",
    "hard_violations": [],
    "physics": "모든 인물의 발이 땅에 닿아 있거나 자연스러운 걷는 자세로 지지되고 있으며, 공중에 뜬 객체가 없음."
   },
   {
    "label": "B",
    "direction": "인물들은 골목길을 가로질러 오른쪽에서 왼쪽으로 이동하며 왼쪽 측면을 응시하고 있음.",
    "built_space": "참조 이미지의 초가집, 돌담 구조 및 골목길 배경과 일치함.",
    "entities": "4명의 인물이 등장함. 왼쪽 안규홍은 복면과 자루를 멤. 가운데 이관회는 복면을 쓰고 짐가방을 들고 있음. 오른쪽 앞 복만은 복면, 자루, 소총을 소지함. 오른쪽 뒤 염재보는 복면과 소총을 소지함.",
    "hard_violations": [
     "[gemini-pro] 안규홍(왼쪽 인물)의 두 발이 모두 땅에서 떨어져 아무런 지지 없이 허공에 부양해 있음 (Physics)",
     "[gpt-high] 안규홍의 두 짚신이 모두 흙길에서 떨어져 보이며 몸을 지지하는 발이나 다른 접촉점이 없다. 도약의 발구름 없이 무거운 자루를 멘 보행 자세 그대로 떠 있어, 한 발만 들어 걷는 동작으로 성립하지 않는다."
    ],
    "physics": "가운데와 오른쪽 인물들은 땅에 지지되어 서 있으나, 왼쪽 안규홍은 두 발이 땅과 그림자에서 완전히 분리된 채 공중에 떠 있음."
   }
  ],
  "cross_model_order": {
   "policy": "cross_model_order_v2_slot0_fwd_slot1_rev_parallel2",
   "slots": [
    {
     "model": "gemini-pro",
     "order": "forward",
     "display_to_canonical": {
      "A": "A",
      "B": "B"
     },
     "raw": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 4,
        "verdict_ko": "안규홍의 복면과 자루가 누락되고 인물 간 소품 배치가 뒤섞였으나, 치명적인 물리적 오류가 없어 B보다 우수합니다."
       },
       {
        "label": "B",
        "score": 2,
        "verdict_ko": "소품과 복장 배치는 비교적 프롬프트에 부합하나, 안규홍이 물리적 지지 없이 허공에 떠 있는 치명적인 오류가 발생했습니다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "인물들은 골목길을 따라 카메라에서 멀어지는 방향으로 걷고 있으며, 우측의 복만은 몸을 틀어 뒤쪽을 바라보고 있음.",
        "built_space": "참조 이미지의 초가집, 돌담, 흙길 등 골목길의 구조와 일치함.",
        "entities": "4명의 인물이 등장함. 맨 앞의 안규홍(흰 옷)은 복면과 자루가 없음. 두 번째 이관회는 복면을 쓰고 자루를 메고 있으며 소총을 소지함. 세 번째 염재보는 복면을 쓰고 큰 자루를 멤. 네 번째 복만은 복면을 쓰고 소총을 들고 있음.",
        "hard_violations": [],
        "physics": "모든 인물의 발이 땅에 닿아 있거나 자연스러운 걷는 자세로 지지되고 있으며, 공중에 뜬 객체가 없음."
       },
       {
        "label": "B",
        "direction": "인물들은 골목길을 가로질러 오른쪽에서 왼쪽으로 이동하며 왼쪽 측면을 응시하고 있음.",
        "built_space": "참조 이미지의 초가집, 돌담 구조 및 골목길 배경과 일치함.",
        "entities": "4명의 인물이 등장함. 왼쪽 안규홍은 복면과 자루를 멤. 가운데 이관회는 복면을 쓰고 짐가방을 들고 있음. 오른쪽 앞 복만은 복면, 자루, 소총을 소지함. 오른쪽 뒤 염재보는 복면과 소총을 소지함.",
        "hard_violations": [
         "안규홍(왼쪽 인물)의 두 발이 모두 땅에서 떨어져 아무런 지지 없이 허공에 부양해 있음 (Physics)"
        ],
        "physics": "가운데와 오른쪽 인물들은 땅에 지지되어 서 있으나, 왼쪽 안규홍은 두 발이 땅과 그림자에서 완전히 분리된 채 공중에 떠 있음."
       }
      ],
      "all_candidates_fail": false
     },
     "normalized": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 4,
        "verdict_ko": "안규홍의 복면과 자루가 누락되고 인물 간 소품 배치가 뒤섞였으나, 치명적인 물리적 오류가 없어 B보다 우수합니다."
       },
       {
        "label": "B",
        "score": 2,
        "verdict_ko": "소품과 복장 배치는 비교적 프롬프트에 부합하나, 안규홍이 물리적 지지 없이 허공에 떠 있는 치명적인 오류가 발생했습니다."
       }
      ],
      "all_candidates_fail": false,
      "readings": [
       {
        "label": "A",
        "direction": "인물들은 골목길을 따라 카메라에서 멀어지는 방향으로 걷고 있으며, 우측의 복만은 몸을 틀어 뒤쪽을 바라보고 있음.",
        "built_space": "참조 이미지의 초가집, 돌담, 흙길 등 골목길의 구조와 일치함.",
        "entities": "4명의 인물이 등장함. 맨 앞의 안규홍(흰 옷)은 복면과 자루가 없음. 두 번째 이관회는 복면을 쓰고 자루를 메고 있으며 소총을 소지함. 세 번째 염재보는 복면을 쓰고 큰 자루를 멤. 네 번째 복만은 복면을 쓰고 소총을 들고 있음.",
        "hard_violations": [],
        "physics": "모든 인물의 발이 땅에 닿아 있거나 자연스러운 걷는 자세로 지지되고 있으며, 공중에 뜬 객체가 없음."
       },
       {
        "label": "B",
        "direction": "인물들은 골목길을 가로질러 오른쪽에서 왼쪽으로 이동하며 왼쪽 측면을 응시하고 있음.",
        "built_space": "참조 이미지의 초가집, 돌담 구조 및 골목길 배경과 일치함.",
        "entities": "4명의 인물이 등장함. 왼쪽 안규홍은 복면과 자루를 멤. 가운데 이관회는 복면을 쓰고 짐가방을 들고 있음. 오른쪽 앞 복만은 복면, 자루, 소총을 소지함. 오른쪽 뒤 염재보는 복면과 소총을 소지함.",
        "hard_violations": [
         "안규홍(왼쪽 인물)의 두 발이 모두 땅에서 떨어져 아무런 지지 없이 허공에 부양해 있음 (Physics)"
        ],
        "physics": "가운데와 오른쪽 인물들은 땅에 지지되어 서 있으나, 왼쪽 안규홍은 두 발이 땅과 그림자에서 완전히 분리된 채 공중에 떠 있음."
       }
      ]
     },
     "ok": true
    },
    {
     "model": "gpt-high",
     "order": "reverse",
     "display_to_canonical": {
      "A": "B",
      "B": "A"
     },
     "raw": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 2,
        "verdict_ko": "장소와 야간 분위기는 맞지만, 안규홍이 지지 없이 떠 있고 일행이 골목 안쪽이 아닌 화면 왼쪽 앞을 향해 후방 사선 추적 구도를 어긴다."
       },
       {
        "label": "B",
        "score": 7,
        "verdict_ko": "골목 안쪽으로 향하는 후방 사선 와이드 구도와 지면에 지지된 보행이 우세하지만, 안규홍의 자루 누락과 이관회에게 배치된 총은 불일치한다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "흰옷 남성과 앞쪽 두 남성의 얼굴·몸은 대체로 화면 왼쪽 앞을 향하며, 갈색 저고리 남성은 발앞 땅을 내려다본다. 일행 뒤에 골목의 연장부가 놓여, 상단 중앙의 골목 안쪽으로 함께 이동하는 방향과 반대다. 앞쪽 장총의 총구는 왼쪽 아래 땅을, 뒤쪽 남성의 총열은 오른쪽 위를 향한다. 특정 사람을 겨누지는 않는다.",
        "built_space": "왼쪽에 작은 초가지붕을 얹은 양짝 사립문 한 곳, 그 뒤 좁은 출입구 한 곳, 중앙 뒤에 마루가 있는 초가 한 채, 오른쪽 뒤에 목문 한 곳이 보인다. 오른쪽 돌담의 기와 덮개와 흙길도 참조와 대체로 맞는다. 네 사람은 모두 길 안에 있지만, 앞모습 위주의 배치여서 요구된 짐의 후면과 후방 추적 관계가 성립하지 않는다. 불가능한 반사나 명백한 시설 중복은 없다.",
        "entities": "복면을 쓴 성인 남성 네 명이며 추가 인물이나 읽을 수 있는 글자는 없다. 흰옷과 묶은 검은 머리의 안규홍, 갈색 저고리와 밝은 짧은 바지의 이관회는 참조 의상과 가깝다. 어두운 짧은 옷의 염재보와 긴 겉옷의 복만도 구분되지만 두 사람의 참조 모자는 없다. 얼굴이 가려져 정확한 얼굴 동일성, 상처와 수염은 확인하기 어렵고 숨긴 옥반지는 평가할 수 없다. 안규홍의 어깨 자루, 이관회의 손보따리, 뒤쪽 남성의 자루와 총, 앞쪽 긴 겉옷 남성의 장총이 보인다. 자루는 거친 천 재질이지만 참조의 성긴 황갈색 마대보다 매끈하다.",
        "hard_violations": [
         "안규홍의 두 짚신이 모두 흙길에서 떨어져 보이며 몸을 지지하는 발이나 다른 접촉점이 없다. 도약의 발구름 없이 무거운 자루를 멘 보행 자세 그대로 떠 있어, 한 발만 들어 걷는 동작으로 성립하지 않는다."
        ],
        "physics": "안규홍은 한쪽 무릎을 들었지만 반대쪽 짚신 아래에도 간격이 보여, 몸을 지지하는 것이 없다. 자루 자체는 어깨와 자루 목을 잡은 양손으로 지지된다. 이관회의 보따리는 손이 손잡이를 잡아 매달고 있으며 두 발은 지면에 놓인다. 앞쪽 장총은 손으로 받치고, 뒤쪽 짐과 총은 몸에 밀착된 어깨끈으로 운반된다. 나머지 인물은 지면 지지가 있으나 은밀히 발을 내딛는 순간보다는 정지한 자세에 가깝다."
       },
       {
        "label": "B",
        "direction": "흰옷 남성은 등을 보이며 골목 안쪽으로 향하고, 갈색 옷 남성과 푸른 옷 남성도 같은 진행 방향으로 몸을 돌린다. 오른쪽 긴 겉옷 남성은 머리를 왼쪽의 동료와 진행로 쪽으로 돌리고 있다. 길은 오른쪽 아래에서 중앙 위쪽으로 이어져 후방 사선 이동 구도에 가깝다. 오른쪽 장총의 총구는 오른쪽 위로, 갈색 옷 남성 등에 멘 총의 총열은 왼쪽 위로 향하며 조준 행동은 아니다.",
        "built_space": "왼쪽의 초가지붕 사립문 한 곳, 중앙의 좁은 출입구 한 곳, 뒤쪽 마루가 있는 초가 한 채와 오른쪽 후방 목문 한 곳이 보인다. 양쪽 돌·흙담과 기와 덮개, 흙길의 재료와 배치는 참조 장소에 가깝다. 네 사람은 담 사이 길 위에 있으며 안규홍은 다른 세 명보다 앞서 있다. 와이드 프레임에 전신과 진행할 길이 함께 들어오고, 짐은 주로 뒤와 옆에서 보인다. 명백한 시설 중복이나 불가능한 반사는 없다.",
        "entities": "추가 인물 없이 남성 네 명이 보인다. 흰옷의 안규홍, 갈색 옷의 이관회, 푸른 계열 짧은 옷과 각반의 염재보, 긴 어두운 겉옷의 복만으로 읽힌다. 염재보와 복만의 참조 모자는 빠졌다. 뒤의 세 명은 얼굴을 천으로 가렸고, 안규홍은 완전한 뒷모습이라 복면과 얼굴 상처를 확인할 수 없다. 수염과 숨긴 옥반지도 가려져 평가할 수 없다. 안규홍의 등과 양어깨가 드러나지만 씨나락 자루는 없다. 이관회에게는 큰 천보따리와 멘 총이 있고 염재보에게는 둥근 곡식 자루가 있어, 총의 인물별 배치가 요구와 다르다. 복만은 장총을 들지만 화승 장치의 세부는 판별하기 어렵다. 읽을 수 있는 글자는 없다.",
        "hard_violations": [],
        "physics": "안규홍은 한 발을 뒤로 들고 반대쪽 발로 길을 딛는 보행으로 읽힌다. 이관회도 한쪽 짚신을 들어 밑창을 보이고 다른 발로 지면을 지지한다. 염재보는 지면에 발을 두며, 복만은 한 발을 앞으로 들고 반대쪽 발로 선다. 몸 전체가 지지 없이 뜬 인물은 없다. 이관회의 짐과 총은 어깨에 걸린 끈으로, 염재보의 등짐은 어깨끈과 등 접촉으로 지지된다. 복만의 총은 양손이 받친다. 다만 염재보는 손을 내리고 있어 짐을 손으로 붙잡으라는 조건은 충족하지 않는다."
       }
      ],
      "all_candidates_fail": false
     },
     "normalized": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 2,
        "verdict_ko": "장소와 야간 분위기는 맞지만, 안규홍이 지지 없이 떠 있고 일행이 골목 안쪽이 아닌 화면 왼쪽 앞을 향해 후방 사선 추적 구도를 어긴다."
       },
       {
        "label": "A",
        "score": 7,
        "verdict_ko": "골목 안쪽으로 향하는 후방 사선 와이드 구도와 지면에 지지된 보행이 우세하지만, 안규홍의 자루 누락과 이관회에게 배치된 총은 불일치한다."
       }
      ],
      "all_candidates_fail": false,
      "readings": [
       {
        "label": "B",
        "direction": "흰옷 남성과 앞쪽 두 남성의 얼굴·몸은 대체로 화면 왼쪽 앞을 향하며, 갈색 저고리 남성은 발앞 땅을 내려다본다. 일행 뒤에 골목의 연장부가 놓여, 상단 중앙의 골목 안쪽으로 함께 이동하는 방향과 반대다. 앞쪽 장총의 총구는 왼쪽 아래 땅을, 뒤쪽 남성의 총열은 오른쪽 위를 향한다. 특정 사람을 겨누지는 않는다.",
        "built_space": "왼쪽에 작은 초가지붕을 얹은 양짝 사립문 한 곳, 그 뒤 좁은 출입구 한 곳, 중앙 뒤에 마루가 있는 초가 한 채, 오른쪽 뒤에 목문 한 곳이 보인다. 오른쪽 돌담의 기와 덮개와 흙길도 참조와 대체로 맞는다. 네 사람은 모두 길 안에 있지만, 앞모습 위주의 배치여서 요구된 짐의 후면과 후방 추적 관계가 성립하지 않는다. 불가능한 반사나 명백한 시설 중복은 없다.",
        "entities": "복면을 쓴 성인 남성 네 명이며 추가 인물이나 읽을 수 있는 글자는 없다. 흰옷과 묶은 검은 머리의 안규홍, 갈색 저고리와 밝은 짧은 바지의 이관회는 참조 의상과 가깝다. 어두운 짧은 옷의 염재보와 긴 겉옷의 복만도 구분되지만 두 사람의 참조 모자는 없다. 얼굴이 가려져 정확한 얼굴 동일성, 상처와 수염은 확인하기 어렵고 숨긴 옥반지는 평가할 수 없다. 안규홍의 어깨 자루, 이관회의 손보따리, 뒤쪽 남성의 자루와 총, 앞쪽 긴 겉옷 남성의 장총이 보인다. 자루는 거친 천 재질이지만 참조의 성긴 황갈색 마대보다 매끈하다.",
        "hard_violations": [
         "안규홍의 두 짚신이 모두 흙길에서 떨어져 보이며 몸을 지지하는 발이나 다른 접촉점이 없다. 도약의 발구름 없이 무거운 자루를 멘 보행 자세 그대로 떠 있어, 한 발만 들어 걷는 동작으로 성립하지 않는다."
        ],
        "physics": "안규홍은 한쪽 무릎을 들었지만 반대쪽 짚신 아래에도 간격이 보여, 몸을 지지하는 것이 없다. 자루 자체는 어깨와 자루 목을 잡은 양손으로 지지된다. 이관회의 보따리는 손이 손잡이를 잡아 매달고 있으며 두 발은 지면에 놓인다. 앞쪽 장총은 손으로 받치고, 뒤쪽 짐과 총은 몸에 밀착된 어깨끈으로 운반된다. 나머지 인물은 지면 지지가 있으나 은밀히 발을 내딛는 순간보다는 정지한 자세에 가깝다."
       },
       {
        "label": "A",
        "direction": "흰옷 남성은 등을 보이며 골목 안쪽으로 향하고, 갈색 옷 남성과 푸른 옷 남성도 같은 진행 방향으로 몸을 돌린다. 오른쪽 긴 겉옷 남성은 머리를 왼쪽의 동료와 진행로 쪽으로 돌리고 있다. 길은 오른쪽 아래에서 중앙 위쪽으로 이어져 후방 사선 이동 구도에 가깝다. 오른쪽 장총의 총구는 오른쪽 위로, 갈색 옷 남성 등에 멘 총의 총열은 왼쪽 위로 향하며 조준 행동은 아니다.",
        "built_space": "왼쪽의 초가지붕 사립문 한 곳, 중앙의 좁은 출입구 한 곳, 뒤쪽 마루가 있는 초가 한 채와 오른쪽 후방 목문 한 곳이 보인다. 양쪽 돌·흙담과 기와 덮개, 흙길의 재료와 배치는 참조 장소에 가깝다. 네 사람은 담 사이 길 위에 있으며 안규홍은 다른 세 명보다 앞서 있다. 와이드 프레임에 전신과 진행할 길이 함께 들어오고, 짐은 주로 뒤와 옆에서 보인다. 명백한 시설 중복이나 불가능한 반사는 없다.",
        "entities": "추가 인물 없이 남성 네 명이 보인다. 흰옷의 안규홍, 갈색 옷의 이관회, 푸른 계열 짧은 옷과 각반의 염재보, 긴 어두운 겉옷의 복만으로 읽힌다. 염재보와 복만의 참조 모자는 빠졌다. 뒤의 세 명은 얼굴을 천으로 가렸고, 안규홍은 완전한 뒷모습이라 복면과 얼굴 상처를 확인할 수 없다. 수염과 숨긴 옥반지도 가려져 평가할 수 없다. 안규홍의 등과 양어깨가 드러나지만 씨나락 자루는 없다. 이관회에게는 큰 천보따리와 멘 총이 있고 염재보에게는 둥근 곡식 자루가 있어, 총의 인물별 배치가 요구와 다르다. 복만은 장총을 들지만 화승 장치의 세부는 판별하기 어렵다. 읽을 수 있는 글자는 없다.",
        "hard_violations": [],
        "physics": "안규홍은 한 발을 뒤로 들고 반대쪽 발로 길을 딛는 보행으로 읽힌다. 이관회도 한쪽 짚신을 들어 밑창을 보이고 다른 발로 지면을 지지한다. 염재보는 지면에 발을 두며, 복만은 한 발을 앞으로 들고 반대쪽 발로 선다. 몸 전체가 지지 없이 뜬 인물은 없다. 이관회의 짐과 총은 어깨에 걸린 끈으로, 염재보의 등짐은 어깨끈과 등 접촉으로 지지된다. 복만의 총은 양손이 받친다. 다만 염재보는 손을 내리고 있어 짐을 손으로 붙잡으라는 조건은 충족하지 않는다."
       }
      ]
     },
     "ok": true
    }
   ],
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "slot_winner_match": true,
   "slot_winner": {
    "gemini-pro": "A",
    "gpt-high": "A"
   },
   "route": "cross_slot_agree"
  },
  "dual": {
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "normalized": {
    "A": 2.0,
    "B": 0.786
   },
   "adjusted": {
    "A": 2.0,
    "B": 0.536
   },
   "violations": {
    "B": [
     "[gemini-pro] 안규홍(왼쪽 인물)의 두 발이 모두 땅에서 떨어져 아무런 지지 없이 허공에 부양해 있음 (Physics)",
     "[gpt-high] 안규홍의 두 짚신이 모두 흙길에서 떨어져 보이며 몸을 지지하는 발이나 다른 접촉점이 없다. 도약의 발구름 없이 무거운 자루를 멘 보행 자세 그대로 떠 있어, 한 발만 들어 걷는 동작으로 성립하지 않는다."
    ]
   },
   "per_model_winner": {
    "gemini-pro": "A",
    "gpt-high": "A"
   },
   "agreed": true
  },
  "totals": {
   "A": 2000,
   "B": 536
  },
  "selected": "A",
  "ranking": [
   "A",
   "B"
  ],
  "verdicts": [
   {
    "label": "A",
    "score": 2000,
    "verdict_ko": "안규홍의 복면과 자루가 누락되고 인물 간 소품 배치가 뒤섞였으나, 치명적인 물리적 오류가 없어 B보다 우수합니다."
   },
   {
    "label": "B",
    "score": 536,
    "verdict_ko": "소품과 복장 배치는 비교적 프롬프트에 부합하나, 안규홍이 물리적 지지 없이 허공에 떠 있는 치명적인 오류가 발생했습니다.  ★위반: [gemini-pro] 안규홍(왼쪽 인물)의 두 발이 모두 땅에서 떨어져 아무런 지지 없이 허공에 부양해 있음 (Physics) / [gpt-high] 안규홍의 두 짚신이 모두 흙길에서 떨어져 보이며 몸을 지지하는 발이나 다른 접촉점이 없다. 도약의 발구름 없이 무거운 자루를 멘 보행 자세 그대로 떠 있어, 한 발만 들어 걷는 동작으로 성립하지 않는다."
   }
  ],
  "refs": [
   {
    "label": "LOCATION STRUCTURE PHOTOGRAPH — the confirmed photograph of this exact place and its fixed structure: it is the SINGLE authority for the location, the structure's shape, proportions, materials, colors, openings and every permanent site detail. Never copy its camera framing, time of day or lighting — the shot text is the authority for those.",
    "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/db00517b-165b-4005-a287-f28936b59b95/images/background_chain/seed_bg_village_lanes_sel.png",
    "asset_id": "aeaa845f-49c4-45b0-8a1f-1c8f50c6daed",
    "role": "location_seed_bg"
   },
   {
    "label": "CHARACTER REFERENCE — 안규홍: the exact person appearing in this shot; match face, hair and build exactly.",
    "path": "<bytes:1028384>",
    "asset_id": "3b5894db-ce55-4f34-b966-5e725c3aa952",
    "role": "character_ref"
   },
   {
    "label": "CHARACTER REFERENCE — 이관회: the exact person appearing in this shot; match face, hair and build exactly.",
    "path": "<bytes:933540>",
    "asset_id": "851b450f-ac84-47fe-ac5a-03a56f63fcae",
    "role": "character_ref"
   },
   {
    "label": "CHARACTER REFERENCE — 염재보: the exact person appearing in this shot; match face, hair and build exactly.",
    "path": "<bytes:964351>",
    "asset_id": "ab287289-d0b7-436d-8f70-68452f898f73",
    "role": "character_ref"
   },
   {
    "label": "CHARACTER REFERENCE — 복만: the exact person appearing in this shot; match face, hair and build exactly.",
    "path": "<bytes:1027479>",
    "asset_id": "b7e69570-d15d-4241-b5ea-4cb8d4ece1cf",
    "role": "character_ref"
   },
   {
    "label": "PROP REFERENCE — 곡식 자루·가마니: the exact object appearing in this shot; match its look, material and wear exactly.",
    "path": "<bytes:1807398>",
    "asset_id": "e30c319e-fdc7-4c01-a301-290ddd45d27a",
    "role": "prop_ref"
   }
  ],
  "critique_skipped": true,
  "shot_run_produced": false,
  "shot_run_uid": "06aa141a-8466-7b72-973b-22e8d8d85d3d",
  "bgfirst": {
   "bg_path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/db00517b-165b-4005-a287-f28936b59b95/images/272bd7af-9294-4490-98c6-88a40e9deacc/scene/recipe/S23sh1__bgfirst_bg.png",
   "bg_asset_id": "ab7d96c5-adea-4469-901a-914aeaa6129e",
   "bg_record_key": "S23sh1::bgfirst_bg",
   "chain_winner": true,
   "authority": "seed_bg"
  },
  "ref_mode": "재투영 배경+콘티+엔티티 (2택1: 체인 승)",
  "share_plan": {
   "ref_plan": "background"
  },
  "lane_policy": "ab_select_ready"
 },
 "S23sh1::cine": {
  "applied": true,
  "attempted_at": "2026-09-09T09:03:04.296487+00:00",
  "fingerprint": "371e99e2e7e450c2e232155e645f0bf94a2af17147131a78633844493b352bc1",
  "fingerprint_version": 2,
  "provider": "mai",
  "endpoint": "openrouter/chat-completions",
  "model": "microsoft/mai-image-2.6",
  "pack": "24.202608252115",
  "source_file": "S23sh1_sel.png",
  "source_sha256": "17c9d6d764318e16692bf1a8b80e8657b2a1046a1a30673a86ca3911a3d6cc67",
  "file": "S23sh1_cine.png",
  "staged_sha256": "c992707459c56b884cfd523dbff601701ceb7ec8e49c8aa015c7c7fd8a9dcc4f",
  "latency_ms": 28425
 },
 "S23sh3::signage": {
  "fp": "bc4120586ffac82b",
  "inscriptions": [],
  "cues": [],
  "dropped": []
 },
 "S23sh3": {
  "input_fingerprint": "52b97a353ce68b52",
  "prompt": "Create ONE FINAL photorealistic live-action film still of the moment below — early 20th-century Korea, Jeolla region; all people are Korean unless explicitly stated otherwise. TIME OF DAY (lock): night.\n\nSHOT TEXT (authoritative, Korean): 어깨에서 내린 씨나락 자루의 밑동이 바닥에 닿은 순간, 자루를 쥐고 있는 안규홍의 두 팔 클로즈업.\n\nLOCATION (lock): The ground immediately outside a village home's rustic wicket gate at night. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: close-up\n- KEY BACKGROUND ELEMENTS: Seed sack (Its base has just touched the ground while both hands still hold it) — The lower side and ground-contacting base are visible between the forearms; used as Provides the precise endpoint of the lowering action without overwhelming the human detail; House gate (The delivery is being placed in front of it) — A cropped exterior portion is visible behind the arms; used as Identifies this as a doorstep delivery rather than an anonymous handling detail; Ground beneath the sack (In contact with the sack's base); used as A narrow visible strip makes the moment of touchdown readable.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Maintain the alley's dim nighttime exposure, preserving quiet tonal distinctions between the hands, sack, and surrounding ground.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The clothing, hair and overall look of 안규홍 — who appear both in that photo and in this shot — are LOCKED to that photo. Anyone else visible in that photo is NOT in this shot: never carry their face, body or clothing onto anyone here. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): A seed-rice sack has reached the ground in front of a house's gate. 안규홍: He is masked and bent over the sack being set down, retaining his facial injuries and concealed jade ring.\n\nTHE HAND THAT IS DOING THIS: the object at the centre of this shot is being held, operated, read, aimed or handed over by 안규홍 right now, so 안규홍's hand — and as much of the wrist and forearm as the framing reaches — is in the frame, gripping or resting on that object exactly the way its real use requires. Match that hand to 안규홍: its size, build, skin, age, grooming, sleeve and anything worn on it belong to that person and to no one else.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 안규홍 (한국인 남성, 청년 나이대, 검은 머리, 단단히 묶은 머리) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.",
  "roll_prompts": {
   "A": "Create ONE FINAL photorealistic live-action film still of the moment below — early 20th-century Korea, Jeolla region; all people are Korean unless explicitly stated otherwise. TIME OF DAY (lock): night.\n\nSHOT TEXT (authoritative, Korean): 어깨에서 내린 씨나락 자루의 밑동이 바닥에 닿은 순간, 자루를 쥐고 있는 안규홍의 두 팔 클로즈업.\n\nLOCATION (lock): The ground immediately outside a village home's rustic wicket gate at night. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: close-up\n- KEY BACKGROUND ELEMENTS: Seed sack (Its base has just touched the ground while both hands still hold it) — The lower side and ground-contacting base are visible between the forearms; used as Provides the precise endpoint of the lowering action without overwhelming the human detail; House gate (The delivery is being placed in front of it) — A cropped exterior portion is visible behind the arms; used as Identifies this as a doorstep delivery rather than an anonymous handling detail; Ground beneath the sack (In contact with the sack's base); used as A narrow visible strip makes the moment of touchdown readable.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Maintain the alley's dim nighttime exposure, preserving quiet tonal distinctions between the hands, sack, and surrounding ground.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The clothing, hair and overall look of 안규홍 — who appear both in that photo and in this shot — are LOCKED to that photo. Anyone else visible in that photo is NOT in this shot: never carry their face, body or clothing onto anyone here. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): A seed-rice sack has reached the ground in front of a house's gate. 안규홍: He is masked and bent over the sack being set down, retaining his facial injuries and concealed jade ring.\n\nTHE HAND THAT IS DOING THIS: the object at the centre of this shot is being held, operated, read, aimed or handed over by 안규홍 right now, so 안규홍's hand — and as much of the wrist and forearm as the framing reaches — is in the frame, gripping or resting on that object exactly the way its real use requires. Match that hand to 안규홍: its size, build, skin, age, grooming, sleeve and anything worn on it belong to that person and to no one else.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 안규홍 (한국인 남성, 청년 나이대, 검은 머리, 단단히 묶은 머리) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.",
   "B": "Create ONE FINAL photorealistic live-action film still of the moment below — early 20th-century Korea, Jeolla region; all people are Korean unless explicitly stated otherwise. TIME OF DAY (lock): night.\n\nSHOT TEXT (authoritative, Korean): 어깨에서 내린 씨나락 자루의 밑동이 바닥에 닿은 순간, 자루를 쥐고 있는 안규홍의 두 팔 클로즈업.\n\nLOCATION (lock): The ground immediately outside a village home's rustic wicket gate at night. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: close-up\n- KEY BACKGROUND ELEMENTS: Seed sack (Its base has just touched the ground while both hands still hold it) — The lower side and ground-contacting base are visible between the forearms; used as Provides the precise endpoint of the lowering action without overwhelming the human detail; House gate (The delivery is being placed in front of it) — A cropped exterior portion is visible behind the arms; used as Identifies this as a doorstep delivery rather than an anonymous handling detail; Ground beneath the sack (In contact with the sack's base); used as A narrow visible strip makes the moment of touchdown readable.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Maintain the alley's dim nighttime exposure, preserving quiet tonal distinctions between the hands, sack, and surrounding ground.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The clothing, hair and overall look of 안규홍 — who appear both in that photo and in this shot — are LOCKED to that photo. Anyone else visible in that photo is NOT in this shot: never carry their face, body or clothing onto anyone here. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): A seed-rice sack has reached the ground in front of a house's gate. 안규홍: He is masked and bent over the sack being set down, retaining his facial injuries and concealed jade ring.\n\nTHE HAND THAT IS DOING THIS: the object at the centre of this shot is being held, operated, read, aimed or handed over by 안규홍 right now, so 안규홍's hand — and as much of the wrist and forearm as the framing reaches — is in the frame, gripping or resting on that object exactly the way its real use requires. Match that hand to 안규홍: its size, build, skin, age, grooming, sleeve and anything worn on it belong to that person and to no one else.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 안규홍 (한국인 남성, 청년 나이대, 검은 머리, 단단히 묶은 머리) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.\n\nCOMPOSITION VARIATION (this candidate only): keep the place, the people and the moment exactly as specified, and keep from the CAMERA & FRAME contract its subject scale, screen placement and key background elements exactly. For this candidate only, its camera angle and height are deliberately left open: choose a camera position distinctly different from the obvious one, as a film director picking a second setup on the same blocking. Stay true to any explicit framing the SHOT TEXT itself dictates."
  },
  "shot_run_spend_attempt_count": 1,
  "readings": [
   {
    "label": "A",
    "direction": "인물의 시선은 바닥에 놓인 곡식 자루를 향하고 있으며, 두 팔은 자루의 윗부분을 향해 뻗어 있습니다.",
    "built_space": "배경에 나무로 엮은 싸리문과 흙담이 보이며, 야간의 어두운 흙바닥이 보입니다. 이전 샷의 장소 구조와 일치합니다.",
    "entities": "안규홍(지정된 흰색 한복, 지시된 복면 착용), 곡식 자루. 지시된 인물과 사물의 조건이 대체로 일치합니다.",
    "hard_violations": [
     "[gemini-pro] 물리적으로 불가능한 해부학 (자루를 쥐고 있는 두 손이 하나로 융합되어 형태가 완전히 뭉개짐)"
    ],
    "physics": "자루는 바닥에 닿아 먼지가 일고 있어 착지 순간을 보여줍니다. 인물은 쪼그려 앉아 두 발로 몸을 지탱하고 있으나, 두 손의 구조가 물리적으로 불가능하게 엉켜 있습니다."
   },
   {
    "label": "B",
    "direction": "인물의 몸과 한쪽 팔이 바닥에 있는 곡식 자루를 향해 있습니다.",
    "built_space": "싸리문과 흙담이 있으나, 벽면에 이전 샷에 존재하지 않던 사각 형태의 램프(조명등)가 임의로 추가되었습니다.",
    "entities": "곡식 자루, 파란색 옷을 입은 인물. 안규홍의 지정된 흰색 의상이 아닌, 이전 샷에 등장한 다른 인물의 파란색 옷을 입고 있어 신원 지시를 위반했습니다.",
    "hard_violations": [
     "[gemini-pro] 발명된 사물 (이전 샷에 없는 벽걸이 조명등 임의 추가)",
     "[gemini-pro] 다른 인물의 의상 차용 (안규홍에게 금지된 이전 샷 엑스트라의 파란색 의상 착용)",
     "[gpt-high] 이전 장면의 문 옆에 없던 점등 등롱을 새 고정 시설로 추가하여, 잠긴 장소의 시설과 야간 조명을 변경했다."
    ],
    "physics": "자루는 땅에 놓여 있고 한 손이 자루 윗부분을 짚고 있습니다. 인물은 다리로 체중을 지탱하며 웅크리고 있으나 지시된 두 팔 중 하나만 보입니다."
   }
  ],
  "cross_model_order": {
   "policy": "cross_model_order_v2_slot0_fwd_slot1_rev_parallel2",
   "slots": [
    {
     "model": "gemini-pro",
     "order": "forward",
     "display_to_canonical": {
      "A": "A",
      "B": "B"
     },
     "raw": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 2,
        "verdict_ko": "지정된 안규홍의 흰색 의상과 바닥에 닿아 먼지가 이는 자루를 표현했으나, 두 손이 융합된 심각한 해부학적 오류가 있으며 '두 팔 클로즈업'이라는 프레이밍 지시를 완전히 위반했습니다."
       },
       {
        "label": "B",
        "score": 1,
        "verdict_ko": "배경에 이전 샷에 없던 조명등을 임의로 추가하고 안규홍에게 엄격히 금지된 다른 인물의 파란색 의상을 입혔으며, 두 팔 클로즈업 지시 역시 어겼습니다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "인물의 시선은 바닥에 놓인 곡식 자루를 향하고 있으며, 두 팔은 자루의 윗부분을 향해 뻗어 있습니다.",
        "built_space": "배경에 나무로 엮은 싸리문과 흙담이 보이며, 야간의 어두운 흙바닥이 보입니다. 이전 샷의 장소 구조와 일치합니다.",
        "entities": "안규홍(지정된 흰색 한복, 지시된 복면 착용), 곡식 자루. 지시된 인물과 사물의 조건이 대체로 일치합니다.",
        "hard_violations": [
         "물리적으로 불가능한 해부학 (자루를 쥐고 있는 두 손이 하나로 융합되어 형태가 완전히 뭉개짐)"
        ],
        "physics": "자루는 바닥에 닿아 먼지가 일고 있어 착지 순간을 보여줍니다. 인물은 쪼그려 앉아 두 발로 몸을 지탱하고 있으나, 두 손의 구조가 물리적으로 불가능하게 엉켜 있습니다."
       },
       {
        "label": "B",
        "direction": "인물의 몸과 한쪽 팔이 바닥에 있는 곡식 자루를 향해 있습니다.",
        "built_space": "싸리문과 흙담이 있으나, 벽면에 이전 샷에 존재하지 않던 사각 형태의 램프(조명등)가 임의로 추가되었습니다.",
        "entities": "곡식 자루, 파란색 옷을 입은 인물. 안규홍의 지정된 흰색 의상이 아닌, 이전 샷에 등장한 다른 인물의 파란색 옷을 입고 있어 신원 지시를 위반했습니다.",
        "hard_violations": [
         "발명된 사물 (이전 샷에 없는 벽걸이 조명등 임의 추가)",
         "다른 인물의 의상 차용 (안규홍에게 금지된 이전 샷 엑스트라의 파란색 의상 착용)"
        ],
        "physics": "자루는 땅에 놓여 있고 한 손이 자루 윗부분을 짚고 있습니다. 인물은 다리로 체중을 지탱하며 웅크리고 있으나 지시된 두 팔 중 하나만 보입니다."
       }
      ],
      "all_candidates_fail": true
     },
     "normalized": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 2,
        "verdict_ko": "지정된 안규홍의 흰색 의상과 바닥에 닿아 먼지가 이는 자루를 표현했으나, 두 손이 융합된 심각한 해부학적 오류가 있으며 '두 팔 클로즈업'이라는 프레이밍 지시를 완전히 위반했습니다."
       },
       {
        "label": "B",
        "score": 1,
        "verdict_ko": "배경에 이전 샷에 없던 조명등을 임의로 추가하고 안규홍에게 엄격히 금지된 다른 인물의 파란색 의상을 입혔으며, 두 팔 클로즈업 지시 역시 어겼습니다."
       }
      ],
      "all_candidates_fail": true,
      "readings": [
       {
        "label": "A",
        "direction": "인물의 시선은 바닥에 놓인 곡식 자루를 향하고 있으며, 두 팔은 자루의 윗부분을 향해 뻗어 있습니다.",
        "built_space": "배경에 나무로 엮은 싸리문과 흙담이 보이며, 야간의 어두운 흙바닥이 보입니다. 이전 샷의 장소 구조와 일치합니다.",
        "entities": "안규홍(지정된 흰색 한복, 지시된 복면 착용), 곡식 자루. 지시된 인물과 사물의 조건이 대체로 일치합니다.",
        "hard_violations": [
         "물리적으로 불가능한 해부학 (자루를 쥐고 있는 두 손이 하나로 융합되어 형태가 완전히 뭉개짐)"
        ],
        "physics": "자루는 바닥에 닿아 먼지가 일고 있어 착지 순간을 보여줍니다. 인물은 쪼그려 앉아 두 발로 몸을 지탱하고 있으나, 두 손의 구조가 물리적으로 불가능하게 엉켜 있습니다."
       },
       {
        "label": "B",
        "direction": "인물의 몸과 한쪽 팔이 바닥에 있는 곡식 자루를 향해 있습니다.",
        "built_space": "싸리문과 흙담이 있으나, 벽면에 이전 샷에 존재하지 않던 사각 형태의 램프(조명등)가 임의로 추가되었습니다.",
        "entities": "곡식 자루, 파란색 옷을 입은 인물. 안규홍의 지정된 흰색 의상이 아닌, 이전 샷에 등장한 다른 인물의 파란색 옷을 입고 있어 신원 지시를 위반했습니다.",
        "hard_violations": [
         "발명된 사물 (이전 샷에 없는 벽걸이 조명등 임의 추가)",
         "다른 인물의 의상 차용 (안규홍에게 금지된 이전 샷 엑스트라의 파란색 의상 착용)"
        ],
        "physics": "자루는 땅에 놓여 있고 한 손이 자루 윗부분을 짚고 있습니다. 인물은 다리로 체중을 지탱하며 웅크리고 있으나 지시된 두 팔 중 하나만 보입니다."
       }
      ]
     },
     "ok": true
    },
    {
     "model": "gpt-high",
     "order": "reverse",
     "display_to_canonical": {
      "A": "B",
      "B": "A"
     },
     "raw": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 3,
        "verdict_ko": "자루의 접지는 명확하지만, 이전 장소에 없던 점등 등롱을 추가했고 안규홍과 다른 짙은 남색 소매 및 자루 중심 구도로 인물·장소·두 팔 클로즈업 지시를 벗어났다."
       },
       {
        "label": "B",
        "score": 8,
        "verdict_ko": "두 손으로 자루를 붙든 채 밑동이 땅에 닿는 순간과 안규홍의 밝은 낡은 소매를 잘 재현했으나, 얼굴·다리와 넓은 배경까지 보여 두 팔 클로즈업보다 구도가 넓다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "오른쪽 위에서 내려온 두 팔이 중앙 자루를 향한다. 뒤쪽 손은 자루 윗부분을 잡고 앞쪽 손은 몸통 측면을 누른다. 얼굴은 잘려 시선은 확인할 수 없다. 자루는 수직으로 서서 밑동이 땅을 향한다.",
        "built_space": "왼쪽에 두 짝으로 구성된 싸리문 한 벌, 그 오른쪽에 목재 문설주 한 개와 돌·흙 담장이 보인다. 문 앞 흙길에 자루가 있고 인물은 그 오른쪽에서 숙이고 있다. 문의 재료는 이전 장면과 비슷하지만, 문설주 옆 담장 상부에 이전 장면에는 없던 켜진 등롱 한 개가 추가되어 고정 시설과 조명이 바뀌었다. 문과 지면이 넓게 노출된다.",
        "entities": "한 사람의 두 손과 팔, 몸통·다리 일부 및 곡식 자루 한 개가 보인다. 손은 성인 남성의 손으로 보이나 얼굴이 없어 한국인 청년 안규홍의 신원은 확인할 수 없다. 짙은 남색 소매는 안규홍의 밝은 낡은 옷과 다르다. 자루는 거친 직물의 물성을 갖췄지만 소품 참조의 끈으로 조인 목과 펼쳐진 입구 대신 윗부분이 접힌 형태다. 얼굴 부상·복면은 프레임 밖이고 옥가락지는 노출되지 않는다. 읽을 수 있는 글자는 없다.",
        "hard_violations": [
         "이전 장면의 문 옆에 없던 점등 등롱을 새 고정 시설로 추가하여, 잠긴 장소의 시설과 야간 조명을 변경했다."
        ],
        "physics": "자루 밑동이 흙바닥에 넓게 닿아 무게를 지탱하고 두 손도 자루에 접촉한다. 뒤쪽에 보이는 신발은 지면을 디디고 있어 몸을 숙인 자세에 지지가 있다. 자루 하단이 눌리고 주름지는 모습은 무게 있는 자루를 내려놓는 동작으로 가능하다. 다만 이미 안정적으로 세워진 모습에 가까워 막 닿은 순간의 구분은 약하다."
       },
       {
        "label": "B",
        "direction": "오른쪽 위의 인물이 얼굴과 시선을 아래쪽 자루로 향한다. 두 팔이 아래와 왼쪽으로 뻗어 두 손으로 자루의 조인 목을 붙들고 있다. 자루 밑동은 바로 아래 흙바닥에 닿아 내려놓는 동작의 종착점이 읽힌다.",
        "built_space": "왼쪽 배경에 두 짝 싸리문 한 벌, 중앙 뒤쪽에 목재 문설주 한 개와 돌·흙 담장이 보인다. 자루는 문 바깥 흙길에 놓이며 인물은 오른쪽에서 몸을 숙인다. 참조의 문·담장 재료와 어두운 야간 분위기가 이어지고 추가 등기구나 중복 시설은 없다. 다만 문의 상당 부분과 넓은 흙바닥을 보여 요구한 잘린 배경과 좁은 지면 띠보다 범위가 넓다.",
        "entities": "한국인 청년 남성 설정에 부합하는 한 인물과 자루 한 개가 보인다. 인물은 밝고 때 묻은 전통 의복과 닳은 소매를 입었고, 부분적으로 보이는 얼굴에는 복면과 눈 주변 상처가 있다. 머리는 천에 가려 묶은 검은 머리 형태를 직접 확인하기 어렵다. 두 손의 크기와 체격은 인물 참조에 대체로 부합한다. 자루의 거친 갈색 직물, 묶인 목과 펼쳐진 입구는 소품 참조와 잘 맞는다. 옥가락지는 드러나지 않으며 다른 인물이나 읽을 수 있는 글자는 없다.",
        "hard_violations": [],
        "physics": "자루 밑동이 흙바닥에 눌려 닿아 있고, 접촉 부근의 작은 먼지 구름이 막 내려놓은 순간을 보강한다. 두 손은 자루 목을 겹쳐 쥐고 있어 물체와 연결되어 있다. 오른쪽에 보이는 발과 짚신이 지면을 디디고 굽힌 다리가 숙인 몸을 지탱한다. 공중에 지지 없이 떠 있는 신체나 물체는 없다."
       }
      ],
      "all_candidates_fail": false
     },
     "normalized": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 3,
        "verdict_ko": "자루의 접지는 명확하지만, 이전 장소에 없던 점등 등롱을 추가했고 안규홍과 다른 짙은 남색 소매 및 자루 중심 구도로 인물·장소·두 팔 클로즈업 지시를 벗어났다."
       },
       {
        "label": "A",
        "score": 8,
        "verdict_ko": "두 손으로 자루를 붙든 채 밑동이 땅에 닿는 순간과 안규홍의 밝은 낡은 소매를 잘 재현했으나, 얼굴·다리와 넓은 배경까지 보여 두 팔 클로즈업보다 구도가 넓다."
       }
      ],
      "all_candidates_fail": false,
      "readings": [
       {
        "label": "B",
        "direction": "오른쪽 위에서 내려온 두 팔이 중앙 자루를 향한다. 뒤쪽 손은 자루 윗부분을 잡고 앞쪽 손은 몸통 측면을 누른다. 얼굴은 잘려 시선은 확인할 수 없다. 자루는 수직으로 서서 밑동이 땅을 향한다.",
        "built_space": "왼쪽에 두 짝으로 구성된 싸리문 한 벌, 그 오른쪽에 목재 문설주 한 개와 돌·흙 담장이 보인다. 문 앞 흙길에 자루가 있고 인물은 그 오른쪽에서 숙이고 있다. 문의 재료는 이전 장면과 비슷하지만, 문설주 옆 담장 상부에 이전 장면에는 없던 켜진 등롱 한 개가 추가되어 고정 시설과 조명이 바뀌었다. 문과 지면이 넓게 노출된다.",
        "entities": "한 사람의 두 손과 팔, 몸통·다리 일부 및 곡식 자루 한 개가 보인다. 손은 성인 남성의 손으로 보이나 얼굴이 없어 한국인 청년 안규홍의 신원은 확인할 수 없다. 짙은 남색 소매는 안규홍의 밝은 낡은 옷과 다르다. 자루는 거친 직물의 물성을 갖췄지만 소품 참조의 끈으로 조인 목과 펼쳐진 입구 대신 윗부분이 접힌 형태다. 얼굴 부상·복면은 프레임 밖이고 옥가락지는 노출되지 않는다. 읽을 수 있는 글자는 없다.",
        "hard_violations": [
         "이전 장면의 문 옆에 없던 점등 등롱을 새 고정 시설로 추가하여, 잠긴 장소의 시설과 야간 조명을 변경했다."
        ],
        "physics": "자루 밑동이 흙바닥에 넓게 닿아 무게를 지탱하고 두 손도 자루에 접촉한다. 뒤쪽에 보이는 신발은 지면을 디디고 있어 몸을 숙인 자세에 지지가 있다. 자루 하단이 눌리고 주름지는 모습은 무게 있는 자루를 내려놓는 동작으로 가능하다. 다만 이미 안정적으로 세워진 모습에 가까워 막 닿은 순간의 구분은 약하다."
       },
       {
        "label": "A",
        "direction": "오른쪽 위의 인물이 얼굴과 시선을 아래쪽 자루로 향한다. 두 팔이 아래와 왼쪽으로 뻗어 두 손으로 자루의 조인 목을 붙들고 있다. 자루 밑동은 바로 아래 흙바닥에 닿아 내려놓는 동작의 종착점이 읽힌다.",
        "built_space": "왼쪽 배경에 두 짝 싸리문 한 벌, 중앙 뒤쪽에 목재 문설주 한 개와 돌·흙 담장이 보인다. 자루는 문 바깥 흙길에 놓이며 인물은 오른쪽에서 몸을 숙인다. 참조의 문·담장 재료와 어두운 야간 분위기가 이어지고 추가 등기구나 중복 시설은 없다. 다만 문의 상당 부분과 넓은 흙바닥을 보여 요구한 잘린 배경과 좁은 지면 띠보다 범위가 넓다.",
        "entities": "한국인 청년 남성 설정에 부합하는 한 인물과 자루 한 개가 보인다. 인물은 밝고 때 묻은 전통 의복과 닳은 소매를 입었고, 부분적으로 보이는 얼굴에는 복면과 눈 주변 상처가 있다. 머리는 천에 가려 묶은 검은 머리 형태를 직접 확인하기 어렵다. 두 손의 크기와 체격은 인물 참조에 대체로 부합한다. 자루의 거친 갈색 직물, 묶인 목과 펼쳐진 입구는 소품 참조와 잘 맞는다. 옥가락지는 드러나지 않으며 다른 인물이나 읽을 수 있는 글자는 없다.",
        "hard_violations": [],
        "physics": "자루 밑동이 흙바닥에 눌려 닿아 있고, 접촉 부근의 작은 먼지 구름이 막 내려놓은 순간을 보강한다. 두 손은 자루 목을 겹쳐 쥐고 있어 물체와 연결되어 있다. 오른쪽에 보이는 발과 짚신이 지면을 디디고 굽힌 다리가 숙인 몸을 지탱한다. 공중에 지지 없이 떠 있는 신체나 물체는 없다."
       }
      ]
     },
     "ok": true
    }
   ],
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "slot_winner_match": true,
   "slot_winner": {
    "gemini-pro": "A",
    "gpt-high": "A"
   },
   "route": "cross_slot_agree"
  },
  "dual": {
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "normalized": {
    "A": 2.0,
    "B": 0.875
   },
   "adjusted": {
    "A": 1.75,
    "B": 0.625
   },
   "violations": {
    "A": [
     "[gemini-pro] 물리적으로 불가능한 해부학 (자루를 쥐고 있는 두 손이 하나로 융합되어 형태가 완전히 뭉개짐)"
    ],
    "B": [
     "[gemini-pro] 발명된 사물 (이전 샷에 없는 벽걸이 조명등 임의 추가)",
     "[gemini-pro] 다른 인물의 의상 차용 (안규홍에게 금지된 이전 샷 엑스트라의 파란색 의상 착용)",
     "[gpt-high] 이전 장면의 문 옆에 없던 점등 등롱을 새 고정 시설로 추가하여, 잠긴 장소의 시설과 야간 조명을 변경했다."
    ]
   },
   "per_model_winner": {
    "gemini-pro": "A",
    "gpt-high": "A"
   },
   "agreed": true
  },
  "totals": {
   "A": 1750,
   "B": 625
  },
  "selected": "A",
  "ranking": [
   "A",
   "B"
  ],
  "verdicts": [
   {
    "label": "A",
    "score": 1750,
    "verdict_ko": "지정된 안규홍의 흰색 의상과 바닥에 닿아 먼지가 이는 자루를 표현했으나, 두 손이 융합된 심각한 해부학적 오류가 있으며 '두 팔 클로즈업'이라는 프레이밍 지시를 완전히 위반했습니다.  ★위반: [gemini-pro] 물리적으로 불가능한 해부학 (자루를 쥐고 있는 두 손이 하나로 융합되어 형태가 완전히 뭉개짐)"
   },
   {
    "label": "B",
    "score": 625,
    "verdict_ko": "배경에 이전 샷에 없던 조명등을 임의로 추가하고 안규홍에게 엄격히 금지된 다른 인물의 파란색 의상을 입혔으며, 두 팔 클로즈업 지시 역시 어겼습니다.  ★위반: [gemini-pro] 발명된 사물 (이전 샷에 없는 벽걸이 조명등 임의 추가) / [gemini-pro] 다른 인물의 의상 차용 (안규홍에게 금지된 이전 샷 엑스트라의 파란색 의상 착용) / [gpt-high] 이전 장면의 문 옆에 없던 점등 등롱을 새 고정 시설로 추가하여, 잠긴 장소의 시설과 야간 조명을 변경했다."
   }
  ],
  "refs": [
   {
    "label": "PREVIOUS SHOT STILL — a visually related earlier shot of this same place: the location's look, materials, fixed features and lighting mood are LOCKED to this photo; never copy its camera framing. The clothing, hair and overall look of 안규홍 — who appear both in that photo and in this shot — are LOCKED to that photo. Anyone else visible in that photo is NOT in this shot: never carry their face, body or clothing onto anyone here. If this photo shows a character who cannot move (dead or unconscious), that character's exact pose, position and orientation are ALSO LOCKED — treat their body as a fixed prop of the set that only the camera moves around.",
    "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/db00517b-165b-4005-a287-f28936b59b95/images/272bd7af-9294-4490-98c6-88a40e9deacc/scene/recipe/S23sh1_sel.png",
    "asset_id": "8c98222a-5d3f-4504-b0b0-f644379664ac",
    "role": "prev_still"
   },
   {
    "label": "CHARACTER REFERENCE — 안규홍: the exact person appearing in this shot; match face, hair and build exactly.",
    "path": "<bytes:1028384>",
    "asset_id": "3b5894db-ce55-4f34-b966-5e725c3aa952",
    "role": "character_ref"
   },
   {
    "label": "PROP REFERENCE — 곡식 자루·가마니: the exact object appearing in this shot; match its look, material and wear exactly.",
    "path": "<bytes:1807398>",
    "asset_id": "e30c319e-fdc7-4c01-a301-290ddd45d27a",
    "role": "prop_ref"
   }
  ],
  "critique_skipped": true,
  "shot_run_produced": false,
  "shot_run_uid": "06aa141d-0d32-7884-ba0f-8a42f02cbdc2",
  "ref_mode": "prev+엔티티",
  "share_plan": {
   "ref_plan": "prev",
   "prev_anchor_tag": "S23sh1"
  }
 },
 "S23sh3::cine": {
  "applied": true,
  "attempted_at": "2026-09-09T03:31:58.617557+00:00",
  "fingerprint": "c0fe2a8d86d7d183a3158932757370ef67e60b06880e11523abb020e9cc36650",
  "fingerprint_version": 2,
  "provider": "mai",
  "endpoint": "openrouter/chat-completions",
  "model": "microsoft/mai-image-2.6",
  "pack": "24.202608252115",
  "source_file": "S23sh3_sel.png",
  "source_sha256": "72ce8bed9f8394facf2b7f8dbaad5bdea60d534f7a9fbb43496900ce21af162e",
  "file": "S23sh3_cine.png",
  "staged_sha256": "99578e1298543efe2c671809ac8978c7ecbbd190737aebc3c2aa0c6dbe91ea3f",
  "latency_ms": 46554
 },
 "S23sh6::signage": {
  "fp": "6fafdb5984ee7a4e",
  "inscriptions": [],
  "cues": [],
  "dropped": []
 },
 "S23sh6": {
  "input_fingerprint": "fd1d8623e4dce8ff",
  "prompt": "Create ONE FINAL photorealistic live-action film still of the moment below — early 20th-century Korea, Jeolla region; all people are Korean unless explicitly stated otherwise. TIME OF DAY (lock): night.\n\nSHOT TEXT (authoritative, Korean): 한 발을 앞으로 내디딘 자세의 안규홍의 넓은 등을 원망스럽게 쳐다보는 이관회의 곁눈질 클로즈업.\n\nLOCATION (lock): The dark village alley continuing past the houses toward the settlement's edge. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: close-up\n- FRAME LAYOUT: Continuous alley between the near profile and the back ahead in the middle-center of the frame, midground.\n- KEY BACKGROUND ELEMENTS: Village alley (Both men continue along the same route) — A narrow visible corridor connects 이관회 in the foreground with 안규홍 farther ahead; used as Proves that the close reaction and departing back occupy one continuous space; 이관회's bundle (Still carried as he walks) — Only a partial edge appears behind his shoulder; used as Retains the physical burden at the edge of the reaction frame.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Continue the dark alley's subdued contrast, allowing the near eye to remain readable without introducing a new or conspicuous light effect.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The clothing, hair and overall look of 안규홍, 이관회 — who appear both in that photo and in this shot — are LOCKED to that photo. Anyone else visible in that photo is NOT in this shot: never carry their face, body or clothing onto anyone here. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The seed-rice sack remains outside one gate, and a rice sack now rests before another house's porch. 안규홍: He walks on masked without the delivered sack, retaining his facial injuries and concealed jade ring. 이관회: He remains masked, rubbing his belly and carrying his travel bundle.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 안규홍 (한국인 남성, 청년 나이대, 검은 머리, 단단히 묶은 머리); 이관회 (한국인 남성, 청년 나이대, 자연스러운 검은 머리) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.",
  "roll_prompts": {
   "A": "Create ONE FINAL photorealistic live-action film still of the moment below — early 20th-century Korea, Jeolla region; all people are Korean unless explicitly stated otherwise. TIME OF DAY (lock): night.\n\nSHOT TEXT (authoritative, Korean): 한 발을 앞으로 내디딘 자세의 안규홍의 넓은 등을 원망스럽게 쳐다보는 이관회의 곁눈질 클로즈업.\n\nLOCATION (lock): The dark village alley continuing past the houses toward the settlement's edge. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: close-up\n- FRAME LAYOUT: Continuous alley between the near profile and the back ahead in the middle-center of the frame, midground.\n- KEY BACKGROUND ELEMENTS: Village alley (Both men continue along the same route) — A narrow visible corridor connects 이관회 in the foreground with 안규홍 farther ahead; used as Proves that the close reaction and departing back occupy one continuous space; 이관회's bundle (Still carried as he walks) — Only a partial edge appears behind his shoulder; used as Retains the physical burden at the edge of the reaction frame.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Continue the dark alley's subdued contrast, allowing the near eye to remain readable without introducing a new or conspicuous light effect.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The clothing, hair and overall look of 안규홍, 이관회 — who appear both in that photo and in this shot — are LOCKED to that photo. Anyone else visible in that photo is NOT in this shot: never carry their face, body or clothing onto anyone here. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The seed-rice sack remains outside one gate, and a rice sack now rests before another house's porch. 안규홍: He walks on masked without the delivered sack, retaining his facial injuries and concealed jade ring. 이관회: He remains masked, rubbing his belly and carrying his travel bundle.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 안규홍 (한국인 남성, 청년 나이대, 검은 머리, 단단히 묶은 머리); 이관회 (한국인 남성, 청년 나이대, 자연스러운 검은 머리) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.",
   "B": "Create ONE FINAL photorealistic live-action film still of the moment below — early 20th-century Korea, Jeolla region; all people are Korean unless explicitly stated otherwise. TIME OF DAY (lock): night.\n\nSHOT TEXT (authoritative, Korean): 한 발을 앞으로 내디딘 자세의 안규홍의 넓은 등을 원망스럽게 쳐다보는 이관회의 곁눈질 클로즈업.\n\nLOCATION (lock): The dark village alley continuing past the houses toward the settlement's edge. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: close-up\n- FRAME LAYOUT: Continuous alley between the near profile and the back ahead in the middle-center of the frame, midground.\n- KEY BACKGROUND ELEMENTS: Village alley (Both men continue along the same route) — A narrow visible corridor connects 이관회 in the foreground with 안규홍 farther ahead; used as Proves that the close reaction and departing back occupy one continuous space; 이관회's bundle (Still carried as he walks) — Only a partial edge appears behind his shoulder; used as Retains the physical burden at the edge of the reaction frame.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Continue the dark alley's subdued contrast, allowing the near eye to remain readable without introducing a new or conspicuous light effect.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The clothing, hair and overall look of 안규홍, 이관회 — who appear both in that photo and in this shot — are LOCKED to that photo. Anyone else visible in that photo is NOT in this shot: never carry their face, body or clothing onto anyone here. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The seed-rice sack remains outside one gate, and a rice sack now rests before another house's porch. 안규홍: He walks on masked without the delivered sack, retaining his facial injuries and concealed jade ring. 이관회: He remains masked, rubbing his belly and carrying his travel bundle.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 안규홍 (한국인 남성, 청년 나이대, 검은 머리, 단단히 묶은 머리); 이관회 (한국인 남성, 청년 나이대, 자연스러운 검은 머리) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.\n\nCOMPOSITION VARIATION (this candidate only): keep the place, the people and the moment exactly as specified, and keep from the CAMERA & FRAME contract its subject scale, screen placement and key background elements exactly. For this candidate only, its camera angle and height are deliberately left open: choose a camera position distinctly different from the obvious one, as a film director picking a second setup on the same blocking. Stay true to any explicit framing the SHOT TEXT itself dictates."
  },
  "shot_run_spend_attempt_count": 1,
  "readings": [
   {
    "label": "A",
    "direction": "전경 우측의 이관회는 곁눈질로 원경 중앙에 있는 안규홍의 등을 쳐다보고 있으며, 안규홍은 앞을 향해 걸어가고 있습니다.",
    "built_space": "레퍼런스와 동일한 형태의 섶문과 흙돌담, 초가집이 있는 좁은 골목길이 정확하게 구현되어 있습니다.",
    "entities": "이관회는 마스크를 쓰고 배를 문지르며 등에 짐 보따리를 메고 있습니다. 안규홍은 마스크 착용 여부는 뒷모습이라 보이지 않으나 흰 옷을 입고 앞으로 한 발 내디디며 걷고 있습니다.",
    "hard_violations": [],
    "physics": "두 인물 모두 땅을 딛고 자연스럽게 서 있거나 걷고 있으며, 이관회의 손은 배 위에 안착해 있고 보따리 역시 어깨에 메달려 물리적으로 자연스럽게 지지되고 있습니다."
   },
   {
    "label": "B",
    "direction": "전경 좌측의 이관회가 고개를 돌려 화면 중앙의 안규홍을 쳐다보고 있으며, 안규홍은 정면을 향해 가만히 서 있습니다.",
    "built_space": "골목길의 분위기는 비슷하나 레퍼런스에 있던 좌측 섶문의 디테일이 생략되었고 벽면의 형태가 변형되었습니다.",
    "entities": "이관회는 마스크를 썼으나 배를 문지르는 동작을 하지 않고 있으며, 짐 보따리는 지시사항과 달리 화면의 큰 비중을 차지하고 있습니다. 안규홍은 한 발을 내디딘 자세가 아니라 두 발을 모으고 서 있습니다.",
    "hard_violations": [],
    "physics": "인물들은 땅에 발을 딛고 서 있으나 안규홍의 뻣뻣한 차려 자세는 걷는 중의 자연스러운 동세가 아닙니다."
   }
  ],
  "cross_model_order": {
   "policy": "cross_model_order_v2_slot0_fwd_slot1_rev_parallel2",
   "slots": [
    {
     "model": "gemini-pro",
     "order": "forward",
     "display_to_canonical": {
      "A": "A",
      "B": "B"
     },
     "raw": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 9,
        "verdict_ko": "프롬프트의 요구사항인 이관회의 배를 문지르는 동작, 안규홍의 앞으로 한 발 내디딘 자세, 어깨 뒤로 살짝 보이는 짐 보따리, 지정된 배경을 모두 훌륭하게 구현했습니다."
       },
       {
        "label": "B",
        "score": 4,
        "verdict_ko": "안규홍이 발을 내디딘 자세가 아닌 차려 자세로 서 있으며, 이관회가 배를 문지르는 동작이 누락되었고 짐 보따리가 어깨 뒤로 살짝 보이는 것이 아니라 화면 밖으로 크게 노출되었습니다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "전경 우측의 이관회는 곁눈질로 원경 중앙에 있는 안규홍의 등을 쳐다보고 있으며, 안규홍은 앞을 향해 걸어가고 있습니다.",
        "built_space": "레퍼런스와 동일한 형태의 섶문과 흙돌담, 초가집이 있는 좁은 골목길이 정확하게 구현되어 있습니다.",
        "entities": "이관회는 마스크를 쓰고 배를 문지르며 등에 짐 보따리를 메고 있습니다. 안규홍은 마스크 착용 여부는 뒷모습이라 보이지 않으나 흰 옷을 입고 앞으로 한 발 내디디며 걷고 있습니다.",
        "hard_violations": [],
        "physics": "두 인물 모두 땅을 딛고 자연스럽게 서 있거나 걷고 있으며, 이관회의 손은 배 위에 안착해 있고 보따리 역시 어깨에 메달려 물리적으로 자연스럽게 지지되고 있습니다."
       },
       {
        "label": "B",
        "direction": "전경 좌측의 이관회가 고개를 돌려 화면 중앙의 안규홍을 쳐다보고 있으며, 안규홍은 정면을 향해 가만히 서 있습니다.",
        "built_space": "골목길의 분위기는 비슷하나 레퍼런스에 있던 좌측 섶문의 디테일이 생략되었고 벽면의 형태가 변형되었습니다.",
        "entities": "이관회는 마스크를 썼으나 배를 문지르는 동작을 하지 않고 있으며, 짐 보따리는 지시사항과 달리 화면의 큰 비중을 차지하고 있습니다. 안규홍은 한 발을 내디딘 자세가 아니라 두 발을 모으고 서 있습니다.",
        "hard_violations": [],
        "physics": "인물들은 땅에 발을 딛고 서 있으나 안규홍의 뻣뻣한 차려 자세는 걷는 중의 자연스러운 동세가 아닙니다."
       }
      ],
      "all_candidates_fail": false
     },
     "normalized": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 9,
        "verdict_ko": "프롬프트의 요구사항인 이관회의 배를 문지르는 동작, 안규홍의 앞으로 한 발 내디딘 자세, 어깨 뒤로 살짝 보이는 짐 보따리, 지정된 배경을 모두 훌륭하게 구현했습니다."
       },
       {
        "label": "B",
        "score": 4,
        "verdict_ko": "안규홍이 발을 내디딘 자세가 아닌 차려 자세로 서 있으며, 이관회가 배를 문지르는 동작이 누락되었고 짐 보따리가 어깨 뒤로 살짝 보이는 것이 아니라 화면 밖으로 크게 노출되었습니다."
       }
      ],
      "all_candidates_fail": false,
      "readings": [
       {
        "label": "A",
        "direction": "전경 우측의 이관회는 곁눈질로 원경 중앙에 있는 안규홍의 등을 쳐다보고 있으며, 안규홍은 앞을 향해 걸어가고 있습니다.",
        "built_space": "레퍼런스와 동일한 형태의 섶문과 흙돌담, 초가집이 있는 좁은 골목길이 정확하게 구현되어 있습니다.",
        "entities": "이관회는 마스크를 쓰고 배를 문지르며 등에 짐 보따리를 메고 있습니다. 안규홍은 마스크 착용 여부는 뒷모습이라 보이지 않으나 흰 옷을 입고 앞으로 한 발 내디디며 걷고 있습니다.",
        "hard_violations": [],
        "physics": "두 인물 모두 땅을 딛고 자연스럽게 서 있거나 걷고 있으며, 이관회의 손은 배 위에 안착해 있고 보따리 역시 어깨에 메달려 물리적으로 자연스럽게 지지되고 있습니다."
       },
       {
        "label": "B",
        "direction": "전경 좌측의 이관회가 고개를 돌려 화면 중앙의 안규홍을 쳐다보고 있으며, 안규홍은 정면을 향해 가만히 서 있습니다.",
        "built_space": "골목길의 분위기는 비슷하나 레퍼런스에 있던 좌측 섶문의 디테일이 생략되었고 벽면의 형태가 변형되었습니다.",
        "entities": "이관회는 마스크를 썼으나 배를 문지르는 동작을 하지 않고 있으며, 짐 보따리는 지시사항과 달리 화면의 큰 비중을 차지하고 있습니다. 안규홍은 한 발을 내디딘 자세가 아니라 두 발을 모으고 서 있습니다.",
        "hard_violations": [],
        "physics": "인물들은 땅에 발을 딛고 서 있으나 안규홍의 뻣뻣한 차려 자세는 걷는 중의 자연스러운 동세가 아닙니다."
       }
      ]
     },
     "ok": true
    },
    {
     "model": "gpt-high",
     "order": "reverse",
     "display_to_canonical": {
      "A": "B",
      "B": "A"
     },
     "raw": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 3,
        "verdict_ko": "전경 눈길이 안규홍의 등보다 카메라 쪽으로 읽히며, 클로즈업 대신 상체와 큰 보따리를 보여주고 두 사람의 머리·복면 설정도 어긋납니다."
       },
       {
        "label": "B",
        "score": 5,
        "verdict_ko": "안규홍 쪽 곁눈질과 이어지는 골목, 한 발 내디딘 동작은 더 명확하지만, 허리까지 넓힌 구도는 요구한 반응 클로즈업이 아니며 머리·복면 설정도 다릅니다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "왼쪽 전경 남성은 고개를 오른쪽으로 돌렸지만 보이는 눈동자는 카메라에 가까운 오른쪽을 향해, 중앙 중경 남성의 등을 겨냥한 곁눈질이 명확하지 않습니다. 중앙 남성은 카메라에 등을 보이고 골목 안쪽을 향합니다. 무기나 방향을 판정할 손도구는 보이지 않습니다.",
        "built_space": "양쪽 돌·흙담 사이에 흙길이 이어지고, 상단에는 여러 초가지붕, 뒤쪽 오른편에는 닫힌 목제 대문 한 곳, 중앙에는 어두운 출입구와 낮은 턱이 보입니다. 전경 남성과 중앙 남성 사이에 같은 골목 바닥이 연결됩니다. 참고의 재료와 야간 분위기는 유사하지만 같은 장소의 구체적인 대문 배치는 B보다 덜 분명합니다. 반사면이나 중복된 고정 설비는 보이지 않습니다.",
        "entities": "한국인 청년 남성으로 보이는 두 사람만 있습니다. 이관회에 해당하는 전경 남성은 갈색의 낡은 옷과 회색 복면을 착용했지만, 자연스러운 짧은 검은 머리 대신 상투와 넓은 이마띠가 있습니다. 보따리는 어깨 뒤 가장자리만이 아니라 화면 왼쪽을 크게 차지합니다. 안규홍에 해당하는 중앙 남성은 흰옷을 입고 짐 없이 등을 보이지만, 단단히 묶은 머리가 아닌 짧은 머리이며 귀와 목 주변에 복면 끈이 확인되지 않습니다. 얼굴 부상과 숨긴 옥반지는 구도상 확인할 수 없습니다. 뒤쪽 낮은 턱 위에는 자루 하나가 보이고, 다른 문 밖의 볍씨 자루는 확인되지 않습니다. 읽을 수 있는 글자는 없습니다.",
        "hard_violations": [],
        "physics": "보따리는 전경 남성의 등과 어깨에 밀착되어 있고 어깨 위 끈이 지지합니다. 두 사람 모두 하체가 잘려 있어 발의 접지는 확인할 수 있으나 없다고 단정할 수는 없습니다. 중앙 남성의 보이는 몸통과 팔은 비교적 대칭적이어서 한 발을 내디딘 순간이라는 단서가 약합니다. 배를 문지르는 손은 프레임 밖입니다. 뒤쪽 자루는 턱에 놓여 있으며, 공중에 뜬 물체나 불가능한 관절은 보이지 않습니다."
       },
       {
        "label": "B",
        "direction": "오른쪽 전경 남성의 얼굴과 눈동자가 화면 왼쪽으로 향해 중앙에서 멀어지는 남성 쪽을 곁눈질하는 관계가 A보다 잘 읽힙니다. 중앙 남성은 등을 보인 채 골목 깊숙한 쪽으로 걷습니다. 전경 남성은 눈을 가늘게 뜨고 눈썹을 긴장시켜 불만 섞인 반응을 보입니다. 무기나 방향성 있는 도구는 없습니다.",
        "built_space": "왼쪽에 초가 차양 아래 갈대살 양짝 대문 한 곳과 두 기둥이 있고, 돌·흙담이 중앙의 좁은 통로로 이어집니다. 뒤쪽에는 초가집 출입구와 낮은 마루가 보입니다. 이 대문과 담, 지붕의 배치는 이전 장면과 직접적으로 대응합니다. 전경 남성 오른쪽과 중앙 보행자 사이에 끊기지 않는 흙길이 있지만, 대문 전체와 전경 허리까지 보여주는 구도는 요구한 클로즈업보다 넓습니다. 불가능한 반사나 설비 중복은 없습니다.",
        "entities": "한국인 청년 남성으로 보이는 두 사람만 있습니다. 전경 이관회는 갈색 저고리, 회색 복면, 여행 보따리와 배 위의 손을 갖추었지만 참고의 자연스러운 검은 머리 대신 상투와 이마띠를 착용했습니다. 보따리는 오른쪽 어깨 뒤에 있으나 단순한 가장자리보다 많이 드러납니다. 중앙 안규홍은 흰옷과 짚신을 착용하고 전달한 자루 없이 걷지만, 묶은 머리가 아닌 짧은 머리이며 뒤통수와 목에 복면 착용의 흔적이 보이지 않습니다. 얼굴 부상과 숨긴 옥반지는 뒷모습으로 확인할 수 없습니다. 집 앞의 두 자루는 이 화면에서 확인되지 않습니다. 읽을 수 있는 글자는 없습니다.",
        "hard_violations": [],
        "physics": "중앙 남성은 한쪽 발을 흙길에 딛고 반대 발의 뒤꿈치를 들어 보행 중의 체중 이동을 보여줍니다. 들린 발에는 지지 다리와 자연스러운 다음 착지 경로가 있어 부유하지 않습니다. 전경 남성의 보따리는 어깨에 걸친 끈과 등에 의해 지지되고, 한 손은 실제로 배에 닿아 있습니다. 전경 발은 화면 밖이며 보이는 신체와 물체에 물리적으로 불가능한 지지는 없습니다."
       }
      ],
      "all_candidates_fail": true
     },
     "normalized": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 3,
        "verdict_ko": "전경 눈길이 안규홍의 등보다 카메라 쪽으로 읽히며, 클로즈업 대신 상체와 큰 보따리를 보여주고 두 사람의 머리·복면 설정도 어긋납니다."
       },
       {
        "label": "A",
        "score": 5,
        "verdict_ko": "안규홍 쪽 곁눈질과 이어지는 골목, 한 발 내디딘 동작은 더 명확하지만, 허리까지 넓힌 구도는 요구한 반응 클로즈업이 아니며 머리·복면 설정도 다릅니다."
       }
      ],
      "all_candidates_fail": true,
      "readings": [
       {
        "label": "B",
        "direction": "왼쪽 전경 남성은 고개를 오른쪽으로 돌렸지만 보이는 눈동자는 카메라에 가까운 오른쪽을 향해, 중앙 중경 남성의 등을 겨냥한 곁눈질이 명확하지 않습니다. 중앙 남성은 카메라에 등을 보이고 골목 안쪽을 향합니다. 무기나 방향을 판정할 손도구는 보이지 않습니다.",
        "built_space": "양쪽 돌·흙담 사이에 흙길이 이어지고, 상단에는 여러 초가지붕, 뒤쪽 오른편에는 닫힌 목제 대문 한 곳, 중앙에는 어두운 출입구와 낮은 턱이 보입니다. 전경 남성과 중앙 남성 사이에 같은 골목 바닥이 연결됩니다. 참고의 재료와 야간 분위기는 유사하지만 같은 장소의 구체적인 대문 배치는 B보다 덜 분명합니다. 반사면이나 중복된 고정 설비는 보이지 않습니다.",
        "entities": "한국인 청년 남성으로 보이는 두 사람만 있습니다. 이관회에 해당하는 전경 남성은 갈색의 낡은 옷과 회색 복면을 착용했지만, 자연스러운 짧은 검은 머리 대신 상투와 넓은 이마띠가 있습니다. 보따리는 어깨 뒤 가장자리만이 아니라 화면 왼쪽을 크게 차지합니다. 안규홍에 해당하는 중앙 남성은 흰옷을 입고 짐 없이 등을 보이지만, 단단히 묶은 머리가 아닌 짧은 머리이며 귀와 목 주변에 복면 끈이 확인되지 않습니다. 얼굴 부상과 숨긴 옥반지는 구도상 확인할 수 없습니다. 뒤쪽 낮은 턱 위에는 자루 하나가 보이고, 다른 문 밖의 볍씨 자루는 확인되지 않습니다. 읽을 수 있는 글자는 없습니다.",
        "hard_violations": [],
        "physics": "보따리는 전경 남성의 등과 어깨에 밀착되어 있고 어깨 위 끈이 지지합니다. 두 사람 모두 하체가 잘려 있어 발의 접지는 확인할 수 있으나 없다고 단정할 수는 없습니다. 중앙 남성의 보이는 몸통과 팔은 비교적 대칭적이어서 한 발을 내디딘 순간이라는 단서가 약합니다. 배를 문지르는 손은 프레임 밖입니다. 뒤쪽 자루는 턱에 놓여 있으며, 공중에 뜬 물체나 불가능한 관절은 보이지 않습니다."
       },
       {
        "label": "A",
        "direction": "오른쪽 전경 남성의 얼굴과 눈동자가 화면 왼쪽으로 향해 중앙에서 멀어지는 남성 쪽을 곁눈질하는 관계가 A보다 잘 읽힙니다. 중앙 남성은 등을 보인 채 골목 깊숙한 쪽으로 걷습니다. 전경 남성은 눈을 가늘게 뜨고 눈썹을 긴장시켜 불만 섞인 반응을 보입니다. 무기나 방향성 있는 도구는 없습니다.",
        "built_space": "왼쪽에 초가 차양 아래 갈대살 양짝 대문 한 곳과 두 기둥이 있고, 돌·흙담이 중앙의 좁은 통로로 이어집니다. 뒤쪽에는 초가집 출입구와 낮은 마루가 보입니다. 이 대문과 담, 지붕의 배치는 이전 장면과 직접적으로 대응합니다. 전경 남성 오른쪽과 중앙 보행자 사이에 끊기지 않는 흙길이 있지만, 대문 전체와 전경 허리까지 보여주는 구도는 요구한 클로즈업보다 넓습니다. 불가능한 반사나 설비 중복은 없습니다.",
        "entities": "한국인 청년 남성으로 보이는 두 사람만 있습니다. 전경 이관회는 갈색 저고리, 회색 복면, 여행 보따리와 배 위의 손을 갖추었지만 참고의 자연스러운 검은 머리 대신 상투와 이마띠를 착용했습니다. 보따리는 오른쪽 어깨 뒤에 있으나 단순한 가장자리보다 많이 드러납니다. 중앙 안규홍은 흰옷과 짚신을 착용하고 전달한 자루 없이 걷지만, 묶은 머리가 아닌 짧은 머리이며 뒤통수와 목에 복면 착용의 흔적이 보이지 않습니다. 얼굴 부상과 숨긴 옥반지는 뒷모습으로 확인할 수 없습니다. 집 앞의 두 자루는 이 화면에서 확인되지 않습니다. 읽을 수 있는 글자는 없습니다.",
        "hard_violations": [],
        "physics": "중앙 남성은 한쪽 발을 흙길에 딛고 반대 발의 뒤꿈치를 들어 보행 중의 체중 이동을 보여줍니다. 들린 발에는 지지 다리와 자연스러운 다음 착지 경로가 있어 부유하지 않습니다. 전경 남성의 보따리는 어깨에 걸친 끈과 등에 의해 지지되고, 한 손은 실제로 배에 닿아 있습니다. 전경 발은 화면 밖이며 보이는 신체와 물체에 물리적으로 불가능한 지지는 없습니다."
       }
      ]
     },
     "ok": true
    }
   ],
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "slot_winner_match": true,
   "slot_winner": {
    "gemini-pro": "A",
    "gpt-high": "A"
   },
   "route": "cross_slot_agree"
  },
  "dual": {
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "normalized": {
    "A": 2.0,
    "B": 1.044
   },
   "adjusted": {
    "A": 2.0,
    "B": 1.044
   },
   "violations": {},
   "per_model_winner": {
    "gemini-pro": "A",
    "gpt-high": "A"
   },
   "agreed": true
  },
  "totals": {
   "A": 2000,
   "B": 1044
  },
  "selected": "A",
  "ranking": [
   "A",
   "B"
  ],
  "verdicts": [
   {
    "label": "A",
    "score": 2000,
    "verdict_ko": "프롬프트의 요구사항인 이관회의 배를 문지르는 동작, 안규홍의 앞으로 한 발 내디딘 자세, 어깨 뒤로 살짝 보이는 짐 보따리, 지정된 배경을 모두 훌륭하게 구현했습니다."
   },
   {
    "label": "B",
    "score": 1044,
    "verdict_ko": "안규홍이 발을 내디딘 자세가 아닌 차려 자세로 서 있으며, 이관회가 배를 문지르는 동작이 누락되었고 짐 보따리가 어깨 뒤로 살짝 보이는 것이 아니라 화면 밖으로 크게 노출되었습니다."
   }
  ],
  "refs": [
   {
    "label": "PREVIOUS SHOT STILL — a visually related earlier shot of this same place: the location's look, materials, fixed features and lighting mood are LOCKED to this photo; never copy its camera framing. The clothing, hair and overall look of 안규홍, 이관회 — who appear both in that photo and in this shot — are LOCKED to that photo. Anyone else visible in that photo is NOT in this shot: never carry their face, body or clothing onto anyone here. If this photo shows a character who cannot move (dead or unconscious), that character's exact pose, position and orientation are ALSO LOCKED — treat their body as a fixed prop of the set that only the camera moves around.",
    "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/db00517b-165b-4005-a287-f28936b59b95/images/272bd7af-9294-4490-98c6-88a40e9deacc/scene/recipe/S23sh1_sel.png",
    "asset_id": "8c98222a-5d3f-4504-b0b0-f644379664ac",
    "role": "prev_still"
   },
   {
    "label": "CHARACTER REFERENCE — 안규홍: the exact person appearing in this shot; match face, hair and build exactly.",
    "path": "<bytes:1028384>",
    "asset_id": "3b5894db-ce55-4f34-b966-5e725c3aa952",
    "role": "character_ref"
   },
   {
    "label": "CHARACTER REFERENCE — 이관회: the exact person appearing in this shot; match face, hair and build exactly.",
    "path": "<bytes:933540>",
    "asset_id": "851b450f-ac84-47fe-ac5a-03a56f63fcae",
    "role": "character_ref"
   }
  ],
  "critique_skipped": true,
  "shot_run_produced": false,
  "shot_run_uid": "06aa141d-0ecf-7627-b09f-180189d2fe36",
  "ref_mode": "prev+엔티티",
  "share_plan": {
   "ref_plan": "prev",
   "prev_anchor_tag": "S23sh1"
  }
 },
 "S23sh6::cine": {
  "applied": true,
  "attempted_at": "2026-09-09T09:03:33.118596+00:00",
  "fingerprint": "5a51ccc0d7a68ec4251fa80a398d4d520de704ebf4e4068f8f4941ffc76a121b",
  "fingerprint_version": 2,
  "provider": "mai",
  "endpoint": "openrouter/chat-completions",
  "model": "microsoft/mai-image-2.6",
  "pack": "24.202608252115",
  "source_file": "S23sh6_sel.png",
  "source_sha256": "5e448ff1106f680247fb9907dc41603549fc1e01292885290e2ae9070cbcb5ca",
  "file": "S23sh6_cine.png",
  "staged_sha256": "c0bc8ad4614f50d81a83f766ae91300af05eeb6a25d3afe240f870b39e7aab02",
  "latency_ms": 26748
 },
 "S24sh2::signage": {
  "fp": "8f0da5791527cf2d",
  "inscriptions": [],
  "cues": [],
  "dropped": []
 },
 "era_assess::a95f6f0865d511a4": {
  "subjects": [
   {
    "subject_native": "20세기 초 한국 장터 국밥집 평상",
    "search_terms_native": [
     "일제강점기 장터 국밥집",
     "근대 장터 주막 평상",
     "옛날 시골 장터 주막",
     "장터 가마솥 국밥 평상"
    ],
    "language_lock_native": "한국어로만 검색해야 하며 다른 언어로 번역하거나 검색어를 덧붙이지 마십시오.",
    "reason_ko": "현대 민속촌 세트나 깔끔한 한옥 식당 가구로 왜곡되기 쉬우므로, 20세기 초 시골 장터 특유의 거칠고 투박한 목재 평상과 주막 가마솥 배치를 확인해야 합니다."
   }
  ],
  "subject_text": "복내장터 국밥집 평상 자리\n가마솥에서 흰 김이 오르는 국밥집. 나무 평상과 구석 자리가 놓이고, 사발과 탁주잔이 평상 위에 놓여 있다.",
  "identity": "canonical",
  "scope_id": "L02",
  "scope_role": "location_exterior",
  "scope_sha": "20346591373194af"
 },
 "era_fail::39401758e6a329db": {
  "stage": "research",
  "subject": "20세기 초 한국 장터 국밥집 평상",
  "terms": [
   "일제강점기 장터 국밥집",
   "근대 장터 주막 평상",
   "옛날 시골 장터 주막",
   "장터 가마솥 국밥 평상"
  ],
  "status": "no_usable",
  "queries": [
   [
    "옛날 시골 장터 주막",
    "장터 가마솥 국밥 평상"
   ]
  ],
  "candidate_urls": [
   "https://ak-d.tripcdn.com/images/1i64c2215bgf4ug9t376D_W_670_10000.jpg?proc=source%2Ftrip",
   "https://gyomubu.or.kr/data/editor/1907/de233fd93d72278edd442c216ef43501_1564018877_5024.jpg",
   "https://tong.visitkorea.or.kr/cms/resource/62/2716162_image2_1.JPG",
   "https://contents.history.go.kr/data/img/km/km_003/km_003_068_01.jpg"
  ],
  "coarse": {
   "eligible": [],
   "chosen_index": 0,
   "reason": "종류·보임·기준을 다 만족하는 후보가 없다 — 이 라운드에선 안 고른다",
   "single_judge": true,
   "rejected_judges": {}
  },
  "verdicts": [
   {
    "index": 1,
    "object_type_match": "no",
    "visible": true,
    "criteria_match": "unsure",
    "similarity": 40
   },
   {
    "index": 2,
    "object_type_match": "no",
    "visible": true,
    "criteria_match": "unsure",
    "similarity": 60
   },
   {
    "index": 3,
    "object_type_match": "yes",
    "visible": true,
    "criteria_match": "unsure",
    "similarity": 95
   },
   {
    "index": 4,
    "object_type_match": "no",
    "visible": false,
    "criteria_match": "unsure",
    "similarity": 50
   }
  ],
  "chosen_reason_ko": "",
  "attempts": 5
 },
 "groupbg::soup_shop_seating": {
  "input_fingerprint": "8def0bb7a0d46e71",
  "meta": {
   "model": "gpt-image-2",
   "size": "1536x864",
   "pack": "11.202607220237",
   "contract": "bgfirst_full_v3",
   "group_sig": {
    "key": "soup_shop_seating",
    "tags": [
     "S24sh2",
     "S24sh7"
    ]
   },
   "context_sig": "5832a4d43299fa9d"
  },
  "prompt": "Create ONE empty live-action location background photograph — NO PEOPLE, no figures, no body parts, no silhouettes, no shadows or reflections of people anywhere.\n\"Empty\" means no people only: KEEP the location's inherent occupants and stock that define the place — animals in an animal shelter, pen or farm, goods and produce in a market, moored boats in a harbour — unless the shot text explicitly removes them.\nThe FIRST attached image is a thin-line storyboard sketch of a shot that happens at this location: use it ONLY as spatial evidence — what this place contains, how its ground, structures and landmarks are arranged and proportioned. Ignore the sketched people and arrows entirely, and do NOT copy its line style: render a fully photographic, physically plausible real place that fits THE LOCATION text below.\nHUMAN-SCALE CALIBRATION: derive every structure's true size from human-scale elements — a door ≈ 2m, a window ≈ 1–1.5m wide, one storey ≈ 2.5–3m; never inflate a small structure or shrink a large one.\n\nTHE LOCATION — early 20th-century Korea, Jeolla region; all people are Korean unless explicitly stated otherwise: The raised wooden dining platform beside a corner seat in a busy daytime market soup eatery; enclosure is not specified.\n\nLOCATION DETAIL (descriptive evidence about this place from the production's location records — evidence, not a staging order): use it to understand what kind of place this is and what it permanently contains. Carry over ONLY its enduring physical features — layout, structures, ground and materials, fixed equipment and installations. Do NOT treat any lighting, weather or time-of-day wording, momentary object states, depicted events or subjective impressions in it as fixed facts of the place; the TIME OF DAY lock below is the sole authority for time and lighting.\n복내장터 국밥집 평상 자리: 장터 한쪽에 위치한 식사 구역으로, 장작불이 타오르는 화덕 위 커다란 가마솥에서 허연 김이 솟아오르고, 주변에 낡은 나무 평상들이 놓여 있다. (특징: 허연 김이 솟는 대형 무쇠 가마솥; 장작 화덕; 손때 묻은 낡은 나무 평상; 토기 사발과 탁주잔)\n\nSCENE EVIDENCE (verbatim quotes from the screenplay about this place — treat them as evidence of what the location physically contains and looks like; stage the PLACE those moments happen in, but do NOT depict the momentary actions, people or staged props themselves):\n- 가마솥에서 허연 김이 솟는 북적이는 국밥집.\n- 구석 자리.\n거친 옷차림의 사내, 오주일이 묵묵히 국밥을 뜬다.\n- 옆 평상. 대장장이와 백정, 나무꾼이 탁주잔을 기울인다.\n\nTIME OF DAY (lock): day.\n\nRender ONE photorealistic empty location photograph, 16:9, neutral enough that every shot of this place can be staged from it later. No readable writing anywhere: surfaces that would carry writing may be present, but stage any wording out of legibility — an oblique angle, distance, shallow focus. No captions, watermarks or overlay text.",
  "effective_prompt": "Create ONE empty live-action location background photograph — NO PEOPLE, no figures, no body parts, no silhouettes, no shadows or reflections of people anywhere.\n\"Empty\" means no people only: KEEP the location's inherent occupants and stock that define the place — animals in an animal shelter, pen or farm, goods and produce in a market, moored boats in a harbour — unless the shot text explicitly removes them.\nThe FIRST attached image is a thin-line storyboard sketch of a shot that happens at this location: use it ONLY as spatial evidence — what this place contains, how its ground, structures and landmarks are arranged and proportioned. Ignore the sketched people and arrows entirely, and do NOT copy its line style: render a fully photographic, physically plausible real place that fits THE LOCATION text below.\nHUMAN-SCALE CALIBRATION: derive every structure's true size from human-scale elements — a door ≈ 2m, a window ≈ 1–1.5m wide, one storey ≈ 2.5–3m; never inflate a small structure or shrink a large one.\n\nTHE LOCATION — early 20th-century Korea, Jeolla region; all people are Korean unless explicitly stated otherwise: The raised wooden dining platform beside a corner seat in a busy daytime market soup eatery; enclosure is not specified.\n\nLOCATION DETAIL (descriptive evidence about this place from the production's location records — evidence, not a staging order): use it to understand what kind of place this is and what it permanently contains. Carry over ONLY its enduring physical features — layout, structures, ground and materials, fixed equipment and installations. Do NOT treat any lighting, weather or time-of-day wording, momentary object states, depicted events or subjective impressions in it as fixed facts of the place; the TIME OF DAY lock below is the sole authority for time and lighting.\n복내장터 국밥집 평상 자리: 장터 한쪽에 위치한 식사 구역으로, 장작불이 타오르는 화덕 위 커다란 가마솥에서 허연 김이 솟아오르고, 주변에 낡은 나무 평상들이 놓여 있다. (특징: 허연 김이 솟는 대형 무쇠 가마솥; 장작 화덕; 손때 묻은 낡은 나무 평상; 토기 사발과 탁주잔)\n\nSCENE EVIDENCE (verbatim quotes from the screenplay about this place — treat them as evidence of what the location physically contains and looks like; stage the PLACE those moments happen in, but do NOT depict the momentary actions, people or staged props themselves):\n- 가마솥에서 허연 김이 솟는 북적이는 국밥집.\n- 구석 자리.\n거친 옷차림의 사내, 오주일이 묵묵히 국밥을 뜬다.\n- 옆 평상. 대장장이와 백정, 나무꾼이 탁주잔을 기울인다.\n\nTIME OF DAY (lock): day.\n\nRender ONE photorealistic empty location photograph, 16:9, neutral enough that every shot of this place can be staged from it later. No readable writing anywhere: surfaces that would carry writing may be present, but stage any wording out of legibility — an oblique angle, distance, shallow focus. No captions, watermarks or overlay text.",
  "bg_path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/db00517b-165b-4005-a287-f28936b59b95/images/272bd7af-9294-4490-98c6-88a40e9deacc/scene/recipe/groupbg_soup_shop_seating_cefa56.png",
  "asset_id": "6bf78bdd-ed4e-47f1-bc6b-c786290e47aa",
  "input_asset_ids": [
   "7acb567e-8915-4455-aca7-3659da9414bb"
  ],
  "origin_tag": "S24sh2",
  "place_text": "The raised wooden dining platform beside a corner seat in a busy daytime market soup eatery; enclosure is not specified.",
  "origin_inputs": {
   "place_text": "The raised wooden dining platform beside a corner seat in a busy daytime market soup eatery; enclosure is not specified.",
   "time_of_day_en": "day",
   "conti_asset_id": "7acb567e-8915-4455-aca7-3659da9414bb"
  }
 },
 "S24sh2::bgfirst_bg": {
  "input_fingerprint": "dbe43ce0d4fab72f",
  "prompt": "Create the EMPTY BACKGROUND PLATE for one film shot — NO PEOPLE, no figures, no body parts, no sketch lines, no arrows anywhere.\n\"Empty\" means no people only: KEEP the location's inherent occupants and stock that define the place — animals in an animal shelter, pen or farm, goods in a market, moored boats in a harbour — unless the shot text explicitly removes them.\nThe FIRST attached image is a thin-line storyboard sketch: use ONLY its camera angle, horizon, perspective and the placement/size of buildings and set masses — ignore the sketched people and arrows entirely. The SECOND attached image (LOCATION PHOTOGRAPH) is the real place: take its architecture, materials, signage and fixed features, and RE-PROJECT them into the sketch's camera. If the photograph's camera differs from the sketch's, the sketch's camera wins.\nHUMAN-SCALE CALIBRATION: derive every structure's true size from human-scale elements — a door ≈ 2m, a window ≈ 1–1.5m wide, one storey ≈ 2.5–3m; never inflate a small structure or shrink a large one.\n\nSHOT TEXT this background must serve (Korean): 나무 평상에 앉아 탁주잔을 공중으로 쨍 부딪치는 대장장이와 백정의 즐거운 상체.\n\nLOCATION (lock): The raised wooden dining platform beside a corner seat in a busy daytime market soup eatery; enclosure is not specified.\n\nTIME OF DAY (lock): day.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: medium shot\n- KEY BACKGROUND ELEMENTS: Wooden seating platform (Occupied by the two men within this crop) — The front edge and a portion of the upper seating surface are visible beneath them; used as Shared physical anchor for the oblique seated two-shot; Takju cups (Raised with their rims touching) — Seen from the side with both rims visible between the men; used as Small central action point linked to two distinct arm gestures.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Use natural ambient light appropriate to the daytime eatery, with restrained saturation and gently open contrast that supports the men's brief conviviality.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nRender ONE photorealistic empty location photograph, 16:9, that this shot can be staged inside later. No readable writing anywhere: surfaces that would carry writing may be present, but stage any wording out of legibility — an oblique angle, distance, shallow focus. No captions, watermarks or overlay text.",
  "effective_prompt": "Create the EMPTY BACKGROUND PLATE for one film shot — NO PEOPLE, no figures, no body parts, no sketch lines, no arrows anywhere.\n\"Empty\" means no people only: KEEP the location's inherent occupants and stock that define the place — animals in an animal shelter, pen or farm, goods in a market, moored boats in a harbour — unless the shot text explicitly removes them.\nThe FIRST attached image is a thin-line storyboard sketch: use ONLY its camera angle, horizon, perspective and the placement/size of buildings and set masses — ignore the sketched people and arrows entirely. The SECOND attached image (LOCATION PHOTOGRAPH) is the real place: take its architecture, materials, signage and fixed features, and RE-PROJECT them into the sketch's camera. If the photograph's camera differs from the sketch's, the sketch's camera wins.\nHUMAN-SCALE CALIBRATION: derive every structure's true size from human-scale elements — a door ≈ 2m, a window ≈ 1–1.5m wide, one storey ≈ 2.5–3m; never inflate a small structure or shrink a large one.\n\nSHOT TEXT this background must serve (Korean): 나무 평상에 앉아 탁주잔을 공중으로 쨍 부딪치는 대장장이와 백정의 즐거운 상체.\n\nLOCATION (lock): The raised wooden dining platform beside a corner seat in a busy daytime market soup eatery; enclosure is not specified.\n\nTIME OF DAY (lock): day.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: medium shot\n- KEY BACKGROUND ELEMENTS: Wooden seating platform (Occupied by the two men within this crop) — The front edge and a portion of the upper seating surface are visible beneath them; used as Shared physical anchor for the oblique seated two-shot; Takju cups (Raised with their rims touching) — Seen from the side with both rims visible between the men; used as Small central action point linked to two distinct arm gestures.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Use natural ambient light appropriate to the daytime eatery, with restrained saturation and gently open contrast that supports the men's brief conviviality.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nRender ONE photorealistic empty location photograph, 16:9, that this shot can be staged inside later. No readable writing anywhere: surfaces that would carry writing may be present, but stage any wording out of legibility — an oblique angle, distance, shallow focus. No captions, watermarks or overlay text.",
  "bg_path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/db00517b-165b-4005-a287-f28936b59b95/images/272bd7af-9294-4490-98c6-88a40e9deacc/scene/recipe/S24sh2__bgfirst_bg.png",
  "asset_id": "7d1839dc-67c7-4bef-845d-acc3564644ee",
  "input_asset_ids": [
   "7acb567e-8915-4455-aca7-3659da9414bb",
   "6bf78bdd-ed4e-47f1-bc6b-c786290e47aa"
  ]
 },
 "S24sh2": {
  "input_fingerprint": "f4ee62ec8936cd1a",
  "prompt": "Create ONE FINAL photorealistic live-action film still of the moment below — early 20th-century Korea, Jeolla region; all people are Korean unless explicitly stated otherwise. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 나무 평상에 앉아 탁주잔을 공중으로 쨍 부딪치는 대장장이와 백정의 즐거운 상체.\n\nLOCATION (lock): The raised wooden dining platform beside a corner seat in a busy daytime market soup eatery; enclosure is not specified. The shot takes place here — the attached LOCATION PHOTOGRAPH shows the exact spot.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: medium shot\n- KEY BACKGROUND ELEMENTS: Wooden seating platform (Occupied by the two men within this crop) — The front edge and a portion of the upper seating surface are visible beneath them; used as Shared physical anchor for the oblique seated two-shot; Takju cups (Raised with their rims touching) — Seen from the side with both rims visible between the men; used as Small central action point linked to two distinct arm gestures.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Use natural ambient light appropriate to the daytime eatery, with restrained saturation and gently open contrast that supports the men's brief conviviality.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): White steam rises from the soup cauldron in the crowded daytime restaurant; rice-soup bowls and rice-wine cups are set out. 대장장이: He sits on the side platform holding a rice-wine cup. 백정: He sits on the side platform holding a rice-wine cup.\n\nTHE HAND THAT IS DOING THIS: the object at the centre of this shot is being held, operated, read, aimed or handed over by 대장장이와 백정 right now, so 대장장이와 백정's hand — and as much of the wrist and forearm as the framing reaches — is in the frame, gripping or resting on that object exactly the way its real use requires. Match that hand to 대장장이와 백정: its size, build, skin, age, grooming, sleeve and anything worn on it belong to that person and to no one else.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 대장장이 (한국인 남성, 성인 나이대, 자연스러운 검은 머리); 백정 (한국인 남성, 성인 나이대, 자연스러운 검은 머리) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.",
  "roll_prompts": {
   "A": "Stage the shot. The FIRST attached image (SHOT BACKGROUND) is the finished empty background of this shot — keep it EXACTLY: its camera, perspective, architecture, lighting and every fixed feature stay untouched. The SECOND attached image (LAYOUT SKETCH) tells you ONLY where the people go: each sketched person's position, screen size, pose and the gaze/motion arrows. Ignore the sketch's background lines. The CHARACTER REFERENCE photographs show the real people.\nPlace the real people into the background at exactly the sketched positions, sizes and poses, following the arrow directions. No sketch lines or arrows may remain.\n\nCreate ONE FINAL photorealistic live-action film still of the moment below — early 20th-century Korea, Jeolla region; all people are Korean unless explicitly stated otherwise. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 나무 평상에 앉아 탁주잔을 공중으로 쨍 부딪치는 대장장이와 백정의 즐거운 상체.\n\nLOCATION (lock): The raised wooden dining platform beside a corner seat in a busy daytime market soup eatery; enclosure is not specified. The shot takes place here — the FIRST attached image (SHOT BACKGROUND) is this exact place, already built: its ground, structures, horizon, materials and lighting are the finished truth of this location and must not be redesigned or replaced. No location photograph is attached — read the place from that image alone, and add no scenery, structure, vehicle or fixture that it does not already show. This lock governs the place only; the figures in the shot follow the staging and pose instructions.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: medium shot\n- KEY BACKGROUND ELEMENTS: Wooden seating platform (Occupied by the two men within this crop) — The front edge and a portion of the upper seating surface are visible beneath them; used as Shared physical anchor for the oblique seated two-shot; Takju cups (Raised with their rims touching) — Seen from the side with both rims visible between the men; used as Small central action point linked to two distinct arm gestures.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Use natural ambient light appropriate to the daytime eatery, with restrained saturation and gently open contrast that supports the men's brief conviviality.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): White steam rises from the soup cauldron in the crowded daytime restaurant; rice-soup bowls and rice-wine cups are set out. 대장장이: He sits on the side platform holding a rice-wine cup. 백정: He sits on the side platform holding a rice-wine cup.\n\nTHE HAND THAT IS DOING THIS: the object at the centre of this shot is being held, operated, read, aimed or handed over by 대장장이와 백정 right now, so 대장장이와 백정's hand — and as much of the wrist and forearm as the framing reaches — is in the frame, gripping or resting on that object exactly the way its real use requires. Match that hand to 대장장이와 백정: its size, build, skin, age, grooming, sleeve and anything worn on it belong to that person and to no one else.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 대장장이 (한국인 남성, 성인 나이대, 자연스러운 검은 머리); 백정 (한국인 남성, 성인 나이대, 자연스러운 검은 머리) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.",
   "B": "Create ONE FINAL photorealistic live-action film still of the moment below — early 20th-century Korea, Jeolla region; all people are Korean unless explicitly stated otherwise. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 나무 평상에 앉아 탁주잔을 공중으로 쨍 부딪치는 대장장이와 백정의 즐거운 상체.\n\nLOCATION (lock): The raised wooden dining platform beside a corner seat in a busy daytime market soup eatery; enclosure is not specified. The shot takes place here — the attached LOCATION PHOTOGRAPH shows the exact spot.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: medium shot\n- KEY BACKGROUND ELEMENTS: Wooden seating platform (Occupied by the two men within this crop) — The front edge and a portion of the upper seating surface are visible beneath them; used as Shared physical anchor for the oblique seated two-shot; Takju cups (Raised with their rims touching) — Seen from the side with both rims visible between the men; used as Small central action point linked to two distinct arm gestures.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Use natural ambient light appropriate to the daytime eatery, with restrained saturation and gently open contrast that supports the men's brief conviviality.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): White steam rises from the soup cauldron in the crowded daytime restaurant; rice-soup bowls and rice-wine cups are set out. 대장장이: He sits on the side platform holding a rice-wine cup. 백정: He sits on the side platform holding a rice-wine cup.\n\nTHE HAND THAT IS DOING THIS: the object at the centre of this shot is being held, operated, read, aimed or handed over by 대장장이와 백정 right now, so 대장장이와 백정's hand — and as much of the wrist and forearm as the framing reaches — is in the frame, gripping or resting on that object exactly the way its real use requires. Match that hand to 대장장이와 백정: its size, build, skin, age, grooming, sleeve and anything worn on it belong to that person and to no one else.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 대장장이 (한국인 남성, 성인 나이대, 자연스러운 검은 머리); 백정 (한국인 남성, 성인 나이대, 자연스러운 검은 머리) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.\n\nCOMPOSITION VARIATION (this candidate only): keep the place, the people and the moment exactly as specified, and keep from the CAMERA & FRAME contract its subject scale, screen placement and key background elements exactly. For this candidate only, its camera angle and height are deliberately left open: choose a camera position distinctly different from the obvious one, as a film director picking a second setup on the same blocking. Stay true to any explicit framing the SHOT TEXT itself dictates."
  },
  "roll_refs": {
   "A": [
    {
     "label": "SHOT BACKGROUND",
     "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/db00517b-165b-4005-a287-f28936b59b95/images/272bd7af-9294-4490-98c6-88a40e9deacc/scene/recipe/S24sh2__bgfirst_bg.png",
     "asset_id": "7d1839dc-67c7-4bef-845d-acc3564644ee",
     "role": "bgfirst_bg"
    },
    {
     "label": "LAYOUT SKETCH (people placement only)",
     "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/db00517b-165b-4005-a287-f28936b59b95/images/272bd7af-9294-4490-98c6-88a40e9deacc/conti/conti_S24sh2.png",
     "asset_id": "7acb567e-8915-4455-aca7-3659da9414bb",
     "role": "conti_light"
    },
    {
     "label": "CHARACTER REFERENCE — 대장장이: the exact person appearing in this shot; match face, hair and build exactly.",
     "path": "<bytes:946438>",
     "asset_id": "a12f8b90-e799-4753-9b7e-e71f4d817db5",
     "role": "character_ref"
    },
    {
     "label": "CHARACTER REFERENCE — 백정: the exact person appearing in this shot; match face, hair and build exactly.",
     "path": "<bytes:980360>",
     "asset_id": "29282584-ae7c-4722-8d77-100c4a5b2a3d",
     "role": "character_ref"
    }
   ],
   "B": [
    {
     "label": "LOCATION PHOTOGRAPH — the exact place of this shot: its architecture, materials, fixed features and lighting mood are spatial truth; stage the moment inside this place. Never copy its camera framing.",
     "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/db00517b-165b-4005-a287-f28936b59b95/images/272bd7af-9294-4490-98c6-88a40e9deacc/scene/recipe/groupbg_soup_shop_seating_cefa56.png",
     "asset_id": "6bf78bdd-ed4e-47f1-bc6b-c786290e47aa",
     "role": "bgfirst_group_bg"
    },
    {
     "label": "CHARACTER REFERENCE — 대장장이: the exact person appearing in this shot; match face, hair and build exactly.",
     "path": "<bytes:946438>",
     "asset_id": "a12f8b90-e799-4753-9b7e-e71f4d817db5",
     "role": "character_ref"
    },
    {
     "label": "CHARACTER REFERENCE — 백정: the exact person appearing in this shot; match face, hair and build exactly.",
     "path": "<bytes:980360>",
     "asset_id": "29282584-ae7c-4722-8d77-100c4a5b2a3d",
     "role": "character_ref"
    }
   ]
  },
  "shot_run_spend_attempt_count": 1,
  "judge_flip_skipped": "judge_fn_owns_order",
  "readings": [
   {
    "label": "A",
    "direction": "두 인물이 서로 마주보며 공중에서 탁주잔을 정확히 맞부딪치고 있음.",
    "built_space": "나무 평상에 앉아 있으나, 레퍼런스의 중앙 화덕 구조가 무시되고 화덕이 인물들 뒤편으로 밀려나 공간이 변형됨.",
    "entities": "대장장이와 백정의 얼굴, 헤어, 복장이 레퍼런스와 완벽히 일치함. 탁주잔이 묘사됨. 추가 인물 없음.",
    "hard_violations": [],
    "physics": "두 인물 모두 평상 위에 안정적으로 앉아 체중이 지지되며, 손으로 잔을 자연스럽게 쥐고 있음."
   },
   {
    "label": "B",
    "direction": "서로 마주보며 시선을 교환하고 탁주잔을 맞부딪치고 있음.",
    "built_space": "레퍼런스의 'ㄷ'자형 평상과 그 사이의 중앙 돌 화덕 구조를 화면 중앙에 매우 정확하게 구현함.",
    "entities": "주요 인물 둘은 레퍼런스와 일치하나, 배경 측면에 프롬프트(PEOPLE 규칙)가 허용하지 않은 다수의 추가 인물들이 존재함.",
    "hard_violations": [
     "[gemini-pro] 지문에 없는 배경 인물 임의 추가 (invented people)",
     "[gpt-high] 왼쪽 배경 식탁에 허용되지 않은 인물 세 명이 추가되어, 대장장이와 백정만 보여야 한다는 인물 제한을 위반한다."
    ],
    "physics": "평상에 정상적으로 앉아 잔을 쥐고 있는 자세가 물리적으로 자연스러움."
   }
  ],
  "cross_model_order": {
   "policy": "cross_model_order_v2_slot0_fwd_slot1_rev_parallel2",
   "slots": [
    {
     "model": "gemini-pro",
     "order": "forward",
     "display_to_canonical": {
      "A": "A",
      "B": "B"
     },
     "raw": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 7,
        "verdict_ko": "미디엄 샷 및 상체 위주의 프레이밍 지시를 어기고 롱샷에 가깝게 연출되었으나, 치명적인 하드 위반 사항이 없어 최종 선택됨."
       },
       {
        "label": "B",
        "score": 3,
        "verdict_ko": "공간 구조와 프레이밍은 더 정확하게 구현했으나, 지문에 명시되지 않은 배경 인물들을 다수 추가하여 하드 위반으로 탈락함."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "두 인물이 서로 마주보며 공중에서 탁주잔을 정확히 맞부딪치고 있음.",
        "built_space": "나무 평상에 앉아 있으나, 레퍼런스의 중앙 화덕 구조가 무시되고 화덕이 인물들 뒤편으로 밀려나 공간이 변형됨.",
        "entities": "대장장이와 백정의 얼굴, 헤어, 복장이 레퍼런스와 완벽히 일치함. 탁주잔이 묘사됨. 추가 인물 없음.",
        "hard_violations": [],
        "physics": "두 인물 모두 평상 위에 안정적으로 앉아 체중이 지지되며, 손으로 잔을 자연스럽게 쥐고 있음."
       },
       {
        "label": "B",
        "direction": "서로 마주보며 시선을 교환하고 탁주잔을 맞부딪치고 있음.",
        "built_space": "레퍼런스의 'ㄷ'자형 평상과 그 사이의 중앙 돌 화덕 구조를 화면 중앙에 매우 정확하게 구현함.",
        "entities": "주요 인물 둘은 레퍼런스와 일치하나, 배경 측면에 프롬프트(PEOPLE 규칙)가 허용하지 않은 다수의 추가 인물들이 존재함.",
        "hard_violations": [
         "지문에 없는 배경 인물 임의 추가 (invented people)"
        ],
        "physics": "평상에 정상적으로 앉아 잔을 쥐고 있는 자세가 물리적으로 자연스러움."
       }
      ],
      "all_candidates_fail": false
     },
     "normalized": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 7,
        "verdict_ko": "미디엄 샷 및 상체 위주의 프레이밍 지시를 어기고 롱샷에 가깝게 연출되었으나, 치명적인 하드 위반 사항이 없어 최종 선택됨."
       },
       {
        "label": "B",
        "score": 3,
        "verdict_ko": "공간 구조와 프레이밍은 더 정확하게 구현했으나, 지문에 명시되지 않은 배경 인물들을 다수 추가하여 하드 위반으로 탈락함."
       }
      ],
      "all_candidates_fail": false,
      "readings": [
       {
        "label": "A",
        "direction": "두 인물이 서로 마주보며 공중에서 탁주잔을 정확히 맞부딪치고 있음.",
        "built_space": "나무 평상에 앉아 있으나, 레퍼런스의 중앙 화덕 구조가 무시되고 화덕이 인물들 뒤편으로 밀려나 공간이 변형됨.",
        "entities": "대장장이와 백정의 얼굴, 헤어, 복장이 레퍼런스와 완벽히 일치함. 탁주잔이 묘사됨. 추가 인물 없음.",
        "hard_violations": [],
        "physics": "두 인물 모두 평상 위에 안정적으로 앉아 체중이 지지되며, 손으로 잔을 자연스럽게 쥐고 있음."
       },
       {
        "label": "B",
        "direction": "서로 마주보며 시선을 교환하고 탁주잔을 맞부딪치고 있음.",
        "built_space": "레퍼런스의 'ㄷ'자형 평상과 그 사이의 중앙 돌 화덕 구조를 화면 중앙에 매우 정확하게 구현함.",
        "entities": "주요 인물 둘은 레퍼런스와 일치하나, 배경 측면에 프롬프트(PEOPLE 규칙)가 허용하지 않은 다수의 추가 인물들이 존재함.",
        "hard_violations": [
         "지문에 없는 배경 인물 임의 추가 (invented people)"
        ],
        "physics": "평상에 정상적으로 앉아 잔을 쥐고 있는 자세가 물리적으로 자연스러움."
       }
      ]
     },
     "ok": true
    },
    {
     "model": "gpt-high",
     "order": "reverse",
     "display_to_canonical": {
      "A": "B",
      "B": "A"
     },
     "raw": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 3,
        "verdict_ko": "즐거운 상체 중심의 미디엄 투숏과 건배 동작은 충실하지만, 배경에 허용되지 않은 인물 세 명을 추가하여 실격이다."
       },
       {
        "label": "B",
        "score": 7,
        "verdict_ko": "두 인물만 평상에 안정적으로 앉아 잔을 맞대고 장소도 잘 유지하지만, 하체와 발까지 보여 주어 요구한 상체 중심 미디엄 숏보다 넓다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "왼쪽 대장장이와 오른쪽 백정이 서로의 얼굴을 보며 웃는다. 각자 잔을 상대방 쪽으로 뻗어 화면 중앙에서 잔의 테두리를 맞댄다. 두 잔의 윗부분이 보이지만 대장장이의 잔은 손에 상당 부분 가려져 있다.",
        "built_space": "두 사람 아래로 하나의 나무 평상 앞면과 상판 일부가 보이고, 뒤에는 가마솥 하나와 장작 화구 하나가 있다. 왼쪽 조리도구 벽, 중앙의 밝은 출입구, 오른쪽 창과 목재 식탁들이 참고 장소와 대응한다. 두 사람은 평상 앞 가장자리에 나란히 앉아 서로 몸을 돌리고 있다. 왼쪽 뒤 식탁에는 별도의 손님 세 명이 있다.",
        "entities": "주인공 두 사람은 검은 머리의 한국인 성인 남성으로 표현되었으며, 대장장이의 남색 작업복과 걷은 소매, 백정의 두건과 해진 어두운 겉옷이 참고와 대체로 맞는다. 손에 든 도기 탁주잔 두 개, 가마솥과 흰 김, 놓인 그릇들이 보인다. 그러나 명시된 두 사람 외에 배경 인물 세 명이 추가되었다. 읽을 수 있는 글자는 보이지 않는다.",
        "hard_violations": [
         "왼쪽 배경 식탁에 허용되지 않은 인물 세 명이 추가되어, 대장장이와 백정만 보여야 한다는 인물 제한을 위반한다."
        ],
        "physics": "두 사람의 엉덩이는 평상 상판에 놓여 있고, 대장장이의 빈손도 상판을 짚고 있다. 각 잔은 해당 인물의 손가락과 손바닥으로 지지되며 팔꿈치를 굽혀 들어 올린 건배 자세가 가능하다. 가마솥은 화덕 위에 놓여 있고 그릇들은 목재 표면에 놓여 있다. 지지 없이 떠 있는 물체나 신체는 보이지 않는다."
       },
       {
        "label": "B",
        "direction": "두 남자는 서로를 바라보며 웃고, 각자 뻗은 팔 끝의 잔 두 개가 중앙에서 테두리를 맞댄다. 양쪽 잔의 테두리와 밝은 내용물이 옆쪽 시점에서 보이며, 두 팔의 건배 제스처가 분명히 구분된다.",
        "built_space": "하나의 넓은 평상과 그 앞면·상판, 뒤와 오른쪽의 연결 난간이 보인다. 중앙 뒤에 가마솥 하나와 화덕 하나, 왼쪽에 조리도구 선반, 중앙에 출입구 하나, 오른쪽에 창과 낮은 식탁들이 배치되어 참고 장소의 주요 구조를 유지한다. 두 사람은 평상 앞쪽의 서로 다른 자리를 차지한다. 다만 상판과 하체를 많이 포함해 상체 중심 미디엄 숏보다 넓은 구도다.",
        "entities": "검은 머리의 한국인 성인 남성 두 명만 보인다. 왼쪽 인물의 얼굴·머리와 남색 작업복, 오른쪽 인물의 두건·해진 겉옷은 각각의 참고 인물과 대체로 부합한다. 손에 든 탁주잔 두 개, 뒤쪽 가마솥, 흰 김과 여러 식기들이 있다. 두 사람 사이의 큰 갈색 음식 용기는 참고 사진의 작은 식기 배치와 다른 요소다. 읽을 수 있는 글자는 없다.",
        "hard_violations": [],
        "physics": "두 인물 모두 엉덩이와 접은 다리가 평상에 지지된다. 대장장이의 한쪽 다리는 앞 가장자리 아래로 내려가고, 백정은 다리를 접어 상판에 올려놓았다. 잔은 각자의 손으로 받치고 잡고 있으며 두 팔의 움직임으로 테두리를 맞댈 수 있다. 가운데 음식 용기와 식기는 상판 또는 쟁반 위에, 가마솥은 화덕 위에 놓여 있다. 지지 없는 부유나 불가능한 자세는 보이지 않는다."
       }
      ],
      "all_candidates_fail": false
     },
     "normalized": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 3,
        "verdict_ko": "즐거운 상체 중심의 미디엄 투숏과 건배 동작은 충실하지만, 배경에 허용되지 않은 인물 세 명을 추가하여 실격이다."
       },
       {
        "label": "A",
        "score": 7,
        "verdict_ko": "두 인물만 평상에 안정적으로 앉아 잔을 맞대고 장소도 잘 유지하지만, 하체와 발까지 보여 주어 요구한 상체 중심 미디엄 숏보다 넓다."
       }
      ],
      "all_candidates_fail": false,
      "readings": [
       {
        "label": "B",
        "direction": "왼쪽 대장장이와 오른쪽 백정이 서로의 얼굴을 보며 웃는다. 각자 잔을 상대방 쪽으로 뻗어 화면 중앙에서 잔의 테두리를 맞댄다. 두 잔의 윗부분이 보이지만 대장장이의 잔은 손에 상당 부분 가려져 있다.",
        "built_space": "두 사람 아래로 하나의 나무 평상 앞면과 상판 일부가 보이고, 뒤에는 가마솥 하나와 장작 화구 하나가 있다. 왼쪽 조리도구 벽, 중앙의 밝은 출입구, 오른쪽 창과 목재 식탁들이 참고 장소와 대응한다. 두 사람은 평상 앞 가장자리에 나란히 앉아 서로 몸을 돌리고 있다. 왼쪽 뒤 식탁에는 별도의 손님 세 명이 있다.",
        "entities": "주인공 두 사람은 검은 머리의 한국인 성인 남성으로 표현되었으며, 대장장이의 남색 작업복과 걷은 소매, 백정의 두건과 해진 어두운 겉옷이 참고와 대체로 맞는다. 손에 든 도기 탁주잔 두 개, 가마솥과 흰 김, 놓인 그릇들이 보인다. 그러나 명시된 두 사람 외에 배경 인물 세 명이 추가되었다. 읽을 수 있는 글자는 보이지 않는다.",
        "hard_violations": [
         "왼쪽 배경 식탁에 허용되지 않은 인물 세 명이 추가되어, 대장장이와 백정만 보여야 한다는 인물 제한을 위반한다."
        ],
        "physics": "두 사람의 엉덩이는 평상 상판에 놓여 있고, 대장장이의 빈손도 상판을 짚고 있다. 각 잔은 해당 인물의 손가락과 손바닥으로 지지되며 팔꿈치를 굽혀 들어 올린 건배 자세가 가능하다. 가마솥은 화덕 위에 놓여 있고 그릇들은 목재 표면에 놓여 있다. 지지 없이 떠 있는 물체나 신체는 보이지 않는다."
       },
       {
        "label": "A",
        "direction": "두 남자는 서로를 바라보며 웃고, 각자 뻗은 팔 끝의 잔 두 개가 중앙에서 테두리를 맞댄다. 양쪽 잔의 테두리와 밝은 내용물이 옆쪽 시점에서 보이며, 두 팔의 건배 제스처가 분명히 구분된다.",
        "built_space": "하나의 넓은 평상과 그 앞면·상판, 뒤와 오른쪽의 연결 난간이 보인다. 중앙 뒤에 가마솥 하나와 화덕 하나, 왼쪽에 조리도구 선반, 중앙에 출입구 하나, 오른쪽에 창과 낮은 식탁들이 배치되어 참고 장소의 주요 구조를 유지한다. 두 사람은 평상 앞쪽의 서로 다른 자리를 차지한다. 다만 상판과 하체를 많이 포함해 상체 중심 미디엄 숏보다 넓은 구도다.",
        "entities": "검은 머리의 한국인 성인 남성 두 명만 보인다. 왼쪽 인물의 얼굴·머리와 남색 작업복, 오른쪽 인물의 두건·해진 겉옷은 각각의 참고 인물과 대체로 부합한다. 손에 든 탁주잔 두 개, 뒤쪽 가마솥, 흰 김과 여러 식기들이 있다. 두 사람 사이의 큰 갈색 음식 용기는 참고 사진의 작은 식기 배치와 다른 요소다. 읽을 수 있는 글자는 없다.",
        "hard_violations": [],
        "physics": "두 인물 모두 엉덩이와 접은 다리가 평상에 지지된다. 대장장이의 한쪽 다리는 앞 가장자리 아래로 내려가고, 백정은 다리를 접어 상판에 올려놓았다. 잔은 각자의 손으로 받치고 잡고 있으며 두 팔의 움직임으로 테두리를 맞댈 수 있다. 가운데 음식 용기와 식기는 상판 또는 쟁반 위에, 가마솥은 화덕 위에 놓여 있다. 지지 없는 부유나 불가능한 자세는 보이지 않는다."
       }
      ]
     },
     "ok": true
    }
   ],
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "slot_winner_match": true,
   "slot_winner": {
    "gemini-pro": "A",
    "gpt-high": "A"
   },
   "route": "cross_slot_agree"
  },
  "dual": {
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "normalized": {
    "A": 2.0,
    "B": 0.857
   },
   "adjusted": {
    "A": 2.0,
    "B": 0.607
   },
   "violations": {
    "B": [
     "[gemini-pro] 지문에 없는 배경 인물 임의 추가 (invented people)",
     "[gpt-high] 왼쪽 배경 식탁에 허용되지 않은 인물 세 명이 추가되어, 대장장이와 백정만 보여야 한다는 인물 제한을 위반한다."
    ]
   },
   "per_model_winner": {
    "gemini-pro": "A",
    "gpt-high": "A"
   },
   "agreed": true
  },
  "totals": {
   "A": 2000,
   "B": 607
  },
  "selected": "A",
  "ranking": [
   "A",
   "B"
  ],
  "verdicts": [
   {
    "label": "A",
    "score": 2000,
    "verdict_ko": "미디엄 샷 및 상체 위주의 프레이밍 지시를 어기고 롱샷에 가깝게 연출되었으나, 치명적인 하드 위반 사항이 없어 최종 선택됨."
   },
   {
    "label": "B",
    "score": 607,
    "verdict_ko": "공간 구조와 프레이밍은 더 정확하게 구현했으나, 지문에 명시되지 않은 배경 인물들을 다수 추가하여 하드 위반으로 탈락함.  ★위반: [gemini-pro] 지문에 없는 배경 인물 임의 추가 (invented people) / [gpt-high] 왼쪽 배경 식탁에 허용되지 않은 인물 세 명이 추가되어, 대장장이와 백정만 보여야 한다는 인물 제한을 위반한다."
   }
  ],
  "refs": [
   {
    "label": "LOCATION PHOTOGRAPH — the exact place of this shot: its architecture, materials, fixed features and lighting mood are spatial truth; stage the moment inside this place. Never copy its camera framing.",
    "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/db00517b-165b-4005-a287-f28936b59b95/images/272bd7af-9294-4490-98c6-88a40e9deacc/scene/recipe/groupbg_soup_shop_seating_cefa56.png",
    "asset_id": "6bf78bdd-ed4e-47f1-bc6b-c786290e47aa",
    "role": "bgfirst_group_bg"
   },
   {
    "label": "CHARACTER REFERENCE — 대장장이: the exact person appearing in this shot; match face, hair and build exactly.",
    "path": "<bytes:946438>",
    "asset_id": "a12f8b90-e799-4753-9b7e-e71f4d817db5",
    "role": "character_ref"
   },
   {
    "label": "CHARACTER REFERENCE — 백정: the exact person appearing in this shot; match face, hair and build exactly.",
    "path": "<bytes:980360>",
    "asset_id": "29282584-ae7c-4722-8d77-100c4a5b2a3d",
    "role": "character_ref"
   }
  ],
  "critique_skipped": true,
  "shot_run_produced": false,
  "shot_run_uid": "06aa141d-1064-7883-a539-1746c96c2522",
  "bgfirst": {
   "bg_path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/db00517b-165b-4005-a287-f28936b59b95/images/272bd7af-9294-4490-98c6-88a40e9deacc/scene/recipe/S24sh2__bgfirst_bg.png",
   "bg_asset_id": "7d1839dc-67c7-4bef-845d-acc3564644ee",
   "bg_record_key": "S24sh2::bgfirst_bg",
   "chain_winner": true,
   "authority": "groupbg",
   "group_key": "soup_shop_seating",
   "groupbg_asset_id": "6bf78bdd-ed4e-47f1-bc6b-c786290e47aa"
  },
  "ref_mode": "재투영 배경+콘티+엔티티 (2택1: 체인 승)",
  "share_plan": {
   "ref_plan": "background"
  }
 },
 "S24sh2::cine": {
  "applied": true,
  "attempted_at": "2026-09-09T09:04:44.769147+00:00",
  "fingerprint": "51ba396937a44bfb24368318ffa78633281c9abc5505e18fa40b297a3306c224",
  "fingerprint_version": 2,
  "provider": "mai",
  "endpoint": "openrouter/chat-completions",
  "model": "microsoft/mai-image-2.6",
  "pack": "24.202608252115",
  "source_file": "S24sh2_sel.png",
  "source_sha256": "7be01d9833939fbb4cf7c14cd0a7a531f20fdfa3611ee770f65b8190899eccd1",
  "file": "S24sh2_cine.png",
  "staged_sha256": "d9af2509d9adc1d4ed059d0b6541d845af24cc602d6c1f0ebf6789eaf30365f7",
  "latency_ms": 29658
 },
 "S24sh7::signage": {
  "fp": "15a06008f9c0f296",
  "inscriptions": [],
  "cues": [],
  "dropped": []
 },
 "S24sh7": {
  "input_fingerprint": "cc5d8c8034534309",
  "prompt": "Create ONE FINAL photorealistic live-action film still of the moment below — early 20th-century Korea, Jeolla region; all people are Korean unless explicitly stated otherwise. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 허공에 멈춘 숟가락을 쥔 채 고개를 확 쳐든 오주일의 굳은 정면.\n\nLOCATION (lock): A corner dining seat in the market soup eatery, immediately beside the drinkers' wooden platform during daytime service. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: close-up\n- KEY BACKGROUND ELEMENTS: Spoon (Suspended in 오주일's hand after his eating motion stops) — Seen obliquely near the lower frame edge; used as A small arrested-action cue beneath the facial reaction; Worn collar (Retains a trace resembling something torn away) — The marked outer portion remains visible below his chin; used as Preserves the screenplay's character evidence without inventing an emblem or inscription; Corner seating area (Occupied by 오주일, apart from the neighboring conversation); used as A narrow background margin retains his separation from the other diners.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Retain the eatery's natural daytime ambient light and muted saturation, allowing stillness and natural facial contrast rather than a lighting change to create tension.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The people visible in that photo are NOT in this shot: never carry their faces, bodies or clothing onto anyone here. Each person in THIS shot is defined solely by the PEOPLE list, their own CHARACTER REFERENCE images and their wardrobe notes. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The soup cauldron continues steaming; bowls and rice-wine cups remain on the dining platforms. 오주일: He remains seated in rough clothes with a removal mark on his worn collar, his spoon suspended.\n\nTHE HAND THAT IS DOING THIS: the object at the centre of this shot is being held, operated, read, aimed or handed over by 오주일 right now, so 오주일's hand — and as much of the wrist and forearm as the framing reaches — is in the frame, gripping or resting on that object exactly the way its real use requires. Match that hand to 오주일: its size, build, skin, age, grooming, sleeve and anything worn on it belong to that person and to no one else.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 오주일 (한국인 남성, 성인 나이대, 자연스러운 검은 머리) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.",
  "roll_prompts": {
   "A": "Create ONE FINAL photorealistic live-action film still of the moment below — early 20th-century Korea, Jeolla region; all people are Korean unless explicitly stated otherwise. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 허공에 멈춘 숟가락을 쥔 채 고개를 확 쳐든 오주일의 굳은 정면.\n\nLOCATION (lock): A corner dining seat in the market soup eatery, immediately beside the drinkers' wooden platform during daytime service. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: close-up\n- KEY BACKGROUND ELEMENTS: Spoon (Suspended in 오주일's hand after his eating motion stops) — Seen obliquely near the lower frame edge; used as A small arrested-action cue beneath the facial reaction; Worn collar (Retains a trace resembling something torn away) — The marked outer portion remains visible below his chin; used as Preserves the screenplay's character evidence without inventing an emblem or inscription; Corner seating area (Occupied by 오주일, apart from the neighboring conversation); used as A narrow background margin retains his separation from the other diners.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Retain the eatery's natural daytime ambient light and muted saturation, allowing stillness and natural facial contrast rather than a lighting change to create tension.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The people visible in that photo are NOT in this shot: never carry their faces, bodies or clothing onto anyone here. Each person in THIS shot is defined solely by the PEOPLE list, their own CHARACTER REFERENCE images and their wardrobe notes. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The soup cauldron continues steaming; bowls and rice-wine cups remain on the dining platforms. 오주일: He remains seated in rough clothes with a removal mark on his worn collar, his spoon suspended.\n\nTHE HAND THAT IS DOING THIS: the object at the centre of this shot is being held, operated, read, aimed or handed over by 오주일 right now, so 오주일's hand — and as much of the wrist and forearm as the framing reaches — is in the frame, gripping or resting on that object exactly the way its real use requires. Match that hand to 오주일: its size, build, skin, age, grooming, sleeve and anything worn on it belong to that person and to no one else.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 오주일 (한국인 남성, 성인 나이대, 자연스러운 검은 머리) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.",
   "B": "Create ONE FINAL photorealistic live-action film still of the moment below — early 20th-century Korea, Jeolla region; all people are Korean unless explicitly stated otherwise. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 허공에 멈춘 숟가락을 쥔 채 고개를 확 쳐든 오주일의 굳은 정면.\n\nLOCATION (lock): A corner dining seat in the market soup eatery, immediately beside the drinkers' wooden platform during daytime service. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: close-up\n- KEY BACKGROUND ELEMENTS: Spoon (Suspended in 오주일's hand after his eating motion stops) — Seen obliquely near the lower frame edge; used as A small arrested-action cue beneath the facial reaction; Worn collar (Retains a trace resembling something torn away) — The marked outer portion remains visible below his chin; used as Preserves the screenplay's character evidence without inventing an emblem or inscription; Corner seating area (Occupied by 오주일, apart from the neighboring conversation); used as A narrow background margin retains his separation from the other diners.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Retain the eatery's natural daytime ambient light and muted saturation, allowing stillness and natural facial contrast rather than a lighting change to create tension.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The people visible in that photo are NOT in this shot: never carry their faces, bodies or clothing onto anyone here. Each person in THIS shot is defined solely by the PEOPLE list, their own CHARACTER REFERENCE images and their wardrobe notes. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The soup cauldron continues steaming; bowls and rice-wine cups remain on the dining platforms. 오주일: He remains seated in rough clothes with a removal mark on his worn collar, his spoon suspended.\n\nTHE HAND THAT IS DOING THIS: the object at the centre of this shot is being held, operated, read, aimed or handed over by 오주일 right now, so 오주일's hand — and as much of the wrist and forearm as the framing reaches — is in the frame, gripping or resting on that object exactly the way its real use requires. Match that hand to 오주일: its size, build, skin, age, grooming, sleeve and anything worn on it belong to that person and to no one else.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 오주일 (한국인 남성, 성인 나이대, 자연스러운 검은 머리) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.\n\nCOMPOSITION VARIATION (this candidate only): keep the place, the people and the moment exactly as specified, and keep from the CAMERA & FRAME contract its subject scale, screen placement and key background elements exactly. For this candidate only, its camera angle and height are deliberately left open: choose a camera position distinctly different from the obvious one, as a film director picking a second setup on the same blocking. Stay true to any explicit framing the SHOT TEXT itself dictates."
  },
  "shot_run_spend_attempt_count": 1,
  "readings": [
   {
    "label": "A",
    "direction": "시선은 카메라 정면을 강렬하게 응시하고 있으며, 오른손에 쥔 숟가락은 화면 하단을 향해 비스듬히 들린 채 멈춰 있음.",
    "built_space": "이전 샷의 주막 평상과 목재 기둥 구조물이 얕은 심도의 배경으로 나타나며, 인물은 구석 자리에 안정적으로 위치함.",
    "entities": "오주일(레퍼런스와 일치하는 얼굴 및 이목구비), 뜯겨나간 흔적이 있는 낡은 깃옷, 숟가락이 정확하게 묘사됨. (프롬프트에서 평상에 두라 한 그릇을 왼손에 든 점은 사소한 차이임)",
    "hard_violations": [],
    "physics": "오른손이 숟가락 손잡이를 단단히 쥐고 허공에 고정되어 있으며, 신체 중심이나 의복의 주름이 자연스러움."
   },
   {
    "label": "B",
    "direction": "시선이 카메라 정면을 향하고 있으며, 숟가락은 가슴 앞 허공에 고정되어 있음.",
    "built_space": "주막의 평상과 가마솥이 배경에 나타나 공간적 배경은 일치함.",
    "entities": "오주일의 얼굴과 낡은 옷, 숟가락이 묘사되었으나, 배경에 샷 텍스트에 명시되지 않은 여러 명의 식객(이전 샷의 인물들과 유사함)이 포함됨.",
    "hard_violations": [
     "[gemini-pro] 등장이 엄격히 금지된 추가 인물 및 이전 샷의 인물 무단 생성 (PEOPLE 제약 위반)",
     "[gpt-high] 오주일만 허용된 장면에 배경 인물 네 명을 추가했으며, 왼쪽의 푸른 옷 인물과 중앙의 두건 쓴 인물은 이전 장면의 인물·의상을 이어받은 것으로 보인다."
    ],
    "physics": "오른손이 숟가락을 쥐고 허공에 떠 있으며, 인물의 자세는 안정적임."
   }
  ],
  "cross_model_order": {
   "policy": "cross_model_order_v2_slot0_fwd_slot1_rev_parallel2",
   "slots": [
    {
     "model": "gemini-pro",
     "order": "forward",
     "display_to_canonical": {
      "A": "A",
      "B": "B"
     },
     "raw": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 9,
        "verdict_ko": "지정된 인물만을 프레임에 담아냈으며, 고개를 든 굳은 표정과 허공에 멈춘 숟가락이라는 핵심 액션을 클로즈업 구도 안에 훌륭하게 구현했습니다."
       },
       {
        "label": "B",
        "score": 3,
        "verdict_ko": "프롬프트에서 엄격하게 금지한 추가 인물(이전 샷의 식객들 포함)을 배경에 무단으로 등장시켜 결정적인 하드 위반을 범했습니다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "시선은 카메라 정면을 강렬하게 응시하고 있으며, 오른손에 쥔 숟가락은 화면 하단을 향해 비스듬히 들린 채 멈춰 있음.",
        "built_space": "이전 샷의 주막 평상과 목재 기둥 구조물이 얕은 심도의 배경으로 나타나며, 인물은 구석 자리에 안정적으로 위치함.",
        "entities": "오주일(레퍼런스와 일치하는 얼굴 및 이목구비), 뜯겨나간 흔적이 있는 낡은 깃옷, 숟가락이 정확하게 묘사됨. (프롬프트에서 평상에 두라 한 그릇을 왼손에 든 점은 사소한 차이임)",
        "hard_violations": [],
        "physics": "오른손이 숟가락 손잡이를 단단히 쥐고 허공에 고정되어 있으며, 신체 중심이나 의복의 주름이 자연스러움."
       },
       {
        "label": "B",
        "direction": "시선이 카메라 정면을 향하고 있으며, 숟가락은 가슴 앞 허공에 고정되어 있음.",
        "built_space": "주막의 평상과 가마솥이 배경에 나타나 공간적 배경은 일치함.",
        "entities": "오주일의 얼굴과 낡은 옷, 숟가락이 묘사되었으나, 배경에 샷 텍스트에 명시되지 않은 여러 명의 식객(이전 샷의 인물들과 유사함)이 포함됨.",
        "hard_violations": [
         "등장이 엄격히 금지된 추가 인물 및 이전 샷의 인물 무단 생성 (PEOPLE 제약 위반)"
        ],
        "physics": "오른손이 숟가락을 쥐고 허공에 떠 있으며, 인물의 자세는 안정적임."
       }
      ],
      "all_candidates_fail": false
     },
     "normalized": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 9,
        "verdict_ko": "지정된 인물만을 프레임에 담아냈으며, 고개를 든 굳은 표정과 허공에 멈춘 숟가락이라는 핵심 액션을 클로즈업 구도 안에 훌륭하게 구현했습니다."
       },
       {
        "label": "B",
        "score": 3,
        "verdict_ko": "프롬프트에서 엄격하게 금지한 추가 인물(이전 샷의 식객들 포함)을 배경에 무단으로 등장시켜 결정적인 하드 위반을 범했습니다."
       }
      ],
      "all_candidates_fail": false,
      "readings": [
       {
        "label": "A",
        "direction": "시선은 카메라 정면을 강렬하게 응시하고 있으며, 오른손에 쥔 숟가락은 화면 하단을 향해 비스듬히 들린 채 멈춰 있음.",
        "built_space": "이전 샷의 주막 평상과 목재 기둥 구조물이 얕은 심도의 배경으로 나타나며, 인물은 구석 자리에 안정적으로 위치함.",
        "entities": "오주일(레퍼런스와 일치하는 얼굴 및 이목구비), 뜯겨나간 흔적이 있는 낡은 깃옷, 숟가락이 정확하게 묘사됨. (프롬프트에서 평상에 두라 한 그릇을 왼손에 든 점은 사소한 차이임)",
        "hard_violations": [],
        "physics": "오른손이 숟가락 손잡이를 단단히 쥐고 허공에 고정되어 있으며, 신체 중심이나 의복의 주름이 자연스러움."
       },
       {
        "label": "B",
        "direction": "시선이 카메라 정면을 향하고 있으며, 숟가락은 가슴 앞 허공에 고정되어 있음.",
        "built_space": "주막의 평상과 가마솥이 배경에 나타나 공간적 배경은 일치함.",
        "entities": "오주일의 얼굴과 낡은 옷, 숟가락이 묘사되었으나, 배경에 샷 텍스트에 명시되지 않은 여러 명의 식객(이전 샷의 인물들과 유사함)이 포함됨.",
        "hard_violations": [
         "등장이 엄격히 금지된 추가 인물 및 이전 샷의 인물 무단 생성 (PEOPLE 제약 위반)"
        ],
        "physics": "오른손이 숟가락을 쥐고 허공에 떠 있으며, 인물의 자세는 안정적임."
       }
      ]
     },
     "ok": true
    },
    {
     "model": "gpt-high",
     "order": "reverse",
     "display_to_canonical": {
      "A": "B",
      "B": "A"
     },
     "raw": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 2,
        "verdict_ko": "굳은 정면과 멈춘 숟가락은 표현했지만, 배경에 허용되지 않은 인물 네 명을 넣어 단독 인물 조건을 결정적으로 위반했다."
       },
       {
        "label": "B",
        "score": 7,
        "verdict_ko": "오주일만 등장하며 정면의 긴장과 손에 멈춘 숟가락을 구현했으나, 국그릇까지 보이는 구도가 다소 넓고 옷깃의 제거 흔적은 불명확하다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "오주일의 얼굴과 두 눈은 거의 렌즈 정면을 향한다. 눈을 크게 뜨고 입을 굳힌 반응은 보이지만 고개를 확 쳐든 동세는 약하다. 숟가락은 화면 왼쪽의 오른손에서 오른쪽 아래로 비스듬히 뻗으며, 오목한 면이 위를 향한 채 입 아래에서 멈춰 있다. 배경 인물들은 오주일이 아니라 식기나 서로 있는 쪽을 향한다.",
        "built_space": "왼쪽에 나무 평상 한 구역과 난간 한 줄, 김이 나는 가마솥 하나, 그 앞의 갈색 항아리 하나와 여러 그릇이 보인다. 오른쪽에는 목재 기둥과 창틀이 이어져 이전 장소의 재료와 낮빛을 상당 부분 유지한다. 그러나 배경이 좁은 여백에 그치지 않고 여러 손님이 앉은 식사 공간을 크게 드러낸다. 오주일은 평상 옆 전경에 있으나 실제 좌면은 화면 밖이다.",
        "entities": "전경에는 검은 짧은 머리의 한국인 성인 남성으로 묘사된 오주일이 있고, 오른손에 금속 숟가락과 국물이 보인다. 얼굴은 참고 인물과 일부 유사하나 정확한 동일성은 약하다. 거친 회갈색 옷과 해진 옷깃은 있지만, 무엇을 뜯어낸 자리인지는 분명하지 않다. 배경에는 왼쪽 가장자리의 부분 인물 한 명과 머리·몸이 보이는 세 명, 총 네 명의 추가 인물이 있다. 읽을 수 있는 글자나 표식은 보이지 않는다.",
        "hard_violations": [
         "오주일만 허용된 장면에 배경 인물 네 명을 추가했으며, 왼쪽의 푸른 옷 인물과 중앙의 두건 쓴 인물은 이전 장면의 인물·의상을 이어받은 것으로 보인다."
        ],
        "physics": "숟가락 자루는 오주일의 오른손 손가락 사이에 실제로 잡혀 있고 손목과 팔뚝이 연결된다. 숟가락의 오목한 면이 위를 향해 국물을 담는 방향도 자연스럽다. 상체는 수직으로 유지되며 공중에 떠 있는 신체는 없다. 배경 인물들은 평상에 앉아 있고 가마솥과 식기들은 화덕이나 목재 면 위에 놓여 있다."
       },
       {
        "label": "B",
        "direction": "오주일의 얼굴은 정면이며 시선은 카메라 부근을 향한다. 벌어진 입과 긴장한 눈썹에서 식사를 멈추고 앞을 올려다본 반응이 읽힌다. 숟가락은 화면 왼쪽의 오른손에서 중앙 아래로 비스듬히 뻗고, 위를 향한 숟가락 머리가 입과 그릇 사이에서 멈춰 있다.",
        "built_space": "뒤에는 넓은 목재 평상 한 구역과 기둥 하나, 오른쪽 위에는 낮은 목재 상이 보인다. 전경 양옆으로 나무 난간이 이어져 구석 좌석에 앉은 배치는 가능하다. 목재의 마모와 자연스러운 낮빛은 참고 장소와 부합하지만, 가마솥이나 고유한 창호가 보이지 않아 동일 장소의 세부 확인은 제한된다. 얼굴뿐 아니라 상체와 국그릇까지 포함해 요청한 밀착 클로즈업보다 다소 넓다.",
        "entities": "확실하게 식별되는 인물은 오주일 한 명뿐이다. 검은 머리의 한국인 성인 남성이라는 조건은 맞으며 얼굴은 참고 인물과 대체로 유사하지만, 드러난 머리는 더 길고 뒤로 넘겨져 있다. 거칠고 해진 어두운 겉옷과 옷깃이 보이지만 옷깃의 특정 제거 흔적은 일반적인 해짐과 구분하기 어렵다. 금속 숟가락 하나와 하단의 국그릇 하나가 보이며, 읽을 수 있는 글자나 새 문장은 없다.",
        "hard_violations": [],
        "physics": "오른손이 숟가락 자루를 감싸 쥐고 있으며 손목과 팔의 연결은 자연스럽다. 왼손은 하단 국그릇의 가장자리를 받치고 있어 그릇도 무지지 상태가 아니다. 식사 중 손을 멈춘 자세로 실현 가능하다. 좌면과 골반은 화면 밖이지만 상체와 주변 난간의 관계에서 앉은 자세가 가능하며, 부유나 불가능한 관절은 보이지 않는다."
       }
      ],
      "all_candidates_fail": false
     },
     "normalized": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 2,
        "verdict_ko": "굳은 정면과 멈춘 숟가락은 표현했지만, 배경에 허용되지 않은 인물 네 명을 넣어 단독 인물 조건을 결정적으로 위반했다."
       },
       {
        "label": "A",
        "score": 7,
        "verdict_ko": "오주일만 등장하며 정면의 긴장과 손에 멈춘 숟가락을 구현했으나, 국그릇까지 보이는 구도가 다소 넓고 옷깃의 제거 흔적은 불명확하다."
       }
      ],
      "all_candidates_fail": false,
      "readings": [
       {
        "label": "B",
        "direction": "오주일의 얼굴과 두 눈은 거의 렌즈 정면을 향한다. 눈을 크게 뜨고 입을 굳힌 반응은 보이지만 고개를 확 쳐든 동세는 약하다. 숟가락은 화면 왼쪽의 오른손에서 오른쪽 아래로 비스듬히 뻗으며, 오목한 면이 위를 향한 채 입 아래에서 멈춰 있다. 배경 인물들은 오주일이 아니라 식기나 서로 있는 쪽을 향한다.",
        "built_space": "왼쪽에 나무 평상 한 구역과 난간 한 줄, 김이 나는 가마솥 하나, 그 앞의 갈색 항아리 하나와 여러 그릇이 보인다. 오른쪽에는 목재 기둥과 창틀이 이어져 이전 장소의 재료와 낮빛을 상당 부분 유지한다. 그러나 배경이 좁은 여백에 그치지 않고 여러 손님이 앉은 식사 공간을 크게 드러낸다. 오주일은 평상 옆 전경에 있으나 실제 좌면은 화면 밖이다.",
        "entities": "전경에는 검은 짧은 머리의 한국인 성인 남성으로 묘사된 오주일이 있고, 오른손에 금속 숟가락과 국물이 보인다. 얼굴은 참고 인물과 일부 유사하나 정확한 동일성은 약하다. 거친 회갈색 옷과 해진 옷깃은 있지만, 무엇을 뜯어낸 자리인지는 분명하지 않다. 배경에는 왼쪽 가장자리의 부분 인물 한 명과 머리·몸이 보이는 세 명, 총 네 명의 추가 인물이 있다. 읽을 수 있는 글자나 표식은 보이지 않는다.",
        "hard_violations": [
         "오주일만 허용된 장면에 배경 인물 네 명을 추가했으며, 왼쪽의 푸른 옷 인물과 중앙의 두건 쓴 인물은 이전 장면의 인물·의상을 이어받은 것으로 보인다."
        ],
        "physics": "숟가락 자루는 오주일의 오른손 손가락 사이에 실제로 잡혀 있고 손목과 팔뚝이 연결된다. 숟가락의 오목한 면이 위를 향해 국물을 담는 방향도 자연스럽다. 상체는 수직으로 유지되며 공중에 떠 있는 신체는 없다. 배경 인물들은 평상에 앉아 있고 가마솥과 식기들은 화덕이나 목재 면 위에 놓여 있다."
       },
       {
        "label": "A",
        "direction": "오주일의 얼굴은 정면이며 시선은 카메라 부근을 향한다. 벌어진 입과 긴장한 눈썹에서 식사를 멈추고 앞을 올려다본 반응이 읽힌다. 숟가락은 화면 왼쪽의 오른손에서 중앙 아래로 비스듬히 뻗고, 위를 향한 숟가락 머리가 입과 그릇 사이에서 멈춰 있다.",
        "built_space": "뒤에는 넓은 목재 평상 한 구역과 기둥 하나, 오른쪽 위에는 낮은 목재 상이 보인다. 전경 양옆으로 나무 난간이 이어져 구석 좌석에 앉은 배치는 가능하다. 목재의 마모와 자연스러운 낮빛은 참고 장소와 부합하지만, 가마솥이나 고유한 창호가 보이지 않아 동일 장소의 세부 확인은 제한된다. 얼굴뿐 아니라 상체와 국그릇까지 포함해 요청한 밀착 클로즈업보다 다소 넓다.",
        "entities": "확실하게 식별되는 인물은 오주일 한 명뿐이다. 검은 머리의 한국인 성인 남성이라는 조건은 맞으며 얼굴은 참고 인물과 대체로 유사하지만, 드러난 머리는 더 길고 뒤로 넘겨져 있다. 거칠고 해진 어두운 겉옷과 옷깃이 보이지만 옷깃의 특정 제거 흔적은 일반적인 해짐과 구분하기 어렵다. 금속 숟가락 하나와 하단의 국그릇 하나가 보이며, 읽을 수 있는 글자나 새 문장은 없다.",
        "hard_violations": [],
        "physics": "오른손이 숟가락 자루를 감싸 쥐고 있으며 손목과 팔의 연결은 자연스럽다. 왼손은 하단 국그릇의 가장자리를 받치고 있어 그릇도 무지지 상태가 아니다. 식사 중 손을 멈춘 자세로 실현 가능하다. 좌면과 골반은 화면 밖이지만 상체와 주변 난간의 관계에서 앉은 자세가 가능하며, 부유나 불가능한 관절은 보이지 않는다."
       }
      ]
     },
     "ok": true
    }
   ],
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "slot_winner_match": true,
   "slot_winner": {
    "gemini-pro": "A",
    "gpt-high": "A"
   },
   "route": "cross_slot_agree"
  },
  "dual": {
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "normalized": {
    "A": 2.0,
    "B": 0.619
   },
   "adjusted": {
    "A": 2.0,
    "B": 0.369
   },
   "violations": {
    "B": [
     "[gemini-pro] 등장이 엄격히 금지된 추가 인물 및 이전 샷의 인물 무단 생성 (PEOPLE 제약 위반)",
     "[gpt-high] 오주일만 허용된 장면에 배경 인물 네 명을 추가했으며, 왼쪽의 푸른 옷 인물과 중앙의 두건 쓴 인물은 이전 장면의 인물·의상을 이어받은 것으로 보인다."
    ]
   },
   "per_model_winner": {
    "gemini-pro": "A",
    "gpt-high": "A"
   },
   "agreed": true
  },
  "totals": {
   "A": 2000,
   "B": 369
  },
  "selected": "A",
  "ranking": [
   "A",
   "B"
  ],
  "verdicts": [
   {
    "label": "A",
    "score": 2000,
    "verdict_ko": "지정된 인물만을 프레임에 담아냈으며, 고개를 든 굳은 표정과 허공에 멈춘 숟가락이라는 핵심 액션을 클로즈업 구도 안에 훌륭하게 구현했습니다."
   },
   {
    "label": "B",
    "score": 369,
    "verdict_ko": "프롬프트에서 엄격하게 금지한 추가 인물(이전 샷의 식객들 포함)을 배경에 무단으로 등장시켜 결정적인 하드 위반을 범했습니다.  ★위반: [gemini-pro] 등장이 엄격히 금지된 추가 인물 및 이전 샷의 인물 무단 생성 (PEOPLE 제약 위반) / [gpt-high] 오주일만 허용된 장면에 배경 인물 네 명을 추가했으며, 왼쪽의 푸른 옷 인물과 중앙의 두건 쓴 인물은 이전 장면의 인물·의상을 이어받은 것으로 보인다."
   }
  ],
  "refs": [
   {
    "label": "PREVIOUS SHOT STILL — a visually related earlier shot of this same place: the location's look, materials, fixed features and lighting mood are LOCKED to this photo; never copy its camera framing. The people visible in that photo are NOT in this shot: never carry their faces, bodies or clothing onto anyone here. Each person in THIS shot is defined solely by the PEOPLE list, their own CHARACTER REFERENCE images and their wardrobe notes. If this photo shows a character who cannot move (dead or unconscious), that character's exact pose, position and orientation are ALSO LOCKED — treat their body as a fixed prop of the set that only the camera moves around.",
    "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/db00517b-165b-4005-a287-f28936b59b95/images/272bd7af-9294-4490-98c6-88a40e9deacc/scene/recipe/S24sh2_sel.png",
    "asset_id": "a47845a6-837b-4967-9805-9f3785827645",
    "role": "prev_still"
   },
   {
    "label": "CHARACTER REFERENCE — 오주일: the exact person appearing in this shot; match face, hair and build exactly.",
    "path": "<bytes:809716>",
    "asset_id": "7c31e825-ceef-4208-915f-71c4f0e15dd6",
    "role": "character_ref"
   }
  ],
  "critique_skipped": true,
  "shot_run_produced": false,
  "shot_run_uid": "06aa141f-d38d-78cb-9982-17fbacab0221",
  "ref_mode": "prev+엔티티",
  "share_plan": {
   "ref_plan": "prev",
   "prev_anchor_tag": "S24sh2"
  }
 },
 "S24sh7::cine": {
  "applied": true,
  "attempted_at": "2026-09-09T03:40:49.850577+00:00",
  "fingerprint": "b01c187dc612ac5be274bf619a9935409a222ddbe65311050055b593471dcce6",
  "fingerprint_version": 2,
  "provider": "mai",
  "endpoint": "openrouter/chat-completions",
  "model": "microsoft/mai-image-2.6",
  "pack": "24.202608252115",
  "source_file": "S24sh7_sel.png",
  "source_sha256": "ed7f6925814b1ea9da5922b844f46b63436d5527d8f3f82ccd04824d4bc4d064",
  "file": "S24sh7_cine.png",
  "staged_sha256": "51d74d1a4cab92ed970f68008ea00ba6f72208e8f95b435182946b8ad55a67e0",
  "latency_ms": 29123
 },
 "S25sh3::signage": {
  "fp": "be7d44ebd6b1ab03",
  "inscriptions": [],
  "cues": [],
  "dropped": []
 },
 "era_assess::b526508d394e43ab": {
  "subjects": [
   {
    "subject_native": "일제강점기 경찰 주재소 집무실 내부",
    "search_terms_native": [
     "일제강점기 주재소 내부",
     "일제강점기 경찰 주재소",
     "근대 관공서 집무실"
    ],
    "language_lock_native": "반드시 한국어로만 검색어를 작성해야 하며 다른 언어로 번역하거나 외래어를 덧붙이지 마십시오.",
    "reason_ko": "일제강점기 주재소 특유의 근대 목조 관공서 양식과 서무 집기 배치가 현대 사무실이나 전통 한옥과 확연히 다르기 때문"
   }
  ],
  "subject_text": "보성 주재소 토미이시 집무실\n정갈한 책상과 의자가 놓인 집무실. 책상 위에 징세 장부와 서류, 주판, 붉은 인주와 도장이 정돈되어 있다.",
  "identity": "canonical",
  "scope_id": "L22",
  "scope_role": "location_interior",
  "scope_sha": "011493550ea7b991"
 },
 "era_fail::7171538d3f3961d2": {
  "stage": "research",
  "subject": "일제강점기 경찰 주재소 집무실 내부",
  "terms": [
   "일제강점기 주재소 내부",
   "일제강점기 경찰 주재소",
   "근대 관공서 집무실"
  ],
  "status": "no_usable",
  "queries": [
   [
    "일제강점기 경찰 주재소 내부 전시",
    "일제강점기 경찰서 집무실 내부 사진"
   ],
   [
    "국가기록원 경찰관 주재소 평면도 사무실",
    "일제강점기 주재소 내부 재현 합천 영상테마파크"
   ]
  ],
  "candidate_urls": [
   "https://d3hwaim9vs2gfj.cloudfront.net/63-3131173118.jpg",
   "https://www.courtmuseum.net/app/tpl/web/images/10/s7.jpg",
   "https://upload.wikimedia.org/wikipedia/commons/a/ac/Interrogation_room.jpg",
   "https://blog.kakaocdn.net/dna/c3LRdN/btru6LSZRZq/AAAAAAAAAAAAAAAAAAAAAAhPXYcFRQKmQUQ7apv4nKs0hQ3JNTuGX7eMtDLWxIUi/img.png?allow_ip=&allow_referer=&credential=yqXZFxpELC7KVnFOS48ylbz2pIh7yKj8&expires=1774969199&signature=%2FinHr628WMofVxGrNEZei561RVY%3D"
  ],
  "coarse": {
   "eligible": [],
   "chosen_index": 0,
   "reason": "종류·보임·기준을 다 만족하는 후보가 없다 — 이 라운드에선 안 고른다",
   "single_judge": true,
   "rejected_judges": {}
  },
  "verdicts": [
   {
    "index": 1,
    "object_type_match": "yes",
    "visible": true,
    "criteria_match": "unsure",
    "similarity": 95
   },
   {
    "index": 2,
    "object_type_match": "yes",
    "visible": true,
    "criteria_match": "unsure",
    "similarity": 60
   },
   {
    "index": 3,
    "object_type_match": "no",
    "visible": true,
    "criteria_match": "unsure",
    "similarity": 10
   },
   {
    "index": 4,
    "object_type_match": "yes",
    "visible": true,
    "criteria_match": "unsure",
    "similarity": 20
   }
  ],
  "chosen_reason_ko": "",
  "attempts": 5
 },
 "S25sh3::bgfirst_bg": {
  "input_fingerprint": "c2d48360ff55c20c",
  "prompt": "Create the EMPTY BACKGROUND PLATE for one film shot — NO PEOPLE, no figures, no body parts, no sketch lines, no arrows anywhere.\n\"Empty\" means no people only: KEEP the location's inherent occupants and stock that define the place — animals in an animal shelter, pen or farm, goods in a market, moored boats in a harbour — unless the shot text explicitly removes them.\nThe FIRST attached image is a thin-line storyboard sketch: use ONLY its camera angle, horizon, perspective and the placement/size of buildings and set masses — ignore the sketched people and arrows entirely. The SECOND attached image (LOCATION PHOTOGRAPH) is the real place: take its architecture, materials, signage and fixed features, and RE-PROJECT them into the sketch's camera. If the photograph's camera differs from the sketch's, the sketch's camera wins.\nHUMAN-SCALE CALIBRATION: derive every structure's true size from human-scale elements — a door ≈ 2m, a window ≈ 1–1.5m wide, one storey ≈ 2.5–3m; never inflate a small structure or shrink a large one.\n\nSHOT TEXT this background must serve (Korean): 책상 맞은편에 두 손을 모은 채 뻣뻣하게 서서 책상 쪽을 내려다보는 배상도의 상체.\n\nLOCATION (lock): The visitor's standing space opposite a tidy desk inside a police outpost office at night, with enough light for ledger work.\n\nTIME OF DAY (lock): night.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: medium shot\n- KEY BACKGROUND ELEMENTS: 책상 (Tidy, separating 배상도 from 토미이시) — The lateral edge runs obliquely across the lower foreground; used as A restrained foreground boundary makes his joined hands and subordinate position readable.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Restrained nighttime interior illumination and muted saturation preserve facial detail while keeping the exchange sober and tense.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nRender ONE photorealistic empty location photograph, 16:9, that this shot can be staged inside later. No readable writing anywhere: surfaces that would carry writing may be present, but stage any wording out of legibility — an oblique angle, distance, shallow focus. No captions, watermarks or overlay text.",
  "effective_prompt": "Create the EMPTY BACKGROUND PLATE for one film shot — NO PEOPLE, no figures, no body parts, no sketch lines, no arrows anywhere.\n\"Empty\" means no people only: KEEP the location's inherent occupants and stock that define the place — animals in an animal shelter, pen or farm, goods in a market, moored boats in a harbour — unless the shot text explicitly removes them.\nThe FIRST attached image is a thin-line storyboard sketch: use ONLY its camera angle, horizon, perspective and the placement/size of buildings and set masses — ignore the sketched people and arrows entirely. The SECOND attached image (LOCATION PHOTOGRAPH) is the real place: take its architecture, materials, signage and fixed features, and RE-PROJECT them into the sketch's camera. If the photograph's camera differs from the sketch's, the sketch's camera wins.\nHUMAN-SCALE CALIBRATION: derive every structure's true size from human-scale elements — a door ≈ 2m, a window ≈ 1–1.5m wide, one storey ≈ 2.5–3m; never inflate a small structure or shrink a large one.\n\nSHOT TEXT this background must serve (Korean): 책상 맞은편에 두 손을 모은 채 뻣뻣하게 서서 책상 쪽을 내려다보는 배상도의 상체.\n\nLOCATION (lock): The visitor's standing space opposite a tidy desk inside a police outpost office at night, with enough light for ledger work.\n\nTIME OF DAY (lock): night.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: medium shot\n- KEY BACKGROUND ELEMENTS: 책상 (Tidy, separating 배상도 from 토미이시) — The lateral edge runs obliquely across the lower foreground; used as A restrained foreground boundary makes his joined hands and subordinate position readable.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Restrained nighttime interior illumination and muted saturation preserve facial detail while keeping the exchange sober and tense.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nRender ONE photorealistic empty location photograph, 16:9, that this shot can be staged inside later. No readable writing anywhere: surfaces that would carry writing may be present, but stage any wording out of legibility — an oblique angle, distance, shallow focus. No captions, watermarks or overlay text.",
  "bg_path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/db00517b-165b-4005-a287-f28936b59b95/images/272bd7af-9294-4490-98c6-88a40e9deacc/scene/recipe/S25sh3__bgfirst_bg.png",
  "asset_id": "8a900505-208c-4ed6-a83a-46156b1a8aa9",
  "input_asset_ids": [
   "85e61bf4-394e-42e3-a697-240357916b83",
   "f1516ce0-832a-42d8-9fc9-aca751ddff41"
  ]
 },
 "S25sh3": {
  "input_fingerprint": "81daa293acd97ecb",
  "prompt": "Create ONE FINAL photorealistic live-action film still of the moment below — early 20th-century Korea, Jeolla region; all people are Korean unless explicitly stated otherwise. TIME OF DAY (lock): night.\n\nSHOT TEXT (authoritative, Korean): 책상 맞은편에 두 손을 모은 채 뻣뻣하게 서서 책상 쪽을 내려다보는 배상도의 상체.\n\nLOCATION (lock): The visitor's standing space opposite a tidy desk inside a police outpost office at night, with enough light for ledger work. The shot takes place here — the attached LOCATION PHOTOGRAPH shows the exact spot.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: medium shot\n- KEY BACKGROUND ELEMENTS: 책상 (Tidy, separating 배상도 from 토미이시) — The lateral edge runs obliquely across the lower foreground; used as A restrained foreground boundary makes his joined hands and subordinate position readable.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Restrained nighttime interior illumination and muted saturation preserve facial detail while keeping the exchange sober and tense.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): Tax ledgers, an abacus and red seal paste are arranged on the neat desk at night. 배상도: He stands on the opposite side of the desk.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 배상도 (한국인 남성, 성인 나이대, 자연스러운 검은 머리) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.",
  "roll_prompts": {
   "A": "Stage the shot. The FIRST attached image (SHOT BACKGROUND) is the finished empty background of this shot — keep it EXACTLY: its camera, perspective, architecture, lighting and every fixed feature stay untouched. The SECOND attached image (LAYOUT SKETCH) tells you ONLY where the people go: each sketched person's position, screen size, pose and the gaze/motion arrows. Ignore the sketch's background lines. The CHARACTER REFERENCE photographs show the real people.\nPlace the real people into the background at exactly the sketched positions, sizes and poses, following the arrow directions. No sketch lines or arrows may remain.\n\nCreate ONE FINAL photorealistic live-action film still of the moment below — early 20th-century Korea, Jeolla region; all people are Korean unless explicitly stated otherwise. TIME OF DAY (lock): night.\n\nSHOT TEXT (authoritative, Korean): 책상 맞은편에 두 손을 모은 채 뻣뻣하게 서서 책상 쪽을 내려다보는 배상도의 상체.\n\nLOCATION (lock): The visitor's standing space opposite a tidy desk inside a police outpost office at night, with enough light for ledger work. The shot takes place here — the FIRST attached image (SHOT BACKGROUND) is this exact place, already built: its ground, structures, horizon, materials and lighting are the finished truth of this location and must not be redesigned or replaced. No location photograph is attached — read the place from that image alone, and add no scenery, structure, vehicle or fixture that it does not already show. This lock governs the place only; the figures in the shot follow the staging and pose instructions.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: medium shot\n- KEY BACKGROUND ELEMENTS: 책상 (Tidy, separating 배상도 from 토미이시) — The lateral edge runs obliquely across the lower foreground; used as A restrained foreground boundary makes his joined hands and subordinate position readable.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Restrained nighttime interior illumination and muted saturation preserve facial detail while keeping the exchange sober and tense.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): Tax ledgers, an abacus and red seal paste are arranged on the neat desk at night. 배상도: He stands on the opposite side of the desk.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 배상도 (한국인 남성, 성인 나이대, 자연스러운 검은 머리) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.",
   "B": "Create ONE FINAL photorealistic live-action film still of the moment below — early 20th-century Korea, Jeolla region; all people are Korean unless explicitly stated otherwise. TIME OF DAY (lock): night.\n\nSHOT TEXT (authoritative, Korean): 책상 맞은편에 두 손을 모은 채 뻣뻣하게 서서 책상 쪽을 내려다보는 배상도의 상체.\n\nLOCATION (lock): The visitor's standing space opposite a tidy desk inside a police outpost office at night, with enough light for ledger work. The shot takes place here — the attached LOCATION PHOTOGRAPH shows the exact spot.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: medium shot\n- KEY BACKGROUND ELEMENTS: 책상 (Tidy, separating 배상도 from 토미이시) — The lateral edge runs obliquely across the lower foreground; used as A restrained foreground boundary makes his joined hands and subordinate position readable.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Restrained nighttime interior illumination and muted saturation preserve facial detail while keeping the exchange sober and tense.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): Tax ledgers, an abacus and red seal paste are arranged on the neat desk at night. 배상도: He stands on the opposite side of the desk.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 배상도 (한국인 남성, 성인 나이대, 자연스러운 검은 머리) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.\n\nCOMPOSITION VARIATION (this candidate only): keep the place, the people and the moment exactly as specified, and keep from the CAMERA & FRAME contract its subject scale, screen placement and key background elements exactly. For this candidate only, its camera angle and height are deliberately left open: choose a camera position distinctly different from the obvious one, as a film director picking a second setup on the same blocking. Stay true to any explicit framing the SHOT TEXT itself dictates."
  },
  "roll_refs": {
   "A": [
    {
     "label": "SHOT BACKGROUND",
     "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/db00517b-165b-4005-a287-f28936b59b95/images/272bd7af-9294-4490-98c6-88a40e9deacc/scene/recipe/S25sh3__bgfirst_bg.png",
     "asset_id": "8a900505-208c-4ed6-a83a-46156b1a8aa9",
     "role": "bgfirst_bg"
    },
    {
     "label": "LAYOUT SKETCH (people placement only)",
     "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/db00517b-165b-4005-a287-f28936b59b95/images/272bd7af-9294-4490-98c6-88a40e9deacc/conti/conti_S25sh3.png",
     "asset_id": "85e61bf4-394e-42e3-a697-240357916b83",
     "role": "conti_light"
    },
    {
     "label": "CHARACTER REFERENCE — 배상도: the exact person appearing in this shot; match face, hair and build exactly.",
     "path": "<bytes:851718>",
     "asset_id": "034f5325-975e-4474-854b-e48359710bf8",
     "role": "character_ref"
    }
   ],
   "B": [
    {
     "label": "LOCATION PHOTOGRAPH — the exact place of this shot: its architecture, materials, fixed features and lighting mood are spatial truth; stage the moment inside this place. Never copy its camera framing.",
     "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/db00517b-165b-4005-a287-f28936b59b95/episodes/272bd7af-9294-4490-98c6-88a40e9deacc/images/background_chain/L22B03.png",
     "asset_id": "f1516ce0-832a-42d8-9fc9-aca751ddff41",
     "role": "location_plate"
    },
    {
     "label": "CHARACTER REFERENCE — 배상도: the exact person appearing in this shot; match face, hair and build exactly.",
     "path": "<bytes:851718>",
     "asset_id": "034f5325-975e-4474-854b-e48359710bf8",
     "role": "character_ref"
    }
   ]
  },
  "shot_run_spend_attempt_count": 1,
  "judge_flip_skipped": "judge_fn_owns_order",
  "initial_roll_all_fail": true,
  "readings": [
   {
    "label": "A",
    "direction": "배상도는 시선을 아래로 향해 책상을 내려다보고 있으며, 두 손을 복부 앞에 가지런히 모으고 있다.",
    "built_space": "인물은 책상 뒤(관리의 자리)에 서 있으며, 화면 앞쪽 하단에 또 다른 평행한 책상이 가로로 놓여 있어 공간 구조가 심하게 왜곡되었다. 프롬프트가 강제한 '사선 구도의 전경 책상 경계'와 '미디엄 샷'을 무시하고 평행 구도의 와이드 샷으로 렌더링되었다.",
    "entities": "배상도의 인상착의(모자, 조끼, 한복, 노리개)는 참조 이미지와 일치한다. 그러나 전경에 놓인 책상 위에 주판 형태의 물건이 좌우로 두 개 놓여 있다.",
    "hard_violations": [
     "[gemini-pro] 중복된 사물 (전경 책상 위에 주판이 2개 배치됨)",
     "[gemini-pro] 프레이밍 및 카메라 시점 위반 (사선 구도 및 미디엄 샷 지시를 완전히 무시하고 가로 평행선 및 와이드 샷으로 처리함)",
     "[gemini-pro] 잘못된 위치 지정 (인물이 방문자 쪽이 아닌 뒷벽과 측면 테이블이 있는 관리 쪽에 서 있음)",
     "[gpt-high] 참조에 하나씩 있는 주 책상과 보조 탁자를 각각 두 개로 중복 생성했다.",
     "[gpt-high] 주판을 두 개로 늘려 참조에 없는 추가 소품을 만들었다.",
     "[gpt-high] 복제한 두 책상 사이에 배상도를 세워 지정된 단일 책상 맞은편 공간을 다른 배치로 대체했다."
    ],
    "physics": "인물은 두 발로 바닥을 지지하고 서 있으며 중력에 어긋나거나 허공에 뜬 물체는 발견되지 않는다."
   },
   {
    "label": "B",
    "direction": "배상도는 시선을 아래로 향해 책상 위의 장부 쪽을 내려다보고 있으며, 두 손을 복부 앞에 단정하게 모으고 있다.",
    "built_space": "전경에 책상이 사선으로 배치되어 미디엄 샷의 경계를 형성하는 구도는 프롬프트의 지시를 정확히 따랐다. 그러나 인물이 방문자 공간이 아닌 의자와 벽보가 있는 관리의 자리에 서 있으며, 창가에 참조에 없는 사각 등불이 놓여 있다.",
    "entities": "배상도의 이목구비와 의상(조끼, 한복)은 참조와 일치하나 모자를 쓰지 않았다. 책상 위에는 장부, 주판, 인주, 도장이 있으나 참조 사진에 없는 호롱불이 임의로 추가되었다.",
    "hard_violations": [
     "[gemini-pro] 물리법칙 위반 (인물 우측 어깨 뒤편 허공에 검은색 모자의 챙 부분이 아무런 지지체 없이 떠 있음)",
     "[gemini-pro] 발명된 사물 (참조 공간에 없는 호롱불과 창가 등불이 임의로 추가됨)",
     "[gemini-pro] 잘못된 위치 지정 (인물이 프롬프트가 지시한 방문자 쪽이 아닌 관리의 의자 쪽에 서 있음)",
     "[gpt-high] 배상도를 방문객 쪽이 아니라 집무 의자가 있는 책상 안쪽에 배치했다.",
     "[gpt-high] 벽 패찰의 한자가 읽을 수 있게 노출되어 글씨 금지 조건을 위반했다.",
     "[gpt-high] 장소 참조에 없는 창가 등과 탁상등을 추가했다."
    ],
    "physics": "인물의 자세는 자연스럽게 바닥을 지지하고 있으나, 우측 어깨 뒤 허공에 검은 모자 형태의 물체가 지지체 없이 둥둥 떠 있다."
   }
  ],
  "cross_model_order": {
   "policy": "cross_model_order_v2_slot0_fwd_slot1_rev_parallel2",
   "slots": [
    {
     "model": "gemini-pro",
     "order": "forward",
     "display_to_canonical": {
      "A": "A",
      "B": "B"
     },
     "raw": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 4,
        "verdict_ko": "지시된 사선 구도의 전경 책상과 미디엄 샷 프레이밍은 훌륭히 구현했으나, 허공에 떠 있는 모자와 임의로 추가된 조명 등 치명적인 오류들이 다수 존재합니다."
       },
       {
        "label": "A",
        "score": 2,
        "verdict_ko": "프레이밍 지시(미디엄 샷, 사선 구도)를 완전히 무시하고 화각을 넓혔으며, 전경의 책상 위에 주판이 두 개로 복제되는 등 공간 구성이 크게 훼손되었습니다."
       }
      ],
      "readings": [
       {
        "label": "B",
        "direction": "배상도는 시선을 아래로 향해 책상 위의 장부 쪽을 내려다보고 있으며, 두 손을 복부 앞에 단정하게 모으고 있다.",
        "built_space": "전경에 책상이 사선으로 배치되어 미디엄 샷의 경계를 형성하는 구도는 프롬프트의 지시를 정확히 따랐다. 그러나 인물이 방문자 공간이 아닌 의자와 벽보가 있는 관리의 자리에 서 있으며, 창가에 참조에 없는 사각 등불이 놓여 있다.",
        "entities": "배상도의 이목구비와 의상(조끼, 한복)은 참조와 일치하나 모자를 쓰지 않았다. 책상 위에는 장부, 주판, 인주, 도장이 있으나 참조 사진에 없는 호롱불이 임의로 추가되었다.",
        "hard_violations": [
         "물리법칙 위반 (인물 우측 어깨 뒤편 허공에 검은색 모자의 챙 부분이 아무런 지지체 없이 떠 있음)",
         "발명된 사물 (참조 공간에 없는 호롱불과 창가 등불이 임의로 추가됨)",
         "잘못된 위치 지정 (인물이 프롬프트가 지시한 방문자 쪽이 아닌 관리의 의자 쪽에 서 있음)"
        ],
        "physics": "인물의 자세는 자연스럽게 바닥을 지지하고 있으나, 우측 어깨 뒤 허공에 검은 모자 형태의 물체가 지지체 없이 둥둥 떠 있다."
       },
       {
        "label": "A",
        "direction": "배상도는 시선을 아래로 향해 책상을 내려다보고 있으며, 두 손을 복부 앞에 가지런히 모으고 있다.",
        "built_space": "인물은 책상 뒤(관리의 자리)에 서 있으며, 화면 앞쪽 하단에 또 다른 평행한 책상이 가로로 놓여 있어 공간 구조가 심하게 왜곡되었다. 프롬프트가 강제한 '사선 구도의 전경 책상 경계'와 '미디엄 샷'을 무시하고 평행 구도의 와이드 샷으로 렌더링되었다.",
        "entities": "배상도의 인상착의(모자, 조끼, 한복, 노리개)는 참조 이미지와 일치한다. 그러나 전경에 놓인 책상 위에 주판 형태의 물건이 좌우로 두 개 놓여 있다.",
        "hard_violations": [
         "중복된 사물 (전경 책상 위에 주판이 2개 배치됨)",
         "프레이밍 및 카메라 시점 위반 (사선 구도 및 미디엄 샷 지시를 완전히 무시하고 가로 평행선 및 와이드 샷으로 처리함)",
         "잘못된 위치 지정 (인물이 방문자 쪽이 아닌 뒷벽과 측면 테이블이 있는 관리 쪽에 서 있음)"
        ],
        "physics": "인물은 두 발로 바닥을 지지하고 서 있으며 중력에 어긋나거나 허공에 뜬 물체는 발견되지 않는다."
       }
      ],
      "all_candidates_fail": true
     },
     "normalized": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 4,
        "verdict_ko": "지시된 사선 구도의 전경 책상과 미디엄 샷 프레이밍은 훌륭히 구현했으나, 허공에 떠 있는 모자와 임의로 추가된 조명 등 치명적인 오류들이 다수 존재합니다."
       },
       {
        "label": "A",
        "score": 2,
        "verdict_ko": "프레이밍 지시(미디엄 샷, 사선 구도)를 완전히 무시하고 화각을 넓혔으며, 전경의 책상 위에 주판이 두 개로 복제되는 등 공간 구성이 크게 훼손되었습니다."
       }
      ],
      "all_candidates_fail": true,
      "readings": [
       {
        "label": "B",
        "direction": "배상도는 시선을 아래로 향해 책상 위의 장부 쪽을 내려다보고 있으며, 두 손을 복부 앞에 단정하게 모으고 있다.",
        "built_space": "전경에 책상이 사선으로 배치되어 미디엄 샷의 경계를 형성하는 구도는 프롬프트의 지시를 정확히 따랐다. 그러나 인물이 방문자 공간이 아닌 의자와 벽보가 있는 관리의 자리에 서 있으며, 창가에 참조에 없는 사각 등불이 놓여 있다.",
        "entities": "배상도의 이목구비와 의상(조끼, 한복)은 참조와 일치하나 모자를 쓰지 않았다. 책상 위에는 장부, 주판, 인주, 도장이 있으나 참조 사진에 없는 호롱불이 임의로 추가되었다.",
        "hard_violations": [
         "물리법칙 위반 (인물 우측 어깨 뒤편 허공에 검은색 모자의 챙 부분이 아무런 지지체 없이 떠 있음)",
         "발명된 사물 (참조 공간에 없는 호롱불과 창가 등불이 임의로 추가됨)",
         "잘못된 위치 지정 (인물이 프롬프트가 지시한 방문자 쪽이 아닌 관리의 의자 쪽에 서 있음)"
        ],
        "physics": "인물의 자세는 자연스럽게 바닥을 지지하고 있으나, 우측 어깨 뒤 허공에 검은 모자 형태의 물체가 지지체 없이 둥둥 떠 있다."
       },
       {
        "label": "A",
        "direction": "배상도는 시선을 아래로 향해 책상을 내려다보고 있으며, 두 손을 복부 앞에 가지런히 모으고 있다.",
        "built_space": "인물은 책상 뒤(관리의 자리)에 서 있으며, 화면 앞쪽 하단에 또 다른 평행한 책상이 가로로 놓여 있어 공간 구조가 심하게 왜곡되었다. 프롬프트가 강제한 '사선 구도의 전경 책상 경계'와 '미디엄 샷'을 무시하고 평행 구도의 와이드 샷으로 렌더링되었다.",
        "entities": "배상도의 인상착의(모자, 조끼, 한복, 노리개)는 참조 이미지와 일치한다. 그러나 전경에 놓인 책상 위에 주판 형태의 물건이 좌우로 두 개 놓여 있다.",
        "hard_violations": [
         "중복된 사물 (전경 책상 위에 주판이 2개 배치됨)",
         "프레이밍 및 카메라 시점 위반 (사선 구도 및 미디엄 샷 지시를 완전히 무시하고 가로 평행선 및 와이드 샷으로 처리함)",
         "잘못된 위치 지정 (인물이 방문자 쪽이 아닌 뒷벽과 측면 테이블이 있는 관리 쪽에 서 있음)"
        ],
        "physics": "인물은 두 발로 바닥을 지지하고 서 있으며 중력에 어긋나거나 허공에 뜬 물체는 발견되지 않는다."
       }
      ]
     },
     "ok": true
    },
    {
     "model": "gpt-high",
     "order": "reverse",
     "display_to_canonical": {
      "A": "B",
      "B": "A"
     },
     "raw": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 3,
        "verdict_ko": "두 손을 모으고 책상을 내려다보는 상체와 사선의 전경 책상은 더 충실하지만, 방문객 위치가 아닌 의자 앞에 서 있고 읽을 수 있는 벽 글씨가 있어 부적격이다."
       },
       {
        "label": "B",
        "score": 2,
        "verdict_ko": "모자와 복장, 모은 손은 잘 맞지만 책상·보조 탁자·주판을 중복 생성해 장소를 바꾸었고, 넓은 공간과 큰 전경이 상체 중심 구도를 약화한다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "고개와 눈이 화면 오른쪽 아래의 책상 상판 및 펼친 장부 쪽을 향한다. 카메라를 보지 않으며, 책상 쪽을 내려다보라는 동작과 맞는다.",
        "built_space": "주 책상 하나, 그 뒤 의자 하나, 격자창 하나가 보이며 목재 기둥과 회벽은 참조와 유사하다. 그러나 배상도는 의자 바로 앞, 즉 참조에서 집무자가 앉는 책상 안쪽에 서 있다. 방문객이 서야 할 반대편 위치가 아니다. 책상 가장자리는 하단을 비스듬히 가로지르지만 상판이 전경의 상당 부분을 차지한다. 참조에 없는 창가 등과 탁상등도 보인다.",
        "entities": "성인 동아시아계 남성 한 명으로 한국인 배상도 설정과 외관상 충돌하지 않는다. 검은 머리, 흰 전통 의복, 검은 조끼, 허리띠와 늘어진 장식은 대체로 맞지만 참조의 검은 모자를 머리에 쓰지 않았고 얼굴의 동일성도 덜 확실하다. 장부 여러 권, 주판 하나, 붉은 인주 그릇 하나와 도장 하나가 보인다. 벽 패찰에는 개별 한자가 식별될 만큼 선명한 글씨가 있다. 토미이시나 다른 사람은 보이지 않는다.",
        "hard_violations": [
         "배상도를 방문객 쪽이 아니라 집무 의자가 있는 책상 안쪽에 배치했다.",
         "벽 패찰의 한자가 읽을 수 있게 노출되어 글씨 금지 조건을 위반했다.",
         "장소 참조에 없는 창가 등과 탁상등을 추가했다."
        ],
        "physics": "상체는 서 있는 사람의 수직 자세로 이어지고 두 손은 아랫배 앞에서 서로 맞잡혀 있다. 발은 프레임 밖이므로 접지는 직접 확인할 수 없지만 공중에 떠 있다는 징후는 없다. 장부·주판·인주·도장·탁상등은 책상에 놓여 있다. 등 뒤의 검은 둥근 형태는 일부가 가려져 정체와 부착 방식을 확정하기 어렵다."
       },
       {
        "label": "B",
        "direction": "고개와 눈이 아래로 향하며 시선은 모은 손 앞쪽의 전경 책상 가장자리와 장부 방향으로 떨어진다. 정면 응시는 아니고 책상을 내려다보는 동작으로 읽힌다.",
        "built_space": "주 책상이 전경과 후경에 하나씩 두 개, 왼쪽 보조 탁자도 앞뒤로 두 개 보인다. 의자 하나는 후경 책상 뒤에 있고, 배상도는 두 책상 사이에 서 있다. 참조의 책상 하나와 보조 탁자 하나인 공간을 복제해 방문객 위치를 만들었다. 격자창 하나와 목재·회벽 재질은 유사하지만, 카메라가 많은 바닥과 후경 가구를 보여 주고 전경 상판도 크게 잡아 절제된 상체 중심 구도에서 벗어난다.",
        "entities": "성인 동아시아계 남성 한 명만 보이며 검은 모자, 흰 전통 의복, 검은 조끼, 허리띠와 장식은 배상도 참조에 더 가깝다. 얼굴은 모자 그늘에 일부 가려져 있다. 전경에는 장부 묶음, 주판 두 개, 붉은 인주 그릇 하나와 도장 하나가 있다. 왼쪽 두 탁자에는 각각 책 더미가 놓였다. 장부에는 격자선만 뚜렷하고 읽을 수 있는 문구는 보이지 않는다.",
        "hard_violations": [
         "참조에 하나씩 있는 주 책상과 보조 탁자를 각각 두 개로 중복 생성했다.",
         "주판을 두 개로 늘려 참조에 없는 추가 소품을 만들었다.",
         "복제한 두 책상 사이에 배상도를 세워 지정된 단일 책상 맞은편 공간을 다른 배치로 대체했다."
        ],
        "physics": "배상도는 곧게 선 몸통과 서로 맞잡은 두 손을 보여 주며 해부학적으로 가능한 정지 자세다. 하체와 발은 전경 책상에 가려졌지만 부유 징후는 없다. 모자는 머리에 얹혀 있고 장식은 허리에서 내려온다. 장부·주판·인주·도장은 상판에, 책 더미는 각 탁자에 지지되어 있다."
       }
      ],
      "all_candidates_fail": true
     },
     "normalized": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 3,
        "verdict_ko": "두 손을 모으고 책상을 내려다보는 상체와 사선의 전경 책상은 더 충실하지만, 방문객 위치가 아닌 의자 앞에 서 있고 읽을 수 있는 벽 글씨가 있어 부적격이다."
       },
       {
        "label": "A",
        "score": 2,
        "verdict_ko": "모자와 복장, 모은 손은 잘 맞지만 책상·보조 탁자·주판을 중복 생성해 장소를 바꾸었고, 넓은 공간과 큰 전경이 상체 중심 구도를 약화한다."
       }
      ],
      "all_candidates_fail": true,
      "readings": [
       {
        "label": "B",
        "direction": "고개와 눈이 화면 오른쪽 아래의 책상 상판 및 펼친 장부 쪽을 향한다. 카메라를 보지 않으며, 책상 쪽을 내려다보라는 동작과 맞는다.",
        "built_space": "주 책상 하나, 그 뒤 의자 하나, 격자창 하나가 보이며 목재 기둥과 회벽은 참조와 유사하다. 그러나 배상도는 의자 바로 앞, 즉 참조에서 집무자가 앉는 책상 안쪽에 서 있다. 방문객이 서야 할 반대편 위치가 아니다. 책상 가장자리는 하단을 비스듬히 가로지르지만 상판이 전경의 상당 부분을 차지한다. 참조에 없는 창가 등과 탁상등도 보인다.",
        "entities": "성인 동아시아계 남성 한 명으로 한국인 배상도 설정과 외관상 충돌하지 않는다. 검은 머리, 흰 전통 의복, 검은 조끼, 허리띠와 늘어진 장식은 대체로 맞지만 참조의 검은 모자를 머리에 쓰지 않았고 얼굴의 동일성도 덜 확실하다. 장부 여러 권, 주판 하나, 붉은 인주 그릇 하나와 도장 하나가 보인다. 벽 패찰에는 개별 한자가 식별될 만큼 선명한 글씨가 있다. 토미이시나 다른 사람은 보이지 않는다.",
        "hard_violations": [
         "배상도를 방문객 쪽이 아니라 집무 의자가 있는 책상 안쪽에 배치했다.",
         "벽 패찰의 한자가 읽을 수 있게 노출되어 글씨 금지 조건을 위반했다.",
         "장소 참조에 없는 창가 등과 탁상등을 추가했다."
        ],
        "physics": "상체는 서 있는 사람의 수직 자세로 이어지고 두 손은 아랫배 앞에서 서로 맞잡혀 있다. 발은 프레임 밖이므로 접지는 직접 확인할 수 없지만 공중에 떠 있다는 징후는 없다. 장부·주판·인주·도장·탁상등은 책상에 놓여 있다. 등 뒤의 검은 둥근 형태는 일부가 가려져 정체와 부착 방식을 확정하기 어렵다."
       },
       {
        "label": "A",
        "direction": "고개와 눈이 아래로 향하며 시선은 모은 손 앞쪽의 전경 책상 가장자리와 장부 방향으로 떨어진다. 정면 응시는 아니고 책상을 내려다보는 동작으로 읽힌다.",
        "built_space": "주 책상이 전경과 후경에 하나씩 두 개, 왼쪽 보조 탁자도 앞뒤로 두 개 보인다. 의자 하나는 후경 책상 뒤에 있고, 배상도는 두 책상 사이에 서 있다. 참조의 책상 하나와 보조 탁자 하나인 공간을 복제해 방문객 위치를 만들었다. 격자창 하나와 목재·회벽 재질은 유사하지만, 카메라가 많은 바닥과 후경 가구를 보여 주고 전경 상판도 크게 잡아 절제된 상체 중심 구도에서 벗어난다.",
        "entities": "성인 동아시아계 남성 한 명만 보이며 검은 모자, 흰 전통 의복, 검은 조끼, 허리띠와 장식은 배상도 참조에 더 가깝다. 얼굴은 모자 그늘에 일부 가려져 있다. 전경에는 장부 묶음, 주판 두 개, 붉은 인주 그릇 하나와 도장 하나가 있다. 왼쪽 두 탁자에는 각각 책 더미가 놓였다. 장부에는 격자선만 뚜렷하고 읽을 수 있는 문구는 보이지 않는다.",
        "hard_violations": [
         "참조에 하나씩 있는 주 책상과 보조 탁자를 각각 두 개로 중복 생성했다.",
         "주판을 두 개로 늘려 참조에 없는 추가 소품을 만들었다.",
         "복제한 두 책상 사이에 배상도를 세워 지정된 단일 책상 맞은편 공간을 다른 배치로 대체했다."
        ],
        "physics": "배상도는 곧게 선 몸통과 서로 맞잡은 두 손을 보여 주며 해부학적으로 가능한 정지 자세다. 하체와 발은 전경 책상에 가려졌지만 부유 징후는 없다. 모자는 머리에 얹혀 있고 장식은 허리에서 내려온다. 장부·주판·인주·도장은 상판에, 책 더미는 각 탁자에 지지되어 있다."
       }
      ]
     },
     "ok": true
    }
   ],
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "slot_winner_match": true,
   "slot_winner": {
    "gemini-pro": "B",
    "gpt-high": "B"
   },
   "route": "cross_slot_agree"
  },
  "dual": {
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "normalized": {
    "A": 1.167,
    "B": 2.0
   },
   "adjusted": {
    "A": 0.917,
    "B": 1.75
   },
   "violations": {
    "B": [
     "[gemini-pro] 물리법칙 위반 (인물 우측 어깨 뒤편 허공에 검은색 모자의 챙 부분이 아무런 지지체 없이 떠 있음)",
     "[gemini-pro] 발명된 사물 (참조 공간에 없는 호롱불과 창가 등불이 임의로 추가됨)",
     "[gemini-pro] 잘못된 위치 지정 (인물이 프롬프트가 지시한 방문자 쪽이 아닌 관리의 의자 쪽에 서 있음)",
     "[gpt-high] 배상도를 방문객 쪽이 아니라 집무 의자가 있는 책상 안쪽에 배치했다.",
     "[gpt-high] 벽 패찰의 한자가 읽을 수 있게 노출되어 글씨 금지 조건을 위반했다.",
     "[gpt-high] 장소 참조에 없는 창가 등과 탁상등을 추가했다."
    ],
    "A": [
     "[gemini-pro] 중복된 사물 (전경 책상 위에 주판이 2개 배치됨)",
     "[gemini-pro] 프레이밍 및 카메라 시점 위반 (사선 구도 및 미디엄 샷 지시를 완전히 무시하고 가로 평행선 및 와이드 샷으로 처리함)",
     "[gemini-pro] 잘못된 위치 지정 (인물이 방문자 쪽이 아닌 뒷벽과 측면 테이블이 있는 관리 쪽에 서 있음)",
     "[gpt-high] 참조에 하나씩 있는 주 책상과 보조 탁자를 각각 두 개로 중복 생성했다.",
     "[gpt-high] 주판을 두 개로 늘려 참조에 없는 추가 소품을 만들었다.",
     "[gpt-high] 복제한 두 책상 사이에 배상도를 세워 지정된 단일 책상 맞은편 공간을 다른 배치로 대체했다."
    ]
   },
   "per_model_winner": {
    "gemini-pro": "B",
    "gpt-high": "B"
   },
   "agreed": true
  },
  "totals": {
   "B": 1750,
   "A": 917
  },
  "selected": "B",
  "ranking": [
   "B",
   "A"
  ],
  "verdicts": [
   {
    "label": "B",
    "score": 1750,
    "verdict_ko": "지시된 사선 구도의 전경 책상과 미디엄 샷 프레이밍은 훌륭히 구현했으나, 허공에 떠 있는 모자와 임의로 추가된 조명 등 치명적인 오류들이 다수 존재합니다.  ★위반: [gemini-pro] 물리법칙 위반 (인물 우측 어깨 뒤편 허공에 검은색 모자의 챙 부분이 아무런 지지체 없이 떠 있음) / [gemini-pro] 발명된 사물 (참조 공간에 없는 호롱불과 창가 등불이 임의로 추가됨) / [gemini-pro] 잘못된 위치 지정 (인물이 프롬프트가 지시한 방문자 쪽이 아닌 관리의 의자 쪽에 서 있음) / [gpt-high] 배상도를 방문객 쪽이 아니라 집무 의자가 있는 책상 안쪽에 배치했다. / [gpt-high] 벽 패찰의 한자가 읽을 수 있게 노출되어 글씨 금지 조건을 위반했다. / [gpt-high] 장소 참조에 없는 창가 등과 탁상등을 추가했다."
   },
   {
    "label": "A",
    "score": 917,
    "verdict_ko": "프레이밍 지시(미디엄 샷, 사선 구도)를 완전히 무시하고 화각을 넓혔으며, 전경의 책상 위에 주판이 두 개로 복제되는 등 공간 구성이 크게 훼손되었습니다.  ★위반: [gemini-pro] 중복된 사물 (전경 책상 위에 주판이 2개 배치됨) / [gemini-pro] 프레이밍 및 카메라 시점 위반 (사선 구도 및 미디엄 샷 지시를 완전히 무시하고 가로 평행선 및 와이드 샷으로 처리함) / [gemini-pro] 잘못된 위치 지정 (인물이 방문자 쪽이 아닌 뒷벽과 측면 테이블이 있는 관리 쪽에 서 있음) / [gpt-high] 참조에 하나씩 있는 주 책상과 보조 탁자를 각각 두 개로 중복 생성했다. / [gpt-high] 주판을 두 개로 늘려 참조에 없는 추가 소품을 만들었다. / [gpt-high] 복제한 두 책상 사이에 배상도를 세워 지정된 단일 책상 맞은편 공간을 다른 배치로 대체했다."
   }
  ],
  "refs": [
   {
    "label": "LOCATION PHOTOGRAPH — the exact place of this shot: its architecture, materials, fixed features and lighting mood are spatial truth; stage the moment inside this place. Never copy its camera framing.",
    "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/db00517b-165b-4005-a287-f28936b59b95/episodes/272bd7af-9294-4490-98c6-88a40e9deacc/images/background_chain/L22B03.png",
    "asset_id": "f1516ce0-832a-42d8-9fc9-aca751ddff41",
    "role": "location_plate"
   },
   {
    "label": "CHARACTER REFERENCE — 배상도: the exact person appearing in this shot; match face, hair and build exactly.",
    "path": "<bytes:851718>",
    "asset_id": "034f5325-975e-4474-854b-e48359710bf8",
    "role": "character_ref"
   }
  ],
  "critique_skipped": true,
  "needs_reshoot": true,
  "shot_run_produced": false,
  "shot_run_uid": "06aa141f-d51a-7921-bc03-4da8c48e59cb",
  "bgfirst": {
   "bg_path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/db00517b-165b-4005-a287-f28936b59b95/images/272bd7af-9294-4490-98c6-88a40e9deacc/scene/recipe/S25sh3__bgfirst_bg.png",
   "bg_asset_id": "8a900505-208c-4ed6-a83a-46156b1a8aa9",
   "bg_record_key": "S25sh3::bgfirst_bg",
   "chain_winner": false,
   "authority": "plate"
  },
  "ref_mode": "플레이트+엔티티 (2택1: 무콘티 승)",
  "share_plan": {
   "ref_plan": "background"
  }
 },
 "S25sh3::cine": {
  "applied": true,
  "attempted_at": "2026-09-09T09:06:05.094701+00:00",
  "fingerprint": "daf10bfc720dfd9a3feec3d026b6f955d4b7467158905442220ff38500971c0a",
  "fingerprint_version": 2,
  "provider": "mai",
  "endpoint": "openrouter/chat-completions",
  "model": "microsoft/mai-image-2.6",
  "pack": "24.202608252115",
  "source_file": "S25sh3_sel.png",
  "source_sha256": "a732ffc9b1cddcde22aeefa6470a21cd3fc55dbeda11bba237badcbb99f13283",
  "file": "S25sh3_cine.png",
  "staged_sha256": "db3fe7b4b97062415b25972043b95e5ce3be923e8de0fcbbf0800f36028a04c1",
  "latency_ms": 29207
 },
 "S25sh8::signage": {
  "fp": "ea511d89434e9de0",
  "inscriptions": [],
  "cues": [],
  "dropped": []
 },
 "S25sh8": {
  "input_fingerprint": "f3d9485197e03b6e",
  "prompt": "Create ONE FINAL photorealistic live-action film still of the moment below — early 20th-century Korea, Jeolla region; all people are Korean unless explicitly stated otherwise. TIME OF DAY (lock): night.\n\nSHOT TEXT (authoritative, Korean): 고개를 든 채 배상도를 향해 서늘한 시선을 던지는 토미이시의 눈빛 클로즈업.\n\nLOCATION (lock): The seated working side of the police outpost office desk, with nighttime light over the open tax ledger. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: close-up\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Maintain the preceding nighttime interior illumination, with naturalistic facial contrast and no heightened lighting effect on his threatening gaze.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The people visible in that photo are NOT in this shot: never carry their faces, bodies or clothing onto anyone here. Each person in THIS shot is defined solely by the PEOPLE list, their own CHARACTER REFERENCE images and their wardrobe notes. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): A red line now crosses one ledger figure; the abacus and red seal paste remain on the neat desk. 토미이시: He remains seated at the ledger, now looking up.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 토미이시 (일본인 남성, 성인 나이대, 자연스러운 검은 머리) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.",
  "roll_prompts": {
   "A": "Create ONE FINAL photorealistic live-action film still of the moment below — early 20th-century Korea, Jeolla region; all people are Korean unless explicitly stated otherwise. TIME OF DAY (lock): night.\n\nSHOT TEXT (authoritative, Korean): 고개를 든 채 배상도를 향해 서늘한 시선을 던지는 토미이시의 눈빛 클로즈업.\n\nLOCATION (lock): The seated working side of the police outpost office desk, with nighttime light over the open tax ledger. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: close-up\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Maintain the preceding nighttime interior illumination, with naturalistic facial contrast and no heightened lighting effect on his threatening gaze.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The people visible in that photo are NOT in this shot: never carry their faces, bodies or clothing onto anyone here. Each person in THIS shot is defined solely by the PEOPLE list, their own CHARACTER REFERENCE images and their wardrobe notes. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): A red line now crosses one ledger figure; the abacus and red seal paste remain on the neat desk. 토미이시: He remains seated at the ledger, now looking up.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 토미이시 (일본인 남성, 성인 나이대, 자연스러운 검은 머리) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.",
   "B": "Create ONE FINAL photorealistic live-action film still of the moment below — early 20th-century Korea, Jeolla region; all people are Korean unless explicitly stated otherwise. TIME OF DAY (lock): night.\n\nSHOT TEXT (authoritative, Korean): 고개를 든 채 배상도를 향해 서늘한 시선을 던지는 토미이시의 눈빛 클로즈업.\n\nLOCATION (lock): The seated working side of the police outpost office desk, with nighttime light over the open tax ledger. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: close-up\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Maintain the preceding nighttime interior illumination, with naturalistic facial contrast and no heightened lighting effect on his threatening gaze.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The people visible in that photo are NOT in this shot: never carry their faces, bodies or clothing onto anyone here. Each person in THIS shot is defined solely by the PEOPLE list, their own CHARACTER REFERENCE images and their wardrobe notes. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): A red line now crosses one ledger figure; the abacus and red seal paste remain on the neat desk. 토미이시: He remains seated at the ledger, now looking up.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 토미이시 (일본인 남성, 성인 나이대, 자연스러운 검은 머리) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.\n\nCOMPOSITION VARIATION (this candidate only): keep the place, the people and the moment exactly as specified, and keep from the CAMERA & FRAME contract its subject scale, screen placement and key background elements exactly. For this candidate only, its camera angle and height are deliberately left open: choose a camera position distinctly different from the obvious one, as a film director picking a second setup on the same blocking. Stay true to any explicit framing the SHOT TEXT itself dictates."
  },
  "shot_run_spend_attempt_count": 1,
  "readings": [
   {
    "label": "A",
    "direction": "화면 왼쪽 전경에 있는 인물을 향해 시선을 고정하고 있습니다.",
    "built_space": "책상과 램프, 장부, 주판, 인주가 배치되어 있으며, 배경의 나무 기둥 등 공간적 요소는 이전 샷의 분위기를 따르고 있습니다.",
    "entities": "주인공의 얼굴이 캐릭터 레퍼런스(토미이시)가 아닌 이전 샷(위치 레퍼런스)에 있던 인물의 얼굴과 완전히 동일하며, 복장 역시 지정된 군복이 아닙니다.",
    "hard_violations": [
     "[gemini-pro] 이전 샷 레퍼런스의 인물 얼굴을 그대로 가져와 사용함 (프롬프트의 엄격한 금지 사항 위반)",
     "[gemini-pro] 지정된 인물(토미이시)의 정체성 및 복장 불일치",
     "[gpt-high] 허용되지 않은 두 번째 남자의 턱과 몸통을 왼쪽 전경에 추가했다.",
     "[gpt-high] 장부의 숫자와 기입 문자를 읽을 수 있게 노출하여 읽을 수 있는 글씨 금지 조건을 위반했다."
    ],
    "physics": "앉아 있는 자세 자체는 무리가 없습니다."
   },
   {
    "label": "B",
    "direction": "카메라 왼쪽 전경에 걸쳐진 인물(배상도)을 향해 고개를 들고 날카로운 시선을 던지고 있습니다.",
    "built_space": "이전 샷과 동일한 나무 벽면을 배경으로 책상 앞에 앉아 있으며, 책상 위에는 지정된 빨간 줄이 그어진 장부, 주판, 인주, 램프가 놓여 있습니다.",
    "entities": "주인공은 캐릭터 레퍼런스로 제공된 토미이시의 얼굴, 머리 스타일, 군복을 정확하게 착용하고 있습니다. 전경에 보이는 인물의 뒷모습은 이전 샷의 인물 복장과 일치합니다.",
    "hard_violations": [
     "[gpt-high] 토미이시만 허용된 화면에 왼쪽 전경 남자의 머리와 어깨를 추가했으며, 이전 장면 인물의 의상까지 재사용한 모습이다."
    ],
    "physics": "책상 앞에 자연스럽게 앉아 고개를 든 자세가 물리적으로 안정적입니다."
   }
  ],
  "cross_model_order": {
   "policy": "cross_model_order_v2_slot0_fwd_slot1_rev_parallel2",
   "slots": [
    {
     "model": "gemini-pro",
     "order": "forward",
     "display_to_canonical": {
      "A": "A",
      "B": "B"
     },
     "raw": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 9,
        "verdict_ko": "샷 텍스트에 묘사된 서늘한 시선을 훌륭하게 연출했으며, 토미이시의 얼굴과 군복, 그리고 책상 위 빨간 줄이 그어진 장부 등 지정된 디테일을 정확하게 구현했습니다."
       },
       {
        "label": "A",
        "score": 2,
        "verdict_ko": "이전 샷 레퍼런스에 등장한 인물의 얼굴을 그대로 복사해 오는 치명적인 지시 위반을 범했으며, 지정된 캐릭터 레퍼런스(토미이시)를 전혀 반영하지 않았습니다."
       }
      ],
      "readings": [
       {
        "label": "B",
        "direction": "카메라 왼쪽 전경에 걸쳐진 인물(배상도)을 향해 고개를 들고 날카로운 시선을 던지고 있습니다.",
        "built_space": "이전 샷과 동일한 나무 벽면을 배경으로 책상 앞에 앉아 있으며, 책상 위에는 지정된 빨간 줄이 그어진 장부, 주판, 인주, 램프가 놓여 있습니다.",
        "entities": "주인공은 캐릭터 레퍼런스로 제공된 토미이시의 얼굴, 머리 스타일, 군복을 정확하게 착용하고 있습니다. 전경에 보이는 인물의 뒷모습은 이전 샷의 인물 복장과 일치합니다.",
        "hard_violations": [],
        "physics": "책상 앞에 자연스럽게 앉아 고개를 든 자세가 물리적으로 안정적입니다."
       },
       {
        "label": "A",
        "direction": "화면 왼쪽 전경에 있는 인물을 향해 시선을 고정하고 있습니다.",
        "built_space": "책상과 램프, 장부, 주판, 인주가 배치되어 있으며, 배경의 나무 기둥 등 공간적 요소는 이전 샷의 분위기를 따르고 있습니다.",
        "entities": "주인공의 얼굴이 캐릭터 레퍼런스(토미이시)가 아닌 이전 샷(위치 레퍼런스)에 있던 인물의 얼굴과 완전히 동일하며, 복장 역시 지정된 군복이 아닙니다.",
        "hard_violations": [
         "이전 샷 레퍼런스의 인물 얼굴을 그대로 가져와 사용함 (프롬프트의 엄격한 금지 사항 위반)",
         "지정된 인물(토미이시)의 정체성 및 복장 불일치"
        ],
        "physics": "앉아 있는 자세 자체는 무리가 없습니다."
       }
      ],
      "all_candidates_fail": false
     },
     "normalized": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 9,
        "verdict_ko": "샷 텍스트에 묘사된 서늘한 시선을 훌륭하게 연출했으며, 토미이시의 얼굴과 군복, 그리고 책상 위 빨간 줄이 그어진 장부 등 지정된 디테일을 정확하게 구현했습니다."
       },
       {
        "label": "A",
        "score": 2,
        "verdict_ko": "이전 샷 레퍼런스에 등장한 인물의 얼굴을 그대로 복사해 오는 치명적인 지시 위반을 범했으며, 지정된 캐릭터 레퍼런스(토미이시)를 전혀 반영하지 않았습니다."
       }
      ],
      "all_candidates_fail": false,
      "readings": [
       {
        "label": "B",
        "direction": "카메라 왼쪽 전경에 걸쳐진 인물(배상도)을 향해 고개를 들고 날카로운 시선을 던지고 있습니다.",
        "built_space": "이전 샷과 동일한 나무 벽면을 배경으로 책상 앞에 앉아 있으며, 책상 위에는 지정된 빨간 줄이 그어진 장부, 주판, 인주, 램프가 놓여 있습니다.",
        "entities": "주인공은 캐릭터 레퍼런스로 제공된 토미이시의 얼굴, 머리 스타일, 군복을 정확하게 착용하고 있습니다. 전경에 보이는 인물의 뒷모습은 이전 샷의 인물 복장과 일치합니다.",
        "hard_violations": [],
        "physics": "책상 앞에 자연스럽게 앉아 고개를 든 자세가 물리적으로 안정적입니다."
       },
       {
        "label": "A",
        "direction": "화면 왼쪽 전경에 있는 인물을 향해 시선을 고정하고 있습니다.",
        "built_space": "책상과 램프, 장부, 주판, 인주가 배치되어 있으며, 배경의 나무 기둥 등 공간적 요소는 이전 샷의 분위기를 따르고 있습니다.",
        "entities": "주인공의 얼굴이 캐릭터 레퍼런스(토미이시)가 아닌 이전 샷(위치 레퍼런스)에 있던 인물의 얼굴과 완전히 동일하며, 복장 역시 지정된 군복이 아닙니다.",
        "hard_violations": [
         "이전 샷 레퍼런스의 인물 얼굴을 그대로 가져와 사용함 (프롬프트의 엄격한 금지 사항 위반)",
         "지정된 인물(토미이시)의 정체성 및 복장 불일치"
        ],
        "physics": "앉아 있는 자세 자체는 무리가 없습니다."
       }
      ]
     },
     "ok": true
    },
    {
     "model": "gpt-high",
     "order": "reverse",
     "display_to_canonical": {
      "A": "B",
      "B": "A"
     },
     "raw": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 3,
        "verdict_ko": "시선과 군복, 상대적으로 큰 얼굴 구도는 B보다 가깝지만, 금지된 상대 인물을 추가했고 토미이시의 참조 얼굴·수염·모자를 재현하지 못했다."
       },
       {
        "label": "B",
        "score": 1,
        "verdict_ko": "올려다보는 동작은 있으나 이전 장면 인물에 가까운 얼굴과 민간복을 사용했고, 추가 인물과 읽히는 장부 글씨까지 보여 눈빛 클로즈업 요구를 크게 벗어났다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "오른쪽 남자는 고개를 돌려 왼쪽 전경 남자의 얼굴을 응시한다. 배상도를 향한 시선으로 해석할 수 있는 구체적 상대는 있지만, 그 상대를 화면에 직접 추가했다. 눈은 자연스러운 홍채와 동공을 유지하며 서늘하게 좁혀져 있다. 무기나 이동 동작은 없다.",
        "built_space": "목재 책상 하나 위에 펼친 장부 하나, 주판 하나, 유리 석유등 하나, 붉은 인주통 하나, 도장 하나와 책 더미가 보인다. 중복된 비품은 없다. 주인공은 책상 오른쪽 작업 면에 앉은 듯하고 상대는 왼쪽 전경을 차지한다. 따뜻한 야간 조명과 낡은 목재는 참조에 가깝지만, 얼굴뿐 아니라 장부와 상대 어깨까지 크게 담아 눈빛 중심의 밀착 구도에서 벗어난다. 반사는 없다.",
        "entities": "주인공은 검은 머리의 성인 동아시아계 남성이며 군복과 갈색 가죽 어깨끈을 착용했다. 국적은 외형만으로 확정할 수 없다. 참조 토미이시의 콧수염·턱수염과 군모가 없고 얼굴 및 머리 모양도 다르며, 군복은 참조보다 밝다. 왼쪽에는 이전 장면의 흰옷·검은 조끼를 연상시키는 별도 인물이 보인다. 장부의 붉은 선은 숫자 하나를 긋는 수준이 아니라 펼친 면을 길게 가로지른다. 글씨는 흐리고 비스듬해 뚜렷하게 판독되지는 않는다.",
        "hard_violations": [
         "토미이시만 허용된 화면에 왼쪽 전경 남자의 머리와 어깨를 추가했으며, 이전 장면 인물의 의상까지 재사용한 모습이다."
        ],
        "physics": "머리와 목, 어깨의 연결 및 고개를 돌리는 자세는 가능하다. 좌석과 하체는 프레임 밖이므로 실제 착석 접점은 확인할 수 없지만 공중에 떠 있다는 증거는 없다. 장부·주판·등잔·인주통·도장·책은 모두 책상에 놓여 지지된다."
       },
       {
        "label": "B",
        "direction": "앉은 남자는 눈을 왼쪽 위로 올려 전경에 선 인물의 얼굴 쪽을 본다. 배상도를 올려다보는 관계는 읽히지만 금지된 상대 신체를 노출했다. 장부는 글자의 윗부분이 앉은 남자 쪽을 향해, 그 남자보다 카메라 쪽 사람이 읽기 좋은 방향으로 펼쳐져 있다.",
        "built_space": "목재 책상 하나, 장부 하나, 석유등 하나, 인주통 하나, 그 뒤 도장 하나, 오른쪽 주판 하나와 가장자리 책 더미들이 보인다. 회벽과 짙은 목재 기둥·가로대는 이전 장소의 재료감을 따른다. 남자는 책상 뒤에 있으나 어깨가 상판 가까이에 놓일 만큼 낮게 앉아 작업 좌석 관계가 어색하다. 전경 인물과 넓은 책상 면을 함께 담아 눈빛 클로즈업보다 훨씬 넓다. 불가능한 반사나 비품 중복은 보이지 않는다.",
        "entities": "중앙 남자는 검은 머리의 젊은 성인 동아시아계 남성이지만, 참조 토미이시의 수염과 군모가 없고 군복 대신 어두운 민간복을 입었다. 얼굴과 가르마는 오히려 이전 장면 남자에 가깝다. 왼쪽 전경에는 별도 성인 남자의 턱과 밝은 옷을 입은 몸통이 추가되어 있다. 주판·인주·등잔은 해당 소품으로 식별된다. 장부에는 붉은 선이 있으나 여러 칸을 가로지르며, 숫자와 한자 형태의 기입이 선명하게 노출된다.",
        "hard_violations": [
         "허용되지 않은 두 번째 남자의 턱과 몸통을 왼쪽 전경에 추가했다.",
         "장부의 숫자와 기입 문자를 읽을 수 있게 노출하여 읽을 수 있는 글씨 금지 조건을 위반했다."
        ],
        "physics": "올려다보는 목과 눈의 자세는 물리적으로 가능하다. 착석 높이는 어색하지만 좌석과 하체가 가려져 있어 지지 없는 부유로 판정할 근거는 없다. 장부와 주판, 인주통, 등잔 및 책 더미는 상판에 놓여 있다. 전경 인물의 지지점 역시 화면 밖이다."
       }
      ],
      "all_candidates_fail": true
     },
     "normalized": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 3,
        "verdict_ko": "시선과 군복, 상대적으로 큰 얼굴 구도는 B보다 가깝지만, 금지된 상대 인물을 추가했고 토미이시의 참조 얼굴·수염·모자를 재현하지 못했다."
       },
       {
        "label": "A",
        "score": 1,
        "verdict_ko": "올려다보는 동작은 있으나 이전 장면 인물에 가까운 얼굴과 민간복을 사용했고, 추가 인물과 읽히는 장부 글씨까지 보여 눈빛 클로즈업 요구를 크게 벗어났다."
       }
      ],
      "all_candidates_fail": true,
      "readings": [
       {
        "label": "B",
        "direction": "오른쪽 남자는 고개를 돌려 왼쪽 전경 남자의 얼굴을 응시한다. 배상도를 향한 시선으로 해석할 수 있는 구체적 상대는 있지만, 그 상대를 화면에 직접 추가했다. 눈은 자연스러운 홍채와 동공을 유지하며 서늘하게 좁혀져 있다. 무기나 이동 동작은 없다.",
        "built_space": "목재 책상 하나 위에 펼친 장부 하나, 주판 하나, 유리 석유등 하나, 붉은 인주통 하나, 도장 하나와 책 더미가 보인다. 중복된 비품은 없다. 주인공은 책상 오른쪽 작업 면에 앉은 듯하고 상대는 왼쪽 전경을 차지한다. 따뜻한 야간 조명과 낡은 목재는 참조에 가깝지만, 얼굴뿐 아니라 장부와 상대 어깨까지 크게 담아 눈빛 중심의 밀착 구도에서 벗어난다. 반사는 없다.",
        "entities": "주인공은 검은 머리의 성인 동아시아계 남성이며 군복과 갈색 가죽 어깨끈을 착용했다. 국적은 외형만으로 확정할 수 없다. 참조 토미이시의 콧수염·턱수염과 군모가 없고 얼굴 및 머리 모양도 다르며, 군복은 참조보다 밝다. 왼쪽에는 이전 장면의 흰옷·검은 조끼를 연상시키는 별도 인물이 보인다. 장부의 붉은 선은 숫자 하나를 긋는 수준이 아니라 펼친 면을 길게 가로지른다. 글씨는 흐리고 비스듬해 뚜렷하게 판독되지는 않는다.",
        "hard_violations": [
         "토미이시만 허용된 화면에 왼쪽 전경 남자의 머리와 어깨를 추가했으며, 이전 장면 인물의 의상까지 재사용한 모습이다."
        ],
        "physics": "머리와 목, 어깨의 연결 및 고개를 돌리는 자세는 가능하다. 좌석과 하체는 프레임 밖이므로 실제 착석 접점은 확인할 수 없지만 공중에 떠 있다는 증거는 없다. 장부·주판·등잔·인주통·도장·책은 모두 책상에 놓여 지지된다."
       },
       {
        "label": "A",
        "direction": "앉은 남자는 눈을 왼쪽 위로 올려 전경에 선 인물의 얼굴 쪽을 본다. 배상도를 올려다보는 관계는 읽히지만 금지된 상대 신체를 노출했다. 장부는 글자의 윗부분이 앉은 남자 쪽을 향해, 그 남자보다 카메라 쪽 사람이 읽기 좋은 방향으로 펼쳐져 있다.",
        "built_space": "목재 책상 하나, 장부 하나, 석유등 하나, 인주통 하나, 그 뒤 도장 하나, 오른쪽 주판 하나와 가장자리 책 더미들이 보인다. 회벽과 짙은 목재 기둥·가로대는 이전 장소의 재료감을 따른다. 남자는 책상 뒤에 있으나 어깨가 상판 가까이에 놓일 만큼 낮게 앉아 작업 좌석 관계가 어색하다. 전경 인물과 넓은 책상 면을 함께 담아 눈빛 클로즈업보다 훨씬 넓다. 불가능한 반사나 비품 중복은 보이지 않는다.",
        "entities": "중앙 남자는 검은 머리의 젊은 성인 동아시아계 남성이지만, 참조 토미이시의 수염과 군모가 없고 군복 대신 어두운 민간복을 입었다. 얼굴과 가르마는 오히려 이전 장면 남자에 가깝다. 왼쪽 전경에는 별도 성인 남자의 턱과 밝은 옷을 입은 몸통이 추가되어 있다. 주판·인주·등잔은 해당 소품으로 식별된다. 장부에는 붉은 선이 있으나 여러 칸을 가로지르며, 숫자와 한자 형태의 기입이 선명하게 노출된다.",
        "hard_violations": [
         "허용되지 않은 두 번째 남자의 턱과 몸통을 왼쪽 전경에 추가했다.",
         "장부의 숫자와 기입 문자를 읽을 수 있게 노출하여 읽을 수 있는 글씨 금지 조건을 위반했다."
        ],
        "physics": "올려다보는 목과 눈의 자세는 물리적으로 가능하다. 착석 높이는 어색하지만 좌석과 하체가 가려져 있어 지지 없는 부유로 판정할 근거는 없다. 장부와 주판, 인주통, 등잔 및 책 더미는 상판에 놓여 있다. 전경 인물의 지지점 역시 화면 밖이다."
       }
      ]
     },
     "ok": true
    }
   ],
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "slot_winner_match": true,
   "slot_winner": {
    "gemini-pro": "B",
    "gpt-high": "B"
   },
   "route": "cross_slot_agree"
  },
  "dual": {
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "normalized": {
    "A": 0.556,
    "B": 2.0
   },
   "adjusted": {
    "A": 0.306,
    "B": 1.75
   },
   "violations": {
    "A": [
     "[gemini-pro] 이전 샷 레퍼런스의 인물 얼굴을 그대로 가져와 사용함 (프롬프트의 엄격한 금지 사항 위반)",
     "[gemini-pro] 지정된 인물(토미이시)의 정체성 및 복장 불일치",
     "[gpt-high] 허용되지 않은 두 번째 남자의 턱과 몸통을 왼쪽 전경에 추가했다.",
     "[gpt-high] 장부의 숫자와 기입 문자를 읽을 수 있게 노출하여 읽을 수 있는 글씨 금지 조건을 위반했다."
    ],
    "B": [
     "[gpt-high] 토미이시만 허용된 화면에 왼쪽 전경 남자의 머리와 어깨를 추가했으며, 이전 장면 인물의 의상까지 재사용한 모습이다."
    ]
   },
   "per_model_winner": {
    "gemini-pro": "B",
    "gpt-high": "B"
   },
   "agreed": true
  },
  "totals": {
   "B": 1750,
   "A": 306
  },
  "selected": "B",
  "ranking": [
   "B",
   "A"
  ],
  "verdicts": [
   {
    "label": "B",
    "score": 1750,
    "verdict_ko": "샷 텍스트에 묘사된 서늘한 시선을 훌륭하게 연출했으며, 토미이시의 얼굴과 군복, 그리고 책상 위 빨간 줄이 그어진 장부 등 지정된 디테일을 정확하게 구현했습니다.  ★위반: [gpt-high] 토미이시만 허용된 화면에 왼쪽 전경 남자의 머리와 어깨를 추가했으며, 이전 장면 인물의 의상까지 재사용한 모습이다."
   },
   {
    "label": "A",
    "score": 306,
    "verdict_ko": "이전 샷 레퍼런스에 등장한 인물의 얼굴을 그대로 복사해 오는 치명적인 지시 위반을 범했으며, 지정된 캐릭터 레퍼런스(토미이시)를 전혀 반영하지 않았습니다.  ★위반: [gemini-pro] 이전 샷 레퍼런스의 인물 얼굴을 그대로 가져와 사용함 (프롬프트의 엄격한 금지 사항 위반) / [gemini-pro] 지정된 인물(토미이시)의 정체성 및 복장 불일치 / [gpt-high] 허용되지 않은 두 번째 남자의 턱과 몸통을 왼쪽 전경에 추가했다. / [gpt-high] 장부의 숫자와 기입 문자를 읽을 수 있게 노출하여 읽을 수 있는 글씨 금지 조건을 위반했다."
   }
  ],
  "refs": [
   {
    "label": "PREVIOUS SHOT STILL — a visually related earlier shot of this same place: the location's look, materials, fixed features and lighting mood are LOCKED to this photo; never copy its camera framing. The people visible in that photo are NOT in this shot: never carry their faces, bodies or clothing onto anyone here. Each person in THIS shot is defined solely by the PEOPLE list, their own CHARACTER REFERENCE images and their wardrobe notes. If this photo shows a character who cannot move (dead or unconscious), that character's exact pose, position and orientation are ALSO LOCKED — treat their body as a fixed prop of the set that only the camera moves around.",
    "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/db00517b-165b-4005-a287-f28936b59b95/images/272bd7af-9294-4490-98c6-88a40e9deacc/scene/recipe/S25sh3_sel.png",
    "asset_id": "e77b68ad-1828-4014-af71-a331a0663cbb",
    "role": "prev_still"
   },
   {
    "label": "CHARACTER REFERENCE — 토미이시: the exact person appearing in this shot; match face, hair and build exactly.",
    "path": "<bytes:797699>",
    "asset_id": "c861ba0f-8750-4262-8324-25cf5c6c9dcd",
    "role": "character_ref"
   }
  ],
  "critique_skipped": true,
  "shot_run_produced": false,
  "shot_run_uid": "06aa1422-fa59-75cc-aa7a-9918202b36e7",
  "ref_mode": "prev+엔티티",
  "share_plan": {
   "ref_plan": "prev",
   "prev_anchor_tag": "S25sh3"
  }
 },
 "S25sh8::cine": {
  "applied": true,
  "attempted_at": "2026-09-09T03:46:22.183864+00:00",
  "fingerprint": "9d8e86fd33c79fe9071b131240c517fa1d81e142de9b225cd8565d8d7a0921a8",
  "fingerprint_version": 2,
  "provider": "mai",
  "endpoint": "openrouter/chat-completions",
  "model": "microsoft/mai-image-2.6",
  "pack": "24.202608252115",
  "source_file": "S25sh8_sel.png",
  "source_sha256": "6cfa45c780884b4468e3df3674355e11593a16f655f073f4dfdbe153cc19b30f",
  "file": "S25sh8_cine.png",
  "staged_sha256": "7a3a7135d1ce033e1d16ad52ee4085a52b3c1169bd440b9cdb32b8f96a28e938",
  "latency_ms": 31989
 },
 "S25sh13::signage": {
  "fp": "e9d11e160c4a661e",
  "inscriptions": [],
  "cues": [
   {
    "text_native": "",
    "source": "scene_text_implied",
    "source_quote": "도장"
   },
   {
    "text_native": "",
    "source": "scene_text_implied",
    "source_quote": "장부"
   }
  ],
  "dropped": []
 },
 "S25sh13": {
  "input_fingerprint": "ba76bb2169edb48d",
  "prompt": "Create ONE FINAL photorealistic live-action film still of the moment below — early 20th-century Korea, Jeolla region; all people are Korean unless explicitly stated otherwise. TIME OF DAY (lock): night.\n\nSHOT TEXT (authoritative, Korean): 붉은 인주가 묻은 도장이 장부 종이 위를 강하게 짓누른 클로즈업.\n\nLOCATION (lock): The illuminated writing surface of the police outpost office desk at night, holding a tax ledger, abacus, and red ink pad. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: close-up\n- KEY BACKGROUND ELEMENTS: 도장 (Red-inked stamping face pressed firmly against the ledger) — The side of the stamp and its contact with the page are visible; the printing face is concealed against the paper; used as The compression point supplies the decisive action without enlarging the object disproportionately; 징세 장부 (Open beneath the stamp, with a number crossed by a red line) — The written face of the open page is visible obliquely, showing numerical entries and the red correction; used as A partial page surrounding the contact point connects the stamp to taxation rather than an abstract impact; 책상 (Tidy beneath the open ledger) — A section of the top and the lateral edge remain visible around the cropped page; used as Provides scale and continuity with the preceding desk-side movement.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Continue the restrained nighttime illumination, allowing the established red ink to register without introducing a new light source or exaggerated contrast.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The people visible in that photo are NOT in this shot: never carry their faces, bodies or clothing onto anyone here. Each person in THIS shot is defined solely by the PEOPLE list, their own CHARACTER REFERENCE images and their wardrobe notes. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The ledger retains the red strike-through and now receives a red seal impression; the abacus and seal paste remain beside it.\n\nNO PEOPLE IN THIS SHOT: the shot text shows only the place and its state — no living person appears in frame, and no face, and no figure standing, sitting or passing through. Do not invent someone to occupy the space.\n\nOne thing this rule does NOT forbid: the hand or arm of whoever is doing what the shot text describes. If the moment is an object being held, carried, operated, read, aimed, worn or handed over, the hand doing it belongs in frame, resting on or gripping that object exactly as its real use requires. An object photographed floating unsupported, with the hand that must be holding it erased, is a worse failure than the person you were avoiding — render the hand.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.",
  "roll_prompts": {
   "A": "Create ONE FINAL photorealistic live-action film still of the moment below — early 20th-century Korea, Jeolla region; all people are Korean unless explicitly stated otherwise. TIME OF DAY (lock): night.\n\nSHOT TEXT (authoritative, Korean): 붉은 인주가 묻은 도장이 장부 종이 위를 강하게 짓누른 클로즈업.\n\nLOCATION (lock): The illuminated writing surface of the police outpost office desk at night, holding a tax ledger, abacus, and red ink pad. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: close-up\n- KEY BACKGROUND ELEMENTS: 도장 (Red-inked stamping face pressed firmly against the ledger) — The side of the stamp and its contact with the page are visible; the printing face is concealed against the paper; used as The compression point supplies the decisive action without enlarging the object disproportionately; 징세 장부 (Open beneath the stamp, with a number crossed by a red line) — The written face of the open page is visible obliquely, showing numerical entries and the red correction; used as A partial page surrounding the contact point connects the stamp to taxation rather than an abstract impact; 책상 (Tidy beneath the open ledger) — A section of the top and the lateral edge remain visible around the cropped page; used as Provides scale and continuity with the preceding desk-side movement.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Continue the restrained nighttime illumination, allowing the established red ink to register without introducing a new light source or exaggerated contrast.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The people visible in that photo are NOT in this shot: never carry their faces, bodies or clothing onto anyone here. Each person in THIS shot is defined solely by the PEOPLE list, their own CHARACTER REFERENCE images and their wardrobe notes. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The ledger retains the red strike-through and now receives a red seal impression; the abacus and seal paste remain beside it.\n\nNO PEOPLE IN THIS SHOT: the shot text shows only the place and its state — no living person appears in frame, and no face, and no figure standing, sitting or passing through. Do not invent someone to occupy the space.\n\nOne thing this rule does NOT forbid: the hand or arm of whoever is doing what the shot text describes. If the moment is an object being held, carried, operated, read, aimed, worn or handed over, the hand doing it belongs in frame, resting on or gripping that object exactly as its real use requires. An object photographed floating unsupported, with the hand that must be holding it erased, is a worse failure than the person you were avoiding — render the hand.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.",
   "B": "Create ONE FINAL photorealistic live-action film still of the moment below — early 20th-century Korea, Jeolla region; all people are Korean unless explicitly stated otherwise. TIME OF DAY (lock): night.\n\nSHOT TEXT (authoritative, Korean): 붉은 인주가 묻은 도장이 장부 종이 위를 강하게 짓누른 클로즈업.\n\nLOCATION (lock): The illuminated writing surface of the police outpost office desk at night, holding a tax ledger, abacus, and red ink pad. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: close-up\n- KEY BACKGROUND ELEMENTS: 도장 (Red-inked stamping face pressed firmly against the ledger) — The side of the stamp and its contact with the page are visible; the printing face is concealed against the paper; used as The compression point supplies the decisive action without enlarging the object disproportionately; 징세 장부 (Open beneath the stamp, with a number crossed by a red line) — The written face of the open page is visible obliquely, showing numerical entries and the red correction; used as A partial page surrounding the contact point connects the stamp to taxation rather than an abstract impact; 책상 (Tidy beneath the open ledger) — A section of the top and the lateral edge remain visible around the cropped page; used as Provides scale and continuity with the preceding desk-side movement.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Continue the restrained nighttime illumination, allowing the established red ink to register without introducing a new light source or exaggerated contrast.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The people visible in that photo are NOT in this shot: never carry their faces, bodies or clothing onto anyone here. Each person in THIS shot is defined solely by the PEOPLE list, their own CHARACTER REFERENCE images and their wardrobe notes. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The ledger retains the red strike-through and now receives a red seal impression; the abacus and seal paste remain beside it.\n\nNO PEOPLE IN THIS SHOT: the shot text shows only the place and its state — no living person appears in frame, and no face, and no figure standing, sitting or passing through. Do not invent someone to occupy the space.\n\nOne thing this rule does NOT forbid: the hand or arm of whoever is doing what the shot text describes. If the moment is an object being held, carried, operated, read, aimed, worn or handed over, the hand doing it belongs in frame, resting on or gripping that object exactly as its real use requires. An object photographed floating unsupported, with the hand that must be holding it erased, is a worse failure than the person you were avoiding — render the hand.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.\n\nCOMPOSITION VARIATION (this candidate only): keep the place, the people and the moment exactly as specified, and keep from the CAMERA & FRAME contract its subject scale, screen placement and key background elements exactly. For this candidate only, its camera angle and height are deliberately left open: choose a camera position distinctly different from the obvious one, as a film director picking a second setup on the same blocking. Stay true to any explicit framing the SHOT TEXT itself dictates."
  },
  "shot_run_spend_attempt_count": 1,
  "initial_roll_all_fail": true,
  "readings": [
   {
    "label": "A",
    "direction": "도장이 종이를 향하나 크게 기울어져 도장면이 노출됨. 짓누르는 동작이 아님.",
    "built_space": "책상 위 장부, 인주, 주판이 놓임. 우측에 사람의 턱과 어깨가 보임.",
    "entities": "도장, 붉은 인주, 주판 일치. 장부에 '징세', '장부' 등 명확히 읽히는 한글이 있음.",
    "hard_violations": [
     "[gemini-pro] 명확히 읽을 수 있는 텍스트 등장",
     "[gemini-pro] 손과 팔 외의 신체 부위(턱, 어깨) 등장",
     "[gpt-high] 장부의 글자와 숫자 및 인장 문양이 판독 가능한 수준으로 드러나 읽을 수 있는 글씨 금지 조건을 위반한다."
    ],
    "physics": "손이 도장을 잡고 있으며 도장의 측면 모서리만 종이에 닿아 있음."
   },
   {
    "label": "B",
    "direction": "도장이 종이를 향하지만 위로 들려 도장면이 카메라에 완전히 노출됨. 짓누르는 동작이 아님.",
    "built_space": "책상 위 장부, 인주, 주판, 램프가 놓임. 손목까지만 프레임에 포함됨.",
    "entities": "도장, 붉은 인주, 주판 일치. 장부의 글씨는 알아볼 수 없게 처리됨.",
    "hard_violations": [
     "[gpt-high] 장부에 판독 가능한 숫자가 노출되어 읽을 수 있는 글씨 금지 조건을 위반한다.",
     "[gpt-high] 사용 중인 도장 외에 배경에 별도의 검은 손잡이 도장이 남아 있어 도장이 중복된다."
    ],
    "physics": "손이 도장을 쥔 채 종이 위 허공에 떠 있음."
   }
  ],
  "cross_model_order": {
   "policy": "cross_model_order_v2_slot0_fwd_slot1_rev_parallel2",
   "slots": [
    {
     "model": "gemini-pro",
     "order": "forward",
     "display_to_canonical": {
      "A": "A",
      "B": "B"
     },
     "raw": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 3,
        "verdict_ko": "도장을 짓누르지 않고 도장면이 노출되었으며, 금지된 텍스트와 신체 부위가 프레임에 들어와 심각한 오류가 발생함."
       },
       {
        "label": "B",
        "score": 4,
        "verdict_ko": "금지 조항(텍스트, 인물 배제)은 지켰으나, 도장을 종이에 밀착해 짓누르는 핵심 묘사에 실패함."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "도장이 종이를 향하나 크게 기울어져 도장면이 노출됨. 짓누르는 동작이 아님.",
        "built_space": "책상 위 장부, 인주, 주판이 놓임. 우측에 사람의 턱과 어깨가 보임.",
        "entities": "도장, 붉은 인주, 주판 일치. 장부에 '징세', '장부' 등 명확히 읽히는 한글이 있음.",
        "hard_violations": [
         "명확히 읽을 수 있는 텍스트 등장",
         "손과 팔 외의 신체 부위(턱, 어깨) 등장"
        ],
        "physics": "손이 도장을 잡고 있으며 도장의 측면 모서리만 종이에 닿아 있음."
       },
       {
        "label": "B",
        "direction": "도장이 종이를 향하지만 위로 들려 도장면이 카메라에 완전히 노출됨. 짓누르는 동작이 아님.",
        "built_space": "책상 위 장부, 인주, 주판, 램프가 놓임. 손목까지만 프레임에 포함됨.",
        "entities": "도장, 붉은 인주, 주판 일치. 장부의 글씨는 알아볼 수 없게 처리됨.",
        "hard_violations": [],
        "physics": "손이 도장을 쥔 채 종이 위 허공에 떠 있음."
       }
      ],
      "all_candidates_fail": true
     },
     "normalized": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 3,
        "verdict_ko": "도장을 짓누르지 않고 도장면이 노출되었으며, 금지된 텍스트와 신체 부위가 프레임에 들어와 심각한 오류가 발생함."
       },
       {
        "label": "B",
        "score": 4,
        "verdict_ko": "금지 조항(텍스트, 인물 배제)은 지켰으나, 도장을 종이에 밀착해 짓누르는 핵심 묘사에 실패함."
       }
      ],
      "all_candidates_fail": true,
      "readings": [
       {
        "label": "A",
        "direction": "도장이 종이를 향하나 크게 기울어져 도장면이 노출됨. 짓누르는 동작이 아님.",
        "built_space": "책상 위 장부, 인주, 주판이 놓임. 우측에 사람의 턱과 어깨가 보임.",
        "entities": "도장, 붉은 인주, 주판 일치. 장부에 '징세', '장부' 등 명확히 읽히는 한글이 있음.",
        "hard_violations": [
         "명확히 읽을 수 있는 텍스트 등장",
         "손과 팔 외의 신체 부위(턱, 어깨) 등장"
        ],
        "physics": "손이 도장을 잡고 있으며 도장의 측면 모서리만 종이에 닿아 있음."
       },
       {
        "label": "B",
        "direction": "도장이 종이를 향하지만 위로 들려 도장면이 카메라에 완전히 노출됨. 짓누르는 동작이 아님.",
        "built_space": "책상 위 장부, 인주, 주판, 램프가 놓임. 손목까지만 프레임에 포함됨.",
        "entities": "도장, 붉은 인주, 주판 일치. 장부의 글씨는 알아볼 수 없게 처리됨.",
        "hard_violations": [],
        "physics": "손이 도장을 쥔 채 종이 위 허공에 떠 있음."
       }
      ]
     },
     "ok": true
    },
    {
     "model": "gpt-high",
     "order": "reverse",
     "display_to_canonical": {
      "A": "B",
      "B": "A"
     },
     "raw": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 2,
        "verdict_ko": "더 밀착된 클로즈업이지만, 커진 원형 도장의 인면이 카메라에 노출되고 장부를 평평하게 누르지 않으며, 읽히는 숫자와 여분 도장까지 보인다."
       },
       {
        "label": "B",
        "score": 3,
        "verdict_ko": "직사각형 도장의 형태와 크기는 참조에 더 가깝지만, 장부를 누르는 순간이 아니라 들어 올린 순간이며 읽을 수 있는 글자와 숫자도 금지 조건을 위반한다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "손이 도장 손잡이를 잡고 장부 쪽으로 기울이고 있지만, 붉은 인면 대부분은 종이가 아니라 카메라를 향한다. 아래쪽 가장자리만 종이에 닿는 듯하여 인면을 감춘 채 강하게 짓누르는 방향과 맞지 않는다. 얼굴이나 시선은 없다.",
        "built_space": "나무 책상 위에 펼친 장부 하나가 전경을 채우고, 왼쪽에 주판 하나, 뒤쪽에 둥근 인주통 하나와 등잔 하나, 오른쪽 뒤에 종이 더미가 있다. 책상 상판과 앞쪽 측면도 보이며 따뜻하고 어두운 조명은 참조와 이어진다. 사용 중인 원형 도장 외에 뒤쪽에 검은 손잡이의 별도 도장이 보인다.",
        "entities": "장부의 붉은 취소선, 주판, 붉은 인주와 낡은 목재 표면은 요구와 맞는다. 사용 중인 도장은 참조의 작고 검은 직사각형 도장과 달리 큰 원형 목제 도장이다. 장부에는 ‘104’처럼 판독 가능한 숫자가 있다. 인물은 손만 보이므로 나이·성별·한국인 여부를 확정할 수 없고, 얼굴이나 몸통은 없다.",
        "hard_violations": [
         "장부에 판독 가능한 숫자가 노출되어 읽을 수 있는 글씨 금지 조건을 위반한다.",
         "사용 중인 도장 외에 배경에 별도의 검은 손잡이 도장이 남아 있어 도장이 중복된다."
        ],
        "physics": "도장은 손가락이 손잡이를 감싸 지지하고, 장부는 책상에 놓여 있다. 주판과 인주통, 등잔도 책상에 지지된다. 부유하거나 지지 없는 물체는 없지만, 비스듬한 도장의 가장자리 접촉으로는 인면 전체를 종이에 강하게 압착하는 동작이 성립하지 않는다."
       },
       {
        "label": "B",
        "direction": "손이 잡은 도장의 인면은 대체로 장부 아래쪽을 향하고, 카메라에는 주로 측면이 보인다. 그러나 도장 아래로 이미 찍힌 붉은 인장이 드러나고 종이와 간격이 있어, 목표 지점을 누르는 순간보다 찍은 뒤 들어 올리는 순간으로 읽힌다. 얼굴이나 시선은 없다.",
        "built_space": "펼친 장부 하나가 책상 중앙에 놓이고, 주판 하나가 왼쪽 뒤, 둥근 인주통 하나가 위쪽에 있다. 오른쪽 뒤에는 종이 더미가 있으며 책상 상판과 오른쪽 앞 측면이 보인다. 참조의 낡은 나무 책상과 절제된 따뜻한 조명은 유지되지만, 접촉점보다 장부 전체를 많이 보여 주어 요구한 클로즈업의 집중도는 낮다.",
        "entities": "검은 손잡이와 직사각형 밑판의 도장은 참조의 형태와 비교적 가깝고 손에 비례한 크기도 자연스럽다. 장부에는 붉은 취소선과 직사각형 인장이 있고 주판과 인주도 남아 있다. 다만 ‘100’ 등의 숫자와 여러 글자가 읽힌다. 손과 소매만 보이고 얼굴이나 몸통은 없으며, 손만으로 인물의 국적·나이·성별은 확인할 수 없다.",
        "hard_violations": [
         "장부의 글자와 숫자 및 인장 문양이 판독 가능한 수준으로 드러나 읽을 수 있는 글씨 금지 조건을 위반한다."
        ],
        "physics": "도장은 손이 손잡이를 확실히 잡아 지지하므로 공중에 있어도 물리적으로 가능하다. 장부와 주판, 인주통은 책상에 놓여 있다. 다만 도장 밑면이 장부에서 떨어져 있어 종이를 강하게 짓누르는 압력이나 접촉은 보이지 않는다."
       }
      ],
      "all_candidates_fail": true
     },
     "normalized": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 2,
        "verdict_ko": "더 밀착된 클로즈업이지만, 커진 원형 도장의 인면이 카메라에 노출되고 장부를 평평하게 누르지 않으며, 읽히는 숫자와 여분 도장까지 보인다."
       },
       {
        "label": "A",
        "score": 3,
        "verdict_ko": "직사각형 도장의 형태와 크기는 참조에 더 가깝지만, 장부를 누르는 순간이 아니라 들어 올린 순간이며 읽을 수 있는 글자와 숫자도 금지 조건을 위반한다."
       }
      ],
      "all_candidates_fail": true,
      "readings": [
       {
        "label": "B",
        "direction": "손이 도장 손잡이를 잡고 장부 쪽으로 기울이고 있지만, 붉은 인면 대부분은 종이가 아니라 카메라를 향한다. 아래쪽 가장자리만 종이에 닿는 듯하여 인면을 감춘 채 강하게 짓누르는 방향과 맞지 않는다. 얼굴이나 시선은 없다.",
        "built_space": "나무 책상 위에 펼친 장부 하나가 전경을 채우고, 왼쪽에 주판 하나, 뒤쪽에 둥근 인주통 하나와 등잔 하나, 오른쪽 뒤에 종이 더미가 있다. 책상 상판과 앞쪽 측면도 보이며 따뜻하고 어두운 조명은 참조와 이어진다. 사용 중인 원형 도장 외에 뒤쪽에 검은 손잡이의 별도 도장이 보인다.",
        "entities": "장부의 붉은 취소선, 주판, 붉은 인주와 낡은 목재 표면은 요구와 맞는다. 사용 중인 도장은 참조의 작고 검은 직사각형 도장과 달리 큰 원형 목제 도장이다. 장부에는 ‘104’처럼 판독 가능한 숫자가 있다. 인물은 손만 보이므로 나이·성별·한국인 여부를 확정할 수 없고, 얼굴이나 몸통은 없다.",
        "hard_violations": [
         "장부에 판독 가능한 숫자가 노출되어 읽을 수 있는 글씨 금지 조건을 위반한다.",
         "사용 중인 도장 외에 배경에 별도의 검은 손잡이 도장이 남아 있어 도장이 중복된다."
        ],
        "physics": "도장은 손가락이 손잡이를 감싸 지지하고, 장부는 책상에 놓여 있다. 주판과 인주통, 등잔도 책상에 지지된다. 부유하거나 지지 없는 물체는 없지만, 비스듬한 도장의 가장자리 접촉으로는 인면 전체를 종이에 강하게 압착하는 동작이 성립하지 않는다."
       },
       {
        "label": "A",
        "direction": "손이 잡은 도장의 인면은 대체로 장부 아래쪽을 향하고, 카메라에는 주로 측면이 보인다. 그러나 도장 아래로 이미 찍힌 붉은 인장이 드러나고 종이와 간격이 있어, 목표 지점을 누르는 순간보다 찍은 뒤 들어 올리는 순간으로 읽힌다. 얼굴이나 시선은 없다.",
        "built_space": "펼친 장부 하나가 책상 중앙에 놓이고, 주판 하나가 왼쪽 뒤, 둥근 인주통 하나가 위쪽에 있다. 오른쪽 뒤에는 종이 더미가 있으며 책상 상판과 오른쪽 앞 측면이 보인다. 참조의 낡은 나무 책상과 절제된 따뜻한 조명은 유지되지만, 접촉점보다 장부 전체를 많이 보여 주어 요구한 클로즈업의 집중도는 낮다.",
        "entities": "검은 손잡이와 직사각형 밑판의 도장은 참조의 형태와 비교적 가깝고 손에 비례한 크기도 자연스럽다. 장부에는 붉은 취소선과 직사각형 인장이 있고 주판과 인주도 남아 있다. 다만 ‘100’ 등의 숫자와 여러 글자가 읽힌다. 손과 소매만 보이고 얼굴이나 몸통은 없으며, 손만으로 인물의 국적·나이·성별은 확인할 수 없다.",
        "hard_violations": [
         "장부의 글자와 숫자 및 인장 문양이 판독 가능한 수준으로 드러나 읽을 수 있는 글씨 금지 조건을 위반한다."
        ],
        "physics": "도장은 손이 손잡이를 확실히 잡아 지지하므로 공중에 있어도 물리적으로 가능하다. 장부와 주판, 인주통은 책상에 놓여 있다. 다만 도장 밑면이 장부에서 떨어져 있어 종이를 강하게 짓누르는 압력이나 접촉은 보이지 않는다."
       }
      ]
     },
     "ok": true
    }
   ],
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "slot_winner_match": false,
   "slot_winner": {
    "gemini-pro": "B",
    "gpt-high": "A"
   },
   "route": "cross_slot_combined"
  },
  "dual": {
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "normalized": {
    "A": 1.75,
    "B": 1.667
   },
   "adjusted": {
    "A": 1.5,
    "B": 1.417
   },
   "violations": {
    "A": [
     "[gemini-pro] 명확히 읽을 수 있는 텍스트 등장",
     "[gemini-pro] 손과 팔 외의 신체 부위(턱, 어깨) 등장",
     "[gpt-high] 장부의 글자와 숫자 및 인장 문양이 판독 가능한 수준으로 드러나 읽을 수 있는 글씨 금지 조건을 위반한다."
    ],
    "B": [
     "[gpt-high] 장부에 판독 가능한 숫자가 노출되어 읽을 수 있는 글씨 금지 조건을 위반한다.",
     "[gpt-high] 사용 중인 도장 외에 배경에 별도의 검은 손잡이 도장이 남아 있어 도장이 중복된다."
    ]
   },
   "per_model_winner": {
    "gemini-pro": "B",
    "gpt-high": "A"
   },
   "agreed": false
  },
  "totals": {
   "A": 1500,
   "B": 1417
  },
  "selected": "A",
  "ranking": [
   "A",
   "B"
  ],
  "verdicts": [
   {
    "label": "A",
    "score": 1500,
    "verdict_ko": "도장을 짓누르지 않고 도장면이 노출되었으며, 금지된 텍스트와 신체 부위가 프레임에 들어와 심각한 오류가 발생함.  ★위반: [gemini-pro] 명확히 읽을 수 있는 텍스트 등장 / [gemini-pro] 손과 팔 외의 신체 부위(턱, 어깨) 등장 / [gpt-high] 장부의 글자와 숫자 및 인장 문양이 판독 가능한 수준으로 드러나 읽을 수 있는 글씨 금지 조건을 위반한다."
   },
   {
    "label": "B",
    "score": 1417,
    "verdict_ko": "금지 조항(텍스트, 인물 배제)은 지켰으나, 도장을 종이에 밀착해 짓누르는 핵심 묘사에 실패함.  ★위반: [gpt-high] 장부에 판독 가능한 숫자가 노출되어 읽을 수 있는 글씨 금지 조건을 위반한다. / [gpt-high] 사용 중인 도장 외에 배경에 별도의 검은 손잡이 도장이 남아 있어 도장이 중복된다."
   }
  ],
  "refs": [
   {
    "label": "PREVIOUS SHOT STILL — a visually related earlier shot of this same place: the location's look, materials, fixed features and lighting mood are LOCKED to this photo; never copy its camera framing. The people visible in that photo are NOT in this shot: never carry their faces, bodies or clothing onto anyone here. Each person in THIS shot is defined solely by the PEOPLE list, their own CHARACTER REFERENCE images and their wardrobe notes. If this photo shows a character who cannot move (dead or unconscious), that character's exact pose, position and orientation are ALSO LOCKED — treat their body as a fixed prop of the set that only the camera moves around.",
    "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/db00517b-165b-4005-a287-f28936b59b95/images/272bd7af-9294-4490-98c6-88a40e9deacc/scene/recipe/S25sh8_sel.png",
    "asset_id": "4e3c699b-e983-48eb-b51b-4981d02a8a63",
    "role": "prev_still"
   }
  ],
  "critique_skipped": true,
  "needs_reshoot": true,
  "shot_run_produced": false,
  "shot_run_uid": "06aa1422-fbf5-758a-b1f8-e2ee8f5daf77",
  "ref_mode": "prev만 (배경 전용·공유 계획)",
  "share_plan": {
   "ref_plan": "prev",
   "prev_anchor_tag": "S25sh8"
  },
  "lane_policy": "share_plan_prev_bgonly"
 },
 "S25sh13::cine": {
  "applied": true,
  "attempted_at": "2026-09-09T03:47:55.238285+00:00",
  "fingerprint": "ce14936d1bbadff3f353710a2bedf361773207f1513efa5ad8cb281122892877",
  "fingerprint_version": 2,
  "provider": "mai",
  "endpoint": "openrouter/chat-completions",
  "model": "microsoft/mai-image-2.6",
  "pack": "24.202608252115",
  "source_file": "S25sh13_sel.png",
  "source_sha256": "22ac0c4a3d566858e774c01dfdcbe575f7d863e2f3b4c80df21b193cfa8fea5b",
  "file": "S25sh13_cine.png",
  "staged_sha256": "17a5564d9f1f031a35349dc936d8a781b7854636da8c79bf4c5f0c2dc054713f",
  "latency_ms": 29584
 },
 "S26sh2::signage": {
  "fp": "e80f0c086fa1ee4e",
  "inscriptions": [],
  "cues": [],
  "dropped": []
 },
 "era_assess::d07a27cc7bd9c865": {
  "subjects": [],
  "subject_text": "파청 산채 공터·훈련장·취사장·막사 주변\n편백숲 속 흙 공터를 중심으로 목책과 움막이 모인 산채. 훈련할 빈터와 가마솥 취사장, 곡식 보관 자리가 나뉘어 있다.",
  "identity": "canonical",
  "scope_id": "L23",
  "scope_role": "location_exterior",
  "scope_sha": "814ed87bef5f0cab"
 },
 "S26sh2::bgfirst_bg": {
  "input_fingerprint": "dc83e44a317c0464",
  "prompt": "Create the EMPTY BACKGROUND PLATE for one film shot — NO PEOPLE, no figures, no body parts, no sketch lines, no arrows anywhere.\n\"Empty\" means no people only: KEEP the location's inherent occupants and stock that define the place — animals in an animal shelter, pen or farm, goods in a market, moored boats in a harbour — unless the shot text explicitly removes them.\nThe FIRST attached image is a thin-line storyboard sketch: use ONLY its camera angle, horizon, perspective and the placement/size of buildings and set masses — ignore the sketched people and arrows entirely. The SECOND attached image (LOCATION PHOTOGRAPH) is the real place: take its architecture, materials, signage and fixed features, and RE-PROJECT them into the sketch's camera. If the photograph's camera differs from the sketch's, the sketch's camera wins.\nHUMAN-SCALE CALIBRATION: derive every structure's true size from human-scale elements — a door ≈ 2m, a window ≈ 1–1.5m wide, one storey ≈ 2.5–3m; never inflate a small structure or shrink a large one.\n\nSHOT TEXT this background must serve (Korean): 공터 맞은편, 낡은 봇짐과 농기구를 손에 쥔 채 주뼛거리며 서 있는 대여섯 명의 새로 온 조선인 남성들 전경.\n\nLOCATION (lock): The newcomers' side of an open mountain-camp clearing, opposite the established members in daylight.\n\nTIME OF DAY (lock): day.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: wide shot\n- KEY BACKGROUND ELEMENTS: 산채 공터 (An open gap separates the newcomers from the established men); used as The unoccupied lower-left area makes the awkward social distance visible; 봇짐 (Held by the newly arrived men); used as Small, separately positioned belongings distinguish the arrivals without creating repeated silhouettes; 낡은 농기구 (Worn farming tools carried by the newcomers) — The tools follow different natural carrying angles beside their holders; used as Readable within the full-body composition as evidence of the men's civilian lives.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Natural daytime illumination, muted saturation and restrained contrast allow tentative human warmth without romanticizing the newcomers.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nRender ONE photorealistic empty location photograph, 16:9, that this shot can be staged inside later. No readable writing anywhere: surfaces that would carry writing may be present, but stage any wording out of legibility — an oblique angle, distance, shallow focus. No captions, watermarks or overlay text.",
  "effective_prompt": "Create the EMPTY BACKGROUND PLATE for one film shot — NO PEOPLE, no figures, no body parts, no sketch lines, no arrows anywhere.\n\"Empty\" means no people only: KEEP the location's inherent occupants and stock that define the place — animals in an animal shelter, pen or farm, goods in a market, moored boats in a harbour — unless the shot text explicitly removes them.\nThe FIRST attached image is a thin-line storyboard sketch: use ONLY its camera angle, horizon, perspective and the placement/size of buildings and set masses — ignore the sketched people and arrows entirely. The SECOND attached image (LOCATION PHOTOGRAPH) is the real place: take its architecture, materials, signage and fixed features, and RE-PROJECT them into the sketch's camera. If the photograph's camera differs from the sketch's, the sketch's camera wins.\nHUMAN-SCALE CALIBRATION: derive every structure's true size from human-scale elements — a door ≈ 2m, a window ≈ 1–1.5m wide, one storey ≈ 2.5–3m; never inflate a small structure or shrink a large one.\n\nSHOT TEXT this background must serve (Korean): 공터 맞은편, 낡은 봇짐과 농기구를 손에 쥔 채 주뼛거리며 서 있는 대여섯 명의 새로 온 조선인 남성들 전경.\n\nLOCATION (lock): The newcomers' side of an open mountain-camp clearing, opposite the established members in daylight.\n\nTIME OF DAY (lock): day.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: wide shot\n- KEY BACKGROUND ELEMENTS: 산채 공터 (An open gap separates the newcomers from the established men); used as The unoccupied lower-left area makes the awkward social distance visible; 봇짐 (Held by the newly arrived men); used as Small, separately positioned belongings distinguish the arrivals without creating repeated silhouettes; 낡은 농기구 (Worn farming tools carried by the newcomers) — The tools follow different natural carrying angles beside their holders; used as Readable within the full-body composition as evidence of the men's civilian lives.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Natural daytime illumination, muted saturation and restrained contrast allow tentative human warmth without romanticizing the newcomers.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nRender ONE photorealistic empty location photograph, 16:9, that this shot can be staged inside later. No readable writing anywhere: surfaces that would carry writing may be present, but stage any wording out of legibility — an oblique angle, distance, shallow focus. No captions, watermarks or overlay text.",
  "bg_path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/db00517b-165b-4005-a287-f28936b59b95/images/272bd7af-9294-4490-98c6-88a40e9deacc/scene/recipe/S26sh2__bgfirst_bg.png",
  "asset_id": "0b28903d-d699-4da7-9123-e64d16e75375",
  "input_asset_ids": [
   "829edf95-8348-46f0-9c37-4fcfa7339808",
   "8c17c7ae-efff-4bb5-b88c-7c5528f00716"
  ]
 },
 "S26sh2": {
  "input_fingerprint": "7578576140220295",
  "prompt": "Create ONE FINAL photorealistic live-action film still of the moment below — early 20th-century Korea, Jeolla region; all people are Korean unless explicitly stated otherwise. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 공터 맞은편, 낡은 봇짐과 농기구를 손에 쥔 채 주뼛거리며 서 있는 대여섯 명의 새로 온 조선인 남성들 전경.\n\nLOCATION (lock): The newcomers' side of an open mountain-camp clearing, opposite the established members in daylight. The shot takes place here — the attached LOCATION STRUCTURE PHOTOGRAPH is the single authority for this exact place — its fixed structure and permanent site details are LOCKED to it. No separate location photograph exists for this place. Build everything else strictly from the location text above and the shot text; the layout sketch (when attached) governs framing and placement only, and the shot text governs time of day, lighting and action.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: wide shot\n- KEY BACKGROUND ELEMENTS: 산채 공터 (An open gap separates the newcomers from the established men); used as The unoccupied lower-left area makes the awkward social distance visible; 봇짐 (Held by the newly arrived men); used as Small, separately positioned belongings distinguish the arrivals without creating repeated silhouettes; 낡은 농기구 (Worn farming tools carried by the newcomers) — The tools follow different natural carrying angles beside their holders; used as Readable within the full-body composition as evidence of the men's civilian lives.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Natural daytime illumination, muted saturation and restrained contrast allow tentative human warmth without romanticizing the newcomers.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): Daylight fills the mountain-camp clearing. 오주일: He stands among the newcomers in rough clothes with the removal mark on his worn collar. 임정현: He stands at the very back of the newcomers. 대장장이: He stands among the newcomers with visibly callused hands. 목수: He stands among the newcomers. 백정: He stands among the newcomers with tense shoulders.\n\nTHE HAND THAT IS DOING THIS: the object at the centre of this shot is being held, operated, read, aimed or handed over by 대여섯 명의 새로 온 조선인 남성들 right now, so 대여섯 명의 새로 온 조선인 남성들's hand — and as much of the wrist and forearm as the framing reaches — is in the frame, gripping or resting on that object exactly the way its real use requires. Match that hand to 대여섯 명의 새로 온 조선인 남성들: its size, build, skin, age, grooming, sleeve and anything worn on it belong to that person and to no one else.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 오주일 (한국인 남성, 성인 나이대, 자연스러운 검은 머리); 임정현 (한국인 남성, 성인 나이대, 자연스러운 검은 머리); 대장장이 (한국인 남성, 성인 나이대, 자연스러운 검은 머리); 백정 (한국인 남성, 성인 나이대, 자연스러운 검은 머리); 목수 (한국인 남성, 성인 나이대, 자연스러운 검은 머리) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.",
  "roll_prompts": {
   "A": "Stage the shot. The FIRST attached image (SHOT BACKGROUND) is the finished empty background of this shot — keep it EXACTLY: its camera, perspective, architecture, lighting and every fixed feature stay untouched. The SECOND attached image (LAYOUT SKETCH) tells you ONLY where the people go: each sketched person's position, screen size, pose and the gaze/motion arrows. Ignore the sketch's background lines. The CHARACTER REFERENCE photographs show the real people.\nPlace the real people into the background at exactly the sketched positions, sizes and poses, following the arrow directions. No sketch lines or arrows may remain.\n\nCreate ONE FINAL photorealistic live-action film still of the moment below — early 20th-century Korea, Jeolla region; all people are Korean unless explicitly stated otherwise. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 공터 맞은편, 낡은 봇짐과 농기구를 손에 쥔 채 주뼛거리며 서 있는 대여섯 명의 새로 온 조선인 남성들 전경.\n\nLOCATION (lock): The newcomers' side of an open mountain-camp clearing, opposite the established members in daylight. The shot takes place here — the FIRST attached image (SHOT BACKGROUND) is this exact place, already built: its ground, structures, horizon, materials and lighting are the finished truth of this location and must not be redesigned or replaced. No location photograph is attached — read the place from that image alone, and add no scenery, structure, vehicle or fixture that it does not already show. This lock governs the place only; the figures in the shot follow the staging and pose instructions.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: wide shot\n- KEY BACKGROUND ELEMENTS: 산채 공터 (An open gap separates the newcomers from the established men); used as The unoccupied lower-left area makes the awkward social distance visible; 봇짐 (Held by the newly arrived men); used as Small, separately positioned belongings distinguish the arrivals without creating repeated silhouettes; 낡은 농기구 (Worn farming tools carried by the newcomers) — The tools follow different natural carrying angles beside their holders; used as Readable within the full-body composition as evidence of the men's civilian lives.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Natural daytime illumination, muted saturation and restrained contrast allow tentative human warmth without romanticizing the newcomers.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): Daylight fills the mountain-camp clearing. 오주일: He stands among the newcomers in rough clothes with the removal mark on his worn collar. 임정현: He stands at the very back of the newcomers. 대장장이: He stands among the newcomers with visibly callused hands. 목수: He stands among the newcomers. 백정: He stands among the newcomers with tense shoulders.\n\nTHE HAND THAT IS DOING THIS: the object at the centre of this shot is being held, operated, read, aimed or handed over by 대여섯 명의 새로 온 조선인 남성들 right now, so 대여섯 명의 새로 온 조선인 남성들's hand — and as much of the wrist and forearm as the framing reaches — is in the frame, gripping or resting on that object exactly the way its real use requires. Match that hand to 대여섯 명의 새로 온 조선인 남성들: its size, build, skin, age, grooming, sleeve and anything worn on it belong to that person and to no one else.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 오주일 (한국인 남성, 성인 나이대, 자연스러운 검은 머리); 임정현 (한국인 남성, 성인 나이대, 자연스러운 검은 머리); 대장장이 (한국인 남성, 성인 나이대, 자연스러운 검은 머리); 백정 (한국인 남성, 성인 나이대, 자연스러운 검은 머리); 목수 (한국인 남성, 성인 나이대, 자연스러운 검은 머리) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.",
   "B": "Create ONE FINAL photorealistic live-action film still of the moment below — early 20th-century Korea, Jeolla region; all people are Korean unless explicitly stated otherwise. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 공터 맞은편, 낡은 봇짐과 농기구를 손에 쥔 채 주뼛거리며 서 있는 대여섯 명의 새로 온 조선인 남성들 전경.\n\nLOCATION (lock): The newcomers' side of an open mountain-camp clearing, opposite the established members in daylight. The shot takes place here — the attached LOCATION STRUCTURE PHOTOGRAPH is the single authority for this exact place — its fixed structure and permanent site details are LOCKED to it. No separate location photograph exists for this place. Build everything else strictly from the location text above and the shot text; the layout sketch (when attached) governs framing and placement only, and the shot text governs time of day, lighting and action.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: wide shot\n- KEY BACKGROUND ELEMENTS: 산채 공터 (An open gap separates the newcomers from the established men); used as The unoccupied lower-left area makes the awkward social distance visible; 봇짐 (Held by the newly arrived men); used as Small, separately positioned belongings distinguish the arrivals without creating repeated silhouettes; 낡은 농기구 (Worn farming tools carried by the newcomers) — The tools follow different natural carrying angles beside their holders; used as Readable within the full-body composition as evidence of the men's civilian lives.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Natural daytime illumination, muted saturation and restrained contrast allow tentative human warmth without romanticizing the newcomers.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): Daylight fills the mountain-camp clearing. 오주일: He stands among the newcomers in rough clothes with the removal mark on his worn collar. 임정현: He stands at the very back of the newcomers. 대장장이: He stands among the newcomers with visibly callused hands. 목수: He stands among the newcomers. 백정: He stands among the newcomers with tense shoulders.\n\nTHE HAND THAT IS DOING THIS: the object at the centre of this shot is being held, operated, read, aimed or handed over by 대여섯 명의 새로 온 조선인 남성들 right now, so 대여섯 명의 새로 온 조선인 남성들's hand — and as much of the wrist and forearm as the framing reaches — is in the frame, gripping or resting on that object exactly the way its real use requires. Match that hand to 대여섯 명의 새로 온 조선인 남성들: its size, build, skin, age, grooming, sleeve and anything worn on it belong to that person and to no one else.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 오주일 (한국인 남성, 성인 나이대, 자연스러운 검은 머리); 임정현 (한국인 남성, 성인 나이대, 자연스러운 검은 머리); 대장장이 (한국인 남성, 성인 나이대, 자연스러운 검은 머리); 백정 (한국인 남성, 성인 나이대, 자연스러운 검은 머리); 목수 (한국인 남성, 성인 나이대, 자연스러운 검은 머리) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.\n\nCOMPOSITION VARIATION (this candidate only): keep the place, the people and the moment exactly as specified, and keep from the CAMERA & FRAME contract its subject scale, screen placement and key background elements exactly. For this candidate only, its camera angle and height are deliberately left open: choose a camera position distinctly different from the obvious one, as a film director picking a second setup on the same blocking. Stay true to any explicit framing the SHOT TEXT itself dictates."
  },
  "roll_refs": {
   "A": [
    {
     "label": "SHOT BACKGROUND",
     "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/db00517b-165b-4005-a287-f28936b59b95/images/272bd7af-9294-4490-98c6-88a40e9deacc/scene/recipe/S26sh2__bgfirst_bg.png",
     "asset_id": "0b28903d-d699-4da7-9123-e64d16e75375",
     "role": "bgfirst_bg"
    },
    {
     "label": "LAYOUT SKETCH (people placement only)",
     "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/db00517b-165b-4005-a287-f28936b59b95/images/272bd7af-9294-4490-98c6-88a40e9deacc/conti/conti_S26sh2.png",
     "asset_id": "829edf95-8348-46f0-9c37-4fcfa7339808",
     "role": "conti_light"
    },
    {
     "label": "CHARACTER REFERENCE — 오주일: the exact person appearing in this shot; match face, hair and build exactly.",
     "path": "<bytes:809716>",
     "asset_id": "7c31e825-ceef-4208-915f-71c4f0e15dd6",
     "role": "character_ref"
    },
    {
     "label": "CHARACTER REFERENCE — 임정현: the exact person appearing in this shot; match face, hair and build exactly.",
     "path": "<bytes:858188>",
     "asset_id": "789c3359-d7f8-4832-9a6f-7e4ad2695c02",
     "role": "character_ref"
    },
    {
     "label": "CHARACTER REFERENCE — 대장장이: the exact person appearing in this shot; match face, hair and build exactly.",
     "path": "<bytes:946438>",
     "asset_id": "a12f8b90-e799-4753-9b7e-e71f4d817db5",
     "role": "character_ref"
    },
    {
     "label": "CHARACTER REFERENCE — 백정: the exact person appearing in this shot; match face, hair and build exactly.",
     "path": "<bytes:980360>",
     "asset_id": "29282584-ae7c-4722-8d77-100c4a5b2a3d",
     "role": "character_ref"
    },
    {
     "label": "CHARACTER REFERENCE — 목수: the exact person appearing in this shot; match face, hair and build exactly.",
     "path": "<bytes:904990>",
     "asset_id": "6b02463a-e818-4e3d-9816-215435a7aa0e",
     "role": "character_ref"
    },
    {
     "label": "PROP REFERENCE — 보자기 꾸러미·봇짐: the exact object appearing in this shot; match its look, material and wear exactly.",
     "path": "<bytes:1465734>",
     "asset_id": "3ca92b24-68b4-4883-b851-cabc06c00d1a",
     "role": "prop_ref"
    }
   ],
   "B": [
    {
     "label": "LOCATION STRUCTURE PHOTOGRAPH — the confirmed photograph of this exact place and its fixed structure: it is the SINGLE authority for the location, the structure's shape, proportions, materials, colors, openings and every permanent site detail. Never copy its camera framing, time of day or lighting — the shot text is the authority for those.",
     "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/db00517b-165b-4005-a287-f28936b59b95/images/background_chain/seed_bg_mountain_camp_sel.png",
     "asset_id": "8c17c7ae-efff-4bb5-b88c-7c5528f00716",
     "role": "location_seed_bg"
    },
    {
     "label": "CHARACTER REFERENCE — 오주일: the exact person appearing in this shot; match face, hair and build exactly.",
     "path": "<bytes:809716>",
     "asset_id": "7c31e825-ceef-4208-915f-71c4f0e15dd6",
     "role": "character_ref"
    },
    {
     "label": "CHARACTER REFERENCE — 임정현: the exact person appearing in this shot; match face, hair and build exactly.",
     "path": "<bytes:858188>",
     "asset_id": "789c3359-d7f8-4832-9a6f-7e4ad2695c02",
     "role": "character_ref"
    },
    {
     "label": "CHARACTER REFERENCE — 대장장이: the exact person appearing in this shot; match face, hair and build exactly.",
     "path": "<bytes:946438>",
     "asset_id": "a12f8b90-e799-4753-9b7e-e71f4d817db5",
     "role": "character_ref"
    },
    {
     "label": "CHARACTER REFERENCE — 백정: the exact person appearing in this shot; match face, hair and build exactly.",
     "path": "<bytes:980360>",
     "asset_id": "29282584-ae7c-4722-8d77-100c4a5b2a3d",
     "role": "character_ref"
    },
    {
     "label": "CHARACTER REFERENCE — 목수: the exact person appearing in this shot; match face, hair and build exactly.",
     "path": "<bytes:904990>",
     "asset_id": "6b02463a-e818-4e3d-9816-215435a7aa0e",
     "role": "character_ref"
    },
    {
     "label": "PROP REFERENCE — 보자기 꾸러미·봇짐: the exact object appearing in this shot; match its look, material and wear exactly.",
     "path": "<bytes:1465734>",
     "asset_id": "3ca92b24-68b4-4883-b851-cabc06c00d1a",
     "role": "prop_ref"
    }
   ]
  },
  "shot_run_spend_attempt_count": 1,
  "judge_flip_skipped": "judge_fn_owns_order",
  "readings": [
   {
    "label": "A",
    "direction": "다섯 명의 남성들이 왼쪽의 산채 건물과 빈 공터를 향해 시선을 두고 나란히 서 있다.",
    "built_space": "레퍼런스 사진과 동일한 초가집 두 채, 바위, 둥근 나무 울타리가 정확한 구조로 배치되어 있으며, 인물들은 공터 우측 가장자리에 알맞게 서 있다.",
    "entities": "5명의 남성이 등장하며, 흰색 한복과 검은 모자, 파란 옷과 밀짚모자 등 레퍼런스의 복장을 잘 반영하고 있다. 여러 인물이 레퍼런스의 초록색 봇짐을 메고 있으며 다양한 농기구(톱, 낫, 곡괭이)를 들고 있다.",
    "hard_violations": [],
    "physics": "인물들은 흙바닥에 안정적으로 발을 딛고 서 있으며, 농기구는 손에 쥐어져 있거나 땅에 자연스럽게 지탱되어 있다. 물리적 오류 없음."
   },
   {
    "label": "B",
    "direction": "다섯 명의 남성들이 정면(카메라 방향)을 향해 평면적으로 서 있다.",
    "built_space": "배경에 나무 울타리와 우측에 초가집 일부가 보이나, 레퍼런스 사진 특유의 원형 공터와 건물 배치 구도와는 일치하지 않는다.",
    "entities": "5명의 남성이 있으나 캐릭터 레퍼런스(동일한 얼굴)와 전혀 다른 새로운 인물들로 구성되었다. 복장도 일치하지 않으며, 지정된 초록색 봇짐 대신 일반적인 마대 자루를 들고 있다.",
    "hard_violations": [
     "[gemini-pro] invented people (레퍼런스와 일치하지 않는 새로운 얼굴의 인물들 생성)"
    ],
    "physics": "인물들은 바닥에 서 있고 손으로 자루와 농기구를 잘 쥐고 있으며, 중력과 그림자 표현에 어색함은 없다."
   }
  ],
  "cross_model_order": {
   "policy": "cross_model_order_v2_slot0_fwd_slot1_rev_parallel2",
   "slots": [
    {
     "model": "gemini-pro",
     "order": "forward",
     "display_to_canonical": {
      "A": "A",
      "B": "B"
     },
     "raw": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 9,
        "verdict_ko": "레퍼런스의 장소 구조와 와이드 샷 구도를 완벽히 구현했으며, 지정된 인물의 복장 특징과 초록색 봇짐, 농기구 배치를 충실히 반영함."
       },
       {
        "label": "B",
        "score": 3,
        "verdict_ko": "지정된 얼굴 레퍼런스와 전혀 다른 인물들을 생성했으며, 지정된 소품(초록색 봇짐)을 누락하고 장소의 전체 구조도 맞지 않음."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "다섯 명의 남성들이 왼쪽의 산채 건물과 빈 공터를 향해 시선을 두고 나란히 서 있다.",
        "built_space": "레퍼런스 사진과 동일한 초가집 두 채, 바위, 둥근 나무 울타리가 정확한 구조로 배치되어 있으며, 인물들은 공터 우측 가장자리에 알맞게 서 있다.",
        "entities": "5명의 남성이 등장하며, 흰색 한복과 검은 모자, 파란 옷과 밀짚모자 등 레퍼런스의 복장을 잘 반영하고 있다. 여러 인물이 레퍼런스의 초록색 봇짐을 메고 있으며 다양한 농기구(톱, 낫, 곡괭이)를 들고 있다.",
        "hard_violations": [],
        "physics": "인물들은 흙바닥에 안정적으로 발을 딛고 서 있으며, 농기구는 손에 쥐어져 있거나 땅에 자연스럽게 지탱되어 있다. 물리적 오류 없음."
       },
       {
        "label": "B",
        "direction": "다섯 명의 남성들이 정면(카메라 방향)을 향해 평면적으로 서 있다.",
        "built_space": "배경에 나무 울타리와 우측에 초가집 일부가 보이나, 레퍼런스 사진 특유의 원형 공터와 건물 배치 구도와는 일치하지 않는다.",
        "entities": "5명의 남성이 있으나 캐릭터 레퍼런스(동일한 얼굴)와 전혀 다른 새로운 인물들로 구성되었다. 복장도 일치하지 않으며, 지정된 초록색 봇짐 대신 일반적인 마대 자루를 들고 있다.",
        "hard_violations": [
         "invented people (레퍼런스와 일치하지 않는 새로운 얼굴의 인물들 생성)"
        ],
        "physics": "인물들은 바닥에 서 있고 손으로 자루와 농기구를 잘 쥐고 있으며, 중력과 그림자 표현에 어색함은 없다."
       }
      ],
      "all_candidates_fail": false
     },
     "normalized": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 9,
        "verdict_ko": "레퍼런스의 장소 구조와 와이드 샷 구도를 완벽히 구현했으며, 지정된 인물의 복장 특징과 초록색 봇짐, 농기구 배치를 충실히 반영함."
       },
       {
        "label": "B",
        "score": 3,
        "verdict_ko": "지정된 얼굴 레퍼런스와 전혀 다른 인물들을 생성했으며, 지정된 소품(초록색 봇짐)을 누락하고 장소의 전체 구조도 맞지 않음."
       }
      ],
      "all_candidates_fail": false,
      "readings": [
       {
        "label": "A",
        "direction": "다섯 명의 남성들이 왼쪽의 산채 건물과 빈 공터를 향해 시선을 두고 나란히 서 있다.",
        "built_space": "레퍼런스 사진과 동일한 초가집 두 채, 바위, 둥근 나무 울타리가 정확한 구조로 배치되어 있으며, 인물들은 공터 우측 가장자리에 알맞게 서 있다.",
        "entities": "5명의 남성이 등장하며, 흰색 한복과 검은 모자, 파란 옷과 밀짚모자 등 레퍼런스의 복장을 잘 반영하고 있다. 여러 인물이 레퍼런스의 초록색 봇짐을 메고 있으며 다양한 농기구(톱, 낫, 곡괭이)를 들고 있다.",
        "hard_violations": [],
        "physics": "인물들은 흙바닥에 안정적으로 발을 딛고 서 있으며, 농기구는 손에 쥐어져 있거나 땅에 자연스럽게 지탱되어 있다. 물리적 오류 없음."
       },
       {
        "label": "B",
        "direction": "다섯 명의 남성들이 정면(카메라 방향)을 향해 평면적으로 서 있다.",
        "built_space": "배경에 나무 울타리와 우측에 초가집 일부가 보이나, 레퍼런스 사진 특유의 원형 공터와 건물 배치 구도와는 일치하지 않는다.",
        "entities": "5명의 남성이 있으나 캐릭터 레퍼런스(동일한 얼굴)와 전혀 다른 새로운 인물들로 구성되었다. 복장도 일치하지 않으며, 지정된 초록색 봇짐 대신 일반적인 마대 자루를 들고 있다.",
        "hard_violations": [
         "invented people (레퍼런스와 일치하지 않는 새로운 얼굴의 인물들 생성)"
        ],
        "physics": "인물들은 바닥에 서 있고 손으로 자루와 농기구를 잘 쥐고 있으며, 중력과 그림자 표현에 어색함은 없다."
       }
      ]
     },
     "ok": true
    },
    {
     "model": "gpt-high",
     "order": "reverse",
     "display_to_canonical": {
      "A": "B",
      "B": "A"
     },
     "raw": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 6,
        "verdict_ko": "다섯 남성의 전신과 왼쪽 아래 공백, 주뼛거리는 정지는 잘 구현했지만 인물별 외형·복장과 봇짐, 산채의 고정 배치가 참조보다 덜 충실하다."
       },
       {
        "label": "B",
        "score": 8,
        "verdict_ko": "산채의 두 초가와 바위·목책 배치, 맞은편을 보는 다섯 남성과 맨 뒤 임정현이 더 충실하지만, 내려놓은 봇짐과 불명확한 오주일 정체성은 감점이다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "다섯 남성은 대체로 화면 왼쪽의 공터 너머, 화면 밖 상대가 있을 방향을 바라본다. 렌즈를 일제히 응시하지 않는다. 긴 농기구는 거의 수직으로 내려가 날이 땅에 닿고, 가운데 남성의 짧은 도끼는 몸 옆에서 왼쪽 아래를 향한다. 도구로 누군가를 겨누거나 이동하는 사람은 없다.",
        "built_space": "남성 다섯 명이 화면 오른쪽 흙바닥에 서 있고 왼쪽 아래는 넓게 비어 있다. 오른쪽 뒤에 초가지붕 한 구간과 어두운 출입구, 뒤쪽 비탈을 따라 이어지는 목책, 중앙의 큰 바위 무리가 보인다. 참조의 인접한 두 초가와 그 앞 바위 배치는 이 화면에서 명확히 재현되지 않으며, 배경은 비탈과 목책 중심으로 바뀌어 보인다. 다만 화면 밖 건물을 누락으로 단정할 수는 없다. 반사면이나 중복된 고정 시설은 보이지 않는다.",
        "entities": "한국인 성인 남성 설정에 부합하는 외형의 남성 다섯 명이며 검거나 짙은 머리다. 그러나 모두 회갈색 계열 옷을 입어 참조의 청색 작업복 대장장이, 밀짚모자 목수, 검은 옷과 머릿수건의 백정, 밝은 긴 도포의 임정현을 개별적으로 식별하기 어렵다. 맨 뒤 남성도 임정현의 참조 복장과 다르다. 오주일의 옷깃 제거 흔적과 대장장이의 굳은살은 확인되지 않는다. 낡은 천 봇짐 네 개와 농기구 두 개가 보이지만 봇짐은 모두 갈색·회색 계열로, 참조의 청록색 천과 꼰 끈은 재현되지 않았다. 낮의 자연광이며 읽을 수 있는 글자는 없다.",
        "hard_violations": [],
        "physics": "다섯 사람 모두 발이 흙바닥에 닿아 체중을 지탱한다. 왼쪽 남성의 봇짐은 어깨에 걸려 있고, 가운데 작은 봇짐은 손에 잡혀 있으며, 오른쪽 두 꾸러미는 손과 팔로 받쳐진다. 긴 농기구는 손과 바닥이 지지하고 짧은 도끼는 손이 자루를 잡고 있다. 지지 없이 뜬 물체는 없다. 다만 왼쪽 남성은 봇짐을 손에 쥐지 않고 어깨에 멘 상태다."
       },
       {
        "label": "B",
        "direction": "다섯 남성의 얼굴과 몸이 화면 왼쪽 안쪽의 빈 공터와 초가 쪽을 향하며, 화면 밖 기존 단원들을 마주하는 방향으로 읽힌다. 긴 자루들은 땅을 향해 비스듬히 내려가고, 한 농기구는 어깨 위에 얹혀 머리 뒤쪽으로 뻗는다. 대장장이로 보이는 남성의 낫은 손 아래에서 안쪽으로 굽고, 백정의 곡괭이는 왼쪽 아래를 향한다. 도구들의 운반 각도가 서로 다르며 공격 동작은 아니다.",
        "built_space": "큰 초가 한 채와 그 오른쪽 작은 초가 한 채, 앞쪽 돌단, 큰 초가 왼쪽의 바위 무리, 뒤쪽 목책 가까이의 둥근 바위 한 개가 보인다. 목책은 공터 뒤와 오른쪽 가장자리를 따라 이어지고 그 너머 산과 계곡이 펼쳐진다. 참조의 주요 건물 수와 상대 배치를 더 잘 유지한다. 남성 다섯 명은 오른쪽 공터 가장자리에 있고 왼쪽 아래와 건물 사이가 비어 있어 사회적 거리가 드러난다. 약간 높은 사선 시점은 허용된 자유 시점에 해당하며 불가능한 반사는 없다.",
        "entities": "한국인 성인 남성 설정에 부합하는 외형의 남성 다섯 명이 보인다. 오른쪽 맨 뒤의 밝은 긴 도포와 상투는 임정현, 그 앞의 검은 옷과 머릿수건은 백정, 짧은 검은 머리와 청색 작업복은 대장장이의 참조 특징에 가깝다. 왼쪽에는 밀짚모자를 쓴 청색 옷 남성이 두 명 있어 목수는 연상되지만 오주일을 확실히 구별하기 어렵고 옷깃 제거 흔적도 확인되지 않는다. 손의 굳은살은 이 크기에서 판독하기 어렵다. 왼쪽 남성의 청록색 둥근 봇짐과 끈은 소품 참조에 가깝다. 다른 등짐들과 바닥에 놓인 꾸러미 두 개, 여러 낡은 농기구도 보인다. 낮의 자연광과 절제된 색조이며 읽을 수 있는 글자는 없다.",
        "hard_violations": [],
        "physics": "다섯 사람 모두 양발 또는 한쪽에 실린 체중으로 땅에 서 있다. 등짐은 어깨를 지나는 끈으로 지지되고, 바닥의 꾸러미 두 개는 흙바닥에 놓여 있다. 농기구 자루는 각 인물의 손에 잡혀 있으며 일부 끝은 땅에 닿고 어깨 위 도구는 손과 어깨가 받친다. 지지 없이 공중에 뜬 몸이나 물체는 없다. 다만 봇짐 대부분을 등에 메거나 땅에 내려놓아, 손에 쥔 채 서 있다는 순간의 행동은 충분히 지키지 못했다."
       }
      ],
      "all_candidates_fail": false
     },
     "normalized": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 6,
        "verdict_ko": "다섯 남성의 전신과 왼쪽 아래 공백, 주뼛거리는 정지는 잘 구현했지만 인물별 외형·복장과 봇짐, 산채의 고정 배치가 참조보다 덜 충실하다."
       },
       {
        "label": "A",
        "score": 8,
        "verdict_ko": "산채의 두 초가와 바위·목책 배치, 맞은편을 보는 다섯 남성과 맨 뒤 임정현이 더 충실하지만, 내려놓은 봇짐과 불명확한 오주일 정체성은 감점이다."
       }
      ],
      "all_candidates_fail": false,
      "readings": [
       {
        "label": "B",
        "direction": "다섯 남성은 대체로 화면 왼쪽의 공터 너머, 화면 밖 상대가 있을 방향을 바라본다. 렌즈를 일제히 응시하지 않는다. 긴 농기구는 거의 수직으로 내려가 날이 땅에 닿고, 가운데 남성의 짧은 도끼는 몸 옆에서 왼쪽 아래를 향한다. 도구로 누군가를 겨누거나 이동하는 사람은 없다.",
        "built_space": "남성 다섯 명이 화면 오른쪽 흙바닥에 서 있고 왼쪽 아래는 넓게 비어 있다. 오른쪽 뒤에 초가지붕 한 구간과 어두운 출입구, 뒤쪽 비탈을 따라 이어지는 목책, 중앙의 큰 바위 무리가 보인다. 참조의 인접한 두 초가와 그 앞 바위 배치는 이 화면에서 명확히 재현되지 않으며, 배경은 비탈과 목책 중심으로 바뀌어 보인다. 다만 화면 밖 건물을 누락으로 단정할 수는 없다. 반사면이나 중복된 고정 시설은 보이지 않는다.",
        "entities": "한국인 성인 남성 설정에 부합하는 외형의 남성 다섯 명이며 검거나 짙은 머리다. 그러나 모두 회갈색 계열 옷을 입어 참조의 청색 작업복 대장장이, 밀짚모자 목수, 검은 옷과 머릿수건의 백정, 밝은 긴 도포의 임정현을 개별적으로 식별하기 어렵다. 맨 뒤 남성도 임정현의 참조 복장과 다르다. 오주일의 옷깃 제거 흔적과 대장장이의 굳은살은 확인되지 않는다. 낡은 천 봇짐 네 개와 농기구 두 개가 보이지만 봇짐은 모두 갈색·회색 계열로, 참조의 청록색 천과 꼰 끈은 재현되지 않았다. 낮의 자연광이며 읽을 수 있는 글자는 없다.",
        "hard_violations": [],
        "physics": "다섯 사람 모두 발이 흙바닥에 닿아 체중을 지탱한다. 왼쪽 남성의 봇짐은 어깨에 걸려 있고, 가운데 작은 봇짐은 손에 잡혀 있으며, 오른쪽 두 꾸러미는 손과 팔로 받쳐진다. 긴 농기구는 손과 바닥이 지지하고 짧은 도끼는 손이 자루를 잡고 있다. 지지 없이 뜬 물체는 없다. 다만 왼쪽 남성은 봇짐을 손에 쥐지 않고 어깨에 멘 상태다."
       },
       {
        "label": "A",
        "direction": "다섯 남성의 얼굴과 몸이 화면 왼쪽 안쪽의 빈 공터와 초가 쪽을 향하며, 화면 밖 기존 단원들을 마주하는 방향으로 읽힌다. 긴 자루들은 땅을 향해 비스듬히 내려가고, 한 농기구는 어깨 위에 얹혀 머리 뒤쪽으로 뻗는다. 대장장이로 보이는 남성의 낫은 손 아래에서 안쪽으로 굽고, 백정의 곡괭이는 왼쪽 아래를 향한다. 도구들의 운반 각도가 서로 다르며 공격 동작은 아니다.",
        "built_space": "큰 초가 한 채와 그 오른쪽 작은 초가 한 채, 앞쪽 돌단, 큰 초가 왼쪽의 바위 무리, 뒤쪽 목책 가까이의 둥근 바위 한 개가 보인다. 목책은 공터 뒤와 오른쪽 가장자리를 따라 이어지고 그 너머 산과 계곡이 펼쳐진다. 참조의 주요 건물 수와 상대 배치를 더 잘 유지한다. 남성 다섯 명은 오른쪽 공터 가장자리에 있고 왼쪽 아래와 건물 사이가 비어 있어 사회적 거리가 드러난다. 약간 높은 사선 시점은 허용된 자유 시점에 해당하며 불가능한 반사는 없다.",
        "entities": "한국인 성인 남성 설정에 부합하는 외형의 남성 다섯 명이 보인다. 오른쪽 맨 뒤의 밝은 긴 도포와 상투는 임정현, 그 앞의 검은 옷과 머릿수건은 백정, 짧은 검은 머리와 청색 작업복은 대장장이의 참조 특징에 가깝다. 왼쪽에는 밀짚모자를 쓴 청색 옷 남성이 두 명 있어 목수는 연상되지만 오주일을 확실히 구별하기 어렵고 옷깃 제거 흔적도 확인되지 않는다. 손의 굳은살은 이 크기에서 판독하기 어렵다. 왼쪽 남성의 청록색 둥근 봇짐과 끈은 소품 참조에 가깝다. 다른 등짐들과 바닥에 놓인 꾸러미 두 개, 여러 낡은 농기구도 보인다. 낮의 자연광과 절제된 색조이며 읽을 수 있는 글자는 없다.",
        "hard_violations": [],
        "physics": "다섯 사람 모두 양발 또는 한쪽에 실린 체중으로 땅에 서 있다. 등짐은 어깨를 지나는 끈으로 지지되고, 바닥의 꾸러미 두 개는 흙바닥에 놓여 있다. 농기구 자루는 각 인물의 손에 잡혀 있으며 일부 끝은 땅에 닿고 어깨 위 도구는 손과 어깨가 받친다. 지지 없이 공중에 뜬 몸이나 물체는 없다. 다만 봇짐 대부분을 등에 메거나 땅에 내려놓아, 손에 쥔 채 서 있다는 순간의 행동은 충분히 지키지 못했다."
       }
      ]
     },
     "ok": true
    }
   ],
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "slot_winner_match": true,
   "slot_winner": {
    "gemini-pro": "A",
    "gpt-high": "A"
   },
   "route": "cross_slot_agree"
  },
  "dual": {
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "normalized": {
    "A": 2.0,
    "B": 1.083
   },
   "adjusted": {
    "A": 2.0,
    "B": 0.833
   },
   "violations": {
    "B": [
     "[gemini-pro] invented people (레퍼런스와 일치하지 않는 새로운 얼굴의 인물들 생성)"
    ]
   },
   "per_model_winner": {
    "gemini-pro": "A",
    "gpt-high": "A"
   },
   "agreed": true
  },
  "totals": {
   "A": 2000,
   "B": 833
  },
  "selected": "A",
  "ranking": [
   "A",
   "B"
  ],
  "verdicts": [
   {
    "label": "A",
    "score": 2000,
    "verdict_ko": "레퍼런스의 장소 구조와 와이드 샷 구도를 완벽히 구현했으며, 지정된 인물의 복장 특징과 초록색 봇짐, 농기구 배치를 충실히 반영함."
   },
   {
    "label": "B",
    "score": 833,
    "verdict_ko": "지정된 얼굴 레퍼런스와 전혀 다른 인물들을 생성했으며, 지정된 소품(초록색 봇짐)을 누락하고 장소의 전체 구조도 맞지 않음.  ★위반: [gemini-pro] invented people (레퍼런스와 일치하지 않는 새로운 얼굴의 인물들 생성)"
   }
  ],
  "refs": [
   {
    "label": "LOCATION STRUCTURE PHOTOGRAPH — the confirmed photograph of this exact place and its fixed structure: it is the SINGLE authority for the location, the structure's shape, proportions, materials, colors, openings and every permanent site detail. Never copy its camera framing, time of day or lighting — the shot text is the authority for those.",
    "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/db00517b-165b-4005-a287-f28936b59b95/images/background_chain/seed_bg_mountain_camp_sel.png",
    "asset_id": "8c17c7ae-efff-4bb5-b88c-7c5528f00716",
    "role": "location_seed_bg"
   },
   {
    "label": "CHARACTER REFERENCE — 오주일: the exact person appearing in this shot; match face, hair and build exactly.",
    "path": "<bytes:809716>",
    "asset_id": "7c31e825-ceef-4208-915f-71c4f0e15dd6",
    "role": "character_ref"
   },
   {
    "label": "CHARACTER REFERENCE — 임정현: the exact person appearing in this shot; match face, hair and build exactly.",
    "path": "<bytes:858188>",
    "asset_id": "789c3359-d7f8-4832-9a6f-7e4ad2695c02",
    "role": "character_ref"
   },
   {
    "label": "CHARACTER REFERENCE — 대장장이: the exact person appearing in this shot; match face, hair and build exactly.",
    "path": "<bytes:946438>",
    "asset_id": "a12f8b90-e799-4753-9b7e-e71f4d817db5",
    "role": "character_ref"
   },
   {
    "label": "CHARACTER REFERENCE — 백정: the exact person appearing in this shot; match face, hair and build exactly.",
    "path": "<bytes:980360>",
    "asset_id": "29282584-ae7c-4722-8d77-100c4a5b2a3d",
    "role": "character_ref"
   },
   {
    "label": "CHARACTER REFERENCE — 목수: the exact person appearing in this shot; match face, hair and build exactly.",
    "path": "<bytes:904990>",
    "asset_id": "6b02463a-e818-4e3d-9816-215435a7aa0e",
    "role": "character_ref"
   },
   {
    "label": "PROP REFERENCE — 보자기 꾸러미·봇짐: the exact object appearing in this shot; match its look, material and wear exactly.",
    "path": "<bytes:1465734>",
    "asset_id": "3ca92b24-68b4-4883-b851-cabc06c00d1a",
    "role": "prop_ref"
   }
  ],
  "critique_skipped": true,
  "shot_run_produced": false,
  "shot_run_uid": "06aa1422-fd8e-7022-bc65-dd6e631043a0",
  "bgfirst": {
   "bg_path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/db00517b-165b-4005-a287-f28936b59b95/images/272bd7af-9294-4490-98c6-88a40e9deacc/scene/recipe/S26sh2__bgfirst_bg.png",
   "bg_asset_id": "0b28903d-d699-4da7-9123-e64d16e75375",
   "bg_record_key": "S26sh2::bgfirst_bg",
   "chain_winner": true,
   "authority": "seed_bg"
  },
  "ref_mode": "재투영 배경+콘티+엔티티 (2택1: 체인 승)",
  "share_plan": {
   "ref_plan": "background"
  },
  "lane_policy": "ab_select_ready"
 },
 "S26sh2::cine": {
  "applied": true,
  "attempted_at": "2026-09-09T03:51:31.775941+00:00",
  "fingerprint": "c7274d62908d35e31191a1ec5f34891d260cf0e41c1bab8a6db3fd5f500e28a8",
  "fingerprint_version": 2,
  "provider": "mai",
  "endpoint": "openrouter/chat-completions",
  "model": "microsoft/mai-image-2.6",
  "pack": "24.202608252115",
  "source_file": "S26sh2_sel.png",
  "source_sha256": "7b0dac63536e7f856f16a257a6bbc60ebe94834c22e08a6f8fb1782f8699842c",
  "file": "S26sh2_cine.png",
  "staged_sha256": "82135919afcfcc5c1264d50a900c89135ed0e815443ac15178ab8280ed2af47c",
  "latency_ms": 32423
 },
 "S26sh10::signage": {
  "fp": "c7c573e584424fc5",
  "inscriptions": [],
  "cues": [],
  "dropped": []
 },
 "S26sh10": {
  "input_fingerprint": "47289f00180fd29d",
  "prompt": "Create ONE FINAL photorealistic live-action film still of the moment below — early 20th-century Korea, Jeolla region; all people are Korean unless explicitly stated otherwise. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 백정의 눈치를 전혀 보지 않고 평온하고 단단한 표정으로 그를 마주 보는 안규홍의 정면.\n\nLOCATION (lock): The central meeting space between two facing groups in the mountain camp's outdoor clearing. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: medium shot\n- KEY BACKGROUND ELEMENTS: 산채 공터 (The conversational gap between the two groups remains open); used as A narrow visible interval between foreground shoulder and speaker keeps the exchange in shared space.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Maintain the clearing's natural daytime illumination and subdued contrast, letting warmth arise from the response rather than a lighting change.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The people visible in that photo are NOT in this shot: never carry their faces, bodies or clothing onto anyone here. Each person in THIS shot is defined solely by the PEOPLE list, their own CHARACTER REFERENCE images and their wardrobe notes. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nPREVIOUS STILL USAGE (follow exactly — what to take from the attached still and what to exclude; it governs place and objects only, never who is in this shot or how the camera sees them): Take the clearing's ground, surrounding vegetation, and daylight visible in the reference. Exclude the newcomers' handheld farm tools and bundles as fixed background decorations.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The daylight camp clearing remains arranged for the meeting; newcomers' bundles and old farm tools are still present. 안규홍: He stands calmly with the jade ring concealed in his clothing.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 안규홍 (한국인 남성, 청년 나이대, 검은 머리, 단단히 묶은 머리) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.",
  "roll_prompts": {
   "A": "Create ONE FINAL photorealistic live-action film still of the moment below — early 20th-century Korea, Jeolla region; all people are Korean unless explicitly stated otherwise. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 백정의 눈치를 전혀 보지 않고 평온하고 단단한 표정으로 그를 마주 보는 안규홍의 정면.\n\nLOCATION (lock): The central meeting space between two facing groups in the mountain camp's outdoor clearing. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: medium shot\n- KEY BACKGROUND ELEMENTS: 산채 공터 (The conversational gap between the two groups remains open); used as A narrow visible interval between foreground shoulder and speaker keeps the exchange in shared space.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Maintain the clearing's natural daytime illumination and subdued contrast, letting warmth arise from the response rather than a lighting change.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The people visible in that photo are NOT in this shot: never carry their faces, bodies or clothing onto anyone here. Each person in THIS shot is defined solely by the PEOPLE list, their own CHARACTER REFERENCE images and their wardrobe notes. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nPREVIOUS STILL USAGE (follow exactly — what to take from the attached still and what to exclude; it governs place and objects only, never who is in this shot or how the camera sees them): Take the clearing's ground, surrounding vegetation, and daylight visible in the reference. Exclude the newcomers' handheld farm tools and bundles as fixed background decorations.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The daylight camp clearing remains arranged for the meeting; newcomers' bundles and old farm tools are still present. 안규홍: He stands calmly with the jade ring concealed in his clothing.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 안규홍 (한국인 남성, 청년 나이대, 검은 머리, 단단히 묶은 머리) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.",
   "B": "Create ONE FINAL photorealistic live-action film still of the moment below — early 20th-century Korea, Jeolla region; all people are Korean unless explicitly stated otherwise. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 백정의 눈치를 전혀 보지 않고 평온하고 단단한 표정으로 그를 마주 보는 안규홍의 정면.\n\nLOCATION (lock): The central meeting space between two facing groups in the mountain camp's outdoor clearing. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: medium shot\n- KEY BACKGROUND ELEMENTS: 산채 공터 (The conversational gap between the two groups remains open); used as A narrow visible interval between foreground shoulder and speaker keeps the exchange in shared space.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Maintain the clearing's natural daytime illumination and subdued contrast, letting warmth arise from the response rather than a lighting change.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The people visible in that photo are NOT in this shot: never carry their faces, bodies or clothing onto anyone here. Each person in THIS shot is defined solely by the PEOPLE list, their own CHARACTER REFERENCE images and their wardrobe notes. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nPREVIOUS STILL USAGE (follow exactly — what to take from the attached still and what to exclude; it governs place and objects only, never who is in this shot or how the camera sees them): Take the clearing's ground, surrounding vegetation, and daylight visible in the reference. Exclude the newcomers' handheld farm tools and bundles as fixed background decorations.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The daylight camp clearing remains arranged for the meeting; newcomers' bundles and old farm tools are still present. 안규홍: He stands calmly with the jade ring concealed in his clothing.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 안규홍 (한국인 남성, 청년 나이대, 검은 머리, 단단히 묶은 머리) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.\n\nCOMPOSITION VARIATION (this candidate only): keep the place, the people and the moment exactly as specified, and keep from the CAMERA & FRAME contract its subject scale, screen placement and key background elements exactly. For this candidate only, its camera angle and height are deliberately left open: choose a camera position distinctly different from the obvious one, as a film director picking a second setup on the same blocking. Stay true to any explicit framing the SHOT TEXT itself dictates."
  },
  "shot_run_spend_attempt_count": 1,
  "readings": [
   {
    "label": "A",
    "direction": "안규홍의 시선이 화면 왼쪽 전경에 위치한 인물의 어깨와 얼굴 쪽을 향하고 있음.",
    "built_space": "산채 공터의 흙바닥, 배경의 초가집, 나무 울타리 등 위치 레퍼런스의 환경이 올바르게 배치됨.",
    "entities": "안규홍의 얼굴, 단단히 묶은 머리, 흙이 묻은 흰색 의상이 레퍼런스와 일치함. 옥가락지는 지시대로 옷 속에 감춰져 보이지 않음. 전경에 백정의 어깨가 프레임에 걸쳐 있음.",
    "hard_violations": [],
    "physics": "땅을 딛고 안정적으로 서 있는 자세가 자연스럽게 표현됨."
   },
   {
    "label": "B",
    "direction": "안규홍의 시선이 화면 왼쪽 전경의 인물을 향하고 있음.",
    "built_space": "산채 공터와 초가집, 나무 울타리가 배경에 있으나 우측 바닥에 짐꾸러미가 남아 있음.",
    "entities": "안규홍의 얼굴은 비슷하나 의상이 갈색으로 바뀌어 레퍼런스와 불일치함. 숨겨야 할 옥가락지가 허리띠 아래로 노출되어 있음. 전경에 백정의 뒷모습이 보임.",
    "hard_violations": [
     "[gpt-high] 참조에 없는 가죽 허리 주머니를 안규홍의 소지품으로 추가했다.",
     "[gpt-high] 고정된 공터 구조에 없는 별도 목책 출입문을 오른쪽 전경에 추가했다."
    ],
    "physics": "자연스럽게 서 있는 자세로 체중이 바닥에 안정적으로 실려 있음."
   }
  ],
  "cross_model_order": {
   "policy": "cross_model_order_v2_slot0_fwd_slot1_rev_parallel2",
   "slots": [
    {
     "model": "gemini-pro",
     "order": "forward",
     "display_to_canonical": {
      "A": "A",
      "B": "B"
     },
     "raw": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 10,
        "verdict_ko": "레퍼런스와 동일한 의상과 외모를 유지하며, 옥가락지를 숨긴 채 전경의 인물(어깨)을 평온하게 마주보는 오버더숄더 숏을 완벽하게 구현했습니다."
       },
       {
        "label": "B",
        "score": 4,
        "verdict_ko": "의상이 레퍼런스와 다르게 갈색으로 변경되었고, 숨겨야 할 옥가락지가 허리춤에 노출되었으며, 배제해야 할 짐꾸러미가 배경에 남아있어 감점되었습니다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "안규홍의 시선이 화면 왼쪽 전경에 위치한 인물의 어깨와 얼굴 쪽을 향하고 있음.",
        "built_space": "산채 공터의 흙바닥, 배경의 초가집, 나무 울타리 등 위치 레퍼런스의 환경이 올바르게 배치됨.",
        "entities": "안규홍의 얼굴, 단단히 묶은 머리, 흙이 묻은 흰색 의상이 레퍼런스와 일치함. 옥가락지는 지시대로 옷 속에 감춰져 보이지 않음. 전경에 백정의 어깨가 프레임에 걸쳐 있음.",
        "hard_violations": [],
        "physics": "땅을 딛고 안정적으로 서 있는 자세가 자연스럽게 표현됨."
       },
       {
        "label": "B",
        "direction": "안규홍의 시선이 화면 왼쪽 전경의 인물을 향하고 있음.",
        "built_space": "산채 공터와 초가집, 나무 울타리가 배경에 있으나 우측 바닥에 짐꾸러미가 남아 있음.",
        "entities": "안규홍의 얼굴은 비슷하나 의상이 갈색으로 바뀌어 레퍼런스와 불일치함. 숨겨야 할 옥가락지가 허리띠 아래로 노출되어 있음. 전경에 백정의 뒷모습이 보임.",
        "hard_violations": [],
        "physics": "자연스럽게 서 있는 자세로 체중이 바닥에 안정적으로 실려 있음."
       }
      ],
      "all_candidates_fail": false
     },
     "normalized": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 10,
        "verdict_ko": "레퍼런스와 동일한 의상과 외모를 유지하며, 옥가락지를 숨긴 채 전경의 인물(어깨)을 평온하게 마주보는 오버더숄더 숏을 완벽하게 구현했습니다."
       },
       {
        "label": "B",
        "score": 4,
        "verdict_ko": "의상이 레퍼런스와 다르게 갈색으로 변경되었고, 숨겨야 할 옥가락지가 허리춤에 노출되었으며, 배제해야 할 짐꾸러미가 배경에 남아있어 감점되었습니다."
       }
      ],
      "all_candidates_fail": false,
      "readings": [
       {
        "label": "A",
        "direction": "안규홍의 시선이 화면 왼쪽 전경에 위치한 인물의 어깨와 얼굴 쪽을 향하고 있음.",
        "built_space": "산채 공터의 흙바닥, 배경의 초가집, 나무 울타리 등 위치 레퍼런스의 환경이 올바르게 배치됨.",
        "entities": "안규홍의 얼굴, 단단히 묶은 머리, 흙이 묻은 흰색 의상이 레퍼런스와 일치함. 옥가락지는 지시대로 옷 속에 감춰져 보이지 않음. 전경에 백정의 어깨가 프레임에 걸쳐 있음.",
        "hard_violations": [],
        "physics": "땅을 딛고 안정적으로 서 있는 자세가 자연스럽게 표현됨."
       },
       {
        "label": "B",
        "direction": "안규홍의 시선이 화면 왼쪽 전경의 인물을 향하고 있음.",
        "built_space": "산채 공터와 초가집, 나무 울타리가 배경에 있으나 우측 바닥에 짐꾸러미가 남아 있음.",
        "entities": "안규홍의 얼굴은 비슷하나 의상이 갈색으로 바뀌어 레퍼런스와 불일치함. 숨겨야 할 옥가락지가 허리띠 아래로 노출되어 있음. 전경에 백정의 뒷모습이 보임.",
        "hard_violations": [],
        "physics": "자연스럽게 서 있는 자세로 체중이 바닥에 안정적으로 실려 있음."
       }
      ]
     },
     "ok": true
    },
    {
     "model": "gpt-high",
     "order": "reverse",
     "display_to_canonical": {
      "A": "B",
      "B": "A"
     },
     "raw": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 3,
        "verdict_ko": "백정을 향한 시선과 허리 위 구도는 적절하지만, 갈색 겉옷·가죽 주머니·노출된 옥색 소품과 새 전경 출입문이 인물 및 장소 고정을 위반한다."
       },
       {
        "label": "B",
        "score": 8,
        "verdict_ko": "정면의 평온하고 단단한 표정, 참조와 가까운 얼굴·의상, 열린 산채 공터를 충실히 구현하지만 시선은 전경 상대보다 렌즈 쪽에 가까워 보인다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "안규홍의 얼굴과 눈은 화면 왼쪽 전경 남자의 얼굴을 향한다. 전경 남자도 안규홍 쪽으로 고개를 돌려 두 사람의 대면 관계가 명확하다. 겨누는 무기나 움직이는 물체는 없다.",
        "built_space": "왼쪽 뒤에 큰 초가지붕과 낮은 초가지붕이 이어지고, 돌기단·바위·후면 목책·숲과 산이 보인다. 두 사람 사이에는 흙바닥 공터가 열린다. 다만 오른쪽 가까이에 굵은 기둥과 가로대가 있는 별도 목책 출입문을 배치했는데, 참조의 해당 공터는 연속된 외곽 목책으로 둘러싸여 있으며 이 전경 출입문은 확인되지 않는다. 반사면은 없다.",
        "entities": "중앙에는 검은 머리를 단단히 묶은 한국인 청년 남성으로 보이는 안규홍이 있고, 왼쪽에는 대사 상대인 백정으로 읽히는 성인 남성의 뒷머리·귀·어깨가 보인다. 안규홍의 얼굴은 참조와 일부 닮았지만 의상은 참조의 때 탄 밝은 저고리와 천 허리띠 대신 갈색 겹옷으로 바뀌었다. 허리에는 참조에 없는 가죽 주머니가 있으며 그 옆에 옥색 고리처럼 보이는 소품이 노출되어, 옥가락지를 옷 속에 감추라는 조건과 어긋난다. 오른쪽 가장자리에는 바닥에 놓인 보따리와 나무 도구 일부가 보인다. 읽을 수 있는 글자는 없다.",
        "hard_violations": [
         "참조에 없는 가죽 허리 주머니를 안규홍의 소지품으로 추가했다.",
         "고정된 공터 구조에 없는 별도 목책 출입문을 오른쪽 전경에 추가했다."
        ],
        "physics": "두 인물은 수직으로 서 있으며 발은 화면 밖이다. 보이는 상체에는 부유하거나 불가능하게 꺾인 부분이 없다. 옷은 어깨에서 내려오고 허리 주머니는 허리띠에 걸려 지지된다. 가장자리의 보따리와 도구는 땅에 놓여 있다. 도약이나 비행을 묘사한 요소는 없다."
       },
       {
        "label": "B",
        "direction": "안규홍은 얼굴을 거의 정면으로 두고 눈도 렌즈 부근을 향한다. 왼쪽 전경 남자는 안규홍을 향해 있지만, 안규홍이 그 남자의 얼굴을 정확히 응시하는지는 불분명하다. 무기나 방향을 갖고 이동하는 물체는 없다.",
        "built_space": "왼쪽 뒤의 큰 초가지붕과 오른쪽으로 이어지는 낮은 초가지붕, 돌기단, 후면의 연속된 목책, 오른쪽 바위, 소나무와 산이 보인다. 참조의 산채 재료와 기본 배치에 부합하며 별도 출입문이나 중복 시설은 없다. 안규홍은 공터에 서 있고 왼쪽 전경 어깨와 그의 몸 사이로 좁은 흙바닥 간격이 보여 대화 공간이 유지된다. 반사면은 없다.",
        "entities": "중앙 인물은 참조와 가까운 얼굴과 체격을 가진 한국인 청년 남성으로 보이며, 검은 머리는 뒤로 단단히 정리되어 있다. 때 탄 밝은 저고리, 가슴의 매듭과 천 허리띠가 참조에 가깝다. 왼쪽에는 백정 역할로 읽히는 성인 남성의 머리와 어깨 일부만 보인다. 옥가락지는 노출되지 않아 은닉 조건에 부합한다. 보따리와 농기구는 이 구도 안에서 확인되지 않지만 이를 보여주기 위해 프레임을 넓힐 필요는 없다. 읽을 수 있는 글자는 없다.",
        "hard_violations": [],
        "physics": "안규홍의 상체와 화면 하단의 손은 자연스럽게 아래로 이어지며, 서 있는 몸의 하부와 발은 잘려 있다. 옷은 어깨와 허리띠에 지지되어 자연스럽게 처진다. 전경 인물도 서 있는 자세의 일부로 읽히며, 지지 없이 떠 있는 신체나 물체는 없다. 다만 양팔을 내린 자세는 다소 대칭적이다."
       }
      ],
      "all_candidates_fail": false
     },
     "normalized": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 3,
        "verdict_ko": "백정을 향한 시선과 허리 위 구도는 적절하지만, 갈색 겉옷·가죽 주머니·노출된 옥색 소품과 새 전경 출입문이 인물 및 장소 고정을 위반한다."
       },
       {
        "label": "A",
        "score": 8,
        "verdict_ko": "정면의 평온하고 단단한 표정, 참조와 가까운 얼굴·의상, 열린 산채 공터를 충실히 구현하지만 시선은 전경 상대보다 렌즈 쪽에 가까워 보인다."
       }
      ],
      "all_candidates_fail": false,
      "readings": [
       {
        "label": "B",
        "direction": "안규홍의 얼굴과 눈은 화면 왼쪽 전경 남자의 얼굴을 향한다. 전경 남자도 안규홍 쪽으로 고개를 돌려 두 사람의 대면 관계가 명확하다. 겨누는 무기나 움직이는 물체는 없다.",
        "built_space": "왼쪽 뒤에 큰 초가지붕과 낮은 초가지붕이 이어지고, 돌기단·바위·후면 목책·숲과 산이 보인다. 두 사람 사이에는 흙바닥 공터가 열린다. 다만 오른쪽 가까이에 굵은 기둥과 가로대가 있는 별도 목책 출입문을 배치했는데, 참조의 해당 공터는 연속된 외곽 목책으로 둘러싸여 있으며 이 전경 출입문은 확인되지 않는다. 반사면은 없다.",
        "entities": "중앙에는 검은 머리를 단단히 묶은 한국인 청년 남성으로 보이는 안규홍이 있고, 왼쪽에는 대사 상대인 백정으로 읽히는 성인 남성의 뒷머리·귀·어깨가 보인다. 안규홍의 얼굴은 참조와 일부 닮았지만 의상은 참조의 때 탄 밝은 저고리와 천 허리띠 대신 갈색 겹옷으로 바뀌었다. 허리에는 참조에 없는 가죽 주머니가 있으며 그 옆에 옥색 고리처럼 보이는 소품이 노출되어, 옥가락지를 옷 속에 감추라는 조건과 어긋난다. 오른쪽 가장자리에는 바닥에 놓인 보따리와 나무 도구 일부가 보인다. 읽을 수 있는 글자는 없다.",
        "hard_violations": [
         "참조에 없는 가죽 허리 주머니를 안규홍의 소지품으로 추가했다.",
         "고정된 공터 구조에 없는 별도 목책 출입문을 오른쪽 전경에 추가했다."
        ],
        "physics": "두 인물은 수직으로 서 있으며 발은 화면 밖이다. 보이는 상체에는 부유하거나 불가능하게 꺾인 부분이 없다. 옷은 어깨에서 내려오고 허리 주머니는 허리띠에 걸려 지지된다. 가장자리의 보따리와 도구는 땅에 놓여 있다. 도약이나 비행을 묘사한 요소는 없다."
       },
       {
        "label": "A",
        "direction": "안규홍은 얼굴을 거의 정면으로 두고 눈도 렌즈 부근을 향한다. 왼쪽 전경 남자는 안규홍을 향해 있지만, 안규홍이 그 남자의 얼굴을 정확히 응시하는지는 불분명하다. 무기나 방향을 갖고 이동하는 물체는 없다.",
        "built_space": "왼쪽 뒤의 큰 초가지붕과 오른쪽으로 이어지는 낮은 초가지붕, 돌기단, 후면의 연속된 목책, 오른쪽 바위, 소나무와 산이 보인다. 참조의 산채 재료와 기본 배치에 부합하며 별도 출입문이나 중복 시설은 없다. 안규홍은 공터에 서 있고 왼쪽 전경 어깨와 그의 몸 사이로 좁은 흙바닥 간격이 보여 대화 공간이 유지된다. 반사면은 없다.",
        "entities": "중앙 인물은 참조와 가까운 얼굴과 체격을 가진 한국인 청년 남성으로 보이며, 검은 머리는 뒤로 단단히 정리되어 있다. 때 탄 밝은 저고리, 가슴의 매듭과 천 허리띠가 참조에 가깝다. 왼쪽에는 백정 역할로 읽히는 성인 남성의 머리와 어깨 일부만 보인다. 옥가락지는 노출되지 않아 은닉 조건에 부합한다. 보따리와 농기구는 이 구도 안에서 확인되지 않지만 이를 보여주기 위해 프레임을 넓힐 필요는 없다. 읽을 수 있는 글자는 없다.",
        "hard_violations": [],
        "physics": "안규홍의 상체와 화면 하단의 손은 자연스럽게 아래로 이어지며, 서 있는 몸의 하부와 발은 잘려 있다. 옷은 어깨와 허리띠에 지지되어 자연스럽게 처진다. 전경 인물도 서 있는 자세의 일부로 읽히며, 지지 없이 떠 있는 신체나 물체는 없다. 다만 양팔을 내린 자세는 다소 대칭적이다."
       }
      ]
     },
     "ok": true
    }
   ],
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "slot_winner_match": true,
   "slot_winner": {
    "gemini-pro": "A",
    "gpt-high": "A"
   },
   "route": "cross_slot_agree"
  },
  "dual": {
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "normalized": {
    "A": 2.0,
    "B": 0.775
   },
   "adjusted": {
    "A": 2.0,
    "B": 0.525
   },
   "violations": {
    "B": [
     "[gpt-high] 참조에 없는 가죽 허리 주머니를 안규홍의 소지품으로 추가했다.",
     "[gpt-high] 고정된 공터 구조에 없는 별도 목책 출입문을 오른쪽 전경에 추가했다."
    ]
   },
   "per_model_winner": {
    "gemini-pro": "A",
    "gpt-high": "A"
   },
   "agreed": true
  },
  "totals": {
   "A": 2000,
   "B": 525
  },
  "selected": "A",
  "ranking": [
   "A",
   "B"
  ],
  "verdicts": [
   {
    "label": "A",
    "score": 2000,
    "verdict_ko": "레퍼런스와 동일한 의상과 외모를 유지하며, 옥가락지를 숨긴 채 전경의 인물(어깨)을 평온하게 마주보는 오버더숄더 숏을 완벽하게 구현했습니다."
   },
   {
    "label": "B",
    "score": 525,
    "verdict_ko": "의상이 레퍼런스와 다르게 갈색으로 변경되었고, 숨겨야 할 옥가락지가 허리춤에 노출되었으며, 배제해야 할 짐꾸러미가 배경에 남아있어 감점되었습니다.  ★위반: [gpt-high] 참조에 없는 가죽 허리 주머니를 안규홍의 소지품으로 추가했다. / [gpt-high] 고정된 공터 구조에 없는 별도 목책 출입문을 오른쪽 전경에 추가했다."
   }
  ],
  "refs": [
   {
    "label": "PREVIOUS SHOT STILL — a visually related earlier shot of this same place: the location's look, materials, fixed features and lighting mood are LOCKED to this photo; never copy its camera framing. The people visible in that photo are NOT in this shot: never carry their faces, bodies or clothing onto anyone here. Each person in THIS shot is defined solely by the PEOPLE list, their own CHARACTER REFERENCE images and their wardrobe notes. If this photo shows a character who cannot move (dead or unconscious), that character's exact pose, position and orientation are ALSO LOCKED — treat their body as a fixed prop of the set that only the camera moves around.",
    "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/db00517b-165b-4005-a287-f28936b59b95/images/272bd7af-9294-4490-98c6-88a40e9deacc/scene/recipe/S26sh2_sel.png",
    "asset_id": "f8beba7b-1428-45c9-96e0-0f6b546f1fe2",
    "role": "prev_still"
   },
   {
    "label": "CHARACTER REFERENCE — 안규홍: the exact person appearing in this shot; match face, hair and build exactly.",
    "path": "<bytes:1028384>",
    "asset_id": "3b5894db-ce55-4f34-b966-5e725c3aa952",
    "role": "character_ref"
   }
  ],
  "critique_skipped": true,
  "shot_run_produced": false,
  "shot_run_uid": "06aa1423-00c6-74ba-9c44-1740fcf98c48",
  "ref_mode": "prev+엔티티",
  "share_plan": {
   "ref_plan": "prev",
   "prev_anchor_tag": "S26sh2"
  }
 },
 "S26sh10::cine": {
  "applied": true,
  "attempted_at": "2026-09-09T03:52:56.253635+00:00",
  "fingerprint": "350a1d97efff06b8e62612e3dacb81255e45fbef76c8357d99954468be753e40",
  "fingerprint_version": 2,
  "provider": "mai",
  "endpoint": "openrouter/chat-completions",
  "model": "microsoft/mai-image-2.6",
  "pack": "24.202608252115",
  "source_file": "S26sh10_sel.png",
  "source_sha256": "024f808186803c90fc500adaed9052008391c396a15bc05bfba76be412638f3d",
  "file": "S26sh10_cine.png",
  "staged_sha256": "a41b8a88360dc138937429cbb67e762c407e41d5f7d5f19cd6757a7c586c90a2",
  "latency_ms": 30654
 },
 "S26sh13::signage": {
  "fp": "82dc0394434eb930",
  "inscriptions": [],
  "cues": [],
  "dropped": []
 },
 "S26sh13": {
  "input_fingerprint": "37e806390eec3ab2",
  "prompt": "Create ONE FINAL photorealistic live-action film still of the moment below — early 20th-century Korea, Jeolla region; all people are Korean unless explicitly stated otherwise. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 무리 뒤쪽에서 사람들을 뚫고 앞으로 성큼 한 걸음 내디뎌, 아직 뒷발이 바닥을 밀어내고 있는 mid-action 자세의 오주일의 당당한 전신.\n\nLOCATION (lock): The rear-to-front opening through the newcomers gathered in the mountain camp's outdoor clearing. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: wide shot\n- FRAME LAYOUT: 오주일 in the middle-right of the frame, midground, moves toward open gap at lower left; open gap at lower left in the lower-left of the frame, foreground.\n- KEY BACKGROUND ELEMENTS: 산채 공터 (Open ground lies ahead of 오주일 between the groups); used as The clear lower-left area gives his forward step a legible destination and preserves contact beneath his trailing foot.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Keep the same natural daytime illumination and muted contrast, avoiding heroic lighting as his purposeful movement establishes authority.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The clothing, hair and overall look of 오주일 — who appear both in that photo and in this shot — are LOCKED to that photo. Anyone else visible in that photo is NOT in this shot: never carry their face, body or clothing onto anyone here. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nPREVIOUS STILL USAGE (follow exactly — what to take from the attached still and what to exclude; it governs place and objects only, never who is in this shot or how the camera sees them): Take the clearing's ground texture, vegetation, and daytime light. Exclude duplicated farm tools or travel bundles introduced by copying the earlier group.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): Bundles and old farm tools remain in the daylight meeting area. 오주일: He steps forward in rough clothes with a removal mark on his worn collar.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 오주일 (한국인 남성, 성인 나이대, 자연스러운 검은 머리) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.",
  "roll_prompts": {
   "A": "Create ONE FINAL photorealistic live-action film still of the moment below — early 20th-century Korea, Jeolla region; all people are Korean unless explicitly stated otherwise. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 무리 뒤쪽에서 사람들을 뚫고 앞으로 성큼 한 걸음 내디뎌, 아직 뒷발이 바닥을 밀어내고 있는 mid-action 자세의 오주일의 당당한 전신.\n\nLOCATION (lock): The rear-to-front opening through the newcomers gathered in the mountain camp's outdoor clearing. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: wide shot\n- FRAME LAYOUT: 오주일 in the middle-right of the frame, midground, moves toward open gap at lower left; open gap at lower left in the lower-left of the frame, foreground.\n- KEY BACKGROUND ELEMENTS: 산채 공터 (Open ground lies ahead of 오주일 between the groups); used as The clear lower-left area gives his forward step a legible destination and preserves contact beneath his trailing foot.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Keep the same natural daytime illumination and muted contrast, avoiding heroic lighting as his purposeful movement establishes authority.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The clothing, hair and overall look of 오주일 — who appear both in that photo and in this shot — are LOCKED to that photo. Anyone else visible in that photo is NOT in this shot: never carry their face, body or clothing onto anyone here. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nPREVIOUS STILL USAGE (follow exactly — what to take from the attached still and what to exclude; it governs place and objects only, never who is in this shot or how the camera sees them): Take the clearing's ground texture, vegetation, and daytime light. Exclude duplicated farm tools or travel bundles introduced by copying the earlier group.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): Bundles and old farm tools remain in the daylight meeting area. 오주일: He steps forward in rough clothes with a removal mark on his worn collar.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 오주일 (한국인 남성, 성인 나이대, 자연스러운 검은 머리) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.",
   "B": "Create ONE FINAL photorealistic live-action film still of the moment below — early 20th-century Korea, Jeolla region; all people are Korean unless explicitly stated otherwise. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 무리 뒤쪽에서 사람들을 뚫고 앞으로 성큼 한 걸음 내디뎌, 아직 뒷발이 바닥을 밀어내고 있는 mid-action 자세의 오주일의 당당한 전신.\n\nLOCATION (lock): The rear-to-front opening through the newcomers gathered in the mountain camp's outdoor clearing. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: wide shot\n- FRAME LAYOUT: 오주일 in the middle-right of the frame, midground, moves toward open gap at lower left; open gap at lower left in the lower-left of the frame, foreground.\n- KEY BACKGROUND ELEMENTS: 산채 공터 (Open ground lies ahead of 오주일 between the groups); used as The clear lower-left area gives his forward step a legible destination and preserves contact beneath his trailing foot.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Keep the same natural daytime illumination and muted contrast, avoiding heroic lighting as his purposeful movement establishes authority.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The clothing, hair and overall look of 오주일 — who appear both in that photo and in this shot — are LOCKED to that photo. Anyone else visible in that photo is NOT in this shot: never carry their face, body or clothing onto anyone here. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nPREVIOUS STILL USAGE (follow exactly — what to take from the attached still and what to exclude; it governs place and objects only, never who is in this shot or how the camera sees them): Take the clearing's ground texture, vegetation, and daytime light. Exclude duplicated farm tools or travel bundles introduced by copying the earlier group.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): Bundles and old farm tools remain in the daylight meeting area. 오주일: He steps forward in rough clothes with a removal mark on his worn collar.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 오주일 (한국인 남성, 성인 나이대, 자연스러운 검은 머리) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.\n\nCOMPOSITION VARIATION (this candidate only): keep the place, the people and the moment exactly as specified, and keep from the CAMERA & FRAME contract its subject scale, screen placement and key background elements exactly. For this candidate only, its camera angle and height are deliberately left open: choose a camera position distinctly different from the obvious one, as a film director picking a second setup on the same blocking. Stay true to any explicit framing the SHOT TEXT itself dictates."
  },
  "shot_run_spend_attempt_count": 1,
  "readings": [
   {
    "label": "A",
    "direction": "오주일의 시선과 진행 방향이 화면 좌측 하단의 열린 공간을 향하고 있습니다.",
    "built_space": "참조 이미지에 등장하는 초가집, 나무 울타리, 좌측의 바위 지형이 올바르게 렌더링되었습니다.",
    "entities": "오주일의 외형과 복장은 참조와 일치합니다. 그러나 바닥에 프롬프트에서 명시적으로 제외하라고 지시한 이전 샷의 초록색 짐꾸러미와 농기구들이 나타나 있습니다.",
    "hard_violations": [
     "[gemini-pro] 오주일의 양발이 모두 바닥에서 떨어져 아무런 지지 없이 허공에 떠 있음 (물리적 불가능)",
     "[gpt-high] 이전 장면에서 제외하라고 한 무리의 특징적인 의상·농기구·보따리 조합을 여러 인물에게 다시 사용해, 이전 군중과 짐을 복사하지 말라는 명시적 금지 조건을 위반한다."
    ],
    "physics": "오주일이 걷는 자세를 취하고 있으나, 양발이 모두 바닥의 그림자와 분리되어 공중에 붕 떠 있어 몸을 지탱하는 곳이 전혀 없습니다."
   },
   {
    "label": "B",
    "direction": "오주일이 화면 좌측 하단의 사람과 사람 사이 빈 공간을 향해 똑바로 걸어가고 있습니다.",
    "built_space": "산채 공터의 배경인 두 채의 초가집, 뒤편의 울타리, 산세와 흙바닥이 참조 이미지의 위치 및 재질과 정확히 일치하게 배치되었습니다.",
    "entities": "오주일의 얼굴, 제복, 소총이 참조 이미지와 일치하며, 프롬프트의 지시대로 이전 샷의 짐꾸러미나 도구 없이 마을 사람들이 새로운 군중을 형성하고 있습니다.",
    "hard_violations": [
     "[gpt-high] 앞으로 내민 신발과 뒤쪽 신발이 모두 지면에서 떨어져 보이는데, 몸은 평범하게 걷는 자세이고 이를 지탱하는 접촉점이나 도약 동작이 없어 공중에 떠 있는 staging이다."
    ],
    "physics": "오주일의 뒷발(왼발)이 바닥에 닿아 지면을 밀어내며 체중을 지탱하고 있고, 오른발이 딛기 위해 나아가는 자세가 물리적으로 완벽하게 구현되었습니다."
   }
  ],
  "cross_model_order": {
   "policy": "cross_model_order_v2_slot0_fwd_slot1_rev_parallel2",
   "slots": [
    {
     "model": "gemini-pro",
     "order": "forward",
     "display_to_canonical": {
      "A": "A",
      "B": "B"
     },
     "raw": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 10,
        "verdict_ko": "지시된 대로 오주일이 뒷발로 땅을 딛고 앞으로 나아가는 역동적인 자세를 완벽히 구현했으며, 요구된 프레이밍과 제외 사항(이전 컷의 짐꾸러미 등)을 정확히 준수했습니다."
       },
       {
        "label": "A",
        "score": 2,
        "verdict_ko": "오주일의 양발이 모두 땅에서 떨어져 허공에 떠 있는 심각한 물리적 오류가 발생했으며, 제외를 지시한 이전 장면의 짐꾸러미가 바닥에 그대로 렌더링되었습니다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "오주일의 시선과 진행 방향이 화면 좌측 하단의 열린 공간을 향하고 있습니다.",
        "built_space": "참조 이미지에 등장하는 초가집, 나무 울타리, 좌측의 바위 지형이 올바르게 렌더링되었습니다.",
        "entities": "오주일의 외형과 복장은 참조와 일치합니다. 그러나 바닥에 프롬프트에서 명시적으로 제외하라고 지시한 이전 샷의 초록색 짐꾸러미와 농기구들이 나타나 있습니다.",
        "hard_violations": [
         "오주일의 양발이 모두 바닥에서 떨어져 아무런 지지 없이 허공에 떠 있음 (물리적 불가능)"
        ],
        "physics": "오주일이 걷는 자세를 취하고 있으나, 양발이 모두 바닥의 그림자와 분리되어 공중에 붕 떠 있어 몸을 지탱하는 곳이 전혀 없습니다."
       },
       {
        "label": "B",
        "direction": "오주일이 화면 좌측 하단의 사람과 사람 사이 빈 공간을 향해 똑바로 걸어가고 있습니다.",
        "built_space": "산채 공터의 배경인 두 채의 초가집, 뒤편의 울타리, 산세와 흙바닥이 참조 이미지의 위치 및 재질과 정확히 일치하게 배치되었습니다.",
        "entities": "오주일의 얼굴, 제복, 소총이 참조 이미지와 일치하며, 프롬프트의 지시대로 이전 샷의 짐꾸러미나 도구 없이 마을 사람들이 새로운 군중을 형성하고 있습니다.",
        "hard_violations": [],
        "physics": "오주일의 뒷발(왼발)이 바닥에 닿아 지면을 밀어내며 체중을 지탱하고 있고, 오른발이 딛기 위해 나아가는 자세가 물리적으로 완벽하게 구현되었습니다."
       }
      ],
      "all_candidates_fail": false
     },
     "normalized": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 10,
        "verdict_ko": "지시된 대로 오주일이 뒷발로 땅을 딛고 앞으로 나아가는 역동적인 자세를 완벽히 구현했으며, 요구된 프레이밍과 제외 사항(이전 컷의 짐꾸러미 등)을 정확히 준수했습니다."
       },
       {
        "label": "A",
        "score": 2,
        "verdict_ko": "오주일의 양발이 모두 땅에서 떨어져 허공에 떠 있는 심각한 물리적 오류가 발생했으며, 제외를 지시한 이전 장면의 짐꾸러미가 바닥에 그대로 렌더링되었습니다."
       }
      ],
      "all_candidates_fail": false,
      "readings": [
       {
        "label": "A",
        "direction": "오주일의 시선과 진행 방향이 화면 좌측 하단의 열린 공간을 향하고 있습니다.",
        "built_space": "참조 이미지에 등장하는 초가집, 나무 울타리, 좌측의 바위 지형이 올바르게 렌더링되었습니다.",
        "entities": "오주일의 외형과 복장은 참조와 일치합니다. 그러나 바닥에 프롬프트에서 명시적으로 제외하라고 지시한 이전 샷의 초록색 짐꾸러미와 농기구들이 나타나 있습니다.",
        "hard_violations": [
         "오주일의 양발이 모두 바닥에서 떨어져 아무런 지지 없이 허공에 떠 있음 (물리적 불가능)"
        ],
        "physics": "오주일이 걷는 자세를 취하고 있으나, 양발이 모두 바닥의 그림자와 분리되어 공중에 붕 떠 있어 몸을 지탱하는 곳이 전혀 없습니다."
       },
       {
        "label": "B",
        "direction": "오주일이 화면 좌측 하단의 사람과 사람 사이 빈 공간을 향해 똑바로 걸어가고 있습니다.",
        "built_space": "산채 공터의 배경인 두 채의 초가집, 뒤편의 울타리, 산세와 흙바닥이 참조 이미지의 위치 및 재질과 정확히 일치하게 배치되었습니다.",
        "entities": "오주일의 얼굴, 제복, 소총이 참조 이미지와 일치하며, 프롬프트의 지시대로 이전 샷의 짐꾸러미나 도구 없이 마을 사람들이 새로운 군중을 형성하고 있습니다.",
        "hard_violations": [],
        "physics": "오주일의 뒷발(왼발)이 바닥에 닿아 지면을 밀어내며 체중을 지탱하고 있고, 오른발이 딛기 위해 나아가는 자세가 물리적으로 완벽하게 구현되었습니다."
       }
      ]
     },
     "ok": true
    },
    {
     "model": "gpt-high",
     "order": "reverse",
     "display_to_canonical": {
      "A": "B",
      "B": "A"
     },
     "raw": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 2,
        "verdict_ko": "중간 오른쪽 배치와 무리 사이 통로는 가깝지만, 두 발이 떠 있는 보행 자세여서 지면을 밀고 있는 뒷발이라는 핵심 조건과 신체 지지를 모두 놓쳤다."
       },
       {
        "label": "B",
        "score": 3,
        "verdict_ko": "앞발의 지면 지지는 A보다 설득력 있지만, 뒷발을 든 달리기 자세와 전경에 커진 인물, 이전 무리의 의상·짐·농기구 재사용으로 재촬영이 필요하다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "오주일의 얼굴과 앞발은 화면 왼쪽 아래 방향을 향하지만 이동축은 하단 중앙에 더 가깝다. 양옆 사람들은 대체로 오주일을 바라본다. 어깨 옆 소총은 위를 향해 세워져 있으며 누구를 겨누지는 않는다.",
        "built_space": "뒤쪽에 큰 초가지붕 한 채와 낮은 초가지붕 한 채, 오른쪽 경계의 목책, 왼쪽 바위 비탈이 보인다. 흙과 잔돌, 소나무와 산 배경은 참조 장소와 대체로 연결된다. 오주일은 중간 오른쪽에 전신으로 배치되고 사람들이 통로 양쪽에 서 있지만, 왼쪽 아래 끝은 가까운 사람들의 몸으로 상당 부분 막혀 있다.",
        "entities": "오주일은 검은 짧은 머리의 성인 동아시아계 남성으로 보이며 참조 얼굴과 대체로 닮았다. 해진 검정 제복, 금속 단추, 허리띠, 밝은 각반과 검정 신발이 보인다. 모자는 없으며 옷깃의 제거 흔적은 명확하지 않다. 주변에는 남녀 무리가 있어 원문의 군중은 구현했다. 이 구도에서 바닥에 남아 있어야 할 보따리와 농기구는 확인되지 않는다. 읽을 수 있는 글자는 없다.",
        "hard_violations": [
         "앞으로 내민 신발과 뒤쪽 신발이 모두 지면에서 떨어져 보이는데, 몸은 평범하게 걷는 자세이고 이를 지탱하는 접촉점이나 도약 동작이 없어 공중에 떠 있는 staging이다."
        ],
        "physics": "앞발 아래에는 흙과 그림자 사이의 간격이 보이고, 뒷발도 들려 있어 몸을 지지하는 발이 없다. 도약을 만든 웅크림이나 강한 밀어내기 없이 걷는 몸이 떠 있다. 따라서 요구된 뒷발의 지면 접촉도 없다. 소총은 어깨끈과 이를 잡은 손으로 지지되며 주변 사람들은 지면에 서 있다."
       },
       {
        "label": "B",
        "direction": "오주일의 얼굴과 진행 방향은 왼쪽 앞쪽의 빈 흙바닥을 향한다. 주변 사람들의 시선도 대체로 그에게 모인다. 소총은 어깨 옆에 세로로 운반되고, 군중의 농기구 날은 아래쪽을 향한다.",
        "built_space": "큰 초가지붕과 그 오른쪽의 낮은 초가지붕, 왼쪽 입석성 바위들, 뒤쪽 목책이 보인다. 장소의 주요 재료와 낮 조명은 참조와 이어진다. 왼쪽 아래 빈터는 명확하지만 군중은 통로 양옆보다 오주일 뒤에 밀집해 있다. 오주일은 중경보다 전경에 크게 나와 지정된 인물 크기와 배치에서 벗어난다.",
        "entities": "오주일은 검은 머리의 성인 동아시아계 남성이며 얼굴, 검정 제복, 단추, 허리띠와 밝은 각반은 인물 참조에 가깝다. 옷깃 제거 흔적은 확실히 식별되지 않는다. 뒤에는 성인 남녀와 어린 인물이 보인다. 이전 장면의 밝은 긴 도포 남성, 푸른 옷에 낫을 든 남성, 검은 머릿수건과 어두운 옷을 입고 농기구를 든 남성의 조합이 다시 나타난다. 바닥에는 녹색 보따리와 갈색 보따리, 오른쪽에 여러 농기구와 짐이 보인다. 읽을 수 있는 글자는 없다.",
        "hard_violations": [
         "이전 장면에서 제외하라고 한 무리의 특징적인 의상·농기구·보따리 조합을 여러 인물에게 다시 사용해, 이전 군중과 짐을 복사하지 말라는 명시적 금지 조건을 위반한다."
        ],
        "physics": "앞으로 나온 신발의 뒤꿈치 부근은 지면과 맞닿아 보이고, 뒤쪽 다리는 무릎을 굽혀 높이 들려 있다. A처럼 지지 없는 보행보다는 착지하는 달리기 동작으로 읽히지만, 뒷발이 아직 바닥을 밀고 있는 순간은 아니다. 소총은 손과 어깨끈으로 지지된다. 바닥의 보따리는 흙 위에 놓이고 농기구는 지면에 기대거나 사람의 손에 잡혀 있다."
       }
      ],
      "all_candidates_fail": true
     },
     "normalized": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 2,
        "verdict_ko": "중간 오른쪽 배치와 무리 사이 통로는 가깝지만, 두 발이 떠 있는 보행 자세여서 지면을 밀고 있는 뒷발이라는 핵심 조건과 신체 지지를 모두 놓쳤다."
       },
       {
        "label": "A",
        "score": 3,
        "verdict_ko": "앞발의 지면 지지는 A보다 설득력 있지만, 뒷발을 든 달리기 자세와 전경에 커진 인물, 이전 무리의 의상·짐·농기구 재사용으로 재촬영이 필요하다."
       }
      ],
      "all_candidates_fail": true,
      "readings": [
       {
        "label": "B",
        "direction": "오주일의 얼굴과 앞발은 화면 왼쪽 아래 방향을 향하지만 이동축은 하단 중앙에 더 가깝다. 양옆 사람들은 대체로 오주일을 바라본다. 어깨 옆 소총은 위를 향해 세워져 있으며 누구를 겨누지는 않는다.",
        "built_space": "뒤쪽에 큰 초가지붕 한 채와 낮은 초가지붕 한 채, 오른쪽 경계의 목책, 왼쪽 바위 비탈이 보인다. 흙과 잔돌, 소나무와 산 배경은 참조 장소와 대체로 연결된다. 오주일은 중간 오른쪽에 전신으로 배치되고 사람들이 통로 양쪽에 서 있지만, 왼쪽 아래 끝은 가까운 사람들의 몸으로 상당 부분 막혀 있다.",
        "entities": "오주일은 검은 짧은 머리의 성인 동아시아계 남성으로 보이며 참조 얼굴과 대체로 닮았다. 해진 검정 제복, 금속 단추, 허리띠, 밝은 각반과 검정 신발이 보인다. 모자는 없으며 옷깃의 제거 흔적은 명확하지 않다. 주변에는 남녀 무리가 있어 원문의 군중은 구현했다. 이 구도에서 바닥에 남아 있어야 할 보따리와 농기구는 확인되지 않는다. 읽을 수 있는 글자는 없다.",
        "hard_violations": [
         "앞으로 내민 신발과 뒤쪽 신발이 모두 지면에서 떨어져 보이는데, 몸은 평범하게 걷는 자세이고 이를 지탱하는 접촉점이나 도약 동작이 없어 공중에 떠 있는 staging이다."
        ],
        "physics": "앞발 아래에는 흙과 그림자 사이의 간격이 보이고, 뒷발도 들려 있어 몸을 지지하는 발이 없다. 도약을 만든 웅크림이나 강한 밀어내기 없이 걷는 몸이 떠 있다. 따라서 요구된 뒷발의 지면 접촉도 없다. 소총은 어깨끈과 이를 잡은 손으로 지지되며 주변 사람들은 지면에 서 있다."
       },
       {
        "label": "A",
        "direction": "오주일의 얼굴과 진행 방향은 왼쪽 앞쪽의 빈 흙바닥을 향한다. 주변 사람들의 시선도 대체로 그에게 모인다. 소총은 어깨 옆에 세로로 운반되고, 군중의 농기구 날은 아래쪽을 향한다.",
        "built_space": "큰 초가지붕과 그 오른쪽의 낮은 초가지붕, 왼쪽 입석성 바위들, 뒤쪽 목책이 보인다. 장소의 주요 재료와 낮 조명은 참조와 이어진다. 왼쪽 아래 빈터는 명확하지만 군중은 통로 양옆보다 오주일 뒤에 밀집해 있다. 오주일은 중경보다 전경에 크게 나와 지정된 인물 크기와 배치에서 벗어난다.",
        "entities": "오주일은 검은 머리의 성인 동아시아계 남성이며 얼굴, 검정 제복, 단추, 허리띠와 밝은 각반은 인물 참조에 가깝다. 옷깃 제거 흔적은 확실히 식별되지 않는다. 뒤에는 성인 남녀와 어린 인물이 보인다. 이전 장면의 밝은 긴 도포 남성, 푸른 옷에 낫을 든 남성, 검은 머릿수건과 어두운 옷을 입고 농기구를 든 남성의 조합이 다시 나타난다. 바닥에는 녹색 보따리와 갈색 보따리, 오른쪽에 여러 농기구와 짐이 보인다. 읽을 수 있는 글자는 없다.",
        "hard_violations": [
         "이전 장면에서 제외하라고 한 무리의 특징적인 의상·농기구·보따리 조합을 여러 인물에게 다시 사용해, 이전 군중과 짐을 복사하지 말라는 명시적 금지 조건을 위반한다."
        ],
        "physics": "앞으로 나온 신발의 뒤꿈치 부근은 지면과 맞닿아 보이고, 뒤쪽 다리는 무릎을 굽혀 높이 들려 있다. A처럼 지지 없는 보행보다는 착지하는 달리기 동작으로 읽히지만, 뒷발이 아직 바닥을 밀고 있는 순간은 아니다. 소총은 손과 어깨끈으로 지지된다. 바닥의 보따리는 흙 위에 놓이고 농기구는 지면에 기대거나 사람의 손에 잡혀 있다."
       }
      ]
     },
     "ok": true
    }
   ],
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "slot_winner_match": false,
   "slot_winner": {
    "gemini-pro": "B",
    "gpt-high": "A"
   },
   "route": "cross_slot_combined"
  },
  "dual": {
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "normalized": {
    "A": 1.2,
    "B": 1.667
   },
   "adjusted": {
    "A": 0.95,
    "B": 1.417
   },
   "violations": {
    "A": [
     "[gemini-pro] 오주일의 양발이 모두 바닥에서 떨어져 아무런 지지 없이 허공에 떠 있음 (물리적 불가능)",
     "[gpt-high] 이전 장면에서 제외하라고 한 무리의 특징적인 의상·농기구·보따리 조합을 여러 인물에게 다시 사용해, 이전 군중과 짐을 복사하지 말라는 명시적 금지 조건을 위반한다."
    ],
    "B": [
     "[gpt-high] 앞으로 내민 신발과 뒤쪽 신발이 모두 지면에서 떨어져 보이는데, 몸은 평범하게 걷는 자세이고 이를 지탱하는 접촉점이나 도약 동작이 없어 공중에 떠 있는 staging이다."
    ]
   },
   "per_model_winner": {
    "gemini-pro": "B",
    "gpt-high": "A"
   },
   "agreed": false
  },
  "totals": {
   "B": 1417,
   "A": 950
  },
  "selected": "B",
  "ranking": [
   "B",
   "A"
  ],
  "verdicts": [
   {
    "label": "B",
    "score": 1417,
    "verdict_ko": "지시된 대로 오주일이 뒷발로 땅을 딛고 앞으로 나아가는 역동적인 자세를 완벽히 구현했으며, 요구된 프레이밍과 제외 사항(이전 컷의 짐꾸러미 등)을 정확히 준수했습니다.  ★위반: [gpt-high] 앞으로 내민 신발과 뒤쪽 신발이 모두 지면에서 떨어져 보이는데, 몸은 평범하게 걷는 자세이고 이를 지탱하는 접촉점이나 도약 동작이 없어 공중에 떠 있는 staging이다."
   },
   {
    "label": "A",
    "score": 950,
    "verdict_ko": "오주일의 양발이 모두 땅에서 떨어져 허공에 떠 있는 심각한 물리적 오류가 발생했으며, 제외를 지시한 이전 장면의 짐꾸러미가 바닥에 그대로 렌더링되었습니다.  ★위반: [gemini-pro] 오주일의 양발이 모두 바닥에서 떨어져 아무런 지지 없이 허공에 떠 있음 (물리적 불가능) / [gpt-high] 이전 장면에서 제외하라고 한 무리의 특징적인 의상·농기구·보따리 조합을 여러 인물에게 다시 사용해, 이전 군중과 짐을 복사하지 말라는 명시적 금지 조건을 위반한다."
   }
  ],
  "refs": [
   {
    "label": "PREVIOUS SHOT STILL — a visually related earlier shot of this same place: the location's look, materials, fixed features and lighting mood are LOCKED to this photo; never copy its camera framing. The clothing, hair and overall look of 오주일 — who appear both in that photo and in this shot — are LOCKED to that photo. Anyone else visible in that photo is NOT in this shot: never carry their face, body or clothing onto anyone here. If this photo shows a character who cannot move (dead or unconscious), that character's exact pose, position and orientation are ALSO LOCKED — treat their body as a fixed prop of the set that only the camera moves around.",
    "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/db00517b-165b-4005-a287-f28936b59b95/images/272bd7af-9294-4490-98c6-88a40e9deacc/scene/recipe/S26sh2_sel.png",
    "asset_id": "f8beba7b-1428-45c9-96e0-0f6b546f1fe2",
    "role": "prev_still"
   },
   {
    "label": "CHARACTER REFERENCE — 오주일: the exact person appearing in this shot; match face, hair and build exactly.",
    "path": "<bytes:809716>",
    "asset_id": "7c31e825-ceef-4208-915f-71c4f0e15dd6",
    "role": "character_ref"
   }
  ],
  "critique_skipped": true,
  "shot_run_produced": false,
  "shot_run_uid": "06aa1423-0264-73aa-833c-32b52652ba0e",
  "ref_mode": "prev+엔티티",
  "share_plan": {
   "ref_plan": "prev",
   "prev_anchor_tag": "S26sh2"
  }
 },
 "S26sh13::cine": {
  "applied": true,
  "attempted_at": "2026-09-09T03:54:30.548792+00:00",
  "fingerprint": "8af53c9962bab896f0cad602a94808ce1faef2d3a6f8a7943c283212cef666cb",
  "fingerprint_version": 2,
  "provider": "mai",
  "endpoint": "openrouter/chat-completions",
  "model": "microsoft/mai-image-2.6",
  "pack": "24.202608252115",
  "source_file": "S26sh13_sel.png",
  "source_sha256": "e5c1b822764272b0051ba7b00d1d8ac3c061cf0e96fe51379a60f20fd94fcc25",
  "file": "S26sh13_cine.png",
  "staged_sha256": "1662c8f7dbe518294eead630b4d8bbc25f9ada21dc9cc320ea935e3fa1d6d647",
  "latency_ms": 31593
 },
 "S27sh3::signage": {
  "fp": "dbb580f1f7cb4aab",
  "inscriptions": [],
  "cues": [],
  "dropped": []
 },
 "S27sh3": {
  "input_fingerprint": "221b6795fd1da231",
  "prompt": "Create ONE FINAL photorealistic live-action film still of the moment below — early 20th-century Korea, Jeolla region; all people are Korean unless explicitly stated otherwise. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 서로 반대 방향으로 몸을 돌린 채 어깨가 맞부딪힌 대장장이와 백정의 당황한 상체 클로즈업.\n\nLOCATION (lock): The outdoor drill formation area in the mountain camp's clearing, with room for ranks to practice turning. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: close-up\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Natural daytime illumination and restrained contrast keep the awkward collision tactile and human rather than broadly comic.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The people visible in that photo are NOT in this shot: never carry their faces, bodies or clothing onto anyone here. Each person in THIS shot is defined solely by the PEOPLE list, their own CHARACTER REFERENCE images and their wardrobe notes. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): Daylight fills the training clearing. 대장장이: He is awkwardly turned out of alignment in the drill. 백정: He is awkwardly turned in the drill formation.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 대장장이 (한국인 남성, 성인 나이대, 자연스러운 검은 머리); 백정 (한국인 남성, 성인 나이대, 자연스러운 검은 머리) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.",
  "roll_prompts": {
   "A": "Create ONE FINAL photorealistic live-action film still of the moment below — early 20th-century Korea, Jeolla region; all people are Korean unless explicitly stated otherwise. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 서로 반대 방향으로 몸을 돌린 채 어깨가 맞부딪힌 대장장이와 백정의 당황한 상체 클로즈업.\n\nLOCATION (lock): The outdoor drill formation area in the mountain camp's clearing, with room for ranks to practice turning. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: close-up\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Natural daytime illumination and restrained contrast keep the awkward collision tactile and human rather than broadly comic.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The people visible in that photo are NOT in this shot: never carry their faces, bodies or clothing onto anyone here. Each person in THIS shot is defined solely by the PEOPLE list, their own CHARACTER REFERENCE images and their wardrobe notes. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): Daylight fills the training clearing. 대장장이: He is awkwardly turned out of alignment in the drill. 백정: He is awkwardly turned in the drill formation.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 대장장이 (한국인 남성, 성인 나이대, 자연스러운 검은 머리); 백정 (한국인 남성, 성인 나이대, 자연스러운 검은 머리) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.",
   "B": "Create ONE FINAL photorealistic live-action film still of the moment below — early 20th-century Korea, Jeolla region; all people are Korean unless explicitly stated otherwise. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 서로 반대 방향으로 몸을 돌린 채 어깨가 맞부딪힌 대장장이와 백정의 당황한 상체 클로즈업.\n\nLOCATION (lock): The outdoor drill formation area in the mountain camp's clearing, with room for ranks to practice turning. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: close-up\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Natural daytime illumination and restrained contrast keep the awkward collision tactile and human rather than broadly comic.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The people visible in that photo are NOT in this shot: never carry their faces, bodies or clothing onto anyone here. Each person in THIS shot is defined solely by the PEOPLE list, their own CHARACTER REFERENCE images and their wardrobe notes. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): Daylight fills the training clearing. 대장장이: He is awkwardly turned out of alignment in the drill. 백정: He is awkwardly turned in the drill formation.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 대장장이 (한국인 남성, 성인 나이대, 자연스러운 검은 머리); 백정 (한국인 남성, 성인 나이대, 자연스러운 검은 머리) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.\n\nCOMPOSITION VARIATION (this candidate only): keep the place, the people and the moment exactly as specified, and keep from the CAMERA & FRAME contract its subject scale, screen placement and key background elements exactly. For this candidate only, its camera angle and height are deliberately left open: choose a camera position distinctly different from the obvious one, as a film director picking a second setup on the same blocking. Stay true to any explicit framing the SHOT TEXT itself dictates."
  },
  "shot_run_spend_attempt_count": 1,
  "readings": [
   {
    "label": "A",
    "direction": "두 남자가 서로 반대 방향으로 고개를 돌려 당황한 표정으로 어딘가를 바라보고 있음.",
    "built_space": "흙바닥으로 된 산채 공터이며, 배경에 초가집, 바위, 나무 울타리, 소나무 숲이 이전 샷 레퍼런스와 일치하게 배치되어 있음.",
    "entities": "대장장이(푸른 계열 옷, 가죽 앞치마 추가됨)와 백정(낡고 해진 옷, 두건)이 레퍼런스의 외모에 맞게 등장하며, 지시대로 그 외의 인물은 전혀 등장하지 않음.",
    "hard_violations": [],
    "physics": "어깨를 부딪힌 충격으로 팔을 뻗고 중심을 잡으려는 듯한 자연스러운 무게 중심 이동이 보이며 두 다리로 지면을 온전히 디디고 있음."
   },
   {
    "label": "B",
    "direction": "두 남자가 등을 맞댄 채 서로 반대 방향으로 고개를 돌려 바라보고 있음.",
    "built_space": "초가집과 산, 흙바닥이 이전 샷의 환경과 잘 일치함.",
    "entities": "대장장이와 백정이 레퍼런스의 복장과 외모를 잘 따르고 있으나, 배경에 프롬프트가 엄격하게 금지한 추가 인물(마을 사람들)이 다수 존재함.",
    "hard_violations": [
     "[gemini-pro] 프롬프트에 명시되지 않은 인물(배경의 사람들) 추가",
     "[gpt-high] 두 주인공 외에는 누구도 등장하지 말라는 지시를 어기고 배경에 여러 명의 추가 인물을 배치했다."
    ],
    "physics": "두 사람 모두 지면에 다리를 대고 서서 몸을 비틀고 있는 자세가 자연스럽게 지탱되고 있음."
   }
  ],
  "cross_model_order": {
   "policy": "cross_model_order_v2_slot0_fwd_slot1_rev_parallel2",
   "slots": [
    {
     "model": "gemini-pro",
     "order": "forward",
     "display_to_canonical": {
      "A": "A",
      "B": "B"
     },
     "raw": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 9,
        "verdict_ko": "명시된 인물 외에 아무도 등장하지 않도록 배경을 철저히 비웠으며, 어깨가 부딪힌 두 사람의 상체 클로즈업과 당황한 표정을 훌륭하게 구현했습니다."
       },
       {
        "label": "B",
        "score": 2,
        "verdict_ko": "주요 인물들의 배치와 표정은 좋으나, 샷 텍스트에 없는 인물들을 추가하지 말라는 지시를 어기고 배경에 다수의 엑스트라를 배치한 치명적인 오류가 있습니다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "두 남자가 서로 반대 방향으로 고개를 돌려 당황한 표정으로 어딘가를 바라보고 있음.",
        "built_space": "흙바닥으로 된 산채 공터이며, 배경에 초가집, 바위, 나무 울타리, 소나무 숲이 이전 샷 레퍼런스와 일치하게 배치되어 있음.",
        "entities": "대장장이(푸른 계열 옷, 가죽 앞치마 추가됨)와 백정(낡고 해진 옷, 두건)이 레퍼런스의 외모에 맞게 등장하며, 지시대로 그 외의 인물은 전혀 등장하지 않음.",
        "hard_violations": [],
        "physics": "어깨를 부딪힌 충격으로 팔을 뻗고 중심을 잡으려는 듯한 자연스러운 무게 중심 이동이 보이며 두 다리로 지면을 온전히 디디고 있음."
       },
       {
        "label": "B",
        "direction": "두 남자가 등을 맞댄 채 서로 반대 방향으로 고개를 돌려 바라보고 있음.",
        "built_space": "초가집과 산, 흙바닥이 이전 샷의 환경과 잘 일치함.",
        "entities": "대장장이와 백정이 레퍼런스의 복장과 외모를 잘 따르고 있으나, 배경에 프롬프트가 엄격하게 금지한 추가 인물(마을 사람들)이 다수 존재함.",
        "hard_violations": [
         "프롬프트에 명시되지 않은 인물(배경의 사람들) 추가"
        ],
        "physics": "두 사람 모두 지면에 다리를 대고 서서 몸을 비틀고 있는 자세가 자연스럽게 지탱되고 있음."
       }
      ],
      "all_candidates_fail": false
     },
     "normalized": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 9,
        "verdict_ko": "명시된 인물 외에 아무도 등장하지 않도록 배경을 철저히 비웠으며, 어깨가 부딪힌 두 사람의 상체 클로즈업과 당황한 표정을 훌륭하게 구현했습니다."
       },
       {
        "label": "B",
        "score": 2,
        "verdict_ko": "주요 인물들의 배치와 표정은 좋으나, 샷 텍스트에 없는 인물들을 추가하지 말라는 지시를 어기고 배경에 다수의 엑스트라를 배치한 치명적인 오류가 있습니다."
       }
      ],
      "all_candidates_fail": false,
      "readings": [
       {
        "label": "A",
        "direction": "두 남자가 서로 반대 방향으로 고개를 돌려 당황한 표정으로 어딘가를 바라보고 있음.",
        "built_space": "흙바닥으로 된 산채 공터이며, 배경에 초가집, 바위, 나무 울타리, 소나무 숲이 이전 샷 레퍼런스와 일치하게 배치되어 있음.",
        "entities": "대장장이(푸른 계열 옷, 가죽 앞치마 추가됨)와 백정(낡고 해진 옷, 두건)이 레퍼런스의 외모에 맞게 등장하며, 지시대로 그 외의 인물은 전혀 등장하지 않음.",
        "hard_violations": [],
        "physics": "어깨를 부딪힌 충격으로 팔을 뻗고 중심을 잡으려는 듯한 자연스러운 무게 중심 이동이 보이며 두 다리로 지면을 온전히 디디고 있음."
       },
       {
        "label": "B",
        "direction": "두 남자가 등을 맞댄 채 서로 반대 방향으로 고개를 돌려 바라보고 있음.",
        "built_space": "초가집과 산, 흙바닥이 이전 샷의 환경과 잘 일치함.",
        "entities": "대장장이와 백정이 레퍼런스의 복장과 외모를 잘 따르고 있으나, 배경에 프롬프트가 엄격하게 금지한 추가 인물(마을 사람들)이 다수 존재함.",
        "hard_violations": [
         "프롬프트에 명시되지 않은 인물(배경의 사람들) 추가"
        ],
        "physics": "두 사람 모두 지면에 다리를 대고 서서 몸을 비틀고 있는 자세가 자연스럽게 지탱되고 있음."
       }
      ]
     },
     "ok": true
    },
    {
     "model": "gpt-high",
     "order": "reverse",
     "display_to_canonical": {
      "A": "B",
      "B": "A"
     },
     "raw": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 2,
        "verdict_ko": "반대 방향으로 돌아 어깨를 맞댄 두 사람은 잘 보이지만, 금지된 배경 인물들을 다수 추가해 실격이다."
       },
       {
        "label": "B",
        "score": 7,
        "verdict_ko": "두 사람만 등장하며 어깨 충돌과 당황한 반응을 자연스럽게 구현했지만, 반대 방향의 몸 회전이 약하고 클로즈업이 다소 넓으며 대장장이의 앞치마가 참조와 다르다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "왼쪽 대장장이의 몸통은 화면 왼쪽으로, 오른쪽 백정의 몸통은 오른쪽으로 돌아 있다. 두 사람은 고개와 눈을 안쪽으로 돌려 서로를 곁눈질하며, 중앙에서 어깨 뒤쪽이 맞닿는다. 반대 방향으로 돌다가 상대를 확인하는 관계가 읽힌다. 무기나 방향을 확인할 휴대 물건은 없다.",
        "built_space": "흙 공터 뒤로 큰 초가지붕 한 채와 오른쪽의 낮은 초가지붕 한 채, 왼쪽 바위 면, 오른쪽 목책, 뒤쪽 숲이 보인다. 참조 장소의 재료와 구성은 대체로 이어지지만 건물들이 화면에서 더 크게 차지한다. 두 주인공은 공터 전경에 서 있으며, 배경에는 왼쪽 두 명, 중앙의 부분적으로 가려진 인물들, 오른쪽 여러 명이 추가되어 있다. 반사면이나 중복된 고정 설비는 보이지 않는다.",
        "entities": "대장장이와 백정은 참조와 유사한 얼굴의 검은 머리 성인 남성으로 표현되며, 한국인 설정과 충돌하는 외형 단서는 없다. 대장장이의 남색 작업복과 걷은 소매, 백정의 갈색 머릿수건과 해지고 붉은 얼룩이 있는 겉옷은 참조에 가깝다. 벌어진 입과 옆눈으로 당황함을 표현한다. 그러나 허용된 두 사람 외에 성인 남녀가 여러 명 등장하며 이전 장면의 훈련 인원 구성을 다시 가져왔다. 읽을 수 있는 글자는 없다.",
        "hard_violations": [
         "두 주인공 외에는 누구도 등장하지 말라는 지시를 어기고 배경에 여러 명의 추가 인물을 배치했다."
        ],
        "physics": "두 주인공의 어깨와 등 위쪽이 접촉하고 옷이 그 접촉을 따라 눌려 있어 실제로 가능한 자세다. 팔은 몸 옆으로 내려와 있으며 몸통은 화면 아래로 이어진다. 발은 잘려 있어 지면 접촉을 직접 확인할 수 없지만 공중에 떠 있다는 단서는 없다. 배경에서 발이 보이는 인물들은 흙바닥에 서 있다. 다만 주인공의 자세는 충돌 순간의 반동보다는 충돌 후 멈춘 상태에 가깝다."
       },
       {
        "label": "B",
        "direction": "왼쪽 대장장이는 고개와 시선을 오른쪽 백정에게 돌리고, 백정은 왼쪽 대장장이를 바라본다. 중앙에서 두 사람의 어깨와 위팔이 맞닿는다. 대장장이의 몸은 약간 왼쪽, 백정의 몸은 약간 오른쪽으로 열려 있으나 둘 다 가슴 앞면이 많이 보여, 서로 반대 방향으로 돌아선 상태는 A보다 약하다. 뻗은 팔은 상대를 가리키기보다 충돌 뒤 균형을 잡는 방향이다.",
        "built_space": "흙 공터, 왼쪽 바위 비탈, 뒤쪽의 큰 초가지붕 한 채와 오른쪽 낮은 초가지붕 한 채, 오른쪽 목책, 숲이 보인다. 두 사람은 같은 공터의 전경에 있고 건물과 목책은 뒤쪽 규모로 유지된다. 추가 인물이나 명백히 중복된 고정 설비는 없다. 반사면은 없다. 인물의 허리와 양팔까지 포함되어 요청한 상체 클로즈업보다는 다소 넓게 잡혔다.",
        "entities": "참조의 대장장이와 백정에 대응하는 성인 남성 두 명만 보인다. 얼굴과 검은 머리, 대장장이의 남색 옷과 걷은 소매, 백정의 갈색 머릿수건과 낡고 얼룩진 겉옷은 대체로 일치한다. 대장장이에게는 참조의 남색 허리 작업포와 달리 목걸이 끈이 있는 갈색 가죽 가슴 앞치마가 보인다. 두 사람의 벌어진 입과 서로를 확인하는 눈빛은 당황한 표정에 맞는다. 별도 도구나 읽을 수 있는 글자는 없다.",
        "hard_violations": [],
        "physics": "두 사람의 어깨와 위팔이 접촉하며, 백정의 팔이 대장장이 팔 앞을 지나가는 겹침은 해부학적으로 가능하다. 네 팔과 손은 각자의 몸에 자연스럽게 연결되어 있고, 벌어진 팔과 굽은 팔꿈치는 갑작스러운 접촉 뒤 균형을 잡는 동작으로 읽힌다. 앞치마는 목 끈과 허리 부분으로 지지된다. 하체와 발은 화면 밖이라 지면 접촉을 확인할 수 없지만, 떠 있는 몸이나 지지 없이 공중에 놓인 물건은 없다."
       }
      ],
      "all_candidates_fail": false
     },
     "normalized": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 2,
        "verdict_ko": "반대 방향으로 돌아 어깨를 맞댄 두 사람은 잘 보이지만, 금지된 배경 인물들을 다수 추가해 실격이다."
       },
       {
        "label": "A",
        "score": 7,
        "verdict_ko": "두 사람만 등장하며 어깨 충돌과 당황한 반응을 자연스럽게 구현했지만, 반대 방향의 몸 회전이 약하고 클로즈업이 다소 넓으며 대장장이의 앞치마가 참조와 다르다."
       }
      ],
      "all_candidates_fail": false,
      "readings": [
       {
        "label": "B",
        "direction": "왼쪽 대장장이의 몸통은 화면 왼쪽으로, 오른쪽 백정의 몸통은 오른쪽으로 돌아 있다. 두 사람은 고개와 눈을 안쪽으로 돌려 서로를 곁눈질하며, 중앙에서 어깨 뒤쪽이 맞닿는다. 반대 방향으로 돌다가 상대를 확인하는 관계가 읽힌다. 무기나 방향을 확인할 휴대 물건은 없다.",
        "built_space": "흙 공터 뒤로 큰 초가지붕 한 채와 오른쪽의 낮은 초가지붕 한 채, 왼쪽 바위 면, 오른쪽 목책, 뒤쪽 숲이 보인다. 참조 장소의 재료와 구성은 대체로 이어지지만 건물들이 화면에서 더 크게 차지한다. 두 주인공은 공터 전경에 서 있으며, 배경에는 왼쪽 두 명, 중앙의 부분적으로 가려진 인물들, 오른쪽 여러 명이 추가되어 있다. 반사면이나 중복된 고정 설비는 보이지 않는다.",
        "entities": "대장장이와 백정은 참조와 유사한 얼굴의 검은 머리 성인 남성으로 표현되며, 한국인 설정과 충돌하는 외형 단서는 없다. 대장장이의 남색 작업복과 걷은 소매, 백정의 갈색 머릿수건과 해지고 붉은 얼룩이 있는 겉옷은 참조에 가깝다. 벌어진 입과 옆눈으로 당황함을 표현한다. 그러나 허용된 두 사람 외에 성인 남녀가 여러 명 등장하며 이전 장면의 훈련 인원 구성을 다시 가져왔다. 읽을 수 있는 글자는 없다.",
        "hard_violations": [
         "두 주인공 외에는 누구도 등장하지 말라는 지시를 어기고 배경에 여러 명의 추가 인물을 배치했다."
        ],
        "physics": "두 주인공의 어깨와 등 위쪽이 접촉하고 옷이 그 접촉을 따라 눌려 있어 실제로 가능한 자세다. 팔은 몸 옆으로 내려와 있으며 몸통은 화면 아래로 이어진다. 발은 잘려 있어 지면 접촉을 직접 확인할 수 없지만 공중에 떠 있다는 단서는 없다. 배경에서 발이 보이는 인물들은 흙바닥에 서 있다. 다만 주인공의 자세는 충돌 순간의 반동보다는 충돌 후 멈춘 상태에 가깝다."
       },
       {
        "label": "A",
        "direction": "왼쪽 대장장이는 고개와 시선을 오른쪽 백정에게 돌리고, 백정은 왼쪽 대장장이를 바라본다. 중앙에서 두 사람의 어깨와 위팔이 맞닿는다. 대장장이의 몸은 약간 왼쪽, 백정의 몸은 약간 오른쪽으로 열려 있으나 둘 다 가슴 앞면이 많이 보여, 서로 반대 방향으로 돌아선 상태는 A보다 약하다. 뻗은 팔은 상대를 가리키기보다 충돌 뒤 균형을 잡는 방향이다.",
        "built_space": "흙 공터, 왼쪽 바위 비탈, 뒤쪽의 큰 초가지붕 한 채와 오른쪽 낮은 초가지붕 한 채, 오른쪽 목책, 숲이 보인다. 두 사람은 같은 공터의 전경에 있고 건물과 목책은 뒤쪽 규모로 유지된다. 추가 인물이나 명백히 중복된 고정 설비는 없다. 반사면은 없다. 인물의 허리와 양팔까지 포함되어 요청한 상체 클로즈업보다는 다소 넓게 잡혔다.",
        "entities": "참조의 대장장이와 백정에 대응하는 성인 남성 두 명만 보인다. 얼굴과 검은 머리, 대장장이의 남색 옷과 걷은 소매, 백정의 갈색 머릿수건과 낡고 얼룩진 겉옷은 대체로 일치한다. 대장장이에게는 참조의 남색 허리 작업포와 달리 목걸이 끈이 있는 갈색 가죽 가슴 앞치마가 보인다. 두 사람의 벌어진 입과 서로를 확인하는 눈빛은 당황한 표정에 맞는다. 별도 도구나 읽을 수 있는 글자는 없다.",
        "hard_violations": [],
        "physics": "두 사람의 어깨와 위팔이 접촉하며, 백정의 팔이 대장장이 팔 앞을 지나가는 겹침은 해부학적으로 가능하다. 네 팔과 손은 각자의 몸에 자연스럽게 연결되어 있고, 벌어진 팔과 굽은 팔꿈치는 갑작스러운 접촉 뒤 균형을 잡는 동작으로 읽힌다. 앞치마는 목 끈과 허리 부분으로 지지된다. 하체와 발은 화면 밖이라 지면 접촉을 확인할 수 없지만, 떠 있는 몸이나 지지 없이 공중에 놓인 물건은 없다."
       }
      ]
     },
     "ok": true
    }
   ],
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "slot_winner_match": true,
   "slot_winner": {
    "gemini-pro": "A",
    "gpt-high": "A"
   },
   "route": "cross_slot_agree"
  },
  "dual": {
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "normalized": {
    "A": 2.0,
    "B": 0.508
   },
   "adjusted": {
    "A": 2.0,
    "B": 0.258
   },
   "violations": {
    "B": [
     "[gemini-pro] 프롬프트에 명시되지 않은 인물(배경의 사람들) 추가",
     "[gpt-high] 두 주인공 외에는 누구도 등장하지 말라는 지시를 어기고 배경에 여러 명의 추가 인물을 배치했다."
    ]
   },
   "per_model_winner": {
    "gemini-pro": "A",
    "gpt-high": "A"
   },
   "agreed": true
  },
  "totals": {
   "A": 2000,
   "B": 258
  },
  "selected": "A",
  "ranking": [
   "A",
   "B"
  ],
  "verdicts": [
   {
    "label": "A",
    "score": 2000,
    "verdict_ko": "명시된 인물 외에 아무도 등장하지 않도록 배경을 철저히 비웠으며, 어깨가 부딪힌 두 사람의 상체 클로즈업과 당황한 표정을 훌륭하게 구현했습니다."
   },
   {
    "label": "B",
    "score": 258,
    "verdict_ko": "주요 인물들의 배치와 표정은 좋으나, 샷 텍스트에 없는 인물들을 추가하지 말라는 지시를 어기고 배경에 다수의 엑스트라를 배치한 치명적인 오류가 있습니다.  ★위반: [gemini-pro] 프롬프트에 명시되지 않은 인물(배경의 사람들) 추가 / [gpt-high] 두 주인공 외에는 누구도 등장하지 말라는 지시를 어기고 배경에 여러 명의 추가 인물을 배치했다."
   }
  ],
  "refs": [
   {
    "label": "PREVIOUS SHOT STILL — a visually related earlier shot of this same place: the location's look, materials, fixed features and lighting mood are LOCKED to this photo; never copy its camera framing. The people visible in that photo are NOT in this shot: never carry their faces, bodies or clothing onto anyone here. Each person in THIS shot is defined solely by the PEOPLE list, their own CHARACTER REFERENCE images and their wardrobe notes. If this photo shows a character who cannot move (dead or unconscious), that character's exact pose, position and orientation are ALSO LOCKED — treat their body as a fixed prop of the set that only the camera moves around.",
    "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/db00517b-165b-4005-a287-f28936b59b95/images/272bd7af-9294-4490-98c6-88a40e9deacc/scene/recipe/S26sh13_sel.png",
    "asset_id": "816c628d-e5dd-4934-af3d-7a0e65fe198d",
    "role": "prev_still"
   },
   {
    "label": "CHARACTER REFERENCE — 대장장이: the exact person appearing in this shot; match face, hair and build exactly.",
    "path": "<bytes:946438>",
    "asset_id": "a12f8b90-e799-4753-9b7e-e71f4d817db5",
    "role": "character_ref"
   },
   {
    "label": "CHARACTER REFERENCE — 백정: the exact person appearing in this shot; match face, hair and build exactly.",
    "path": "<bytes:980360>",
    "asset_id": "29282584-ae7c-4722-8d77-100c4a5b2a3d",
    "role": "character_ref"
   }
  ],
  "critique_skipped": true,
  "shot_run_produced": false,
  "shot_run_uid": "06aa1423-03fc-79bc-9761-9471c6b1003d",
  "ref_mode": "prev+엔티티",
  "share_plan": {
   "ref_plan": "prev",
   "prev_anchor_tag": "S26sh13"
  }
 },
 "S27sh3::cine": {
  "applied": true,
  "attempted_at": "2026-09-09T03:55:58.024153+00:00",
  "fingerprint": "685b331c7d5f39215da9b6ad6a8bd7a306b254625d671ef7756e976889f49930",
  "fingerprint_version": 2,
  "provider": "mai",
  "endpoint": "openrouter/chat-completions",
  "model": "microsoft/mai-image-2.6",
  "pack": "24.202608252115",
  "source_file": "S27sh3_sel.png",
  "source_sha256": "687e3a8247aa02ee86b8866275c259018af6f5e59cf1d6aa3f7b0fd87a253042",
  "file": "S27sh3_cine.png",
  "staged_sha256": "f609780996e311d0780f5d35b1d330bcffc8c06e9469d432db25079d549af054",
  "latency_ms": 32732
 },
 "S27sh21::signage": {
  "fp": "d03e439d58c86fba",
  "inscriptions": [],
  "cues": [],
  "dropped": []
 },
 "groupbg::mountain_camp_hearth": {
  "input_fingerprint": "f8e7b346da8c5c14",
  "meta": {
   "model": "gpt-image-2",
   "size": "1536x864",
   "pack": "11.202607220237",
   "contract": "bgfirst_full_v3",
   "group_sig": {
    "key": "mountain_camp_hearth",
    "tags": [
     "S27sh21",
     "S27sh26",
     "S41sh1",
     "S41sh12",
     "S41sh8",
     "S50sh1",
     "S50sh4",
     "S54sh4",
     "S56sh40",
     "S58sh3",
     "S60sh11",
     "S60sh2",
     "S61sh15",
     "S61sh32",
     "S64sh10",
     "S64sh18",
     "S64sh3",
     "S66sh1",
     "S66sh9"
    ]
   },
   "context_sig": "5522294bc0f77962"
  },
  "prompt": "Create ONE empty live-action location background photograph — NO PEOPLE, no figures, no body parts, no silhouettes, no shadows or reflections of people anywhere.\n\"Empty\" means no people only: KEEP the location's inherent occupants and stock that define the place — animals in an animal shelter, pen or farm, goods and produce in a market, moored boats in a harbour — unless the shot text explicitly removes them.\nThe FIRST attached image is a thin-line storyboard sketch of a shot that happens at this location: use it ONLY as spatial evidence — what this place contains, how its ground, structures and landmarks are arranged and proportioned. Ignore the sketched people and arrows entirely, and do NOT copy its line style: render a fully photographic, physically plausible real place that fits THE LOCATION text below.\nHUMAN-SCALE CALIBRATION: derive every structure's true size from human-scale elements — a door ≈ 2m, a window ≈ 1–1.5m wide, one storey ≈ 2.5–3m; never inflate a small structure or shrink a large one.\n\nTHE LOCATION — early 20th-century Korea, Jeolla region; all people are Korean unless explicitly stated otherwise: The outdoor serving area beside the mountain camp's large cooking pot after training.\n\nLOCATION DETAIL (descriptive evidence about this place from the production's location records — evidence, not a staging order): use it to understand what kind of place this is and what it permanently contains. Carry over ONLY its enduring physical features — layout, structures, ground and materials, fixed equipment and installations. Do NOT treat any lighting, weather or time-of-day wording, momentary object states, depicted events or subjective impressions in it as fixed facts of the place; the TIME OF DAY lock below is the sole authority for time and lighting.\n파청 산채 공터·훈련장·취사장·막사 주변: 산 중턱에 통나무를 세워 만든 조잡한 목책 안의 넓은 흙바닥 공터. 나무와 거적으로 만든 움막과 막사가 흩어져 있고, 돌과 진흙으로 만든 취사용 아궁이와 무쇠 솥, 모닥불 터가 있다. 전투 후에는 솥이 뒤집히고 막사가 불타거나 무너진다. (특징: 통나무를 박아 세운 목책; 넓게 다져진 흙바닥 (공터/훈련장); 돌과 진흙으로 만든 다수의 아궁이와 무쇠솥; 나무와 거적으로 엮어 만든 움막과 막사; 부서진 토기 밥그릇과 뒤집힌 무쇠솥 (전투 후); 불에 탄 나무 잔해와 흩어진 쌀밥 (전투 후))\n\nSCENE EVIDENCE (verbatim quotes from the screenplay about this place — treat them as evidence of what the location physically contains and looks like; stage the PLACE those moments happen in, but do NOT depict the momentary actions, people or staged props themselves):\n- 5. 훈련이 끝나고 가마솥 앞으로 몰려드는 사내들.\n- 군데군데 부상자의 신음. 모닥불 하나가 유난히 낮게 타들어간다.\n- 의병들이 삼삼오오 둘러앉아 밥을 먹는다. 오랜만의 쌀밥.\n- 정현이 밥을 먹다 말고 고개를 든다.\n- 사방에서 일제사격. 밥그릇이 깨진다. 의병 하나가 밥을 씹던 그대로 쓰러진다.\n솥이 뒤집힌다.\n- 뒤집힌 솥. 깨진 밥그릇. 흙과 피에 뒤섞인 쌀밥.\n- 발밑. 깨진 밥그릇. 흙과 피 속에 흩어진 하얀 쌀밥.\n- 어깨가 피로 흥건하다. 숨을 죽인 채 쓰러진 사람들 사이에 엎드려 있다.\n- 아까까지 사람들이 웃으며 먹던 밥그릇과 감자가 피와 흙 속에 흩어져 있다.\n- 재보가 나무에 기대 있던 몸을 일으킨다.\n어깨에서 피가 흐른다. 오주일이 총을 집어 든다.\n\nTIME OF DAY (lock): day.\n\nRender ONE photorealistic empty location photograph, 16:9, neutral enough that every shot of this place can be staged from it later. No readable writing anywhere: surfaces that would carry writing may be present, but stage any wording out of legibility — an oblique angle, distance, shallow focus. No captions, watermarks or overlay text.",
  "effective_prompt": "Create ONE empty live-action location background photograph — NO PEOPLE, no figures, no body parts, no silhouettes, no shadows or reflections of people anywhere.\n\"Empty\" means no people only: KEEP the location's inherent occupants and stock that define the place — animals in an animal shelter, pen or farm, goods and produce in a market, moored boats in a harbour — unless the shot text explicitly removes them.\nThe FIRST attached image is a thin-line storyboard sketch of a shot that happens at this location: use it ONLY as spatial evidence — what this place contains, how its ground, structures and landmarks are arranged and proportioned. Ignore the sketched people and arrows entirely, and do NOT copy its line style: render a fully photographic, physically plausible real place that fits THE LOCATION text below.\nHUMAN-SCALE CALIBRATION: derive every structure's true size from human-scale elements — a door ≈ 2m, a window ≈ 1–1.5m wide, one storey ≈ 2.5–3m; never inflate a small structure or shrink a large one.\n\nTHE LOCATION — early 20th-century Korea, Jeolla region; all people are Korean unless explicitly stated otherwise: The outdoor serving area beside the mountain camp's large cooking pot after training.\n\nLOCATION DETAIL (descriptive evidence about this place from the production's location records — evidence, not a staging order): use it to understand what kind of place this is and what it permanently contains. Carry over ONLY its enduring physical features — layout, structures, ground and materials, fixed equipment and installations. Do NOT treat any lighting, weather or time-of-day wording, momentary object states, depicted events or subjective impressions in it as fixed facts of the place; the TIME OF DAY lock below is the sole authority for time and lighting.\n파청 산채 공터·훈련장·취사장·막사 주변: 산 중턱에 통나무를 세워 만든 조잡한 목책 안의 넓은 흙바닥 공터. 나무와 거적으로 만든 움막과 막사가 흩어져 있고, 돌과 진흙으로 만든 취사용 아궁이와 무쇠 솥, 모닥불 터가 있다. 전투 후에는 솥이 뒤집히고 막사가 불타거나 무너진다. (특징: 통나무를 박아 세운 목책; 넓게 다져진 흙바닥 (공터/훈련장); 돌과 진흙으로 만든 다수의 아궁이와 무쇠솥; 나무와 거적으로 엮어 만든 움막과 막사; 부서진 토기 밥그릇과 뒤집힌 무쇠솥 (전투 후); 불에 탄 나무 잔해와 흩어진 쌀밥 (전투 후))\n\nSCENE EVIDENCE (verbatim quotes from the screenplay about this place — treat them as evidence of what the location physically contains and looks like; stage the PLACE those moments happen in, but do NOT depict the momentary actions, people or staged props themselves):\n- 5. 훈련이 끝나고 가마솥 앞으로 몰려드는 사내들.\n- 군데군데 부상자의 신음. 모닥불 하나가 유난히 낮게 타들어간다.\n- 의병들이 삼삼오오 둘러앉아 밥을 먹는다. 오랜만의 쌀밥.\n- 정현이 밥을 먹다 말고 고개를 든다.\n- 사방에서 일제사격. 밥그릇이 깨진다. 의병 하나가 밥을 씹던 그대로 쓰러진다.\n솥이 뒤집힌다.\n- 뒤집힌 솥. 깨진 밥그릇. 흙과 피에 뒤섞인 쌀밥.\n- 발밑. 깨진 밥그릇. 흙과 피 속에 흩어진 하얀 쌀밥.\n- 어깨가 피로 흥건하다. 숨을 죽인 채 쓰러진 사람들 사이에 엎드려 있다.\n- 아까까지 사람들이 웃으며 먹던 밥그릇과 감자가 피와 흙 속에 흩어져 있다.\n- 재보가 나무에 기대 있던 몸을 일으킨다.\n어깨에서 피가 흐른다. 오주일이 총을 집어 든다.\n\nTIME OF DAY (lock): day.\n\nRender ONE photorealistic empty location photograph, 16:9, neutral enough that every shot of this place can be staged from it later. No readable writing anywhere: surfaces that would carry writing may be present, but stage any wording out of legibility — an oblique angle, distance, shallow focus. No captions, watermarks or overlay text.",
  "bg_path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/db00517b-165b-4005-a287-f28936b59b95/images/272bd7af-9294-4490-98c6-88a40e9deacc/scene/recipe/groupbg_mountain_camp_hearth_81c95f.png",
  "asset_id": "87bcdd2b-607c-40d6-8d7f-f4dfbf3ed054",
  "input_asset_ids": [
   "6b43ab3c-f77b-46d3-9765-83ffa19fa20f"
  ],
  "origin_tag": "S27sh21",
  "place_text": "The outdoor serving area beside the mountain camp's large cooking pot after training.",
  "origin_inputs": {
   "place_text": "The outdoor serving area beside the mountain camp's large cooking pot after training.",
   "time_of_day_en": "day",
   "conti_asset_id": "6b43ab3c-f77b-46d3-9765-83ffa19fa20f"
  }
 },
 "S27sh21::bgfirst_bg": {
  "input_fingerprint": "79ad921e00ca4bd5",
  "prompt": "Create the EMPTY BACKGROUND PLATE for one film shot — NO PEOPLE, no figures, no body parts, no sketch lines, no arrows anywhere.\n\"Empty\" means no people only: KEEP the location's inherent occupants and stock that define the place — animals in an animal shelter, pen or farm, goods in a market, moored boats in a harbour — unless the shot text explicitly removes them.\nThe FIRST attached image is a thin-line storyboard sketch: use ONLY its camera angle, horizon, perspective and the placement/size of buildings and set masses — ignore the sketched people and arrows entirely. The SECOND attached image (LOCATION PHOTOGRAPH) is the real place: take its architecture, materials, signage and fixed features, and RE-PROJECT them into the sketch's camera. If the photograph's camera differs from the sketch's, the sketch's camera wins.\nHUMAN-SCALE CALIBRATION: derive every structure's true size from human-scale elements — a door ≈ 2m, a window ≈ 1–1.5m wide, one storey ≈ 2.5–3m; never inflate a small structure or shrink a large one.\n\nSHOT TEXT this background must serve (Korean): 자신의 토기 사발에 담긴 멀건 국물 같은 죽을 허탈하게 내려다보는 백정의 퀭한 시선.\n\nLOCATION (lock): The outdoor serving area beside the mountain camp's large cooking pot after training.\n\nTIME OF DAY (lock): day.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: close-up\n- KEY BACKGROUND ELEMENTS: 토기 사발 (Held below 백정's face and containing thin porridge) — Its open interior tilts into the camera's downward view; used as A small lower-frame counterpoint to his eyes makes the cause of his disappointment visible; 멀건 죽 (Watery porridge inside the bowl); used as The visible contents ground his reaction in the inadequate meal.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Natural daytime illumination remains subdued and unembellished, preserving detail in his tired face and the thin porridge without adding atmospheric effects.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nRender ONE photorealistic empty location photograph, 16:9, that this shot can be staged inside later. No readable writing anywhere: surfaces that would carry writing may be present, but stage any wording out of legibility — an oblique angle, distance, shallow focus. No captions, watermarks or overlay text.",
  "effective_prompt": "Create the EMPTY BACKGROUND PLATE for one film shot — NO PEOPLE, no figures, no body parts, no sketch lines, no arrows anywhere.\n\"Empty\" means no people only: KEEP the location's inherent occupants and stock that define the place — animals in an animal shelter, pen or farm, goods in a market, moored boats in a harbour — unless the shot text explicitly removes them.\nThe FIRST attached image is a thin-line storyboard sketch: use ONLY its camera angle, horizon, perspective and the placement/size of buildings and set masses — ignore the sketched people and arrows entirely. The SECOND attached image (LOCATION PHOTOGRAPH) is the real place: take its architecture, materials, signage and fixed features, and RE-PROJECT them into the sketch's camera. If the photograph's camera differs from the sketch's, the sketch's camera wins.\nHUMAN-SCALE CALIBRATION: derive every structure's true size from human-scale elements — a door ≈ 2m, a window ≈ 1–1.5m wide, one storey ≈ 2.5–3m; never inflate a small structure or shrink a large one.\n\nSHOT TEXT this background must serve (Korean): 자신의 토기 사발에 담긴 멀건 국물 같은 죽을 허탈하게 내려다보는 백정의 퀭한 시선.\n\nLOCATION (lock): The outdoor serving area beside the mountain camp's large cooking pot after training.\n\nTIME OF DAY (lock): day.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: close-up\n- KEY BACKGROUND ELEMENTS: 토기 사발 (Held below 백정's face and containing thin porridge) — Its open interior tilts into the camera's downward view; used as A small lower-frame counterpoint to his eyes makes the cause of his disappointment visible; 멀건 죽 (Watery porridge inside the bowl); used as The visible contents ground his reaction in the inadequate meal.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Natural daytime illumination remains subdued and unembellished, preserving detail in his tired face and the thin porridge without adding atmospheric effects.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nRender ONE photorealistic empty location photograph, 16:9, that this shot can be staged inside later. No readable writing anywhere: surfaces that would carry writing may be present, but stage any wording out of legibility — an oblique angle, distance, shallow focus. No captions, watermarks or overlay text.",
  "bg_path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/db00517b-165b-4005-a287-f28936b59b95/images/272bd7af-9294-4490-98c6-88a40e9deacc/scene/recipe/S27sh21__bgfirst_bg.png",
  "asset_id": "e40cc56e-348b-46ad-a7ae-b303c53153e0",
  "input_asset_ids": [
   "6b43ab3c-f77b-46d3-9765-83ffa19fa20f",
   "87bcdd2b-607c-40d6-8d7f-f4dfbf3ed054"
  ]
 },
 "S27sh21": {
  "input_fingerprint": "ca1ff3d8a2fc85b1",
  "prompt": "Create ONE FINAL photorealistic live-action film still of the moment below — early 20th-century Korea, Jeolla region; all people are Korean unless explicitly stated otherwise. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 자신의 토기 사발에 담긴 멀건 국물 같은 죽을 허탈하게 내려다보는 백정의 퀭한 시선.\n\nLOCATION (lock): The outdoor serving area beside the mountain camp's large cooking pot after training. The shot takes place here — the attached LOCATION PHOTOGRAPH shows the exact spot.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: close-up\n- KEY BACKGROUND ELEMENTS: 토기 사발 (Held below 백정's face and containing thin porridge) — Its open interior tilts into the camera's downward view; used as A small lower-frame counterpoint to his eyes makes the cause of his disappointment visible; 멀건 죽 (Watery porridge inside the bowl); used as The visible contents ground his reaction in the inadequate meal.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Natural daytime illumination remains subdued and unembellished, preserving detail in his tired face and the thin porridge without adding atmospheric effects.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The cauldron lid is open, exposing watery porridge; serving bowls are filled with the same thin food. 백정: He holds his bowl and looks down into the watery porridge.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 백정 (한국인 남성, 성인 나이대, 자연스러운 검은 머리) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.",
  "roll_prompts": {
   "A": "Stage the shot. The FIRST attached image (SHOT BACKGROUND) is the finished empty background of this shot — keep it EXACTLY: its camera, perspective, architecture, lighting and every fixed feature stay untouched. The SECOND attached image (LAYOUT SKETCH) tells you ONLY where the people go: each sketched person's position, screen size, pose and the gaze/motion arrows. Ignore the sketch's background lines. The CHARACTER REFERENCE photographs show the real people.\nPlace the real people into the background at exactly the sketched positions, sizes and poses, following the arrow directions. No sketch lines or arrows may remain.\n\nCreate ONE FINAL photorealistic live-action film still of the moment below — early 20th-century Korea, Jeolla region; all people are Korean unless explicitly stated otherwise. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 자신의 토기 사발에 담긴 멀건 국물 같은 죽을 허탈하게 내려다보는 백정의 퀭한 시선.\n\nLOCATION (lock): The outdoor serving area beside the mountain camp's large cooking pot after training. The shot takes place here — the FIRST attached image (SHOT BACKGROUND) is this exact place, already built: its ground, structures, horizon, materials and lighting are the finished truth of this location and must not be redesigned or replaced. No location photograph is attached — read the place from that image alone, and add no scenery, structure, vehicle or fixture that it does not already show. This lock governs the place only; the figures in the shot follow the staging and pose instructions.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: close-up\n- KEY BACKGROUND ELEMENTS: 토기 사발 (Held below 백정's face and containing thin porridge) — Its open interior tilts into the camera's downward view; used as A small lower-frame counterpoint to his eyes makes the cause of his disappointment visible; 멀건 죽 (Watery porridge inside the bowl); used as The visible contents ground his reaction in the inadequate meal.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Natural daytime illumination remains subdued and unembellished, preserving detail in his tired face and the thin porridge without adding atmospheric effects.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The cauldron lid is open, exposing watery porridge; serving bowls are filled with the same thin food. 백정: He holds his bowl and looks down into the watery porridge.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 백정 (한국인 남성, 성인 나이대, 자연스러운 검은 머리) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.",
   "B": "Create ONE FINAL photorealistic live-action film still of the moment below — early 20th-century Korea, Jeolla region; all people are Korean unless explicitly stated otherwise. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 자신의 토기 사발에 담긴 멀건 국물 같은 죽을 허탈하게 내려다보는 백정의 퀭한 시선.\n\nLOCATION (lock): The outdoor serving area beside the mountain camp's large cooking pot after training. The shot takes place here — the attached LOCATION PHOTOGRAPH shows the exact spot.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: close-up\n- KEY BACKGROUND ELEMENTS: 토기 사발 (Held below 백정's face and containing thin porridge) — Its open interior tilts into the camera's downward view; used as A small lower-frame counterpoint to his eyes makes the cause of his disappointment visible; 멀건 죽 (Watery porridge inside the bowl); used as The visible contents ground his reaction in the inadequate meal.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Natural daytime illumination remains subdued and unembellished, preserving detail in his tired face and the thin porridge without adding atmospheric effects.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The cauldron lid is open, exposing watery porridge; serving bowls are filled with the same thin food. 백정: He holds his bowl and looks down into the watery porridge.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 백정 (한국인 남성, 성인 나이대, 자연스러운 검은 머리) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.\n\nCOMPOSITION VARIATION (this candidate only): keep the place, the people and the moment exactly as specified, and keep from the CAMERA & FRAME contract its subject scale, screen placement and key background elements exactly. For this candidate only, its camera angle and height are deliberately left open: choose a camera position distinctly different from the obvious one, as a film director picking a second setup on the same blocking. Stay true to any explicit framing the SHOT TEXT itself dictates."
  },
  "roll_refs": {
   "A": [
    {
     "label": "SHOT BACKGROUND",
     "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/db00517b-165b-4005-a287-f28936b59b95/images/272bd7af-9294-4490-98c6-88a40e9deacc/scene/recipe/S27sh21__bgfirst_bg.png",
     "asset_id": "e40cc56e-348b-46ad-a7ae-b303c53153e0",
     "role": "bgfirst_bg"
    },
    {
     "label": "LAYOUT SKETCH (people placement only)",
     "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/db00517b-165b-4005-a287-f28936b59b95/images/272bd7af-9294-4490-98c6-88a40e9deacc/conti/conti_S27sh21.png",
     "asset_id": "6b43ab3c-f77b-46d3-9765-83ffa19fa20f",
     "role": "conti_light"
    },
    {
     "label": "CHARACTER REFERENCE — 백정: the exact person appearing in this shot; match face, hair and build exactly.",
     "path": "<bytes:980360>",
     "asset_id": "29282584-ae7c-4722-8d77-100c4a5b2a3d",
     "role": "character_ref"
    }
   ],
   "B": [
    {
     "label": "LOCATION PHOTOGRAPH — the exact place of this shot: its architecture, materials, fixed features and lighting mood are spatial truth; stage the moment inside this place. Never copy its camera framing.",
     "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/db00517b-165b-4005-a287-f28936b59b95/images/272bd7af-9294-4490-98c6-88a40e9deacc/scene/recipe/groupbg_mountain_camp_hearth_81c95f.png",
     "asset_id": "87bcdd2b-607c-40d6-8d7f-f4dfbf3ed054",
     "role": "bgfirst_group_bg"
    },
    {
     "label": "CHARACTER REFERENCE — 백정: the exact person appearing in this shot; match face, hair and build exactly.",
     "path": "<bytes:980360>",
     "asset_id": "29282584-ae7c-4722-8d77-100c4a5b2a3d",
     "role": "character_ref"
    }
   ]
  },
  "shot_run_spend_attempt_count": 1,
  "judge_flip_skipped": "judge_fn_owns_order",
  "readings": [
   {
    "label": "A",
    "direction": "백정의 시선이 양손에 든 토기 사발을 향해 아래로 자연스럽게 향하고 있다.",
    "built_space": "야외 배식 공간과 커다란 가마솥 등 지시된 장소의 요소들을 잘 갖추고 있으나, 배경 우측과 오두막 근처에 프롬프트에 지시되지 않은 두 명의 인물이 묘사되어 있다.",
    "entities": "백정의 외모와 의상은 레퍼런스와 일치하며, 손에 든 토기 사발과 그 안의 멀건 죽 역시 잘 묘사되었다. 하지만 등장해서는 안 될 엑스트라 인물들이 존재한다.",
    "hard_violations": [
     "[gemini-pro] invented people (지시문에 없는 배경 인물 2명 임의 추가)",
     "[gpt-high] 백정만 보여야 하는 장면에 배경 인물 두 명을 추가했다."
    ],
    "physics": "양손을 사용해 사발의 무게를 안정적으로 받치고 들고 있다."
   },
   {
    "label": "B",
    "direction": "백정의 퀭한 시선이 자신의 토기 사발 안의 멀건 죽을 향해 정확히 아래로 향하고 있다.",
    "built_space": "레퍼런스의 야외 배식 공간을 정확히 재현했으며, 지시문에 맞게 뒤편의 가마솥 뚜껑이 열려 바닥에 기대어 있는 모습이 묘사되었다. 추가된 인물 없이 지정된 인물만 존재한다.",
    "entities": "백정의 얼굴과 낡은 의상이 레퍼런스와 잘 일치하며(머리띠는 제외됨), 질감이 살아있는 낡은 토기 사발과 묽은 죽이 프롬프트의 묘사와 부합한다.",
    "hard_violations": [],
    "physics": "두 손으로 사발을 감싸 쥐고 안정적으로 공중에 지탱하고 있다."
   }
  ],
  "cross_model_order": {
   "policy": "cross_model_order_v2_slot0_fwd_slot1_rev_parallel2",
   "slots": [
    {
     "model": "gemini-pro",
     "order": "forward",
     "display_to_canonical": {
      "A": "A",
      "B": "B"
     },
     "raw": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 9,
        "verdict_ko": "지시된 클로즈업 앵글 속에서 멀건 죽이 담긴 사발을 허탈하게 내려다보는 백정의 표정과 시선, 뒤편에 뚜껑이 열린 가마솥까지 프롬프트의 요구사항을 정확하고 충실하게 구현했습니다."
       },
       {
        "label": "A",
        "score": 3,
        "verdict_ko": "백정의 묘사와 들고 있는 사발의 표현은 훌륭하나, 샷 텍스트에 없는 인물들을 배경에 임의로 추가하여 엄격한 프롬프트 규칙을 위반했습니다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "백정의 시선이 양손에 든 토기 사발을 향해 아래로 자연스럽게 향하고 있다.",
        "built_space": "야외 배식 공간과 커다란 가마솥 등 지시된 장소의 요소들을 잘 갖추고 있으나, 배경 우측과 오두막 근처에 프롬프트에 지시되지 않은 두 명의 인물이 묘사되어 있다.",
        "entities": "백정의 외모와 의상은 레퍼런스와 일치하며, 손에 든 토기 사발과 그 안의 멀건 죽 역시 잘 묘사되었다. 하지만 등장해서는 안 될 엑스트라 인물들이 존재한다.",
        "hard_violations": [
         "invented people (지시문에 없는 배경 인물 2명 임의 추가)"
        ],
        "physics": "양손을 사용해 사발의 무게를 안정적으로 받치고 들고 있다."
       },
       {
        "label": "B",
        "direction": "백정의 퀭한 시선이 자신의 토기 사발 안의 멀건 죽을 향해 정확히 아래로 향하고 있다.",
        "built_space": "레퍼런스의 야외 배식 공간을 정확히 재현했으며, 지시문에 맞게 뒤편의 가마솥 뚜껑이 열려 바닥에 기대어 있는 모습이 묘사되었다. 추가된 인물 없이 지정된 인물만 존재한다.",
        "entities": "백정의 얼굴과 낡은 의상이 레퍼런스와 잘 일치하며(머리띠는 제외됨), 질감이 살아있는 낡은 토기 사발과 묽은 죽이 프롬프트의 묘사와 부합한다.",
        "hard_violations": [],
        "physics": "두 손으로 사발을 감싸 쥐고 안정적으로 공중에 지탱하고 있다."
       }
      ],
      "all_candidates_fail": false
     },
     "normalized": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 9,
        "verdict_ko": "지시된 클로즈업 앵글 속에서 멀건 죽이 담긴 사발을 허탈하게 내려다보는 백정의 표정과 시선, 뒤편에 뚜껑이 열린 가마솥까지 프롬프트의 요구사항을 정확하고 충실하게 구현했습니다."
       },
       {
        "label": "A",
        "score": 3,
        "verdict_ko": "백정의 묘사와 들고 있는 사발의 표현은 훌륭하나, 샷 텍스트에 없는 인물들을 배경에 임의로 추가하여 엄격한 프롬프트 규칙을 위반했습니다."
       }
      ],
      "all_candidates_fail": false,
      "readings": [
       {
        "label": "A",
        "direction": "백정의 시선이 양손에 든 토기 사발을 향해 아래로 자연스럽게 향하고 있다.",
        "built_space": "야외 배식 공간과 커다란 가마솥 등 지시된 장소의 요소들을 잘 갖추고 있으나, 배경 우측과 오두막 근처에 프롬프트에 지시되지 않은 두 명의 인물이 묘사되어 있다.",
        "entities": "백정의 외모와 의상은 레퍼런스와 일치하며, 손에 든 토기 사발과 그 안의 멀건 죽 역시 잘 묘사되었다. 하지만 등장해서는 안 될 엑스트라 인물들이 존재한다.",
        "hard_violations": [
         "invented people (지시문에 없는 배경 인물 2명 임의 추가)"
        ],
        "physics": "양손을 사용해 사발의 무게를 안정적으로 받치고 들고 있다."
       },
       {
        "label": "B",
        "direction": "백정의 퀭한 시선이 자신의 토기 사발 안의 멀건 죽을 향해 정확히 아래로 향하고 있다.",
        "built_space": "레퍼런스의 야외 배식 공간을 정확히 재현했으며, 지시문에 맞게 뒤편의 가마솥 뚜껑이 열려 바닥에 기대어 있는 모습이 묘사되었다. 추가된 인물 없이 지정된 인물만 존재한다.",
        "entities": "백정의 얼굴과 낡은 의상이 레퍼런스와 잘 일치하며(머리띠는 제외됨), 질감이 살아있는 낡은 토기 사발과 묽은 죽이 프롬프트의 묘사와 부합한다.",
        "hard_violations": [],
        "physics": "두 손으로 사발을 감싸 쥐고 안정적으로 공중에 지탱하고 있다."
       }
      ]
     },
     "ok": true
    },
    {
     "model": "gpt-high",
     "order": "reverse",
     "display_to_canonical": {
      "A": "B",
      "B": "A"
     },
     "raw": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 8,
        "verdict_ko": "추가 인물 없이 백정의 허탈한 하향 시선과 묽은 죽을 클로즈업으로 연결하지만, 사발의 비중이 다소 크고 참조의 머리쓰개가 빠졌다."
       },
       {
        "label": "B",
        "score": 2,
        "verdict_ko": "사발을 내려다보는 행동과 머리쓰개는 맞지만, 금지된 배경 인물 두 명을 추가해 장면의 필수 조건을 위반한다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "백정의 두 눈은 렌즈가 아니라 얼굴 아래 사발의 내부를 향한다. 고개는 조금 숙였고 처진 눈꺼풀과 이완된 입이 허탈함을 표현한다. 사발 입구는 위로 열려 있어 내려다보는 카메라에도 묽은 죽이 보인다.",
        "built_space": "흙마당 뒤로 왼쪽 움막, 중앙의 작은 차양, 오른쪽 긴 초가지붕 건물과 목책이 보여 참조 장소의 배치를 대체로 유지한다. 오른쪽에는 열린 솥 하나와 그 옆에 비스듬히 기대 놓은 뚜껑 하나가 보인다. 백정은 그 앞에 있으며 솥이나 건물과 충돌하지 않는다. 솥 내부는 어두워 죽이 담겼는지는 확인되지 않는다.",
        "entities": "성인 한국인 남성으로 설정에 부합하는 외형의 인물 한 명만 보인다. 검은 머리와 해진 어두운 겹옷은 맞지만, 참조의 검은 머리쓰개가 없고 얼굴 인상도 완전히 같지는 않다. 갈색 토기 사발 안에는 물기가 많고 낟알이 드문 죽이 있다. 사발은 얼굴 아래에 있으나 작은 하단 보조 요소라기에는 화면 비중이 크다. 눈은 정상적인 홍채와 동공을 유지하며 읽을 수 있는 글자는 없다.",
        "hard_violations": [],
        "physics": "사발 아래 화면 가장자리에서 손 일부가 보이며 그릇 밑을 받친다. 죽은 그릇 바닥에 수평에 가까운 액면으로 고여 있다. 솥은 바닥의 돌에 지지되고 뚜껑은 지면과 솥 옆에 기대어 있다. 하체는 프레임 밖이므로 발의 접지는 확인할 수 없지만, 보이는 상체에 부유하거나 불가능한 자세는 없다."
       },
       {
        "label": "B",
        "direction": "백정은 고개와 눈을 아래로 향해 양손에 든 사발의 죽을 바라본다. 사발 입구는 위를 향하고 카메라에서도 내부가 보인다. 배경 중앙 남성은 자신이 든 그릇 쪽을, 오른쪽 남성은 손앞 작업면 쪽을 내려다본다.",
        "built_space": "흙마당, 목책, 왼쪽 움막 일부, 중앙 차양과 오른쪽 초가 건물이 참조 장소를 따른다. 중경에 흙 화덕이 있고 오른쪽에는 큰 금속 뚜껑 하나가 돌에 기대어 있다. 오른쪽 전경에는 토기 사발이 들어 있는 큰 금속 용기가 별도로 놓여 화면을 차지한다. 배경 중앙의 남성 한 명은 마당에 서 있고 다른 한 명은 오른쪽 건물 앞 작업대에 서 있다. 열린 큰 솥에 담긴 죽은 명확히 보이지 않는다.",
        "entities": "백정 외에 배경 성인 남성 두 명이 추가되어 허용 인물 수를 위반한다. 백정의 검은 머리쓰개, 검은 머리, 해진 어두운 옷과 얼굴 인상은 참조에 비교적 가깝다. 양손에 든 토기 사발에는 묽고 낟알이 적은 죽이 보인다. 정상적인 눈과 자연스러운 피로 표정이 있으며 읽을 수 있는 글자는 없다. 낮이지만 얼굴과 배경의 직사광선 대비는 요청한 억제된 조명보다 강하다.",
        "hard_violations": [
         "백정만 보여야 하는 장면에 배경 인물 두 명을 추가했다."
        ],
        "physics": "백정은 양손으로 사발의 양옆과 밑을 받치고 있어 지지가 명확하며 액체도 자연스럽게 고여 있다. 큰 뚜껑은 지면의 돌에 기대고, 전경 금속 용기는 낮은 받침 위에 놓여 있으며 그 안의 사발도 용기 내부에 지지된다. 배경 중앙 남성은 발을 땅에 딛고 손으로 그릇을 들고 있다. 오른쪽 남성은 하체 일부가 가려졌지만 작업면에 손을 댄 상체 자세는 가능하다."
       }
      ],
      "all_candidates_fail": false
     },
     "normalized": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 8,
        "verdict_ko": "추가 인물 없이 백정의 허탈한 하향 시선과 묽은 죽을 클로즈업으로 연결하지만, 사발의 비중이 다소 크고 참조의 머리쓰개가 빠졌다."
       },
       {
        "label": "A",
        "score": 2,
        "verdict_ko": "사발을 내려다보는 행동과 머리쓰개는 맞지만, 금지된 배경 인물 두 명을 추가해 장면의 필수 조건을 위반한다."
       }
      ],
      "all_candidates_fail": false,
      "readings": [
       {
        "label": "B",
        "direction": "백정의 두 눈은 렌즈가 아니라 얼굴 아래 사발의 내부를 향한다. 고개는 조금 숙였고 처진 눈꺼풀과 이완된 입이 허탈함을 표현한다. 사발 입구는 위로 열려 있어 내려다보는 카메라에도 묽은 죽이 보인다.",
        "built_space": "흙마당 뒤로 왼쪽 움막, 중앙의 작은 차양, 오른쪽 긴 초가지붕 건물과 목책이 보여 참조 장소의 배치를 대체로 유지한다. 오른쪽에는 열린 솥 하나와 그 옆에 비스듬히 기대 놓은 뚜껑 하나가 보인다. 백정은 그 앞에 있으며 솥이나 건물과 충돌하지 않는다. 솥 내부는 어두워 죽이 담겼는지는 확인되지 않는다.",
        "entities": "성인 한국인 남성으로 설정에 부합하는 외형의 인물 한 명만 보인다. 검은 머리와 해진 어두운 겹옷은 맞지만, 참조의 검은 머리쓰개가 없고 얼굴 인상도 완전히 같지는 않다. 갈색 토기 사발 안에는 물기가 많고 낟알이 드문 죽이 있다. 사발은 얼굴 아래에 있으나 작은 하단 보조 요소라기에는 화면 비중이 크다. 눈은 정상적인 홍채와 동공을 유지하며 읽을 수 있는 글자는 없다.",
        "hard_violations": [],
        "physics": "사발 아래 화면 가장자리에서 손 일부가 보이며 그릇 밑을 받친다. 죽은 그릇 바닥에 수평에 가까운 액면으로 고여 있다. 솥은 바닥의 돌에 지지되고 뚜껑은 지면과 솥 옆에 기대어 있다. 하체는 프레임 밖이므로 발의 접지는 확인할 수 없지만, 보이는 상체에 부유하거나 불가능한 자세는 없다."
       },
       {
        "label": "A",
        "direction": "백정은 고개와 눈을 아래로 향해 양손에 든 사발의 죽을 바라본다. 사발 입구는 위를 향하고 카메라에서도 내부가 보인다. 배경 중앙 남성은 자신이 든 그릇 쪽을, 오른쪽 남성은 손앞 작업면 쪽을 내려다본다.",
        "built_space": "흙마당, 목책, 왼쪽 움막 일부, 중앙 차양과 오른쪽 초가 건물이 참조 장소를 따른다. 중경에 흙 화덕이 있고 오른쪽에는 큰 금속 뚜껑 하나가 돌에 기대어 있다. 오른쪽 전경에는 토기 사발이 들어 있는 큰 금속 용기가 별도로 놓여 화면을 차지한다. 배경 중앙의 남성 한 명은 마당에 서 있고 다른 한 명은 오른쪽 건물 앞 작업대에 서 있다. 열린 큰 솥에 담긴 죽은 명확히 보이지 않는다.",
        "entities": "백정 외에 배경 성인 남성 두 명이 추가되어 허용 인물 수를 위반한다. 백정의 검은 머리쓰개, 검은 머리, 해진 어두운 옷과 얼굴 인상은 참조에 비교적 가깝다. 양손에 든 토기 사발에는 묽고 낟알이 적은 죽이 보인다. 정상적인 눈과 자연스러운 피로 표정이 있으며 읽을 수 있는 글자는 없다. 낮이지만 얼굴과 배경의 직사광선 대비는 요청한 억제된 조명보다 강하다.",
        "hard_violations": [
         "백정만 보여야 하는 장면에 배경 인물 두 명을 추가했다."
        ],
        "physics": "백정은 양손으로 사발의 양옆과 밑을 받치고 있어 지지가 명확하며 액체도 자연스럽게 고여 있다. 큰 뚜껑은 지면의 돌에 기대고, 전경 금속 용기는 낮은 받침 위에 놓여 있으며 그 안의 사발도 용기 내부에 지지된다. 배경 중앙 남성은 발을 땅에 딛고 손으로 그릇을 들고 있다. 오른쪽 남성은 하체 일부가 가려졌지만 작업면에 손을 댄 상체 자세는 가능하다."
       }
      ]
     },
     "ok": true
    }
   ],
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "slot_winner_match": true,
   "slot_winner": {
    "gemini-pro": "B",
    "gpt-high": "B"
   },
   "route": "cross_slot_agree"
  },
  "dual": {
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "normalized": {
    "A": 0.583,
    "B": 2.0
   },
   "adjusted": {
    "A": 0.333,
    "B": 2.0
   },
   "violations": {
    "A": [
     "[gemini-pro] invented people (지시문에 없는 배경 인물 2명 임의 추가)",
     "[gpt-high] 백정만 보여야 하는 장면에 배경 인물 두 명을 추가했다."
    ]
   },
   "per_model_winner": {
    "gemini-pro": "B",
    "gpt-high": "B"
   },
   "agreed": true
  },
  "totals": {
   "B": 2000,
   "A": 333
  },
  "selected": "B",
  "ranking": [
   "B",
   "A"
  ],
  "verdicts": [
   {
    "label": "B",
    "score": 2000,
    "verdict_ko": "지시된 클로즈업 앵글 속에서 멀건 죽이 담긴 사발을 허탈하게 내려다보는 백정의 표정과 시선, 뒤편에 뚜껑이 열린 가마솥까지 프롬프트의 요구사항을 정확하고 충실하게 구현했습니다."
   },
   {
    "label": "A",
    "score": 333,
    "verdict_ko": "백정의 묘사와 들고 있는 사발의 표현은 훌륭하나, 샷 텍스트에 없는 인물들을 배경에 임의로 추가하여 엄격한 프롬프트 규칙을 위반했습니다.  ★위반: [gemini-pro] invented people (지시문에 없는 배경 인물 2명 임의 추가) / [gpt-high] 백정만 보여야 하는 장면에 배경 인물 두 명을 추가했다."
   }
  ],
  "refs": [
   {
    "label": "LOCATION PHOTOGRAPH — the exact place of this shot: its architecture, materials, fixed features and lighting mood are spatial truth; stage the moment inside this place. Never copy its camera framing.",
    "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/db00517b-165b-4005-a287-f28936b59b95/images/272bd7af-9294-4490-98c6-88a40e9deacc/scene/recipe/groupbg_mountain_camp_hearth_81c95f.png",
    "asset_id": "87bcdd2b-607c-40d6-8d7f-f4dfbf3ed054",
    "role": "bgfirst_group_bg"
   },
   {
    "label": "CHARACTER REFERENCE — 백정: the exact person appearing in this shot; match face, hair and build exactly.",
    "path": "<bytes:980360>",
    "asset_id": "29282584-ae7c-4722-8d77-100c4a5b2a3d",
    "role": "character_ref"
   }
  ],
  "critique_skipped": true,
  "shot_run_produced": false,
  "shot_run_uid": "06aa1423-058e-7347-b688-ef71491d08ee",
  "bgfirst": {
   "bg_path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/db00517b-165b-4005-a287-f28936b59b95/images/272bd7af-9294-4490-98c6-88a40e9deacc/scene/recipe/S27sh21__bgfirst_bg.png",
   "bg_asset_id": "e40cc56e-348b-46ad-a7ae-b303c53153e0",
   "bg_record_key": "S27sh21::bgfirst_bg",
   "chain_winner": false,
   "authority": "groupbg",
   "group_key": "mountain_camp_hearth",
   "groupbg_asset_id": "87bcdd2b-607c-40d6-8d7f-f4dfbf3ed054"
  },
  "ref_mode": "그룹 배경+엔티티 (2택1: 무콘티 승)",
  "share_plan": {
   "ref_plan": "background"
  }
 },
 "S27sh21::cine": {
  "applied": true,
  "attempted_at": "2026-09-09T04:01:13.465292+00:00",
  "fingerprint": "28846089409ff24a3d498b17111c64343b043ccbda460cabe19a11a87e16eee0",
  "fingerprint_version": 2,
  "provider": "mai",
  "endpoint": "openrouter/chat-completions",
  "model": "microsoft/mai-image-2.6",
  "pack": "24.202608252115",
  "source_file": "S27sh21_sel.png",
  "source_sha256": "f110948c8e07ee2742f26e111eba7d09015fc3962f82185dea8a49e447aa5d5e",
  "file": "S27sh21_cine.png",
  "staged_sha256": "65cc7aa5ab5984443c5f98b1a9cad5456e7fa7ed98ee6ab80ce93310a3cfa030",
  "latency_ms": 27595
 },
 "S27sh26::signage": {
  "fp": "679a904cd1604f43",
  "inscriptions": [],
  "cues": [],
  "dropped": []
 },
 "S27sh26": {
  "input_fingerprint": "eadf43d094e40ae5",
  "prompt": "Create ONE FINAL photorealistic live-action film still of the moment below — early 20th-century Korea, Jeolla region; all people are Korean unless explicitly stated otherwise. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 깊은 근심에 잠겨 미간이 깊게 팬 안규홍의 무거운 정면 클로즈업.\n\nLOCATION (lock): The edge of the mountain camp's outdoor meal area, slightly apart from the men eating beside the cooking pot. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: close-up\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Maintain natural daytime illumination with restrained saturation and sober facial contrast, allowing the burden to emerge without a dramatic lighting change.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The people visible in that photo are NOT in this shot: never carry their faces, bodies or clothing onto anyone here. Each person in THIS shot is defined solely by the PEOPLE list, their own CHARACTER REFERENCE images and their wardrobe notes. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The open cauldron contains watery porridge; the provisions ledger remains open at the rice calculation. 안규홍: He stands apart watching the meal, with the jade ring concealed.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 안규홍 (한국인 남성, 청년 나이대, 검은 머리, 단단히 묶은 머리) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.",
  "roll_prompts": {
   "A": "Create ONE FINAL photorealistic live-action film still of the moment below — early 20th-century Korea, Jeolla region; all people are Korean unless explicitly stated otherwise. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 깊은 근심에 잠겨 미간이 깊게 팬 안규홍의 무거운 정면 클로즈업.\n\nLOCATION (lock): The edge of the mountain camp's outdoor meal area, slightly apart from the men eating beside the cooking pot. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: close-up\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Maintain natural daytime illumination with restrained saturation and sober facial contrast, allowing the burden to emerge without a dramatic lighting change.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The people visible in that photo are NOT in this shot: never carry their faces, bodies or clothing onto anyone here. Each person in THIS shot is defined solely by the PEOPLE list, their own CHARACTER REFERENCE images and their wardrobe notes. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The open cauldron contains watery porridge; the provisions ledger remains open at the rice calculation. 안규홍: He stands apart watching the meal, with the jade ring concealed.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 안규홍 (한국인 남성, 청년 나이대, 검은 머리, 단단히 묶은 머리) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.",
   "B": "Create ONE FINAL photorealistic live-action film still of the moment below — early 20th-century Korea, Jeolla region; all people are Korean unless explicitly stated otherwise. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 깊은 근심에 잠겨 미간이 깊게 팬 안규홍의 무거운 정면 클로즈업.\n\nLOCATION (lock): The edge of the mountain camp's outdoor meal area, slightly apart from the men eating beside the cooking pot. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: close-up\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Maintain natural daytime illumination with restrained saturation and sober facial contrast, allowing the burden to emerge without a dramatic lighting change.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The people visible in that photo are NOT in this shot: never carry their faces, bodies or clothing onto anyone here. Each person in THIS shot is defined solely by the PEOPLE list, their own CHARACTER REFERENCE images and their wardrobe notes. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The open cauldron contains watery porridge; the provisions ledger remains open at the rice calculation. 안규홍: He stands apart watching the meal, with the jade ring concealed.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 안규홍 (한국인 남성, 청년 나이대, 검은 머리, 단단히 묶은 머리) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.\n\nCOMPOSITION VARIATION (this candidate only): keep the place, the people and the moment exactly as specified, and keep from the CAMERA & FRAME contract its subject scale, screen placement and key background elements exactly. For this candidate only, its camera angle and height are deliberately left open: choose a camera position distinctly different from the obvious one, as a film director picking a second setup on the same blocking. Stay true to any explicit framing the SHOT TEXT itself dictates."
  },
  "shot_run_spend_attempt_count": 1,
  "readings": [
   {
    "label": "A",
    "direction": "안규홍은 정면을 향하되 시선을 약간 아래로 두고 있으며, 배경의 인물들은 각자의 식사(그릇)에 시선을 고정하고 있음.",
    "built_space": "산채의 야외 식사 공간으로, 흙바닥 위에 큰 가마솥이 놓여 있고 뒤편에 초가 구조물과 목책이 배치됨.",
    "entities": "안규홍(한국인 남성)이 흰색 한복을 입고 등장하나 미간의 주름 묘사가 약함. 배경에 식사 중인 여러 남성들이 있으나 지시된 '장부'는 보이지 않음.",
    "hard_violations": [
     "[gpt-high] 안규홍만 허용된 숏에 최소 네 명의 식사 중인 배경 인물을 추가했다."
    ],
    "physics": "안규홍은 바닥에 딛고 서 있으며, 배경 인물들은 바닥에 앉아 손으로 그릇과 숟가락을 쥐고 있음. 가마솥과 소품들의 물리적 지지 상태는 정상적임."
   },
   {
    "label": "B",
    "direction": "안규홍은 정면으로 카메라 렌즈를 강렬하게 응시하고 있으며, 배경의 인물들은 각각 장부와 식사에 시선을 두고 있음.",
    "built_space": "산채의 야외 공간으로, 돌출된 화덕 위에 올려진 대형 가마솥과 배경의 초가 구조물, 목책 등이 이전 샷의 환경적 연속성을 잘 유지하며 배치됨.",
    "entities": "안규홍(한국인 남성)은 짙은 근심과 깊게 팬 미간을 잘 보여주며, 배경에는 식사 중인 남성과 무릎에 장부를 펼친 남성이 정확히 등장함.",
    "hard_violations": [
     "[gpt-high] 숏에 허용되지 않은 식사 중인 배경 남자 두 명을 추가했다.",
     "[gpt-high] 명시적으로 재사용을 금지한 이전 숏 인물의 얼굴과 짙은 누더기 의상을 주인공에게 옮겨 안규홍을 대체했다."
    ],
    "physics": "안규홍은 화면 중앙에 안정적으로 서 있으며, 배경 인물들은 바닥에 앉아 손으로 젓가락과 장부를 자연스럽게 지지하고 있고 가마솥은 돌 위에 제대로 안착됨."
   }
  ],
  "cross_model_order": {
   "policy": "cross_model_order_v2_slot0_fwd_slot1_rev_parallel2",
   "slots": [
    {
     "model": "gemini-pro",
     "order": "forward",
     "display_to_canonical": {
      "A": "A",
      "B": "B"
     },
     "raw": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 9,
        "verdict_ko": "지시된 '깊게 팬 미간'과 '정면 클로즈업' 프레이밍을 매우 사실적이고 강렬하게 구현했으며, 배경에 '펼쳐진 장부'를 배치해 디테일한 묘사까지 훌륭히 살렸습니다."
       },
       {
        "label": "A",
        "score": 7,
        "verdict_ko": "주요 인물의 표정과 프레이밍이 다소 평이하고 넓게 잡혔으며, 프롬프트에 명시된 소품인 '장부'가 배경에 전혀 묘사되지 않았습니다."
       }
      ],
      "readings": [
       {
        "label": "B",
        "direction": "안규홍은 정면으로 카메라 렌즈를 강렬하게 응시하고 있으며, 배경의 인물들은 각각 장부와 식사에 시선을 두고 있음.",
        "built_space": "산채의 야외 공간으로, 돌출된 화덕 위에 올려진 대형 가마솥과 배경의 초가 구조물, 목책 등이 이전 샷의 환경적 연속성을 잘 유지하며 배치됨.",
        "entities": "안규홍(한국인 남성)은 짙은 근심과 깊게 팬 미간을 잘 보여주며, 배경에는 식사 중인 남성과 무릎에 장부를 펼친 남성이 정확히 등장함.",
        "hard_violations": [],
        "physics": "안규홍은 화면 중앙에 안정적으로 서 있으며, 배경 인물들은 바닥에 앉아 손으로 젓가락과 장부를 자연스럽게 지지하고 있고 가마솥은 돌 위에 제대로 안착됨."
       },
       {
        "label": "A",
        "direction": "안규홍은 정면을 향하되 시선을 약간 아래로 두고 있으며, 배경의 인물들은 각자의 식사(그릇)에 시선을 고정하고 있음.",
        "built_space": "산채의 야외 식사 공간으로, 흙바닥 위에 큰 가마솥이 놓여 있고 뒤편에 초가 구조물과 목책이 배치됨.",
        "entities": "안규홍(한국인 남성)이 흰색 한복을 입고 등장하나 미간의 주름 묘사가 약함. 배경에 식사 중인 여러 남성들이 있으나 지시된 '장부'는 보이지 않음.",
        "hard_violations": [],
        "physics": "안규홍은 바닥에 딛고 서 있으며, 배경 인물들은 바닥에 앉아 손으로 그릇과 숟가락을 쥐고 있음. 가마솥과 소품들의 물리적 지지 상태는 정상적임."
       }
      ],
      "all_candidates_fail": false
     },
     "normalized": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 9,
        "verdict_ko": "지시된 '깊게 팬 미간'과 '정면 클로즈업' 프레이밍을 매우 사실적이고 강렬하게 구현했으며, 배경에 '펼쳐진 장부'를 배치해 디테일한 묘사까지 훌륭히 살렸습니다."
       },
       {
        "label": "A",
        "score": 7,
        "verdict_ko": "주요 인물의 표정과 프레이밍이 다소 평이하고 넓게 잡혔으며, 프롬프트에 명시된 소품인 '장부'가 배경에 전혀 묘사되지 않았습니다."
       }
      ],
      "all_candidates_fail": false,
      "readings": [
       {
        "label": "B",
        "direction": "안규홍은 정면으로 카메라 렌즈를 강렬하게 응시하고 있으며, 배경의 인물들은 각각 장부와 식사에 시선을 두고 있음.",
        "built_space": "산채의 야외 공간으로, 돌출된 화덕 위에 올려진 대형 가마솥과 배경의 초가 구조물, 목책 등이 이전 샷의 환경적 연속성을 잘 유지하며 배치됨.",
        "entities": "안규홍(한국인 남성)은 짙은 근심과 깊게 팬 미간을 잘 보여주며, 배경에는 식사 중인 남성과 무릎에 장부를 펼친 남성이 정확히 등장함.",
        "hard_violations": [],
        "physics": "안규홍은 화면 중앙에 안정적으로 서 있으며, 배경 인물들은 바닥에 앉아 손으로 젓가락과 장부를 자연스럽게 지지하고 있고 가마솥은 돌 위에 제대로 안착됨."
       },
       {
        "label": "A",
        "direction": "안규홍은 정면을 향하되 시선을 약간 아래로 두고 있으며, 배경의 인물들은 각자의 식사(그릇)에 시선을 고정하고 있음.",
        "built_space": "산채의 야외 식사 공간으로, 흙바닥 위에 큰 가마솥이 놓여 있고 뒤편에 초가 구조물과 목책이 배치됨.",
        "entities": "안규홍(한국인 남성)이 흰색 한복을 입고 등장하나 미간의 주름 묘사가 약함. 배경에 식사 중인 여러 남성들이 있으나 지시된 '장부'는 보이지 않음.",
        "hard_violations": [],
        "physics": "안규홍은 바닥에 딛고 서 있으며, 배경 인물들은 바닥에 앉아 손으로 그릇과 숟가락을 쥐고 있음. 가마솥과 소품들의 물리적 지지 상태는 정상적임."
       }
      ]
     },
     "ok": true
    },
    {
     "model": "gpt-high",
     "order": "reverse",
     "display_to_canonical": {
      "A": "B",
      "B": "A"
     },
     "raw": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 1,
        "verdict_ko": "정면의 깊은 미간 주름과 밀착 클로즈업은 맞지만, 안규홍 대신 이전 숏 인물의 얼굴·옷을 재사용하고 금지된 배경 인물까지 추가했다."
       },
       {
        "label": "B",
        "score": 3,
        "verdict_ko": "안규홍의 인상·묶은 머리·밝은 옷과 근심 어린 정면 표정은 더 충실하지만, 구도가 다소 넓고 금지된 식사 인물들을 추가하여 재촬영이 필요하다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "주인공은 얼굴을 정면으로 두고 시선을 렌즈 부근의 약간 아래로 향한다. 뒤쪽 두 남자는 각자 손에 든 그릇을 내려다본다. 주인공이 뒤의 식사를 지켜보는 시선 관계는 드러나지 않지만, 무거운 정면 표정과 깊게 접힌 미간은 명확하다.",
        "built_space": "흙바닥, 뒤쪽 목책, 좌우 초가 건물과 산이 보인다. 오른쪽에는 열린 가마솥 한 개와 비스듬한 뚜껑 한 개, 하부 받침돌이 있으며 장소의 재료와 낮빛은 이전 숏과 유사하다. 주인공 뒤 가마솥 양옆에 식사하는 남자 두 명이 배치되어 있다. 반사상이나 중복 가마솥은 없다.",
        "entities": "전경에는 검은 머리의 성인 동아시아계 남성 한 명이 있으나, 얼굴과 해진 짙은 옷이 안규홍 참조보다 이전 숏 남성과 일치한다. 안규홍의 밝은 저고리와 단단히 묶은 머리는 재현되지 않았다. 배경 남자 두 명은 출연이 허용되지 않은 인물이다. 그릇과 식사 도구, 열린 가마솥, 펼친 장부처럼 보이는 물체가 보이지만 솥 속 죽과 장부 내용은 판별되지 않는다. 옥가락지는 노출되지 않고 읽을 수 있는 글자는 없다.",
        "hard_violations": [
         "숏에 허용되지 않은 식사 중인 배경 남자 두 명을 추가했다.",
         "명시적으로 재사용을 금지한 이전 숏 인물의 얼굴과 짙은 누더기 의상을 주인공에게 옮겨 안규홍을 대체했다."
        ],
        "physics": "전경 인물의 머리와 어깨 연결은 자연스럽고 하체는 클로즈업 밖이다. 배경 남자들은 지면 가까이 앉거나 웅크린 자세이며 그릇과 식사 도구를 손으로 잡고 있다. 가마솥은 돌 위에 놓이고 뚜껑은 솥 옆에 기대어 있다. 명확히 떠 있거나 지지 없는 물체는 없다."
       },
       {
        "label": "B",
        "direction": "안규홍은 정면을 향한 채 눈을 약간 아래로 내리깔고 미간을 찌푸린다. 렌즈를 멍하게 응시하기보다 생각에 잠긴 표정이다. 뒤쪽에서 얼굴이 식별되는 네 남자는 각각 그릇이나 입으로 가져가는 식사 도구를 향한다. 안규홍의 시선이 뒤쪽 식사에 직접 닿는 모습은 아니다.",
        "built_space": "흙마당 뒤로 목책과 산, 중앙 왼쪽 초가 한 채, 오른쪽 긴 초가 한 채 및 왼쪽 끝 지붕 일부가 보인다. 오른쪽 아래 열린 가마솥 한 개가 돌 받침 위에 놓여 있다. 솥뚜껑은 확인되지 않는다. 주인공을 중앙 전경에 두고 식사 인물들을 바로 뒤에 배열해, 홀로 떨어진 인상은 약하다. 머리부터 가슴까지 포함하여 A보다 얼굴 밀착도가 낮다.",
        "entities": "주인공은 청년 동아시아계 남성으로, 참조 안규홍과 유사한 얼굴 윤곽, 검은 묶은 머리, 때 묻은 밝은 교차깃 저고리를 갖춘다. 이전 숏의 짙은 누더기 옷을 주인공에게 옮기지 않았다. 배경에는 최소 네 명의 별도 식사 인물과 각자의 그릇·도구가 있어 인물 제한을 위반한다. 가마솥 속 죽과 장부는 확인되지 않지만 이 클로즈업에서 반드시 노출할 필요는 없다. 옥가락지와 읽을 수 있는 글자는 보이지 않는다.",
        "hard_violations": [
         "안규홍만 허용된 숏에 최소 네 명의 식사 중인 배경 인물을 추가했다."
        ],
        "physics": "안규홍의 목과 상체는 자연스럽게 이어지고 하체는 프레임 밖이다. 배경 인물들은 지면 가까이 낮게 앉은 자세이며 그릇을 손으로 받치고 다른 손으로 식사 도구를 잡는다. 가마솥 아래에는 돌 지지가 보인다. 공중에 지지 없이 떠 있는 신체나 물체, 명백히 불가능한 자세는 없다."
       }
      ],
      "all_candidates_fail": true
     },
     "normalized": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 1,
        "verdict_ko": "정면의 깊은 미간 주름과 밀착 클로즈업은 맞지만, 안규홍 대신 이전 숏 인물의 얼굴·옷을 재사용하고 금지된 배경 인물까지 추가했다."
       },
       {
        "label": "A",
        "score": 3,
        "verdict_ko": "안규홍의 인상·묶은 머리·밝은 옷과 근심 어린 정면 표정은 더 충실하지만, 구도가 다소 넓고 금지된 식사 인물들을 추가하여 재촬영이 필요하다."
       }
      ],
      "all_candidates_fail": true,
      "readings": [
       {
        "label": "B",
        "direction": "주인공은 얼굴을 정면으로 두고 시선을 렌즈 부근의 약간 아래로 향한다. 뒤쪽 두 남자는 각자 손에 든 그릇을 내려다본다. 주인공이 뒤의 식사를 지켜보는 시선 관계는 드러나지 않지만, 무거운 정면 표정과 깊게 접힌 미간은 명확하다.",
        "built_space": "흙바닥, 뒤쪽 목책, 좌우 초가 건물과 산이 보인다. 오른쪽에는 열린 가마솥 한 개와 비스듬한 뚜껑 한 개, 하부 받침돌이 있으며 장소의 재료와 낮빛은 이전 숏과 유사하다. 주인공 뒤 가마솥 양옆에 식사하는 남자 두 명이 배치되어 있다. 반사상이나 중복 가마솥은 없다.",
        "entities": "전경에는 검은 머리의 성인 동아시아계 남성 한 명이 있으나, 얼굴과 해진 짙은 옷이 안규홍 참조보다 이전 숏 남성과 일치한다. 안규홍의 밝은 저고리와 단단히 묶은 머리는 재현되지 않았다. 배경 남자 두 명은 출연이 허용되지 않은 인물이다. 그릇과 식사 도구, 열린 가마솥, 펼친 장부처럼 보이는 물체가 보이지만 솥 속 죽과 장부 내용은 판별되지 않는다. 옥가락지는 노출되지 않고 읽을 수 있는 글자는 없다.",
        "hard_violations": [
         "숏에 허용되지 않은 식사 중인 배경 남자 두 명을 추가했다.",
         "명시적으로 재사용을 금지한 이전 숏 인물의 얼굴과 짙은 누더기 의상을 주인공에게 옮겨 안규홍을 대체했다."
        ],
        "physics": "전경 인물의 머리와 어깨 연결은 자연스럽고 하체는 클로즈업 밖이다. 배경 남자들은 지면 가까이 앉거나 웅크린 자세이며 그릇과 식사 도구를 손으로 잡고 있다. 가마솥은 돌 위에 놓이고 뚜껑은 솥 옆에 기대어 있다. 명확히 떠 있거나 지지 없는 물체는 없다."
       },
       {
        "label": "A",
        "direction": "안규홍은 정면을 향한 채 눈을 약간 아래로 내리깔고 미간을 찌푸린다. 렌즈를 멍하게 응시하기보다 생각에 잠긴 표정이다. 뒤쪽에서 얼굴이 식별되는 네 남자는 각각 그릇이나 입으로 가져가는 식사 도구를 향한다. 안규홍의 시선이 뒤쪽 식사에 직접 닿는 모습은 아니다.",
        "built_space": "흙마당 뒤로 목책과 산, 중앙 왼쪽 초가 한 채, 오른쪽 긴 초가 한 채 및 왼쪽 끝 지붕 일부가 보인다. 오른쪽 아래 열린 가마솥 한 개가 돌 받침 위에 놓여 있다. 솥뚜껑은 확인되지 않는다. 주인공을 중앙 전경에 두고 식사 인물들을 바로 뒤에 배열해, 홀로 떨어진 인상은 약하다. 머리부터 가슴까지 포함하여 A보다 얼굴 밀착도가 낮다.",
        "entities": "주인공은 청년 동아시아계 남성으로, 참조 안규홍과 유사한 얼굴 윤곽, 검은 묶은 머리, 때 묻은 밝은 교차깃 저고리를 갖춘다. 이전 숏의 짙은 누더기 옷을 주인공에게 옮기지 않았다. 배경에는 최소 네 명의 별도 식사 인물과 각자의 그릇·도구가 있어 인물 제한을 위반한다. 가마솥 속 죽과 장부는 확인되지 않지만 이 클로즈업에서 반드시 노출할 필요는 없다. 옥가락지와 읽을 수 있는 글자는 보이지 않는다.",
        "hard_violations": [
         "안규홍만 허용된 숏에 최소 네 명의 식사 중인 배경 인물을 추가했다."
        ],
        "physics": "안규홍의 목과 상체는 자연스럽게 이어지고 하체는 프레임 밖이다. 배경 인물들은 지면 가까이 낮게 앉은 자세이며 그릇을 손으로 받치고 다른 손으로 식사 도구를 잡는다. 가마솥 아래에는 돌 지지가 보인다. 공중에 지지 없이 떠 있는 신체나 물체, 명백히 불가능한 자세는 없다."
       }
      ]
     },
     "ok": true
    }
   ],
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "slot_winner_match": false,
   "slot_winner": {
    "gemini-pro": "B",
    "gpt-high": "A"
   },
   "route": "cross_slot_combined"
  },
  "dual": {
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "normalized": {
    "A": 1.778,
    "B": 1.333
   },
   "adjusted": {
    "A": 1.528,
    "B": 1.083
   },
   "violations": {
    "B": [
     "[gpt-high] 숏에 허용되지 않은 식사 중인 배경 남자 두 명을 추가했다.",
     "[gpt-high] 명시적으로 재사용을 금지한 이전 숏 인물의 얼굴과 짙은 누더기 의상을 주인공에게 옮겨 안규홍을 대체했다."
    ],
    "A": [
     "[gpt-high] 안규홍만 허용된 숏에 최소 네 명의 식사 중인 배경 인물을 추가했다."
    ]
   },
   "per_model_winner": {
    "gemini-pro": "B",
    "gpt-high": "A"
   },
   "agreed": false
  },
  "totals": {
   "B": 1083,
   "A": 1528
  },
  "selected": "A",
  "ranking": [
   "A",
   "B"
  ],
  "verdicts": [
   {
    "label": "B",
    "score": 1083,
    "verdict_ko": "지시된 '깊게 팬 미간'과 '정면 클로즈업' 프레이밍을 매우 사실적이고 강렬하게 구현했으며, 배경에 '펼쳐진 장부'를 배치해 디테일한 묘사까지 훌륭히 살렸습니다.  ★위반: [gpt-high] 숏에 허용되지 않은 식사 중인 배경 남자 두 명을 추가했다. / [gpt-high] 명시적으로 재사용을 금지한 이전 숏 인물의 얼굴과 짙은 누더기 의상을 주인공에게 옮겨 안규홍을 대체했다."
   },
   {
    "label": "A",
    "score": 1528,
    "verdict_ko": "주요 인물의 표정과 프레이밍이 다소 평이하고 넓게 잡혔으며, 프롬프트에 명시된 소품인 '장부'가 배경에 전혀 묘사되지 않았습니다.  ★위반: [gpt-high] 안규홍만 허용된 숏에 최소 네 명의 식사 중인 배경 인물을 추가했다."
   }
  ],
  "refs": [
   {
    "label": "PREVIOUS SHOT STILL — a visually related earlier shot of this same place: the location's look, materials, fixed features and lighting mood are LOCKED to this photo; never copy its camera framing. The people visible in that photo are NOT in this shot: never carry their faces, bodies or clothing onto anyone here. Each person in THIS shot is defined solely by the PEOPLE list, their own CHARACTER REFERENCE images and their wardrobe notes. If this photo shows a character who cannot move (dead or unconscious), that character's exact pose, position and orientation are ALSO LOCKED — treat their body as a fixed prop of the set that only the camera moves around.",
    "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/db00517b-165b-4005-a287-f28936b59b95/images/272bd7af-9294-4490-98c6-88a40e9deacc/scene/recipe/S27sh21_sel.png",
    "asset_id": "90605a36-00a6-4660-8958-5ddd53aec808",
    "role": "prev_still"
   },
   {
    "label": "CHARACTER REFERENCE — 안규홍: the exact person appearing in this shot; match face, hair and build exactly.",
    "path": "<bytes:1028384>",
    "asset_id": "3b5894db-ce55-4f34-b966-5e725c3aa952",
    "role": "character_ref"
   }
  ],
  "critique_skipped": true,
  "shot_run_produced": false,
  "shot_run_uid": "06aa1423-09cc-7825-ba7d-a96ade1ff8f3",
  "ref_mode": "prev+엔티티",
  "share_plan": {
   "ref_plan": "prev",
   "prev_anchor_tag": "S27sh21"
  }
 },
 "S27sh26::cine": {
  "applied": true,
  "attempted_at": "2026-09-09T04:02:55.549121+00:00",
  "fingerprint": "15dd293fee994f987c88bb68ef927c08fa2da6dbc50834f90afdd94f8411fb22",
  "fingerprint_version": 2,
  "provider": "mai",
  "endpoint": "openrouter/chat-completions",
  "model": "microsoft/mai-image-2.6",
  "pack": "24.202608252115",
  "source_file": "S27sh26_sel.png",
  "source_sha256": "ec804f6f015abab1ac3d4a6577a61a14eaad531a6e1946ebc8331a10dc63ed8a",
  "file": "S27sh26_cine.png",
  "staged_sha256": "c8ef074637a6df5ed5798ddb76b3a27ff539a1dba46eb1f4d11a117d466ded6a",
  "latency_ms": 30613
 },
 "S28sh4::signage": {
  "fp": "85e77c2543c11fa9",
  "inscriptions": [],
  "cues": [],
  "dropped": []
 },
 "S28sh4": {
  "input_fingerprint": "acbaaacf8e94d416",
  "prompt": "Create ONE FINAL photorealistic live-action film still of the moment below — early 20th-century Korea, Jeolla region; all people are Korean unless explicitly stated otherwise. TIME OF DAY (lock): night.\n\nSHOT TEXT (authoritative, Korean): 허리를 꼿꼿이 세운 채 마루 위 양반들을 똑바로 응시하는 안규홍의 정면.\n\nLOCATION (lock): The nighttime courtyard below the raised guest-wing veranda of a traditional estate. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: medium shot\n- FRAME LAYOUT: 안규홍 in the middle-left of the frame, midground; 이관회 in the middle-right of the frame, background; courtyard-facing porch edge beside the seated listener in the lower-center of the frame, foreground.\n- KEY BACKGROUND ELEMENTS: 사랑채 마루 (Occupied by the seated listeners above the courtyard) — The courtyard-facing edge cuts obliquely through the lower foreground beneath 박제현's shoulder; used as Establishes the listeners' elevated position without turning the appeal into a frontal portrait; 마당 (안규홍 addresses the porch from below, with 이관회 several steps behind him); used as A visible background interval separates the two men and clarifies the direction of the appeal.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Restrained nighttime illumination with muted saturation and naturalistic contrast keeps the appeal dignified while preserving the tension of the unequal positions.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The people visible in that photo are NOT in this shot: never carry their faces, bodies or clothing onto anyone here. Each person in THIS shot is defined solely by the PEOPLE list, their own CHARACTER REFERENCE images and their wardrobe notes. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The grain-store door still bears the red seizure seal in the dark yard. 안규홍: He stands below the veranda with the jade ring concealed in his clothing.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 안규홍 (한국인 남성, 청년 나이대, 검은 머리, 단단히 묶은 머리) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.",
  "roll_prompts": {
   "A": "Create ONE FINAL photorealistic live-action film still of the moment below — early 20th-century Korea, Jeolla region; all people are Korean unless explicitly stated otherwise. TIME OF DAY (lock): night.\n\nSHOT TEXT (authoritative, Korean): 허리를 꼿꼿이 세운 채 마루 위 양반들을 똑바로 응시하는 안규홍의 정면.\n\nLOCATION (lock): The nighttime courtyard below the raised guest-wing veranda of a traditional estate. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: medium shot\n- FRAME LAYOUT: 안규홍 in the middle-left of the frame, midground; 이관회 in the middle-right of the frame, background; courtyard-facing porch edge beside the seated listener in the lower-center of the frame, foreground.\n- KEY BACKGROUND ELEMENTS: 사랑채 마루 (Occupied by the seated listeners above the courtyard) — The courtyard-facing edge cuts obliquely through the lower foreground beneath 박제현's shoulder; used as Establishes the listeners' elevated position without turning the appeal into a frontal portrait; 마당 (안규홍 addresses the porch from below, with 이관회 several steps behind him); used as A visible background interval separates the two men and clarifies the direction of the appeal.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Restrained nighttime illumination with muted saturation and naturalistic contrast keeps the appeal dignified while preserving the tension of the unequal positions.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The people visible in that photo are NOT in this shot: never carry their faces, bodies or clothing onto anyone here. Each person in THIS shot is defined solely by the PEOPLE list, their own CHARACTER REFERENCE images and their wardrobe notes. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The grain-store door still bears the red seizure seal in the dark yard. 안규홍: He stands below the veranda with the jade ring concealed in his clothing.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 안규홍 (한국인 남성, 청년 나이대, 검은 머리, 단단히 묶은 머리) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.",
   "B": "Create ONE FINAL photorealistic live-action film still of the moment below — early 20th-century Korea, Jeolla region; all people are Korean unless explicitly stated otherwise. TIME OF DAY (lock): night.\n\nSHOT TEXT (authoritative, Korean): 허리를 꼿꼿이 세운 채 마루 위 양반들을 똑바로 응시하는 안규홍의 정면.\n\nLOCATION (lock): The nighttime courtyard below the raised guest-wing veranda of a traditional estate. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: medium shot\n- FRAME LAYOUT: 안규홍 in the middle-left of the frame, midground; 이관회 in the middle-right of the frame, background; courtyard-facing porch edge beside the seated listener in the lower-center of the frame, foreground.\n- KEY BACKGROUND ELEMENTS: 사랑채 마루 (Occupied by the seated listeners above the courtyard) — The courtyard-facing edge cuts obliquely through the lower foreground beneath 박제현's shoulder; used as Establishes the listeners' elevated position without turning the appeal into a frontal portrait; 마당 (안규홍 addresses the porch from below, with 이관회 several steps behind him); used as A visible background interval separates the two men and clarifies the direction of the appeal.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Restrained nighttime illumination with muted saturation and naturalistic contrast keeps the appeal dignified while preserving the tension of the unequal positions.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The people visible in that photo are NOT in this shot: never carry their faces, bodies or clothing onto anyone here. Each person in THIS shot is defined solely by the PEOPLE list, their own CHARACTER REFERENCE images and their wardrobe notes. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The grain-store door still bears the red seizure seal in the dark yard. 안규홍: He stands below the veranda with the jade ring concealed in his clothing.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 안규홍 (한국인 남성, 청년 나이대, 검은 머리, 단단히 묶은 머리) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.\n\nCOMPOSITION VARIATION (this candidate only): keep the place, the people and the moment exactly as specified, and keep from the CAMERA & FRAME contract its subject scale, screen placement and key background elements exactly. For this candidate only, its camera angle and height are deliberately left open: choose a camera position distinctly different from the obvious one, as a film director picking a second setup on the same blocking. Stay true to any explicit framing the SHOT TEXT itself dictates."
  },
  "shot_run_spend_attempt_count": 1,
  "initial_roll_all_fail": true,
  "readings": [
   {
    "label": "A",
    "direction": "안규홍은 서서 화면 우측 상단을 향해 시선을 두고 있으며, 이는 카메라 주변 마루에 앉아 있는 양반들을 향한 시선으로 적절히 기능합니다.",
    "built_space": "카메라가 지시대로 마루 위에 위치하여 나무 바닥 가장자리가 전경 하단을 사선으로 가로지릅니다. 배경에는 레퍼런스의 돌담과 붉은 종이가 붙은 곡간 문이 정확한 위치에 구현되었으나, 마당 한가운데에 레퍼런스에 없는 큰 나무 상자와 등불들이 놓여 있습니다.",
    "entities": "안규홍은 흰색 낡은 한복, 짚신, 묶은 머리 등 캐릭터 레퍼런스를 정확히 반영한 채 정면을 향해 있습니다. 배경 우측에는 이관회로 보이는 인물이 서 있고, 전경 좌우에는 지시된 청자(양반들)가 마루에 앉아 있습니다.",
    "hard_violations": [
     "[gemini-pro] 마당 중앙에 프롬프트와 레퍼런스에 없는 커다란 나무 상자와 등불을 임의로 추가함 (발명된 사물).",
     "[gpt-high] 원본에서 비어 있던 마당 전경에 요청되지 않은 대형 목제 궤짝을 추가해 장면의 주요 소품과 공간 점유를 바꾸었다.",
     "[gpt-high] 장소와 조명 연속성이 고정되어 있는데 원본에 없던 등불 세 개를 추가했다."
    ],
    "physics": "안규홍과 배경 인물은 흙바닥 위에 안정적으로 서 있으며, 전경의 인물들 또한 마루 위에 체중을 싣고 자연스럽게 앉아 있어 지지 상태가 정상적입니다."
   },
   {
    "label": "B",
    "direction": "안규홍은 화면 좌측 배경에 위치한 마루를 향해 측면으로 서 있으며, 전경의 인물은 안규홍을 등 뒤나 측면에서 바라보는 구도입니다.",
    "built_space": "마루 가장자리가 전경을 가로지르는 필수 구도를 무시하고, 마루를 좌측 배경에 배치했습니다. 우측 전경에는 레퍼런스 공간에 존재하지 않는 탁자와 같은 나무 구조물이 배치되어 공간을 왜곡합니다.",
    "entities": "안규홍의 의상과 머리 모양은 레퍼런스와 일치하지만, 샷 텍스트가 명시적으로 요구한 '정면'이 아닌 측면(프로필)을 보여줍니다. 배경 마루와 마당, 전경에 여러 인물이 배치되어 있습니다.",
    "hard_violations": [
     "[gemini-pro] 카메라 위치를 마루 위에서 마당으로 임의로 변경하여 '마루 가장자리가 전경을 가로지르는' 필수 구도를 완전히 파괴함 (스테이징 위반).",
     "[gemini-pro] 안규홍의 정면을 요구한 샷 텍스트를 무시하고 측면을 그림.",
     "[gemini-pro] 전경에 레퍼런스에 없는 탁자 구조물을 임의로 생성함 (발명된 사물).",
     "[gpt-high] 응시받아야 할 마루 위 양반 세 명을 안규홍의 등 뒤에 배치하여, 핵심 대면 장면과 맞지 않는 청자 위치를 만들었다."
    ],
    "physics": "인물들은 바닥과 마루, 탁자 구조물에 각각 정상적으로 닿아 있으며 물리적으로 공중에 떠 있거나 지탱되지 않은 객체는 없습니다."
   }
  ],
  "cross_model_order": {
   "policy": "cross_model_order_v2_slot0_fwd_slot1_rev_parallel2",
   "slots": [
    {
     "model": "gemini-pro",
     "order": "forward",
     "display_to_canonical": {
      "A": "A",
      "B": "B"
     },
     "raw": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 4,
        "verdict_ko": "지시된 복잡한 카메라 위치(마루 위에서 마당을 내려다보는 구도)와 인물 배치를 훌륭하게 구현했으나, 마당 한가운데에 지시되지 않은 큰 나무 상자와 등불을 임의로 생성하여 치명적인 위반을 범했습니다."
       },
       {
        "label": "B",
        "score": 2,
        "verdict_ko": "카메라 위치를 마당으로 임의 변경해 전경을 가로지르는 마루라는 필수 구도를 파괴했으며, 안규홍의 정면을 요구한 지시를 무시하고 측면을 그렸습니다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "안규홍은 서서 화면 우측 상단을 향해 시선을 두고 있으며, 이는 카메라 주변 마루에 앉아 있는 양반들을 향한 시선으로 적절히 기능합니다.",
        "built_space": "카메라가 지시대로 마루 위에 위치하여 나무 바닥 가장자리가 전경 하단을 사선으로 가로지릅니다. 배경에는 레퍼런스의 돌담과 붉은 종이가 붙은 곡간 문이 정확한 위치에 구현되었으나, 마당 한가운데에 레퍼런스에 없는 큰 나무 상자와 등불들이 놓여 있습니다.",
        "entities": "안규홍은 흰색 낡은 한복, 짚신, 묶은 머리 등 캐릭터 레퍼런스를 정확히 반영한 채 정면을 향해 있습니다. 배경 우측에는 이관회로 보이는 인물이 서 있고, 전경 좌우에는 지시된 청자(양반들)가 마루에 앉아 있습니다.",
        "hard_violations": [
         "마당 중앙에 프롬프트와 레퍼런스에 없는 커다란 나무 상자와 등불을 임의로 추가함 (발명된 사물)."
        ],
        "physics": "안규홍과 배경 인물은 흙바닥 위에 안정적으로 서 있으며, 전경의 인물들 또한 마루 위에 체중을 싣고 자연스럽게 앉아 있어 지지 상태가 정상적입니다."
       },
       {
        "label": "B",
        "direction": "안규홍은 화면 좌측 배경에 위치한 마루를 향해 측면으로 서 있으며, 전경의 인물은 안규홍을 등 뒤나 측면에서 바라보는 구도입니다.",
        "built_space": "마루 가장자리가 전경을 가로지르는 필수 구도를 무시하고, 마루를 좌측 배경에 배치했습니다. 우측 전경에는 레퍼런스 공간에 존재하지 않는 탁자와 같은 나무 구조물이 배치되어 공간을 왜곡합니다.",
        "entities": "안규홍의 의상과 머리 모양은 레퍼런스와 일치하지만, 샷 텍스트가 명시적으로 요구한 '정면'이 아닌 측면(프로필)을 보여줍니다. 배경 마루와 마당, 전경에 여러 인물이 배치되어 있습니다.",
        "hard_violations": [
         "카메라 위치를 마루 위에서 마당으로 임의로 변경하여 '마루 가장자리가 전경을 가로지르는' 필수 구도를 완전히 파괴함 (스테이징 위반).",
         "안규홍의 정면을 요구한 샷 텍스트를 무시하고 측면을 그림.",
         "전경에 레퍼런스에 없는 탁자 구조물을 임의로 생성함 (발명된 사물)."
        ],
        "physics": "인물들은 바닥과 마루, 탁자 구조물에 각각 정상적으로 닿아 있으며 물리적으로 공중에 떠 있거나 지탱되지 않은 객체는 없습니다."
       }
      ],
      "all_candidates_fail": true
     },
     "normalized": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 4,
        "verdict_ko": "지시된 복잡한 카메라 위치(마루 위에서 마당을 내려다보는 구도)와 인물 배치를 훌륭하게 구현했으나, 마당 한가운데에 지시되지 않은 큰 나무 상자와 등불을 임의로 생성하여 치명적인 위반을 범했습니다."
       },
       {
        "label": "B",
        "score": 2,
        "verdict_ko": "카메라 위치를 마당으로 임의 변경해 전경을 가로지르는 마루라는 필수 구도를 파괴했으며, 안규홍의 정면을 요구한 지시를 무시하고 측면을 그렸습니다."
       }
      ],
      "all_candidates_fail": true,
      "readings": [
       {
        "label": "A",
        "direction": "안규홍은 서서 화면 우측 상단을 향해 시선을 두고 있으며, 이는 카메라 주변 마루에 앉아 있는 양반들을 향한 시선으로 적절히 기능합니다.",
        "built_space": "카메라가 지시대로 마루 위에 위치하여 나무 바닥 가장자리가 전경 하단을 사선으로 가로지릅니다. 배경에는 레퍼런스의 돌담과 붉은 종이가 붙은 곡간 문이 정확한 위치에 구현되었으나, 마당 한가운데에 레퍼런스에 없는 큰 나무 상자와 등불들이 놓여 있습니다.",
        "entities": "안규홍은 흰색 낡은 한복, 짚신, 묶은 머리 등 캐릭터 레퍼런스를 정확히 반영한 채 정면을 향해 있습니다. 배경 우측에는 이관회로 보이는 인물이 서 있고, 전경 좌우에는 지시된 청자(양반들)가 마루에 앉아 있습니다.",
        "hard_violations": [
         "마당 중앙에 프롬프트와 레퍼런스에 없는 커다란 나무 상자와 등불을 임의로 추가함 (발명된 사물)."
        ],
        "physics": "안규홍과 배경 인물은 흙바닥 위에 안정적으로 서 있으며, 전경의 인물들 또한 마루 위에 체중을 싣고 자연스럽게 앉아 있어 지지 상태가 정상적입니다."
       },
       {
        "label": "B",
        "direction": "안규홍은 화면 좌측 배경에 위치한 마루를 향해 측면으로 서 있으며, 전경의 인물은 안규홍을 등 뒤나 측면에서 바라보는 구도입니다.",
        "built_space": "마루 가장자리가 전경을 가로지르는 필수 구도를 무시하고, 마루를 좌측 배경에 배치했습니다. 우측 전경에는 레퍼런스 공간에 존재하지 않는 탁자와 같은 나무 구조물이 배치되어 공간을 왜곡합니다.",
        "entities": "안규홍의 의상과 머리 모양은 레퍼런스와 일치하지만, 샷 텍스트가 명시적으로 요구한 '정면'이 아닌 측면(프로필)을 보여줍니다. 배경 마루와 마당, 전경에 여러 인물이 배치되어 있습니다.",
        "hard_violations": [
         "카메라 위치를 마루 위에서 마당으로 임의로 변경하여 '마루 가장자리가 전경을 가로지르는' 필수 구도를 완전히 파괴함 (스테이징 위반).",
         "안규홍의 정면을 요구한 샷 텍스트를 무시하고 측면을 그림.",
         "전경에 레퍼런스에 없는 탁자 구조물을 임의로 생성함 (발명된 사물)."
        ],
        "physics": "인물들은 바닥과 마루, 탁자 구조물에 각각 정상적으로 닿아 있으며 물리적으로 공중에 떠 있거나 지탱되지 않은 객체는 없습니다."
       }
      ]
     },
     "ok": true
    },
    {
     "model": "gpt-high",
     "order": "reverse",
     "display_to_canonical": {
      "A": "B",
      "B": "A"
     },
     "raw": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 2,
        "verdict_ko": "안규홍을 전신 측면으로 잡고 양반 세 명을 그의 뒤에 배치해, 정면 미디엄 숏과 마루 위 청자를 똑바로 응시하는 관계를 모두 벗어났다."
       },
       {
        "label": "B",
        "score": 3,
        "verdict_ko": "안규홍의 정면성과 인물 간 간격, 기존 장소는 더 가깝지만, 미디엄보다 넓은 구도와 새로 추가된 대형 궤짝·등불 때문에 충실한 실현은 아니다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "안규홍은 몸과 얼굴을 화면 오른쪽으로 돌리고 수평에 가까운 시선을 보낸다. 오른쪽 전경 청자 방향으로는 읽히지만, 왼쪽 마루에 앉은 양반 세 명은 그의 등 뒤에 있어 응시 대상이 아니다. 뒤쪽 남성은 안규홍 쪽을 보고, 왼쪽 양반들도 마당을 향한다. 무기나 이동 동작은 없다.",
        "built_space": "왼쪽에 목조 기둥과 창호가 연속된 높은 마루, 그 위 청자 세 명이 보인다. 별도의 마루 면이 아래쪽 전경을 비스듬히 가르고 그 위 청자의 뒷머리와 어깨가 크게 잡힌다. 안규홍은 왼쪽 중경의 흙바닥, 동행 남성은 오른쪽 배경에 서 있다. 붉은 봉인이 붙은 문 한 곳은 오른쪽 끝에 있다. 원본의 목재·돌 기단·흙마당은 유지하지만, 청자들을 안규홍 앞과 뒤로 나누어 놓았다. 작은 불빛 두 곳과 그릇도 보인다.",
        "entities": "안규홍은 검은 머리를 뒤로 묶은 한국인 청년 남성으로 보이고, 때 묻은 밝은 저고리·허리띠·바지·짚신은 인물 참고와 대체로 맞는다. 다만 얼굴이 측면이어서 정확한 정체성 비교는 제한된다. 오른쪽 배경에는 중년 남성 한 명, 왼쪽 마루에는 전통 복식의 남성 세 명, 전경에는 남성 청자 한 명의 뒷모습이 보인다. 봉인 문은 존재하고 글자는 읽히지 않는다. 옥가락지는 드러나지 않아 은닉 조건과 충돌하지 않는다.",
        "hard_violations": [
         "응시받아야 할 마루 위 양반 세 명을 안규홍의 등 뒤에 배치하여, 핵심 대면 장면과 맞지 않는 청자 위치를 만들었다."
        ],
        "physics": "안규홍의 두 발은 흙바닥에 닿고 몸무게를 지탱한다. 뒤쪽 남성도 지면에 서 있으며, 마루의 세 청자는 엉덩이와 접힌 다리를 마루에 둔다. 전경 청자의 하체 접점은 가려졌지만 마루 위에 앉은 상체로 읽힌다. 불빛과 그릇은 마루 표면에 놓여 있다. 공중에 떠 있는 몸이나 물체는 보이지 않는다."
       },
       {
        "label": "B",
        "direction": "안규홍은 상체를 거의 정면으로 두고 턱과 눈을 화면 오른쪽 위로 향한다. 오른쪽 전경에 앉은 청자의 화면 밖 얼굴 쪽을 바라보는 것으로 읽혀 높은 마루를 향한 호소 방향이 A보다 분명하다. 다만 청자의 얼굴이 잘려 정확한 눈맞춤은 확인할 수 없다. 배경 남성은 안규홍과 전경 마루 쪽을 향한다. 무기나 이동 동작은 없다.",
        "built_space": "원본과 닮은 굵은 목조 기둥이 왼쪽에 있고, 기와지붕을 얹은 봉인 문 한 곳과 낮은 돌담이 배경에 유지된다. 마루 가장자리는 아래 전경을 대각선으로 가른다. 안규홍은 왼쪽 중경의 마당에, 동행 남성은 간격을 두고 오른쪽 배경에 있다. 왼쪽 전경에 청자 두 명의 일부, 오른쪽 전경에 청자 한 명의 일부가 보인다. 그러나 마당 전경에 큰 궤짝 한 개와 등불 두 개, 기둥 옆에 매단 등불 한 개가 추가되어 원본의 비어 있던 마당과 조명 상태가 바뀌었다.",
        "entities": "안규홍은 참고와 유사한 얼굴의 한국인 청년 남성으로 보이며, 검은 머리와 오염된 밝은 저고리·천 허리띠·바지가 맞는다. 뒤로 묶은 머리 형태는 이 각도에서 명확하지 않다. 배경 남성은 머리띠를 두른 중년 남성으로 보인다. 전경 청자들은 전통 복식 차림이지만 얼굴이 잘려 나이와 정체성은 확인하기 어렵다. 붉은 봉인 문은 유지되고 판독 가능한 글자는 없다. 옥가락지는 노출되지 않는다. 큰 목제 궤짝과 세 등불은 요청이나 장소 참고에 없는 추가 소품이다.",
        "hard_violations": [
         "원본에서 비어 있던 마당 전경에 요청되지 않은 대형 목제 궤짝을 추가해 장면의 주요 소품과 공간 점유를 바꾸었다.",
         "장소와 조명 연속성이 고정되어 있는데 원본에 없던 등불 세 개를 추가했다."
        ],
        "physics": "안규홍의 발은 전경 마루에 가려져 있지만 두 다리는 지면으로 이어지는 정상적인 입식 자세다. 배경 남성의 발은 흙바닥에 닿는다. 왼쪽 청자의 신발은 마루에 놓이고, 오른쪽 청자의 손은 자신의 무릎 위에 얹혀 있다. 궤짝과 바닥 등불은 흙바닥에 놓여 있으며, 위쪽 등불은 처마 쪽에 매달린 배치로 보인다. 지지 없이 떠 있는 몸이나 물체는 보이지 않는다."
       }
      ],
      "all_candidates_fail": true
     },
     "normalized": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 2,
        "verdict_ko": "안규홍을 전신 측면으로 잡고 양반 세 명을 그의 뒤에 배치해, 정면 미디엄 숏과 마루 위 청자를 똑바로 응시하는 관계를 모두 벗어났다."
       },
       {
        "label": "A",
        "score": 3,
        "verdict_ko": "안규홍의 정면성과 인물 간 간격, 기존 장소는 더 가깝지만, 미디엄보다 넓은 구도와 새로 추가된 대형 궤짝·등불 때문에 충실한 실현은 아니다."
       }
      ],
      "all_candidates_fail": true,
      "readings": [
       {
        "label": "B",
        "direction": "안규홍은 몸과 얼굴을 화면 오른쪽으로 돌리고 수평에 가까운 시선을 보낸다. 오른쪽 전경 청자 방향으로는 읽히지만, 왼쪽 마루에 앉은 양반 세 명은 그의 등 뒤에 있어 응시 대상이 아니다. 뒤쪽 남성은 안규홍 쪽을 보고, 왼쪽 양반들도 마당을 향한다. 무기나 이동 동작은 없다.",
        "built_space": "왼쪽에 목조 기둥과 창호가 연속된 높은 마루, 그 위 청자 세 명이 보인다. 별도의 마루 면이 아래쪽 전경을 비스듬히 가르고 그 위 청자의 뒷머리와 어깨가 크게 잡힌다. 안규홍은 왼쪽 중경의 흙바닥, 동행 남성은 오른쪽 배경에 서 있다. 붉은 봉인이 붙은 문 한 곳은 오른쪽 끝에 있다. 원본의 목재·돌 기단·흙마당은 유지하지만, 청자들을 안규홍 앞과 뒤로 나누어 놓았다. 작은 불빛 두 곳과 그릇도 보인다.",
        "entities": "안규홍은 검은 머리를 뒤로 묶은 한국인 청년 남성으로 보이고, 때 묻은 밝은 저고리·허리띠·바지·짚신은 인물 참고와 대체로 맞는다. 다만 얼굴이 측면이어서 정확한 정체성 비교는 제한된다. 오른쪽 배경에는 중년 남성 한 명, 왼쪽 마루에는 전통 복식의 남성 세 명, 전경에는 남성 청자 한 명의 뒷모습이 보인다. 봉인 문은 존재하고 글자는 읽히지 않는다. 옥가락지는 드러나지 않아 은닉 조건과 충돌하지 않는다.",
        "hard_violations": [
         "응시받아야 할 마루 위 양반 세 명을 안규홍의 등 뒤에 배치하여, 핵심 대면 장면과 맞지 않는 청자 위치를 만들었다."
        ],
        "physics": "안규홍의 두 발은 흙바닥에 닿고 몸무게를 지탱한다. 뒤쪽 남성도 지면에 서 있으며, 마루의 세 청자는 엉덩이와 접힌 다리를 마루에 둔다. 전경 청자의 하체 접점은 가려졌지만 마루 위에 앉은 상체로 읽힌다. 불빛과 그릇은 마루 표면에 놓여 있다. 공중에 떠 있는 몸이나 물체는 보이지 않는다."
       },
       {
        "label": "A",
        "direction": "안규홍은 상체를 거의 정면으로 두고 턱과 눈을 화면 오른쪽 위로 향한다. 오른쪽 전경에 앉은 청자의 화면 밖 얼굴 쪽을 바라보는 것으로 읽혀 높은 마루를 향한 호소 방향이 A보다 분명하다. 다만 청자의 얼굴이 잘려 정확한 눈맞춤은 확인할 수 없다. 배경 남성은 안규홍과 전경 마루 쪽을 향한다. 무기나 이동 동작은 없다.",
        "built_space": "원본과 닮은 굵은 목조 기둥이 왼쪽에 있고, 기와지붕을 얹은 봉인 문 한 곳과 낮은 돌담이 배경에 유지된다. 마루 가장자리는 아래 전경을 대각선으로 가른다. 안규홍은 왼쪽 중경의 마당에, 동행 남성은 간격을 두고 오른쪽 배경에 있다. 왼쪽 전경에 청자 두 명의 일부, 오른쪽 전경에 청자 한 명의 일부가 보인다. 그러나 마당 전경에 큰 궤짝 한 개와 등불 두 개, 기둥 옆에 매단 등불 한 개가 추가되어 원본의 비어 있던 마당과 조명 상태가 바뀌었다.",
        "entities": "안규홍은 참고와 유사한 얼굴의 한국인 청년 남성으로 보이며, 검은 머리와 오염된 밝은 저고리·천 허리띠·바지가 맞는다. 뒤로 묶은 머리 형태는 이 각도에서 명확하지 않다. 배경 남성은 머리띠를 두른 중년 남성으로 보인다. 전경 청자들은 전통 복식 차림이지만 얼굴이 잘려 나이와 정체성은 확인하기 어렵다. 붉은 봉인 문은 유지되고 판독 가능한 글자는 없다. 옥가락지는 노출되지 않는다. 큰 목제 궤짝과 세 등불은 요청이나 장소 참고에 없는 추가 소품이다.",
        "hard_violations": [
         "원본에서 비어 있던 마당 전경에 요청되지 않은 대형 목제 궤짝을 추가해 장면의 주요 소품과 공간 점유를 바꾸었다.",
         "장소와 조명 연속성이 고정되어 있는데 원본에 없던 등불 세 개를 추가했다."
        ],
        "physics": "안규홍의 발은 전경 마루에 가려져 있지만 두 다리는 지면으로 이어지는 정상적인 입식 자세다. 배경 남성의 발은 흙바닥에 닿는다. 왼쪽 청자의 신발은 마루에 놓이고, 오른쪽 청자의 손은 자신의 무릎 위에 얹혀 있다. 궤짝과 바닥 등불은 흙바닥에 놓여 있으며, 위쪽 등불은 처마 쪽에 매달린 배치로 보인다. 지지 없이 떠 있는 몸이나 물체는 보이지 않는다."
       }
      ]
     },
     "ok": true
    }
   ],
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "slot_winner_match": true,
   "slot_winner": {
    "gemini-pro": "A",
    "gpt-high": "A"
   },
   "route": "cross_slot_agree"
  },
  "dual": {
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "normalized": {
    "A": 2.0,
    "B": 1.167
   },
   "adjusted": {
    "A": 1.75,
    "B": 0.917
   },
   "violations": {
    "A": [
     "[gemini-pro] 마당 중앙에 프롬프트와 레퍼런스에 없는 커다란 나무 상자와 등불을 임의로 추가함 (발명된 사물).",
     "[gpt-high] 원본에서 비어 있던 마당 전경에 요청되지 않은 대형 목제 궤짝을 추가해 장면의 주요 소품과 공간 점유를 바꾸었다.",
     "[gpt-high] 장소와 조명 연속성이 고정되어 있는데 원본에 없던 등불 세 개를 추가했다."
    ],
    "B": [
     "[gemini-pro] 카메라 위치를 마루 위에서 마당으로 임의로 변경하여 '마루 가장자리가 전경을 가로지르는' 필수 구도를 완전히 파괴함 (스테이징 위반).",
     "[gemini-pro] 안규홍의 정면을 요구한 샷 텍스트를 무시하고 측면을 그림.",
     "[gemini-pro] 전경에 레퍼런스에 없는 탁자 구조물을 임의로 생성함 (발명된 사물).",
     "[gpt-high] 응시받아야 할 마루 위 양반 세 명을 안규홍의 등 뒤에 배치하여, 핵심 대면 장면과 맞지 않는 청자 위치를 만들었다."
    ]
   },
   "per_model_winner": {
    "gemini-pro": "A",
    "gpt-high": "A"
   },
   "agreed": true
  },
  "totals": {
   "A": 1750,
   "B": 917
  },
  "selected": "A",
  "ranking": [
   "A",
   "B"
  ],
  "verdicts": [
   {
    "label": "A",
    "score": 1750,
    "verdict_ko": "지시된 복잡한 카메라 위치(마루 위에서 마당을 내려다보는 구도)와 인물 배치를 훌륭하게 구현했으나, 마당 한가운데에 지시되지 않은 큰 나무 상자와 등불을 임의로 생성하여 치명적인 위반을 범했습니다.  ★위반: [gemini-pro] 마당 중앙에 프롬프트와 레퍼런스에 없는 커다란 나무 상자와 등불을 임의로 추가함 (발명된 사물). / [gpt-high] 원본에서 비어 있던 마당 전경에 요청되지 않은 대형 목제 궤짝을 추가해 장면의 주요 소품과 공간 점유를 바꾸었다. / [gpt-high] 장소와 조명 연속성이 고정되어 있는데 원본에 없던 등불 세 개를 추가했다."
   },
   {
    "label": "B",
    "score": 917,
    "verdict_ko": "카메라 위치를 마당으로 임의 변경해 전경을 가로지르는 마루라는 필수 구도를 파괴했으며, 안규홍의 정면을 요구한 지시를 무시하고 측면을 그렸습니다.  ★위반: [gemini-pro] 카메라 위치를 마루 위에서 마당으로 임의로 변경하여 '마루 가장자리가 전경을 가로지르는' 필수 구도를 완전히 파괴함 (스테이징 위반). / [gemini-pro] 안규홍의 정면을 요구한 샷 텍스트를 무시하고 측면을 그림. / [gemini-pro] 전경에 레퍼런스에 없는 탁자 구조물을 임의로 생성함 (발명된 사물). / [gpt-high] 응시받아야 할 마루 위 양반 세 명을 안규홍의 등 뒤에 배치하여, 핵심 대면 장면과 맞지 않는 청자 위치를 만들었다."
   }
  ],
  "refs": [
   {
    "label": "PREVIOUS SHOT STILL — a visually related earlier shot of this same place: the location's look, materials, fixed features and lighting mood are LOCKED to this photo; never copy its camera framing. The people visible in that photo are NOT in this shot: never carry their faces, bodies or clothing onto anyone here. Each person in THIS shot is defined solely by the PEOPLE list, their own CHARACTER REFERENCE images and their wardrobe notes. If this photo shows a character who cannot move (dead or unconscious), that character's exact pose, position and orientation are ALSO LOCKED — treat their body as a fixed prop of the set that only the camera moves around.",
    "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/db00517b-165b-4005-a287-f28936b59b95/images/272bd7af-9294-4490-98c6-88a40e9deacc/scene/recipe/S20sh8_sel.png",
    "asset_id": "091fe1fa-0a25-4822-ac58-9415addc7820",
    "role": "prev_still"
   },
   {
    "label": "CHARACTER REFERENCE — 안규홍: the exact person appearing in this shot; match face, hair and build exactly.",
    "path": "<bytes:1028384>",
    "asset_id": "3b5894db-ce55-4f34-b966-5e725c3aa952",
    "role": "character_ref"
   }
  ],
  "critique_skipped": true,
  "needs_reshoot": true,
  "shot_run_produced": false,
  "shot_run_uid": "06aa1423-0b6e-786c-9b8e-d570a2c12c70",
  "ref_mode": "prev+엔티티",
  "share_plan": {
   "ref_plan": "prev",
   "prev_anchor_tag": "S20sh8"
  },
  "lane_policy": "ab_select_bypass:prev"
 },
 "S28sh4::cine": {
  "applied": true,
  "attempted_at": "2026-09-09T04:04:54.893214+00:00",
  "fingerprint": "9e9476288332516b78ac9a04dd4d428c9a8080887c8975caca202ab949daed5d",
  "fingerprint_version": 2,
  "provider": "mai",
  "endpoint": "openrouter/chat-completions",
  "model": "microsoft/mai-image-2.6",
  "pack": "24.202608252115",
  "source_file": "S28sh4_sel.png",
  "source_sha256": "671ead3efc8ca746821cfc0c6641c1d8aa0f91c44d7f5d53d87b9b75b0b03ff0",
  "file": "S28sh4_cine.png",
  "staged_sha256": "9ff9195a0b8d990d281db14b298d68c486bf6f1666717f19eafa32aa523f42f1",
  "latency_ms": 31275
 },
 "S28sh18::signage": {
  "fp": "b80348e74cc6de86",
  "inscriptions": [],
  "cues": [
   {
    "text_native": "",
    "source": "scene_text_implied",
    "source_quote": "붉은 압류 봉인지"
   }
  ],
  "dropped": []
 },
 "S28sh18": {
  "input_fingerprint": "3bb5b4662592323b",
  "prompt": "Create ONE FINAL photorealistic live-action film still of the moment below — early 20th-century Korea, Jeolla region; all people are Korean unless explicitly stated otherwise. TIME OF DAY (lock): night.\n\nSHOT TEXT (authoritative, Korean): 어두운 곳간 문 한가운데 선명하게 붙은 붉은 압류 봉인지 클로즈업.\n\nLOCATION (lock): The exterior surface of the estate's closed granary door in the dark courtyard, marked with a red confiscation seal. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: close-up\n- KEY BACKGROUND ELEMENTS: 압류 봉인지 (Attached prominently at the center of the storehouse door) — The outward-facing red seizure notice is visible to the camera; do not invent readable wording; used as Central focal detail explaining the refusal; 곳간 문 (Bearing the seizure seal) — Its seal-bearing face lies nearly parallel to the image plane; used as Provides immediate spatial context around the seal without introducing another focal point.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Keep the nighttime door subdued while preserving the red seal's legibility through restrained tonal separation and naturalistic contrast.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The people visible in that photo are NOT in this shot: never carry their faces, bodies or clothing onto anyone here. Each person in THIS shot is defined solely by the PEOPLE list, their own CHARACTER REFERENCE images and their wardrobe notes. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The red seizure seal remains intact on the closed grain-store door.\n\nNO PEOPLE IN THIS SHOT: the shot text shows only the place and its state — no living person appears in frame, and no face, and no figure standing, sitting or passing through. Do not invent someone to occupy the space.\n\nOne thing this rule does NOT forbid: the hand or arm of whoever is doing what the shot text describes. If the moment is an object being held, carried, operated, read, aimed, worn or handed over, the hand doing it belongs in frame, resting on or gripping that object exactly as its real use requires. An object photographed floating unsupported, with the hand that must be holding it erased, is a worse failure than the person you were avoiding — render the hand.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.",
  "roll_prompts": {
   "A": "Create ONE FINAL photorealistic live-action film still of the moment below — early 20th-century Korea, Jeolla region; all people are Korean unless explicitly stated otherwise. TIME OF DAY (lock): night.\n\nSHOT TEXT (authoritative, Korean): 어두운 곳간 문 한가운데 선명하게 붙은 붉은 압류 봉인지 클로즈업.\n\nLOCATION (lock): The exterior surface of the estate's closed granary door in the dark courtyard, marked with a red confiscation seal. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: close-up\n- KEY BACKGROUND ELEMENTS: 압류 봉인지 (Attached prominently at the center of the storehouse door) — The outward-facing red seizure notice is visible to the camera; do not invent readable wording; used as Central focal detail explaining the refusal; 곳간 문 (Bearing the seizure seal) — Its seal-bearing face lies nearly parallel to the image plane; used as Provides immediate spatial context around the seal without introducing another focal point.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Keep the nighttime door subdued while preserving the red seal's legibility through restrained tonal separation and naturalistic contrast.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The people visible in that photo are NOT in this shot: never carry their faces, bodies or clothing onto anyone here. Each person in THIS shot is defined solely by the PEOPLE list, their own CHARACTER REFERENCE images and their wardrobe notes. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The red seizure seal remains intact on the closed grain-store door.\n\nNO PEOPLE IN THIS SHOT: the shot text shows only the place and its state — no living person appears in frame, and no face, and no figure standing, sitting or passing through. Do not invent someone to occupy the space.\n\nOne thing this rule does NOT forbid: the hand or arm of whoever is doing what the shot text describes. If the moment is an object being held, carried, operated, read, aimed, worn or handed over, the hand doing it belongs in frame, resting on or gripping that object exactly as its real use requires. An object photographed floating unsupported, with the hand that must be holding it erased, is a worse failure than the person you were avoiding — render the hand.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.",
   "B": "Create ONE FINAL photorealistic live-action film still of the moment below — early 20th-century Korea, Jeolla region; all people are Korean unless explicitly stated otherwise. TIME OF DAY (lock): night.\n\nSHOT TEXT (authoritative, Korean): 어두운 곳간 문 한가운데 선명하게 붙은 붉은 압류 봉인지 클로즈업.\n\nLOCATION (lock): The exterior surface of the estate's closed granary door in the dark courtyard, marked with a red confiscation seal. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: close-up\n- KEY BACKGROUND ELEMENTS: 압류 봉인지 (Attached prominently at the center of the storehouse door) — The outward-facing red seizure notice is visible to the camera; do not invent readable wording; used as Central focal detail explaining the refusal; 곳간 문 (Bearing the seizure seal) — Its seal-bearing face lies nearly parallel to the image plane; used as Provides immediate spatial context around the seal without introducing another focal point.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Keep the nighttime door subdued while preserving the red seal's legibility through restrained tonal separation and naturalistic contrast.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The people visible in that photo are NOT in this shot: never carry their faces, bodies or clothing onto anyone here. Each person in THIS shot is defined solely by the PEOPLE list, their own CHARACTER REFERENCE images and their wardrobe notes. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The red seizure seal remains intact on the closed grain-store door.\n\nNO PEOPLE IN THIS SHOT: the shot text shows only the place and its state — no living person appears in frame, and no face, and no figure standing, sitting or passing through. Do not invent someone to occupy the space.\n\nOne thing this rule does NOT forbid: the hand or arm of whoever is doing what the shot text describes. If the moment is an object being held, carried, operated, read, aimed, worn or handed over, the hand doing it belongs in frame, resting on or gripping that object exactly as its real use requires. An object photographed floating unsupported, with the hand that must be holding it erased, is a worse failure than the person you were avoiding — render the hand.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.\n\nCOMPOSITION VARIATION (this candidate only): keep the place, the people and the moment exactly as specified, and keep from the CAMERA & FRAME contract its subject scale, screen placement and key background elements exactly. For this candidate only, its camera angle and height are deliberately left open: choose a camera position distinctly different from the obvious one, as a film director picking a second setup on the same blocking. Stay true to any explicit framing the SHOT TEXT itself dictates."
  },
  "shot_run_spend_attempt_count": 1,
  "readings": [
   {
    "label": "A",
    "direction": "카메라가 어두운 곳간 문 중앙에 붙은 붉은 봉인지를 정면으로 향해 바라봄.",
    "built_space": "세로로 된 나무 판자 틈이 있는 어두운 목재 문 표면.",
    "entities": "검은색 원형 도장 무늬와 읽을 수 있는 한글 텍스트(단어 파편)가 적힌 붉은색 종이 봉인지.",
    "hard_violations": [
     "[gpt-high] 봉인지에 판독 가능한 낱글자가 노출되어, 읽을 수 있는 글자를 전혀 허용하지 않는 조건을 위반한다."
    ],
    "physics": "봉인지가 문에 붙어 있으나, 종이 표면에 벗겨지거나 마른 잎사귀 같은 이질적인 부스러기들이 붙어 있음."
   },
   {
    "label": "B",
    "direction": "카메라가 어두운 곳간 문 중앙에 붙은 붉은 봉인지를 정면으로 향해 바라봄.",
    "built_space": "세로로 된 나무 판자 틈이 있는 어두운 목재 문 표면.",
    "entities": "검은색 원형 도장 무늬와 읽을 수 있는 한글 텍스트가 적힌 붉은색 종이 봉인지.",
    "hard_violations": [
     "[gpt-high] 봉인지에 판독 가능한 낱글자가 노출되어, 읽을 수 있는 글자를 전혀 허용하지 않는 조건을 위반한다."
    ],
    "physics": "봉인지가 구겨진 형태를 유지하며 나무 문 표면에 자연스럽게 밀착되어 있음."
   }
  ],
  "cross_model_order": {
   "policy": "cross_model_order_v2_slot0_fwd_slot1_rev_parallel2",
   "slots": [
    {
     "model": "gemini-pro",
     "order": "forward",
     "display_to_canonical": {
      "A": "A",
      "B": "B"
     },
     "raw": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 6,
        "verdict_ko": "종이의 구김과 표면 질감 표현이 매우 사실적이고 조명 분위기가 우수하나, 읽을 수 없는 글씨를 배제하라는 지시를 어기고 뚜렷한 한글('압류봉인' 등)이 나타났습니다."
       },
       {
        "label": "A",
        "score": 4,
        "verdict_ko": "문과 봉인지의 구도는 지시와 일치하나, 읽을 수 없는 글씨 조건을 어기고 한글(‘물’, ‘무’ 등)이 나타났으며 봉인지 표면에 정체불명의 부스러기 질감이 있어 부자연스럽습니다."
       }
      ],
      "readings": [
       {
        "label": "B",
        "direction": "카메라가 어두운 곳간 문 중앙에 붙은 붉은 봉인지를 정면으로 향해 바라봄.",
        "built_space": "세로로 된 나무 판자 틈이 있는 어두운 목재 문 표면.",
        "entities": "검은색 원형 도장 무늬와 읽을 수 있는 한글 텍스트가 적힌 붉은색 종이 봉인지.",
        "hard_violations": [],
        "physics": "봉인지가 구겨진 형태를 유지하며 나무 문 표면에 자연스럽게 밀착되어 있음."
       },
       {
        "label": "A",
        "direction": "카메라가 어두운 곳간 문 중앙에 붙은 붉은 봉인지를 정면으로 향해 바라봄.",
        "built_space": "세로로 된 나무 판자 틈이 있는 어두운 목재 문 표면.",
        "entities": "검은색 원형 도장 무늬와 읽을 수 있는 한글 텍스트(단어 파편)가 적힌 붉은색 종이 봉인지.",
        "hard_violations": [],
        "physics": "봉인지가 문에 붙어 있으나, 종이 표면에 벗겨지거나 마른 잎사귀 같은 이질적인 부스러기들이 붙어 있음."
       }
      ],
      "all_candidates_fail": false
     },
     "normalized": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 6,
        "verdict_ko": "종이의 구김과 표면 질감 표현이 매우 사실적이고 조명 분위기가 우수하나, 읽을 수 없는 글씨를 배제하라는 지시를 어기고 뚜렷한 한글('압류봉인' 등)이 나타났습니다."
       },
       {
        "label": "A",
        "score": 4,
        "verdict_ko": "문과 봉인지의 구도는 지시와 일치하나, 읽을 수 없는 글씨 조건을 어기고 한글(‘물’, ‘무’ 등)이 나타났으며 봉인지 표면에 정체불명의 부스러기 질감이 있어 부자연스럽습니다."
       }
      ],
      "all_candidates_fail": false,
      "readings": [
       {
        "label": "B",
        "direction": "카메라가 어두운 곳간 문 중앙에 붙은 붉은 봉인지를 정면으로 향해 바라봄.",
        "built_space": "세로로 된 나무 판자 틈이 있는 어두운 목재 문 표면.",
        "entities": "검은색 원형 도장 무늬와 읽을 수 있는 한글 텍스트가 적힌 붉은색 종이 봉인지.",
        "hard_violations": [],
        "physics": "봉인지가 구겨진 형태를 유지하며 나무 문 표면에 자연스럽게 밀착되어 있음."
       },
       {
        "label": "A",
        "direction": "카메라가 어두운 곳간 문 중앙에 붙은 붉은 봉인지를 정면으로 향해 바라봄.",
        "built_space": "세로로 된 나무 판자 틈이 있는 어두운 목재 문 표면.",
        "entities": "검은색 원형 도장 무늬와 읽을 수 있는 한글 텍스트(단어 파편)가 적힌 붉은색 종이 봉인지.",
        "hard_violations": [],
        "physics": "봉인지가 문에 붙어 있으나, 종이 표면에 벗겨지거나 마른 잎사귀 같은 이질적인 부스러기들이 붙어 있음."
       }
      ]
     },
     "ok": true
    },
    {
     "model": "gpt-high",
     "order": "reverse",
     "display_to_canonical": {
      "A": "B",
      "B": "A"
     },
     "raw": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 3,
        "verdict_ko": "붉은 봉인지의 온전한 부착 상태와 야간 목재 질감은 충실하지만, 판독 가능한 낱글자가 노출되고 정면·중앙 구도는 B보다 덜 정확하다."
       },
       {
        "label": "B",
        "score": 4,
        "verdict_ko": "문과 거의 평행한 정면 시점의 중앙 클로즈업이 더 정확하지만, 읽을 수 있는 낱글자와 봉인지의 벗겨진 부분 때문에 그대로 사용하기는 어렵다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "붉은 봉인지의 인쇄면이 카메라를 향한다. 문짝과 봉인지는 화면에 대해 조금 기울었지만 거의 평행하며, 사람의 시선이나 움직이는 물체는 없다.",
        "built_space": "낡은 세로 판재로 된 양쪽 문짝과 중앙 맞댐부가 보이고, 하단 왼쪽에는 가로 보강대 일부가 보인다. 붉은 봉인지 한 장이 중앙 이음매를 가로지른다. 참고 이미지의 어두운 목재 문과 부착 위치가 대체로 유지되며, 나머지 문 부속은 클로즈업 밖에 있다.",
        "entities": "사람 없이 목재 문과 붉은 직사각형 종이 봉인지가 보인다. 봉인지에는 참고 이미지처럼 원형 인장 두 개와 검은 필획이 있다. 종이는 주름지고 낡았지만 이어져 있다. 다만 상단의 '압'처럼 개별 글자로 읽히는 표기가 있어 문자 판독 금지 조건을 충족하지 못한다.",
        "hard_violations": [
         "봉인지에 판독 가능한 낱글자가 노출되어, 읽을 수 있는 글자를 전혀 허용하지 않는 조건을 위반한다."
        ],
        "physics": "봉인지의 양쪽 부분이 각 문짝 표면에 붙어 중앙 틈을 건넌다. 종이의 주름과 가장자리 들뜸은 부착된 종이에서 가능한 상태이며, 지지 없이 떠 있는 물체는 없다."
       },
       {
        "label": "B",
        "direction": "봉인지의 인쇄면이 카메라를 정면으로 향한다. 문짝 면은 화면과 거의 평행하고 중앙 이음매는 수직이다. 시선, 무기 또는 이동 방향을 판단할 대상은 없다.",
        "built_space": "양쪽 목재 문짝, 중앙의 두 세로 테두리, 하단 좌우의 가로 보강대가 보인다. 봉인지 한 장이 중앙 틈 위에 수평으로 붙어 화면 중심을 차지한다. 참고 이미지의 문 재료와 좌우 보강대 배치를 유지하며, 중복된 부속이나 불가능한 공간은 보이지 않는다.",
        "entities": "사람은 없고, 어두운 목재 문과 붉은 종이 봉인지가 있다. 원형 인장 두 개는 참고 이미지와 대응한다. 봉인지 여러 곳에 표면이 벗겨지거나 뜯긴 듯한 밝은 조각이 있어 온전한 상태의 연속성이 약해진다. 검은 표기 중 일부는 한글 낱글자로 읽힐 만큼 드러나 있다.",
        "hard_violations": [
         "봉인지에 판독 가능한 낱글자가 노출되어, 읽을 수 있는 글자를 전혀 허용하지 않는 조건을 위반한다."
        ],
        "physics": "종이는 좌우 문짝에 부착되어 중앙 틈을 가로지르며 지지된다. 벗겨진 부분도 종이 표면이나 가장자리에 연결되어 있고, 공중에 독립적으로 떠 있는 물체는 없다."
       }
      ],
      "all_candidates_fail": true
     },
     "normalized": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 3,
        "verdict_ko": "붉은 봉인지의 온전한 부착 상태와 야간 목재 질감은 충실하지만, 판독 가능한 낱글자가 노출되고 정면·중앙 구도는 B보다 덜 정확하다."
       },
       {
        "label": "A",
        "score": 4,
        "verdict_ko": "문과 거의 평행한 정면 시점의 중앙 클로즈업이 더 정확하지만, 읽을 수 있는 낱글자와 봉인지의 벗겨진 부분 때문에 그대로 사용하기는 어렵다."
       }
      ],
      "all_candidates_fail": true,
      "readings": [
       {
        "label": "B",
        "direction": "붉은 봉인지의 인쇄면이 카메라를 향한다. 문짝과 봉인지는 화면에 대해 조금 기울었지만 거의 평행하며, 사람의 시선이나 움직이는 물체는 없다.",
        "built_space": "낡은 세로 판재로 된 양쪽 문짝과 중앙 맞댐부가 보이고, 하단 왼쪽에는 가로 보강대 일부가 보인다. 붉은 봉인지 한 장이 중앙 이음매를 가로지른다. 참고 이미지의 어두운 목재 문과 부착 위치가 대체로 유지되며, 나머지 문 부속은 클로즈업 밖에 있다.",
        "entities": "사람 없이 목재 문과 붉은 직사각형 종이 봉인지가 보인다. 봉인지에는 참고 이미지처럼 원형 인장 두 개와 검은 필획이 있다. 종이는 주름지고 낡았지만 이어져 있다. 다만 상단의 '압'처럼 개별 글자로 읽히는 표기가 있어 문자 판독 금지 조건을 충족하지 못한다.",
        "hard_violations": [
         "봉인지에 판독 가능한 낱글자가 노출되어, 읽을 수 있는 글자를 전혀 허용하지 않는 조건을 위반한다."
        ],
        "physics": "봉인지의 양쪽 부분이 각 문짝 표면에 붙어 중앙 틈을 건넌다. 종이의 주름과 가장자리 들뜸은 부착된 종이에서 가능한 상태이며, 지지 없이 떠 있는 물체는 없다."
       },
       {
        "label": "A",
        "direction": "봉인지의 인쇄면이 카메라를 정면으로 향한다. 문짝 면은 화면과 거의 평행하고 중앙 이음매는 수직이다. 시선, 무기 또는 이동 방향을 판단할 대상은 없다.",
        "built_space": "양쪽 목재 문짝, 중앙의 두 세로 테두리, 하단 좌우의 가로 보강대가 보인다. 봉인지 한 장이 중앙 틈 위에 수평으로 붙어 화면 중심을 차지한다. 참고 이미지의 문 재료와 좌우 보강대 배치를 유지하며, 중복된 부속이나 불가능한 공간은 보이지 않는다.",
        "entities": "사람은 없고, 어두운 목재 문과 붉은 종이 봉인지가 있다. 원형 인장 두 개는 참고 이미지와 대응한다. 봉인지 여러 곳에 표면이 벗겨지거나 뜯긴 듯한 밝은 조각이 있어 온전한 상태의 연속성이 약해진다. 검은 표기 중 일부는 한글 낱글자로 읽힐 만큼 드러나 있다.",
        "hard_violations": [
         "봉인지에 판독 가능한 낱글자가 노출되어, 읽을 수 있는 글자를 전혀 허용하지 않는 조건을 위반한다."
        ],
        "physics": "종이는 좌우 문짝에 부착되어 중앙 틈을 가로지르며 지지된다. 벗겨진 부분도 종이 표면이나 가장자리에 연결되어 있고, 공중에 독립적으로 떠 있는 물체는 없다."
       }
      ]
     },
     "ok": true
    }
   ],
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "slot_winner_match": false,
   "slot_winner": {
    "gemini-pro": "B",
    "gpt-high": "A"
   },
   "route": "cross_slot_combined"
  },
  "dual": {
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "normalized": {
    "A": 1.667,
    "B": 1.75
   },
   "adjusted": {
    "A": 1.417,
    "B": 1.5
   },
   "violations": {
    "B": [
     "[gpt-high] 봉인지에 판독 가능한 낱글자가 노출되어, 읽을 수 있는 글자를 전혀 허용하지 않는 조건을 위반한다."
    ],
    "A": [
     "[gpt-high] 봉인지에 판독 가능한 낱글자가 노출되어, 읽을 수 있는 글자를 전혀 허용하지 않는 조건을 위반한다."
    ]
   },
   "per_model_winner": {
    "gemini-pro": "B",
    "gpt-high": "A"
   },
   "agreed": false
  },
  "totals": {
   "B": 1500,
   "A": 1417
  },
  "selected": "B",
  "ranking": [
   "B",
   "A"
  ],
  "verdicts": [
   {
    "label": "B",
    "score": 1500,
    "verdict_ko": "종이의 구김과 표면 질감 표현이 매우 사실적이고 조명 분위기가 우수하나, 읽을 수 없는 글씨를 배제하라는 지시를 어기고 뚜렷한 한글('압류봉인' 등)이 나타났습니다.  ★위반: [gpt-high] 봉인지에 판독 가능한 낱글자가 노출되어, 읽을 수 있는 글자를 전혀 허용하지 않는 조건을 위반한다."
   },
   {
    "label": "A",
    "score": 1417,
    "verdict_ko": "문과 봉인지의 구도는 지시와 일치하나, 읽을 수 없는 글씨 조건을 어기고 한글(‘물’, ‘무’ 등)이 나타났으며 봉인지 표면에 정체불명의 부스러기 질감이 있어 부자연스럽습니다.  ★위반: [gpt-high] 봉인지에 판독 가능한 낱글자가 노출되어, 읽을 수 있는 글자를 전혀 허용하지 않는 조건을 위반한다."
   }
  ],
  "refs": [
   {
    "label": "PREVIOUS SHOT STILL — a visually related earlier shot of this same place: the location's look, materials, fixed features and lighting mood are LOCKED to this photo; never copy its camera framing. The people visible in that photo are NOT in this shot: never carry their faces, bodies or clothing onto anyone here. Each person in THIS shot is defined solely by the PEOPLE list, their own CHARACTER REFERENCE images and their wardrobe notes. If this photo shows a character who cannot move (dead or unconscious), that character's exact pose, position and orientation are ALSO LOCKED — treat their body as a fixed prop of the set that only the camera moves around.",
    "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/db00517b-165b-4005-a287-f28936b59b95/images/272bd7af-9294-4490-98c6-88a40e9deacc/scene/recipe/S20sh2_sel.png",
    "asset_id": "b1a50fdb-e98b-4950-b7fd-a377f3e046a9",
    "role": "prev_still"
   }
  ],
  "critique_skipped": true,
  "shot_run_produced": false,
  "shot_run_uid": "06aa1423-0cff-7855-b9bc-f240b0e3cf92",
  "ref_mode": "prev만 (배경 전용·공유 계획)",
  "share_plan": {
   "ref_plan": "prev",
   "prev_anchor_tag": "S20sh2"
  },
  "lane_policy": "ab_select_bypass:bg_only:share_plan_prev_bgonly"
 },
 "S28sh18::cine": {
  "applied": true,
  "attempted_at": "2026-09-09T04:06:17.413543+00:00",
  "fingerprint": "65e492a5de8bf91a02691862cb806e8a9566ca18d8ab32fe9f20fb7b11d4f71b",
  "fingerprint_version": 2,
  "provider": "mai",
  "endpoint": "openrouter/chat-completions",
  "model": "microsoft/mai-image-2.6",
  "pack": "24.202608252115",
  "source_file": "S28sh18_sel.png",
  "source_sha256": "821ed2d40e5406bfbdd4a221d6389968a5002214a7386ca12256fb1b3f5c5c2a",
  "file": "S28sh18_cine.png",
  "staged_sha256": "7de6c48b913e2a0789f51fb727b7ca69bab12c7c70c8860cd9809664f1c264cc",
  "latency_ms": 28943
 },
 "S28sh21::signage": {
  "fp": "d7958ad7bc528e7b",
  "inscriptions": [],
  "cues": [],
  "dropped": []
 },
 "S28sh21": {
  "input_fingerprint": "613b974af1cae57b",
  "prompt": "Create ONE FINAL photorealistic live-action film still of the moment below — early 20th-century Korea, Jeolla region; all people are Korean unless explicitly stated otherwise. TIME OF DAY (lock): night.\n\nSHOT TEXT (authoritative, Korean): 어깨를 축 늘어뜨린 채 대문 쪽을 향해 몸을 돌려 선 안규홍의 씁쓸한 뒷모습.\n\nLOCATION (lock): The open courtyard route leading away from the estate's guest wing toward the main gate at night. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: medium shot\n- KEY BACKGROUND ELEMENTS: 마당 (Being crossed as 안규홍 leaves the seated gathering); used as Open space beside his turned body gives the departure somewhere to continue.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Maintain subdued nighttime ambient illumination and restrained contrast, allowing the dropped shoulder line to remain readable without theatrical backlighting.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The clothing, hair and overall look of 안규홍 — who appear both in that photo and in this shot — are LOCKED to that photo. Anyone else visible in that photo is NOT in this shot: never carry their face, body or clothing onto anyone here. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The grain-store door retains its red seizure seal; the yard remains dark. 안규홍: He turns to leave after bowing his head, with the jade ring concealed.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 안규홍 (한국인 남성, 청년 나이대, 검은 머리, 단단히 묶은 머리) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.",
  "roll_prompts": {
   "A": "Create ONE FINAL photorealistic live-action film still of the moment below — early 20th-century Korea, Jeolla region; all people are Korean unless explicitly stated otherwise. TIME OF DAY (lock): night.\n\nSHOT TEXT (authoritative, Korean): 어깨를 축 늘어뜨린 채 대문 쪽을 향해 몸을 돌려 선 안규홍의 씁쓸한 뒷모습.\n\nLOCATION (lock): The open courtyard route leading away from the estate's guest wing toward the main gate at night. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: medium shot\n- KEY BACKGROUND ELEMENTS: 마당 (Being crossed as 안규홍 leaves the seated gathering); used as Open space beside his turned body gives the departure somewhere to continue.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Maintain subdued nighttime ambient illumination and restrained contrast, allowing the dropped shoulder line to remain readable without theatrical backlighting.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The clothing, hair and overall look of 안규홍 — who appear both in that photo and in this shot — are LOCKED to that photo. Anyone else visible in that photo is NOT in this shot: never carry their face, body or clothing onto anyone here. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The grain-store door retains its red seizure seal; the yard remains dark. 안규홍: He turns to leave after bowing his head, with the jade ring concealed.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 안규홍 (한국인 남성, 청년 나이대, 검은 머리, 단단히 묶은 머리) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.",
   "B": "Create ONE FINAL photorealistic live-action film still of the moment below — early 20th-century Korea, Jeolla region; all people are Korean unless explicitly stated otherwise. TIME OF DAY (lock): night.\n\nSHOT TEXT (authoritative, Korean): 어깨를 축 늘어뜨린 채 대문 쪽을 향해 몸을 돌려 선 안규홍의 씁쓸한 뒷모습.\n\nLOCATION (lock): The open courtyard route leading away from the estate's guest wing toward the main gate at night. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: medium shot\n- KEY BACKGROUND ELEMENTS: 마당 (Being crossed as 안규홍 leaves the seated gathering); used as Open space beside his turned body gives the departure somewhere to continue.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Maintain subdued nighttime ambient illumination and restrained contrast, allowing the dropped shoulder line to remain readable without theatrical backlighting.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The clothing, hair and overall look of 안규홍 — who appear both in that photo and in this shot — are LOCKED to that photo. Anyone else visible in that photo is NOT in this shot: never carry their face, body or clothing onto anyone here. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The grain-store door retains its red seizure seal; the yard remains dark. 안규홍: He turns to leave after bowing his head, with the jade ring concealed.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 안규홍 (한국인 남성, 청년 나이대, 검은 머리, 단단히 묶은 머리) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.\n\nCOMPOSITION VARIATION (this candidate only): keep the place, the people and the moment exactly as specified, and keep from the CAMERA & FRAME contract its subject scale, screen placement and key background elements exactly. For this candidate only, its camera angle and height are deliberately left open: choose a camera position distinctly different from the obvious one, as a film director picking a second setup on the same blocking. Stay true to any explicit framing the SHOT TEXT itself dictates."
  },
  "shot_run_spend_attempt_count": 1,
  "readings": [
   {
    "label": "A",
    "direction": "안규홍이 대문 쪽을 향해 서 있으며 고개를 살짝 숙이고 있음.",
    "built_space": "마당, 왼편의 객사 마루, 붉은 부적이 붙은 배경의 닫힌 대문이 보임. 이전 샷 마당 중앙에 위치했던 나무 궤짝이 누락됨.",
    "entities": "안규홍의 뒷모습. 그러나 기준 이미지나 이전 샷에 없던 머리띠(다른 인물의 소품)를 착용하고 있으며, 허리띠가 누락됨.",
    "hard_violations": [
     "[gemini-pro] 이전 샷의 다른 인물이 착용했던 머리띠가 안규홍에게 잘못 적용됨 (다른 인물의 의상 전이 금지 지시 위반)"
    ],
    "physics": "두 다리로 땅을 안정적으로 딛고 서 있으며 지지 상태가 자연스러움."
   },
   {
    "label": "B",
    "direction": "안규홍이 대문 쪽을 향해 몸을 돌려 서서 고개를 약간 숙이고 있음.",
    "built_space": "마당과 객사 건물, 붉은 부적이 붙은 대문이 구현됨. 이전 샷의 나무 궤짝이 누락되었고, 바닥에 있어야 할 랜턴 중 하나가 마루 위로 위치가 변경됨.",
    "entities": "안규홍의 뒷모습. 단단히 묶은 흑발과 등 뒤로 묶인 천 허리띠 등 기준 이미지의 외양 특징을 올바르게 반영함.",
    "hard_violations": [],
    "physics": "두 다리로 땅을 딛고 서 있으며 체중이 자연스럽게 실려 있음. 늘어뜨린 양팔과 손의 묘사도 물리적으로 타당함."
   }
  ],
  "cross_model_order": {
   "policy": "cross_model_order_v2_slot0_fwd_slot1_rev_parallel2",
   "slots": [
    {
     "model": "gemini-pro",
     "order": "forward",
     "display_to_canonical": {
      "A": "A",
      "B": "B"
     },
     "raw": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 6,
        "verdict_ko": "프레이밍이 미디엄 샷이 아닌 풀 샷으로 넓어지고 마당의 궤짝이 누락되었으나, 안규홍의 헤어스타일과 의상을 기준 이미지에 맞게 정확히 재현하여 규정을 준수했습니다."
       },
       {
        "label": "A",
        "score": 2,
        "verdict_ko": "미디엄 샷의 프레이밍은 더 근접하게 연출했으나, 이전 샷의 다른 인물이 쓰던 머리띠를 안규홍에게 입혀 의상 전이 금지 지시를 명백히 위반했습니다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "안규홍이 대문 쪽을 향해 서 있으며 고개를 살짝 숙이고 있음.",
        "built_space": "마당, 왼편의 객사 마루, 붉은 부적이 붙은 배경의 닫힌 대문이 보임. 이전 샷 마당 중앙에 위치했던 나무 궤짝이 누락됨.",
        "entities": "안규홍의 뒷모습. 그러나 기준 이미지나 이전 샷에 없던 머리띠(다른 인물의 소품)를 착용하고 있으며, 허리띠가 누락됨.",
        "hard_violations": [
         "이전 샷의 다른 인물이 착용했던 머리띠가 안규홍에게 잘못 적용됨 (다른 인물의 의상 전이 금지 지시 위반)"
        ],
        "physics": "두 다리로 땅을 안정적으로 딛고 서 있으며 지지 상태가 자연스러움."
       },
       {
        "label": "B",
        "direction": "안규홍이 대문 쪽을 향해 몸을 돌려 서서 고개를 약간 숙이고 있음.",
        "built_space": "마당과 객사 건물, 붉은 부적이 붙은 대문이 구현됨. 이전 샷의 나무 궤짝이 누락되었고, 바닥에 있어야 할 랜턴 중 하나가 마루 위로 위치가 변경됨.",
        "entities": "안규홍의 뒷모습. 단단히 묶은 흑발과 등 뒤로 묶인 천 허리띠 등 기준 이미지의 외양 특징을 올바르게 반영함.",
        "hard_violations": [],
        "physics": "두 다리로 땅을 딛고 서 있으며 체중이 자연스럽게 실려 있음. 늘어뜨린 양팔과 손의 묘사도 물리적으로 타당함."
       }
      ],
      "all_candidates_fail": false
     },
     "normalized": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 6,
        "verdict_ko": "프레이밍이 미디엄 샷이 아닌 풀 샷으로 넓어지고 마당의 궤짝이 누락되었으나, 안규홍의 헤어스타일과 의상을 기준 이미지에 맞게 정확히 재현하여 규정을 준수했습니다."
       },
       {
        "label": "A",
        "score": 2,
        "verdict_ko": "미디엄 샷의 프레이밍은 더 근접하게 연출했으나, 이전 샷의 다른 인물이 쓰던 머리띠를 안규홍에게 입혀 의상 전이 금지 지시를 명백히 위반했습니다."
       }
      ],
      "all_candidates_fail": false,
      "readings": [
       {
        "label": "A",
        "direction": "안규홍이 대문 쪽을 향해 서 있으며 고개를 살짝 숙이고 있음.",
        "built_space": "마당, 왼편의 객사 마루, 붉은 부적이 붙은 배경의 닫힌 대문이 보임. 이전 샷 마당 중앙에 위치했던 나무 궤짝이 누락됨.",
        "entities": "안규홍의 뒷모습. 그러나 기준 이미지나 이전 샷에 없던 머리띠(다른 인물의 소품)를 착용하고 있으며, 허리띠가 누락됨.",
        "hard_violations": [
         "이전 샷의 다른 인물이 착용했던 머리띠가 안규홍에게 잘못 적용됨 (다른 인물의 의상 전이 금지 지시 위반)"
        ],
        "physics": "두 다리로 땅을 안정적으로 딛고 서 있으며 지지 상태가 자연스러움."
       },
       {
        "label": "B",
        "direction": "안규홍이 대문 쪽을 향해 몸을 돌려 서서 고개를 약간 숙이고 있음.",
        "built_space": "마당과 객사 건물, 붉은 부적이 붙은 대문이 구현됨. 이전 샷의 나무 궤짝이 누락되었고, 바닥에 있어야 할 랜턴 중 하나가 마루 위로 위치가 변경됨.",
        "entities": "안규홍의 뒷모습. 단단히 묶은 흑발과 등 뒤로 묶인 천 허리띠 등 기준 이미지의 외양 특징을 올바르게 반영함.",
        "hard_violations": [],
        "physics": "두 다리로 땅을 딛고 서 있으며 체중이 자연스럽게 실려 있음. 늘어뜨린 양팔과 손의 묘사도 물리적으로 타당함."
       }
      ]
     },
     "ok": true
    },
    {
     "model": "gpt-high",
     "order": "reverse",
     "display_to_canonical": {
      "A": "B",
      "B": "A"
     },
     "raw": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 6,
        "verdict_ko": "머리와 허리띠는 이전 컷에 가깝지만, 전신을 보여 주는 구도와 걷는 듯한 발동작이 지정된 미디엄 숏과 돌아서 선 순간에서 벗어난다."
       },
       {
        "label": "B",
        "score": 7,
        "verdict_ko": "추가된 머리띠와 사라진 허리띠는 연속성 오류지만, 미디엄 숏에 가까운 뒷모습과 처진 어깨, 옆으로 열린 마당이 핵심 연출을 더 충실히 구현한다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "인물은 카메라에 등을 보이고 화면 오른쪽 뒤의 붉은 봉인 문 쪽으로 몸을 틀었다. 고개는 숙여져 있어 시선은 문 자체보다 앞쪽 땅을 향하는 것으로 보인다. 떠나는 방향은 읽히지만, 봉인된 문과 별개인 대문은 화면에서 확인되지 않는다.",
        "built_space": "왼쪽에 목조 건물과 돌 기단, 뒤에 낮은 담장, 오른쪽 뒤에 기와지붕을 얹은 닫힌 양개문 한 곳이 있다. 문 중앙에는 붉은 봉인 하나가 있고, 왼쪽 처마의 매달린 등 하나와 기단 옆 바닥등 하나가 보인다. 인물은 기단 아래 흙마당에 있으며 오른쪽으로 빈 이동 공간이 남는다. 이전 컷의 주요 재료와 배치는 대체로 유지되지만, 인물을 발끝까지 담아 미디엄 숏보다 넓다.",
        "entities": "젊은 남성 한 명만 보이며 다른 사람은 없다. 얼굴은 거의 숨겨져 정확한 얼굴 동일성은 확인할 수 없지만, 검은 머리와 체격은 참고 인물에 대체로 부합한다. 때 탄 회백색 저고리와 바지, 감은 허리띠, 짚신은 참고 의상과 가깝다. 옥가락지는 노출되지 않는다. 붉은 봉인에는 읽을 수 있는 글자가 없고 별도의 자막이나 표식도 없다.",
        "hard_violations": [],
        "physics": "한쪽 발이 흙바닥에 체중을 싣고 다른 발이 뒤따르는 듯한 자세로, 지면 지지가 있다. 팔과 빈손은 자연스럽게 아래로 내려오고 옷도 중력 방향으로 처진다. 공중에 뜬 신체나 물체는 없다. 다만 발이 엇갈린 모습은 돌아서 멈춘 순간보다 이미 걷기 시작한 동작에 가깝다."
       },
       {
        "label": "B",
        "direction": "인물은 등을 보인 채 화면 중앙 오른쪽의 봉인 문 방향으로 몸을 돌리고 고개를 낮춘다. 보이는 얼굴 방향은 그 문으로 이어지는 마당 바닥을 향한다. 인물 오른쪽의 넓은 빈 마당이 퇴장 방향을 제공하지만, 별도의 대문 위치까지 확정할 수는 없다.",
        "built_space": "왼쪽에 목조 건물의 툇마루와 돌 기단, 뒤에 담장과 기와지붕 양개문 한 곳, 오른쪽 가장자리에 건물 일부가 보인다. 붉은 봉인 하나, 왼쪽의 매달린 등 하나, 바닥등 두 개가 보이며 등 수는 이전 컷에서 확인되는 세 개와 맞는다. 인물은 마당 전경에 서 있고 건축물은 뒤로 물러나 있다. 오른쪽 건물은 이전 컷에서 확인되지 않는 부분이지만 명백한 구조적 모순은 보이지 않는다. 허벅지에서 잘린 구도는 엄밀한 허리 위 미디엄 숏보다는 조금 넓지만 A보다 요구에 가깝다.",
        "entities": "젊은 남성 한 명의 뒷모습이며 다른 인물은 없다. 검은 머리와 체격은 대체로 맞지만, 참고에 없는 베이지색 머리띠가 추가되어 머리의 연속성이 달라진다. 회백색의 낡은 저고리와 바지는 유사하나, 화면에 보이는 허리에 참고의 감은 띠가 없고 등에는 덧댄 천이 두드러진다. 얼굴은 돌아가 있어 정확한 동일성은 검증할 수 없다. 옥가락지는 보이지 않고 봉인이나 다른 표면에도 읽을 수 있는 글자는 없다.",
        "hard_violations": [],
        "physics": "하체와 발의 접지점은 화면 아래로 잘려 직접 확인할 수 없지만, 몸통과 골반은 마당에 서 있는 사람의 자연스러운 정렬을 보인다. 공중에 떠 있다는 징후는 없다. 낮춘 머리, 둥글게 처진 어깨, 늘어진 팔과 손은 고개를 숙인 뒤 돌아서 선 동작으로 가능하다. 바닥등은 땅에 놓여 있고 처마등은 위에서 매달려 있다."
       }
      ],
      "all_candidates_fail": false
     },
     "normalized": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 6,
        "verdict_ko": "머리와 허리띠는 이전 컷에 가깝지만, 전신을 보여 주는 구도와 걷는 듯한 발동작이 지정된 미디엄 숏과 돌아서 선 순간에서 벗어난다."
       },
       {
        "label": "A",
        "score": 7,
        "verdict_ko": "추가된 머리띠와 사라진 허리띠는 연속성 오류지만, 미디엄 숏에 가까운 뒷모습과 처진 어깨, 옆으로 열린 마당이 핵심 연출을 더 충실히 구현한다."
       }
      ],
      "all_candidates_fail": false,
      "readings": [
       {
        "label": "B",
        "direction": "인물은 카메라에 등을 보이고 화면 오른쪽 뒤의 붉은 봉인 문 쪽으로 몸을 틀었다. 고개는 숙여져 있어 시선은 문 자체보다 앞쪽 땅을 향하는 것으로 보인다. 떠나는 방향은 읽히지만, 봉인된 문과 별개인 대문은 화면에서 확인되지 않는다.",
        "built_space": "왼쪽에 목조 건물과 돌 기단, 뒤에 낮은 담장, 오른쪽 뒤에 기와지붕을 얹은 닫힌 양개문 한 곳이 있다. 문 중앙에는 붉은 봉인 하나가 있고, 왼쪽 처마의 매달린 등 하나와 기단 옆 바닥등 하나가 보인다. 인물은 기단 아래 흙마당에 있으며 오른쪽으로 빈 이동 공간이 남는다. 이전 컷의 주요 재료와 배치는 대체로 유지되지만, 인물을 발끝까지 담아 미디엄 숏보다 넓다.",
        "entities": "젊은 남성 한 명만 보이며 다른 사람은 없다. 얼굴은 거의 숨겨져 정확한 얼굴 동일성은 확인할 수 없지만, 검은 머리와 체격은 참고 인물에 대체로 부합한다. 때 탄 회백색 저고리와 바지, 감은 허리띠, 짚신은 참고 의상과 가깝다. 옥가락지는 노출되지 않는다. 붉은 봉인에는 읽을 수 있는 글자가 없고 별도의 자막이나 표식도 없다.",
        "hard_violations": [],
        "physics": "한쪽 발이 흙바닥에 체중을 싣고 다른 발이 뒤따르는 듯한 자세로, 지면 지지가 있다. 팔과 빈손은 자연스럽게 아래로 내려오고 옷도 중력 방향으로 처진다. 공중에 뜬 신체나 물체는 없다. 다만 발이 엇갈린 모습은 돌아서 멈춘 순간보다 이미 걷기 시작한 동작에 가깝다."
       },
       {
        "label": "A",
        "direction": "인물은 등을 보인 채 화면 중앙 오른쪽의 봉인 문 방향으로 몸을 돌리고 고개를 낮춘다. 보이는 얼굴 방향은 그 문으로 이어지는 마당 바닥을 향한다. 인물 오른쪽의 넓은 빈 마당이 퇴장 방향을 제공하지만, 별도의 대문 위치까지 확정할 수는 없다.",
        "built_space": "왼쪽에 목조 건물의 툇마루와 돌 기단, 뒤에 담장과 기와지붕 양개문 한 곳, 오른쪽 가장자리에 건물 일부가 보인다. 붉은 봉인 하나, 왼쪽의 매달린 등 하나, 바닥등 두 개가 보이며 등 수는 이전 컷에서 확인되는 세 개와 맞는다. 인물은 마당 전경에 서 있고 건축물은 뒤로 물러나 있다. 오른쪽 건물은 이전 컷에서 확인되지 않는 부분이지만 명백한 구조적 모순은 보이지 않는다. 허벅지에서 잘린 구도는 엄밀한 허리 위 미디엄 숏보다는 조금 넓지만 A보다 요구에 가깝다.",
        "entities": "젊은 남성 한 명의 뒷모습이며 다른 인물은 없다. 검은 머리와 체격은 대체로 맞지만, 참고에 없는 베이지색 머리띠가 추가되어 머리의 연속성이 달라진다. 회백색의 낡은 저고리와 바지는 유사하나, 화면에 보이는 허리에 참고의 감은 띠가 없고 등에는 덧댄 천이 두드러진다. 얼굴은 돌아가 있어 정확한 동일성은 검증할 수 없다. 옥가락지는 보이지 않고 봉인이나 다른 표면에도 읽을 수 있는 글자는 없다.",
        "hard_violations": [],
        "physics": "하체와 발의 접지점은 화면 아래로 잘려 직접 확인할 수 없지만, 몸통과 골반은 마당에 서 있는 사람의 자연스러운 정렬을 보인다. 공중에 떠 있다는 징후는 없다. 낮춘 머리, 둥글게 처진 어깨, 늘어진 팔과 손은 고개를 숙인 뒤 돌아서 선 동작으로 가능하다. 바닥등은 땅에 놓여 있고 처마등은 위에서 매달려 있다."
       }
      ]
     },
     "ok": true
    }
   ],
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "slot_winner_match": false,
   "slot_winner": {
    "gemini-pro": "B",
    "gpt-high": "A"
   },
   "route": "cross_slot_combined"
  },
  "dual": {
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "normalized": {
    "A": 1.333,
    "B": 1.857
   },
   "adjusted": {
    "A": 1.083,
    "B": 1.857
   },
   "violations": {
    "A": [
     "[gemini-pro] 이전 샷의 다른 인물이 착용했던 머리띠가 안규홍에게 잘못 적용됨 (다른 인물의 의상 전이 금지 지시 위반)"
    ]
   },
   "per_model_winner": {
    "gemini-pro": "B",
    "gpt-high": "A"
   },
   "agreed": false
  },
  "totals": {
   "B": 1857,
   "A": 1083
  },
  "selected": "B",
  "ranking": [
   "B",
   "A"
  ],
  "verdicts": [
   {
    "label": "B",
    "score": 1857,
    "verdict_ko": "프레이밍이 미디엄 샷이 아닌 풀 샷으로 넓어지고 마당의 궤짝이 누락되었으나, 안규홍의 헤어스타일과 의상을 기준 이미지에 맞게 정확히 재현하여 규정을 준수했습니다."
   },
   {
    "label": "A",
    "score": 1083,
    "verdict_ko": "미디엄 샷의 프레이밍은 더 근접하게 연출했으나, 이전 샷의 다른 인물이 쓰던 머리띠를 안규홍에게 입혀 의상 전이 금지 지시를 명백히 위반했습니다.  ★위반: [gemini-pro] 이전 샷의 다른 인물이 착용했던 머리띠가 안규홍에게 잘못 적용됨 (다른 인물의 의상 전이 금지 지시 위반)"
   }
  ],
  "refs": [
   {
    "label": "PREVIOUS SHOT STILL — a visually related earlier shot of this same place: the location's look, materials, fixed features and lighting mood are LOCKED to this photo; never copy its camera framing. The clothing, hair and overall look of 안규홍 — who appear both in that photo and in this shot — are LOCKED to that photo. Anyone else visible in that photo is NOT in this shot: never carry their face, body or clothing onto anyone here. If this photo shows a character who cannot move (dead or unconscious), that character's exact pose, position and orientation are ALSO LOCKED — treat their body as a fixed prop of the set that only the camera moves around.",
    "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/db00517b-165b-4005-a287-f28936b59b95/images/272bd7af-9294-4490-98c6-88a40e9deacc/scene/recipe/S28sh4_sel.png",
    "asset_id": "06deec8a-4d51-4b76-9261-88b8f9fbee59",
    "role": "prev_still"
   },
   {
    "label": "CHARACTER REFERENCE — 안규홍: the exact person appearing in this shot; match face, hair and build exactly.",
    "path": "<bytes:1028384>",
    "asset_id": "3b5894db-ce55-4f34-b966-5e725c3aa952",
    "role": "character_ref"
   }
  ],
  "critique_skipped": true,
  "shot_run_produced": false,
  "shot_run_uid": "06aa1423-0e8c-7a2d-83e3-13c26e39c358",
  "ref_mode": "prev+엔티티",
  "share_plan": {
   "ref_plan": "prev",
   "prev_anchor_tag": "S28sh4"
  }
 },
 "S28sh21::cine": {
  "applied": true,
  "attempted_at": "2026-09-09T04:08:04.248687+00:00",
  "fingerprint": "bc2ff3714fd5a883bbf209646cb1f9327cf58bef55a6afe4acf2ce57a46920eb",
  "fingerprint_version": 2,
  "provider": "mai",
  "endpoint": "openrouter/chat-completions",
  "model": "microsoft/mai-image-2.6",
  "pack": "24.202608252115",
  "source_file": "S28sh21_sel.png",
  "source_sha256": "6e054c653069739ca02d7018e79ea11ec6a03f1a3f45ff9a0c7b62801767cd51",
  "file": "S28sh21_cine.png",
  "staged_sha256": "8a56dc4807366a750d555a65097a0b6c3a10e62e0148c4f0dd98c6c9c6087ffd",
  "latency_ms": 24702
 },
 "S29sh2::signage": {
  "fp": "eabac90c7bddb30b",
  "inscriptions": [],
  "cues": [],
  "dropped": []
 },
 "S29sh2": {
  "input_fingerprint": "187ce29eff81b777",
  "prompt": "Create ONE FINAL photorealistic live-action film still of the moment below — early 20th-century Korea, Jeolla region; all people are Korean unless explicitly stated otherwise. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 공터 쪽으로 쏜살같이 뛰어 들어오며 한 발은 땅을 딛고 한 발은 허공에 뜬 mid-stride 자세로 입을 벌리고 있는 달래의 역동적인 찰나.\n\nLOCATION (lock): The outdoor approach opening into the mountain camp's training clearing in daylight. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: wide shot\n- FRAME LAYOUT: Open clearing ahead of 달래's rightward run in the middle-right of the frame, midground.\n- KEY BACKGROUND ELEMENTS: 산채 공터 (달래 is entering the area where the men have been training); used as Visible ground beneath both feet and open space to the right establish her running direction.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Use natural daytime ambient light with muted saturation and clear, unembellished contrast that preserves the urgency of her expression.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The people visible in that photo are NOT in this shot: never carry their faces, bodies or clothing onto anyone here. Each person in THIS shot is defined solely by the PEOPLE list, their own CHARACTER REFERENCE images and their wardrobe notes. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): Daylight fills the camp's training clearing. 달래: She runs in breathless.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 달래 (한국인 여성, 젊은 성인, 자연스러운 검은 머리) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.",
  "roll_prompts": {
   "A": "Create ONE FINAL photorealistic live-action film still of the moment below — early 20th-century Korea, Jeolla region; all people are Korean unless explicitly stated otherwise. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 공터 쪽으로 쏜살같이 뛰어 들어오며 한 발은 땅을 딛고 한 발은 허공에 뜬 mid-stride 자세로 입을 벌리고 있는 달래의 역동적인 찰나.\n\nLOCATION (lock): The outdoor approach opening into the mountain camp's training clearing in daylight. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: wide shot\n- FRAME LAYOUT: Open clearing ahead of 달래's rightward run in the middle-right of the frame, midground.\n- KEY BACKGROUND ELEMENTS: 산채 공터 (달래 is entering the area where the men have been training); used as Visible ground beneath both feet and open space to the right establish her running direction.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Use natural daytime ambient light with muted saturation and clear, unembellished contrast that preserves the urgency of her expression.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The people visible in that photo are NOT in this shot: never carry their faces, bodies or clothing onto anyone here. Each person in THIS shot is defined solely by the PEOPLE list, their own CHARACTER REFERENCE images and their wardrobe notes. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): Daylight fills the camp's training clearing. 달래: She runs in breathless.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 달래 (한국인 여성, 젊은 성인, 자연스러운 검은 머리) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.",
   "B": "Create ONE FINAL photorealistic live-action film still of the moment below — early 20th-century Korea, Jeolla region; all people are Korean unless explicitly stated otherwise. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 공터 쪽으로 쏜살같이 뛰어 들어오며 한 발은 땅을 딛고 한 발은 허공에 뜬 mid-stride 자세로 입을 벌리고 있는 달래의 역동적인 찰나.\n\nLOCATION (lock): The outdoor approach opening into the mountain camp's training clearing in daylight. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: wide shot\n- FRAME LAYOUT: Open clearing ahead of 달래's rightward run in the middle-right of the frame, midground.\n- KEY BACKGROUND ELEMENTS: 산채 공터 (달래 is entering the area where the men have been training); used as Visible ground beneath both feet and open space to the right establish her running direction.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Use natural daytime ambient light with muted saturation and clear, unembellished contrast that preserves the urgency of her expression.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The people visible in that photo are NOT in this shot: never carry their faces, bodies or clothing onto anyone here. Each person in THIS shot is defined solely by the PEOPLE list, their own CHARACTER REFERENCE images and their wardrobe notes. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): Daylight fills the camp's training clearing. 달래: She runs in breathless.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 달래 (한국인 여성, 젊은 성인, 자연스러운 검은 머리) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.\n\nCOMPOSITION VARIATION (this candidate only): keep the place, the people and the moment exactly as specified, and keep from the CAMERA & FRAME contract its subject scale, screen placement and key background elements exactly. For this candidate only, its camera angle and height are deliberately left open: choose a camera position distinctly different from the obvious one, as a film director picking a second setup on the same blocking. Stay true to any explicit framing the SHOT TEXT itself dictates."
  },
  "shot_run_spend_attempt_count": 1,
  "readings": [
   {
    "label": "A",
    "direction": "달래가 프레임 우측의 열린 공터를 향해 시선을 두고 달려가는 방향이 명확함.",
    "built_space": "왼쪽의 바위 경사면, 중앙 배경의 초가집, 오른쪽의 나무 울타리 등 레퍼런스의 공간적 특징과 배치가 정확하게 유지됨.",
    "entities": "지정된 달래 캐릭터가 녹색 저고리와 검은 치마를 입고 입을 벌린 역동적인 표정으로 묘사됨.",
    "hard_violations": [],
    "physics": "달리는 역동적인 자세가 자연스럽지만, 프롬프트의 '한 발은 땅을 딛고'라는 지시와 달리 두 발이 모두 지면에서 떨어져 허공에 떠 있는 체공 상태임."
   },
   {
    "label": "B",
    "direction": "달래가 우측으로 달려가며 시선을 앞으로 향하고 있음.",
    "built_space": "배경의 초가집과 바위는 존재하나, 기준 이미지나 프롬프트에 없는 거대한 울타리가 우측 근경을 크게 차지하며 공간을 왜곡함.",
    "entities": "달래가 지정된 의상을 입고 입을 벌린 채 묘사됨.",
    "hard_violations": [],
    "physics": "달리는 자세이나 A와 마찬가지로 두 발이 모두 지면에서 떨어져 있어 텍스트의 지면 접촉 지시를 위반함."
   }
  ],
  "cross_model_order": {
   "policy": "cross_model_order_v2_slot0_fwd_slot1_rev_parallel2",
   "slots": [
    {
     "model": "gemini-pro",
     "order": "forward",
     "display_to_canonical": {
      "A": "A",
      "B": "B"
     },
     "raw": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 8,
        "verdict_ko": "기준 이미지의 산채 공터 배경과 프레이밍 지시를 완벽하게 따랐으나, 두 발이 모두 허공에 떠 있어 '한 발은 땅을 딛고'라는 세부 포즈 지시를 놓친 점이 아쉽습니다."
       },
       {
        "label": "B",
        "score": 5,
        "verdict_ko": "포즈 지시를 어기고 두 발이 허공에 떠 있을 뿐만 아니라, 우측 전경에 거대한 나무 울타리를 임의로 추가하여 텍스트가 요구한 탁 트인 공간감과 배경 스케일을 크게 훼손했습니다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "달래가 프레임 우측의 열린 공터를 향해 시선을 두고 달려가는 방향이 명확함.",
        "built_space": "왼쪽의 바위 경사면, 중앙 배경의 초가집, 오른쪽의 나무 울타리 등 레퍼런스의 공간적 특징과 배치가 정확하게 유지됨.",
        "entities": "지정된 달래 캐릭터가 녹색 저고리와 검은 치마를 입고 입을 벌린 역동적인 표정으로 묘사됨.",
        "hard_violations": [],
        "physics": "달리는 역동적인 자세가 자연스럽지만, 프롬프트의 '한 발은 땅을 딛고'라는 지시와 달리 두 발이 모두 지면에서 떨어져 허공에 떠 있는 체공 상태임."
       },
       {
        "label": "B",
        "direction": "달래가 우측으로 달려가며 시선을 앞으로 향하고 있음.",
        "built_space": "배경의 초가집과 바위는 존재하나, 기준 이미지나 프롬프트에 없는 거대한 울타리가 우측 근경을 크게 차지하며 공간을 왜곡함.",
        "entities": "달래가 지정된 의상을 입고 입을 벌린 채 묘사됨.",
        "hard_violations": [],
        "physics": "달리는 자세이나 A와 마찬가지로 두 발이 모두 지면에서 떨어져 있어 텍스트의 지면 접촉 지시를 위반함."
       }
      ],
      "all_candidates_fail": false
     },
     "normalized": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 8,
        "verdict_ko": "기준 이미지의 산채 공터 배경과 프레이밍 지시를 완벽하게 따랐으나, 두 발이 모두 허공에 떠 있어 '한 발은 땅을 딛고'라는 세부 포즈 지시를 놓친 점이 아쉽습니다."
       },
       {
        "label": "B",
        "score": 5,
        "verdict_ko": "포즈 지시를 어기고 두 발이 허공에 떠 있을 뿐만 아니라, 우측 전경에 거대한 나무 울타리를 임의로 추가하여 텍스트가 요구한 탁 트인 공간감과 배경 스케일을 크게 훼손했습니다."
       }
      ],
      "all_candidates_fail": false,
      "readings": [
       {
        "label": "A",
        "direction": "달래가 프레임 우측의 열린 공터를 향해 시선을 두고 달려가는 방향이 명확함.",
        "built_space": "왼쪽의 바위 경사면, 중앙 배경의 초가집, 오른쪽의 나무 울타리 등 레퍼런스의 공간적 특징과 배치가 정확하게 유지됨.",
        "entities": "지정된 달래 캐릭터가 녹색 저고리와 검은 치마를 입고 입을 벌린 역동적인 표정으로 묘사됨.",
        "hard_violations": [],
        "physics": "달리는 역동적인 자세가 자연스럽지만, 프롬프트의 '한 발은 땅을 딛고'라는 지시와 달리 두 발이 모두 지면에서 떨어져 허공에 떠 있는 체공 상태임."
       },
       {
        "label": "B",
        "direction": "달래가 우측으로 달려가며 시선을 앞으로 향하고 있음.",
        "built_space": "배경의 초가집과 바위는 존재하나, 기준 이미지나 프롬프트에 없는 거대한 울타리가 우측 근경을 크게 차지하며 공간을 왜곡함.",
        "entities": "달래가 지정된 의상을 입고 입을 벌린 채 묘사됨.",
        "hard_violations": [],
        "physics": "달리는 자세이나 A와 마찬가지로 두 발이 모두 지면에서 떨어져 있어 텍스트의 지면 접촉 지시를 위반함."
       }
      ]
     },
     "ok": true
    },
    {
     "model": "gpt-high",
     "order": "reverse",
     "display_to_canonical": {
      "A": "B",
      "B": "A"
     },
     "raw": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 3,
        "verdict_ko": "입을 벌린 달리기는 보이지만 진행 방향이 오른쪽이 아닌 왼쪽 앞이며, 두 발 모두 떠 있고 오른쪽 전경 울타리가 요구된 열린 공터 구도를 약화한다."
       },
       {
        "label": "B",
        "score": 4,
        "verdict_ko": "산채 배경과 오른쪽의 열린 공간은 A보다 충실하지만, 왼쪽 앞으로 달리는 방향과 양발이 뜬 순간은 핵심 지시를 충족하지 못한다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "얼굴과 시선은 화면 왼쪽 앞의 화면 밖 지점을 향한다. 앞발도 왼쪽 아래로 뻗고 뒷발은 오른쪽으로 접혀 있어 카메라 쪽으로 가까워지며 왼쪽으로 달리는 모습이다. 지정된 화면 오른쪽 공터로 진입하는 방향은 아니다. 무기나 겨냥하는 소품은 없다.",
        "built_space": "왼쪽 바위 비탈, 뒤쪽 큰 초가 한 채와 그 오른쪽 낮은 초가 한 채, 뒤쪽 목책이 보인다. 큰 초가에는 어두운 출입구와 작은 창이 있고, 앞에는 낮은 목제 걸상 하나가 있다. 오른쪽 전경에는 별도의 목책 구간이 크게 들어온다. 흙바닥과 숲, 초가 재료는 장소 참조와 유사하지만 전경 목책이 오른쪽의 열린 공간을 막는다. 달래는 중앙 오른쪽에 전신으로 있으나 중경 인물치고 화면에서 크게 보인다. 반사는 없다.",
        "entities": "젊은 동아시아계 여성 한 명만 있으며 달래의 연령·성별 및 한국인 설정과 외관상 충돌하지 않는다. 얼굴은 참조와 대체로 유사하고 검은 머리, 연녹색 저고리, 흰 깃, 검은 긴 치마와 낮은 신발이 맞는다. 참조의 가슴 앞으로 내려온 긴 땋은 머리는 보이지 않고 뒤로 묶은 머리처럼 보이며, 붉은 장식도 허리 뒤쪽에 조금 보인다. 입을 벌린 다급한 표정은 맞고 이전 장면의 남성이나 읽을 수 있는 글자는 없다.",
        "hard_violations": [],
        "physics": "앞쪽 신발 밑창과 지면 사이에 틈이 있고 뒤쪽 발도 높이 들려 있어 현재 땅에 닿은 발은 없다. 다만 접힌 뒷다리와 앞으로 뻗은 착지 다리, 달리는 팔 동작은 직전 지면 박차기로 생긴 짧은 체공으로 설명 가능하다. 앞발이 바로 아래 흙바닥에 착지할 수 있어 불가능한 부유는 아니지만, 반드시 한 발을 딛고 있어야 하는 지정 순간과 다르다. 옷은 몸에 걸려 있고 치맛자락은 달리기 동작에 따라 들려 있다."
       },
       {
        "label": "B",
        "direction": "시선과 열린 입이 향한 얼굴은 화면 왼쪽 앞을 바라본다. 앞다리는 왼쪽 아래로 나가고 뒷다리는 오른쪽 뒤로 접혀 있어 이동도 왼쪽 전방으로 읽힌다. 화면 오른쪽에 빈 공터는 있으나 달래가 그쪽으로 달려 들어가지는 않는다. 방향을 가진 별도 소품은 없다.",
        "built_space": "왼쪽 바위 비탈과 숲, 뒤쪽 큰 초가 한 채와 오른쪽의 낮은 초가 한 채, 뒤편 목책이 보이며 맨 오른쪽 가장자리에는 지붕 일부가 걸린다. 큰 초가 앞에는 가지와 장작 더미가 있다. 달래는 중앙 오른쪽에 전신으로 배치되고 발 아래 흙바닥과 오른쪽 공터가 끊기지 않고 보인다. A의 전경 목책이 없어 열린 훈련장 구성과 참조의 공간감에 더 가깝지만 인물은 여전히 중경 와이드숏치고 큰 편이다. 반사는 없다.",
        "entities": "젊은 동아시아계 여성 한 명으로 달래의 연령·성별 및 한국인 설정과 외관상 충돌하지 않는다. 얼굴과 체형은 참조에 대체로 가깝고 연녹색 저고리, 흰 깃, 검은 치마, 검은 계열 신발이 일치한다. 검은 머리는 뒤로 묶여 보이며 참조의 긴 앞쪽 땋은 머리는 확인되지 않는다. 붉은 장식 일부는 몸 뒤에 보인다. 입을 크게 벌린 숨 가쁜 표정은 지시와 맞는다. 다른 사람이나 읽을 수 있는 글자는 없다.",
        "hard_violations": [],
        "physics": "앞쪽 신발은 밑창이 보인 채 지면 위에 떠 있고, 뒤쪽 신발 역시 공중에 있어 한 발의 지면 접촉이 없다. 굽힌 뒷무릎, 앞으로 나온 착지 다리와 교차하는 팔 동작은 달리기의 정상적인 체공 단계로 가능하며 앞발 아래에 착지할 흙바닥도 있다. 따라서 근거 없는 부유나 불가능한 해부학은 아니지만 요청한 한 발 지지 자세에는 실패한다. 치마와 옷고름의 움직임은 몸에 고정된 천이 달리면서 흔들리는 것으로 설명된다."
       }
      ],
      "all_candidates_fail": true
     },
     "normalized": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 3,
        "verdict_ko": "입을 벌린 달리기는 보이지만 진행 방향이 오른쪽이 아닌 왼쪽 앞이며, 두 발 모두 떠 있고 오른쪽 전경 울타리가 요구된 열린 공터 구도를 약화한다."
       },
       {
        "label": "A",
        "score": 4,
        "verdict_ko": "산채 배경과 오른쪽의 열린 공간은 A보다 충실하지만, 왼쪽 앞으로 달리는 방향과 양발이 뜬 순간은 핵심 지시를 충족하지 못한다."
       }
      ],
      "all_candidates_fail": true,
      "readings": [
       {
        "label": "B",
        "direction": "얼굴과 시선은 화면 왼쪽 앞의 화면 밖 지점을 향한다. 앞발도 왼쪽 아래로 뻗고 뒷발은 오른쪽으로 접혀 있어 카메라 쪽으로 가까워지며 왼쪽으로 달리는 모습이다. 지정된 화면 오른쪽 공터로 진입하는 방향은 아니다. 무기나 겨냥하는 소품은 없다.",
        "built_space": "왼쪽 바위 비탈, 뒤쪽 큰 초가 한 채와 그 오른쪽 낮은 초가 한 채, 뒤쪽 목책이 보인다. 큰 초가에는 어두운 출입구와 작은 창이 있고, 앞에는 낮은 목제 걸상 하나가 있다. 오른쪽 전경에는 별도의 목책 구간이 크게 들어온다. 흙바닥과 숲, 초가 재료는 장소 참조와 유사하지만 전경 목책이 오른쪽의 열린 공간을 막는다. 달래는 중앙 오른쪽에 전신으로 있으나 중경 인물치고 화면에서 크게 보인다. 반사는 없다.",
        "entities": "젊은 동아시아계 여성 한 명만 있으며 달래의 연령·성별 및 한국인 설정과 외관상 충돌하지 않는다. 얼굴은 참조와 대체로 유사하고 검은 머리, 연녹색 저고리, 흰 깃, 검은 긴 치마와 낮은 신발이 맞는다. 참조의 가슴 앞으로 내려온 긴 땋은 머리는 보이지 않고 뒤로 묶은 머리처럼 보이며, 붉은 장식도 허리 뒤쪽에 조금 보인다. 입을 벌린 다급한 표정은 맞고 이전 장면의 남성이나 읽을 수 있는 글자는 없다.",
        "hard_violations": [],
        "physics": "앞쪽 신발 밑창과 지면 사이에 틈이 있고 뒤쪽 발도 높이 들려 있어 현재 땅에 닿은 발은 없다. 다만 접힌 뒷다리와 앞으로 뻗은 착지 다리, 달리는 팔 동작은 직전 지면 박차기로 생긴 짧은 체공으로 설명 가능하다. 앞발이 바로 아래 흙바닥에 착지할 수 있어 불가능한 부유는 아니지만, 반드시 한 발을 딛고 있어야 하는 지정 순간과 다르다. 옷은 몸에 걸려 있고 치맛자락은 달리기 동작에 따라 들려 있다."
       },
       {
        "label": "A",
        "direction": "시선과 열린 입이 향한 얼굴은 화면 왼쪽 앞을 바라본다. 앞다리는 왼쪽 아래로 나가고 뒷다리는 오른쪽 뒤로 접혀 있어 이동도 왼쪽 전방으로 읽힌다. 화면 오른쪽에 빈 공터는 있으나 달래가 그쪽으로 달려 들어가지는 않는다. 방향을 가진 별도 소품은 없다.",
        "built_space": "왼쪽 바위 비탈과 숲, 뒤쪽 큰 초가 한 채와 오른쪽의 낮은 초가 한 채, 뒤편 목책이 보이며 맨 오른쪽 가장자리에는 지붕 일부가 걸린다. 큰 초가 앞에는 가지와 장작 더미가 있다. 달래는 중앙 오른쪽에 전신으로 배치되고 발 아래 흙바닥과 오른쪽 공터가 끊기지 않고 보인다. A의 전경 목책이 없어 열린 훈련장 구성과 참조의 공간감에 더 가깝지만 인물은 여전히 중경 와이드숏치고 큰 편이다. 반사는 없다.",
        "entities": "젊은 동아시아계 여성 한 명으로 달래의 연령·성별 및 한국인 설정과 외관상 충돌하지 않는다. 얼굴과 체형은 참조에 대체로 가깝고 연녹색 저고리, 흰 깃, 검은 치마, 검은 계열 신발이 일치한다. 검은 머리는 뒤로 묶여 보이며 참조의 긴 앞쪽 땋은 머리는 확인되지 않는다. 붉은 장식 일부는 몸 뒤에 보인다. 입을 크게 벌린 숨 가쁜 표정은 지시와 맞는다. 다른 사람이나 읽을 수 있는 글자는 없다.",
        "hard_violations": [],
        "physics": "앞쪽 신발은 밑창이 보인 채 지면 위에 떠 있고, 뒤쪽 신발 역시 공중에 있어 한 발의 지면 접촉이 없다. 굽힌 뒷무릎, 앞으로 나온 착지 다리와 교차하는 팔 동작은 달리기의 정상적인 체공 단계로 가능하며 앞발 아래에 착지할 흙바닥도 있다. 따라서 근거 없는 부유나 불가능한 해부학은 아니지만 요청한 한 발 지지 자세에는 실패한다. 치마와 옷고름의 움직임은 몸에 고정된 천이 달리면서 흔들리는 것으로 설명된다."
       }
      ]
     },
     "ok": true
    }
   ],
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "slot_winner_match": true,
   "slot_winner": {
    "gemini-pro": "A",
    "gpt-high": "A"
   },
   "route": "cross_slot_agree"
  },
  "dual": {
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "normalized": {
    "A": 2.0,
    "B": 1.375
   },
   "adjusted": {
    "A": 2.0,
    "B": 1.375
   },
   "violations": {},
   "per_model_winner": {
    "gemini-pro": "A",
    "gpt-high": "A"
   },
   "agreed": true
  },
  "totals": {
   "A": 2000,
   "B": 1375
  },
  "selected": "A",
  "ranking": [
   "A",
   "B"
  ],
  "verdicts": [
   {
    "label": "A",
    "score": 2000,
    "verdict_ko": "기준 이미지의 산채 공터 배경과 프레이밍 지시를 완벽하게 따랐으나, 두 발이 모두 허공에 떠 있어 '한 발은 땅을 딛고'라는 세부 포즈 지시를 놓친 점이 아쉽습니다."
   },
   {
    "label": "B",
    "score": 1375,
    "verdict_ko": "포즈 지시를 어기고 두 발이 허공에 떠 있을 뿐만 아니라, 우측 전경에 거대한 나무 울타리를 임의로 추가하여 텍스트가 요구한 탁 트인 공간감과 배경 스케일을 크게 훼손했습니다."
   }
  ],
  "refs": [
   {
    "label": "PREVIOUS SHOT STILL — a visually related earlier shot of this same place: the location's look, materials, fixed features and lighting mood are LOCKED to this photo; never copy its camera framing. The people visible in that photo are NOT in this shot: never carry their faces, bodies or clothing onto anyone here. Each person in THIS shot is defined solely by the PEOPLE list, their own CHARACTER REFERENCE images and their wardrobe notes. If this photo shows a character who cannot move (dead or unconscious), that character's exact pose, position and orientation are ALSO LOCKED — treat their body as a fixed prop of the set that only the camera moves around.",
    "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/db00517b-165b-4005-a287-f28936b59b95/images/272bd7af-9294-4490-98c6-88a40e9deacc/scene/recipe/S27sh3_sel.png",
    "asset_id": "61d5bc30-646e-459b-9130-0692fd87bcb2",
    "role": "prev_still"
   },
   {
    "label": "CHARACTER REFERENCE — 달래: the exact person appearing in this shot; match face, hair and build exactly.",
    "path": "<bytes:879791>",
    "asset_id": "83024e69-b375-4f3e-98ec-820d0e2ff1f0",
    "role": "character_ref"
   }
  ],
  "critique_skipped": true,
  "shot_run_produced": false,
  "shot_run_uid": "06aa1423-1015-7358-b038-936df0804c5c",
  "ref_mode": "prev+엔티티",
  "share_plan": {
   "ref_plan": "prev",
   "prev_anchor_tag": "S27sh3"
  }
 },
 "S29sh2::cine": {
  "applied": true,
  "attempted_at": "2026-09-09T04:09:31.421694+00:00",
  "fingerprint": "8db8653641dddbc850b4cb8624a91b64e5fbe0de4707b94f62c35962b637d318",
  "fingerprint_version": 2,
  "provider": "mai",
  "endpoint": "openrouter/chat-completions",
  "model": "microsoft/mai-image-2.6",
  "pack": "24.202608252115",
  "source_file": "S29sh2_sel.png",
  "source_sha256": "e5a0b79b4a1ae3f390443a1224add0e83bb5de494b034ab43bf8efca462ae2bf",
  "file": "S29sh2_cine.png",
  "staged_sha256": "1e38bb103615acd4ebc0ef491761c2b006771ee7b58312c8a28aea3f92318cdb",
  "latency_ms": 25472
 },
 "S29sh6::signage": {
  "fp": "069a2c72d5c9256f",
  "inscriptions": [],
  "cues": [],
  "dropped": []
 },
 "S29sh6": {
  "input_fingerprint": "174c12a38c3fc71f",
  "prompt": "Create ONE FINAL photorealistic live-action film still of the moment below — early 20th-century Korea, Jeolla region; all people are Korean unless explicitly stated otherwise. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 미간을 찌푸린 채 눈빛이 매섭게 굳어진 안규홍의 정면 클로즈업.\n\nLOCATION (lock): The mountain camp's open training ground, where the interrupted drill group has gathered to hear news. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: close-up\n- KEY BACKGROUND ELEMENTS: 산채 공터 (The conversation has interrupted training); used as A softly resolved background margin maintains the location without distracting from his face.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Maintain the same natural daytime illumination and muted palette, preserving ordinary facial texture as warmth leaves his expression.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The people visible in that photo are NOT in this shot: never carry their faces, bodies or clothing onto anyone here. Each person in THIS shot is defined solely by the PEOPLE list, their own CHARACTER REFERENCE images and their wardrobe notes. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nPREVIOUS STILL USAGE (follow exactly — what to take from the attached still and what to exclude; it governs place and objects only, never who is in this shot or how the camera sees them): Take the training clearing's ground, vegetation, and daylight visible in the reference. Exclude the messenger's running action and any dust kicked up only during her arrival.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The training branch has been set down in the camp clearing. 안규홍: He has stopped training, his expression serious; the jade ring remains concealed.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 안규홍 (한국인 남성, 청년 나이대, 검은 머리, 단단히 묶은 머리) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.",
  "roll_prompts": {
   "A": "Create ONE FINAL photorealistic live-action film still of the moment below — early 20th-century Korea, Jeolla region; all people are Korean unless explicitly stated otherwise. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 미간을 찌푸린 채 눈빛이 매섭게 굳어진 안규홍의 정면 클로즈업.\n\nLOCATION (lock): The mountain camp's open training ground, where the interrupted drill group has gathered to hear news. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: close-up\n- KEY BACKGROUND ELEMENTS: 산채 공터 (The conversation has interrupted training); used as A softly resolved background margin maintains the location without distracting from his face.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Maintain the same natural daytime illumination and muted palette, preserving ordinary facial texture as warmth leaves his expression.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The people visible in that photo are NOT in this shot: never carry their faces, bodies or clothing onto anyone here. Each person in THIS shot is defined solely by the PEOPLE list, their own CHARACTER REFERENCE images and their wardrobe notes. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nPREVIOUS STILL USAGE (follow exactly — what to take from the attached still and what to exclude; it governs place and objects only, never who is in this shot or how the camera sees them): Take the training clearing's ground, vegetation, and daylight visible in the reference. Exclude the messenger's running action and any dust kicked up only during her arrival.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The training branch has been set down in the camp clearing. 안규홍: He has stopped training, his expression serious; the jade ring remains concealed.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 안규홍 (한국인 남성, 청년 나이대, 검은 머리, 단단히 묶은 머리) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.",
   "B": "Create ONE FINAL photorealistic live-action film still of the moment below — early 20th-century Korea, Jeolla region; all people are Korean unless explicitly stated otherwise. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 미간을 찌푸린 채 눈빛이 매섭게 굳어진 안규홍의 정면 클로즈업.\n\nLOCATION (lock): The mountain camp's open training ground, where the interrupted drill group has gathered to hear news. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: close-up\n- KEY BACKGROUND ELEMENTS: 산채 공터 (The conversation has interrupted training); used as A softly resolved background margin maintains the location without distracting from his face.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Maintain the same natural daytime illumination and muted palette, preserving ordinary facial texture as warmth leaves his expression.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The people visible in that photo are NOT in this shot: never carry their faces, bodies or clothing onto anyone here. Each person in THIS shot is defined solely by the PEOPLE list, their own CHARACTER REFERENCE images and their wardrobe notes. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nPREVIOUS STILL USAGE (follow exactly — what to take from the attached still and what to exclude; it governs place and objects only, never who is in this shot or how the camera sees them): Take the training clearing's ground, vegetation, and daylight visible in the reference. Exclude the messenger's running action and any dust kicked up only during her arrival.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The training branch has been set down in the camp clearing. 안규홍: He has stopped training, his expression serious; the jade ring remains concealed.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 안규홍 (한국인 남성, 청년 나이대, 검은 머리, 단단히 묶은 머리) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.\n\nCOMPOSITION VARIATION (this candidate only): keep the place, the people and the moment exactly as specified, and keep from the CAMERA & FRAME contract its subject scale, screen placement and key background elements exactly. For this candidate only, its camera angle and height are deliberately left open: choose a camera position distinctly different from the obvious one, as a film director picking a second setup on the same blocking. Stay true to any explicit framing the SHOT TEXT itself dictates."
  },
  "shot_run_spend_attempt_count": 1,
  "readings": [
   {
    "label": "A",
    "direction": "남성의 시선과 얼굴 방향이 카메라 렌즈를 완벽하게 정면으로 향하고 있습니다.",
    "built_space": "야외 산채 공터의 모습으로, 참조 이미지와 동일하게 왼쪽에 바위 언덕, 중앙에 초가집들, 오른쪽에 나무 울타리가 흐릿한 배경(아웃포커싱)으로 정확히 배치되어 있습니다.",
    "entities": "캐릭터 참조 이미지와 일치하는 이목구비, 묶은 머리, 더러워진 흰색 한복을 입은 안규홍 한 명만 프레임 중앙에 등장하며, 프롬프트대로 미간을 찌푸린 강렬한 표정을 짓고 있습니다.",
    "hard_violations": [],
    "physics": "어깨와 가슴의 자세로 보아 중력에 맞게 지면에 안정적으로 서 있는 상태입니다."
   },
   {
    "label": "B",
    "direction": "남성의 시선은 카메라를 향하고 있으나, 어깨와 고개가 약간 틀어진 채 비스듬한 정면을 응시하고 있습니다.",
    "built_space": "참조 이미지와 일치하는 바위 언덕, 초가집, 울타리 등의 배경 요소가 흐릿하게 올바른 위치에 자리 잡고 있습니다.",
    "entities": "안규홍의 외모, 헤어스타일, 의상이 참조 이미지와 일치하며 미간을 찌푸린 표정을 보여줍니다.",
    "hard_violations": [],
    "physics": "비스듬히 선 자세로 지면에 안정적으로 서입는 것으로 보입니다."
   }
  ],
  "cross_model_order": {
   "policy": "cross_model_order_v2_slot0_fwd_slot1_rev_parallel2",
   "slots": [
    {
     "model": "gemini-pro",
     "order": "forward",
     "display_to_canonical": {
      "A": "A",
      "B": "B"
     },
     "raw": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 10,
        "verdict_ko": "프롬프트가 요구한 '정면 클로즈업'과 미간을 찌푸린 매서운 표정을 완벽한 대칭적 정면 구도로 정확하게 구현했으며, 배경 위치도 훌륭하게 유지되었습니다."
       },
       {
        "label": "B",
        "score": 9,
        "verdict_ko": "인물의 표정과 배경 묘사는 우수하나, 몸과 고개가 약간 틀어져 있어 요구된 '정면' 구도의 직관성이 A 후보에 비해 다소 떨어집니다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "남성의 시선과 얼굴 방향이 카메라 렌즈를 완벽하게 정면으로 향하고 있습니다.",
        "built_space": "야외 산채 공터의 모습으로, 참조 이미지와 동일하게 왼쪽에 바위 언덕, 중앙에 초가집들, 오른쪽에 나무 울타리가 흐릿한 배경(아웃포커싱)으로 정확히 배치되어 있습니다.",
        "entities": "캐릭터 참조 이미지와 일치하는 이목구비, 묶은 머리, 더러워진 흰색 한복을 입은 안규홍 한 명만 프레임 중앙에 등장하며, 프롬프트대로 미간을 찌푸린 강렬한 표정을 짓고 있습니다.",
        "hard_violations": [],
        "physics": "어깨와 가슴의 자세로 보아 중력에 맞게 지면에 안정적으로 서 있는 상태입니다."
       },
       {
        "label": "B",
        "direction": "남성의 시선은 카메라를 향하고 있으나, 어깨와 고개가 약간 틀어진 채 비스듬한 정면을 응시하고 있습니다.",
        "built_space": "참조 이미지와 일치하는 바위 언덕, 초가집, 울타리 등의 배경 요소가 흐릿하게 올바른 위치에 자리 잡고 있습니다.",
        "entities": "안규홍의 외모, 헤어스타일, 의상이 참조 이미지와 일치하며 미간을 찌푸린 표정을 보여줍니다.",
        "hard_violations": [],
        "physics": "비스듬히 선 자세로 지면에 안정적으로 서입는 것으로 보입니다."
       }
      ],
      "all_candidates_fail": false
     },
     "normalized": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 10,
        "verdict_ko": "프롬프트가 요구한 '정면 클로즈업'과 미간을 찌푸린 매서운 표정을 완벽한 대칭적 정면 구도로 정확하게 구현했으며, 배경 위치도 훌륭하게 유지되었습니다."
       },
       {
        "label": "B",
        "score": 9,
        "verdict_ko": "인물의 표정과 배경 묘사는 우수하나, 몸과 고개가 약간 틀어져 있어 요구된 '정면' 구도의 직관성이 A 후보에 비해 다소 떨어집니다."
       }
      ],
      "all_candidates_fail": false,
      "readings": [
       {
        "label": "A",
        "direction": "남성의 시선과 얼굴 방향이 카메라 렌즈를 완벽하게 정면으로 향하고 있습니다.",
        "built_space": "야외 산채 공터의 모습으로, 참조 이미지와 동일하게 왼쪽에 바위 언덕, 중앙에 초가집들, 오른쪽에 나무 울타리가 흐릿한 배경(아웃포커싱)으로 정확히 배치되어 있습니다.",
        "entities": "캐릭터 참조 이미지와 일치하는 이목구비, 묶은 머리, 더러워진 흰색 한복을 입은 안규홍 한 명만 프레임 중앙에 등장하며, 프롬프트대로 미간을 찌푸린 강렬한 표정을 짓고 있습니다.",
        "hard_violations": [],
        "physics": "어깨와 가슴의 자세로 보아 중력에 맞게 지면에 안정적으로 서 있는 상태입니다."
       },
       {
        "label": "B",
        "direction": "남성의 시선은 카메라를 향하고 있으나, 어깨와 고개가 약간 틀어진 채 비스듬한 정면을 응시하고 있습니다.",
        "built_space": "참조 이미지와 일치하는 바위 언덕, 초가집, 울타리 등의 배경 요소가 흐릿하게 올바른 위치에 자리 잡고 있습니다.",
        "entities": "안규홍의 외모, 헤어스타일, 의상이 참조 이미지와 일치하며 미간을 찌푸린 표정을 보여줍니다.",
        "hard_violations": [],
        "physics": "비스듬히 선 자세로 지면에 안정적으로 서입는 것으로 보입니다."
       }
      ]
     },
     "ok": true
    },
    {
     "model": "gpt-high",
     "order": "reverse",
     "display_to_canonical": {
      "A": "B",
      "B": "A"
     },
     "raw": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 8,
        "verdict_ko": "정면의 찌푸린 미간과 굳은 눈빛은 충실하지만, B보다 얼굴이 작고 오른쪽으로 치우쳐 공터 배경의 비중이 크다."
       },
       {
        "label": "B",
        "score": 9,
        "verdict_ko": "얼굴 중심의 더 밀착된 정면 클로즈업으로 매섭게 굳은 눈빛과 미간의 긴장을 강조하고, 산채는 흐린 주변 배경으로 유지한다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "얼굴은 거의 정면이고 양눈은 카메라 바로 앞쪽을 응시한다. 화면 안에 시선의 상대는 없으며, 별도의 표적을 지정하지 않은 정면 클로즈업 지시와 맞는다. 겨누는 무기나 이동 동작은 없다.",
        "built_space": "오른쪽에 놓인 인물의 머리와 윗가슴이 보이고 왼쪽 배경 여백이 넓다. 뒤에는 큰 초가 한 채, 인물 오른쪽 뒤의 작은 초가 지붕 일부, 맨 오른쪽 가장자리의 지붕 일부가 보인다. 왼쪽 바위 비탈, 흙 공터, 뒤쪽 목책과 숲이 이전 장소와 이어진다. 배경은 흐리지만 비교적 넓고 구체적으로 드러난다. 반사면이나 중복된 시설은 보이지 않는다.",
        "entities": "젊은 한국인 남성으로 설정된 안규홍 한 명만 보인다. 검은 머리를 뒤로 팽팽하게 넘긴 형태, 얼굴 윤곽과 이목구비, 때 묻은 미색 옷깃이 인물 참조와 대체로 맞는다. 뒤쪽 머리 매듭은 가려져 확인할 수 없다. 미간이 깊게 접히고 입술이 닫혀 있으며 눈은 정상적인 홍채와 동공을 유지한다. 손, 훈련용 나뭇가지와 옥가락지는 구도 밖이므로 내려놓은 상태나 은닉 여부를 직접 확인할 수 없다. 추가 인물과 읽을 수 있는 글자는 없다.",
        "hard_violations": [],
        "physics": "머리는 목과 어깨에 자연스럽게 연결되고 옷은 어깨와 몸통에 걸쳐 아래로 처진다. 하체와 발의 지면 접촉은 클로즈업 밖이지만, 보이는 상체에 부유나 불가능한 자세의 징후는 없다. 배경 건물과 바위는 지면에 놓여 있고 공중에 뜬 물체나 도착 장면의 먼지는 없다."
       },
       {
        "label": "B",
        "direction": "얼굴과 어깨가 정면에 가깝고 양눈은 카메라 바로 앞쪽을 날카롭게 응시한다. 화면 안의 다른 인물을 보는 구도는 아니다. 정면 클로즈업과 굳어진 눈빛이라는 지시에 부합하며, 방향을 확인할 무기나 이동체는 없다.",
        "built_space": "인물의 얼굴이 화면 중앙을 크게 차지하고 아래쪽에는 어깨와 옷깃만 남는다. 배경에는 왼쪽 뒤 큰 초가의 지붕 일부, 오른쪽 뒤 작은 초가 한 채, 맨 오른쪽 끝 초가 일부가 보인다. 왼쪽 바위 비탈과 뒤쪽 목책, 숲, 흙 공터가 흐리게 이어져 이전 장소의 재료와 지형을 유지한다. 배경 시설은 인물보다 작고 멀리 놓여 있으며 불가능한 반사나 명백한 시설 중복은 없다.",
        "entities": "안규홍에 해당하는 젊은 한국인 남성 한 명만 보인다. 참조의 검은색으로 단단히 넘긴 머리, 얼굴형과 이목구비, 낡고 때 묻은 미색 옷이 대체로 유지된다. 머리의 실제 매듭은 뒤에 가려져 있다. 모인 눈썹과 깊어진 미간, 긴장된 입매가 심각하고 매서운 표정을 만든다. 안구는 자연스럽고 피부 질감도 실제 얼굴처럼 보인다. 손과 나뭇가지, 옥가락지는 프레임에 없으며 추가 인물이나 읽을 수 있는 문자는 없다.",
        "hard_violations": [],
        "physics": "목이 머리를 받치고 양어깨가 옷을 지지하는 자연스러운 정지 자세다. 발과 지면의 접촉은 프레임 밖이라 확인할 수 없으나 상체의 연결과 무게감에 이상은 없다. 들고 있거나 날아가는 물체가 없고, 배경 건물과 바위는 지면에 안정적으로 놓여 있다."
       }
      ],
      "all_candidates_fail": false
     },
     "normalized": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 8,
        "verdict_ko": "정면의 찌푸린 미간과 굳은 눈빛은 충실하지만, B보다 얼굴이 작고 오른쪽으로 치우쳐 공터 배경의 비중이 크다."
       },
       {
        "label": "A",
        "score": 9,
        "verdict_ko": "얼굴 중심의 더 밀착된 정면 클로즈업으로 매섭게 굳은 눈빛과 미간의 긴장을 강조하고, 산채는 흐린 주변 배경으로 유지한다."
       }
      ],
      "all_candidates_fail": false,
      "readings": [
       {
        "label": "B",
        "direction": "얼굴은 거의 정면이고 양눈은 카메라 바로 앞쪽을 응시한다. 화면 안에 시선의 상대는 없으며, 별도의 표적을 지정하지 않은 정면 클로즈업 지시와 맞는다. 겨누는 무기나 이동 동작은 없다.",
        "built_space": "오른쪽에 놓인 인물의 머리와 윗가슴이 보이고 왼쪽 배경 여백이 넓다. 뒤에는 큰 초가 한 채, 인물 오른쪽 뒤의 작은 초가 지붕 일부, 맨 오른쪽 가장자리의 지붕 일부가 보인다. 왼쪽 바위 비탈, 흙 공터, 뒤쪽 목책과 숲이 이전 장소와 이어진다. 배경은 흐리지만 비교적 넓고 구체적으로 드러난다. 반사면이나 중복된 시설은 보이지 않는다.",
        "entities": "젊은 한국인 남성으로 설정된 안규홍 한 명만 보인다. 검은 머리를 뒤로 팽팽하게 넘긴 형태, 얼굴 윤곽과 이목구비, 때 묻은 미색 옷깃이 인물 참조와 대체로 맞는다. 뒤쪽 머리 매듭은 가려져 확인할 수 없다. 미간이 깊게 접히고 입술이 닫혀 있으며 눈은 정상적인 홍채와 동공을 유지한다. 손, 훈련용 나뭇가지와 옥가락지는 구도 밖이므로 내려놓은 상태나 은닉 여부를 직접 확인할 수 없다. 추가 인물과 읽을 수 있는 글자는 없다.",
        "hard_violations": [],
        "physics": "머리는 목과 어깨에 자연스럽게 연결되고 옷은 어깨와 몸통에 걸쳐 아래로 처진다. 하체와 발의 지면 접촉은 클로즈업 밖이지만, 보이는 상체에 부유나 불가능한 자세의 징후는 없다. 배경 건물과 바위는 지면에 놓여 있고 공중에 뜬 물체나 도착 장면의 먼지는 없다."
       },
       {
        "label": "A",
        "direction": "얼굴과 어깨가 정면에 가깝고 양눈은 카메라 바로 앞쪽을 날카롭게 응시한다. 화면 안의 다른 인물을 보는 구도는 아니다. 정면 클로즈업과 굳어진 눈빛이라는 지시에 부합하며, 방향을 확인할 무기나 이동체는 없다.",
        "built_space": "인물의 얼굴이 화면 중앙을 크게 차지하고 아래쪽에는 어깨와 옷깃만 남는다. 배경에는 왼쪽 뒤 큰 초가의 지붕 일부, 오른쪽 뒤 작은 초가 한 채, 맨 오른쪽 끝 초가 일부가 보인다. 왼쪽 바위 비탈과 뒤쪽 목책, 숲, 흙 공터가 흐리게 이어져 이전 장소의 재료와 지형을 유지한다. 배경 시설은 인물보다 작고 멀리 놓여 있으며 불가능한 반사나 명백한 시설 중복은 없다.",
        "entities": "안규홍에 해당하는 젊은 한국인 남성 한 명만 보인다. 참조의 검은색으로 단단히 넘긴 머리, 얼굴형과 이목구비, 낡고 때 묻은 미색 옷이 대체로 유지된다. 머리의 실제 매듭은 뒤에 가려져 있다. 모인 눈썹과 깊어진 미간, 긴장된 입매가 심각하고 매서운 표정을 만든다. 안구는 자연스럽고 피부 질감도 실제 얼굴처럼 보인다. 손과 나뭇가지, 옥가락지는 프레임에 없으며 추가 인물이나 읽을 수 있는 문자는 없다.",
        "hard_violations": [],
        "physics": "목이 머리를 받치고 양어깨가 옷을 지지하는 자연스러운 정지 자세다. 발과 지면의 접촉은 프레임 밖이라 확인할 수 없으나 상체의 연결과 무게감에 이상은 없다. 들고 있거나 날아가는 물체가 없고, 배경 건물과 바위는 지면에 안정적으로 놓여 있다."
       }
      ]
     },
     "ok": true
    }
   ],
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "slot_winner_match": true,
   "slot_winner": {
    "gemini-pro": "A",
    "gpt-high": "A"
   },
   "route": "cross_slot_agree"
  },
  "dual": {
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "normalized": {
    "A": 2.0,
    "B": 1.789
   },
   "adjusted": {
    "A": 2.0,
    "B": 1.789
   },
   "violations": {},
   "per_model_winner": {
    "gemini-pro": "A",
    "gpt-high": "A"
   },
   "agreed": true
  },
  "totals": {
   "A": 2000,
   "B": 1789
  },
  "selected": "A",
  "ranking": [
   "A",
   "B"
  ],
  "verdicts": [
   {
    "label": "A",
    "score": 2000,
    "verdict_ko": "프롬프트가 요구한 '정면 클로즈업'과 미간을 찌푸린 매서운 표정을 완벽한 대칭적 정면 구도로 정확하게 구현했으며, 배경 위치도 훌륭하게 유지되었습니다."
   },
   {
    "label": "B",
    "score": 1789,
    "verdict_ko": "인물의 표정과 배경 묘사는 우수하나, 몸과 고개가 약간 틀어져 있어 요구된 '정면' 구도의 직관성이 A 후보에 비해 다소 떨어집니다."
   }
  ],
  "refs": [
   {
    "label": "PREVIOUS SHOT STILL — a visually related earlier shot of this same place: the location's look, materials, fixed features and lighting mood are LOCKED to this photo; never copy its camera framing. The people visible in that photo are NOT in this shot: never carry their faces, bodies or clothing onto anyone here. Each person in THIS shot is defined solely by the PEOPLE list, their own CHARACTER REFERENCE images and their wardrobe notes. If this photo shows a character who cannot move (dead or unconscious), that character's exact pose, position and orientation are ALSO LOCKED — treat their body as a fixed prop of the set that only the camera moves around.",
    "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/db00517b-165b-4005-a287-f28936b59b95/images/272bd7af-9294-4490-98c6-88a40e9deacc/scene/recipe/S29sh2_sel.png",
    "asset_id": "035ce225-e75f-4055-9bb3-ca1a292beba1",
    "role": "prev_still"
   },
   {
    "label": "CHARACTER REFERENCE — 안규홍: the exact person appearing in this shot; match face, hair and build exactly.",
    "path": "<bytes:1028384>",
    "asset_id": "3b5894db-ce55-4f34-b966-5e725c3aa952",
    "role": "character_ref"
   }
  ],
  "critique_skipped": true,
  "shot_run_produced": false,
  "shot_run_uid": "06aa1423-11a4-70a6-9cec-7144fff50e92",
  "ref_mode": "prev+엔티티",
  "share_plan": {
   "ref_plan": "prev",
   "prev_anchor_tag": "S29sh2"
  }
 },
 "S29sh6::cine": {
  "applied": true,
  "attempted_at": "2026-09-09T04:10:41.902140+00:00",
  "fingerprint": "e196c152add2a587eb5cbd2373becaadf27d70ca838ea38793b51553bb405a71",
  "fingerprint_version": 2,
  "provider": "mai",
  "endpoint": "openrouter/chat-completions",
  "model": "microsoft/mai-image-2.6",
  "pack": "24.202608252115",
  "source_file": "S29sh6_sel.png",
  "source_sha256": "d77ed75c51d2c124aba6ce63192136bd959e7f0ab44807b3dcf532067bcc6d71",
  "file": "S29sh6_cine.png",
  "staged_sha256": "a3c7a6ea5f0cb6e9f0f98095d90e2b30833b93c0195ecbff9bf770cf8d01c74b",
  "latency_ms": 26356
 },
 "S29sh12::signage": {
  "fp": "05ef46d92d59cd80",
  "inscriptions": [],
  "cues": [],
  "dropped": []
 },
 "S29sh12": {
  "input_fingerprint": "cfd9ee3f3e121bed",
  "prompt": "Create ONE FINAL photorealistic live-action film still of the moment below — early 20th-century Korea, Jeolla region; all people are Korean unless explicitly stated otherwise. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 안규홍을 향해 일제히 묵묵한 시선을 보내는 오주일과 복만, 낡은 옷차림의 조선인 의병들의 전경.\n\nLOCATION (lock): The outdoor gathering space within the mountain camp's training clearing. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: wide shot\n- KEY BACKGROUND ELEMENTS: 훈련하던 공터 (Training has stopped while the men listen); used as Visible intervals between the men establish a gathering rather than a uniform formation; worn clothing and distinct shoulder angles separate individuals.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Keep the daytime ambient light and restrained saturation continuous, with sufficient tonal openness to distinguish individual faces throughout the gathering.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The people visible in that photo are NOT in this shot: never carry their faces, bodies or clothing onto anyone here. Each person in THIS shot is defined solely by the PEOPLE list, their own CHARACTER REFERENCE images and their wardrobe notes. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nPREVIOUS STILL USAGE (follow exactly — what to take from the attached still and what to exclude; it governs place and objects only, never who is in this shot or how the camera sees them): Take the clearing's dirt surface, vegetation, and daytime lighting. Exclude the arrival's transient dust and do not duplicate the messenger across the gathering.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The training branch remains where it was set down in the daylight clearing. 오주일: He stands with the branch no longer in hand. 복만: He remains standing with his matchlock.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 오주일 (한국인 남성, 성인 나이대, 자연스러운 검은 머리); 복만 (한국인 남성, 성인 나이대, 거칠게 자란 수염, 자연스러운 검은 머리) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.",
  "roll_prompts": {
   "A": "Create ONE FINAL photorealistic live-action film still of the moment below — early 20th-century Korea, Jeolla region; all people are Korean unless explicitly stated otherwise. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 안규홍을 향해 일제히 묵묵한 시선을 보내는 오주일과 복만, 낡은 옷차림의 조선인 의병들의 전경.\n\nLOCATION (lock): The outdoor gathering space within the mountain camp's training clearing. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: wide shot\n- KEY BACKGROUND ELEMENTS: 훈련하던 공터 (Training has stopped while the men listen); used as Visible intervals between the men establish a gathering rather than a uniform formation; worn clothing and distinct shoulder angles separate individuals.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Keep the daytime ambient light and restrained saturation continuous, with sufficient tonal openness to distinguish individual faces throughout the gathering.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The people visible in that photo are NOT in this shot: never carry their faces, bodies or clothing onto anyone here. Each person in THIS shot is defined solely by the PEOPLE list, their own CHARACTER REFERENCE images and their wardrobe notes. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nPREVIOUS STILL USAGE (follow exactly — what to take from the attached still and what to exclude; it governs place and objects only, never who is in this shot or how the camera sees them): Take the clearing's dirt surface, vegetation, and daytime lighting. Exclude the arrival's transient dust and do not duplicate the messenger across the gathering.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The training branch remains where it was set down in the daylight clearing. 오주일: He stands with the branch no longer in hand. 복만: He remains standing with his matchlock.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 오주일 (한국인 남성, 성인 나이대, 자연스러운 검은 머리); 복만 (한국인 남성, 성인 나이대, 거칠게 자란 수염, 자연스러운 검은 머리) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.",
   "B": "Create ONE FINAL photorealistic live-action film still of the moment below — early 20th-century Korea, Jeolla region; all people are Korean unless explicitly stated otherwise. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 안규홍을 향해 일제히 묵묵한 시선을 보내는 오주일과 복만, 낡은 옷차림의 조선인 의병들의 전경.\n\nLOCATION (lock): The outdoor gathering space within the mountain camp's training clearing. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: wide shot\n- KEY BACKGROUND ELEMENTS: 훈련하던 공터 (Training has stopped while the men listen); used as Visible intervals between the men establish a gathering rather than a uniform formation; worn clothing and distinct shoulder angles separate individuals.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Keep the daytime ambient light and restrained saturation continuous, with sufficient tonal openness to distinguish individual faces throughout the gathering.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The people visible in that photo are NOT in this shot: never carry their faces, bodies or clothing onto anyone here. Each person in THIS shot is defined solely by the PEOPLE list, their own CHARACTER REFERENCE images and their wardrobe notes. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nPREVIOUS STILL USAGE (follow exactly — what to take from the attached still and what to exclude; it governs place and objects only, never who is in this shot or how the camera sees them): Take the clearing's dirt surface, vegetation, and daytime lighting. Exclude the arrival's transient dust and do not duplicate the messenger across the gathering.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The training branch remains where it was set down in the daylight clearing. 오주일: He stands with the branch no longer in hand. 복만: He remains standing with his matchlock.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 오주일 (한국인 남성, 성인 나이대, 자연스러운 검은 머리); 복만 (한국인 남성, 성인 나이대, 거칠게 자란 수염, 자연스러운 검은 머리) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.\n\nCOMPOSITION VARIATION (this candidate only): keep the place, the people and the moment exactly as specified, and keep from the CAMERA & FRAME contract its subject scale, screen placement and key background elements exactly. For this candidate only, its camera angle and height are deliberately left open: choose a camera position distinctly different from the obvious one, as a film director picking a second setup on the same blocking. Stay true to any explicit framing the SHOT TEXT itself dictates."
  },
  "shot_run_spend_attempt_count": 1,
  "readings": [
   {
    "label": "A",
    "direction": "오주일, 복만과 의병들이 화면 우측 전경에 있는 초록색 옷을 입은 인물을 향해 일제히 시선을 보내고 있음. 시선의 대상은 우측 전경의 초록색 옷 인물임.",
    "built_space": "참조 이미지에 등장한 흙바닥 공터, 초가집, 바위, 뒷산의 소나무 등 야외 훈련장의 고정된 구조물과 지형이 올바르게 배치되어 있음.",
    "entities": "오주일(중앙, 푸른 제복)은 참조와 일치함. 복만(우측, 검은 옷)은 얼굴은 비슷하나 참조의 갓을 쓰지 않음. 프롬프트가 엄격히 금지했음에도 전경 인물이 이전 샷 전령의 초록색 상의를 그대로 입고 있음. 바닥에 나뭇가지가 배치됨.",
    "hard_violations": [
     "[gemini-pro] 금지된 복장 복제 (이전 샷 전령의 초록색 옷을 전경 인물에게 적용함)",
     "[gemini-pro] 물리적으로 불가능한 해부학 (전경 인물의 기형적으로 작은 오른손)",
     "[gpt-high] 이전 장면 인물의 의복을 누구에게도 옮기지 말라는 명시적 금지와 달리, 오른쪽 전경의 별도 인물에 전령의 연녹색 윗옷·검정 하의 복장을 재사용했다."
    ],
    "physics": "전경에 있는 초록색 옷 인물의 오른손이 물리적으로 불가능할 정도로 작게 왜곡됨. 오주일의 어깨에 멘 소총 하단이 팔과 융합된 것처럼 보임. 인물들의 발은 땅에 지탱되어 있음."
   },
   {
    "label": "B",
    "direction": "오주일, 복만과 의병들이 화면 좌측 전경에 등을 보이고 서 있는 두 인물(안규홍 일행으로 추정)을 향해 묵묵한 시선을 보내고 있음. 시선의 대상은 좌측 전경의 인물들임.",
    "built_space": "초가집, 울타리, 바위, 소나무 등 참조 이미지의 야외 공터 배경 요소들이 정확한 위치와 비례로 구현되어 있음.",
    "entities": "오주일(중앙, 푸른 제복)과 복만(우측, 검은 옷과 갓) 모두 참조 이미지와 외형 및 복장이 정확히 일치함. 의병들도 낡은 옷차림으로 적절히 등장함. 그러나 오주일 앞 바닥에 있어야 할 나뭇가지가 없음.",
    "hard_violations": [
     "[gemini-pro] 발명된 사물 (프롬프트에 명시된 훈련용 나뭇가지 대신 화승총이 바닥에 뜬금없이 놓여 있음)"
    ],
    "physics": "복만이 화승총을 두 손으로 자연스럽게 쥐고 있으며, 바닥에 놓인 총과 서 있는 모든 인물이 땅에 올바르게 지탱되어 있음. 시선과 자세가 물리적인 제약 안에서 자연스러움."
   }
  ],
  "cross_model_order": {
   "policy": "cross_model_order_v2_slot0_fwd_slot1_rev_parallel2",
   "slots": [
    {
     "model": "gemini-pro",
     "order": "forward",
     "display_to_canonical": {
      "A": "A",
      "B": "B"
     },
     "raw": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 3,
        "verdict_ko": "이전 샷의 전령 복장을 사용하지 말라는 네거티브 지시를 위반하여 전경 인물에게 입혔으며, 해당 인물의 손 등 해부학적 오류가 매우 뚜렷합니다."
       },
       {
        "label": "B",
        "score": 8,
        "verdict_ko": "오주일과 복만의 외형을 참조와 완벽하게 일치시키고 긴장감 있는 시선 처리를 잘 연출했으나, 바닥에 명시된 나뭇가지 대신 총을 배치한 점이 아쉽습니다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "오주일, 복만과 의병들이 화면 우측 전경에 있는 초록색 옷을 입은 인물을 향해 일제히 시선을 보내고 있음. 시선의 대상은 우측 전경의 초록색 옷 인물임.",
        "built_space": "참조 이미지에 등장한 흙바닥 공터, 초가집, 바위, 뒷산의 소나무 등 야외 훈련장의 고정된 구조물과 지형이 올바르게 배치되어 있음.",
        "entities": "오주일(중앙, 푸른 제복)은 참조와 일치함. 복만(우측, 검은 옷)은 얼굴은 비슷하나 참조의 갓을 쓰지 않음. 프롬프트가 엄격히 금지했음에도 전경 인물이 이전 샷 전령의 초록색 상의를 그대로 입고 있음. 바닥에 나뭇가지가 배치됨.",
        "hard_violations": [
         "금지된 복장 복제 (이전 샷 전령의 초록색 옷을 전경 인물에게 적용함)",
         "물리적으로 불가능한 해부학 (전경 인물의 기형적으로 작은 오른손)"
        ],
        "physics": "전경에 있는 초록색 옷 인물의 오른손이 물리적으로 불가능할 정도로 작게 왜곡됨. 오주일의 어깨에 멘 소총 하단이 팔과 융합된 것처럼 보임. 인물들의 발은 땅에 지탱되어 있음."
       },
       {
        "label": "B",
        "direction": "오주일, 복만과 의병들이 화면 좌측 전경에 등을 보이고 서 있는 두 인물(안규홍 일행으로 추정)을 향해 묵묵한 시선을 보내고 있음. 시선의 대상은 좌측 전경의 인물들임.",
        "built_space": "초가집, 울타리, 바위, 소나무 등 참조 이미지의 야외 공터 배경 요소들이 정확한 위치와 비례로 구현되어 있음.",
        "entities": "오주일(중앙, 푸른 제복)과 복만(우측, 검은 옷과 갓) 모두 참조 이미지와 외형 및 복장이 정확히 일치함. 의병들도 낡은 옷차림으로 적절히 등장함. 그러나 오주일 앞 바닥에 있어야 할 나뭇가지가 없음.",
        "hard_violations": [
         "발명된 사물 (프롬프트에 명시된 훈련용 나뭇가지 대신 화승총이 바닥에 뜬금없이 놓여 있음)"
        ],
        "physics": "복만이 화승총을 두 손으로 자연스럽게 쥐고 있으며, 바닥에 놓인 총과 서 있는 모든 인물이 땅에 올바르게 지탱되어 있음. 시선과 자세가 물리적인 제약 안에서 자연스러움."
       }
      ],
      "all_candidates_fail": false
     },
     "normalized": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 3,
        "verdict_ko": "이전 샷의 전령 복장을 사용하지 말라는 네거티브 지시를 위반하여 전경 인물에게 입혔으며, 해당 인물의 손 등 해부학적 오류가 매우 뚜렷합니다."
       },
       {
        "label": "B",
        "score": 8,
        "verdict_ko": "오주일과 복만의 외형을 참조와 완벽하게 일치시키고 긴장감 있는 시선 처리를 잘 연출했으나, 바닥에 명시된 나뭇가지 대신 총을 배치한 점이 아쉽습니다."
       }
      ],
      "all_candidates_fail": false,
      "readings": [
       {
        "label": "A",
        "direction": "오주일, 복만과 의병들이 화면 우측 전경에 있는 초록색 옷을 입은 인물을 향해 일제히 시선을 보내고 있음. 시선의 대상은 우측 전경의 초록색 옷 인물임.",
        "built_space": "참조 이미지에 등장한 흙바닥 공터, 초가집, 바위, 뒷산의 소나무 등 야외 훈련장의 고정된 구조물과 지형이 올바르게 배치되어 있음.",
        "entities": "오주일(중앙, 푸른 제복)은 참조와 일치함. 복만(우측, 검은 옷)은 얼굴은 비슷하나 참조의 갓을 쓰지 않음. 프롬프트가 엄격히 금지했음에도 전경 인물이 이전 샷 전령의 초록색 상의를 그대로 입고 있음. 바닥에 나뭇가지가 배치됨.",
        "hard_violations": [
         "금지된 복장 복제 (이전 샷 전령의 초록색 옷을 전경 인물에게 적용함)",
         "물리적으로 불가능한 해부학 (전경 인물의 기형적으로 작은 오른손)"
        ],
        "physics": "전경에 있는 초록색 옷 인물의 오른손이 물리적으로 불가능할 정도로 작게 왜곡됨. 오주일의 어깨에 멘 소총 하단이 팔과 융합된 것처럼 보임. 인물들의 발은 땅에 지탱되어 있음."
       },
       {
        "label": "B",
        "direction": "오주일, 복만과 의병들이 화면 좌측 전경에 등을 보이고 서 있는 두 인물(안규홍 일행으로 추정)을 향해 묵묵한 시선을 보내고 있음. 시선의 대상은 좌측 전경의 인물들임.",
        "built_space": "초가집, 울타리, 바위, 소나무 등 참조 이미지의 야외 공터 배경 요소들이 정확한 위치와 비례로 구현되어 있음.",
        "entities": "오주일(중앙, 푸른 제복)과 복만(우측, 검은 옷과 갓) 모두 참조 이미지와 외형 및 복장이 정확히 일치함. 의병들도 낡은 옷차림으로 적절히 등장함. 그러나 오주일 앞 바닥에 있어야 할 나뭇가지가 없음.",
        "hard_violations": [
         "발명된 사물 (프롬프트에 명시된 훈련용 나뭇가지 대신 화승총이 바닥에 뜬금없이 놓여 있음)"
        ],
        "physics": "복만이 화승총을 두 손으로 자연스럽게 쥐고 있으며, 바닥에 놓인 총과 서 있는 모든 인물이 땅에 올바르게 지탱되어 있음. 시선과 자세가 물리적인 제약 안에서 자연스러움."
       }
      ]
     },
     "ok": true
    },
    {
     "model": "gpt-high",
     "order": "reverse",
     "display_to_canonical": {
      "A": "B",
      "B": "A"
     },
     "raw": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 5,
        "verdict_ko": "넓은 공터의 불규칙한 모임과 두 주인공의 참조 복장은 잘 맞지만, 전경 인물들 때문에 시선의 단일 목표가 불분명하고 내려놓은 훈련용 나뭇가지도 확인되지 않는다."
       },
       {
        "label": "B",
        "score": 3,
        "verdict_ko": "오른쪽 인물에게 모이는 시선과 땅에 놓인 나뭇가지는 명확하지만, 이전 장면의 연녹색·검정 복장을 전경 인물에 재사용하고 두 주인공의 지정 모자를 누락했다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "오주일과 복만, 뒤쪽 의병들은 대체로 화면 왼쪽 전경을 바라본다. 그곳에는 등을 보인 남자가 두 명 있어 누구를 안규홍으로 바라보는지 특정하기 어렵다. 전경 두 사람은 모임 쪽으로 몸을 돌려 다른 사람들과 동일한 방향의 시선을 보내지는 않는다. 복만의 총구는 오른쪽 위, 회청색 옷을 입은 의병의 어깨 옆 공간을 향하며 조준 자세는 아니다. 바닥 총의 총구는 왼쪽 아래를 향한다.",
        "built_space": "돌이 섞인 흙 공터, 왼쪽 바위 비탈, 뒤쪽 숲이 보인다. 뒤 오른쪽에는 큰 초가 한 채와 벽 앞 장작더미가 있고, 왼쪽 위에도 초가 일부가 보인다. 참조 장소의 재료와 낮 조명은 대체로 이어지지만 건물과 바위의 정확한 대응은 확정하기 어렵다. 사람들은 공터에 서로 다른 간격으로 서 있어 정렬된 훈련 대형보다는 모임으로 읽힌다. 반사면이나 중복된 고정 설비 문제는 보이지 않는다.",
        "entities": "오주일은 젊은 성인 동아시아계 남성으로, 참조의 검은 제모와 기운 군복, 허리띠, 밝은 각반을 유지하며 얼굴도 대체로 닮았다. 복만은 검은 수염의 성인 동아시아계 남성으로, 참조의 챙 있는 모자와 낡은 긴 겉옷을 착용하고 긴 총을 들었다. 주변에는 낡은 조선식 옷을 입은 남성 의병들이 있으며 이전 장면의 여성 전령은 보이지 않는다. 오주일의 손은 비어 있지만, 발치 물체는 훈련용 나뭇가지가 아니라 개머리판과 총열이 있는 소총으로 보인다. 읽을 수 있는 글자는 없다.",
        "hard_violations": [],
        "physics": "전신이 보이는 인물들은 발로 흙바닥을 딛고 서 있으며 공중에 뜬 몸은 없다. 전경에서 하체가 잘린 두 사람도 서 있는 상체로 자연스럽게 이어진다. 복만은 양손으로 총의 앞부분과 개머리판 쪽을 받치고 있다. 내려놓은 소총은 흙바닥에 닿아 있으며, 지지 없는 물체나 명백히 불가능한 관절은 보이지 않는다."
       },
       {
        "label": "B",
        "direction": "오주일과 복만의 얼굴과 시선은 오른쪽 전경의 연녹색 옷 남자를 향한다. 뒤쪽 의병들도 대체로 그쪽을 보아 듣는 대상은 A보다 명확하지만, 그 인물을 안규홍으로 확인할 근거는 없다. 전경 남자는 왼쪽의 두 주인공을 마주 본다. 복만의 총구는 위쪽을 향하고 개머리판은 바닥에 놓여 있으며, 오주일의 어깨 뒤 총구도 위를 향한다. 다른 의병들의 총 역시 사람을 겨누는 자세는 아니다.",
        "built_space": "흙과 자갈이 섞인 공터 뒤로 바위 비탈과 숲이 이어진다. 왼쪽 중앙의 큰 초가 한 채, 중앙 뒤의 낮은 초가 한 채, 오른쪽 가장자리의 초가 한 채가 보이고 오른쪽 뒤에는 목책이 있다. 참조의 공터 재질과 식생, 낮의 부드러운 빛을 잘 따른다. 두 주인공은 앞쪽 중앙, 의병들은 간격을 두고 뒤쪽에 서 있으나 오른쪽 전경 인물이 모임의 상당 부분을 가린다. 불가능한 반사나 명백한 설비 중복은 없다.",
        "entities": "오주일은 참조와 비슷한 젊은 성인 동아시아계 남성이며 기운 군복과 각반을 착용하지만 지정된 검은 제모가 없다. 복만은 검은 수염과 낡은 긴 겉옷을 유지하지만 참조의 챙 있는 모자 대신 머리띠와 상투를 했다. 복만에게는 긴 총이 있고 오주일에게도 어깨에 멘 총이 보인다. 왼쪽 아래에는 실제 가지 형태의 훈련용 나뭇가지가 놓여 있다. 오른쪽 전경 남자는 두 지정 주인공과 다른 인물이며, 연녹색 윗옷과 검정 하의 조합이 이전 장면 전령의 복장을 재사용한 것으로 읽힌다. 주변 의병들은 낡은 조선식 옷차림이며 읽을 수 있는 글자는 없다.",
        "hard_violations": [
         "이전 장면 인물의 의복을 누구에게도 옮기지 말라는 명시적 금지와 달리, 오른쪽 전경의 별도 인물에 전령의 연녹색 윗옷·검정 하의 복장을 재사용했다."
        ],
        "physics": "두 주인공과 뒤쪽 의병들은 발로 지면을 딛고 있다. 복만의 총은 개머리판이 땅에 닿고 손이 몸통을 잡아 지지한다. 오주일의 총은 어깨를 지나는 멜빵으로 지지되며, 주변 의병들의 총도 손 또는 지면의 받침이 보인다. 나뭇가지는 흙 위에 놓여 있다. 전경 인물의 발은 프레임 밖이지만 보이는 자세에서 부유나 불가능한 동작은 발견되지 않는다."
       }
      ],
      "all_candidates_fail": true
     },
     "normalized": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 5,
        "verdict_ko": "넓은 공터의 불규칙한 모임과 두 주인공의 참조 복장은 잘 맞지만, 전경 인물들 때문에 시선의 단일 목표가 불분명하고 내려놓은 훈련용 나뭇가지도 확인되지 않는다."
       },
       {
        "label": "A",
        "score": 3,
        "verdict_ko": "오른쪽 인물에게 모이는 시선과 땅에 놓인 나뭇가지는 명확하지만, 이전 장면의 연녹색·검정 복장을 전경 인물에 재사용하고 두 주인공의 지정 모자를 누락했다."
       }
      ],
      "all_candidates_fail": true,
      "readings": [
       {
        "label": "B",
        "direction": "오주일과 복만, 뒤쪽 의병들은 대체로 화면 왼쪽 전경을 바라본다. 그곳에는 등을 보인 남자가 두 명 있어 누구를 안규홍으로 바라보는지 특정하기 어렵다. 전경 두 사람은 모임 쪽으로 몸을 돌려 다른 사람들과 동일한 방향의 시선을 보내지는 않는다. 복만의 총구는 오른쪽 위, 회청색 옷을 입은 의병의 어깨 옆 공간을 향하며 조준 자세는 아니다. 바닥 총의 총구는 왼쪽 아래를 향한다.",
        "built_space": "돌이 섞인 흙 공터, 왼쪽 바위 비탈, 뒤쪽 숲이 보인다. 뒤 오른쪽에는 큰 초가 한 채와 벽 앞 장작더미가 있고, 왼쪽 위에도 초가 일부가 보인다. 참조 장소의 재료와 낮 조명은 대체로 이어지지만 건물과 바위의 정확한 대응은 확정하기 어렵다. 사람들은 공터에 서로 다른 간격으로 서 있어 정렬된 훈련 대형보다는 모임으로 읽힌다. 반사면이나 중복된 고정 설비 문제는 보이지 않는다.",
        "entities": "오주일은 젊은 성인 동아시아계 남성으로, 참조의 검은 제모와 기운 군복, 허리띠, 밝은 각반을 유지하며 얼굴도 대체로 닮았다. 복만은 검은 수염의 성인 동아시아계 남성으로, 참조의 챙 있는 모자와 낡은 긴 겉옷을 착용하고 긴 총을 들었다. 주변에는 낡은 조선식 옷을 입은 남성 의병들이 있으며 이전 장면의 여성 전령은 보이지 않는다. 오주일의 손은 비어 있지만, 발치 물체는 훈련용 나뭇가지가 아니라 개머리판과 총열이 있는 소총으로 보인다. 읽을 수 있는 글자는 없다.",
        "hard_violations": [],
        "physics": "전신이 보이는 인물들은 발로 흙바닥을 딛고 서 있으며 공중에 뜬 몸은 없다. 전경에서 하체가 잘린 두 사람도 서 있는 상체로 자연스럽게 이어진다. 복만은 양손으로 총의 앞부분과 개머리판 쪽을 받치고 있다. 내려놓은 소총은 흙바닥에 닿아 있으며, 지지 없는 물체나 명백히 불가능한 관절은 보이지 않는다."
       },
       {
        "label": "A",
        "direction": "오주일과 복만의 얼굴과 시선은 오른쪽 전경의 연녹색 옷 남자를 향한다. 뒤쪽 의병들도 대체로 그쪽을 보아 듣는 대상은 A보다 명확하지만, 그 인물을 안규홍으로 확인할 근거는 없다. 전경 남자는 왼쪽의 두 주인공을 마주 본다. 복만의 총구는 위쪽을 향하고 개머리판은 바닥에 놓여 있으며, 오주일의 어깨 뒤 총구도 위를 향한다. 다른 의병들의 총 역시 사람을 겨누는 자세는 아니다.",
        "built_space": "흙과 자갈이 섞인 공터 뒤로 바위 비탈과 숲이 이어진다. 왼쪽 중앙의 큰 초가 한 채, 중앙 뒤의 낮은 초가 한 채, 오른쪽 가장자리의 초가 한 채가 보이고 오른쪽 뒤에는 목책이 있다. 참조의 공터 재질과 식생, 낮의 부드러운 빛을 잘 따른다. 두 주인공은 앞쪽 중앙, 의병들은 간격을 두고 뒤쪽에 서 있으나 오른쪽 전경 인물이 모임의 상당 부분을 가린다. 불가능한 반사나 명백한 설비 중복은 없다.",
        "entities": "오주일은 참조와 비슷한 젊은 성인 동아시아계 남성이며 기운 군복과 각반을 착용하지만 지정된 검은 제모가 없다. 복만은 검은 수염과 낡은 긴 겉옷을 유지하지만 참조의 챙 있는 모자 대신 머리띠와 상투를 했다. 복만에게는 긴 총이 있고 오주일에게도 어깨에 멘 총이 보인다. 왼쪽 아래에는 실제 가지 형태의 훈련용 나뭇가지가 놓여 있다. 오른쪽 전경 남자는 두 지정 주인공과 다른 인물이며, 연녹색 윗옷과 검정 하의 조합이 이전 장면 전령의 복장을 재사용한 것으로 읽힌다. 주변 의병들은 낡은 조선식 옷차림이며 읽을 수 있는 글자는 없다.",
        "hard_violations": [
         "이전 장면 인물의 의복을 누구에게도 옮기지 말라는 명시적 금지와 달리, 오른쪽 전경의 별도 인물에 전령의 연녹색 윗옷·검정 하의 복장을 재사용했다."
        ],
        "physics": "두 주인공과 뒤쪽 의병들은 발로 지면을 딛고 있다. 복만의 총은 개머리판이 땅에 닿고 손이 몸통을 잡아 지지한다. 오주일의 총은 어깨를 지나는 멜빵으로 지지되며, 주변 의병들의 총도 손 또는 지면의 받침이 보인다. 나뭇가지는 흙 위에 놓여 있다. 전경 인물의 발은 프레임 밖이지만 보이는 자세에서 부유나 불가능한 동작은 발견되지 않는다."
       }
      ]
     },
     "ok": true
    }
   ],
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "slot_winner_match": true,
   "slot_winner": {
    "gemini-pro": "B",
    "gpt-high": "B"
   },
   "route": "cross_slot_agree"
  },
  "dual": {
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "normalized": {
    "A": 0.975,
    "B": 2.0
   },
   "adjusted": {
    "A": 0.725,
    "B": 1.75
   },
   "violations": {
    "A": [
     "[gemini-pro] 금지된 복장 복제 (이전 샷 전령의 초록색 옷을 전경 인물에게 적용함)",
     "[gemini-pro] 물리적으로 불가능한 해부학 (전경 인물의 기형적으로 작은 오른손)",
     "[gpt-high] 이전 장면 인물의 의복을 누구에게도 옮기지 말라는 명시적 금지와 달리, 오른쪽 전경의 별도 인물에 전령의 연녹색 윗옷·검정 하의 복장을 재사용했다."
    ],
    "B": [
     "[gemini-pro] 발명된 사물 (프롬프트에 명시된 훈련용 나뭇가지 대신 화승총이 바닥에 뜬금없이 놓여 있음)"
    ]
   },
   "per_model_winner": {
    "gemini-pro": "B",
    "gpt-high": "B"
   },
   "agreed": true
  },
  "totals": {
   "A": 725,
   "B": 1750
  },
  "selected": "B",
  "ranking": [
   "B",
   "A"
  ],
  "verdicts": [
   {
    "label": "A",
    "score": 725,
    "verdict_ko": "이전 샷의 전령 복장을 사용하지 말라는 네거티브 지시를 위반하여 전경 인물에게 입혔으며, 해당 인물의 손 등 해부학적 오류가 매우 뚜렷합니다.  ★위반: [gemini-pro] 금지된 복장 복제 (이전 샷 전령의 초록색 옷을 전경 인물에게 적용함) / [gemini-pro] 물리적으로 불가능한 해부학 (전경 인물의 기형적으로 작은 오른손) / [gpt-high] 이전 장면 인물의 의복을 누구에게도 옮기지 말라는 명시적 금지와 달리, 오른쪽 전경의 별도 인물에 전령의 연녹색 윗옷·검정 하의 복장을 재사용했다."
   },
   {
    "label": "B",
    "score": 1750,
    "verdict_ko": "오주일과 복만의 외형을 참조와 완벽하게 일치시키고 긴장감 있는 시선 처리를 잘 연출했으나, 바닥에 명시된 나뭇가지 대신 총을 배치한 점이 아쉽습니다.  ★위반: [gemini-pro] 발명된 사물 (프롬프트에 명시된 훈련용 나뭇가지 대신 화승총이 바닥에 뜬금없이 놓여 있음)"
   }
  ],
  "refs": [
   {
    "label": "PREVIOUS SHOT STILL — a visually related earlier shot of this same place: the location's look, materials, fixed features and lighting mood are LOCKED to this photo; never copy its camera framing. The people visible in that photo are NOT in this shot: never carry their faces, bodies or clothing onto anyone here. Each person in THIS shot is defined solely by the PEOPLE list, their own CHARACTER REFERENCE images and their wardrobe notes. If this photo shows a character who cannot move (dead or unconscious), that character's exact pose, position and orientation are ALSO LOCKED — treat their body as a fixed prop of the set that only the camera moves around.",
    "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/db00517b-165b-4005-a287-f28936b59b95/images/272bd7af-9294-4490-98c6-88a40e9deacc/scene/recipe/S29sh2_sel.png",
    "asset_id": "035ce225-e75f-4055-9bb3-ca1a292beba1",
    "role": "prev_still"
   },
   {
    "label": "CHARACTER REFERENCE — 오주일: the exact person appearing in this shot; match face, hair and build exactly.",
    "path": "<bytes:809716>",
    "asset_id": "7c31e825-ceef-4208-915f-71c4f0e15dd6",
    "role": "character_ref"
   },
   {
    "label": "CHARACTER REFERENCE — 복만: the exact person appearing in this shot; match face, hair and build exactly.",
    "path": "<bytes:1027479>",
    "asset_id": "b7e69570-d15d-4241-b5ea-4cb8d4ece1cf",
    "role": "character_ref"
   }
  ],
  "critique_skipped": true,
  "shot_run_produced": false,
  "shot_run_uid": "06aa1423-133f-7d3f-b337-ddf197e968ec",
  "ref_mode": "prev+엔티티",
  "share_plan": {
   "ref_plan": "prev",
   "prev_anchor_tag": "S29sh2"
  }
 },
 "S29sh12::cine": {
  "applied": true,
  "attempted_at": "2026-09-09T04:12:49.596371+00:00",
  "fingerprint": "242e84eabebf9f223916fd1e3d0a89466862b81f8fd224a0006d03ec49fef7d6",
  "fingerprint_version": 2,
  "provider": "mai",
  "endpoint": "openrouter/chat-completions",
  "model": "microsoft/mai-image-2.6",
  "pack": "24.202608252115",
  "source_file": "S29sh12_sel.png",
  "source_sha256": "8e5a0e151ce8c07417b14f42c73b16df025a3314d3a7071ffd5622eaae315f83",
  "file": "S29sh12_cine.png",
  "staged_sha256": "711cd18ab5013dc3a360ecc61e794b4b0e19ff6d18945ec1ecf086ee89907f4b",
  "latency_ms": 32025
 },
 "S30sh21::signage": {
  "fp": "feed246aeed2fc20",
  "inscriptions": [],
  "cues": [],
  "dropped": []
 },
 "S30sh21": {
  "input_fingerprint": "40b8718137b36a03",
  "prompt": "Create ONE FINAL photorealistic live-action film still of the moment below — early 20th-century Korea, Jeolla region; all people are Korean unless explicitly stated otherwise. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 텅 빈 장터 한복판으로 성큼 걸어 나온 안규홍의 꼿꼿하고 당당한 전신.\n\nLOCATION (lock): The cleared center of the market's outdoor dirt thoroughfare, beside the earthenware stalls. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: wide shot\n- FRAME LAYOUT: Cleared passage ahead toward the off-screen military policemen in the middle-right of the frame, midground.\n- KEY BACKGROUND ELEMENTS: 장터 중앙 통로 (Cleared by marketgoers before 안규홍 steps into it); used as Provides uninterrupted full-body visibility and open space toward the confrontation; 옹기 좌판 (Still intact before the first blow) — Seen obliquely along the far edge of the passage; used as A peripheral location reference that anticipates the fight without competing with 안규홍.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Use natural daytime ambient illumination with restrained saturation and grounded contrast, avoiding a heroic spotlight or sudden tonal shift.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The clothing, hair and overall look of 안규홍 — who appear both in that photo and in this shot — are LOCKED to that photo. Anyone else visible in that photo is NOT in this shot: never carry their face, body or clothing onto anyone here. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nPREVIOUS STILL USAGE (follow exactly — what to take from the attached still and what to exclude; it governs place and objects only, never who is in this shot or how the camera sees them): Take the sunlit dirt ground and the pottery-market surroundings visible in the reference. Exclude the earlier walking action and any props not present in the current confrontation.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The carrying frame has been set down in the market center; soup-stall smoke continues rising. 안규홍: He steps forward in a porter's clothes without the frame on his back, retaining the concealed jade ring.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 안규홍 (한국인 남성, 청년 나이대, 검은 머리, 단단히 묶은 머리) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.",
  "roll_prompts": {
   "A": "Create ONE FINAL photorealistic live-action film still of the moment below — early 20th-century Korea, Jeolla region; all people are Korean unless explicitly stated otherwise. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 텅 빈 장터 한복판으로 성큼 걸어 나온 안규홍의 꼿꼿하고 당당한 전신.\n\nLOCATION (lock): The cleared center of the market's outdoor dirt thoroughfare, beside the earthenware stalls. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: wide shot\n- FRAME LAYOUT: Cleared passage ahead toward the off-screen military policemen in the middle-right of the frame, midground.\n- KEY BACKGROUND ELEMENTS: 장터 중앙 통로 (Cleared by marketgoers before 안규홍 steps into it); used as Provides uninterrupted full-body visibility and open space toward the confrontation; 옹기 좌판 (Still intact before the first blow) — Seen obliquely along the far edge of the passage; used as A peripheral location reference that anticipates the fight without competing with 안규홍.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Use natural daytime ambient illumination with restrained saturation and grounded contrast, avoiding a heroic spotlight or sudden tonal shift.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The clothing, hair and overall look of 안규홍 — who appear both in that photo and in this shot — are LOCKED to that photo. Anyone else visible in that photo is NOT in this shot: never carry their face, body or clothing onto anyone here. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nPREVIOUS STILL USAGE (follow exactly — what to take from the attached still and what to exclude; it governs place and objects only, never who is in this shot or how the camera sees them): Take the sunlit dirt ground and the pottery-market surroundings visible in the reference. Exclude the earlier walking action and any props not present in the current confrontation.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The carrying frame has been set down in the market center; soup-stall smoke continues rising. 안규홍: He steps forward in a porter's clothes without the frame on his back, retaining the concealed jade ring.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 안규홍 (한국인 남성, 청년 나이대, 검은 머리, 단단히 묶은 머리) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.",
   "B": "Create ONE FINAL photorealistic live-action film still of the moment below — early 20th-century Korea, Jeolla region; all people are Korean unless explicitly stated otherwise. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 텅 빈 장터 한복판으로 성큼 걸어 나온 안규홍의 꼿꼿하고 당당한 전신.\n\nLOCATION (lock): The cleared center of the market's outdoor dirt thoroughfare, beside the earthenware stalls. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: wide shot\n- FRAME LAYOUT: Cleared passage ahead toward the off-screen military policemen in the middle-right of the frame, midground.\n- KEY BACKGROUND ELEMENTS: 장터 중앙 통로 (Cleared by marketgoers before 안규홍 steps into it); used as Provides uninterrupted full-body visibility and open space toward the confrontation; 옹기 좌판 (Still intact before the first blow) — Seen obliquely along the far edge of the passage; used as A peripheral location reference that anticipates the fight without competing with 안규홍.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Use natural daytime ambient illumination with restrained saturation and grounded contrast, avoiding a heroic spotlight or sudden tonal shift.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The clothing, hair and overall look of 안규홍 — who appear both in that photo and in this shot — are LOCKED to that photo. Anyone else visible in that photo is NOT in this shot: never carry their face, body or clothing onto anyone here. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nPREVIOUS STILL USAGE (follow exactly — what to take from the attached still and what to exclude; it governs place and objects only, never who is in this shot or how the camera sees them): Take the sunlit dirt ground and the pottery-market surroundings visible in the reference. Exclude the earlier walking action and any props not present in the current confrontation.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The carrying frame has been set down in the market center; soup-stall smoke continues rising. 안규홍: He steps forward in a porter's clothes without the frame on his back, retaining the concealed jade ring.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 안규홍 (한국인 남성, 청년 나이대, 검은 머리, 단단히 묶은 머리) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.\n\nCOMPOSITION VARIATION (this candidate only): keep the place, the people and the moment exactly as specified, and keep from the CAMERA & FRAME contract its subject scale, screen placement and key background elements exactly. For this candidate only, its camera angle and height are deliberately left open: choose a camera position distinctly different from the obvious one, as a film director picking a second setup on the same blocking. Stay true to any explicit framing the SHOT TEXT itself dictates."
  },
  "shot_run_spend_attempt_count": 1,
  "initial_roll_all_fail": true,
  "readings": [
   {
    "label": "A",
    "direction": "안규홍은 정면을 향해 성큼 걸어오고 있으며, 시선은 앞으로 고정되어 있음. 화면에 등장한 헌병들은 안규홍을 주시함.",
    "built_space": "흙길 통로와 양옆의 옹기 좌판이 구현됨. 기와문과 산이 안규홍의 뒤편에 위치하여, 레퍼런스에서 설정된 안규홍의 이동 방향을 정확히 유지한 채 카메라만 앞쪽으로 이동한 올바른 공간 구성을 보여줌.",
    "entities": "안규홍의 의상과 당당한 전신 외모는 레퍼런스를 잘 반영함. 그러나 '텅 빈 장터'라는 샷 텍스트와 다른 인물 추가를 엄격히 금지한 지시(never anyone else)에도 불구하고 다수의 헌병과 장터 사람들이 프레임에 포함됨.",
    "hard_violations": [
     "[gemini-pro] 지시사항에 명시적으로 금지된 다수의 추가 인물(헌병 및 상인들) 등장",
     "[gpt-high] 허용되지 않은 헌병 신체를 왼쪽 전경에 크게 넣고, 중경 오른쪽에도 헌병 세 명을 등장시켰다.",
     "[gpt-high] 안규홍만 보여야 하는 장면에 여러 민간인을 추가했다."
    ],
    "physics": "안규홍이 걷는 자세와 무게 중심이 자연스러우며 두 발이 지면에 올바르게 밀착됨."
   },
   {
    "label": "B",
    "direction": "안규홍이 카메라를 등지고 서 있으며, 멀리 떨어진 헌병들을 향해 있음.",
    "built_space": "옹기 좌판과 흙길 통로가 보이나, 레퍼런스에서 안규홍의 등 뒤에 있던 기와문과 산이 정면에 위치함. 이는 안규홍이 왔던 길로 되돌아간 것이 아니라면 성립할 수 없는 180도 뒤집힌 심각한 공간 설정 오류임.",
    "entities": "안규홍의 전면은 보이지 않고 뒷모습만 묘사됨. 샷 텍스트에 없는 헌병과 장터 사람들이 배경에 다수 추가됨.",
    "hard_violations": [
     "[gemini-pro] 지시사항에 명시적으로 금지된 다수의 추가 인물 등장",
     "[gemini-pro] 레퍼런스의 이동 방향과 모순되는 공간 구조 뒤집힘 오류",
     "[gpt-high] 화면 밖에 있어야 할 헌병 두 명을 중경에 직접 등장시켰다.",
     "[gpt-high] 안규홍 외에는 허용되지 않았는데 좌판 주변과 후경에 여러 민간인을 추가했다."
    ],
    "physics": "두 발로 지면에 안정적으로 서 있음."
   }
  ],
  "cross_model_order": {
   "policy": "cross_model_order_v2_slot0_fwd_slot1_rev_parallel2",
   "slots": [
    {
     "model": "gemini-pro",
     "order": "forward",
     "display_to_canonical": {
      "A": "A",
      "B": "B"
     },
     "raw": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 3,
        "verdict_ko": "공간의 방향성과 안규홍의 당당한 전신 묘사는 훌륭하나, 텅 빈 장터여야 함에도 다수의 엑스트라가 등장한 치명적인 위반이 있습니다."
       },
       {
        "label": "B",
        "score": 1,
        "verdict_ko": "추가 인물 등장이라는 위반에 더해, 안규홍의 이동 방향이 레퍼런스와 정반대로 뒤집히는 심각한 공간 설정 오류가 발생했습니다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "안규홍은 정면을 향해 성큼 걸어오고 있으며, 시선은 앞으로 고정되어 있음. 화면에 등장한 헌병들은 안규홍을 주시함.",
        "built_space": "흙길 통로와 양옆의 옹기 좌판이 구현됨. 기와문과 산이 안규홍의 뒤편에 위치하여, 레퍼런스에서 설정된 안규홍의 이동 방향을 정확히 유지한 채 카메라만 앞쪽으로 이동한 올바른 공간 구성을 보여줌.",
        "entities": "안규홍의 의상과 당당한 전신 외모는 레퍼런스를 잘 반영함. 그러나 '텅 빈 장터'라는 샷 텍스트와 다른 인물 추가를 엄격히 금지한 지시(never anyone else)에도 불구하고 다수의 헌병과 장터 사람들이 프레임에 포함됨.",
        "hard_violations": [
         "지시사항에 명시적으로 금지된 다수의 추가 인물(헌병 및 상인들) 등장"
        ],
        "physics": "안규홍이 걷는 자세와 무게 중심이 자연스러우며 두 발이 지면에 올바르게 밀착됨."
       },
       {
        "label": "B",
        "direction": "안규홍이 카메라를 등지고 서 있으며, 멀리 떨어진 헌병들을 향해 있음.",
        "built_space": "옹기 좌판과 흙길 통로가 보이나, 레퍼런스에서 안규홍의 등 뒤에 있던 기와문과 산이 정면에 위치함. 이는 안규홍이 왔던 길로 되돌아간 것이 아니라면 성립할 수 없는 180도 뒤집힌 심각한 공간 설정 오류임.",
        "entities": "안규홍의 전면은 보이지 않고 뒷모습만 묘사됨. 샷 텍스트에 없는 헌병과 장터 사람들이 배경에 다수 추가됨.",
        "hard_violations": [
         "지시사항에 명시적으로 금지된 다수의 추가 인물 등장",
         "레퍼런스의 이동 방향과 모순되는 공간 구조 뒤집힘 오류"
        ],
        "physics": "두 발로 지면에 안정적으로 서 있음."
       }
      ],
      "all_candidates_fail": true
     },
     "normalized": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 3,
        "verdict_ko": "공간의 방향성과 안규홍의 당당한 전신 묘사는 훌륭하나, 텅 빈 장터여야 함에도 다수의 엑스트라가 등장한 치명적인 위반이 있습니다."
       },
       {
        "label": "B",
        "score": 1,
        "verdict_ko": "추가 인물 등장이라는 위반에 더해, 안규홍의 이동 방향이 레퍼런스와 정반대로 뒤집히는 심각한 공간 설정 오류가 발생했습니다."
       }
      ],
      "all_candidates_fail": true,
      "readings": [
       {
        "label": "A",
        "direction": "안규홍은 정면을 향해 성큼 걸어오고 있으며, 시선은 앞으로 고정되어 있음. 화면에 등장한 헌병들은 안규홍을 주시함.",
        "built_space": "흙길 통로와 양옆의 옹기 좌판이 구현됨. 기와문과 산이 안규홍의 뒤편에 위치하여, 레퍼런스에서 설정된 안규홍의 이동 방향을 정확히 유지한 채 카메라만 앞쪽으로 이동한 올바른 공간 구성을 보여줌.",
        "entities": "안규홍의 의상과 당당한 전신 외모는 레퍼런스를 잘 반영함. 그러나 '텅 빈 장터'라는 샷 텍스트와 다른 인물 추가를 엄격히 금지한 지시(never anyone else)에도 불구하고 다수의 헌병과 장터 사람들이 프레임에 포함됨.",
        "hard_violations": [
         "지시사항에 명시적으로 금지된 다수의 추가 인물(헌병 및 상인들) 등장"
        ],
        "physics": "안규홍이 걷는 자세와 무게 중심이 자연스러우며 두 발이 지면에 올바르게 밀착됨."
       },
       {
        "label": "B",
        "direction": "안규홍이 카메라를 등지고 서 있으며, 멀리 떨어진 헌병들을 향해 있음.",
        "built_space": "옹기 좌판과 흙길 통로가 보이나, 레퍼런스에서 안규홍의 등 뒤에 있던 기와문과 산이 정면에 위치함. 이는 안규홍이 왔던 길로 되돌아간 것이 아니라면 성립할 수 없는 180도 뒤집힌 심각한 공간 설정 오류임.",
        "entities": "안규홍의 전면은 보이지 않고 뒷모습만 묘사됨. 샷 텍스트에 없는 헌병과 장터 사람들이 배경에 다수 추가됨.",
        "hard_violations": [
         "지시사항에 명시적으로 금지된 다수의 추가 인물 등장",
         "레퍼런스의 이동 방향과 모순되는 공간 구조 뒤집힘 오류"
        ],
        "physics": "두 발로 지면에 안정적으로 서 있음."
       }
      ]
     },
     "ok": true
    },
    {
     "model": "gpt-high",
     "order": "reverse",
     "display_to_canonical": {
      "A": "B",
      "B": "A"
     },
     "raw": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 3,
        "verdict_ko": "추가 인물로 실격이지만, 꼿꼿한 전신과 중경 오른쪽을 향해 열린 장터 통로는 B보다 지시 구도에 가깝다."
       },
       {
        "label": "B",
        "score": 2,
        "verdict_ko": "얼굴과 전진 동작은 잘 맞지만, 전경의 헌병 신체와 여러 추가 인물, 크게 부각된 옹기 때문에 단독 전신과 빈 대치 공간이라는 구도를 더 크게 위반한다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "안규홍은 카메라에 등을 보이고 중경 오른쪽의 헌병 두 명을 향한다. 얼굴이 가려져 눈동자 방향은 확인할 수 없지만 머리와 몸의 방향은 대치 상대와 일치한다. 헌병들도 안규홍 쪽을 바라본다. 보이는 장총은 총구가 아래로 내려가 있으며 사람을 겨누지 않는다. 다만 대치 상대는 지시와 달리 화면 안에 있다.",
        "built_space": "흙길 한 줄이 중앙에서 뒤쪽으로 이어지고, 양옆에 좌판 열이 하나씩 있다. 왼쪽에는 여러 옹기가 놓인 목제 진열대와 차양 지주, 오른쪽에는 낮은 진열대와 철망 우리가 보인다. 후경에는 기와 대문 한 곳과 그 앞 계단이 있다. 흙벽·목재·차양·기와의 재료와 낮 조명은 참조와 유사하다. 안규홍의 전신 앞에서 헌병들까지 통로가 열려 있으며, 옹기는 주로 가장자리에 배치된다. 반사면에 따른 광학적 모순은 없다.",
        "entities": "중앙 인물은 검은 머리를 단단히 묶은 청년 남성으로, 체격과 때 묻은 미색 저고리·바지·허리끈·짚신이 안규홍 참조와 대체로 맞는다. 뒷모습이라 얼굴 동일성은 확인할 수 없다. 헌병 두 명과 좌판 주변의 민간인 여러 명이 추가되어 안규홍만 보여야 한다는 조건을 어긴다. 옹기는 파손되지 않았고 등에 지게는 없다. 내려놓은 지게와 국밥 좌판의 연기는 뚜렷하게 확인되지 않으며, 숨긴 옥가락지는 보이지 않아 판정할 수 없다. 읽을 수 있는 글자는 없다.",
        "hard_violations": [
         "화면 밖에 있어야 할 헌병 두 명을 중경에 직접 등장시켰다.",
         "안규홍 외에는 허용되지 않았는데 좌판 주변과 후경에 여러 민간인을 추가했다."
        ],
        "physics": "안규홍은 두 발을 흙바닥에 딛고 서 있으며 몸의 무게가 자연스럽게 지지된다. 다만 보폭과 체중 이동이 작아 성큼 걸어 나온 순간보다는 멈춰 선 자세에 가깝다. 다른 인물들도 지면에 서 있고, 장총은 손과 멜빵으로 지지된다. 옹기는 진열대 또는 땅에 놓이고 차양은 목제 지주가 받친다. 지지 없이 떠 있는 몸이나 물건은 없다."
       },
       {
        "label": "B",
        "direction": "안규홍은 카메라 쪽으로 걸어오며 정면 가까이를 응시한다. 왼쪽 전경의 헌병과 대치하는 구도로 읽히지만, 눈길이 그 인물에게 정확히 닿는지는 불분명하다. 요구된 중경 오른쪽의 화면 밖 상대를 향하는 방향은 아니다. 오른쪽 헌병들은 안규홍 또는 카메라 쪽을 보고, 보이는 장총은 아래로 향한다.",
        "built_space": "중앙 흙길 양쪽으로 좌판 열이 이어지고, 왼쪽에는 흙벽과 짚 차양, 오른쪽에는 기와 처마와 천 차양이 보인다. 후경에는 기와 대문 한 곳이 있다. 철망 우리 한 개와 여러 목제 진열대가 오른쪽에 놓인다. 참조의 장터 재료와 산 배경은 이어지지만, 오른쪽 옹기 무리가 전경을 크게 차지해 먼 가장자리의 주변 요소라는 요구와 다르다. 왼쪽 전경은 헌병의 몸이 가리고 중경 오른쪽도 헌병들이 점유하여, 지정된 빈 대치 공간이 성립하지 않는다.",
        "entities": "안규홍은 참조와 유사한 얼굴·청년 남성 체격·뒤로 묶은 검은 머리이며, 미색의 낡은 저고리와 바지, 허리끈, 짚신도 잘 맞는다. 그러나 왼쪽 전경의 제복 인물 일부와 오른쪽의 헌병 세 명, 후경의 여러 민간인이 함께 등장한다. 옹기는 온전하며 등에 지게가 없다. 왼쪽 바닥에는 내려놓은 운반틀로 보이는 물건이 있고 좌판에서 연기가 올라와 상태 연속성 일부를 반영한다. 숨긴 옥가락지는 확인되지 않으며 읽을 수 있는 글자는 없다.",
        "hard_violations": [
         "허용되지 않은 헌병 신체를 왼쪽 전경에 크게 넣고, 중경 오른쪽에도 헌병 세 명을 등장시켰다.",
         "안규홍만 보여야 하는 장면에 여러 민간인을 추가했다."
        ],
        "physics": "안규홍은 한 발을 앞에 딛고 반대 발을 뒤에 둔 채 전진하며, 지면 접촉과 체중 이동이 가능한 보행 자세다. 팔도 몸 옆에서 자연스럽게 움직인다. 다른 인물들의 몸은 지면에 지지되고, 옹기와 우리는 진열대에 놓여 있다. 운반틀은 바닥에 내려져 있고 차양은 지주와 건물에 연결된다. 연기는 좌판 부근에서 위로 퍼지며, 지지 없이 떠 있는 물체나 신체는 없다."
       }
      ],
      "all_candidates_fail": true
     },
     "normalized": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 3,
        "verdict_ko": "추가 인물로 실격이지만, 꼿꼿한 전신과 중경 오른쪽을 향해 열린 장터 통로는 B보다 지시 구도에 가깝다."
       },
       {
        "label": "A",
        "score": 2,
        "verdict_ko": "얼굴과 전진 동작은 잘 맞지만, 전경의 헌병 신체와 여러 추가 인물, 크게 부각된 옹기 때문에 단독 전신과 빈 대치 공간이라는 구도를 더 크게 위반한다."
       }
      ],
      "all_candidates_fail": true,
      "readings": [
       {
        "label": "B",
        "direction": "안규홍은 카메라에 등을 보이고 중경 오른쪽의 헌병 두 명을 향한다. 얼굴이 가려져 눈동자 방향은 확인할 수 없지만 머리와 몸의 방향은 대치 상대와 일치한다. 헌병들도 안규홍 쪽을 바라본다. 보이는 장총은 총구가 아래로 내려가 있으며 사람을 겨누지 않는다. 다만 대치 상대는 지시와 달리 화면 안에 있다.",
        "built_space": "흙길 한 줄이 중앙에서 뒤쪽으로 이어지고, 양옆에 좌판 열이 하나씩 있다. 왼쪽에는 여러 옹기가 놓인 목제 진열대와 차양 지주, 오른쪽에는 낮은 진열대와 철망 우리가 보인다. 후경에는 기와 대문 한 곳과 그 앞 계단이 있다. 흙벽·목재·차양·기와의 재료와 낮 조명은 참조와 유사하다. 안규홍의 전신 앞에서 헌병들까지 통로가 열려 있으며, 옹기는 주로 가장자리에 배치된다. 반사면에 따른 광학적 모순은 없다.",
        "entities": "중앙 인물은 검은 머리를 단단히 묶은 청년 남성으로, 체격과 때 묻은 미색 저고리·바지·허리끈·짚신이 안규홍 참조와 대체로 맞는다. 뒷모습이라 얼굴 동일성은 확인할 수 없다. 헌병 두 명과 좌판 주변의 민간인 여러 명이 추가되어 안규홍만 보여야 한다는 조건을 어긴다. 옹기는 파손되지 않았고 등에 지게는 없다. 내려놓은 지게와 국밥 좌판의 연기는 뚜렷하게 확인되지 않으며, 숨긴 옥가락지는 보이지 않아 판정할 수 없다. 읽을 수 있는 글자는 없다.",
        "hard_violations": [
         "화면 밖에 있어야 할 헌병 두 명을 중경에 직접 등장시켰다.",
         "안규홍 외에는 허용되지 않았는데 좌판 주변과 후경에 여러 민간인을 추가했다."
        ],
        "physics": "안규홍은 두 발을 흙바닥에 딛고 서 있으며 몸의 무게가 자연스럽게 지지된다. 다만 보폭과 체중 이동이 작아 성큼 걸어 나온 순간보다는 멈춰 선 자세에 가깝다. 다른 인물들도 지면에 서 있고, 장총은 손과 멜빵으로 지지된다. 옹기는 진열대 또는 땅에 놓이고 차양은 목제 지주가 받친다. 지지 없이 떠 있는 몸이나 물건은 없다."
       },
       {
        "label": "A",
        "direction": "안규홍은 카메라 쪽으로 걸어오며 정면 가까이를 응시한다. 왼쪽 전경의 헌병과 대치하는 구도로 읽히지만, 눈길이 그 인물에게 정확히 닿는지는 불분명하다. 요구된 중경 오른쪽의 화면 밖 상대를 향하는 방향은 아니다. 오른쪽 헌병들은 안규홍 또는 카메라 쪽을 보고, 보이는 장총은 아래로 향한다.",
        "built_space": "중앙 흙길 양쪽으로 좌판 열이 이어지고, 왼쪽에는 흙벽과 짚 차양, 오른쪽에는 기와 처마와 천 차양이 보인다. 후경에는 기와 대문 한 곳이 있다. 철망 우리 한 개와 여러 목제 진열대가 오른쪽에 놓인다. 참조의 장터 재료와 산 배경은 이어지지만, 오른쪽 옹기 무리가 전경을 크게 차지해 먼 가장자리의 주변 요소라는 요구와 다르다. 왼쪽 전경은 헌병의 몸이 가리고 중경 오른쪽도 헌병들이 점유하여, 지정된 빈 대치 공간이 성립하지 않는다.",
        "entities": "안규홍은 참조와 유사한 얼굴·청년 남성 체격·뒤로 묶은 검은 머리이며, 미색의 낡은 저고리와 바지, 허리끈, 짚신도 잘 맞는다. 그러나 왼쪽 전경의 제복 인물 일부와 오른쪽의 헌병 세 명, 후경의 여러 민간인이 함께 등장한다. 옹기는 온전하며 등에 지게가 없다. 왼쪽 바닥에는 내려놓은 운반틀로 보이는 물건이 있고 좌판에서 연기가 올라와 상태 연속성 일부를 반영한다. 숨긴 옥가락지는 확인되지 않으며 읽을 수 있는 글자는 없다.",
        "hard_violations": [
         "허용되지 않은 헌병 신체를 왼쪽 전경에 크게 넣고, 중경 오른쪽에도 헌병 세 명을 등장시켰다.",
         "안규홍만 보여야 하는 장면에 여러 민간인을 추가했다."
        ],
        "physics": "안규홍은 한 발을 앞에 딛고 반대 발을 뒤에 둔 채 전진하며, 지면 접촉과 체중 이동이 가능한 보행 자세다. 팔도 몸 옆에서 자연스럽게 움직인다. 다른 인물들의 몸은 지면에 지지되고, 옹기와 우리는 진열대에 놓여 있다. 운반틀은 바닥에 내려져 있고 차양은 지주와 건물에 연결된다. 연기는 좌판 부근에서 위로 퍼지며, 지지 없이 떠 있는 물체나 신체는 없다."
       }
      ]
     },
     "ok": true
    }
   ],
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "slot_winner_match": false,
   "slot_winner": {
    "gemini-pro": "A",
    "gpt-high": "B"
   },
   "route": "cross_slot_combined"
  },
  "dual": {
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "normalized": {
    "A": 1.667,
    "B": 1.333
   },
   "adjusted": {
    "A": 1.417,
    "B": 1.083
   },
   "violations": {
    "A": [
     "[gemini-pro] 지시사항에 명시적으로 금지된 다수의 추가 인물(헌병 및 상인들) 등장",
     "[gpt-high] 허용되지 않은 헌병 신체를 왼쪽 전경에 크게 넣고, 중경 오른쪽에도 헌병 세 명을 등장시켰다.",
     "[gpt-high] 안규홍만 보여야 하는 장면에 여러 민간인을 추가했다."
    ],
    "B": [
     "[gemini-pro] 지시사항에 명시적으로 금지된 다수의 추가 인물 등장",
     "[gemini-pro] 레퍼런스의 이동 방향과 모순되는 공간 구조 뒤집힘 오류",
     "[gpt-high] 화면 밖에 있어야 할 헌병 두 명을 중경에 직접 등장시켰다.",
     "[gpt-high] 안규홍 외에는 허용되지 않았는데 좌판 주변과 후경에 여러 민간인을 추가했다."
    ]
   },
   "per_model_winner": {
    "gemini-pro": "A",
    "gpt-high": "B"
   },
   "agreed": false
  },
  "totals": {
   "A": 1417,
   "B": 1083
  },
  "selected": "A",
  "ranking": [
   "A",
   "B"
  ],
  "verdicts": [
   {
    "label": "A",
    "score": 1417,
    "verdict_ko": "공간의 방향성과 안규홍의 당당한 전신 묘사는 훌륭하나, 텅 빈 장터여야 함에도 다수의 엑스트라가 등장한 치명적인 위반이 있습니다.  ★위반: [gemini-pro] 지시사항에 명시적으로 금지된 다수의 추가 인물(헌병 및 상인들) 등장 / [gpt-high] 허용되지 않은 헌병 신체를 왼쪽 전경에 크게 넣고, 중경 오른쪽에도 헌병 세 명을 등장시켰다. / [gpt-high] 안규홍만 보여야 하는 장면에 여러 민간인을 추가했다."
   },
   {
    "label": "B",
    "score": 1083,
    "verdict_ko": "추가 인물 등장이라는 위반에 더해, 안규홍의 이동 방향이 레퍼런스와 정반대로 뒤집히는 심각한 공간 설정 오류가 발생했습니다.  ★위반: [gemini-pro] 지시사항에 명시적으로 금지된 다수의 추가 인물 등장 / [gemini-pro] 레퍼런스의 이동 방향과 모순되는 공간 구조 뒤집힘 오류 / [gpt-high] 화면 밖에 있어야 할 헌병 두 명을 중경에 직접 등장시켰다. / [gpt-high] 안규홍 외에는 허용되지 않았는데 좌판 주변과 후경에 여러 민간인을 추가했다."
   }
  ],
  "refs": [
   {
    "label": "PREVIOUS SHOT STILL — a visually related earlier shot of this same place: the location's look, materials, fixed features and lighting mood are LOCKED to this photo; never copy its camera framing. The clothing, hair and overall look of 안규홍 — who appear both in that photo and in this shot — are LOCKED to that photo. Anyone else visible in that photo is NOT in this shot: never carry their face, body or clothing onto anyone here. If this photo shows a character who cannot move (dead or unconscious), that character's exact pose, position and orientation are ALSO LOCKED — treat their body as a fixed prop of the set that only the camera moves around.",
    "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/db00517b-165b-4005-a287-f28936b59b95/images/272bd7af-9294-4490-98c6-88a40e9deacc/scene/recipe/S1sh10_sel.png",
    "asset_id": "f1ae73c0-86b8-44fb-90bd-89aa539b1b61",
    "role": "prev_still"
   },
   {
    "label": "CHARACTER REFERENCE — 안규홍: the exact person appearing in this shot; match face, hair and build exactly.",
    "path": "<bytes:1028384>",
    "asset_id": "3b5894db-ce55-4f34-b966-5e725c3aa952",
    "role": "character_ref"
   }
  ],
  "critique_skipped": true,
  "needs_reshoot": true,
  "shot_run_produced": false,
  "shot_run_uid": "06aa1423-14ec-7a9a-895c-4d42437eba0e",
  "ref_mode": "prev+엔티티",
  "share_plan": {
   "ref_plan": "prev",
   "prev_anchor_tag": "S1sh10"
  }
 },
 "S30sh21::cine": {
  "applied": true,
  "attempted_at": "2026-09-09T04:14:24.966408+00:00",
  "fingerprint": "40b32116990aa11e2a983249e101b40a45a25bfcee37e1837fece46ac51affe3",
  "fingerprint_version": 2,
  "provider": "mai",
  "endpoint": "openrouter/chat-completions",
  "model": "microsoft/mai-image-2.6",
  "pack": "24.202608252115",
  "source_file": "S30sh21_sel.png",
  "source_sha256": "89d7e8cd3fe31849ec045b385cac78b9849c158e6b4a63e008a100383e09ec79",
  "file": "S30sh21_cine.png",
  "staged_sha256": "6dacf3534a6800977ee8a62d670a6fba1f64155dd564b6d819227868229ba78b",
  "latency_ms": 28254
 },
 "S30sh38::signage": {
  "fp": "09d7bd1c304721c0",
  "inscriptions": [],
  "cues": [],
  "dropped": []
 },
 "S30sh38": {
  "input_fingerprint": "45f245d73f4c5689",
  "prompt": "Create ONE FINAL photorealistic live-action film still of the moment below — early 20th-century Korea, Jeolla region; all people are Korean unless explicitly stated otherwise. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 처음 사람을 찌른 충격에 총을 쥔 채 뻣뻣하게 굳어버린 안규홍의 하얗게 질린 얼굴.\n\nLOCATION (lock): The outdoor fighting space beside the market's earthenware stalls, on the dirt thoroughfare occupied by the prisoner escort. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: close-up\n- KEY BACKGROUND ELEMENTS: 빼앗은 소총 (Still gripped after the bayonet thrust) — Only the section held by his hands crosses the lower frame; the bayonet remains outside the crop; used as Connects the frozen expression to the action without obscuring his face or exaggerating the weapon's scale; 장터 흙바닥 (Visible beyond the confrontation); used as A softly resolved lower background keeps the close view anchored to the physical location.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Preserve the existing daytime light and naturalistic contrast, letting his pallor register through performance and complexion rather than a new color cast.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The clothing, hair and overall look of 안규홍 — who appear both in that photo and in this shot — are LOCKED to that photo. Anyone else visible in that photo is NOT in this shot: never carry their face, body or clothing onto anyone here. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nPREVIOUS STILL USAGE (follow exactly — what to take from the attached still and what to exclude; it governs place and objects only, never who is in this shot or how the camera sees them): Take the market's fixed stall structures, dirt surface, and harsh daylight. Exclude the previously intact pottery display wherever the collision has now broken it.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The carrying frame remains on the dirt; smashed pottery now litters the struck stall. 안규홍: He stands frozen holding the captured bayoneted rifle; the jade ring remains concealed.\n\nTHE HAND THAT IS DOING THIS: the object at the centre of this shot is being held, operated, read, aimed or handed over by 안규홍 right now, so 안규홍's hand — and as much of the wrist and forearm as the framing reaches — is in the frame, gripping or resting on that object exactly the way its real use requires. Match that hand to 안규홍: its size, build, skin, age, grooming, sleeve and anything worn on it belong to that person and to no one else.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 안규홍 (한국인 남성, 청년 나이대, 검은 머리, 단단히 묶은 머리) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.",
  "roll_prompts": {
   "A": "Create ONE FINAL photorealistic live-action film still of the moment below — early 20th-century Korea, Jeolla region; all people are Korean unless explicitly stated otherwise. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 처음 사람을 찌른 충격에 총을 쥔 채 뻣뻣하게 굳어버린 안규홍의 하얗게 질린 얼굴.\n\nLOCATION (lock): The outdoor fighting space beside the market's earthenware stalls, on the dirt thoroughfare occupied by the prisoner escort. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: close-up\n- KEY BACKGROUND ELEMENTS: 빼앗은 소총 (Still gripped after the bayonet thrust) — Only the section held by his hands crosses the lower frame; the bayonet remains outside the crop; used as Connects the frozen expression to the action without obscuring his face or exaggerating the weapon's scale; 장터 흙바닥 (Visible beyond the confrontation); used as A softly resolved lower background keeps the close view anchored to the physical location.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Preserve the existing daytime light and naturalistic contrast, letting his pallor register through performance and complexion rather than a new color cast.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The clothing, hair and overall look of 안규홍 — who appear both in that photo and in this shot — are LOCKED to that photo. Anyone else visible in that photo is NOT in this shot: never carry their face, body or clothing onto anyone here. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nPREVIOUS STILL USAGE (follow exactly — what to take from the attached still and what to exclude; it governs place and objects only, never who is in this shot or how the camera sees them): Take the market's fixed stall structures, dirt surface, and harsh daylight. Exclude the previously intact pottery display wherever the collision has now broken it.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The carrying frame remains on the dirt; smashed pottery now litters the struck stall. 안규홍: He stands frozen holding the captured bayoneted rifle; the jade ring remains concealed.\n\nTHE HAND THAT IS DOING THIS: the object at the centre of this shot is being held, operated, read, aimed or handed over by 안규홍 right now, so 안규홍's hand — and as much of the wrist and forearm as the framing reaches — is in the frame, gripping or resting on that object exactly the way its real use requires. Match that hand to 안규홍: its size, build, skin, age, grooming, sleeve and anything worn on it belong to that person and to no one else.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 안규홍 (한국인 남성, 청년 나이대, 검은 머리, 단단히 묶은 머리) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.",
   "B": "Create ONE FINAL photorealistic live-action film still of the moment below — early 20th-century Korea, Jeolla region; all people are Korean unless explicitly stated otherwise. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 처음 사람을 찌른 충격에 총을 쥔 채 뻣뻣하게 굳어버린 안규홍의 하얗게 질린 얼굴.\n\nLOCATION (lock): The outdoor fighting space beside the market's earthenware stalls, on the dirt thoroughfare occupied by the prisoner escort. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: close-up\n- KEY BACKGROUND ELEMENTS: 빼앗은 소총 (Still gripped after the bayonet thrust) — Only the section held by his hands crosses the lower frame; the bayonet remains outside the crop; used as Connects the frozen expression to the action without obscuring his face or exaggerating the weapon's scale; 장터 흙바닥 (Visible beyond the confrontation); used as A softly resolved lower background keeps the close view anchored to the physical location.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Preserve the existing daytime light and naturalistic contrast, letting his pallor register through performance and complexion rather than a new color cast.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The clothing, hair and overall look of 안규홍 — who appear both in that photo and in this shot — are LOCKED to that photo. Anyone else visible in that photo is NOT in this shot: never carry their face, body or clothing onto anyone here. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nPREVIOUS STILL USAGE (follow exactly — what to take from the attached still and what to exclude; it governs place and objects only, never who is in this shot or how the camera sees them): Take the market's fixed stall structures, dirt surface, and harsh daylight. Exclude the previously intact pottery display wherever the collision has now broken it.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The carrying frame remains on the dirt; smashed pottery now litters the struck stall. 안규홍: He stands frozen holding the captured bayoneted rifle; the jade ring remains concealed.\n\nTHE HAND THAT IS DOING THIS: the object at the centre of this shot is being held, operated, read, aimed or handed over by 안규홍 right now, so 안규홍's hand — and as much of the wrist and forearm as the framing reaches — is in the frame, gripping or resting on that object exactly the way its real use requires. Match that hand to 안규홍: its size, build, skin, age, grooming, sleeve and anything worn on it belong to that person and to no one else.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 안규홍 (한국인 남성, 청년 나이대, 검은 머리, 단단히 묶은 머리) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.\n\nCOMPOSITION VARIATION (this candidate only): keep the place, the people and the moment exactly as specified, and keep from the CAMERA & FRAME contract its subject scale, screen placement and key background elements exactly. For this candidate only, its camera angle and height are deliberately left open: choose a camera position distinctly different from the obvious one, as a film director picking a second setup on the same blocking. Stay true to any explicit framing the SHOT TEXT itself dictates."
  },
  "shot_run_spend_attempt_count": 1,
  "readings": [
   {
    "label": "A",
    "direction": "안규홍은 넋이 나간 창백한 표정으로 화면 우측 밖의 대상을 응시하고 있으며, 쥐고 있는 소총의 총구 또한 우측을 향하고 있습니다.",
    "built_space": "초점이 흐려진 장터 흙바닥을 배경으로 목조 가판대와 온전한 항아리들이 일부 보입니다. 요구된 흙바닥은 보이나, 깨진 도자기 파편은 이 프레임 내에 나타나지 않았습니다.",
    "entities": "안규홍의 얼굴, 헤어스타일, 체형이 레퍼런스와 정확히 일치하며 한복의 깃(좌임)도 올바르게 묘사되었습니다. 빼앗은 소총을 두 손으로 단단히 쥐고 있으며, 총의 몸통 일부만 하단 프레임에 걸쳐 있고 대검은 프레임 밖에 위치해 지시를 잘 따랐습니다.",
    "hard_violations": [],
    "physics": "오른손으로 방아쇠 쪽을, 왼손으로 앞쪽 총열을 받치고 있어 소총의 무게를 지탱하는 파지법이 물리적으로 자연스럽고 안정적입니다."
   },
   {
    "label": "B",
    "direction": "안규홍은 충격을 받은 표정으로 화면 좌측 밖을 응시하고 있으며, 소총의 총구도 좌측을 향하고 있습니다.",
    "built_space": "장터 가판대와 흙바닥이 배경으로 나타나며, 특히 화면 좌측 바닥에 깨진 도자기 파편들이 널려 있어 샷 텍스트의 배경 묘사를 잘 반영했습니다.",
    "entities": "소총이 하단 프레임에 걸쳐 등장하나, 안규홍의 뺨에 레퍼런스에 없는 뚜렷한 점이 생겼고 한복 깃이 반대로(우임) 겹쳐진 채 묘사되어 복장 고증이 틀렸습니다.",
    "hard_violations": [
     "[gemini-pro] 물리적으로 불가능한 신체 구조: 방아쇠를 잡고 있는 뒤쪽 손(왼팔)의 손목이 카메라 쪽(가까운 면)에 있으면서 손가락이 앞으로 감싸는 형태를 취하고 있어, 왼팔에 오른손이 달려 있는 해부학적 오류가 발생했습니다."
    ],
    "physics": "카메라 쪽에 가까운 오른팔이 앞쪽 총열을 받치고 있으나, 방아쇠를 당기는 왼쪽 팔의 손목과 손가락 방향이 꼬여 있어 정상적으로 총을 지탱하고 조작할 수 없는 형태입니다."
   }
  ],
  "cross_model_order": {
   "policy": "cross_model_order_v2_slot0_fwd_slot1_rev_parallel2",
   "slots": [
    {
     "model": "gemini-pro",
     "order": "forward",
     "display_to_canonical": {
      "A": "A",
      "B": "B"
     },
     "raw": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 8,
        "verdict_ko": "레퍼런스와 일치하는 인물 외모와 올바른 한복 착장을 보여주며, 사람을 처음 찌른 충격에 굳은 표정과 소총 파지를 지시된 클로즈업 앵글 안에서 훌륭히 구현했습니다."
       },
       {
        "label": "B",
        "score": 2,
        "verdict_ko": "바닥에 깨진 도자기 파편을 지시대로 묘사했으나, 한복 깃이 반대로 뒤집히고 방아쇠를 잡은 뒤쪽 손의 해부학적 구조가 불가능하게 왜곡되어 치명적인 오류를 범했습니다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "안규홍은 넋이 나간 창백한 표정으로 화면 우측 밖의 대상을 응시하고 있으며, 쥐고 있는 소총의 총구 또한 우측을 향하고 있습니다.",
        "built_space": "초점이 흐려진 장터 흙바닥을 배경으로 목조 가판대와 온전한 항아리들이 일부 보입니다. 요구된 흙바닥은 보이나, 깨진 도자기 파편은 이 프레임 내에 나타나지 않았습니다.",
        "entities": "안규홍의 얼굴, 헤어스타일, 체형이 레퍼런스와 정확히 일치하며 한복의 깃(좌임)도 올바르게 묘사되었습니다. 빼앗은 소총을 두 손으로 단단히 쥐고 있으며, 총의 몸통 일부만 하단 프레임에 걸쳐 있고 대검은 프레임 밖에 위치해 지시를 잘 따랐습니다.",
        "hard_violations": [],
        "physics": "오른손으로 방아쇠 쪽을, 왼손으로 앞쪽 총열을 받치고 있어 소총의 무게를 지탱하는 파지법이 물리적으로 자연스럽고 안정적입니다."
       },
       {
        "label": "B",
        "direction": "안규홍은 충격을 받은 표정으로 화면 좌측 밖을 응시하고 있으며, 소총의 총구도 좌측을 향하고 있습니다.",
        "built_space": "장터 가판대와 흙바닥이 배경으로 나타나며, 특히 화면 좌측 바닥에 깨진 도자기 파편들이 널려 있어 샷 텍스트의 배경 묘사를 잘 반영했습니다.",
        "entities": "소총이 하단 프레임에 걸쳐 등장하나, 안규홍의 뺨에 레퍼런스에 없는 뚜렷한 점이 생겼고 한복 깃이 반대로(우임) 겹쳐진 채 묘사되어 복장 고증이 틀렸습니다.",
        "hard_violations": [
         "물리적으로 불가능한 신체 구조: 방아쇠를 잡고 있는 뒤쪽 손(왼팔)의 손목이 카메라 쪽(가까운 면)에 있으면서 손가락이 앞으로 감싸는 형태를 취하고 있어, 왼팔에 오른손이 달려 있는 해부학적 오류가 발생했습니다."
        ],
        "physics": "카메라 쪽에 가까운 오른팔이 앞쪽 총열을 받치고 있으나, 방아쇠를 당기는 왼쪽 팔의 손목과 손가락 방향이 꼬여 있어 정상적으로 총을 지탱하고 조작할 수 없는 형태입니다."
       }
      ],
      "all_candidates_fail": false
     },
     "normalized": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 8,
        "verdict_ko": "레퍼런스와 일치하는 인물 외모와 올바른 한복 착장을 보여주며, 사람을 처음 찌른 충격에 굳은 표정과 소총 파지를 지시된 클로즈업 앵글 안에서 훌륭히 구현했습니다."
       },
       {
        "label": "B",
        "score": 2,
        "verdict_ko": "바닥에 깨진 도자기 파편을 지시대로 묘사했으나, 한복 깃이 반대로 뒤집히고 방아쇠를 잡은 뒤쪽 손의 해부학적 구조가 불가능하게 왜곡되어 치명적인 오류를 범했습니다."
       }
      ],
      "all_candidates_fail": false,
      "readings": [
       {
        "label": "A",
        "direction": "안규홍은 넋이 나간 창백한 표정으로 화면 우측 밖의 대상을 응시하고 있으며, 쥐고 있는 소총의 총구 또한 우측을 향하고 있습니다.",
        "built_space": "초점이 흐려진 장터 흙바닥을 배경으로 목조 가판대와 온전한 항아리들이 일부 보입니다. 요구된 흙바닥은 보이나, 깨진 도자기 파편은 이 프레임 내에 나타나지 않았습니다.",
        "entities": "안규홍의 얼굴, 헤어스타일, 체형이 레퍼런스와 정확히 일치하며 한복의 깃(좌임)도 올바르게 묘사되었습니다. 빼앗은 소총을 두 손으로 단단히 쥐고 있으며, 총의 몸통 일부만 하단 프레임에 걸쳐 있고 대검은 프레임 밖에 위치해 지시를 잘 따랐습니다.",
        "hard_violations": [],
        "physics": "오른손으로 방아쇠 쪽을, 왼손으로 앞쪽 총열을 받치고 있어 소총의 무게를 지탱하는 파지법이 물리적으로 자연스럽고 안정적입니다."
       },
       {
        "label": "B",
        "direction": "안규홍은 충격을 받은 표정으로 화면 좌측 밖을 응시하고 있으며, 소총의 총구도 좌측을 향하고 있습니다.",
        "built_space": "장터 가판대와 흙바닥이 배경으로 나타나며, 특히 화면 좌측 바닥에 깨진 도자기 파편들이 널려 있어 샷 텍스트의 배경 묘사를 잘 반영했습니다.",
        "entities": "소총이 하단 프레임에 걸쳐 등장하나, 안규홍의 뺨에 레퍼런스에 없는 뚜렷한 점이 생겼고 한복 깃이 반대로(우임) 겹쳐진 채 묘사되어 복장 고증이 틀렸습니다.",
        "hard_violations": [
         "물리적으로 불가능한 신체 구조: 방아쇠를 잡고 있는 뒤쪽 손(왼팔)의 손목이 카메라 쪽(가까운 면)에 있으면서 손가락이 앞으로 감싸는 형태를 취하고 있어, 왼팔에 오른손이 달려 있는 해부학적 오류가 발생했습니다."
        ],
        "physics": "카메라 쪽에 가까운 오른팔이 앞쪽 총열을 받치고 있으나, 방아쇠를 당기는 왼쪽 팔의 손목과 손가락 방향이 꼬여 있어 정상적으로 총을 지탱하고 조작할 수 없는 형태입니다."
       }
      ]
     },
     "ok": true
    },
    {
     "model": "gpt-high",
     "order": "reverse",
     "display_to_canonical": {
      "A": "B",
      "B": "A"
     },
     "raw": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 7,
        "verdict_ko": "창백한 충격과 깨진 옹기는 잘 보이지만, 소총을 어깨에 붙여 조준하는 듯한 자세와 크게 드러난 무기가 얼굴 중심 클로즈업의 지시에서 벗어납니다."
       },
       {
        "label": "B",
        "score": 8,
        "verdict_ko": "굳어버린 얼굴을 중심으로 두고 손에 쥔 소총을 하단에 제한해 핵심 구도에 더 충실하나, 충돌로 깨진 옹기 잔해는 뚜렷하지 않습니다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "시선과 소총의 총열은 모두 화면 왼쪽 밖을 향하며, 그 방향의 상대는 보이지 않습니다. 총검은 프레임 밖입니다. 개머리판을 어깨 쪽에 붙이고 눈을 총열 방향에 둔 모습은 찌른 직후 멈춘 상태보다 다시 조준하는 상태에 가깝게 읽힙니다.",
        "built_space": "인물은 흙길에 있고, 왼쪽 뒤로 목재 기둥 여러 개와 차양, 옹기 진열대, 흙벽 및 기와지붕이 이어집니다. 하단 왼쪽에는 깨진 옹기 조각들이 있고 그 뒤 진열대에는 온전한 옹기가 남아 있습니다. 뒤쪽에 지게처럼 보이는 틀 하나가 있으나 흐려 확정하기 어렵습니다. 참조의 시장 재료와 낮빛은 유지되지만 고정 시설의 정확한 동일 배치는 확인하기 어렵습니다. 반사면이나 명백히 중복된 고정 시설은 없습니다.",
        "entities": "보이는 사람은 안규홍 한 명뿐이며, 검은 머리를 뒤로 단단히 묶은 청년 남성의 외형과 때 탄 밝은색 저고리가 참조와 대체로 맞습니다. 얼굴은 창백하고 입이 벌어져 있으며 눈동자는 정상입니다. 목재와 금속으로 된 소총, 이를 잡은 두 손, 흙바닥과 옹기 파편이 보입니다. 옥가락지는 노출되지 않습니다. 읽을 수 있는 글자는 없습니다.",
        "hard_violations": [],
        "physics": "앞손이 소총의 목재 아랫부분을 받치고 뒷손이 방아쇠 부근을 잡으며 개머리판은 어깨와 가슴 쪽에 닿아 있습니다. 멜빵은 총의 연결부에서 아래로 늘어집니다. 총이나 손에 지지 없는 부유는 없습니다. 다리와 발은 크롭 밖이므로 하체의 체중 배분은 판단할 수 없습니다. 옹기와 파편은 진열대 또는 땅에 놓여 있습니다."
       },
       {
        "label": "B",
        "direction": "시선은 렌즈를 약간 비껴 화면 오른쪽 밖을 향하고 소총 총열도 오른쪽 밖으로 이어집니다. 화면 안에 겨냥한 상대는 없으며 총검 역시 보이지 않습니다. 눈을 조준선에 맞추지 않고 총을 가슴 아래에 붙잡고 있어, 찌른 뒤 충격으로 동작이 멎은 순간에 더 가깝습니다.",
        "built_space": "얼굴 뒤와 양옆으로 흙길이 넓게 흐려져 보이고, 상단에는 흙벽과 목재 문, 낮은 진열대 및 차양 지지대가 있습니다. 오른쪽 뒤에는 입구가 카메라 쪽으로 향한 넘어진 큰 옹기 하나와 그 뒤의 세워진 옹기들이 보입니다. 왼쪽 가장자리에는 바닥에 기대어진 운반틀처럼 보이는 목조 물체 하나가 일부 들어옵니다. 깨진 옹기 파편은 명확하게 식별되지 않습니다. 참조와 재료 및 낮빛은 유사하지만 시설의 정확한 동일성까지 확인되지는 않습니다.",
        "entities": "안규홍 한 명만 보이며 청년 남성의 얼굴, 뒤로 묶은 검은 머리, 체격과 때 탄 밝은색 저고리가 참조에 가깝습니다. 창백한 얼굴과 벌어진 입, 멈춘 시선으로 충격을 표현하며 눈의 해부학적 변형은 없습니다. 소총의 손잡이와 앞쪽 목재 부분을 본인의 두 손으로 잡고 있고 무기는 하단을 가로지릅니다. 옥가락지는 보이지 않으며 글자나 다른 사람도 없습니다.",
        "hard_violations": [],
        "physics": "뒷손은 개머리판 목과 방아쇠 부근을 감싸고 앞손은 총의 앞쪽 목재를 받쳐 소총의 무게를 지탱합니다. 팔꿈치가 굽혀진 채 총을 몸 앞에 고정한 자세는 동작 중 굳어버린 상태로 가능합니다. 멜빵은 연결부에서 아래로 늘어지고, 배경의 넘어진 옹기는 땅에 닿아 있습니다. 발은 프레임 밖이라 지면 접촉은 확인할 수 있지만이 아니라 판단할 수 없으며, 보이는 범위에는 지지 없는 신체나 물체가 없습니다."
       }
      ],
      "all_candidates_fail": false
     },
     "normalized": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 7,
        "verdict_ko": "창백한 충격과 깨진 옹기는 잘 보이지만, 소총을 어깨에 붙여 조준하는 듯한 자세와 크게 드러난 무기가 얼굴 중심 클로즈업의 지시에서 벗어납니다."
       },
       {
        "label": "A",
        "score": 8,
        "verdict_ko": "굳어버린 얼굴을 중심으로 두고 손에 쥔 소총을 하단에 제한해 핵심 구도에 더 충실하나, 충돌로 깨진 옹기 잔해는 뚜렷하지 않습니다."
       }
      ],
      "all_candidates_fail": false,
      "readings": [
       {
        "label": "B",
        "direction": "시선과 소총의 총열은 모두 화면 왼쪽 밖을 향하며, 그 방향의 상대는 보이지 않습니다. 총검은 프레임 밖입니다. 개머리판을 어깨 쪽에 붙이고 눈을 총열 방향에 둔 모습은 찌른 직후 멈춘 상태보다 다시 조준하는 상태에 가깝게 읽힙니다.",
        "built_space": "인물은 흙길에 있고, 왼쪽 뒤로 목재 기둥 여러 개와 차양, 옹기 진열대, 흙벽 및 기와지붕이 이어집니다. 하단 왼쪽에는 깨진 옹기 조각들이 있고 그 뒤 진열대에는 온전한 옹기가 남아 있습니다. 뒤쪽에 지게처럼 보이는 틀 하나가 있으나 흐려 확정하기 어렵습니다. 참조의 시장 재료와 낮빛은 유지되지만 고정 시설의 정확한 동일 배치는 확인하기 어렵습니다. 반사면이나 명백히 중복된 고정 시설은 없습니다.",
        "entities": "보이는 사람은 안규홍 한 명뿐이며, 검은 머리를 뒤로 단단히 묶은 청년 남성의 외형과 때 탄 밝은색 저고리가 참조와 대체로 맞습니다. 얼굴은 창백하고 입이 벌어져 있으며 눈동자는 정상입니다. 목재와 금속으로 된 소총, 이를 잡은 두 손, 흙바닥과 옹기 파편이 보입니다. 옥가락지는 노출되지 않습니다. 읽을 수 있는 글자는 없습니다.",
        "hard_violations": [],
        "physics": "앞손이 소총의 목재 아랫부분을 받치고 뒷손이 방아쇠 부근을 잡으며 개머리판은 어깨와 가슴 쪽에 닿아 있습니다. 멜빵은 총의 연결부에서 아래로 늘어집니다. 총이나 손에 지지 없는 부유는 없습니다. 다리와 발은 크롭 밖이므로 하체의 체중 배분은 판단할 수 없습니다. 옹기와 파편은 진열대 또는 땅에 놓여 있습니다."
       },
       {
        "label": "A",
        "direction": "시선은 렌즈를 약간 비껴 화면 오른쪽 밖을 향하고 소총 총열도 오른쪽 밖으로 이어집니다. 화면 안에 겨냥한 상대는 없으며 총검 역시 보이지 않습니다. 눈을 조준선에 맞추지 않고 총을 가슴 아래에 붙잡고 있어, 찌른 뒤 충격으로 동작이 멎은 순간에 더 가깝습니다.",
        "built_space": "얼굴 뒤와 양옆으로 흙길이 넓게 흐려져 보이고, 상단에는 흙벽과 목재 문, 낮은 진열대 및 차양 지지대가 있습니다. 오른쪽 뒤에는 입구가 카메라 쪽으로 향한 넘어진 큰 옹기 하나와 그 뒤의 세워진 옹기들이 보입니다. 왼쪽 가장자리에는 바닥에 기대어진 운반틀처럼 보이는 목조 물체 하나가 일부 들어옵니다. 깨진 옹기 파편은 명확하게 식별되지 않습니다. 참조와 재료 및 낮빛은 유사하지만 시설의 정확한 동일성까지 확인되지는 않습니다.",
        "entities": "안규홍 한 명만 보이며 청년 남성의 얼굴, 뒤로 묶은 검은 머리, 체격과 때 탄 밝은색 저고리가 참조에 가깝습니다. 창백한 얼굴과 벌어진 입, 멈춘 시선으로 충격을 표현하며 눈의 해부학적 변형은 없습니다. 소총의 손잡이와 앞쪽 목재 부분을 본인의 두 손으로 잡고 있고 무기는 하단을 가로지릅니다. 옥가락지는 보이지 않으며 글자나 다른 사람도 없습니다.",
        "hard_violations": [],
        "physics": "뒷손은 개머리판 목과 방아쇠 부근을 감싸고 앞손은 총의 앞쪽 목재를 받쳐 소총의 무게를 지탱합니다. 팔꿈치가 굽혀진 채 총을 몸 앞에 고정한 자세는 동작 중 굳어버린 상태로 가능합니다. 멜빵은 연결부에서 아래로 늘어지고, 배경의 넘어진 옹기는 땅에 닿아 있습니다. 발은 프레임 밖이라 지면 접촉은 확인할 수 있지만이 아니라 판단할 수 없으며, 보이는 범위에는 지지 없는 신체나 물체가 없습니다."
       }
      ]
     },
     "ok": true
    }
   ],
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "slot_winner_match": true,
   "slot_winner": {
    "gemini-pro": "A",
    "gpt-high": "A"
   },
   "route": "cross_slot_agree"
  },
  "dual": {
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "normalized": {
    "A": 2.0,
    "B": 1.125
   },
   "adjusted": {
    "A": 2.0,
    "B": 0.875
   },
   "violations": {
    "B": [
     "[gemini-pro] 물리적으로 불가능한 신체 구조: 방아쇠를 잡고 있는 뒤쪽 손(왼팔)의 손목이 카메라 쪽(가까운 면)에 있으면서 손가락이 앞으로 감싸는 형태를 취하고 있어, 왼팔에 오른손이 달려 있는 해부학적 오류가 발생했습니다."
    ]
   },
   "per_model_winner": {
    "gemini-pro": "A",
    "gpt-high": "A"
   },
   "agreed": true
  },
  "totals": {
   "A": 2000,
   "B": 875
  },
  "selected": "A",
  "ranking": [
   "A",
   "B"
  ],
  "verdicts": [
   {
    "label": "A",
    "score": 2000,
    "verdict_ko": "레퍼런스와 일치하는 인물 외모와 올바른 한복 착장을 보여주며, 사람을 처음 찌른 충격에 굳은 표정과 소총 파지를 지시된 클로즈업 앵글 안에서 훌륭히 구현했습니다."
   },
   {
    "label": "B",
    "score": 875,
    "verdict_ko": "바닥에 깨진 도자기 파편을 지시대로 묘사했으나, 한복 깃이 반대로 뒤집히고 방아쇠를 잡은 뒤쪽 손의 해부학적 구조가 불가능하게 왜곡되어 치명적인 오류를 범했습니다.  ★위반: [gemini-pro] 물리적으로 불가능한 신체 구조: 방아쇠를 잡고 있는 뒤쪽 손(왼팔)의 손목이 카메라 쪽(가까운 면)에 있으면서 손가락이 앞으로 감싸는 형태를 취하고 있어, 왼팔에 오른손이 달려 있는 해부학적 오류가 발생했습니다."
   }
  ],
  "refs": [
   {
    "label": "PREVIOUS SHOT STILL — a visually related earlier shot of this same place: the location's look, materials, fixed features and lighting mood are LOCKED to this photo; never copy its camera framing. The clothing, hair and overall look of 안규홍 — who appear both in that photo and in this shot — are LOCKED to that photo. Anyone else visible in that photo is NOT in this shot: never carry their face, body or clothing onto anyone here. If this photo shows a character who cannot move (dead or unconscious), that character's exact pose, position and orientation are ALSO LOCKED — treat their body as a fixed prop of the set that only the camera moves around.",
    "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/db00517b-165b-4005-a287-f28936b59b95/images/272bd7af-9294-4490-98c6-88a40e9deacc/scene/recipe/S30sh21_sel.png",
    "asset_id": "a85537f6-51ad-4b4a-87c2-935ebeb672b4",
    "role": "prev_still"
   },
   {
    "label": "CHARACTER REFERENCE — 안규홍: the exact person appearing in this shot; match face, hair and build exactly.",
    "path": "<bytes:1028384>",
    "asset_id": "3b5894db-ce55-4f34-b966-5e725c3aa952",
    "role": "character_ref"
   }
  ],
  "critique_skipped": true,
  "shot_run_produced": false,
  "shot_run_uid": "06aa1423-168c-7c25-8567-644842634515",
  "ref_mode": "prev+엔티티",
  "share_plan": {
   "ref_plan": "prev",
   "prev_anchor_tag": "S30sh21"
  }
 },
 "S30sh38::cine": {
  "applied": false,
  "attempted_at": "2026-09-09T09:06:36.261776+00:00",
  "fingerprint": "54efb586022c31a1453f13e42e3d7e3781450a1de4381a5c12776f920afcad97",
  "fingerprint_version": 2,
  "provider": "mai",
  "endpoint": "openrouter/chat-completions",
  "model": "microsoft/mai-image-2.6",
  "pack": "24.202608252115",
  "source_file": "S30sh38_sel.png",
  "source_sha256": "7b33733d54403bffe679b3f559cb159fac0cc1e8f07b47136fd11ab8ba3573ae",
  "error": "RuntimeError: moderation blocked: {\"message\": \"Response content blocked by label 'MultiSeverity_ViolenceScore'.\", \"code\": 400}",
  "moderation_refusals": 2,
  "declined": true,
  "declined_reason": "moderation"
 },
 "S30sh50::signage": {
  "fp": "bcb6e8a7b56f9d42",
  "inscriptions": [],
  "cues": [],
  "dropped": []
 },
 "S30sh50": {
  "input_fingerprint": "2dab77d4f31aa28c",
  "prompt": "Create ONE FINAL photorealistic live-action film still of the moment below — early 20th-century Korea, Jeolla region; all people are Korean unless explicitly stated otherwise. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 흙바닥에 널브러진 일본 헌병들만 남겨진 텅 빈 장터의 스산한 전경.\n\nLOCATION (lock): The deserted dirt confrontation area beside the market stalls, with the subdued escort left behind near the rear-alley escape route. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: wide shot\n- KEY BACKGROUND ELEMENTS: 장터 흙바닥과 중앙 통로 (Only the fallen military policemen remain in the visible passage); used as Broad negative space separates the bodies and makes the rescuers' absence readable; 옹기 좌판과 깨진 옹기 (Disturbed and broken during the fight) — The stall is seen obliquely from above along the passage edge; used as Peripheral evidence of the brief violence, kept subordinate to the emptied space.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Retain unembellished daytime ambient light, muted saturation, and sober contrast without darkening the aftermath artificially.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The people visible in that photo are NOT in this shot: never carry their faces, bodies or clothing onto anyone here. Each person in THIS shot is defined solely by the PEOPLE list, their own CHARACTER REFERENCE images and their wardrobe notes. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nPREVIOUS STILL USAGE (follow exactly — what to take from the attached still and what to exclude; it governs place and objects only, never who is in this shot or how the camera sees them): Take the market-stall structures, dirt ground, and daytime lighting. Exclude the rescuers and prisoners who have left, and do not restore pottery broken during the fight.\n\nIMMOBILE CHARACTER POSE — CANONICAL (identical wherever this character appears in ANY panel; on any conflict THIS POSE WINS): His dead body is collapsed on the market's dirt ground, bleeding and remaining where he fell. The screenplay does not specify his torso orientation, head angle, or the placement of his arms and legs.\n\nIMMOBILE BODIES OBEY GRAVITY: a person who is dead or unconscious\nexerts NO muscular effort. Every part of their body — head, torso,\narms, hands, fingers, legs — rests fully on whatever supports it\n(floor, wall, furniture, their own lap) and hangs or slumps with\ngravity. NEVER show any part of an immobile person's body lifted,\nraised, held up in the air, or posed as if presenting something:\nan object in their grip stays clenched in a hand that itself lies\nfallen on a support — the hand does not hold the object up. If the\ncanonical pose leaves a body part unspecified, resolve it as the\nmost gravity-compliant, fully-supported position.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): Broken pottery and cut binding ropes remain in the market; the carrying frame has not been retrieved. 헌병(안규홍에게 총검으로 찔리는 헌병): His dead body is collapsed on the market's dirt ground with a bleeding chest wound.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 헌병(안규홍에게 총검으로 찔리는 헌병) (일본인 남성, 성인 나이대, 자연스러운 검은 머리) — wearing: 카키색 일본 헌병대 일반병 제복. 붉은색 테두리의 헌병 모자와 가죽 탄띠, 각반, 군화를 착용했다. — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.",
  "roll_prompts": {
   "A": "Create ONE FINAL photorealistic live-action film still of the moment below — early 20th-century Korea, Jeolla region; all people are Korean unless explicitly stated otherwise. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 흙바닥에 널브러진 일본 헌병들만 남겨진 텅 빈 장터의 스산한 전경.\n\nLOCATION (lock): The deserted dirt confrontation area beside the market stalls, with the subdued escort left behind near the rear-alley escape route. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: wide shot\n- KEY BACKGROUND ELEMENTS: 장터 흙바닥과 중앙 통로 (Only the fallen military policemen remain in the visible passage); used as Broad negative space separates the bodies and makes the rescuers' absence readable; 옹기 좌판과 깨진 옹기 (Disturbed and broken during the fight) — The stall is seen obliquely from above along the passage edge; used as Peripheral evidence of the brief violence, kept subordinate to the emptied space.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Retain unembellished daytime ambient light, muted saturation, and sober contrast without darkening the aftermath artificially.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The people visible in that photo are NOT in this shot: never carry their faces, bodies or clothing onto anyone here. Each person in THIS shot is defined solely by the PEOPLE list, their own CHARACTER REFERENCE images and their wardrobe notes. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nPREVIOUS STILL USAGE (follow exactly — what to take from the attached still and what to exclude; it governs place and objects only, never who is in this shot or how the camera sees them): Take the market-stall structures, dirt ground, and daytime lighting. Exclude the rescuers and prisoners who have left, and do not restore pottery broken during the fight.\n\nIMMOBILE CHARACTER POSE — CANONICAL (identical wherever this character appears in ANY panel; on any conflict THIS POSE WINS): His dead body is collapsed on the market's dirt ground, bleeding and remaining where he fell. The screenplay does not specify his torso orientation, head angle, or the placement of his arms and legs.\n\nIMMOBILE BODIES OBEY GRAVITY: a person who is dead or unconscious\nexerts NO muscular effort. Every part of their body — head, torso,\narms, hands, fingers, legs — rests fully on whatever supports it\n(floor, wall, furniture, their own lap) and hangs or slumps with\ngravity. NEVER show any part of an immobile person's body lifted,\nraised, held up in the air, or posed as if presenting something:\nan object in their grip stays clenched in a hand that itself lies\nfallen on a support — the hand does not hold the object up. If the\ncanonical pose leaves a body part unspecified, resolve it as the\nmost gravity-compliant, fully-supported position.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): Broken pottery and cut binding ropes remain in the market; the carrying frame has not been retrieved. 헌병(안규홍에게 총검으로 찔리는 헌병): His dead body is collapsed on the market's dirt ground with a bleeding chest wound.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 헌병(안규홍에게 총검으로 찔리는 헌병) (일본인 남성, 성인 나이대, 자연스러운 검은 머리) — wearing: 카키색 일본 헌병대 일반병 제복. 붉은색 테두리의 헌병 모자와 가죽 탄띠, 각반, 군화를 착용했다. — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.",
   "B": "Create ONE FINAL photorealistic live-action film still of the moment below — early 20th-century Korea, Jeolla region; all people are Korean unless explicitly stated otherwise. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 흙바닥에 널브러진 일본 헌병들만 남겨진 텅 빈 장터의 스산한 전경.\n\nLOCATION (lock): The deserted dirt confrontation area beside the market stalls, with the subdued escort left behind near the rear-alley escape route. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: wide shot\n- KEY BACKGROUND ELEMENTS: 장터 흙바닥과 중앙 통로 (Only the fallen military policemen remain in the visible passage); used as Broad negative space separates the bodies and makes the rescuers' absence readable; 옹기 좌판과 깨진 옹기 (Disturbed and broken during the fight) — The stall is seen obliquely from above along the passage edge; used as Peripheral evidence of the brief violence, kept subordinate to the emptied space.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Retain unembellished daytime ambient light, muted saturation, and sober contrast without darkening the aftermath artificially.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The people visible in that photo are NOT in this shot: never carry their faces, bodies or clothing onto anyone here. Each person in THIS shot is defined solely by the PEOPLE list, their own CHARACTER REFERENCE images and their wardrobe notes. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nPREVIOUS STILL USAGE (follow exactly — what to take from the attached still and what to exclude; it governs place and objects only, never who is in this shot or how the camera sees them): Take the market-stall structures, dirt ground, and daytime lighting. Exclude the rescuers and prisoners who have left, and do not restore pottery broken during the fight.\n\nIMMOBILE CHARACTER POSE — CANONICAL (identical wherever this character appears in ANY panel; on any conflict THIS POSE WINS): His dead body is collapsed on the market's dirt ground, bleeding and remaining where he fell. The screenplay does not specify his torso orientation, head angle, or the placement of his arms and legs.\n\nIMMOBILE BODIES OBEY GRAVITY: a person who is dead or unconscious\nexerts NO muscular effort. Every part of their body — head, torso,\narms, hands, fingers, legs — rests fully on whatever supports it\n(floor, wall, furniture, their own lap) and hangs or slumps with\ngravity. NEVER show any part of an immobile person's body lifted,\nraised, held up in the air, or posed as if presenting something:\nan object in their grip stays clenched in a hand that itself lies\nfallen on a support — the hand does not hold the object up. If the\ncanonical pose leaves a body part unspecified, resolve it as the\nmost gravity-compliant, fully-supported position.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): Broken pottery and cut binding ropes remain in the market; the carrying frame has not been retrieved. 헌병(안규홍에게 총검으로 찔리는 헌병): His dead body is collapsed on the market's dirt ground with a bleeding chest wound.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 헌병(안규홍에게 총검으로 찔리는 헌병) (일본인 남성, 성인 나이대, 자연스러운 검은 머리) — wearing: 카키색 일본 헌병대 일반병 제복. 붉은색 테두리의 헌병 모자와 가죽 탄띠, 각반, 군화를 착용했다. — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.\n\nCOMPOSITION VARIATION (this candidate only): keep the place, the people and the moment exactly as specified, and keep from the CAMERA & FRAME contract its subject scale, screen placement and key background elements exactly. For this candidate only, its camera angle and height are deliberately left open: choose a camera position distinctly different from the obvious one, as a film director picking a second setup on the same blocking. Stay true to any explicit framing the SHOT TEXT itself dictates."
  },
  "shot_run_spend_attempt_count": 1,
  "readings": [
   {
    "label": "A",
    "direction": "카메라는 장터 중앙 통로를 따라 원경의 문을 향하고 있으며, 시체들은 바닥에 다양한 방향으로 널브러져 있음.",
    "built_space": "레퍼런스와 일치하는 장터 구조. 흙바닥 중앙 통로를 기준으로 우측에 옹기들이 배열되어 있고, 바닥에 깨진 옹기 조각과 밧줄이 흩어져 있음.",
    "entities": "가슴에 피를 흘리며 카키색 제복과 붉은 테두리 모자를 쓴 일본 헌병들의 시체만 존재함. 그 외의 인물은 없음.",
    "hard_violations": [],
    "physics": "모든 시체와 사물들이 흙바닥 위에 중력에 맞게 자연스럽게 안착되어 있음."
   },
   {
    "label": "B",
    "direction": "카메라는 장터 중앙 통로를 따라 원경을 향함. 전경의 헌병 시체는 하늘을 향해 누워 있음.",
    "built_space": "우측의 옹기 배열, 좌측의 쓰러진 지게, 바닥의 깨진 옹기 조각 등 장터의 전반적인 구조와 소품 배치가 레퍼런스와 일치함.",
    "entities": "전경에 지시된 일본 헌병 시체가 있으나, 중경에 흰옷을 입은 신원 미상의 시체들이 있고, 원경의 구조물 아래에 앉아있는 살아있는 인물이 발견됨.",
    "hard_violations": [
     "[gemini-pro] 프롬프트에 없는 인물 임의 추가 (원경에 앉아있는 생존 인물 및 중경의 흰옷 입은 시체들)",
     "[gpt-high] 뒤쪽 대문 오른편 바구니 주변에 쓰러진 헌병이 아닌 앉거나 웅크린 인물들이 남아 있어, 허용되지 않은 사람들이 장면에 추가되었다."
    ],
    "physics": "시체와 지게, 옹기 등 모든 객체가 흙바닥에 안정적으로 지탱되어 있음."
   }
  ],
  "cross_model_order": {
   "policy": "cross_model_order_v2_slot0_fwd_slot1_rev_parallel2",
   "slots": [
    {
     "model": "gemini-pro",
     "order": "forward",
     "display_to_canonical": {
      "A": "A",
      "B": "B"
     },
     "raw": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 9,
        "verdict_ko": "프롬프트에 명시된 대로 텅 빈 장터에 쓰러진 일본 헌병들의 모습만 정확히 묘사하였으며, 배경에 다른 인물이나 생존자를 추가하지 않아 샷 텍스트와 제한 사항을 충실히 준수했습니다."
       },
       {
        "label": "B",
        "score": 4,
        "verdict_ko": "쓰러진 일본 헌병 외에 흰옷을 입은 다른 시체들이 포함되어 있고, 원경에 앉아있는 생존 인물까지 등장하여 '일본 헌병들만 남겨진 텅 빈 장터'라는 핵심 지시를 명백히 위반했습니다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "카메라는 장터 중앙 통로를 따라 원경의 문을 향하고 있으며, 시체들은 바닥에 다양한 방향으로 널브러져 있음.",
        "built_space": "레퍼런스와 일치하는 장터 구조. 흙바닥 중앙 통로를 기준으로 우측에 옹기들이 배열되어 있고, 바닥에 깨진 옹기 조각과 밧줄이 흩어져 있음.",
        "entities": "가슴에 피를 흘리며 카키색 제복과 붉은 테두리 모자를 쓴 일본 헌병들의 시체만 존재함. 그 외의 인물은 없음.",
        "hard_violations": [],
        "physics": "모든 시체와 사물들이 흙바닥 위에 중력에 맞게 자연스럽게 안착되어 있음."
       },
       {
        "label": "B",
        "direction": "카메라는 장터 중앙 통로를 따라 원경을 향함. 전경의 헌병 시체는 하늘을 향해 누워 있음.",
        "built_space": "우측의 옹기 배열, 좌측의 쓰러진 지게, 바닥의 깨진 옹기 조각 등 장터의 전반적인 구조와 소품 배치가 레퍼런스와 일치함.",
        "entities": "전경에 지시된 일본 헌병 시체가 있으나, 중경에 흰옷을 입은 신원 미상의 시체들이 있고, 원경의 구조물 아래에 앉아있는 살아있는 인물이 발견됨.",
        "hard_violations": [
         "프롬프트에 없는 인물 임의 추가 (원경에 앉아있는 생존 인물 및 중경의 흰옷 입은 시체들)"
        ],
        "physics": "시체와 지게, 옹기 등 모든 객체가 흙바닥에 안정적으로 지탱되어 있음."
       }
      ],
      "all_candidates_fail": false
     },
     "normalized": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 9,
        "verdict_ko": "프롬프트에 명시된 대로 텅 빈 장터에 쓰러진 일본 헌병들의 모습만 정확히 묘사하였으며, 배경에 다른 인물이나 생존자를 추가하지 않아 샷 텍스트와 제한 사항을 충실히 준수했습니다."
       },
       {
        "label": "B",
        "score": 4,
        "verdict_ko": "쓰러진 일본 헌병 외에 흰옷을 입은 다른 시체들이 포함되어 있고, 원경에 앉아있는 생존 인물까지 등장하여 '일본 헌병들만 남겨진 텅 빈 장터'라는 핵심 지시를 명백히 위반했습니다."
       }
      ],
      "all_candidates_fail": false,
      "readings": [
       {
        "label": "A",
        "direction": "카메라는 장터 중앙 통로를 따라 원경의 문을 향하고 있으며, 시체들은 바닥에 다양한 방향으로 널브러져 있음.",
        "built_space": "레퍼런스와 일치하는 장터 구조. 흙바닥 중앙 통로를 기준으로 우측에 옹기들이 배열되어 있고, 바닥에 깨진 옹기 조각과 밧줄이 흩어져 있음.",
        "entities": "가슴에 피를 흘리며 카키색 제복과 붉은 테두리 모자를 쓴 일본 헌병들의 시체만 존재함. 그 외의 인물은 없음.",
        "hard_violations": [],
        "physics": "모든 시체와 사물들이 흙바닥 위에 중력에 맞게 자연스럽게 안착되어 있음."
       },
       {
        "label": "B",
        "direction": "카메라는 장터 중앙 통로를 따라 원경을 향함. 전경의 헌병 시체는 하늘을 향해 누워 있음.",
        "built_space": "우측의 옹기 배열, 좌측의 쓰러진 지게, 바닥의 깨진 옹기 조각 등 장터의 전반적인 구조와 소품 배치가 레퍼런스와 일치함.",
        "entities": "전경에 지시된 일본 헌병 시체가 있으나, 중경에 흰옷을 입은 신원 미상의 시체들이 있고, 원경의 구조물 아래에 앉아있는 살아있는 인물이 발견됨.",
        "hard_violations": [
         "프롬프트에 없는 인물 임의 추가 (원경에 앉아있는 생존 인물 및 중경의 흰옷 입은 시체들)"
        ],
        "physics": "시체와 지게, 옹기 등 모든 객체가 흙바닥에 안정적으로 지탱되어 있음."
       }
      ]
     },
     "ok": true
    },
    {
     "model": "gpt-high",
     "order": "reverse",
     "display_to_canonical": {
      "A": "B",
      "B": "A"
     },
     "raw": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 3,
        "verdict_ko": "넓은 빈 통로와 내려다본 옹기 좌판은 지시대로지만, 후경에 앉아 있는 인물들이 남아 있어 쓰러진 헌병들만 남겨야 한다는 핵심 조건을 위반한다."
       },
       {
        "label": "B",
        "score": 8,
        "verdict_ko": "쓰러진 군복 인물들 사이의 빈 흙길을 보여주는 낮의 와이드숏으로, 구조자들의 부재와 전투 뒤의 흔적을 더 충실하게 구현한다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "전경 헌병은 머리가 화면 오른쪽 위, 발이 왼쪽 아래를 향하도록 누워 있고 얼굴은 위를 향한다. 다른 쓰러진 인물들도 통로 가장자리에 가로놓여 있다. 누군가를 겨누는 총구나 능동적인 시선은 보이지 않는다.",
        "built_space": "중앙 흙길 양쪽에 목조 점포와 좌판이 이어지고, 왼쪽에는 거친 차양과 지주들, 오른쪽에는 밝은 천 차양 하나와 옹기 진열대가 있다. 오른쪽에는 철망 상자 하나, 통로 왼쪽에는 원뿔형 운반 틀 하나가 보인다. 사선으로 내려다본 좌판과 넓은 빈 바닥은 요구에 맞지만, 뒤쪽 대문 오른편에 사람들이 남아 있어 장터가 완전히 비어 있지 않다.",
        "entities": "전경에는 가슴에 피가 묻은 성인 동아시아계 남성 헌병이 보이며 카키 군복, 테두리 있는 군모, 가죽 탄띠와 각반을 착용했다. 일본인 여부는 외모만으로 확정할 수 없다. 그 밖에 쓰러진 몸 세 구가 보이는데 오른쪽 중경 인물의 밝은 옷은 헌병 군복으로 명확히 읽히지 않는다. 후경 바구니 주변에는 앉거나 웅크린 인물들이 추가로 보인다. 깨진 옹기, 바닥의 밧줄과 남겨진 운반 틀은 확인되며 읽을 수 있는 글자는 없다.",
        "hard_violations": [
         "뒤쪽 대문 오른편 바구니 주변에 쓰러진 헌병이 아닌 앉거나 웅크린 인물들이 남아 있어, 허용되지 않은 사람들이 장면에 추가되었다."
        ],
        "physics": "전경 시신의 머리와 몸통, 양팔과 다리는 흙바닥에 놓여 있으며 손이 물건을 들어 올리고 있지 않다. 나머지 쓰러진 몸들도 지면에 지지된다. 옹기는 좌판 위에, 파편과 밧줄 및 운반 틀은 바닥에 놓여 있다. 명백하게 공중에 떠 있는 몸이나 물체는 보이지 않는다."
       },
       {
        "label": "B",
        "direction": "앞쪽 왼편 시신은 머리가 왼쪽, 발이 오른쪽을 향하고, 앞쪽 오른편의 흉부 출혈 시신은 그 반대 방향으로 누워 있다. 중경의 몸들도 통로를 가로지르는 방향으로 쓰러져 있다. 얼굴은 위나 옆으로 돌아가 있으며 카메라나 특정 인물을 능동적으로 바라보거나 겨누는 동작은 없다.",
        "built_space": "빈 중앙 흙길 양옆으로 목조 점포와 좌판이 이어진다. 왼쪽의 거친 차양과 지주들, 오른쪽의 밝은 천 차양 하나, 옹기 진열대와 철망 상자 하나가 기준 장소의 주요 구성을 따른다. 쓰러진 몸은 전경 두 구, 중경 중앙과 오른쪽에 각 한 구, 왼쪽 좌판 아래 일부 가려진 한 구로 보인다. 몸들 사이와 뒤쪽에 빈 통로가 남고, 오른쪽 좌판의 윗면도 비스듬히 보인다. 기준보다 중앙 대문이 더 뚜렷하지만 중복 설비나 불가능한 반사는 없다.",
        "entities": "보이는 인물들은 모두 쓰러져 있다. 전경 두 명은 성인 동아시아계 남성으로 카키 군복, 군모, 가죽 탄띠와 각반을 착용하며 오른쪽 인물에게 가슴 출혈이 있다. 중경 두 명도 옅은 군복과 허리띠를 착용한 모습이다. 왼쪽 좌판 아래 몸은 가려져 세부 신원을 확인하기 어렵다. 일본인이라는 국적과 검은 머리는 이 거리와 모자 때문에 확정하기 어렵다. 온전한 옹기와 깨진 파편, 바닥의 밧줄은 보이나 남겨진 운반 틀은 뚜렷하게 식별되지 않는다. 읽을 수 있는 글자는 없다.",
        "hard_violations": [],
        "physics": "전경 두 시신의 머리, 몸통, 팔과 다리는 지면에 놓여 있다. 중경 왼쪽 인물은 옆으로 접힌 자세로 바닥과 자기 몸에 지지되며, 손이나 머리를 능동적으로 들고 있다고 볼 명확한 징후는 없다. 다른 몸들도 바닥에 누워 있다. 옹기는 진열대가 받치고 파편과 밧줄은 흙바닥에 놓여 있어 무지지 부유는 보이지 않는다."
       }
      ],
      "all_candidates_fail": false
     },
     "normalized": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 3,
        "verdict_ko": "넓은 빈 통로와 내려다본 옹기 좌판은 지시대로지만, 후경에 앉아 있는 인물들이 남아 있어 쓰러진 헌병들만 남겨야 한다는 핵심 조건을 위반한다."
       },
       {
        "label": "A",
        "score": 8,
        "verdict_ko": "쓰러진 군복 인물들 사이의 빈 흙길을 보여주는 낮의 와이드숏으로, 구조자들의 부재와 전투 뒤의 흔적을 더 충실하게 구현한다."
       }
      ],
      "all_candidates_fail": false,
      "readings": [
       {
        "label": "B",
        "direction": "전경 헌병은 머리가 화면 오른쪽 위, 발이 왼쪽 아래를 향하도록 누워 있고 얼굴은 위를 향한다. 다른 쓰러진 인물들도 통로 가장자리에 가로놓여 있다. 누군가를 겨누는 총구나 능동적인 시선은 보이지 않는다.",
        "built_space": "중앙 흙길 양쪽에 목조 점포와 좌판이 이어지고, 왼쪽에는 거친 차양과 지주들, 오른쪽에는 밝은 천 차양 하나와 옹기 진열대가 있다. 오른쪽에는 철망 상자 하나, 통로 왼쪽에는 원뿔형 운반 틀 하나가 보인다. 사선으로 내려다본 좌판과 넓은 빈 바닥은 요구에 맞지만, 뒤쪽 대문 오른편에 사람들이 남아 있어 장터가 완전히 비어 있지 않다.",
        "entities": "전경에는 가슴에 피가 묻은 성인 동아시아계 남성 헌병이 보이며 카키 군복, 테두리 있는 군모, 가죽 탄띠와 각반을 착용했다. 일본인 여부는 외모만으로 확정할 수 없다. 그 밖에 쓰러진 몸 세 구가 보이는데 오른쪽 중경 인물의 밝은 옷은 헌병 군복으로 명확히 읽히지 않는다. 후경 바구니 주변에는 앉거나 웅크린 인물들이 추가로 보인다. 깨진 옹기, 바닥의 밧줄과 남겨진 운반 틀은 확인되며 읽을 수 있는 글자는 없다.",
        "hard_violations": [
         "뒤쪽 대문 오른편 바구니 주변에 쓰러진 헌병이 아닌 앉거나 웅크린 인물들이 남아 있어, 허용되지 않은 사람들이 장면에 추가되었다."
        ],
        "physics": "전경 시신의 머리와 몸통, 양팔과 다리는 흙바닥에 놓여 있으며 손이 물건을 들어 올리고 있지 않다. 나머지 쓰러진 몸들도 지면에 지지된다. 옹기는 좌판 위에, 파편과 밧줄 및 운반 틀은 바닥에 놓여 있다. 명백하게 공중에 떠 있는 몸이나 물체는 보이지 않는다."
       },
       {
        "label": "A",
        "direction": "앞쪽 왼편 시신은 머리가 왼쪽, 발이 오른쪽을 향하고, 앞쪽 오른편의 흉부 출혈 시신은 그 반대 방향으로 누워 있다. 중경의 몸들도 통로를 가로지르는 방향으로 쓰러져 있다. 얼굴은 위나 옆으로 돌아가 있으며 카메라나 특정 인물을 능동적으로 바라보거나 겨누는 동작은 없다.",
        "built_space": "빈 중앙 흙길 양옆으로 목조 점포와 좌판이 이어진다. 왼쪽의 거친 차양과 지주들, 오른쪽의 밝은 천 차양 하나, 옹기 진열대와 철망 상자 하나가 기준 장소의 주요 구성을 따른다. 쓰러진 몸은 전경 두 구, 중경 중앙과 오른쪽에 각 한 구, 왼쪽 좌판 아래 일부 가려진 한 구로 보인다. 몸들 사이와 뒤쪽에 빈 통로가 남고, 오른쪽 좌판의 윗면도 비스듬히 보인다. 기준보다 중앙 대문이 더 뚜렷하지만 중복 설비나 불가능한 반사는 없다.",
        "entities": "보이는 인물들은 모두 쓰러져 있다. 전경 두 명은 성인 동아시아계 남성으로 카키 군복, 군모, 가죽 탄띠와 각반을 착용하며 오른쪽 인물에게 가슴 출혈이 있다. 중경 두 명도 옅은 군복과 허리띠를 착용한 모습이다. 왼쪽 좌판 아래 몸은 가려져 세부 신원을 확인하기 어렵다. 일본인이라는 국적과 검은 머리는 이 거리와 모자 때문에 확정하기 어렵다. 온전한 옹기와 깨진 파편, 바닥의 밧줄은 보이나 남겨진 운반 틀은 뚜렷하게 식별되지 않는다. 읽을 수 있는 글자는 없다.",
        "hard_violations": [],
        "physics": "전경 두 시신의 머리, 몸통, 팔과 다리는 지면에 놓여 있다. 중경 왼쪽 인물은 옆으로 접힌 자세로 바닥과 자기 몸에 지지되며, 손이나 머리를 능동적으로 들고 있다고 볼 명확한 징후는 없다. 다른 몸들도 바닥에 누워 있다. 옹기는 진열대가 받치고 파편과 밧줄은 흙바닥에 놓여 있어 무지지 부유는 보이지 않는다."
       }
      ]
     },
     "ok": true
    }
   ],
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "slot_winner_match": true,
   "slot_winner": {
    "gemini-pro": "A",
    "gpt-high": "A"
   },
   "route": "cross_slot_agree"
  },
  "dual": {
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "normalized": {
    "A": 2.0,
    "B": 0.819
   },
   "adjusted": {
    "A": 2.0,
    "B": 0.569
   },
   "violations": {
    "B": [
     "[gemini-pro] 프롬프트에 없는 인물 임의 추가 (원경에 앉아있는 생존 인물 및 중경의 흰옷 입은 시체들)",
     "[gpt-high] 뒤쪽 대문 오른편 바구니 주변에 쓰러진 헌병이 아닌 앉거나 웅크린 인물들이 남아 있어, 허용되지 않은 사람들이 장면에 추가되었다."
    ]
   },
   "per_model_winner": {
    "gemini-pro": "A",
    "gpt-high": "A"
   },
   "agreed": true
  },
  "totals": {
   "A": 2000,
   "B": 569
  },
  "selected": "A",
  "ranking": [
   "A",
   "B"
  ],
  "verdicts": [
   {
    "label": "A",
    "score": 2000,
    "verdict_ko": "프롬프트에 명시된 대로 텅 빈 장터에 쓰러진 일본 헌병들의 모습만 정확히 묘사하였으며, 배경에 다른 인물이나 생존자를 추가하지 않아 샷 텍스트와 제한 사항을 충실히 준수했습니다."
   },
   {
    "label": "B",
    "score": 569,
    "verdict_ko": "쓰러진 일본 헌병 외에 흰옷을 입은 다른 시체들이 포함되어 있고, 원경에 앉아있는 생존 인물까지 등장하여 '일본 헌병들만 남겨진 텅 빈 장터'라는 핵심 지시를 명백히 위반했습니다.  ★위반: [gemini-pro] 프롬프트에 없는 인물 임의 추가 (원경에 앉아있는 생존 인물 및 중경의 흰옷 입은 시체들) / [gpt-high] 뒤쪽 대문 오른편 바구니 주변에 쓰러진 헌병이 아닌 앉거나 웅크린 인물들이 남아 있어, 허용되지 않은 사람들이 장면에 추가되었다."
   }
  ],
  "refs": [
   {
    "label": "PREVIOUS SHOT STILL — a visually related earlier shot of this same place: the location's look, materials, fixed features and lighting mood are LOCKED to this photo; never copy its camera framing. The people visible in that photo are NOT in this shot: never carry their faces, bodies or clothing onto anyone here. Each person in THIS shot is defined solely by the PEOPLE list, their own CHARACTER REFERENCE images and their wardrobe notes. If this photo shows a character who cannot move (dead or unconscious), that character's exact pose, position and orientation are ALSO LOCKED — treat their body as a fixed prop of the set that only the camera moves around.",
    "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/db00517b-165b-4005-a287-f28936b59b95/images/272bd7af-9294-4490-98c6-88a40e9deacc/scene/recipe/S30sh21_sel.png",
    "asset_id": "a85537f6-51ad-4b4a-87c2-935ebeb672b4",
    "role": "prev_still"
   }
  ],
  "critique_skipped": true,
  "shot_run_produced": false,
  "shot_run_uid": "06aa1423-fb4f-7e15-b3dc-8f4cb82e0b20",
  "ref_mode": "prev+엔티티",
  "share_plan": {
   "ref_plan": "prev",
   "prev_anchor_tag": "S30sh21"
  },
  "lane_policy": "ab_select_bypass:prev"
 },
 "S30sh50::cine": {
  "applied": false,
  "attempted_at": "2026-09-09T09:06:54.728580+00:00",
  "fingerprint": "8ccaf9bc183239a2c5bcfdd7eb6ba7f64d0b447e574b344d87c5f8127ec17798",
  "fingerprint_version": 2,
  "provider": "mai",
  "endpoint": "openrouter/chat-completions",
  "model": "microsoft/mai-image-2.6",
  "pack": "24.202608252115",
  "source_file": "S30sh50_sel.png",
  "source_sha256": "de615d897afd6a3ae427dd8dcaec423b5465329391caa118eb004af7aff9997f",
  "error": "RuntimeError: moderation blocked: {\"message\": \"Response content blocked by label 'MultiSeverity_ViolenceScore'.\", \"code\": 400}",
  "moderation_refusals": 2,
  "declined": true,
  "declined_reason": "moderation"
 },
 "S31sh1::signage": {
  "fp": "009915c4df3a766f",
  "inscriptions": [],
  "cues": [],
  "dropped": []
 },
 "era_assess::0b176419a3216c6d": {
  "subjects": [],
  "subject_text": "해안 바위지대·모래밭\n거친 파도가 부서지는 해안 바위지대. 큰 바위 뒤로 모래밭이 이어지고, 바위 가장자리에 흰 포말이 번진다.",
  "identity": "canonical",
  "scope_id": "L28",
  "scope_role": "location_exterior",
  "scope_sha": "8b89132d8ec90742"
 },
 "groupbg::coastal_rock_shelter": {
  "input_fingerprint": "eca5595bc9a9299a",
  "meta": {
   "model": "gpt-image-2",
   "size": "1536x864",
   "pack": "11.202607220237",
   "contract": "bgfirst_full_v3",
   "group_sig": {
    "key": "coastal_rock_shelter",
    "tags": [
     "S31sh1",
     "S31sh12",
     "S31sh17"
    ]
   },
   "context_sig": "78377eb92684e5c6"
  },
  "prompt": "Create ONE empty live-action location background photograph — NO PEOPLE, no figures, no body parts, no silhouettes, no shadows or reflections of people anywhere.\n\"Empty\" means no people only: KEEP the location's inherent occupants and stock that define the place — animals in an animal shelter, pen or farm, goods and produce in a market, moored boats in a harbour — unless the shot text explicitly removes them.\nThe FIRST attached image is a thin-line storyboard sketch of a shot that happens at this location: use it ONLY as spatial evidence — what this place contains, how its ground, structures and landmarks are arranged and proportioned. Ignore the sketched people and arrows entirely, and do NOT copy its line style: render a fully photographic, physically plausible real place that fits THE LOCATION text below.\nHUMAN-SCALE CALIBRATION: derive every structure's true size from human-scale elements — a door ≈ 2m, a window ≈ 1–1.5m wide, one storey ≈ 2.5–3m; never inflate a small structure or shrink a large one.\n\nTHE LOCATION — early 20th-century Korea, Jeolla region; all people are Korean unless explicitly stated otherwise: The sandy refuge immediately behind large coastal rocks, beside rough breaking surf.\n\nLOCATION DETAIL (descriptive evidence about this place from the production's location records — evidence, not a staging order): use it to understand what kind of place this is and what it permanently contains. Carry over ONLY its enduring physical features — layout, structures, ground and materials, fixed equipment and installations. Do NOT treat any lighting, weather or time-of-day wording, momentary object states, depicted events or subjective impressions in it as fixed facts of the place; the TIME OF DAY lock below is the sole authority for time and lighting.\n해안 바위지대·모래밭: 거친 파도가 밀려와 부서지는 넓은 해안가 모래밭. 사람들이 숨거나 쉴 수 있는 큼지막한 갯바위들이 늘어선 구역이다. (특징: 거친 파도가 치는 바다; 넓은 모래사장; 곳곳에 솟아오른 커다란 해안 바위들)\n\nSCENE EVIDENCE (verbatim quotes from the screenplay about this place — treat them as evidence of what the location physically contains and looks like; stage the PLACE those moments happen in, but do NOT depict the momentary actions, people or staged props themselves):\n- 거친 파도 소리.\n규홍 일행이 바위 뒤로 뛰어들어온다. 구해낸 포로 셋도 함께다.\n- 사람들이 모래위에 주저않아 거친 숨을 몰아쉰다.\n\nTIME OF DAY (lock): afternoon.\n\nRender ONE photorealistic empty location photograph, 16:9, neutral enough that every shot of this place can be staged from it later. No readable writing anywhere: surfaces that would carry writing may be present, but stage any wording out of legibility — an oblique angle, distance, shallow focus. No captions, watermarks or overlay text.",
  "effective_prompt": "Create ONE empty live-action location background photograph — NO PEOPLE, no figures, no body parts, no silhouettes, no shadows or reflections of people anywhere.\n\"Empty\" means no people only: KEEP the location's inherent occupants and stock that define the place — animals in an animal shelter, pen or farm, goods and produce in a market, moored boats in a harbour — unless the shot text explicitly removes them.\nThe FIRST attached image is a thin-line storyboard sketch of a shot that happens at this location: use it ONLY as spatial evidence — what this place contains, how its ground, structures and landmarks are arranged and proportioned. Ignore the sketched people and arrows entirely, and do NOT copy its line style: render a fully photographic, physically plausible real place that fits THE LOCATION text below.\nHUMAN-SCALE CALIBRATION: derive every structure's true size from human-scale elements — a door ≈ 2m, a window ≈ 1–1.5m wide, one storey ≈ 2.5–3m; never inflate a small structure or shrink a large one.\n\nTHE LOCATION — early 20th-century Korea, Jeolla region; all people are Korean unless explicitly stated otherwise: The sandy refuge immediately behind large coastal rocks, beside rough breaking surf.\n\nLOCATION DETAIL (descriptive evidence about this place from the production's location records — evidence, not a staging order): use it to understand what kind of place this is and what it permanently contains. Carry over ONLY its enduring physical features — layout, structures, ground and materials, fixed equipment and installations. Do NOT treat any lighting, weather or time-of-day wording, momentary object states, depicted events or subjective impressions in it as fixed facts of the place; the TIME OF DAY lock below is the sole authority for time and lighting.\n해안 바위지대·모래밭: 거친 파도가 밀려와 부서지는 넓은 해안가 모래밭. 사람들이 숨거나 쉴 수 있는 큼지막한 갯바위들이 늘어선 구역이다. (특징: 거친 파도가 치는 바다; 넓은 모래사장; 곳곳에 솟아오른 커다란 해안 바위들)\n\nSCENE EVIDENCE (verbatim quotes from the screenplay about this place — treat them as evidence of what the location physically contains and looks like; stage the PLACE those moments happen in, but do NOT depict the momentary actions, people or staged props themselves):\n- 거친 파도 소리.\n규홍 일행이 바위 뒤로 뛰어들어온다. 구해낸 포로 셋도 함께다.\n- 사람들이 모래위에 주저않아 거친 숨을 몰아쉰다.\n\nTIME OF DAY (lock): afternoon.\n\nRender ONE photorealistic empty location photograph, 16:9, neutral enough that every shot of this place can be staged from it later. No readable writing anywhere: surfaces that would carry writing may be present, but stage any wording out of legibility — an oblique angle, distance, shallow focus. No captions, watermarks or overlay text.",
  "bg_path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/db00517b-165b-4005-a287-f28936b59b95/images/272bd7af-9294-4490-98c6-88a40e9deacc/scene/recipe/groupbg_coastal_rock_shelter_9ea29b.png",
  "asset_id": "c54772cc-ea97-4590-93d2-3432743dcf57",
  "input_asset_ids": [
   "34d329c4-6159-44b7-bf3f-98f6943cb851"
  ],
  "origin_tag": "S31sh1",
  "place_text": "The sandy refuge immediately behind large coastal rocks, beside rough breaking surf.",
  "origin_inputs": {
   "place_text": "The sandy refuge immediately behind large coastal rocks, beside rough breaking surf.",
   "time_of_day_en": "afternoon",
   "conti_asset_id": "34d329c4-6159-44b7-bf3f-98f6943cb851"
  }
 },
 "S31sh1::bgfirst_bg": {
  "input_fingerprint": "73a70ebddc7c1797",
  "prompt": "Create the EMPTY BACKGROUND PLATE for one film shot — NO PEOPLE, no figures, no body parts, no sketch lines, no arrows anywhere.\n\"Empty\" means no people only: KEEP the location's inherent occupants and stock that define the place — animals in an animal shelter, pen or farm, goods in a market, moored boats in a harbour — unless the shot text explicitly removes them.\nThe FIRST attached image is a thin-line storyboard sketch: use ONLY its camera angle, horizon, perspective and the placement/size of buildings and set masses — ignore the sketched people and arrows entirely. The SECOND attached image (LOCATION PHOTOGRAPH) is the real place: take its architecture, materials, signage and fixed features, and RE-PROJECT them into the sketch's camera. If the photograph's camera differs from the sketch's, the sketch's camera wins.\nHUMAN-SCALE CALIBRATION: derive every structure's true size from human-scale elements — a door ≈ 2m, a window ≈ 1–1.5m wide, one storey ≈ 2.5–3m; never inflate a small structure or shrink a large one.\n\nSHOT TEXT this background must serve (Korean): 거친 파도가 치는 해안가 거대한 바위 뒤로 한쪽 발을 딛고 뒷발로 모래를 차올리며 달려 들어오는 미드 액션, 앞을 향해 상체를 기울인 안규홍, 이관회, 염재보, 의병 포로1, 의병 포로2, 의병 포로3의 역동적인 전경.\n\nLOCATION (lock): The sandy refuge immediately behind large coastal rocks, beside rough breaking surf.\n\nTIME OF DAY (lock): afternoon.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: wide shot\n- FRAME LAYOUT: Sheltered arrival space behind the coastal rock in the middle-right of the frame, midground.\n- KEY BACKGROUND ELEMENTS: 해안 바위 (Sheltering the arriving group) — The shelter-facing side appears at the right edge, with the arrival route visible beside it; used as Defines the destination without hiding the runners or overwhelming their scale; 모래 바닥 (Disturbed beneath the arriving feet); used as Separates planted and airborne feet and makes the staggered stride phases readable; 해안 파도 (Breaking behind the rocks); used as A limited background opening establishes the coast without competing with the arrival.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Use natural afternoon ambient light with restrained saturation and sufficient tonal openness to distinguish the overlapping bodies without polishing away their exhaustion.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nRender ONE photorealistic empty location photograph, 16:9, that this shot can be staged inside later. No readable writing anywhere: surfaces that would carry writing may be present, but stage any wording out of legibility — an oblique angle, distance, shallow focus. No captions, watermarks or overlay text.",
  "effective_prompt": "Create the EMPTY BACKGROUND PLATE for one film shot — NO PEOPLE, no figures, no body parts, no sketch lines, no arrows anywhere.\n\"Empty\" means no people only: KEEP the location's inherent occupants and stock that define the place — animals in an animal shelter, pen or farm, goods in a market, moored boats in a harbour — unless the shot text explicitly removes them.\nThe FIRST attached image is a thin-line storyboard sketch: use ONLY its camera angle, horizon, perspective and the placement/size of buildings and set masses — ignore the sketched people and arrows entirely. The SECOND attached image (LOCATION PHOTOGRAPH) is the real place: take its architecture, materials, signage and fixed features, and RE-PROJECT them into the sketch's camera. If the photograph's camera differs from the sketch's, the sketch's camera wins.\nHUMAN-SCALE CALIBRATION: derive every structure's true size from human-scale elements — a door ≈ 2m, a window ≈ 1–1.5m wide, one storey ≈ 2.5–3m; never inflate a small structure or shrink a large one.\n\nSHOT TEXT this background must serve (Korean): 거친 파도가 치는 해안가 거대한 바위 뒤로 한쪽 발을 딛고 뒷발로 모래를 차올리며 달려 들어오는 미드 액션, 앞을 향해 상체를 기울인 안규홍, 이관회, 염재보, 의병 포로1, 의병 포로2, 의병 포로3의 역동적인 전경.\n\nLOCATION (lock): The sandy refuge immediately behind large coastal rocks, beside rough breaking surf.\n\nTIME OF DAY (lock): afternoon.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: wide shot\n- FRAME LAYOUT: Sheltered arrival space behind the coastal rock in the middle-right of the frame, midground.\n- KEY BACKGROUND ELEMENTS: 해안 바위 (Sheltering the arriving group) — The shelter-facing side appears at the right edge, with the arrival route visible beside it; used as Defines the destination without hiding the runners or overwhelming their scale; 모래 바닥 (Disturbed beneath the arriving feet); used as Separates planted and airborne feet and makes the staggered stride phases readable; 해안 파도 (Breaking behind the rocks); used as A limited background opening establishes the coast without competing with the arrival.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Use natural afternoon ambient light with restrained saturation and sufficient tonal openness to distinguish the overlapping bodies without polishing away their exhaustion.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nRender ONE photorealistic empty location photograph, 16:9, that this shot can be staged inside later. No readable writing anywhere: surfaces that would carry writing may be present, but stage any wording out of legibility — an oblique angle, distance, shallow focus. No captions, watermarks or overlay text.",
  "bg_path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/db00517b-165b-4005-a287-f28936b59b95/images/272bd7af-9294-4490-98c6-88a40e9deacc/scene/recipe/S31sh1__bgfirst_bg.png",
  "asset_id": "b1a5f517-7294-4a62-bdb4-1e6a1c57553e",
  "input_asset_ids": [
   "34d329c4-6159-44b7-bf3f-98f6943cb851",
   "c54772cc-ea97-4590-93d2-3432743dcf57"
  ]
 },
 "S31sh1": {
  "input_fingerprint": "cd1f955083c0d02b",
  "prompt": "Create ONE FINAL photorealistic live-action film still of the moment below — early 20th-century Korea, Jeolla region; all people are Korean unless explicitly stated otherwise. TIME OF DAY (lock): afternoon.\n\nSHOT TEXT (authoritative, Korean): 거친 파도가 치는 해안가 거대한 바위 뒤로 한쪽 발을 딛고 뒷발로 모래를 차올리며 달려 들어오는 미드 액션, 앞을 향해 상체를 기울인 안규홍, 이관회, 염재보, 의병 포로1, 의병 포로2, 의병 포로3의 역동적인 전경.\n\nLOCATION (lock): The sandy refuge immediately behind large coastal rocks, beside rough breaking surf. The shot takes place here — the attached LOCATION PHOTOGRAPH shows the exact spot.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: wide shot\n- FRAME LAYOUT: Sheltered arrival space behind the coastal rock in the middle-right of the frame, midground.\n- KEY BACKGROUND ELEMENTS: 해안 바위 (Sheltering the arriving group) — The shelter-facing side appears at the right edge, with the arrival route visible beside it; used as Defines the destination without hiding the runners or overwhelming their scale; 모래 바닥 (Disturbed beneath the arriving feet); used as Separates planted and airborne feet and makes the staggered stride phases readable; 해안 파도 (Breaking behind the rocks); used as A limited background opening establishes the coast without competing with the arrival.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Use natural afternoon ambient light with restrained saturation and sufficient tonal openness to distinguish the overlapping bodies without polishing away their exhaustion.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): Rough surf breaks beside the coastal rocks and sand. 안규홍: He arrives without his carrying frame, with split, bleeding knuckles and the jade ring concealed. 이관회: He arrives breathless, retaining the sickle used in the rescue. 염재보: He arrives carrying a rescued prisoner piggyback, with his matchlock retained. 의병 포로1: He is freed from his binding rope and arrives exhausted. 의병 포로2: He is freed from his binding rope and arrives exhausted. 의병 포로3: He cannot walk unaided and arrives being carried.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 안규홍 (한국인 남성, 청년 나이대, 검은 머리, 단단히 묶은 머리); 이관회 (한국인 남성, 청년 나이대, 자연스러운 검은 머리); 염재보 (한국인 남성, 성인 나이대, 자연스러운 검은 머리); 의병 포로1 (한국인 남성, 성인 나이대, 자연스러운 검은 머리); 의병 포로2 (한국인 남성, 성인 나이대, 자연스러운 검은 머리); 의병 포로3 (한국인 남성, 성인 나이대, 자연스러운 검은 머리) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.",
  "roll_prompts": {
   "A": "Stage the shot. The FIRST attached image (SHOT BACKGROUND) is the finished empty background of this shot — keep it EXACTLY: its camera, perspective, architecture, lighting and every fixed feature stay untouched. The SECOND attached image (LAYOUT SKETCH) tells you ONLY where the people go: each sketched person's position, screen size, pose and the gaze/motion arrows. Ignore the sketch's background lines. The CHARACTER REFERENCE photographs show the real people.\nPlace the real people into the background at exactly the sketched positions, sizes and poses, following the arrow directions. No sketch lines or arrows may remain.\n\nCreate ONE FINAL photorealistic live-action film still of the moment below — early 20th-century Korea, Jeolla region; all people are Korean unless explicitly stated otherwise. TIME OF DAY (lock): afternoon.\n\nSHOT TEXT (authoritative, Korean): 거친 파도가 치는 해안가 거대한 바위 뒤로 한쪽 발을 딛고 뒷발로 모래를 차올리며 달려 들어오는 미드 액션, 앞을 향해 상체를 기울인 안규홍, 이관회, 염재보, 의병 포로1, 의병 포로2, 의병 포로3의 역동적인 전경.\n\nLOCATION (lock): The sandy refuge immediately behind large coastal rocks, beside rough breaking surf. The shot takes place here — the FIRST attached image (SHOT BACKGROUND) is this exact place, already built: its ground, structures, horizon, materials and lighting are the finished truth of this location and must not be redesigned or replaced. No location photograph is attached — read the place from that image alone, and add no scenery, structure, vehicle or fixture that it does not already show. This lock governs the place only; the figures in the shot follow the staging and pose instructions.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: wide shot\n- FRAME LAYOUT: Sheltered arrival space behind the coastal rock in the middle-right of the frame, midground.\n- KEY BACKGROUND ELEMENTS: 해안 바위 (Sheltering the arriving group) — The shelter-facing side appears at the right edge, with the arrival route visible beside it; used as Defines the destination without hiding the runners or overwhelming their scale; 모래 바닥 (Disturbed beneath the arriving feet); used as Separates planted and airborne feet and makes the staggered stride phases readable; 해안 파도 (Breaking behind the rocks); used as A limited background opening establishes the coast without competing with the arrival.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Use natural afternoon ambient light with restrained saturation and sufficient tonal openness to distinguish the overlapping bodies without polishing away their exhaustion.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): Rough surf breaks beside the coastal rocks and sand. 안규홍: He arrives without his carrying frame, with split, bleeding knuckles and the jade ring concealed. 이관회: He arrives breathless, retaining the sickle used in the rescue. 염재보: He arrives carrying a rescued prisoner piggyback, with his matchlock retained. 의병 포로1: He is freed from his binding rope and arrives exhausted. 의병 포로2: He is freed from his binding rope and arrives exhausted. 의병 포로3: He cannot walk unaided and arrives being carried.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 안규홍 (한국인 남성, 청년 나이대, 검은 머리, 단단히 묶은 머리); 이관회 (한국인 남성, 청년 나이대, 자연스러운 검은 머리); 염재보 (한국인 남성, 성인 나이대, 자연스러운 검은 머리); 의병 포로1 (한국인 남성, 성인 나이대, 자연스러운 검은 머리); 의병 포로2 (한국인 남성, 성인 나이대, 자연스러운 검은 머리); 의병 포로3 (한국인 남성, 성인 나이대, 자연스러운 검은 머리) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.",
   "B": "Create ONE FINAL photorealistic live-action film still of the moment below — early 20th-century Korea, Jeolla region; all people are Korean unless explicitly stated otherwise. TIME OF DAY (lock): afternoon.\n\nSHOT TEXT (authoritative, Korean): 거친 파도가 치는 해안가 거대한 바위 뒤로 한쪽 발을 딛고 뒷발로 모래를 차올리며 달려 들어오는 미드 액션, 앞을 향해 상체를 기울인 안규홍, 이관회, 염재보, 의병 포로1, 의병 포로2, 의병 포로3의 역동적인 전경.\n\nLOCATION (lock): The sandy refuge immediately behind large coastal rocks, beside rough breaking surf. The shot takes place here — the attached LOCATION PHOTOGRAPH shows the exact spot.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: wide shot\n- FRAME LAYOUT: Sheltered arrival space behind the coastal rock in the middle-right of the frame, midground.\n- KEY BACKGROUND ELEMENTS: 해안 바위 (Sheltering the arriving group) — The shelter-facing side appears at the right edge, with the arrival route visible beside it; used as Defines the destination without hiding the runners or overwhelming their scale; 모래 바닥 (Disturbed beneath the arriving feet); used as Separates planted and airborne feet and makes the staggered stride phases readable; 해안 파도 (Breaking behind the rocks); used as A limited background opening establishes the coast without competing with the arrival.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Use natural afternoon ambient light with restrained saturation and sufficient tonal openness to distinguish the overlapping bodies without polishing away their exhaustion.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): Rough surf breaks beside the coastal rocks and sand. 안규홍: He arrives without his carrying frame, with split, bleeding knuckles and the jade ring concealed. 이관회: He arrives breathless, retaining the sickle used in the rescue. 염재보: He arrives carrying a rescued prisoner piggyback, with his matchlock retained. 의병 포로1: He is freed from his binding rope and arrives exhausted. 의병 포로2: He is freed from his binding rope and arrives exhausted. 의병 포로3: He cannot walk unaided and arrives being carried.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 안규홍 (한국인 남성, 청년 나이대, 검은 머리, 단단히 묶은 머리); 이관회 (한국인 남성, 청년 나이대, 자연스러운 검은 머리); 염재보 (한국인 남성, 성인 나이대, 자연스러운 검은 머리); 의병 포로1 (한국인 남성, 성인 나이대, 자연스러운 검은 머리); 의병 포로2 (한국인 남성, 성인 나이대, 자연스러운 검은 머리); 의병 포로3 (한국인 남성, 성인 나이대, 자연스러운 검은 머리) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.\n\nCOMPOSITION VARIATION (this candidate only): keep the place, the people and the moment exactly as specified, and keep from the CAMERA & FRAME contract its subject scale, screen placement and key background elements exactly. For this candidate only, its camera angle and height are deliberately left open: choose a camera position distinctly different from the obvious one, as a film director picking a second setup on the same blocking. Stay true to any explicit framing the SHOT TEXT itself dictates."
  },
  "roll_refs": {
   "A": [
    {
     "label": "SHOT BACKGROUND",
     "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/db00517b-165b-4005-a287-f28936b59b95/images/272bd7af-9294-4490-98c6-88a40e9deacc/scene/recipe/S31sh1__bgfirst_bg.png",
     "asset_id": "b1a5f517-7294-4a62-bdb4-1e6a1c57553e",
     "role": "bgfirst_bg"
    },
    {
     "label": "LAYOUT SKETCH (people placement only)",
     "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/db00517b-165b-4005-a287-f28936b59b95/images/272bd7af-9294-4490-98c6-88a40e9deacc/conti/conti_S31sh1.png",
     "asset_id": "34d329c4-6159-44b7-bf3f-98f6943cb851",
     "role": "conti_light"
    },
    {
     "label": "CHARACTER REFERENCE — 안규홍: the exact person appearing in this shot; match face, hair and build exactly.",
     "path": "<bytes:1028384>",
     "asset_id": "3b5894db-ce55-4f34-b966-5e725c3aa952",
     "role": "character_ref"
    },
    {
     "label": "CHARACTER REFERENCE — 이관회: the exact person appearing in this shot; match face, hair and build exactly.",
     "path": "<bytes:933540>",
     "asset_id": "851b450f-ac84-47fe-ac5a-03a56f63fcae",
     "role": "character_ref"
    },
    {
     "label": "CHARACTER REFERENCE — 염재보: the exact person appearing in this shot; match face, hair and build exactly.",
     "path": "<bytes:964351>",
     "asset_id": "ab287289-d0b7-436d-8f70-68452f898f73",
     "role": "character_ref"
    },
    {
     "label": "CHARACTER REFERENCE — 의병 포로1: the exact person appearing in this shot; match face, hair and build exactly.",
     "path": "<bytes:910581>",
     "asset_id": "5a2bba8f-883f-4d9b-9003-b67ab053e656",
     "role": "character_ref"
    },
    {
     "label": "CHARACTER REFERENCE — 의병 포로2: the exact person appearing in this shot; match face, hair and build exactly.",
     "path": "<bytes:867759>",
     "asset_id": "e56e0b15-c7ee-4e69-8d88-8e0c77016d3c",
     "role": "character_ref"
    },
    {
     "label": "CHARACTER REFERENCE — 의병 포로3: the exact person appearing in this shot; match face, hair and build exactly.",
     "path": "<bytes:1497882>",
     "asset_id": "cb73bd5d-620b-48c3-9664-22790ab19cd2",
     "role": "character_ref"
    }
   ],
   "B": [
    {
     "label": "LOCATION PHOTOGRAPH — the exact place of this shot: its architecture, materials, fixed features and lighting mood are spatial truth; stage the moment inside this place. Never copy its camera framing.",
     "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/db00517b-165b-4005-a287-f28936b59b95/images/272bd7af-9294-4490-98c6-88a40e9deacc/scene/recipe/groupbg_coastal_rock_shelter_9ea29b.png",
     "asset_id": "c54772cc-ea97-4590-93d2-3432743dcf57",
     "role": "bgfirst_group_bg"
    },
    {
     "label": "CHARACTER REFERENCE — 안규홍: the exact person appearing in this shot; match face, hair and build exactly.",
     "path": "<bytes:1028384>",
     "asset_id": "3b5894db-ce55-4f34-b966-5e725c3aa952",
     "role": "character_ref"
    },
    {
     "label": "CHARACTER REFERENCE — 이관회: the exact person appearing in this shot; match face, hair and build exactly.",
     "path": "<bytes:933540>",
     "asset_id": "851b450f-ac84-47fe-ac5a-03a56f63fcae",
     "role": "character_ref"
    },
    {
     "label": "CHARACTER REFERENCE — 염재보: the exact person appearing in this shot; match face, hair and build exactly.",
     "path": "<bytes:964351>",
     "asset_id": "ab287289-d0b7-436d-8f70-68452f898f73",
     "role": "character_ref"
    },
    {
     "label": "CHARACTER REFERENCE — 의병 포로1: the exact person appearing in this shot; match face, hair and build exactly.",
     "path": "<bytes:910581>",
     "asset_id": "5a2bba8f-883f-4d9b-9003-b67ab053e656",
     "role": "character_ref"
    },
    {
     "label": "CHARACTER REFERENCE — 의병 포로2: the exact person appearing in this shot; match face, hair and build exactly.",
     "path": "<bytes:867759>",
     "asset_id": "e56e0b15-c7ee-4e69-8d88-8e0c77016d3c",
     "role": "character_ref"
    },
    {
     "label": "CHARACTER REFERENCE — 의병 포로3: the exact person appearing in this shot; match face, hair and build exactly.",
     "path": "<bytes:1497882>",
     "asset_id": "cb73bd5d-620b-48c3-9664-22790ab19cd2",
     "role": "character_ref"
    }
   ]
  },
  "shot_run_spend_attempt_count": 1,
  "judge_flip_skipped": "judge_fn_owns_order",
  "readings": [
   {
    "label": "A",
    "direction": "인물들이 정면이 아닌 화면 우측의 바위벽을 향해 측면으로 나란히 달리고 있습니다.",
    "built_space": "좌측에 파도치는 바다, 우측에 거대한 해안 바위가 위치한 모래사장 배경이 적절히 배치되어 있습니다.",
    "entities": "안규홍, 낫을 든 이관회, 포로를 업고 화승총을 든 염재보는 확인되나, 중간에 달리는 두 인물은 레퍼런스가 요구한 의병 포로 1(가슴이 찢어진 옷)과 의병 포로 2(맨가슴과 피 묻은 바지)의 외형을 전혀 따르지 않았습니다.",
    "hard_violations": [
     "[gemini-pro] 지정된 인물 외형 누락 및 임의의 인물 생성 (의병 포로 1, 의병 포로 2의 특징이 없음)"
    ],
    "physics": "달리며 모래를 차올리는 동작과 발이 땅에 닿아 지탱하는 모습 등은 물리적으로 자연스럽습니다."
   },
   {
    "label": "B",
    "direction": "지문에 명시된 대로 6명의 인물 모두가 카메라가 있는 앞(정면)을 향해 상체를 기울인 채 달려 들어오고 있습니다.",
    "built_space": "우측 프레임 가장자리에 거대한 해안 바위가 은신처의 경계를 형성하고, 뒤쪽으로 파도치는 바다가 보여 요구된 구도와 훌륭하게 일치합니다.",
    "entities": "피 흘리는 주먹의 안규홍, 낫을 쥔 이관회, 포로를 업고 화승총을 든 염재보, 찢어진 가슴 옷의 의병 포로 1, 앞섶을 풀어 헤친 맨가슴의 의병 포로 2, 업혀 있는 의병 포로 3 등 모든 인물이 레퍼런스 및 지문과 정확히 일치합니다.",
    "hard_violations": [],
    "physics": "뛰어오며 뒷발로 모래를 강하게 차올리는 역동적인 뜀박질의 무게 중심과, 포로를 업은 채 달리는 염재보의 신체 지탱이 매우 자연스럽게 묘사되었습니다."
   }
  ],
  "cross_model_order": {
   "policy": "cross_model_order_v2_slot0_fwd_slot1_rev_parallel2",
   "slots": [
    {
     "model": "gemini-pro",
     "order": "forward",
     "display_to_canonical": {
      "A": "A",
      "B": "B"
     },
     "raw": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 10,
        "verdict_ko": "모든 등장인물의 세부적인 외형(피 묻은 주먹, 찢어진 옷, 맨가슴 등)과 소지품, 업힌 상태를 완벽하게 반영하며 정면으로 달려오는 역동적인 구도를 훌륭하게 구현했습니다."
       },
       {
        "label": "A",
        "score": 3,
        "verdict_ko": "정면이 아닌 우측 바위를 향해 달리고 있으며, 의병 포로 1, 2의 지정된 복장 및 외형적 특징이 전혀 반영되지 않은 임의의 인물이 생성되었습니다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "인물들이 정면이 아닌 화면 우측의 바위벽을 향해 측면으로 나란히 달리고 있습니다.",
        "built_space": "좌측에 파도치는 바다, 우측에 거대한 해안 바위가 위치한 모래사장 배경이 적절히 배치되어 있습니다.",
        "entities": "안규홍, 낫을 든 이관회, 포로를 업고 화승총을 든 염재보는 확인되나, 중간에 달리는 두 인물은 레퍼런스가 요구한 의병 포로 1(가슴이 찢어진 옷)과 의병 포로 2(맨가슴과 피 묻은 바지)의 외형을 전혀 따르지 않았습니다.",
        "hard_violations": [
         "지정된 인물 외형 누락 및 임의의 인물 생성 (의병 포로 1, 의병 포로 2의 특징이 없음)"
        ],
        "physics": "달리며 모래를 차올리는 동작과 발이 땅에 닿아 지탱하는 모습 등은 물리적으로 자연스럽습니다."
       },
       {
        "label": "B",
        "direction": "지문에 명시된 대로 6명의 인물 모두가 카메라가 있는 앞(정면)을 향해 상체를 기울인 채 달려 들어오고 있습니다.",
        "built_space": "우측 프레임 가장자리에 거대한 해안 바위가 은신처의 경계를 형성하고, 뒤쪽으로 파도치는 바다가 보여 요구된 구도와 훌륭하게 일치합니다.",
        "entities": "피 흘리는 주먹의 안규홍, 낫을 쥔 이관회, 포로를 업고 화승총을 든 염재보, 찢어진 가슴 옷의 의병 포로 1, 앞섶을 풀어 헤친 맨가슴의 의병 포로 2, 업혀 있는 의병 포로 3 등 모든 인물이 레퍼런스 및 지문과 정확히 일치합니다.",
        "hard_violations": [],
        "physics": "뛰어오며 뒷발로 모래를 강하게 차올리는 역동적인 뜀박질의 무게 중심과, 포로를 업은 채 달리는 염재보의 신체 지탱이 매우 자연스럽게 묘사되었습니다."
       }
      ],
      "all_candidates_fail": false
     },
     "normalized": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 10,
        "verdict_ko": "모든 등장인물의 세부적인 외형(피 묻은 주먹, 찢어진 옷, 맨가슴 등)과 소지품, 업힌 상태를 완벽하게 반영하며 정면으로 달려오는 역동적인 구도를 훌륭하게 구현했습니다."
       },
       {
        "label": "A",
        "score": 3,
        "verdict_ko": "정면이 아닌 우측 바위를 향해 달리고 있으며, 의병 포로 1, 2의 지정된 복장 및 외형적 특징이 전혀 반영되지 않은 임의의 인물이 생성되었습니다."
       }
      ],
      "all_candidates_fail": false,
      "readings": [
       {
        "label": "A",
        "direction": "인물들이 정면이 아닌 화면 우측의 바위벽을 향해 측면으로 나란히 달리고 있습니다.",
        "built_space": "좌측에 파도치는 바다, 우측에 거대한 해안 바위가 위치한 모래사장 배경이 적절히 배치되어 있습니다.",
        "entities": "안규홍, 낫을 든 이관회, 포로를 업고 화승총을 든 염재보는 확인되나, 중간에 달리는 두 인물은 레퍼런스가 요구한 의병 포로 1(가슴이 찢어진 옷)과 의병 포로 2(맨가슴과 피 묻은 바지)의 외형을 전혀 따르지 않았습니다.",
        "hard_violations": [
         "지정된 인물 외형 누락 및 임의의 인물 생성 (의병 포로 1, 의병 포로 2의 특징이 없음)"
        ],
        "physics": "달리며 모래를 차올리는 동작과 발이 땅에 닿아 지탱하는 모습 등은 물리적으로 자연스럽습니다."
       },
       {
        "label": "B",
        "direction": "지문에 명시된 대로 6명의 인물 모두가 카메라가 있는 앞(정면)을 향해 상체를 기울인 채 달려 들어오고 있습니다.",
        "built_space": "우측 프레임 가장자리에 거대한 해안 바위가 은신처의 경계를 형성하고, 뒤쪽으로 파도치는 바다가 보여 요구된 구도와 훌륭하게 일치합니다.",
        "entities": "피 흘리는 주먹의 안규홍, 낫을 쥔 이관회, 포로를 업고 화승총을 든 염재보, 찢어진 가슴 옷의 의병 포로 1, 앞섶을 풀어 헤친 맨가슴의 의병 포로 2, 업혀 있는 의병 포로 3 등 모든 인물이 레퍼런스 및 지문과 정확히 일치합니다.",
        "hard_violations": [],
        "physics": "뛰어오며 뒷발로 모래를 강하게 차올리는 역동적인 뜀박질의 무게 중심과, 포로를 업은 채 달리는 염재보의 신체 지탱이 매우 자연스럽게 묘사되었습니다."
       }
      ]
     },
     "ok": true
    },
    {
     "model": "gpt-high",
     "order": "reverse",
     "display_to_canonical": {
      "A": "B",
      "B": "A"
     },
     "raw": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 6,
        "verdict_ko": "여섯 인물과 업힌 포로·낫·장총은 구현했지만, 카메라 정면으로 달려오는 큰 인물 배치와 넓은 파도 배경이 중경 오른쪽 바위 뒤로 진입하는 지정 구도를 약화한다."
       },
       {
        "label": "B",
        "score": 8,
        "verdict_ko": "오른쪽 바위 곁으로 달려드는 와이드 구도와 전방으로 기울인 상체·엇갈린 보폭·모래 차올림이 더 정확하지만, 포로들의 구별되는 복식과 머리 장식은 덜 충실하다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "안규홍은 화면 중앙에서 카메라 쪽으로 달리며 정면을 본다. 이관회와 두 보행 포로도 대체로 카메라 앞쪽 모래밭을 향하고, 염재보와 업힌 포로는 앞쪽 아래를 본다. 오른쪽 바위 뒤의 도착 공간으로 방향을 꺾는 동작은 드러나지 않는다. 이관회의 낫날은 아래로, 염재보 앞의 장총 총구는 화면 위쪽으로 향하며 누군가를 겨누지는 않는다.",
        "built_space": "인공 구조물이나 고정 설비는 없다. 오른쪽 가장자리의 큰 암벽, 발밑 모래밭, 뒤쪽 암초와 소나무는 참조 해안의 재료와 지형을 따른다. 다만 파도가 왼쪽과 중앙 배경을 크게 차지하고, 인물들이 중앙 전경까지 나와 있어 중경 오른쪽의 피신 도착 공간보다 정면 달리기가 강조된다.",
        "entities": "성인 또는 청년의 한국인 남성으로 읽히는 여섯 명이 있으며 추가 인물은 없다. 안규홍의 묶은 검은 머리·밝은 옷·피 묻은 손은 보이고 지게와 노출된 옥반지는 없다. 이관회는 갈색 저고리와 밝은 바지를 입고 낫을 쥔다. 염재보는 어두운 옷과 각반을 착용하고 장총을 지닌 채 상처 입고 머리를 동여맨 포로3을 업는다. 포로1의 가슴이 찢어진 밝은 옷, 포로2의 머리띠·벌어진 저고리·손상된 무릎은 참조와 대응한다. 염재보와 포로1의 참조 모자는 없고, 안규홍의 저고리는 참조보다 크게 벌어져 있다. 포로들의 결박과 읽을 수 있는 글자는 보이지 않는다.",
        "hard_violations": [],
        "physics": "이관회와 뒤쪽 보행 포로들은 모래에 닿은 지지발과 들린 반대발이 구별된다. 안규홍의 두 발은 지면에서 잠깐 떨어진 달리기 체공 단계로 보이며, 엇갈린 다리와 바로 아래 착지면이 있어 근거 없는 부유는 아니다. 다만 상체 기울기와 모래를 뒤로 차는 힘은 비교적 약하다. 포로3은 염재보의 등에 몸을 기대고 어깨 둘레로 팔을 걸며 다리를 양옆으로 늘어뜨린다. 염재보의 몸과 팔이 그를 받치고, 장총은 몸 앞에서 손과 팔에 걸쳐 지지된다. 낫은 이관회의 손에 잡혀 있다."
       },
       {
        "label": "B",
        "direction": "안규홍을 선두로 네 보행 인물과 포로를 업은 염재보가 화면 왼쪽에서 오른쪽 암벽 곁 모래 공간으로 달린다. 보행 인물들의 얼굴과 시선도 오른쪽 진입 방향을 향하며, 업힌 포로는 아래로 고개를 떨어뜨린다. 바위 뒤 최종 진입로는 완전히 드러나지 않지만 목적지 방향은 A보다 분명하다. 이관회의 낫은 손 앞에서 날이 위로 굽어 있고, 염재보의 총구는 오른쪽 아래 모래 쪽을 향한다.",
        "built_space": "인공 구조물이나 고정 설비는 없다. 오른쪽 가장자리의 큰 암벽과 밑동 바위, 중앙의 모래 이동로, 왼쪽 바다와 뒤편 암초·소나무 사면이 참조 장소에 더 가깝게 연결된다. 인물들은 전신이 보이는 와이드 숏의 중경에 배치되고 선두는 중경 오른쪽 바위 곁에 도착한다. 바다도 보이지만 A처럼 큰 쇄파가 인물 뒤를 압도하지 않는다.",
        "entities": "총 여섯 명의 한국인 청년·성인 남성으로 읽히며, 다섯 명이 달리고 한 명이 업혀 있다. 선두 안규홍의 묶은 머리·밝은 옷·붉게 상처 난 주먹, 이관회의 갈색 저고리와 손에 든 낫, 염재보의 어두운 옷·각반·장총은 역할과 맞는다. 포로3은 머리 붕대와 해진 옷을 입고 업혀 있다. 가운데 두 보행 포로는 모두 비슷한 회갈색 옷으로 보여 포로1의 가슴 찢김과 포로2의 머리띠·노출된 가슴·피 묻은 무릎이라는 참조상 차이가 잘 재현되지 않았다. 염재보와 포로1의 모자도 없다. 지게·노출된 옥반지·포로 결박·읽을 수 있는 글자는 보이지 않는다.",
        "hard_violations": [],
        "physics": "달리는 인물들은 상체를 진행 방향으로 기울이고 한쪽 다리를 뒤로 뻗거나 접는다. 여러 지지발이 모래를 누르고 그 뒤로 모래가 솟아, 추진과 다음 착지가 자연스럽게 연결된다. 염재보의 짧은 체공도 벌어진 보폭과 바로 아래 착지 가능한 모래밭으로 설명된다. 업힌 포로는 염재보의 등에 밀착하고 팔을 어깨 앞으로 늘어뜨리며, 염재보가 뒤로 뻗은 손으로 다리를 받친다. 장총은 앞쪽 손과 팔에 지지되고 낫도 손잡이를 쥔 손에 연결되어 있다. 지지나 달리기 동작으로 설명되지 않는 부유는 보이지 않는다."
       }
      ],
      "all_candidates_fail": false
     },
     "normalized": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 6,
        "verdict_ko": "여섯 인물과 업힌 포로·낫·장총은 구현했지만, 카메라 정면으로 달려오는 큰 인물 배치와 넓은 파도 배경이 중경 오른쪽 바위 뒤로 진입하는 지정 구도를 약화한다."
       },
       {
        "label": "A",
        "score": 8,
        "verdict_ko": "오른쪽 바위 곁으로 달려드는 와이드 구도와 전방으로 기울인 상체·엇갈린 보폭·모래 차올림이 더 정확하지만, 포로들의 구별되는 복식과 머리 장식은 덜 충실하다."
       }
      ],
      "all_candidates_fail": false,
      "readings": [
       {
        "label": "B",
        "direction": "안규홍은 화면 중앙에서 카메라 쪽으로 달리며 정면을 본다. 이관회와 두 보행 포로도 대체로 카메라 앞쪽 모래밭을 향하고, 염재보와 업힌 포로는 앞쪽 아래를 본다. 오른쪽 바위 뒤의 도착 공간으로 방향을 꺾는 동작은 드러나지 않는다. 이관회의 낫날은 아래로, 염재보 앞의 장총 총구는 화면 위쪽으로 향하며 누군가를 겨누지는 않는다.",
        "built_space": "인공 구조물이나 고정 설비는 없다. 오른쪽 가장자리의 큰 암벽, 발밑 모래밭, 뒤쪽 암초와 소나무는 참조 해안의 재료와 지형을 따른다. 다만 파도가 왼쪽과 중앙 배경을 크게 차지하고, 인물들이 중앙 전경까지 나와 있어 중경 오른쪽의 피신 도착 공간보다 정면 달리기가 강조된다.",
        "entities": "성인 또는 청년의 한국인 남성으로 읽히는 여섯 명이 있으며 추가 인물은 없다. 안규홍의 묶은 검은 머리·밝은 옷·피 묻은 손은 보이고 지게와 노출된 옥반지는 없다. 이관회는 갈색 저고리와 밝은 바지를 입고 낫을 쥔다. 염재보는 어두운 옷과 각반을 착용하고 장총을 지닌 채 상처 입고 머리를 동여맨 포로3을 업는다. 포로1의 가슴이 찢어진 밝은 옷, 포로2의 머리띠·벌어진 저고리·손상된 무릎은 참조와 대응한다. 염재보와 포로1의 참조 모자는 없고, 안규홍의 저고리는 참조보다 크게 벌어져 있다. 포로들의 결박과 읽을 수 있는 글자는 보이지 않는다.",
        "hard_violations": [],
        "physics": "이관회와 뒤쪽 보행 포로들은 모래에 닿은 지지발과 들린 반대발이 구별된다. 안규홍의 두 발은 지면에서 잠깐 떨어진 달리기 체공 단계로 보이며, 엇갈린 다리와 바로 아래 착지면이 있어 근거 없는 부유는 아니다. 다만 상체 기울기와 모래를 뒤로 차는 힘은 비교적 약하다. 포로3은 염재보의 등에 몸을 기대고 어깨 둘레로 팔을 걸며 다리를 양옆으로 늘어뜨린다. 염재보의 몸과 팔이 그를 받치고, 장총은 몸 앞에서 손과 팔에 걸쳐 지지된다. 낫은 이관회의 손에 잡혀 있다."
       },
       {
        "label": "A",
        "direction": "안규홍을 선두로 네 보행 인물과 포로를 업은 염재보가 화면 왼쪽에서 오른쪽 암벽 곁 모래 공간으로 달린다. 보행 인물들의 얼굴과 시선도 오른쪽 진입 방향을 향하며, 업힌 포로는 아래로 고개를 떨어뜨린다. 바위 뒤 최종 진입로는 완전히 드러나지 않지만 목적지 방향은 A보다 분명하다. 이관회의 낫은 손 앞에서 날이 위로 굽어 있고, 염재보의 총구는 오른쪽 아래 모래 쪽을 향한다.",
        "built_space": "인공 구조물이나 고정 설비는 없다. 오른쪽 가장자리의 큰 암벽과 밑동 바위, 중앙의 모래 이동로, 왼쪽 바다와 뒤편 암초·소나무 사면이 참조 장소에 더 가깝게 연결된다. 인물들은 전신이 보이는 와이드 숏의 중경에 배치되고 선두는 중경 오른쪽 바위 곁에 도착한다. 바다도 보이지만 A처럼 큰 쇄파가 인물 뒤를 압도하지 않는다.",
        "entities": "총 여섯 명의 한국인 청년·성인 남성으로 읽히며, 다섯 명이 달리고 한 명이 업혀 있다. 선두 안규홍의 묶은 머리·밝은 옷·붉게 상처 난 주먹, 이관회의 갈색 저고리와 손에 든 낫, 염재보의 어두운 옷·각반·장총은 역할과 맞는다. 포로3은 머리 붕대와 해진 옷을 입고 업혀 있다. 가운데 두 보행 포로는 모두 비슷한 회갈색 옷으로 보여 포로1의 가슴 찢김과 포로2의 머리띠·노출된 가슴·피 묻은 무릎이라는 참조상 차이가 잘 재현되지 않았다. 염재보와 포로1의 모자도 없다. 지게·노출된 옥반지·포로 결박·읽을 수 있는 글자는 보이지 않는다.",
        "hard_violations": [],
        "physics": "달리는 인물들은 상체를 진행 방향으로 기울이고 한쪽 다리를 뒤로 뻗거나 접는다. 여러 지지발이 모래를 누르고 그 뒤로 모래가 솟아, 추진과 다음 착지가 자연스럽게 연결된다. 염재보의 짧은 체공도 벌어진 보폭과 바로 아래 착지 가능한 모래밭으로 설명된다. 업힌 포로는 염재보의 등에 밀착하고 팔을 어깨 앞으로 늘어뜨리며, 염재보가 뒤로 뻗은 손으로 다리를 받친다. 장총은 앞쪽 손과 팔에 지지되고 낫도 손잡이를 쥔 손에 연결되어 있다. 지지나 달리기 동작으로 설명되지 않는 부유는 보이지 않는다."
       }
      ]
     },
     "ok": true
    }
   ],
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "slot_winner_match": false,
   "slot_winner": {
    "gemini-pro": "B",
    "gpt-high": "A"
   },
   "route": "cross_slot_combined"
  },
  "dual": {
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "normalized": {
    "A": 1.3,
    "B": 1.75
   },
   "adjusted": {
    "A": 1.05,
    "B": 1.75
   },
   "violations": {
    "A": [
     "[gemini-pro] 지정된 인물 외형 누락 및 임의의 인물 생성 (의병 포로 1, 의병 포로 2의 특징이 없음)"
    ]
   },
   "per_model_winner": {
    "gemini-pro": "B",
    "gpt-high": "A"
   },
   "agreed": false
  },
  "totals": {
   "B": 1750,
   "A": 1050
  },
  "selected": "B",
  "ranking": [
   "B",
   "A"
  ],
  "verdicts": [
   {
    "label": "B",
    "score": 1750,
    "verdict_ko": "모든 등장인물의 세부적인 외형(피 묻은 주먹, 찢어진 옷, 맨가슴 등)과 소지품, 업힌 상태를 완벽하게 반영하며 정면으로 달려오는 역동적인 구도를 훌륭하게 구현했습니다."
   },
   {
    "label": "A",
    "score": 1050,
    "verdict_ko": "정면이 아닌 우측 바위를 향해 달리고 있으며, 의병 포로 1, 2의 지정된 복장 및 외형적 특징이 전혀 반영되지 않은 임의의 인물이 생성되었습니다.  ★위반: [gemini-pro] 지정된 인물 외형 누락 및 임의의 인물 생성 (의병 포로 1, 의병 포로 2의 특징이 없음)"
   }
  ],
  "refs": [
   {
    "label": "LOCATION PHOTOGRAPH — the exact place of this shot: its architecture, materials, fixed features and lighting mood are spatial truth; stage the moment inside this place. Never copy its camera framing.",
    "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/db00517b-165b-4005-a287-f28936b59b95/images/272bd7af-9294-4490-98c6-88a40e9deacc/scene/recipe/groupbg_coastal_rock_shelter_9ea29b.png",
    "asset_id": "c54772cc-ea97-4590-93d2-3432743dcf57",
    "role": "bgfirst_group_bg"
   },
   {
    "label": "CHARACTER REFERENCE — 안규홍: the exact person appearing in this shot; match face, hair and build exactly.",
    "path": "<bytes:1028384>",
    "asset_id": "3b5894db-ce55-4f34-b966-5e725c3aa952",
    "role": "character_ref"
   },
   {
    "label": "CHARACTER REFERENCE — 이관회: the exact person appearing in this shot; match face, hair and build exactly.",
    "path": "<bytes:933540>",
    "asset_id": "851b450f-ac84-47fe-ac5a-03a56f63fcae",
    "role": "character_ref"
   },
   {
    "label": "CHARACTER REFERENCE — 염재보: the exact person appearing in this shot; match face, hair and build exactly.",
    "path": "<bytes:964351>",
    "asset_id": "ab287289-d0b7-436d-8f70-68452f898f73",
    "role": "character_ref"
   },
   {
    "label": "CHARACTER REFERENCE — 의병 포로1: the exact person appearing in this shot; match face, hair and build exactly.",
    "path": "<bytes:910581>",
    "asset_id": "5a2bba8f-883f-4d9b-9003-b67ab053e656",
    "role": "character_ref"
   },
   {
    "label": "CHARACTER REFERENCE — 의병 포로2: the exact person appearing in this shot; match face, hair and build exactly.",
    "path": "<bytes:867759>",
    "asset_id": "e56e0b15-c7ee-4e69-8d88-8e0c77016d3c",
    "role": "character_ref"
   },
   {
    "label": "CHARACTER REFERENCE — 의병 포로3: the exact person appearing in this shot; match face, hair and build exactly.",
    "path": "<bytes:1497882>",
    "asset_id": "cb73bd5d-620b-48c3-9664-22790ab19cd2",
    "role": "character_ref"
   }
  ],
  "critique_skipped": true,
  "shot_run_produced": false,
  "shot_run_uid": "06aa1425-16db-7ec2-a6f1-275b820570b5",
  "bgfirst": {
   "bg_path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/db00517b-165b-4005-a287-f28936b59b95/images/272bd7af-9294-4490-98c6-88a40e9deacc/scene/recipe/S31sh1__bgfirst_bg.png",
   "bg_asset_id": "b1a5f517-7294-4a62-bdb4-1e6a1c57553e",
   "bg_record_key": "S31sh1::bgfirst_bg",
   "chain_winner": false,
   "authority": "groupbg",
   "group_key": "coastal_rock_shelter",
   "groupbg_asset_id": "c54772cc-ea97-4590-93d2-3432743dcf57"
  },
  "ref_mode": "그룹 배경+엔티티 (2택1: 무콘티 승)",
  "share_plan": {
   "ref_plan": "background"
  }
 },
 "S31sh1::cine": {
  "applied": false,
  "attempted_at": "2026-09-09T09:07:22.742685+00:00",
  "fingerprint": "69d229e4b1596f12df4d1f17344c389229bc95352ded8f2979b0804e7dc5209b",
  "fingerprint_version": 2,
  "provider": "mai",
  "endpoint": "openrouter/chat-completions",
  "model": "microsoft/mai-image-2.6",
  "pack": "24.202608252115",
  "source_file": "S31sh1_sel.png",
  "source_sha256": "274b282fdae72f3fb934348c88c15d5ac0444519d30bac93df714691fbea0588",
  "error": "RuntimeError: moderation blocked: {\"message\": \"Input content violated mainline safety policies.\", \"code\": 400}",
  "moderation_refusals": 2,
  "declined": true,
  "declined_reason": "moderation"
 },
 "S31sh12::signage": {
  "fp": "9a973ab73eee659f",
  "inscriptions": [],
  "cues": [],
  "dropped": []
 },
 "S31sh12": {
  "input_fingerprint": "7b4e8fc98e9ad151",
  "prompt": "Create ONE FINAL photorealistic live-action film still of the moment below — early 20th-century Korea, Jeolla region; all people are Korean unless explicitly stated otherwise. TIME OF DAY (lock): afternoon.\n\nSHOT TEXT (authoritative, Korean): 상처에 손을 얹은 채 고개를 들어 안규홍을 향해 또렷한 시선을 고정한 의병4의 굳은 정면.\n\nLOCATION (lock): The sheltered sand behind the coastal boulders, where the rescued prisoners' wounds are being tended. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: medium shot\n- KEY BACKGROUND ELEMENTS: 바위 뒤 모래 바닥 (The rescued group has settled here to recover and attend to wounds); used as Visible around the attending hand and seated prisoner, grounding the exchange in ongoing care.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Maintain natural afternoon illumination with restrained saturation and gentle tonal separation, allowing human warmth to emerge without a sentimental lighting change.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The people visible in that photo are NOT in this shot: never carry their faces, bodies or clothing onto anyone here. Each person in THIS shot is defined solely by the PEOPLE list, their own CHARACTER REFERENCE images and their wardrobe notes. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nPREVIOUS STILL USAGE (follow exactly — what to take from the attached still and what to exclude; it governs place and objects only, never who is in this shot or how the camera sees them): Take the large coastal rocks, sand, sea colors, and afternoon light. Exclude sand thrown up during the arrival and do not freeze an earlier breaking wave in place.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): Rough surf continues behind the rocks; torn cloth is being used for improvised dressings. 의병4: He is crouched in a wound-checking posture.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 의병4 (한국인 남성, 성인 나이대, 자연스러운 검은 머리) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.",
  "roll_prompts": {
   "A": "Create ONE FINAL photorealistic live-action film still of the moment below — early 20th-century Korea, Jeolla region; all people are Korean unless explicitly stated otherwise. TIME OF DAY (lock): afternoon.\n\nSHOT TEXT (authoritative, Korean): 상처에 손을 얹은 채 고개를 들어 안규홍을 향해 또렷한 시선을 고정한 의병4의 굳은 정면.\n\nLOCATION (lock): The sheltered sand behind the coastal boulders, where the rescued prisoners' wounds are being tended. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: medium shot\n- KEY BACKGROUND ELEMENTS: 바위 뒤 모래 바닥 (The rescued group has settled here to recover and attend to wounds); used as Visible around the attending hand and seated prisoner, grounding the exchange in ongoing care.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Maintain natural afternoon illumination with restrained saturation and gentle tonal separation, allowing human warmth to emerge without a sentimental lighting change.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The people visible in that photo are NOT in this shot: never carry their faces, bodies or clothing onto anyone here. Each person in THIS shot is defined solely by the PEOPLE list, their own CHARACTER REFERENCE images and their wardrobe notes. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nPREVIOUS STILL USAGE (follow exactly — what to take from the attached still and what to exclude; it governs place and objects only, never who is in this shot or how the camera sees them): Take the large coastal rocks, sand, sea colors, and afternoon light. Exclude sand thrown up during the arrival and do not freeze an earlier breaking wave in place.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): Rough surf continues behind the rocks; torn cloth is being used for improvised dressings. 의병4: He is crouched in a wound-checking posture.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 의병4 (한국인 남성, 성인 나이대, 자연스러운 검은 머리) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.",
   "B": "Create ONE FINAL photorealistic live-action film still of the moment below — early 20th-century Korea, Jeolla region; all people are Korean unless explicitly stated otherwise. TIME OF DAY (lock): afternoon.\n\nSHOT TEXT (authoritative, Korean): 상처에 손을 얹은 채 고개를 들어 안규홍을 향해 또렷한 시선을 고정한 의병4의 굳은 정면.\n\nLOCATION (lock): The sheltered sand behind the coastal boulders, where the rescued prisoners' wounds are being tended. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: medium shot\n- KEY BACKGROUND ELEMENTS: 바위 뒤 모래 바닥 (The rescued group has settled here to recover and attend to wounds); used as Visible around the attending hand and seated prisoner, grounding the exchange in ongoing care.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Maintain natural afternoon illumination with restrained saturation and gentle tonal separation, allowing human warmth to emerge without a sentimental lighting change.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The people visible in that photo are NOT in this shot: never carry their faces, bodies or clothing onto anyone here. Each person in THIS shot is defined solely by the PEOPLE list, their own CHARACTER REFERENCE images and their wardrobe notes. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nPREVIOUS STILL USAGE (follow exactly — what to take from the attached still and what to exclude; it governs place and objects only, never who is in this shot or how the camera sees them): Take the large coastal rocks, sand, sea colors, and afternoon light. Exclude sand thrown up during the arrival and do not freeze an earlier breaking wave in place.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): Rough surf continues behind the rocks; torn cloth is being used for improvised dressings. 의병4: He is crouched in a wound-checking posture.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 의병4 (한국인 남성, 성인 나이대, 자연스러운 검은 머리) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.\n\nCOMPOSITION VARIATION (this candidate only): keep the place, the people and the moment exactly as specified, and keep from the CAMERA & FRAME contract its subject scale, screen placement and key background elements exactly. For this candidate only, its camera angle and height are deliberately left open: choose a camera position distinctly different from the obvious one, as a film director picking a second setup on the same blocking. Stay true to any explicit framing the SHOT TEXT itself dictates."
  },
  "shot_run_spend_attempt_count": 1,
  "initial_roll_all_fail": true,
  "readings": [
   {
    "label": "A",
    "direction": "중앙의 인물이 상처에 두 손을 얹은 채 카메라 정면을 또렷하게 응시하고 있음(안규홍이 렌즈 너머에 있다고 설정된 시선).",
    "built_space": "해안가 바위 사이의 모래 바닥이며, 뒤편으로 파도치는 바다가 보임.",
    "entities": "레퍼런스와 일치하는 의병4(두건 및 복장)가 중앙에 쪼그려 있으나, 프롬프트 지시와 달리 배경에 5명의 추가 인물들이 상처를 치료하는 모습으로 등장함.",
    "hard_violations": [
     "[gemini-pro] 지정되지 않은 추가 인물 다수 등장 (배경의 5명)",
     "[gpt-high] 의병4 외에 왼쪽 세 명과 오른쪽 두 명, 총 다섯 명의 금지된 추가 인물을 배치했다.",
     "[gpt-high] 이전 장면에서 제외하도록 지정한 인물들의 외형과 복장을 배경에 재사용했다."
    ],
    "physics": "인물들이 모래 바닥에 체중을 싣고 안정적으로 앉거나 쪼그려 있음."
   },
   {
    "label": "B",
    "direction": "중앙의 인물이 상처에 손을 얹고 화면 우측 전경에 있는 다른 인물을 향해 시선을 던지고 있음.",
    "built_space": "해안가 거대한 바위 밑 모래 바닥이며, 바위가 인물들 뒤로 넓게 자리 잡고 있음.",
    "entities": "중심 인물이 의병4의 레퍼런스(두건 착용 등)와 일치하지 않고 이전 샷의 다른 인물과 유사함. 또한 화면 전체에 걸쳐 프롬프트에 허용되지 않은 인물 3명이 추가로 등장함.",
    "hard_violations": [
     "[gemini-pro] 지정되지 않은 추가 인물 등장 (화면 안의 다른 3명)",
     "[gemini-pro] 캐릭터 외형 불일치 및 이전 샷 인물 차용",
     "[gpt-high] 의병4만 허용된 화면에 전경 남성 한 명과 배경 남성 세 명을 추가했다.",
     "[gpt-high] 이전 장면 인물의 얼굴과 의상을 가져오지 말라는 명시적 금지에도 중앙 인물과 배경 인물들에 그 외형을 재사용했다."
    ],
    "physics": "인물들이 모래 바닥에 쪼그려 앉아있거나 바위에 기대어 있음."
   }
  ],
  "cross_model_order": {
   "policy": "cross_model_order_v2_slot0_fwd_slot1_rev_parallel2",
   "slots": [
    {
     "model": "gemini-pro",
     "order": "forward",
     "display_to_canonical": {
      "A": "A",
      "B": "B"
     },
     "raw": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 0,
        "verdict_ko": "의병4의 외형과 정면 구도는 훌륭하게 구현했으나, 프롬프트가 엄격히 금지한 추가 인물들(배경의 5명)이 묘사되어 치명적 규정 위반입니다."
       },
       {
        "label": "B",
        "score": 0,
        "verdict_ko": "프롬프트가 금지한 추가 인물들이 다수 등장할 뿐만 아니라, 중심 인물 역시 지정된 의병4가 아닌 이전 컷의 인물로 그려져 치명적 규정 위반입니다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "중앙의 인물이 상처에 두 손을 얹은 채 카메라 정면을 또렷하게 응시하고 있음(안규홍이 렌즈 너머에 있다고 설정된 시선).",
        "built_space": "해안가 바위 사이의 모래 바닥이며, 뒤편으로 파도치는 바다가 보임.",
        "entities": "레퍼런스와 일치하는 의병4(두건 및 복장)가 중앙에 쪼그려 있으나, 프롬프트 지시와 달리 배경에 5명의 추가 인물들이 상처를 치료하는 모습으로 등장함.",
        "hard_violations": [
         "지정되지 않은 추가 인물 다수 등장 (배경의 5명)"
        ],
        "physics": "인물들이 모래 바닥에 체중을 싣고 안정적으로 앉거나 쪼그려 있음."
       },
       {
        "label": "B",
        "direction": "중앙의 인물이 상처에 손을 얹고 화면 우측 전경에 있는 다른 인물을 향해 시선을 던지고 있음.",
        "built_space": "해안가 거대한 바위 밑 모래 바닥이며, 바위가 인물들 뒤로 넓게 자리 잡고 있음.",
        "entities": "중심 인물이 의병4의 레퍼런스(두건 착용 등)와 일치하지 않고 이전 샷의 다른 인물과 유사함. 또한 화면 전체에 걸쳐 프롬프트에 허용되지 않은 인물 3명이 추가로 등장함.",
        "hard_violations": [
         "지정되지 않은 추가 인물 등장 (화면 안의 다른 3명)",
         "캐릭터 외형 불일치 및 이전 샷 인물 차용"
        ],
        "physics": "인물들이 모래 바닥에 쪼그려 앉아있거나 바위에 기대어 있음."
       }
      ],
      "all_candidates_fail": true
     },
     "normalized": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 0,
        "verdict_ko": "의병4의 외형과 정면 구도는 훌륭하게 구현했으나, 프롬프트가 엄격히 금지한 추가 인물들(배경의 5명)이 묘사되어 치명적 규정 위반입니다."
       },
       {
        "label": "B",
        "score": 0,
        "verdict_ko": "프롬프트가 금지한 추가 인물들이 다수 등장할 뿐만 아니라, 중심 인물 역시 지정된 의병4가 아닌 이전 컷의 인물로 그려져 치명적 규정 위반입니다."
       }
      ],
      "all_candidates_fail": true,
      "readings": [
       {
        "label": "A",
        "direction": "중앙의 인물이 상처에 두 손을 얹은 채 카메라 정면을 또렷하게 응시하고 있음(안규홍이 렌즈 너머에 있다고 설정된 시선).",
        "built_space": "해안가 바위 사이의 모래 바닥이며, 뒤편으로 파도치는 바다가 보임.",
        "entities": "레퍼런스와 일치하는 의병4(두건 및 복장)가 중앙에 쪼그려 있으나, 프롬프트 지시와 달리 배경에 5명의 추가 인물들이 상처를 치료하는 모습으로 등장함.",
        "hard_violations": [
         "지정되지 않은 추가 인물 다수 등장 (배경의 5명)"
        ],
        "physics": "인물들이 모래 바닥에 체중을 싣고 안정적으로 앉거나 쪼그려 있음."
       },
       {
        "label": "B",
        "direction": "중앙의 인물이 상처에 손을 얹고 화면 우측 전경에 있는 다른 인물을 향해 시선을 던지고 있음.",
        "built_space": "해안가 거대한 바위 밑 모래 바닥이며, 바위가 인물들 뒤로 넓게 자리 잡고 있음.",
        "entities": "중심 인물이 의병4의 레퍼런스(두건 착용 등)와 일치하지 않고 이전 샷의 다른 인물과 유사함. 또한 화면 전체에 걸쳐 프롬프트에 허용되지 않은 인물 3명이 추가로 등장함.",
        "hard_violations": [
         "지정되지 않은 추가 인물 등장 (화면 안의 다른 3명)",
         "캐릭터 외형 불일치 및 이전 샷 인물 차용"
        ],
        "physics": "인물들이 모래 바닥에 쪼그려 앉아있거나 바위에 기대어 있음."
       }
      ]
     },
     "ok": true
    },
    {
     "model": "gpt-high",
     "order": "reverse",
     "display_to_canonical": {
      "A": "B",
      "B": "A"
     },
     "raw": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 1,
        "verdict_ko": "상처를 짚고 상대를 보는 동작은 있으나, 의병4 대신 이전 장면 인물의 외형을 사용하고 금지된 추가 인물 네 명을 보여 실격입니다."
       },
       {
        "label": "B",
        "score": 3,
        "verdict_ko": "의병4의 두건·복장과 상처에 손을 얹은 굳은 정면은 더 충실하지만, 이전 장면에서 가져온 듯한 추가 인물 다섯 명 때문에 실격입니다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "중앙 남성은 고개를 들고 화면 오른쪽 전경 남성의 얼굴 쪽을 바라본다. 시선의 상대는 분명하지만 이 전경 인물을 안규홍으로 확인할 근거가 없으며, 허용된 등장인물도 아니다. 왼쪽 배경 두 남성은 손과 다리의 처치 부위를 내려다보고, 오른쪽 뒤 남성도 아래를 본다. 무기나 이동하는 물체는 없다.",
        "built_space": "인공 구조물이나 고정 설비는 없다. 오른쪽의 큰 해안 암벽, 왼쪽 바다와 포말, 아래 모래 바닥은 장소 참고와 잘 연결된다. 중앙에 낮게 앉은 남성 한 명, 오른쪽 전경에 일부만 보이는 남성 한 명, 뒤쪽 모래와 바위 곁에 남성 세 명으로 총 다섯 명이 보인다. 정면 단독 미디엄 숏보다는 상대 어깨를 포함한 대화 구도에 가깝다.",
        "entities": "중앙 인물은 검은 머리의 한국인 성인 남성으로 읽히지만, 인물 참고의 검은 두건과 검은 허리띠가 없고 얼굴도 다르다. 오히려 이전 장면의 가슴이 찢어진 밝은 옷을 입은 남성과 닮았다. 무릎 아래에는 피 묻은 천 드레싱이 있고 손이 그 주변을 감싼다. 배경 인물들의 얼굴과 옷 역시 제외하라고 한 이전 장면을 재사용한 모습이다. 읽을 수 있는 글자는 없다.",
        "hard_violations": [
         "의병4만 허용된 화면에 전경 남성 한 명과 배경 남성 세 명을 추가했다.",
         "이전 장면 인물의 얼굴과 의상을 가져오지 말라는 명시적 금지에도 중앙 인물과 배경 인물들에 그 외형을 재사용했다."
        ],
        "physics": "중앙 남성은 다리를 굽혀 낮게 앉아 있고 하체가 모래에 닿는 배치로 읽힌다. 한 손은 무릎에, 다른 손은 드레싱 옆에 접촉하며 천은 다리를 둘러 지지된다. 배경 인물들도 엉덩이·무릎·다리가 모래에 닿아 있다. 전경 남성의 하체는 화면 밖이므로 지지 상태를 단정할 수 없지만, 부유나 불가능한 자세는 보이지 않는다."
       },
       {
        "label": "B",
        "direction": "중앙 남성은 얼굴을 정면으로 들고 렌즈 부근을 또렷하게 응시한다. 화면 밖 안규홍을 카메라 축에 두었다면 성립하지만, 이미지 자체에는 그 상대를 확인할 단서가 없다. 왼쪽 인물들은 붕대를 감는 손과 다리를 내려다보고, 오른쪽 앞 남성도 처치 부위를 본다. 오른쪽 뒤 남성은 왼쪽을 향한다. 무기나 이동 물체는 없다.",
        "built_space": "인공 구조물이나 고정 설비는 없다. 양옆의 큰 바위 사이로 바다와 거친 포말이 보이고 전경은 모래다. 참고 장소의 재질과 해안 환경은 유지하지만 바위 사이 통로를 정면에서 보는 배치로 바뀌었다. 중앙 남성 한 명, 왼쪽 세 명, 오른쪽 두 명으로 총 여섯 명이 앉아 있다. 중앙 인물의 얼굴부터 상처를 누르는 손까지 보여 정면 미디엄 숏의 핵심을 담는다.",
        "entities": "중앙 인물은 한국인 성인 남성으로 읽히며 검은 두건, 해진 밝은 저고리, 검은 허리띠가 인물 참고와 잘 맞는다. 얼굴은 참고보다 다소 나이 들어 보이지만 A보다 정체성이 가깝다. 피 묻은 찢어진 천이 다리에 감겨 있고 양손이 그 위에 놓여 있다. 나머지 다섯 남성은 허용되지 않은 등장인물이며 일부는 이전 장면 인물의 얼굴과 복장을 닮았다. 오른쪽 모래에는 천을 담은 얕은 나무 용기 두 개와 화면 끝의 바구니 하나가 보인다. 읽을 수 있는 글자는 없다.",
        "hard_violations": [
         "의병4 외에 왼쪽 세 명과 오른쪽 두 명, 총 다섯 명의 금지된 추가 인물을 배치했다.",
         "이전 장면에서 제외하도록 지정한 인물들의 외형과 복장을 배경에 재사용했다."
        ],
        "physics": "중앙 남성은 한쪽 다리를 접고 다른 쪽 다리를 전경으로 내민 낮은 앉은 자세이며, 모래에 놓인 하체가 몸을 지지한다. 깊이 웅크린 자세보다는 앉아서 자신의 상처를 확인하는 자세에 가깝다. 양손이 드레싱에 실제로 닿고 천은 다리에 감겨 있다. 배경 인물들의 엉덩이와 다리도 모래에 놓이며 처치하는 손은 붕대에 접촉한다. 나무 용기와 바구니는 모래 위에 놓여 있고, 지지 없이 떠 있는 물체나 몸은 없다."
       }
      ],
      "all_candidates_fail": true
     },
     "normalized": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 1,
        "verdict_ko": "상처를 짚고 상대를 보는 동작은 있으나, 의병4 대신 이전 장면 인물의 외형을 사용하고 금지된 추가 인물 네 명을 보여 실격입니다."
       },
       {
        "label": "A",
        "score": 3,
        "verdict_ko": "의병4의 두건·복장과 상처에 손을 얹은 굳은 정면은 더 충실하지만, 이전 장면에서 가져온 듯한 추가 인물 다섯 명 때문에 실격입니다."
       }
      ],
      "all_candidates_fail": true,
      "readings": [
       {
        "label": "B",
        "direction": "중앙 남성은 고개를 들고 화면 오른쪽 전경 남성의 얼굴 쪽을 바라본다. 시선의 상대는 분명하지만 이 전경 인물을 안규홍으로 확인할 근거가 없으며, 허용된 등장인물도 아니다. 왼쪽 배경 두 남성은 손과 다리의 처치 부위를 내려다보고, 오른쪽 뒤 남성도 아래를 본다. 무기나 이동하는 물체는 없다.",
        "built_space": "인공 구조물이나 고정 설비는 없다. 오른쪽의 큰 해안 암벽, 왼쪽 바다와 포말, 아래 모래 바닥은 장소 참고와 잘 연결된다. 중앙에 낮게 앉은 남성 한 명, 오른쪽 전경에 일부만 보이는 남성 한 명, 뒤쪽 모래와 바위 곁에 남성 세 명으로 총 다섯 명이 보인다. 정면 단독 미디엄 숏보다는 상대 어깨를 포함한 대화 구도에 가깝다.",
        "entities": "중앙 인물은 검은 머리의 한국인 성인 남성으로 읽히지만, 인물 참고의 검은 두건과 검은 허리띠가 없고 얼굴도 다르다. 오히려 이전 장면의 가슴이 찢어진 밝은 옷을 입은 남성과 닮았다. 무릎 아래에는 피 묻은 천 드레싱이 있고 손이 그 주변을 감싼다. 배경 인물들의 얼굴과 옷 역시 제외하라고 한 이전 장면을 재사용한 모습이다. 읽을 수 있는 글자는 없다.",
        "hard_violations": [
         "의병4만 허용된 화면에 전경 남성 한 명과 배경 남성 세 명을 추가했다.",
         "이전 장면 인물의 얼굴과 의상을 가져오지 말라는 명시적 금지에도 중앙 인물과 배경 인물들에 그 외형을 재사용했다."
        ],
        "physics": "중앙 남성은 다리를 굽혀 낮게 앉아 있고 하체가 모래에 닿는 배치로 읽힌다. 한 손은 무릎에, 다른 손은 드레싱 옆에 접촉하며 천은 다리를 둘러 지지된다. 배경 인물들도 엉덩이·무릎·다리가 모래에 닿아 있다. 전경 남성의 하체는 화면 밖이므로 지지 상태를 단정할 수 없지만, 부유나 불가능한 자세는 보이지 않는다."
       },
       {
        "label": "A",
        "direction": "중앙 남성은 얼굴을 정면으로 들고 렌즈 부근을 또렷하게 응시한다. 화면 밖 안규홍을 카메라 축에 두었다면 성립하지만, 이미지 자체에는 그 상대를 확인할 단서가 없다. 왼쪽 인물들은 붕대를 감는 손과 다리를 내려다보고, 오른쪽 앞 남성도 처치 부위를 본다. 오른쪽 뒤 남성은 왼쪽을 향한다. 무기나 이동 물체는 없다.",
        "built_space": "인공 구조물이나 고정 설비는 없다. 양옆의 큰 바위 사이로 바다와 거친 포말이 보이고 전경은 모래다. 참고 장소의 재질과 해안 환경은 유지하지만 바위 사이 통로를 정면에서 보는 배치로 바뀌었다. 중앙 남성 한 명, 왼쪽 세 명, 오른쪽 두 명으로 총 여섯 명이 앉아 있다. 중앙 인물의 얼굴부터 상처를 누르는 손까지 보여 정면 미디엄 숏의 핵심을 담는다.",
        "entities": "중앙 인물은 한국인 성인 남성으로 읽히며 검은 두건, 해진 밝은 저고리, 검은 허리띠가 인물 참고와 잘 맞는다. 얼굴은 참고보다 다소 나이 들어 보이지만 A보다 정체성이 가깝다. 피 묻은 찢어진 천이 다리에 감겨 있고 양손이 그 위에 놓여 있다. 나머지 다섯 남성은 허용되지 않은 등장인물이며 일부는 이전 장면 인물의 얼굴과 복장을 닮았다. 오른쪽 모래에는 천을 담은 얕은 나무 용기 두 개와 화면 끝의 바구니 하나가 보인다. 읽을 수 있는 글자는 없다.",
        "hard_violations": [
         "의병4 외에 왼쪽 세 명과 오른쪽 두 명, 총 다섯 명의 금지된 추가 인물을 배치했다.",
         "이전 장면에서 제외하도록 지정한 인물들의 외형과 복장을 배경에 재사용했다."
        ],
        "physics": "중앙 남성은 한쪽 다리를 접고 다른 쪽 다리를 전경으로 내민 낮은 앉은 자세이며, 모래에 놓인 하체가 몸을 지지한다. 깊이 웅크린 자세보다는 앉아서 자신의 상처를 확인하는 자세에 가깝다. 양손이 드레싱에 실제로 닿고 천은 다리에 감겨 있다. 배경 인물들의 엉덩이와 다리도 모래에 놓이며 처치하는 손은 붕대에 접촉한다. 나무 용기와 바구니는 모래 위에 놓여 있고, 지지 없이 떠 있는 물체나 몸은 없다."
       }
      ]
     },
     "ok": true
    }
   ],
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "slot_winner_match": true,
   "slot_winner": {
    "gemini-pro": "A",
    "gpt-high": "A"
   },
   "route": "cross_slot_agree"
  },
  "dual": {
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "normalized": {
    "A": 1.0,
    "B": 0.333
   },
   "adjusted": {
    "A": 0.75,
    "B": 0.083
   },
   "violations": {
    "A": [
     "[gemini-pro] 지정되지 않은 추가 인물 다수 등장 (배경의 5명)",
     "[gpt-high] 의병4 외에 왼쪽 세 명과 오른쪽 두 명, 총 다섯 명의 금지된 추가 인물을 배치했다.",
     "[gpt-high] 이전 장면에서 제외하도록 지정한 인물들의 외형과 복장을 배경에 재사용했다."
    ],
    "B": [
     "[gemini-pro] 지정되지 않은 추가 인물 등장 (화면 안의 다른 3명)",
     "[gemini-pro] 캐릭터 외형 불일치 및 이전 샷 인물 차용",
     "[gpt-high] 의병4만 허용된 화면에 전경 남성 한 명과 배경 남성 세 명을 추가했다.",
     "[gpt-high] 이전 장면 인물의 얼굴과 의상을 가져오지 말라는 명시적 금지에도 중앙 인물과 배경 인물들에 그 외형을 재사용했다."
    ]
   },
   "per_model_winner": {
    "gemini-pro": "A",
    "gpt-high": "A"
   },
   "agreed": true
  },
  "totals": {
   "A": 750,
   "B": 83
  },
  "selected": "A",
  "ranking": [
   "A",
   "B"
  ],
  "verdicts": [
   {
    "label": "A",
    "score": 750,
    "verdict_ko": "의병4의 외형과 정면 구도는 훌륭하게 구현했으나, 프롬프트가 엄격히 금지한 추가 인물들(배경의 5명)이 묘사되어 치명적 규정 위반입니다.  ★위반: [gemini-pro] 지정되지 않은 추가 인물 다수 등장 (배경의 5명) / [gpt-high] 의병4 외에 왼쪽 세 명과 오른쪽 두 명, 총 다섯 명의 금지된 추가 인물을 배치했다. / [gpt-high] 이전 장면에서 제외하도록 지정한 인물들의 외형과 복장을 배경에 재사용했다."
   },
   {
    "label": "B",
    "score": 83,
    "verdict_ko": "프롬프트가 금지한 추가 인물들이 다수 등장할 뿐만 아니라, 중심 인물 역시 지정된 의병4가 아닌 이전 컷의 인물로 그려져 치명적 규정 위반입니다.  ★위반: [gemini-pro] 지정되지 않은 추가 인물 등장 (화면 안의 다른 3명) / [gemini-pro] 캐릭터 외형 불일치 및 이전 샷 인물 차용 / [gpt-high] 의병4만 허용된 화면에 전경 남성 한 명과 배경 남성 세 명을 추가했다. / [gpt-high] 이전 장면 인물의 얼굴과 의상을 가져오지 말라는 명시적 금지에도 중앙 인물과 배경 인물들에 그 외형을 재사용했다."
   }
  ],
  "refs": [
   {
    "label": "PREVIOUS SHOT STILL — a visually related earlier shot of this same place: the location's look, materials, fixed features and lighting mood are LOCKED to this photo; never copy its camera framing. The people visible in that photo are NOT in this shot: never carry their faces, bodies or clothing onto anyone here. Each person in THIS shot is defined solely by the PEOPLE list, their own CHARACTER REFERENCE images and their wardrobe notes. If this photo shows a character who cannot move (dead or unconscious), that character's exact pose, position and orientation are ALSO LOCKED — treat their body as a fixed prop of the set that only the camera moves around.",
    "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/db00517b-165b-4005-a287-f28936b59b95/images/272bd7af-9294-4490-98c6-88a40e9deacc/scene/recipe/S31sh1_sel.png",
    "asset_id": "2524449e-8132-4898-a5ff-7c2d5994ddf0",
    "role": "prev_still"
   },
   {
    "label": "CHARACTER REFERENCE — 의병4: the exact person appearing in this shot; match face, hair and build exactly.",
    "path": "<bytes:851066>",
    "asset_id": "64222903-a4aa-4ec7-8c4f-6a6771e8a1fa",
    "role": "character_ref"
   }
  ],
  "critique_skipped": true,
  "needs_reshoot": true,
  "shot_run_produced": false,
  "shot_run_uid": "06aa1426-0a1e-7ba6-b045-fe5d848eefb2",
  "ref_mode": "prev+엔티티",
  "share_plan": {
   "ref_plan": "prev",
   "prev_anchor_tag": "S31sh1"
  }
 },
 "S31sh12::cine": {
  "applied": false,
  "attempted_at": "2026-09-09T09:07:46.715086+00:00",
  "fingerprint": "534b0315d993cca979eee82b4f931fb36856b3c95e14e8ac33bf4986bbdeb9c8",
  "fingerprint_version": 2,
  "provider": "mai",
  "endpoint": "openrouter/chat-completions",
  "model": "microsoft/mai-image-2.6",
  "pack": "24.202608252115",
  "source_file": "S31sh12_sel.png",
  "source_sha256": "149a3e69065ff8f9e3a412adbbaa26d3bb0a66f5ae1c8030f4724721b8d0f325",
  "error": "RuntimeError: moderation blocked: {\"message\": \"Response content blocked by label 'MultiSeverity_ViolenceScore'.\", \"code\": 400}",
  "moderation_refusals": 2,
  "declined": true,
  "declined_reason": "moderation"
 },
 "S31sh17::signage": {
  "fp": "f29656c61a72140f",
  "inscriptions": [],
  "cues": [],
  "dropped": []
 },
 "S31sh17": {
  "input_fingerprint": "912d0404f8801499",
  "prompt": "Create ONE FINAL photorealistic live-action film still of the moment below — early 20th-century Korea, Jeolla region; all people are Korean unless explicitly stated otherwise. TIME OF DAY (lock): afternoon.\n\nSHOT TEXT (authoritative, Korean): 부서지는 파도를 배경으로 바위 뒤에서 입을 벌려 왁자지껄 웃음을 터뜨린 채 멈춘 안규홍 무리의 훈훈한 전경.\n\nLOCATION (lock): The outdoor resting pocket behind large shoreline rocks, with breaking waves beyond. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: wide shot\n- KEY BACKGROUND ELEMENTS: Coastal rocks (The group has taken shelter behind them) — Their group-facing sides border the foreground; used as Frame the gathering without hiding the rescued prisoners; Breaking waves (Breaking behind the laughing men); used as Maintain the moving background across the upper portion of the composition; Sand (Occupied by the seated prisoners and reclining 이관회); used as Make the different resting postures and group spacing legible.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Natural afternoon light, muted saturation and restrained contrast preserve the men's relief without romanticizing their injuries.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The clothing, hair and overall look of 안규홍, 염재보, 의병 포로1, 의병 포로2, 의병 포로3, 이관회 — who appear both in that photo and in this shot — are LOCKED to that photo. Anyone else visible in that photo is NOT in this shot: never carry their face, body or clothing onto anyone here. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nPREVIOUS STILL USAGE (follow exactly — what to take from the attached still and what to exclude; it governs place and objects only, never who is in this shot or how the camera sees them): Take the coastal rock surfaces, sandy ground, and afternoon sea light. Exclude the arrival's airborne sand and the exact transient shape of the earlier surf.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): Waves continue breaking behind the rocky shelter; improvised cloth dressings remain in use. 안규홍: His knuckles remain split but wiped of blood; the jade ring remains concealed. 이관회: He remains resting on the sand. 염재보: He remains in the refuge retaining his matchlock. 의병 포로1: He remains freed in the shelter after the escape. 의병 포로2: He remains freed in the shelter after the escape. 의병 포로3: He remains freed but unable to walk unaided. 의병4: He remains in the first-aid area after checking a wound.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 안규홍 (한국인 남성, 청년 나이대, 검은 머리, 단단히 묶은 머리); 이관회 (한국인 남성, 청년 나이대, 자연스러운 검은 머리); 염재보 (한국인 남성, 성인 나이대, 자연스러운 검은 머리); 의병 포로1 (한국인 남성, 성인 나이대, 자연스러운 검은 머리); 의병 포로2 (한국인 남성, 성인 나이대, 자연스러운 검은 머리); 의병 포로3 (한국인 남성, 성인 나이대, 자연스러운 검은 머리); 의병4 (한국인 남성, 성인 나이대, 자연스러운 검은 머리) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.",
  "roll_prompts": {
   "A": "Create ONE FINAL photorealistic live-action film still of the moment below — early 20th-century Korea, Jeolla region; all people are Korean unless explicitly stated otherwise. TIME OF DAY (lock): afternoon.\n\nSHOT TEXT (authoritative, Korean): 부서지는 파도를 배경으로 바위 뒤에서 입을 벌려 왁자지껄 웃음을 터뜨린 채 멈춘 안규홍 무리의 훈훈한 전경.\n\nLOCATION (lock): The outdoor resting pocket behind large shoreline rocks, with breaking waves beyond. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: wide shot\n- KEY BACKGROUND ELEMENTS: Coastal rocks (The group has taken shelter behind them) — Their group-facing sides border the foreground; used as Frame the gathering without hiding the rescued prisoners; Breaking waves (Breaking behind the laughing men); used as Maintain the moving background across the upper portion of the composition; Sand (Occupied by the seated prisoners and reclining 이관회); used as Make the different resting postures and group spacing legible.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Natural afternoon light, muted saturation and restrained contrast preserve the men's relief without romanticizing their injuries.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The clothing, hair and overall look of 안규홍, 염재보, 의병 포로1, 의병 포로2, 의병 포로3, 이관회 — who appear both in that photo and in this shot — are LOCKED to that photo. Anyone else visible in that photo is NOT in this shot: never carry their face, body or clothing onto anyone here. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nPREVIOUS STILL USAGE (follow exactly — what to take from the attached still and what to exclude; it governs place and objects only, never who is in this shot or how the camera sees them): Take the coastal rock surfaces, sandy ground, and afternoon sea light. Exclude the arrival's airborne sand and the exact transient shape of the earlier surf.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): Waves continue breaking behind the rocky shelter; improvised cloth dressings remain in use. 안규홍: His knuckles remain split but wiped of blood; the jade ring remains concealed. 이관회: He remains resting on the sand. 염재보: He remains in the refuge retaining his matchlock. 의병 포로1: He remains freed in the shelter after the escape. 의병 포로2: He remains freed in the shelter after the escape. 의병 포로3: He remains freed but unable to walk unaided. 의병4: He remains in the first-aid area after checking a wound.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 안규홍 (한국인 남성, 청년 나이대, 검은 머리, 단단히 묶은 머리); 이관회 (한국인 남성, 청년 나이대, 자연스러운 검은 머리); 염재보 (한국인 남성, 성인 나이대, 자연스러운 검은 머리); 의병 포로1 (한국인 남성, 성인 나이대, 자연스러운 검은 머리); 의병 포로2 (한국인 남성, 성인 나이대, 자연스러운 검은 머리); 의병 포로3 (한국인 남성, 성인 나이대, 자연스러운 검은 머리); 의병4 (한국인 남성, 성인 나이대, 자연스러운 검은 머리) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.",
   "B": "Create ONE FINAL photorealistic live-action film still of the moment below — early 20th-century Korea, Jeolla region; all people are Korean unless explicitly stated otherwise. TIME OF DAY (lock): afternoon.\n\nSHOT TEXT (authoritative, Korean): 부서지는 파도를 배경으로 바위 뒤에서 입을 벌려 왁자지껄 웃음을 터뜨린 채 멈춘 안규홍 무리의 훈훈한 전경.\n\nLOCATION (lock): The outdoor resting pocket behind large shoreline rocks, with breaking waves beyond. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: wide shot\n- KEY BACKGROUND ELEMENTS: Coastal rocks (The group has taken shelter behind them) — Their group-facing sides border the foreground; used as Frame the gathering without hiding the rescued prisoners; Breaking waves (Breaking behind the laughing men); used as Maintain the moving background across the upper portion of the composition; Sand (Occupied by the seated prisoners and reclining 이관회); used as Make the different resting postures and group spacing legible.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Natural afternoon light, muted saturation and restrained contrast preserve the men's relief without romanticizing their injuries.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The clothing, hair and overall look of 안규홍, 염재보, 의병 포로1, 의병 포로2, 의병 포로3, 이관회 — who appear both in that photo and in this shot — are LOCKED to that photo. Anyone else visible in that photo is NOT in this shot: never carry their face, body or clothing onto anyone here. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nPREVIOUS STILL USAGE (follow exactly — what to take from the attached still and what to exclude; it governs place and objects only, never who is in this shot or how the camera sees them): Take the coastal rock surfaces, sandy ground, and afternoon sea light. Exclude the arrival's airborne sand and the exact transient shape of the earlier surf.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): Waves continue breaking behind the rocky shelter; improvised cloth dressings remain in use. 안규홍: His knuckles remain split but wiped of blood; the jade ring remains concealed. 이관회: He remains resting on the sand. 염재보: He remains in the refuge retaining his matchlock. 의병 포로1: He remains freed in the shelter after the escape. 의병 포로2: He remains freed in the shelter after the escape. 의병 포로3: He remains freed but unable to walk unaided. 의병4: He remains in the first-aid area after checking a wound.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 안규홍 (한국인 남성, 청년 나이대, 검은 머리, 단단히 묶은 머리); 이관회 (한국인 남성, 청년 나이대, 자연스러운 검은 머리); 염재보 (한국인 남성, 성인 나이대, 자연스러운 검은 머리); 의병 포로1 (한국인 남성, 성인 나이대, 자연스러운 검은 머리); 의병 포로2 (한국인 남성, 성인 나이대, 자연스러운 검은 머리); 의병 포로3 (한국인 남성, 성인 나이대, 자연스러운 검은 머리); 의병4 (한국인 남성, 성인 나이대, 자연스러운 검은 머리) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.\n\nCOMPOSITION VARIATION (this candidate only): keep the place, the people and the moment exactly as specified, and keep from the CAMERA & FRAME contract its subject scale, screen placement and key background elements exactly. For this candidate only, its camera angle and height are deliberately left open: choose a camera position distinctly different from the obvious one, as a film director picking a second setup on the same blocking. Stay true to any explicit framing the SHOT TEXT itself dictates."
  },
  "shot_run_spend_attempt_count": 1,
  "readings": [
   {
    "label": "A",
    "direction": "인물들이 모래사장에 둥글게 모여 앉아 서로를 마주보며 왁자지껄하게 웃고 있다.",
    "built_space": "해안가 바위들이 무리를 감싸고 있으며, 뒤편으로는 하얀 파도가 부서지는 바다가 배경으로 펼쳐져 있다.",
    "entities": "안규홍, 이관회, 염재보, 의병 포로1, 의병 포로2, 의병 포로3이 레퍼런스와 일치하는 복장으로 등장하나, 의병4는 프레임 내에 보이지 않는다.",
    "hard_violations": [
     "[gemini-pro] 안규홍이 쪼그려 앉은 자세에서 두 다리 사이에 바닥을 딛고 있는 세 번째 짚신(발)이 생성됨",
     "[gemini-pro] 염재보의 왼손이 자신의 무릎 위와 왼쪽 바위 위 두 곳에 중복되어 나타남",
     "[gemini-pro] 의병 포로1의 오른쪽 손목과 손가락 부위가 심하게 뭉개지고 해부학적으로 불가능한 구조로 렌더링됨"
    ],
    "physics": "안규홍 밑에 누구의 것인지 알 수 없는 발이 바닥을 딛고 있으며, 염재보가 쥔 화승총의 아랫부분이 허공이나 다리와 불명확하게 융합되어 있다. 신체가 중력이나 구조에 맞지 않게 중복 파생되어 있다."
   },
   {
    "label": "B",
    "direction": "안규홍은 정면 하늘을 향해, 나머지 인물들은 서로를 향해 고개를 돌려 환하게 웃음을 터뜨리고 있다.",
    "built_space": "바위가 양옆과 뒤를 감싸는 형태의 모래사장 쉼터이며, 바위 너머로 거친 파도가 치고 있어 지정된 공간과 일치한다.",
    "entities": "안규홍, 이관회, 염재보, 의병 포로1, 의병 포로2, 의병 포로3, 그리고 등지고 앉은 의병4까지 프롬프트가 지시한 모든 인물이 레퍼런스의 복장 및 외형을 유지한 채 등장한다.",
    "hard_violations": [
     "[gpt-high] 모래 위에 앉도록 지정된 포로2와 포로3을 각각 오른쪽과 왼쪽 바위 위에 앉혀 명시된 휴식 배치를 변경했다."
    ],
    "physics": "모든 인물이 모래사장이나 바위 위에 앉거나 서서 체중을 자연스럽게 지탱하고 있다. 염재보가 쥔 화승총 역시 오른손의 악력에 의해 제대로 지지되어 있으며 허공에 뜬 부자연스러운 물체나 기형적 신체가 없다."
   }
  ],
  "cross_model_order": {
   "policy": "cross_model_order_v2_slot0_fwd_slot1_rev_parallel2",
   "slots": [
    {
     "model": "gemini-pro",
     "order": "forward",
     "display_to_canonical": {
      "A": "A",
      "B": "B"
     },
     "raw": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 9,
        "verdict_ko": "프롬프트가 요구한 모든 인물의 신원과 복장을 정확히 구현하였으며, 치명적인 물리적 결함 없이 무리가 환하게 웃고 있는 전경을 자연스럽고 훌륭하게 연출함."
       },
       {
        "label": "A",
        "score": 0,
        "verdict_ko": "안규홍의 두 다리 사이에 정체불명의 세 번째 발이 존재하며, 다른 인물들의 손이 중복되거나 기형적으로 나타나는 등 심각한 신체 렌더링 오류가 발생함."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "인물들이 모래사장에 둥글게 모여 앉아 서로를 마주보며 왁자지껄하게 웃고 있다.",
        "built_space": "해안가 바위들이 무리를 감싸고 있으며, 뒤편으로는 하얀 파도가 부서지는 바다가 배경으로 펼쳐져 있다.",
        "entities": "안규홍, 이관회, 염재보, 의병 포로1, 의병 포로2, 의병 포로3이 레퍼런스와 일치하는 복장으로 등장하나, 의병4는 프레임 내에 보이지 않는다.",
        "hard_violations": [
         "안규홍이 쪼그려 앉은 자세에서 두 다리 사이에 바닥을 딛고 있는 세 번째 짚신(발)이 생성됨",
         "염재보의 왼손이 자신의 무릎 위와 왼쪽 바위 위 두 곳에 중복되어 나타남",
         "의병 포로1의 오른쪽 손목과 손가락 부위가 심하게 뭉개지고 해부학적으로 불가능한 구조로 렌더링됨"
        ],
        "physics": "안규홍 밑에 누구의 것인지 알 수 없는 발이 바닥을 딛고 있으며, 염재보가 쥔 화승총의 아랫부분이 허공이나 다리와 불명확하게 융합되어 있다. 신체가 중력이나 구조에 맞지 않게 중복 파생되어 있다."
       },
       {
        "label": "B",
        "direction": "안규홍은 정면 하늘을 향해, 나머지 인물들은 서로를 향해 고개를 돌려 환하게 웃음을 터뜨리고 있다.",
        "built_space": "바위가 양옆과 뒤를 감싸는 형태의 모래사장 쉼터이며, 바위 너머로 거친 파도가 치고 있어 지정된 공간과 일치한다.",
        "entities": "안규홍, 이관회, 염재보, 의병 포로1, 의병 포로2, 의병 포로3, 그리고 등지고 앉은 의병4까지 프롬프트가 지시한 모든 인물이 레퍼런스의 복장 및 외형을 유지한 채 등장한다.",
        "hard_violations": [],
        "physics": "모든 인물이 모래사장이나 바위 위에 앉거나 서서 체중을 자연스럽게 지탱하고 있다. 염재보가 쥔 화승총 역시 오른손의 악력에 의해 제대로 지지되어 있으며 허공에 뜬 부자연스러운 물체나 기형적 신체가 없다."
       }
      ],
      "all_candidates_fail": false
     },
     "normalized": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 9,
        "verdict_ko": "프롬프트가 요구한 모든 인물의 신원과 복장을 정확히 구현하였으며, 치명적인 물리적 결함 없이 무리가 환하게 웃고 있는 전경을 자연스럽고 훌륭하게 연출함."
       },
       {
        "label": "A",
        "score": 0,
        "verdict_ko": "안규홍의 두 다리 사이에 정체불명의 세 번째 발이 존재하며, 다른 인물들의 손이 중복되거나 기형적으로 나타나는 등 심각한 신체 렌더링 오류가 발생함."
       }
      ],
      "all_candidates_fail": false,
      "readings": [
       {
        "label": "A",
        "direction": "인물들이 모래사장에 둥글게 모여 앉아 서로를 마주보며 왁자지껄하게 웃고 있다.",
        "built_space": "해안가 바위들이 무리를 감싸고 있으며, 뒤편으로는 하얀 파도가 부서지는 바다가 배경으로 펼쳐져 있다.",
        "entities": "안규홍, 이관회, 염재보, 의병 포로1, 의병 포로2, 의병 포로3이 레퍼런스와 일치하는 복장으로 등장하나, 의병4는 프레임 내에 보이지 않는다.",
        "hard_violations": [
         "안규홍이 쪼그려 앉은 자세에서 두 다리 사이에 바닥을 딛고 있는 세 번째 짚신(발)이 생성됨",
         "염재보의 왼손이 자신의 무릎 위와 왼쪽 바위 위 두 곳에 중복되어 나타남",
         "의병 포로1의 오른쪽 손목과 손가락 부위가 심하게 뭉개지고 해부학적으로 불가능한 구조로 렌더링됨"
        ],
        "physics": "안규홍 밑에 누구의 것인지 알 수 없는 발이 바닥을 딛고 있으며, 염재보가 쥔 화승총의 아랫부분이 허공이나 다리와 불명확하게 융합되어 있다. 신체가 중력이나 구조에 맞지 않게 중복 파생되어 있다."
       },
       {
        "label": "B",
        "direction": "안규홍은 정면 하늘을 향해, 나머지 인물들은 서로를 향해 고개를 돌려 환하게 웃음을 터뜨리고 있다.",
        "built_space": "바위가 양옆과 뒤를 감싸는 형태의 모래사장 쉼터이며, 바위 너머로 거친 파도가 치고 있어 지정된 공간과 일치한다.",
        "entities": "안규홍, 이관회, 염재보, 의병 포로1, 의병 포로2, 의병 포로3, 그리고 등지고 앉은 의병4까지 프롬프트가 지시한 모든 인물이 레퍼런스의 복장 및 외형을 유지한 채 등장한다.",
        "hard_violations": [],
        "physics": "모든 인물이 모래사장이나 바위 위에 앉거나 서서 체중을 자연스럽게 지탱하고 있다. 염재보가 쥔 화승총 역시 오른손의 악력에 의해 제대로 지지되어 있으며 허공에 뜬 부자연스러운 물체나 기형적 신체가 없다."
       }
      ]
     },
     "ok": true
    },
    {
     "model": "gpt-high",
     "order": "reverse",
     "display_to_canonical": {
      "A": "B",
      "B": "A"
     },
     "raw": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 6,
        "verdict_ko": "파도 앞에서 웃는 무리의 전경은 구현했지만, 포로들을 바위에 앉히고 이관회를 곧게 앉혀 지정된 모래 위 휴식 배치와 달라졌다."
       },
       {
        "label": "B",
        "score": 8,
        "verdict_ko": "양옆 바위가 감싸는 모래 쉼터, 기대어 쉬는 이관회와 앉은 포로들, 상단의 쇄파와 집단 웃음을 더 충실히 구현했으나 염재보의 머리와 안규홍 손의 혈흔은 연속성이 부족하다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "이관회와 포로1은 중앙 안규홍 쪽으로 얼굴을 들어 웃고, 왼쪽 뒤 포로3과 오른쪽 포로2도 모임 안쪽을 향한다. 염재보는 왼쪽 아래 무리를 보고, 등을 보인 의병4는 안규홍을 향한다. 안규홍은 거의 카메라 정면으로 입을 벌리고 웃는다. 염재보의 총구는 오른쪽 위 허공을 향하며 사람을 겨누지 않는다.",
        "built_space": "인공 구조물이나 고정 설비, 반사는 없다. 중앙의 큰 바위 덩어리 하나, 왼쪽 암석군과 오른쪽 암벽이 모임을 둘러싸며 상단에 부서지는 파도가 보인다. 안규홍과 염재보는 서 있고, 이관회·포로1·의병4는 모래에 앉아 있다. 포로2는 오른쪽 낮은 바위에, 포로3은 왼쪽 바위에 걸터앉아 있어 포로들이 모래를 점유한다는 배치와 다르다. 중앙 바위가 모임 뒤를 크게 채우지만 얼굴을 가리지는 않는다.",
        "entities": "총 일곱 명으로, 허용된 인물들의 의상에 대응하는 한국인 설정의 청년·성인 남성들로 보인다. 안규홍의 묶은 검은 머리와 밝은 옷은 유사하지만 이전 장면의 벌어진 앞섶은 닫혔다. 이관회의 갈색 저고리, 염재보의 검은 옷과 장총, 포로1의 찢어진 가슴 부분, 포로2의 머리띠와 무릎 상처, 포로3의 해진 갈색 옷이 식별된다. 포로2의 짙은 각반과 신발은 이전 모습과 다르다. 의병4는 검은 머릿수건과 허리띠가 보이나 뒷모습이라 얼굴은 확인할 수 없다. 머리에 감은 천은 있으나 추가 응급 드레싱은 뚜렷하지 않다. 옥반지와 읽을 수 있는 문자는 보이지 않는다.",
        "hard_violations": [
         "모래 위에 앉도록 지정된 포로2와 포로3을 각각 오른쪽과 왼쪽 바위 위에 앉혀 명시된 휴식 배치를 변경했다."
        ],
        "physics": "서 있는 두 사람은 발로 모래를 딛고 있다. 모래에 앉은 사람들은 엉덩이와 다리로, 바위에 앉은 포로들은 바위와 내려놓은 다리로 지지된다. 염재보의 장총은 손에 잡혀 있다. 지지 없이 떠 있는 사람이나 물체는 없다. 다만 이관회는 기대어 눕기보다 상체를 세워 앉아 있어 지정된 휴식 자세와 다르다."
       },
       {
        "label": "B",
        "direction": "왼쪽 이관회와 포로1, 뒤쪽 포로3은 중앙의 웃는 안규홍 쪽을 향한다. 오른쪽 염재보와 포로2도 모임 안쪽을 보고 웃는다. 안규홍은 눈을 가늘게 감고 얼굴을 앞으로 향한 채 입을 크게 벌린다. 염재보의 총구는 위쪽 허공을 향하고 어느 동료도 겨누지 않는다. 파도는 뒤쪽 바다에서 해안으로 밀려와 부서진다.",
        "built_space": "인공 설비나 반사는 없다. 왼쪽 전경의 바위와 오른쪽 연속 암벽이 가운데 모래 쉼터를 감싸고, 상단에는 바다와 쇄파가 넓게 이어진다. 이관회는 왼쪽 바위에 등을 기대고 모래 위로 다리를 뻗는다. 포로1은 왼쪽 중앙 모래에 앉고, 포로3은 그 뒤에 낮게 머문다. 안규홍은 중앙에 쪼그려 앉고, 염재보는 오른쪽 낮은 암석에 앉으며 포로2는 오른쪽 모래에서 암벽에 기대 쉰다. 바위가 포로들의 모습을 가리지 않아 서로 다른 휴식 자세가 읽힌다.",
        "entities": "여섯 명의 한국인 설정에 부합하는 청년·성인 남성이 보이며 안규홍, 이관회, 염재보와 포로 세 명으로 식별된다. 의병4는 보이지 않는다. 안규홍의 얼굴과 묶은 머리, 이관회의 갈색 저고리와 밝은 바지, 포로1의 가슴이 찢어진 옷, 포로2의 머리띠와 무릎이 해진 회색 옷, 포로3의 낡은 갈색 옷은 참고와 대체로 대응한다. 염재보의 검은 옷과 장총은 유지되지만 머리띠와 솟은 상투는 이전 장면의 머리와 다르다. 손목 등에 감은 천 드레싱이 명확하다. 안규홍의 든 손에는 붉은 혈흔이 남아 닦인 손이라는 조건에 덜 맞는다. 옥반지나 읽을 수 있는 문자는 보이지 않는다.",
        "hard_violations": [],
        "physics": "이관회와 포로2는 엉덩이와 다리를 모래에 대고 등을 바위에 기대므로 체중 지지가 자연스럽다. 포로1도 모래에 앉아 있고, 뒤의 포로3은 낮게 앉아 있어 혼자 걷는 모습이 아니다. 안규홍은 양발로 모래를 딛고 무릎을 굽혀 쪼그려 앉아 있다. 염재보는 낮은 바위와 발로 지지되며, 장총은 손으로 붙잡고 개머리판을 모래에 댄다. 떠 있거나 지지점이 없는 몸과 물체는 없다."
       }
      ],
      "all_candidates_fail": false
     },
     "normalized": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 6,
        "verdict_ko": "파도 앞에서 웃는 무리의 전경은 구현했지만, 포로들을 바위에 앉히고 이관회를 곧게 앉혀 지정된 모래 위 휴식 배치와 달라졌다."
       },
       {
        "label": "A",
        "score": 8,
        "verdict_ko": "양옆 바위가 감싸는 모래 쉼터, 기대어 쉬는 이관회와 앉은 포로들, 상단의 쇄파와 집단 웃음을 더 충실히 구현했으나 염재보의 머리와 안규홍 손의 혈흔은 연속성이 부족하다."
       }
      ],
      "all_candidates_fail": false,
      "readings": [
       {
        "label": "B",
        "direction": "이관회와 포로1은 중앙 안규홍 쪽으로 얼굴을 들어 웃고, 왼쪽 뒤 포로3과 오른쪽 포로2도 모임 안쪽을 향한다. 염재보는 왼쪽 아래 무리를 보고, 등을 보인 의병4는 안규홍을 향한다. 안규홍은 거의 카메라 정면으로 입을 벌리고 웃는다. 염재보의 총구는 오른쪽 위 허공을 향하며 사람을 겨누지 않는다.",
        "built_space": "인공 구조물이나 고정 설비, 반사는 없다. 중앙의 큰 바위 덩어리 하나, 왼쪽 암석군과 오른쪽 암벽이 모임을 둘러싸며 상단에 부서지는 파도가 보인다. 안규홍과 염재보는 서 있고, 이관회·포로1·의병4는 모래에 앉아 있다. 포로2는 오른쪽 낮은 바위에, 포로3은 왼쪽 바위에 걸터앉아 있어 포로들이 모래를 점유한다는 배치와 다르다. 중앙 바위가 모임 뒤를 크게 채우지만 얼굴을 가리지는 않는다.",
        "entities": "총 일곱 명으로, 허용된 인물들의 의상에 대응하는 한국인 설정의 청년·성인 남성들로 보인다. 안규홍의 묶은 검은 머리와 밝은 옷은 유사하지만 이전 장면의 벌어진 앞섶은 닫혔다. 이관회의 갈색 저고리, 염재보의 검은 옷과 장총, 포로1의 찢어진 가슴 부분, 포로2의 머리띠와 무릎 상처, 포로3의 해진 갈색 옷이 식별된다. 포로2의 짙은 각반과 신발은 이전 모습과 다르다. 의병4는 검은 머릿수건과 허리띠가 보이나 뒷모습이라 얼굴은 확인할 수 없다. 머리에 감은 천은 있으나 추가 응급 드레싱은 뚜렷하지 않다. 옥반지와 읽을 수 있는 문자는 보이지 않는다.",
        "hard_violations": [
         "모래 위에 앉도록 지정된 포로2와 포로3을 각각 오른쪽과 왼쪽 바위 위에 앉혀 명시된 휴식 배치를 변경했다."
        ],
        "physics": "서 있는 두 사람은 발로 모래를 딛고 있다. 모래에 앉은 사람들은 엉덩이와 다리로, 바위에 앉은 포로들은 바위와 내려놓은 다리로 지지된다. 염재보의 장총은 손에 잡혀 있다. 지지 없이 떠 있는 사람이나 물체는 없다. 다만 이관회는 기대어 눕기보다 상체를 세워 앉아 있어 지정된 휴식 자세와 다르다."
       },
       {
        "label": "A",
        "direction": "왼쪽 이관회와 포로1, 뒤쪽 포로3은 중앙의 웃는 안규홍 쪽을 향한다. 오른쪽 염재보와 포로2도 모임 안쪽을 보고 웃는다. 안규홍은 눈을 가늘게 감고 얼굴을 앞으로 향한 채 입을 크게 벌린다. 염재보의 총구는 위쪽 허공을 향하고 어느 동료도 겨누지 않는다. 파도는 뒤쪽 바다에서 해안으로 밀려와 부서진다.",
        "built_space": "인공 설비나 반사는 없다. 왼쪽 전경의 바위와 오른쪽 연속 암벽이 가운데 모래 쉼터를 감싸고, 상단에는 바다와 쇄파가 넓게 이어진다. 이관회는 왼쪽 바위에 등을 기대고 모래 위로 다리를 뻗는다. 포로1은 왼쪽 중앙 모래에 앉고, 포로3은 그 뒤에 낮게 머문다. 안규홍은 중앙에 쪼그려 앉고, 염재보는 오른쪽 낮은 암석에 앉으며 포로2는 오른쪽 모래에서 암벽에 기대 쉰다. 바위가 포로들의 모습을 가리지 않아 서로 다른 휴식 자세가 읽힌다.",
        "entities": "여섯 명의 한국인 설정에 부합하는 청년·성인 남성이 보이며 안규홍, 이관회, 염재보와 포로 세 명으로 식별된다. 의병4는 보이지 않는다. 안규홍의 얼굴과 묶은 머리, 이관회의 갈색 저고리와 밝은 바지, 포로1의 가슴이 찢어진 옷, 포로2의 머리띠와 무릎이 해진 회색 옷, 포로3의 낡은 갈색 옷은 참고와 대체로 대응한다. 염재보의 검은 옷과 장총은 유지되지만 머리띠와 솟은 상투는 이전 장면의 머리와 다르다. 손목 등에 감은 천 드레싱이 명확하다. 안규홍의 든 손에는 붉은 혈흔이 남아 닦인 손이라는 조건에 덜 맞는다. 옥반지나 읽을 수 있는 문자는 보이지 않는다.",
        "hard_violations": [],
        "physics": "이관회와 포로2는 엉덩이와 다리를 모래에 대고 등을 바위에 기대므로 체중 지지가 자연스럽다. 포로1도 모래에 앉아 있고, 뒤의 포로3은 낮게 앉아 있어 혼자 걷는 모습이 아니다. 안규홍은 양발로 모래를 딛고 무릎을 굽혀 쪼그려 앉아 있다. 염재보는 낮은 바위와 발로 지지되며, 장총은 손으로 붙잡고 개머리판을 모래에 댄다. 떠 있거나 지지점이 없는 몸과 물체는 없다."
       }
      ]
     },
     "ok": true
    }
   ],
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "slot_winner_match": false,
   "slot_winner": {
    "gemini-pro": "B",
    "gpt-high": "A"
   },
   "route": "cross_slot_combined"
  },
  "dual": {
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "normalized": {
    "A": 1.0,
    "B": 1.75
   },
   "adjusted": {
    "A": 0.75,
    "B": 1.5
   },
   "violations": {
    "A": [
     "[gemini-pro] 안규홍이 쪼그려 앉은 자세에서 두 다리 사이에 바닥을 딛고 있는 세 번째 짚신(발)이 생성됨",
     "[gemini-pro] 염재보의 왼손이 자신의 무릎 위와 왼쪽 바위 위 두 곳에 중복되어 나타남",
     "[gemini-pro] 의병 포로1의 오른쪽 손목과 손가락 부위가 심하게 뭉개지고 해부학적으로 불가능한 구조로 렌더링됨"
    ],
    "B": [
     "[gpt-high] 모래 위에 앉도록 지정된 포로2와 포로3을 각각 오른쪽과 왼쪽 바위 위에 앉혀 명시된 휴식 배치를 변경했다."
    ]
   },
   "per_model_winner": {
    "gemini-pro": "B",
    "gpt-high": "A"
   },
   "agreed": false
  },
  "totals": {
   "B": 1500,
   "A": 750
  },
  "selected": "B",
  "ranking": [
   "B",
   "A"
  ],
  "verdicts": [
   {
    "label": "B",
    "score": 1500,
    "verdict_ko": "프롬프트가 요구한 모든 인물의 신원과 복장을 정확히 구현하였으며, 치명적인 물리적 결함 없이 무리가 환하게 웃고 있는 전경을 자연스럽고 훌륭하게 연출함.  ★위반: [gpt-high] 모래 위에 앉도록 지정된 포로2와 포로3을 각각 오른쪽과 왼쪽 바위 위에 앉혀 명시된 휴식 배치를 변경했다."
   },
   {
    "label": "A",
    "score": 750,
    "verdict_ko": "안규홍의 두 다리 사이에 정체불명의 세 번째 발이 존재하며, 다른 인물들의 손이 중복되거나 기형적으로 나타나는 등 심각한 신체 렌더링 오류가 발생함.  ★위반: [gemini-pro] 안규홍이 쪼그려 앉은 자세에서 두 다리 사이에 바닥을 딛고 있는 세 번째 짚신(발)이 생성됨 / [gemini-pro] 염재보의 왼손이 자신의 무릎 위와 왼쪽 바위 위 두 곳에 중복되어 나타남 / [gemini-pro] 의병 포로1의 오른쪽 손목과 손가락 부위가 심하게 뭉개지고 해부학적으로 불가능한 구조로 렌더링됨"
   }
  ],
  "refs": [
   {
    "label": "PREVIOUS SHOT STILL — a visually related earlier shot of this same place: the location's look, materials, fixed features and lighting mood are LOCKED to this photo; never copy its camera framing. The clothing, hair and overall look of 안규홍, 염재보, 의병 포로1, 의병 포로2, 의병 포로3, 이관회 — who appear both in that photo and in this shot — are LOCKED to that photo. Anyone else visible in that photo is NOT in this shot: never carry their face, body or clothing onto anyone here. If this photo shows a character who cannot move (dead or unconscious), that character's exact pose, position and orientation are ALSO LOCKED — treat their body as a fixed prop of the set that only the camera moves around.",
    "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/db00517b-165b-4005-a287-f28936b59b95/images/272bd7af-9294-4490-98c6-88a40e9deacc/scene/recipe/S31sh1_sel.png",
    "asset_id": "2524449e-8132-4898-a5ff-7c2d5994ddf0",
    "role": "prev_still"
   },
   {
    "label": "CHARACTER REFERENCE — 안규홍: the exact person appearing in this shot; match face, hair and build exactly.",
    "path": "<bytes:1028384>",
    "asset_id": "3b5894db-ce55-4f34-b966-5e725c3aa952",
    "role": "character_ref"
   },
   {
    "label": "CHARACTER REFERENCE — 이관회: the exact person appearing in this shot; match face, hair and build exactly.",
    "path": "<bytes:933540>",
    "asset_id": "851b450f-ac84-47fe-ac5a-03a56f63fcae",
    "role": "character_ref"
   },
   {
    "label": "CHARACTER REFERENCE — 염재보: the exact person appearing in this shot; match face, hair and build exactly.",
    "path": "<bytes:964351>",
    "asset_id": "ab287289-d0b7-436d-8f70-68452f898f73",
    "role": "character_ref"
   },
   {
    "label": "CHARACTER REFERENCE — 의병 포로1: the exact person appearing in this shot; match face, hair and build exactly.",
    "path": "<bytes:910581>",
    "asset_id": "5a2bba8f-883f-4d9b-9003-b67ab053e656",
    "role": "character_ref"
   },
   {
    "label": "CHARACTER REFERENCE — 의병 포로2: the exact person appearing in this shot; match face, hair and build exactly.",
    "path": "<bytes:867759>",
    "asset_id": "e56e0b15-c7ee-4e69-8d88-8e0c77016d3c",
    "role": "character_ref"
   },
   {
    "label": "CHARACTER REFERENCE — 의병 포로3: the exact person appearing in this shot; match face, hair and build exactly.",
    "path": "<bytes:936113>",
    "asset_id": "ac401254-a302-4663-98a7-cccfef519c1f",
    "role": "character_ref"
   },
   {
    "label": "CHARACTER REFERENCE — 의병4: the exact person appearing in this shot; match face, hair and build exactly.",
    "path": "<bytes:851066>",
    "asset_id": "64222903-a4aa-4ec7-8c4f-6a6771e8a1fa",
    "role": "character_ref"
   }
  ],
  "critique_skipped": true,
  "shot_run_produced": false,
  "shot_run_uid": "06aa1427-2128-7a0d-b45f-f719f31ca348",
  "ref_mode": "prev+엔티티",
  "share_plan": {
   "ref_plan": "prev",
   "prev_anchor_tag": "S31sh1"
  }
 },
 "S31sh17::cine": {
  "applied": true,
  "attempted_at": "2026-09-09T04:26:18.793394+00:00",
  "fingerprint": "a7f0a4fea2602d2bb61b58a5636f338b9401b6e6470bfdee93b8ed2e2a2c42d6",
  "fingerprint_version": 2,
  "provider": "mai",
  "endpoint": "openrouter/chat-completions",
  "model": "microsoft/mai-image-2.6",
  "pack": "24.202608252115",
  "source_file": "S31sh17_sel.png",
  "source_sha256": "6660dbff667118ce4f29837cd37b098824fb859f6d0f8388443ca1c569ad43af",
  "file": "S31sh17_cine.png",
  "staged_sha256": "64d30113cc5a89a46de9cd8ca97b2c1a5379ead21484118caf5d1c2cc72d03b7",
  "latency_ms": 29985
 },
 "S32sh1::signage": {
  "fp": "eed56d0c07b9adb2",
  "inscriptions": [],
  "cues": [],
  "dropped": []
 },
 "S32sh1": {
  "input_fingerprint": "9616a5b609a60cb7",
  "prompt": "Create ONE FINAL photorealistic live-action film still of the moment below — early 20th-century Korea, Jeolla region; all people are Korean unless explicitly stated otherwise. TIME OF DAY (lock): night.\n\nSHOT TEXT (authoritative, Korean): 어두운 주재소 집무실, 벽 쪽에 주눅 들어 선 상처투성이 일본 헌병 둘과 흰 천이 덮인 시신, 그리고 굳어 있는 배상도의 전경.\n\nLOCATION (lock): The open floor beside the wall inside a dim police outpost office at night, accommodating two survivors and a covered body. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: wide shot\n- KEY BACKGROUND ELEMENTS: Office table (Documents are being handled by 토미이시) — An oblique view reveals the top and the entrance-facing edge; used as Establish the route toward the body while occupying less than two-fifths of the image; White covering cloth (Still covering the dead military policeman); used as Give the lower center a sober visual anchor without revealing the face; Office wall (배상도 and the injured policemen wait beside it); used as Hold the figures in a shared background plane that can be recognized during the later assault; Documents (Being turned through; the later designation has not yet been written) — Page faces are angled upward, but their contents are not legible at this distance; used as Connect 토미이시's attention to the table without prematurely revealing the final text.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Keep the nighttime office dim with restrained saturation and readable shadow detail, introducing no unestablished visible light source.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The clothing, hair and overall look of 배상도 — who appear both in that photo and in this shot — are LOCKED to that photo. Anyone else visible in that photo is NOT in this shot: never carry their face, body or clothing onto anyone here. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nIMMOBILE CHARACTER POSE — CANONICAL (identical wherever this character appears in ANY panel; on any conflict THIS POSE WINS): His motionless corpse is covered with a white cloth in the office. The screenplay does not identify the supporting surface, whether the body is face-up or face-down, or the positions of the head, arms, and legs beneath the cloth.\n\nIMMOBILE BODIES OBEY GRAVITY: a person who is dead or unconscious\nexerts NO muscular effort. Every part of their body — head, torso,\narms, hands, fingers, legs — rests fully on whatever supports it\n(floor, wall, furniture, their own lap) and hangs or slumps with\ngravity. NEVER show any part of an immobile person's body lifted,\nraised, held up in the air, or posed as if presenting something:\nan object in their grip stays clenched in a hand that itself lies\nfallen on a support — the hand does not hold the object up. If the\ncanonical pose leaves a body part unspecified, resolve it as the\nmost gravity-compliant, fully-supported position.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): Documents lie on the office table at night; a white shroud covers the brought-in body. 배상도: He stands rigidly by the wall.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 배상도 (한국인 남성, 성인 나이대, 자연스러운 검은 머리) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.",
  "roll_prompts": {
   "A": "Create ONE FINAL photorealistic live-action film still of the moment below — early 20th-century Korea, Jeolla region; all people are Korean unless explicitly stated otherwise. TIME OF DAY (lock): night.\n\nSHOT TEXT (authoritative, Korean): 어두운 주재소 집무실, 벽 쪽에 주눅 들어 선 상처투성이 일본 헌병 둘과 흰 천이 덮인 시신, 그리고 굳어 있는 배상도의 전경.\n\nLOCATION (lock): The open floor beside the wall inside a dim police outpost office at night, accommodating two survivors and a covered body. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: wide shot\n- KEY BACKGROUND ELEMENTS: Office table (Documents are being handled by 토미이시) — An oblique view reveals the top and the entrance-facing edge; used as Establish the route toward the body while occupying less than two-fifths of the image; White covering cloth (Still covering the dead military policeman); used as Give the lower center a sober visual anchor without revealing the face; Office wall (배상도 and the injured policemen wait beside it); used as Hold the figures in a shared background plane that can be recognized during the later assault; Documents (Being turned through; the later designation has not yet been written) — Page faces are angled upward, but their contents are not legible at this distance; used as Connect 토미이시's attention to the table without prematurely revealing the final text.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Keep the nighttime office dim with restrained saturation and readable shadow detail, introducing no unestablished visible light source.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The clothing, hair and overall look of 배상도 — who appear both in that photo and in this shot — are LOCKED to that photo. Anyone else visible in that photo is NOT in this shot: never carry their face, body or clothing onto anyone here. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nIMMOBILE CHARACTER POSE — CANONICAL (identical wherever this character appears in ANY panel; on any conflict THIS POSE WINS): His motionless corpse is covered with a white cloth in the office. The screenplay does not identify the supporting surface, whether the body is face-up or face-down, or the positions of the head, arms, and legs beneath the cloth.\n\nIMMOBILE BODIES OBEY GRAVITY: a person who is dead or unconscious\nexerts NO muscular effort. Every part of their body — head, torso,\narms, hands, fingers, legs — rests fully on whatever supports it\n(floor, wall, furniture, their own lap) and hangs or slumps with\ngravity. NEVER show any part of an immobile person's body lifted,\nraised, held up in the air, or posed as if presenting something:\nan object in their grip stays clenched in a hand that itself lies\nfallen on a support — the hand does not hold the object up. If the\ncanonical pose leaves a body part unspecified, resolve it as the\nmost gravity-compliant, fully-supported position.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): Documents lie on the office table at night; a white shroud covers the brought-in body. 배상도: He stands rigidly by the wall.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 배상도 (한국인 남성, 성인 나이대, 자연스러운 검은 머리) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.",
   "B": "Create ONE FINAL photorealistic live-action film still of the moment below — early 20th-century Korea, Jeolla region; all people are Korean unless explicitly stated otherwise. TIME OF DAY (lock): night.\n\nSHOT TEXT (authoritative, Korean): 어두운 주재소 집무실, 벽 쪽에 주눅 들어 선 상처투성이 일본 헌병 둘과 흰 천이 덮인 시신, 그리고 굳어 있는 배상도의 전경.\n\nLOCATION (lock): The open floor beside the wall inside a dim police outpost office at night, accommodating two survivors and a covered body. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: wide shot\n- KEY BACKGROUND ELEMENTS: Office table (Documents are being handled by 토미이시) — An oblique view reveals the top and the entrance-facing edge; used as Establish the route toward the body while occupying less than two-fifths of the image; White covering cloth (Still covering the dead military policeman); used as Give the lower center a sober visual anchor without revealing the face; Office wall (배상도 and the injured policemen wait beside it); used as Hold the figures in a shared background plane that can be recognized during the later assault; Documents (Being turned through; the later designation has not yet been written) — Page faces are angled upward, but their contents are not legible at this distance; used as Connect 토미이시's attention to the table without prematurely revealing the final text.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Keep the nighttime office dim with restrained saturation and readable shadow detail, introducing no unestablished visible light source.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The clothing, hair and overall look of 배상도 — who appear both in that photo and in this shot — are LOCKED to that photo. Anyone else visible in that photo is NOT in this shot: never carry their face, body or clothing onto anyone here. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nIMMOBILE CHARACTER POSE — CANONICAL (identical wherever this character appears in ANY panel; on any conflict THIS POSE WINS): His motionless corpse is covered with a white cloth in the office. The screenplay does not identify the supporting surface, whether the body is face-up or face-down, or the positions of the head, arms, and legs beneath the cloth.\n\nIMMOBILE BODIES OBEY GRAVITY: a person who is dead or unconscious\nexerts NO muscular effort. Every part of their body — head, torso,\narms, hands, fingers, legs — rests fully on whatever supports it\n(floor, wall, furniture, their own lap) and hangs or slumps with\ngravity. NEVER show any part of an immobile person's body lifted,\nraised, held up in the air, or posed as if presenting something:\nan object in their grip stays clenched in a hand that itself lies\nfallen on a support — the hand does not hold the object up. If the\ncanonical pose leaves a body part unspecified, resolve it as the\nmost gravity-compliant, fully-supported position.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): Documents lie on the office table at night; a white shroud covers the brought-in body. 배상도: He stands rigidly by the wall.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 배상도 (한국인 남성, 성인 나이대, 자연스러운 검은 머리) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.\n\nCOMPOSITION VARIATION (this candidate only): keep the place, the people and the moment exactly as specified, and keep from the CAMERA & FRAME contract its subject scale, screen placement and key background elements exactly. For this candidate only, its camera angle and height are deliberately left open: choose a camera position distinctly different from the obvious one, as a film director picking a second setup on the same blocking. Stay true to any explicit framing the SHOT TEXT itself dictates."
  },
  "shot_run_spend_attempt_count": 1,
  "readings": [
   {
    "label": "A",
    "direction": "토미이시의 시선과 두 손은 책상 위 문서로 향하고 있음. 벽 쪽에 선 배상도와 두 일본 헌병의 시선은 굳은 채로 정면 또는 아래를 향함.",
    "built_space": "이전 샷과 동일한 주재소 집무실 내부. 왼쪽에 책상, 호롱불, 주판 등이 위치하고 토미이시가 의자에 앉아 있음. 뒤쪽 벽면을 따라 배상도와 헌병들이 서 있으며 공간적 배치가 지시문과 일치함.",
    "entities": "토미이시(일본 장교), 상처투성이 헌병 2명, 흰 천이 덮인 시신 모두 프롬프트 설명에 완벽히 부합함. 배상도는 레퍼런스와 얼굴은 일치하나 이전 샷에서 착용하지 않았던 모자를 쓰고 있음. 문서 내용은 의도대로 흐릿하게 처리됨.",
    "hard_violations": [
     "[gpt-high] 이전 숏에서 확립된 책상 하나의 집무 공간을 별개의 책상 두 개로 늘려 고정된 장소 구성을 변경했다.",
     "[gpt-high] 벽패에 판독 가능한 한자가 남아 있어, 화면 어디에도 읽을 수 있는 글자를 두지 말라는 명시적 금지에 위배된다."
    ],
    "physics": "모든 인물이 바닥과 의자에 안정적으로 지지되어 있으며, 문서와 시신 모두 중력에 맞게 자연스럽게 놓이거나 들려 있음."
   },
   {
    "label": "B",
    "direction": "토미이시는 책상 뒤에 서서 손에 든 문서를 내려다봄. 배상도는 정면을 응시하고, 헌병들은 눈을 감고 고개를 숙인 채 서 있음.",
    "built_space": "주재소 집무실 내부 구조와 책상 위 소품들이 이전 샷과 잘 일치함. 인물들은 프롬프트 요구대로 벽 쪽에 나란히 서서 배경을 공유함.",
    "entities": "배상도가 모자를 쓰지 않아 이전 샷의 상태와 완벽히 연결됨. 헌병들과 시신도 적절히 묘사되었으나, 토미이시가 들고 있는 문서에 선명한 글꼴과 원형 다이어그램이 생성되어 '내용을 읽을 수 없어야 한다'는 지시를 심각하게 위반함.",
    "hard_violations": [
     "[gemini-pro] leaked markers/diagrams/text: 프롬프트에서 명시적으로 금지한 '읽을 수 있는 글씨'와 다이어그램이 문서 표면에 선명하게 생성됨.",
     "[gpt-high] 벽패에 판독 가능한 한자가 남아 있어, 화면 어디에도 읽을 수 있는 글자를 두지 말라는 명시적 금지에 위배된다."
    ],
    "physics": "서 있는 인물들의 발과 바닥의 접촉이 자연스러우며, 토미이시가 두 손으로 문서를 받치고 있는 자세 등 물리적 오류가 없음."
   }
  ],
  "cross_model_order": {
   "policy": "cross_model_order_v2_slot0_fwd_slot1_rev_parallel2",
   "slots": [
    {
     "model": "gemini-pro",
     "order": "forward",
     "display_to_canonical": {
      "A": "A",
      "B": "B"
     },
     "raw": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 8,
        "verdict_ko": "문서의 내용을 흐릿하게 처리하여 '읽을 수 없게 연출하라'는 지시를 훌륭히 따랐으나, 배상도가 이전 샷(모자 미착용)과 달리 모자를 쓰고 등장한 점이 유일한 감점 요인입니다."
       },
       {
        "label": "B",
        "score": 4,
        "verdict_ko": "배상도의 복장은 이전 샷과 잘 일치하지만, 문서 표면에 선명하고 읽을 수 있는 글씨와 다이어그램이 그대로 노출되어 프롬프트의 가장 중요한 묘사 및 금지 지시를 정면으로 위반했습니다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "토미이시의 시선과 두 손은 책상 위 문서로 향하고 있음. 벽 쪽에 선 배상도와 두 일본 헌병의 시선은 굳은 채로 정면 또는 아래를 향함.",
        "built_space": "이전 샷과 동일한 주재소 집무실 내부. 왼쪽에 책상, 호롱불, 주판 등이 위치하고 토미이시가 의자에 앉아 있음. 뒤쪽 벽면을 따라 배상도와 헌병들이 서 있으며 공간적 배치가 지시문과 일치함.",
        "entities": "토미이시(일본 장교), 상처투성이 헌병 2명, 흰 천이 덮인 시신 모두 프롬프트 설명에 완벽히 부합함. 배상도는 레퍼런스와 얼굴은 일치하나 이전 샷에서 착용하지 않았던 모자를 쓰고 있음. 문서 내용은 의도대로 흐릿하게 처리됨.",
        "hard_violations": [],
        "physics": "모든 인물이 바닥과 의자에 안정적으로 지지되어 있으며, 문서와 시신 모두 중력에 맞게 자연스럽게 놓이거나 들려 있음."
       },
       {
        "label": "B",
        "direction": "토미이시는 책상 뒤에 서서 손에 든 문서를 내려다봄. 배상도는 정면을 응시하고, 헌병들은 눈을 감고 고개를 숙인 채 서 있음.",
        "built_space": "주재소 집무실 내부 구조와 책상 위 소품들이 이전 샷과 잘 일치함. 인물들은 프롬프트 요구대로 벽 쪽에 나란히 서서 배경을 공유함.",
        "entities": "배상도가 모자를 쓰지 않아 이전 샷의 상태와 완벽히 연결됨. 헌병들과 시신도 적절히 묘사되었으나, 토미이시가 들고 있는 문서에 선명한 글꼴과 원형 다이어그램이 생성되어 '내용을 읽을 수 없어야 한다'는 지시를 심각하게 위반함.",
        "hard_violations": [
         "leaked markers/diagrams/text: 프롬프트에서 명시적으로 금지한 '읽을 수 있는 글씨'와 다이어그램이 문서 표면에 선명하게 생성됨."
        ],
        "physics": "서 있는 인물들의 발과 바닥의 접촉이 자연스러우며, 토미이시가 두 손으로 문서를 받치고 있는 자세 등 물리적 오류가 없음."
       }
      ],
      "all_candidates_fail": false
     },
     "normalized": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 8,
        "verdict_ko": "문서의 내용을 흐릿하게 처리하여 '읽을 수 없게 연출하라'는 지시를 훌륭히 따랐으나, 배상도가 이전 샷(모자 미착용)과 달리 모자를 쓰고 등장한 점이 유일한 감점 요인입니다."
       },
       {
        "label": "B",
        "score": 4,
        "verdict_ko": "배상도의 복장은 이전 샷과 잘 일치하지만, 문서 표면에 선명하고 읽을 수 있는 글씨와 다이어그램이 그대로 노출되어 프롬프트의 가장 중요한 묘사 및 금지 지시를 정면으로 위반했습니다."
       }
      ],
      "all_candidates_fail": false,
      "readings": [
       {
        "label": "A",
        "direction": "토미이시의 시선과 두 손은 책상 위 문서로 향하고 있음. 벽 쪽에 선 배상도와 두 일본 헌병의 시선은 굳은 채로 정면 또는 아래를 향함.",
        "built_space": "이전 샷과 동일한 주재소 집무실 내부. 왼쪽에 책상, 호롱불, 주판 등이 위치하고 토미이시가 의자에 앉아 있음. 뒤쪽 벽면을 따라 배상도와 헌병들이 서 있으며 공간적 배치가 지시문과 일치함.",
        "entities": "토미이시(일본 장교), 상처투성이 헌병 2명, 흰 천이 덮인 시신 모두 프롬프트 설명에 완벽히 부합함. 배상도는 레퍼런스와 얼굴은 일치하나 이전 샷에서 착용하지 않았던 모자를 쓰고 있음. 문서 내용은 의도대로 흐릿하게 처리됨.",
        "hard_violations": [],
        "physics": "모든 인물이 바닥과 의자에 안정적으로 지지되어 있으며, 문서와 시신 모두 중력에 맞게 자연스럽게 놓이거나 들려 있음."
       },
       {
        "label": "B",
        "direction": "토미이시는 책상 뒤에 서서 손에 든 문서를 내려다봄. 배상도는 정면을 응시하고, 헌병들은 눈을 감고 고개를 숙인 채 서 있음.",
        "built_space": "주재소 집무실 내부 구조와 책상 위 소품들이 이전 샷과 잘 일치함. 인물들은 프롬프트 요구대로 벽 쪽에 나란히 서서 배경을 공유함.",
        "entities": "배상도가 모자를 쓰지 않아 이전 샷의 상태와 완벽히 연결됨. 헌병들과 시신도 적절히 묘사되었으나, 토미이시가 들고 있는 문서에 선명한 글꼴과 원형 다이어그램이 생성되어 '내용을 읽을 수 없어야 한다'는 지시를 심각하게 위반함.",
        "hard_violations": [
         "leaked markers/diagrams/text: 프롬프트에서 명시적으로 금지한 '읽을 수 있는 글씨'와 다이어그램이 문서 표면에 선명하게 생성됨."
        ],
        "physics": "서 있는 인물들의 발과 바닥의 접촉이 자연스러우며, 토미이시가 두 손으로 문서를 받치고 있는 자세 등 물리적 오류가 없음."
       }
      ]
     },
     "ok": true
    },
    {
     "model": "gpt-high",
     "order": "reverse",
     "display_to_canonical": {
      "A": "B",
      "B": "A"
     },
     "raw": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 4,
        "verdict_ko": "단일 책상과 벽가 인물들의 와이드 배치는 더 충실하지만, 벽패에 읽을 수 있는 문자가 남아 있어 최종 사용에는 부적합하다."
       },
       {
        "label": "B",
        "score": 2,
        "verdict_ko": "부상자들의 위축된 연기는 적절하지만, 책상을 두 개로 늘린 공간 변형과 읽히는 벽패가 치명적이며 배상도의 모자도 이전 숏과 다르다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "왼쪽의 토미이시는 손으로 넘기는 문서를 내려다보고 있으며, 펼친 페이지도 자신의 눈 쪽으로 기울어 있다. 배상도는 왼쪽 아래의 책상·시신 방향을 바라본다. 오른쪽 헌병 둘은 고개를 기울이거나 숙여 시선을 낮추고 있다. 카메라를 정면으로 응시하는 인물이나 겨누는 무기는 없다.",
        "built_space": "목재 골조와 낡은 회벽, 격자창 하나, 벽패 하나가 보인다. 왼쪽 전경에 책상 하나가 있고 상판과 앞면이 사선으로 드러나며 화면의 약 5분의 2 미만을 차지한다. 책상 위에는 유리 석유등 하나, 주판 하나, 인주통 하나, 도장 하나와 여러 문서 묶음이 있다. 창 아래 등롱과 의자는 이 구도에서 확인되지 않는다. 배상도와 부상자 둘은 오른쪽 벽 앞의 같은 깊이에 서 있고, 덮인 시신은 그들 앞 바닥의 하단 중앙에서 오른쪽으로 놓여 있다. 책상에서 시신으로 이어지는 공간이 확보되어 있다. 반사상은 없다.",
        "entities": "살아 있는 성인 남성 네 명과 흰 천으로 완전히 덮인 시신 한 구가 보인다. 토미이시는 구도 지시에서 명시된 문서 취급 역할을 수행한다. 헌병 둘은 동아시아계 외양에 군복을 입었으며 얼굴 상처와 붕대가 있지만, 국적 자체는 외양만으로 확인할 수 없다. 배상도는 젊은 동아시아계 남성으로 검은 가르마 머리, 흰옷, 검은 조끼와 허리 장식이 이전 숏에 가깝고 모자를 쓰지 않았다. 시신의 얼굴은 노출되지 않았다. 밤의 어둡고 절제된 색감은 맞지만 벽패의 한자들은 판독 가능하다.",
        "hard_violations": [
         "벽패에 판독 가능한 한자가 남아 있어, 화면 어디에도 읽을 수 있는 글자를 두지 말라는 명시적 금지에 위배된다."
        ],
        "physics": "서 있는 네 사람은 바닥에 체중을 싣고 있으며 부상자들의 기울어진 머리와 처진 팔도 가능한 자세다. 토미이시의 양손이 문서를 잡고 나머지 종이는 책상에 받쳐져 있다. 시신은 바닥에 누워 있고 흰 천은 몸의 윤곽과 바닥에 걸쳐 늘어진다. 오른쪽 천의 돌출은 덮인 발 부위로 설명 가능하다. 지지 없이 떠 있는 신체나 물체는 없다."
       },
       {
        "label": "B",
        "direction": "왼쪽에 앉은 토미이시는 손에 든 페이지를 내려다보며, 문서는 그가 읽을 수 있는 방향으로 펼쳐져 있다. 배상도는 왼쪽 아래를 보고, 헌병 둘은 고개를 깊이 숙여 시신 주변 바닥을 향한다. 부상자들의 시선과 몸짓은 주눅 든 상태에 부합한다. 겨누는 무기나 이동 중인 몸은 없다.",
        "built_space": "낡은 회벽과 목재 골조, 격자창 하나, 벽패 하나, 창 아래 등롱 하나가 보인다. 그러나 전경의 문서 작업 책상과 뒤쪽의 석유등이 놓인 책상이 별개로 두 개 있으며, 토미이시가 앉은 의자와 뒤쪽 빈 의자도 각각 하나씩 보인다. 이전 숏의 단일 책상 공간이 두 책상 공간으로 바뀌었다. 석유등 하나, 주판 하나, 도장 하나, 인주통 하나와 문서 묶음들이 두 책상으로 나뉘어 있다. 배상도와 헌병 둘은 벽 앞에 나란히 서고, 시신은 앞쪽 바닥에 놓여 있다. 반사상은 없다.",
        "entities": "살아 있는 성인 남성 네 명과 천으로 덮인 시신 한 구가 보인다. 군복을 입은 헌병 둘의 얼굴과 옷에 상처·혈흔·붕대가 뚜렷하다. 이들은 동아시아계 외양이지만 일본인 여부는 시각적으로 확정할 수 없다. 배상도의 흰옷과 검은 조끼, 허리 장식은 맞으나, 머리에 쓴 검은 모자는 모자를 벗은 이전 숏의 외양 잠금과 다르다. 문서 취급 인물은 토미이시의 역할에 해당한다. 시신의 얼굴은 가려져 있고 야간 분위기도 유지된다. 벽패에는 읽을 수 있는 한자가 있다.",
        "hard_violations": [
         "이전 숏에서 확립된 책상 하나의 집무 공간을 별개의 책상 두 개로 늘려 고정된 장소 구성을 변경했다.",
         "벽패에 판독 가능한 한자가 남아 있어, 화면 어디에도 읽을 수 있는 글자를 두지 말라는 명시적 금지에 위배된다."
        ],
        "physics": "토미이시는 의자 좌판에 앉고 등받이는 몸 뒤에 있으며, 양손으로 책상 위 문서의 페이지를 넘긴다. 벽가 세 사람은 발을 바닥에 붙이고 서 있으며 헌병들의 숙인 상체와 늘어진 팔도 중력에 맞는다. 시신과 흰 천은 바닥에 지지되어 있고 천 끝도 바닥에 늘어져 있다. 책상 위 소품 역시 상판에 놓여 있다. 지지 없이 떠 있는 신체나 물체는 없다."
       }
      ],
      "all_candidates_fail": true
     },
     "normalized": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 4,
        "verdict_ko": "단일 책상과 벽가 인물들의 와이드 배치는 더 충실하지만, 벽패에 읽을 수 있는 문자가 남아 있어 최종 사용에는 부적합하다."
       },
       {
        "label": "A",
        "score": 2,
        "verdict_ko": "부상자들의 위축된 연기는 적절하지만, 책상을 두 개로 늘린 공간 변형과 읽히는 벽패가 치명적이며 배상도의 모자도 이전 숏과 다르다."
       }
      ],
      "all_candidates_fail": true,
      "readings": [
       {
        "label": "B",
        "direction": "왼쪽의 토미이시는 손으로 넘기는 문서를 내려다보고 있으며, 펼친 페이지도 자신의 눈 쪽으로 기울어 있다. 배상도는 왼쪽 아래의 책상·시신 방향을 바라본다. 오른쪽 헌병 둘은 고개를 기울이거나 숙여 시선을 낮추고 있다. 카메라를 정면으로 응시하는 인물이나 겨누는 무기는 없다.",
        "built_space": "목재 골조와 낡은 회벽, 격자창 하나, 벽패 하나가 보인다. 왼쪽 전경에 책상 하나가 있고 상판과 앞면이 사선으로 드러나며 화면의 약 5분의 2 미만을 차지한다. 책상 위에는 유리 석유등 하나, 주판 하나, 인주통 하나, 도장 하나와 여러 문서 묶음이 있다. 창 아래 등롱과 의자는 이 구도에서 확인되지 않는다. 배상도와 부상자 둘은 오른쪽 벽 앞의 같은 깊이에 서 있고, 덮인 시신은 그들 앞 바닥의 하단 중앙에서 오른쪽으로 놓여 있다. 책상에서 시신으로 이어지는 공간이 확보되어 있다. 반사상은 없다.",
        "entities": "살아 있는 성인 남성 네 명과 흰 천으로 완전히 덮인 시신 한 구가 보인다. 토미이시는 구도 지시에서 명시된 문서 취급 역할을 수행한다. 헌병 둘은 동아시아계 외양에 군복을 입었으며 얼굴 상처와 붕대가 있지만, 국적 자체는 외양만으로 확인할 수 없다. 배상도는 젊은 동아시아계 남성으로 검은 가르마 머리, 흰옷, 검은 조끼와 허리 장식이 이전 숏에 가깝고 모자를 쓰지 않았다. 시신의 얼굴은 노출되지 않았다. 밤의 어둡고 절제된 색감은 맞지만 벽패의 한자들은 판독 가능하다.",
        "hard_violations": [
         "벽패에 판독 가능한 한자가 남아 있어, 화면 어디에도 읽을 수 있는 글자를 두지 말라는 명시적 금지에 위배된다."
        ],
        "physics": "서 있는 네 사람은 바닥에 체중을 싣고 있으며 부상자들의 기울어진 머리와 처진 팔도 가능한 자세다. 토미이시의 양손이 문서를 잡고 나머지 종이는 책상에 받쳐져 있다. 시신은 바닥에 누워 있고 흰 천은 몸의 윤곽과 바닥에 걸쳐 늘어진다. 오른쪽 천의 돌출은 덮인 발 부위로 설명 가능하다. 지지 없이 떠 있는 신체나 물체는 없다."
       },
       {
        "label": "A",
        "direction": "왼쪽에 앉은 토미이시는 손에 든 페이지를 내려다보며, 문서는 그가 읽을 수 있는 방향으로 펼쳐져 있다. 배상도는 왼쪽 아래를 보고, 헌병 둘은 고개를 깊이 숙여 시신 주변 바닥을 향한다. 부상자들의 시선과 몸짓은 주눅 든 상태에 부합한다. 겨누는 무기나 이동 중인 몸은 없다.",
        "built_space": "낡은 회벽과 목재 골조, 격자창 하나, 벽패 하나, 창 아래 등롱 하나가 보인다. 그러나 전경의 문서 작업 책상과 뒤쪽의 석유등이 놓인 책상이 별개로 두 개 있으며, 토미이시가 앉은 의자와 뒤쪽 빈 의자도 각각 하나씩 보인다. 이전 숏의 단일 책상 공간이 두 책상 공간으로 바뀌었다. 석유등 하나, 주판 하나, 도장 하나, 인주통 하나와 문서 묶음들이 두 책상으로 나뉘어 있다. 배상도와 헌병 둘은 벽 앞에 나란히 서고, 시신은 앞쪽 바닥에 놓여 있다. 반사상은 없다.",
        "entities": "살아 있는 성인 남성 네 명과 천으로 덮인 시신 한 구가 보인다. 군복을 입은 헌병 둘의 얼굴과 옷에 상처·혈흔·붕대가 뚜렷하다. 이들은 동아시아계 외양이지만 일본인 여부는 시각적으로 확정할 수 없다. 배상도의 흰옷과 검은 조끼, 허리 장식은 맞으나, 머리에 쓴 검은 모자는 모자를 벗은 이전 숏의 외양 잠금과 다르다. 문서 취급 인물은 토미이시의 역할에 해당한다. 시신의 얼굴은 가려져 있고 야간 분위기도 유지된다. 벽패에는 읽을 수 있는 한자가 있다.",
        "hard_violations": [
         "이전 숏에서 확립된 책상 하나의 집무 공간을 별개의 책상 두 개로 늘려 고정된 장소 구성을 변경했다.",
         "벽패에 판독 가능한 한자가 남아 있어, 화면 어디에도 읽을 수 있는 글자를 두지 말라는 명시적 금지에 위배된다."
        ],
        "physics": "토미이시는 의자 좌판에 앉고 등받이는 몸 뒤에 있으며, 양손으로 책상 위 문서의 페이지를 넘긴다. 벽가 세 사람은 발을 바닥에 붙이고 서 있으며 헌병들의 숙인 상체와 늘어진 팔도 중력에 맞는다. 시신과 흰 천은 바닥에 지지되어 있고 천 끝도 바닥에 늘어져 있다. 책상 위 소품 역시 상판에 놓여 있다. 지지 없이 떠 있는 신체나 물체는 없다."
       }
      ]
     },
     "ok": true
    }
   ],
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "slot_winner_match": false,
   "slot_winner": {
    "gemini-pro": "A",
    "gpt-high": "B"
   },
   "route": "cross_slot_combined"
  },
  "dual": {
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "normalized": {
    "A": 1.5,
    "B": 1.5
   },
   "adjusted": {
    "A": 1.25,
    "B": 1.25
   },
   "violations": {
    "B": [
     "[gemini-pro] leaked markers/diagrams/text: 프롬프트에서 명시적으로 금지한 '읽을 수 있는 글씨'와 다이어그램이 문서 표면에 선명하게 생성됨.",
     "[gpt-high] 벽패에 판독 가능한 한자가 남아 있어, 화면 어디에도 읽을 수 있는 글자를 두지 말라는 명시적 금지에 위배된다."
    ],
    "A": [
     "[gpt-high] 이전 숏에서 확립된 책상 하나의 집무 공간을 별개의 책상 두 개로 늘려 고정된 장소 구성을 변경했다.",
     "[gpt-high] 벽패에 판독 가능한 한자가 남아 있어, 화면 어디에도 읽을 수 있는 글자를 두지 말라는 명시적 금지에 위배된다."
    ]
   },
   "per_model_winner": {
    "gemini-pro": "A",
    "gpt-high": "B"
   },
   "agreed": false
  },
  "totals": {
   "A": 1250,
   "B": 1250
  },
  "selected": "A",
  "ranking": [
   "A",
   "B"
  ],
  "verdicts": [
   {
    "label": "A",
    "score": 1250,
    "verdict_ko": "문서의 내용을 흐릿하게 처리하여 '읽을 수 없게 연출하라'는 지시를 훌륭히 따랐으나, 배상도가 이전 샷(모자 미착용)과 달리 모자를 쓰고 등장한 점이 유일한 감점 요인입니다.  ★위반: [gpt-high] 이전 숏에서 확립된 책상 하나의 집무 공간을 별개의 책상 두 개로 늘려 고정된 장소 구성을 변경했다. / [gpt-high] 벽패에 판독 가능한 한자가 남아 있어, 화면 어디에도 읽을 수 있는 글자를 두지 말라는 명시적 금지에 위배된다."
   },
   {
    "label": "B",
    "score": 1250,
    "verdict_ko": "배상도의 복장은 이전 샷과 잘 일치하지만, 문서 표면에 선명하고 읽을 수 있는 글씨와 다이어그램이 그대로 노출되어 프롬프트의 가장 중요한 묘사 및 금지 지시를 정면으로 위반했습니다.  ★위반: [gemini-pro] leaked markers/diagrams/text: 프롬프트에서 명시적으로 금지한 '읽을 수 있는 글씨'와 다이어그램이 문서 표면에 선명하게 생성됨. / [gpt-high] 벽패에 판독 가능한 한자가 남아 있어, 화면 어디에도 읽을 수 있는 글자를 두지 말라는 명시적 금지에 위배된다."
   }
  ],
  "refs": [
   {
    "label": "PREVIOUS SHOT STILL — a visually related earlier shot of this same place: the location's look, materials, fixed features and lighting mood are LOCKED to this photo; never copy its camera framing. The clothing, hair and overall look of 배상도 — who appear both in that photo and in this shot — are LOCKED to that photo. Anyone else visible in that photo is NOT in this shot: never carry their face, body or clothing onto anyone here. If this photo shows a character who cannot move (dead or unconscious), that character's exact pose, position and orientation are ALSO LOCKED — treat their body as a fixed prop of the set that only the camera moves around.",
    "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/db00517b-165b-4005-a287-f28936b59b95/images/272bd7af-9294-4490-98c6-88a40e9deacc/scene/recipe/S25sh3_sel.png",
    "asset_id": "e77b68ad-1828-4014-af71-a331a0663cbb",
    "role": "prev_still"
   },
   {
    "label": "CHARACTER REFERENCE — 배상도: the exact person appearing in this shot; match face, hair and build exactly.",
    "path": "<bytes:851718>",
    "asset_id": "034f5325-975e-4474-854b-e48359710bf8",
    "role": "character_ref"
   }
  ],
  "critique_skipped": true,
  "shot_run_produced": false,
  "shot_run_uid": "06aa1427-22eb-7d09-85d8-512ea93ef90a",
  "ref_mode": "prev+엔티티",
  "share_plan": {
   "ref_plan": "prev",
   "prev_anchor_tag": "S25sh3"
  }
 },
 "S32sh1::cine": {
  "applied": false,
  "attempted_at": "2026-09-09T09:08:05.922264+00:00",
  "fingerprint": "21c423d4cfaef73e8268c2cc7e0ca722b92978abbce70de627fc55d9685d21fe",
  "fingerprint_version": 2,
  "provider": "mai",
  "endpoint": "openrouter/chat-completions",
  "model": "microsoft/mai-image-2.6",
  "pack": "24.202608252115",
  "source_file": "S32sh1_sel.png",
  "source_sha256": "691084ad969e12835d84eebaf5fe5c22b3cf4f3d6082da6b86817f90dfa3e6a5",
  "error": "RuntimeError: moderation blocked: {\"message\": \"Response content blocked by label 'MultiSeverity_ViolenceScore'.\", \"code\": 400}",
  "moderation_refusals": 2,
  "declined": true,
  "declined_reason": "moderation"
 },
 "S32sh10::signage": {
  "fp": "2bfe8693ae0c4857",
  "inscriptions": [],
  "cues": [],
  "dropped": []
 },
 "S32sh10": {
  "input_fingerprint": "0d28f55d1e36d331",
  "prompt": "Create ONE FINAL photorealistic live-action film still of the moment below — early 20th-century Korea, Jeolla region; all people are Korean unless explicitly stated otherwise. TIME OF DAY (lock): night.\n\nSHOT TEXT (authoritative, Korean): 일본 헌병의 멱살을 움켜쥐고 벽에 강하게 밀어붙인 채 고정된 토미이시의 핏대 선 얼굴.\n\nLOCATION (lock): The wall beside the desk inside the dim nighttime police office, with standing space between the wall and furniture. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: close-up\n- KEY BACKGROUND ELEMENTS: Office wall (The military policeman is pinned against it); used as A narrow visible strip behind the shoulder proves physical confinement.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Maintain the office's dim nighttime exposure and sober, naturalistic contrast, allowing facial strain and the gripping hand to remain readable.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The people visible in that photo are NOT in this shot: never carry their faces, bodies or clothing onto anyone here. Each person in THIS shot is defined solely by the PEOPLE list, their own CHARACTER REFERENCE images and their wardrobe notes. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nPREVIOUS STILL USAGE (follow exactly — what to take from the attached still and what to exclude; it governs place and objects only, never who is in this shot or how the camera sees them): Take the office wall materials, adjacent furnishings, and subdued nighttime lighting. Exclude the untouched arrangement of papers or objects disturbed by the impact against the desk.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The shroud has been pulled back to expose the dead face; documents remain on the table. 토미이시: He stands in an aggressive gripping posture with his arms extended.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 토미이시 (일본인 남성, 성인 나이대, 자연스러운 검은 머리) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.",
  "roll_prompts": {
   "A": "Create ONE FINAL photorealistic live-action film still of the moment below — early 20th-century Korea, Jeolla region; all people are Korean unless explicitly stated otherwise. TIME OF DAY (lock): night.\n\nSHOT TEXT (authoritative, Korean): 일본 헌병의 멱살을 움켜쥐고 벽에 강하게 밀어붙인 채 고정된 토미이시의 핏대 선 얼굴.\n\nLOCATION (lock): The wall beside the desk inside the dim nighttime police office, with standing space between the wall and furniture. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: close-up\n- KEY BACKGROUND ELEMENTS: Office wall (The military policeman is pinned against it); used as A narrow visible strip behind the shoulder proves physical confinement.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Maintain the office's dim nighttime exposure and sober, naturalistic contrast, allowing facial strain and the gripping hand to remain readable.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The people visible in that photo are NOT in this shot: never carry their faces, bodies or clothing onto anyone here. Each person in THIS shot is defined solely by the PEOPLE list, their own CHARACTER REFERENCE images and their wardrobe notes. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nPREVIOUS STILL USAGE (follow exactly — what to take from the attached still and what to exclude; it governs place and objects only, never who is in this shot or how the camera sees them): Take the office wall materials, adjacent furnishings, and subdued nighttime lighting. Exclude the untouched arrangement of papers or objects disturbed by the impact against the desk.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The shroud has been pulled back to expose the dead face; documents remain on the table. 토미이시: He stands in an aggressive gripping posture with his arms extended.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 토미이시 (일본인 남성, 성인 나이대, 자연스러운 검은 머리) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.",
   "B": "Create ONE FINAL photorealistic live-action film still of the moment below — early 20th-century Korea, Jeolla region; all people are Korean unless explicitly stated otherwise. TIME OF DAY (lock): night.\n\nSHOT TEXT (authoritative, Korean): 일본 헌병의 멱살을 움켜쥐고 벽에 강하게 밀어붙인 채 고정된 토미이시의 핏대 선 얼굴.\n\nLOCATION (lock): The wall beside the desk inside the dim nighttime police office, with standing space between the wall and furniture. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: close-up\n- KEY BACKGROUND ELEMENTS: Office wall (The military policeman is pinned against it); used as A narrow visible strip behind the shoulder proves physical confinement.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Maintain the office's dim nighttime exposure and sober, naturalistic contrast, allowing facial strain and the gripping hand to remain readable.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The people visible in that photo are NOT in this shot: never carry their faces, bodies or clothing onto anyone here. Each person in THIS shot is defined solely by the PEOPLE list, their own CHARACTER REFERENCE images and their wardrobe notes. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nPREVIOUS STILL USAGE (follow exactly — what to take from the attached still and what to exclude; it governs place and objects only, never who is in this shot or how the camera sees them): Take the office wall materials, adjacent furnishings, and subdued nighttime lighting. Exclude the untouched arrangement of papers or objects disturbed by the impact against the desk.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The shroud has been pulled back to expose the dead face; documents remain on the table. 토미이시: He stands in an aggressive gripping posture with his arms extended.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 토미이시 (일본인 남성, 성인 나이대, 자연스러운 검은 머리) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.\n\nCOMPOSITION VARIATION (this candidate only): keep the place, the people and the moment exactly as specified, and keep from the CAMERA & FRAME contract its subject scale, screen placement and key background elements exactly. For this candidate only, its camera angle and height are deliberately left open: choose a camera position distinctly different from the obvious one, as a film director picking a second setup on the same blocking. Stay true to any explicit framing the SHOT TEXT itself dictates."
  },
  "shot_run_spend_attempt_count": 1,
  "readings": [
   {
    "label": "A",
    "direction": "토미이시의 시선은 화면 오른쪽 바깥을 강렬하게 응시하고 있다. 화면 우측 하단에서 뻗어 나온 손(오른손 형태)이 토미이시의 왼쪽 멱살(붉은 깃 부분)을 강하게 움켜쥐고 있다.",
    "built_space": "조명이 어두운 야간의 사무실 내부. 토미이시의 왼쪽 어깨 바로 뒤로 벽면이 좁게 보여 그가 벽에 밀어붙여진 상태임을 증명한다. 배경에는 레퍼런스와 동일한 창살 있는 창문과 불이 켜진 램프가 정확히 배치되어 있다.",
    "entities": "프레임 중앙에 보이는 얼굴은 콧수염과 턱수염, 이목구비가 캐릭터 레퍼런스의 토미이시와 완벽히 일치하며, 착용하고 있는 갈색 헌병 제복 역시 지시사항과 부합한다.",
    "hard_violations": [
     "[gpt-high] 토미이시가 상대 헌병을 벽에 고정해야 하는데, 토미이시로 읽히는 군복 남자 자신이 벽에 붙어 자기 목 아래 옷깃을 잡힌 구도로 배치되어 가해자와 피압박자의 핵심 위치·행동 관계가 뒤바뀐다."
    ],
    "physics": "인물은 두 발로 서서 벽에 기댄 채 신체를 지탱하고 있으며, 멱살을 꽉 틀어쥔 손에 의해 팽팽한 힘이 가해져 자세가 고정되어 있다. 허공에 떠 있거나 물리적으로 지탱되지 않은 요소는 없다."
   },
   {
    "label": "B",
    "direction": "왼쪽에 있는 남자의 시선이 오른쪽 헌병의 얼굴을 향해 똑바로 꽂히고 있으며, 남자의 양손은 헌병의 멱살을 쥐고 벽 쪽으로 밀어붙이고 있다.",
    "built_space": "어두운 사무실 실내. 오른쪽 헌병의 등 뒤로 벽면이 위치해 물리적 압박감을 주며, 배경 좌측으로 램프와 창문이 있다. 두 인물이 가구와 벽 사이의 공간에 자연스럽게 자리 잡고 있다.",
    "entities": "왼쪽에서 멱살을 쥐고 있는 남자는 레퍼런스의 토미이시와 전혀 다른 얼굴(수염 없음)과 사복(검은 양복)을 착용한 창조된 인물이다. 오른쪽 인물은 일본 헌병 제복을 입고 있으나 얼굴이 명확히 드러나지 않는다.",
    "hard_violations": [
     "[gemini-pro] invented people (프롬프트에 허용되지 않은 완전히 새로운 외모와 복장의 인물이 등장함)"
    ],
    "physics": "두 사람 모두 바닥에 발을 단단히 딛고 서로의 힘과 체중을 이용해 신체를 지탱하고 있으며, 멱살을 잡고 벽으로 밀어붙이는 물리적 상호작용이 자연스럽다."
   }
  ],
  "cross_model_order": {
   "policy": "cross_model_order_v2_slot0_fwd_slot1_rev_parallel2",
   "slots": [
    {
     "model": "gemini-pro",
     "order": "forward",
     "display_to_canonical": {
      "A": "A",
      "B": "B"
     },
     "raw": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 7,
        "verdict_ko": "지시된 클로즈업 프레이밍, 좁은 배경 벽의 배치, 그리고 토미이시의 캐릭터 레퍼런스(얼굴과 제복)를 매우 정확하게 구현했으나, 주체적인 멱살잡이 행동이 오히려 자신이 멱살을 잡힌(혹은 스스로 잡은) 수동적 형태로 어색하게 연출된 점이 감점 요인입니다."
       },
       {
        "label": "B",
        "score": 2,
        "verdict_ko": "지시된 클로즈업을 무시하고 샷을 넓게 잡았으며, 무엇보다 프롬프트에 명시되지 않은 전혀 새로운 외모와 사복 차림의 인물을 메인 주체로 등장시켜 치명적인 규칙 위반(invented people)을 범했습니다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "토미이시의 시선은 화면 오른쪽 바깥을 강렬하게 응시하고 있다. 화면 우측 하단에서 뻗어 나온 손(오른손 형태)이 토미이시의 왼쪽 멱살(붉은 깃 부분)을 강하게 움켜쥐고 있다.",
        "built_space": "조명이 어두운 야간의 사무실 내부. 토미이시의 왼쪽 어깨 바로 뒤로 벽면이 좁게 보여 그가 벽에 밀어붙여진 상태임을 증명한다. 배경에는 레퍼런스와 동일한 창살 있는 창문과 불이 켜진 램프가 정확히 배치되어 있다.",
        "entities": "프레임 중앙에 보이는 얼굴은 콧수염과 턱수염, 이목구비가 캐릭터 레퍼런스의 토미이시와 완벽히 일치하며, 착용하고 있는 갈색 헌병 제복 역시 지시사항과 부합한다.",
        "hard_violations": [],
        "physics": "인물은 두 발로 서서 벽에 기댄 채 신체를 지탱하고 있으며, 멱살을 꽉 틀어쥔 손에 의해 팽팽한 힘이 가해져 자세가 고정되어 있다. 허공에 떠 있거나 물리적으로 지탱되지 않은 요소는 없다."
       },
       {
        "label": "B",
        "direction": "왼쪽에 있는 남자의 시선이 오른쪽 헌병의 얼굴을 향해 똑바로 꽂히고 있으며, 남자의 양손은 헌병의 멱살을 쥐고 벽 쪽으로 밀어붙이고 있다.",
        "built_space": "어두운 사무실 실내. 오른쪽 헌병의 등 뒤로 벽면이 위치해 물리적 압박감을 주며, 배경 좌측으로 램프와 창문이 있다. 두 인물이 가구와 벽 사이의 공간에 자연스럽게 자리 잡고 있다.",
        "entities": "왼쪽에서 멱살을 쥐고 있는 남자는 레퍼런스의 토미이시와 전혀 다른 얼굴(수염 없음)과 사복(검은 양복)을 착용한 창조된 인물이다. 오른쪽 인물은 일본 헌병 제복을 입고 있으나 얼굴이 명확히 드러나지 않는다.",
        "hard_violations": [
         "invented people (프롬프트에 허용되지 않은 완전히 새로운 외모와 복장의 인물이 등장함)"
        ],
        "physics": "두 사람 모두 바닥에 발을 단단히 딛고 서로의 힘과 체중을 이용해 신체를 지탱하고 있으며, 멱살을 잡고 벽으로 밀어붙이는 물리적 상호작용이 자연스럽다."
       }
      ],
      "all_candidates_fail": false
     },
     "normalized": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 7,
        "verdict_ko": "지시된 클로즈업 프레이밍, 좁은 배경 벽의 배치, 그리고 토미이시의 캐릭터 레퍼런스(얼굴과 제복)를 매우 정확하게 구현했으나, 주체적인 멱살잡이 행동이 오히려 자신이 멱살을 잡힌(혹은 스스로 잡은) 수동적 형태로 어색하게 연출된 점이 감점 요인입니다."
       },
       {
        "label": "B",
        "score": 2,
        "verdict_ko": "지시된 클로즈업을 무시하고 샷을 넓게 잡았으며, 무엇보다 프롬프트에 명시되지 않은 전혀 새로운 외모와 사복 차림의 인물을 메인 주체로 등장시켜 치명적인 규칙 위반(invented people)을 범했습니다."
       }
      ],
      "all_candidates_fail": false,
      "readings": [
       {
        "label": "A",
        "direction": "토미이시의 시선은 화면 오른쪽 바깥을 강렬하게 응시하고 있다. 화면 우측 하단에서 뻗어 나온 손(오른손 형태)이 토미이시의 왼쪽 멱살(붉은 깃 부분)을 강하게 움켜쥐고 있다.",
        "built_space": "조명이 어두운 야간의 사무실 내부. 토미이시의 왼쪽 어깨 바로 뒤로 벽면이 좁게 보여 그가 벽에 밀어붙여진 상태임을 증명한다. 배경에는 레퍼런스와 동일한 창살 있는 창문과 불이 켜진 램프가 정확히 배치되어 있다.",
        "entities": "프레임 중앙에 보이는 얼굴은 콧수염과 턱수염, 이목구비가 캐릭터 레퍼런스의 토미이시와 완벽히 일치하며, 착용하고 있는 갈색 헌병 제복 역시 지시사항과 부합한다.",
        "hard_violations": [],
        "physics": "인물은 두 발로 서서 벽에 기댄 채 신체를 지탱하고 있으며, 멱살을 꽉 틀어쥔 손에 의해 팽팽한 힘이 가해져 자세가 고정되어 있다. 허공에 떠 있거나 물리적으로 지탱되지 않은 요소는 없다."
       },
       {
        "label": "B",
        "direction": "왼쪽에 있는 남자의 시선이 오른쪽 헌병의 얼굴을 향해 똑바로 꽂히고 있으며, 남자의 양손은 헌병의 멱살을 쥐고 벽 쪽으로 밀어붙이고 있다.",
        "built_space": "어두운 사무실 실내. 오른쪽 헌병의 등 뒤로 벽면이 위치해 물리적 압박감을 주며, 배경 좌측으로 램프와 창문이 있다. 두 인물이 가구와 벽 사이의 공간에 자연스럽게 자리 잡고 있다.",
        "entities": "왼쪽에서 멱살을 쥐고 있는 남자는 레퍼런스의 토미이시와 전혀 다른 얼굴(수염 없음)과 사복(검은 양복)을 착용한 창조된 인물이다. 오른쪽 인물은 일본 헌병 제복을 입고 있으나 얼굴이 명확히 드러나지 않는다.",
        "hard_violations": [
         "invented people (프롬프트에 허용되지 않은 완전히 새로운 외모와 복장의 인물이 등장함)"
        ],
        "physics": "두 사람 모두 바닥에 발을 단단히 딛고 서로의 힘과 체중을 이용해 신체를 지탱하고 있으며, 멱살을 잡고 벽으로 밀어붙이는 물리적 상호작용이 자연스럽다."
       }
      ]
     },
     "ok": true
    },
    {
     "model": "gpt-high",
     "order": "reverse",
     "display_to_canonical": {
      "A": "B",
      "B": "A"
     },
     "raw": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 6,
        "verdict_ko": "헌병의 멱살을 잡아 벽에 고정한 관계와 핏대 선 얼굴은 분명하지만, 토미이시의 얼굴·군복이 참조와 다르고 얼굴 중심 클로즈업보다 넓다."
       },
       {
        "label": "B",
        "score": 3,
        "verdict_ko": "얼굴 클로즈업과 군복은 더 가깝지만, 토미이시가 상대 헌병을 벽에 밀어붙이는 대신 자신이 벽에 몰려 옷깃을 잡힌 듯 보여 핵심 행동 관계가 뒤바뀐다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "왼쪽 남자는 오른쪽 헌병의 얼굴을 똑바로 노려보고, 손으로 헌병의 목 아래 옷깃을 움켜쥔다. 헌병은 턱을 뒤로 들고 왼쪽 남자를 향한다. 밀어붙이는 방향은 헌병 뒤의 오른쪽 벽으로 이어져 지시된 표적과 일치한다.",
        "built_space": "오른쪽에 회벽과 세로 목재 기둥이 있으며 헌병의 머리·등 뒤로 좁은 벽면이 보인다. 왼쪽 배경에는 격자창 하나, 창가 사각등 하나, 책상 일부와 그 위 유리 석유등 하나가 보인다. 회벽·어두운 목재·야간 조명은 장소 참조와 대체로 맞고, 두 사람은 가구 옆 벽 앞에 있다. 중복 설비나 반사는 보이지 않는다.",
        "entities": "성인 동아시아계 남성 두 명이 보이며, 외형만으로 일본 국적은 확인할 수 없다. 잡는 남자는 검은 가르마 머리와 짧은 수염이 있지만 토미이시 참조의 얼굴과 정확히 일치하지 않고, 참조의 갈색 군복 대신 짙은 민간 재킷과 밝은 속옷을 입었다. 잡힌 남자는 짧은 검은 머리, 콧수염, 견장과 붉은 깃이 있는 군복으로 헌병 역할을 읽을 수 있다. 읽을 수 있는 글자는 없다. 시신과 서류는 이 구도에서 확인되지 않는다.",
        "hard_violations": [],
        "physics": "잡는 손가락이 실제로 헌병의 옷깃을 감싸고 천이 당겨져 접힌다. 헌병의 뒤통수와 등은 벽에 밀착되어 압박을 받는 자세가 가능하다. 발은 화면 밖이지만 보이는 상체에 부유나 불가능한 지지는 없다. 다만 잡는 팔은 상당히 굽어 있어 팔을 뻗은 상태라는 연속성 지시에는 덜 맞는다."
       },
       {
        "label": "B",
        "direction": "군복 남자의 눈은 화면 오른쪽 가까운 바깥을 향하지만, 그 시선의 상대는 보이지 않는다. 손은 화면에 나온 남자 자신의 목 아래 군복 깃에 걸려 있다. 별도 헌병을 향한 뻗은 팔이나 그를 벽으로 미는 방향은 확인되지 않는다.",
        "built_space": "오른쪽 회벽에 화면의 유일한 남자 어깨가 붙어 있다. 왼쪽 배경에는 격자창 하나, 흐릿한 창가 조명 하나, 의자 등받이 하나, 책상 일부와 석유등 하나가 보인다. 재료와 어두운 야간 분위기는 참조에 가깝지만, 벽에 구속된 사람이 상대 헌병이라는 공간 관계가 성립하지 않는다. 중복 설비나 불가능한 반사는 보이지 않는다.",
        "entities": "짧은 검은 머리와 콧수염·턱수염을 가진 성인 동아시아계 군복 남자 한 명의 얼굴과 상체가 보인다. 갈색 군복, 붉은 견장, 가죽 사선 끈은 토미이시 참조에 A보다 가깝지만, 얼굴은 더 나이 들고 머리 형태도 다르다. 별도 상대 헌병의 머리나 몸은 식별되지 않는다. 눈을 크게 뜬 표정이며 자연스러운 홍채와 동공은 보인다. 읽을 수 있는 글자는 없고 시신과 서류는 구도 밖이다.",
        "hard_violations": [
         "토미이시가 상대 헌병을 벽에 고정해야 하는데, 토미이시로 읽히는 군복 남자 자신이 벽에 붙어 자기 목 아래 옷깃을 잡힌 구도로 배치되어 가해자와 피압박자의 핵심 위치·행동 관계가 뒤바뀐다."
        ],
        "physics": "손은 군복 깃을 실제로 쥐고 천을 구기며, 손목과 팔은 화면 아래로 이어진다. 남자의 어깨는 벽에 닿아 있어 지지 없는 부유는 아니다. 그러나 보이는 접촉은 자기 옷깃을 쥐거나 화면 밖 누군가에게 잡힌 자세로 읽히며, 상대 헌병을 뻗은 팔로 밀어 고정하는 동작을 입증하지 못한다."
       }
      ],
      "all_candidates_fail": false
     },
     "normalized": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 6,
        "verdict_ko": "헌병의 멱살을 잡아 벽에 고정한 관계와 핏대 선 얼굴은 분명하지만, 토미이시의 얼굴·군복이 참조와 다르고 얼굴 중심 클로즈업보다 넓다."
       },
       {
        "label": "A",
        "score": 3,
        "verdict_ko": "얼굴 클로즈업과 군복은 더 가깝지만, 토미이시가 상대 헌병을 벽에 밀어붙이는 대신 자신이 벽에 몰려 옷깃을 잡힌 듯 보여 핵심 행동 관계가 뒤바뀐다."
       }
      ],
      "all_candidates_fail": false,
      "readings": [
       {
        "label": "B",
        "direction": "왼쪽 남자는 오른쪽 헌병의 얼굴을 똑바로 노려보고, 손으로 헌병의 목 아래 옷깃을 움켜쥔다. 헌병은 턱을 뒤로 들고 왼쪽 남자를 향한다. 밀어붙이는 방향은 헌병 뒤의 오른쪽 벽으로 이어져 지시된 표적과 일치한다.",
        "built_space": "오른쪽에 회벽과 세로 목재 기둥이 있으며 헌병의 머리·등 뒤로 좁은 벽면이 보인다. 왼쪽 배경에는 격자창 하나, 창가 사각등 하나, 책상 일부와 그 위 유리 석유등 하나가 보인다. 회벽·어두운 목재·야간 조명은 장소 참조와 대체로 맞고, 두 사람은 가구 옆 벽 앞에 있다. 중복 설비나 반사는 보이지 않는다.",
        "entities": "성인 동아시아계 남성 두 명이 보이며, 외형만으로 일본 국적은 확인할 수 없다. 잡는 남자는 검은 가르마 머리와 짧은 수염이 있지만 토미이시 참조의 얼굴과 정확히 일치하지 않고, 참조의 갈색 군복 대신 짙은 민간 재킷과 밝은 속옷을 입었다. 잡힌 남자는 짧은 검은 머리, 콧수염, 견장과 붉은 깃이 있는 군복으로 헌병 역할을 읽을 수 있다. 읽을 수 있는 글자는 없다. 시신과 서류는 이 구도에서 확인되지 않는다.",
        "hard_violations": [],
        "physics": "잡는 손가락이 실제로 헌병의 옷깃을 감싸고 천이 당겨져 접힌다. 헌병의 뒤통수와 등은 벽에 밀착되어 압박을 받는 자세가 가능하다. 발은 화면 밖이지만 보이는 상체에 부유나 불가능한 지지는 없다. 다만 잡는 팔은 상당히 굽어 있어 팔을 뻗은 상태라는 연속성 지시에는 덜 맞는다."
       },
       {
        "label": "A",
        "direction": "군복 남자의 눈은 화면 오른쪽 가까운 바깥을 향하지만, 그 시선의 상대는 보이지 않는다. 손은 화면에 나온 남자 자신의 목 아래 군복 깃에 걸려 있다. 별도 헌병을 향한 뻗은 팔이나 그를 벽으로 미는 방향은 확인되지 않는다.",
        "built_space": "오른쪽 회벽에 화면의 유일한 남자 어깨가 붙어 있다. 왼쪽 배경에는 격자창 하나, 흐릿한 창가 조명 하나, 의자 등받이 하나, 책상 일부와 석유등 하나가 보인다. 재료와 어두운 야간 분위기는 참조에 가깝지만, 벽에 구속된 사람이 상대 헌병이라는 공간 관계가 성립하지 않는다. 중복 설비나 불가능한 반사는 보이지 않는다.",
        "entities": "짧은 검은 머리와 콧수염·턱수염을 가진 성인 동아시아계 군복 남자 한 명의 얼굴과 상체가 보인다. 갈색 군복, 붉은 견장, 가죽 사선 끈은 토미이시 참조에 A보다 가깝지만, 얼굴은 더 나이 들고 머리 형태도 다르다. 별도 상대 헌병의 머리나 몸은 식별되지 않는다. 눈을 크게 뜬 표정이며 자연스러운 홍채와 동공은 보인다. 읽을 수 있는 글자는 없고 시신과 서류는 구도 밖이다.",
        "hard_violations": [
         "토미이시가 상대 헌병을 벽에 고정해야 하는데, 토미이시로 읽히는 군복 남자 자신이 벽에 붙어 자기 목 아래 옷깃을 잡힌 구도로 배치되어 가해자와 피압박자의 핵심 위치·행동 관계가 뒤바뀐다."
        ],
        "physics": "손은 군복 깃을 실제로 쥐고 천을 구기며, 손목과 팔은 화면 아래로 이어진다. 남자의 어깨는 벽에 닿아 있어 지지 없는 부유는 아니다. 그러나 보이는 접촉은 자기 옷깃을 쥐거나 화면 밖 누군가에게 잡힌 자세로 읽히며, 상대 헌병을 뻗은 팔로 밀어 고정하는 동작을 입증하지 못한다."
       }
      ]
     },
     "ok": true
    }
   ],
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "slot_winner_match": false,
   "slot_winner": {
    "gemini-pro": "A",
    "gpt-high": "B"
   },
   "route": "cross_slot_combined"
  },
  "dual": {
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "normalized": {
    "A": 1.5,
    "B": 1.286
   },
   "adjusted": {
    "A": 1.25,
    "B": 1.036
   },
   "violations": {
    "B": [
     "[gemini-pro] invented people (프롬프트에 허용되지 않은 완전히 새로운 외모와 복장의 인물이 등장함)"
    ],
    "A": [
     "[gpt-high] 토미이시가 상대 헌병을 벽에 고정해야 하는데, 토미이시로 읽히는 군복 남자 자신이 벽에 붙어 자기 목 아래 옷깃을 잡힌 구도로 배치되어 가해자와 피압박자의 핵심 위치·행동 관계가 뒤바뀐다."
    ]
   },
   "per_model_winner": {
    "gemini-pro": "A",
    "gpt-high": "B"
   },
   "agreed": false
  },
  "totals": {
   "A": 1250,
   "B": 1036
  },
  "selected": "A",
  "ranking": [
   "A",
   "B"
  ],
  "verdicts": [
   {
    "label": "A",
    "score": 1250,
    "verdict_ko": "지시된 클로즈업 프레이밍, 좁은 배경 벽의 배치, 그리고 토미이시의 캐릭터 레퍼런스(얼굴과 제복)를 매우 정확하게 구현했으나, 주체적인 멱살잡이 행동이 오히려 자신이 멱살을 잡힌(혹은 스스로 잡은) 수동적 형태로 어색하게 연출된 점이 감점 요인입니다.  ★위반: [gpt-high] 토미이시가 상대 헌병을 벽에 고정해야 하는데, 토미이시로 읽히는 군복 남자 자신이 벽에 붙어 자기 목 아래 옷깃을 잡힌 구도로 배치되어 가해자와 피압박자의 핵심 위치·행동 관계가 뒤바뀐다."
   },
   {
    "label": "B",
    "score": 1036,
    "verdict_ko": "지시된 클로즈업을 무시하고 샷을 넓게 잡았으며, 무엇보다 프롬프트에 명시되지 않은 전혀 새로운 외모와 사복 차림의 인물을 메인 주체로 등장시켜 치명적인 규칙 위반(invented people)을 범했습니다.  ★위반: [gemini-pro] invented people (프롬프트에 허용되지 않은 완전히 새로운 외모와 복장의 인물이 등장함)"
   }
  ],
  "refs": [
   {
    "label": "PREVIOUS SHOT STILL — a visually related earlier shot of this same place: the location's look, materials, fixed features and lighting mood are LOCKED to this photo; never copy its camera framing. The people visible in that photo are NOT in this shot: never carry their faces, bodies or clothing onto anyone here. Each person in THIS shot is defined solely by the PEOPLE list, their own CHARACTER REFERENCE images and their wardrobe notes. If this photo shows a character who cannot move (dead or unconscious), that character's exact pose, position and orientation are ALSO LOCKED — treat their body as a fixed prop of the set that only the camera moves around.",
    "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/db00517b-165b-4005-a287-f28936b59b95/images/272bd7af-9294-4490-98c6-88a40e9deacc/scene/recipe/S32sh1_sel.png",
    "asset_id": "f8afb6f8-8cdb-4280-b8d1-e2171ae12871",
    "role": "prev_still"
   },
   {
    "label": "CHARACTER REFERENCE — 토미이시: the exact person appearing in this shot; match face, hair and build exactly.",
    "path": "<bytes:797699>",
    "asset_id": "c861ba0f-8750-4262-8324-25cf5c6c9dcd",
    "role": "character_ref"
   }
  ],
  "critique_skipped": true,
  "shot_run_produced": false,
  "shot_run_uid": "06aa1428-3daf-7fe2-9aee-34166fd3c511",
  "ref_mode": "prev+엔티티",
  "share_plan": {
   "ref_plan": "prev",
   "prev_anchor_tag": "S32sh1"
  }
 },
 "S32sh10::cine": {
  "applied": true,
  "attempted_at": "2026-09-09T04:31:01.833896+00:00",
  "fingerprint": "47ace807390d2a9d1794a198599629b336a73117e293e09fa9ab4ee3a74f2c1e",
  "fingerprint_version": 2,
  "provider": "mai",
  "endpoint": "openrouter/chat-completions",
  "model": "microsoft/mai-image-2.6",
  "pack": "24.202608252115",
  "source_file": "S32sh10_sel.png",
  "source_sha256": "c18cd2f73d44181e267534d98439cc9d50ed6a3248262f22ed5a54d38901c909",
  "file": "S32sh10_cine.png",
  "staged_sha256": "a39c73f65b8267e680b35893b5700b0c3a9d47188fd7acf86780ef0d56fa39a7",
  "latency_ms": 26895
 },
 "S32sh22::signage": {
  "fp": "b14cdef648b6b038",
  "inscriptions": [
   {
    "text_native": "巨魁 安圭洪",
    "source": "scene_text_quoted",
    "reason_ko": "서류 종이에 붓글씨로 적혀 있는 대상의 이름과 칭호이다.",
    "source_quote": "巨魁 安圭洪"
   }
  ],
  "cues": [],
  "dropped": []
 },
 "S32sh22": {
  "input_fingerprint": "89c458cc48ea6297",
  "prompt": "Create ONE FINAL photorealistic live-action film still of the moment below — early 20th-century Korea, Jeolla region; all people are Korean unless explicitly stated otherwise. TIME OF DAY (lock): night.\n\nSHOT TEXT (authoritative, Korean): 텅 빈 집무실 테이블 위, '巨魁 安圭洪' 붓글씨가 짙게 적힌 서류 종이 클로즈업.\n\nLOCATION (lock): The paperwork surface of a table inside the dim police outpost office at night, lit sufficiently to show freshly written brush characters. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: close-up\n- FRAME LAYOUT: Completed document with 巨魁 安圭洪 facing upward in the middle-center of the frame, midground.\n- KEY BACKGROUND ELEMENTS: Completed document (The designation and name have been written in dark brush strokes) — The written face points directly upward toward the camera, with '巨魁 安圭洪' legible in its reading orientation; used as Primary focal detail, isolated from all faces and hands; Table edge and surrounding tabletop (Supporting the completed document) — The camera sees the top and its boundary from directly above; used as Provide a restrained border around the page rather than an oversized furniture foreground; Office floor (Visible beyond the cropped table edge); used as Keep the document grounded in the office and preserve realistic framing proportions.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Continue the restrained nighttime office exposure while preserving clear separation between the dark brush lettering and the paper.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established — the place, its fixed features and wear, each person's clothing and state — persists. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The paper on the table now reads '巨魁 安圭洪' in dark ink; the shroud remains drawn back.\n\nPEOPLE: the SHOT TEXT alone decides who is visible in this shot. People known to appear somewhere in this scene: 배상도 (한국인 남성, 성인 나이대, 자연스러운 검은 머리); 토미이시 (일본인 남성, 성인 나이대, 자연스러운 검은 머리). That list is scene-level, not a cast list for this frame — it may name someone this shot does not show, and it may omit someone this shot does show. If the shot text names a person who is not on the list, draw that person exactly as the shot text describes them; the list does not override the shot text. Never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person.\n\nWORDS TO RENDER (authoritative — the scene itself calls for these; render each as period-real physical lettering in the native script, exactly as written; add no other readable text anywhere):\n- \"巨魁 安圭洪\"\n\nThe WORDS TO RENDER above are the only readable writing in this image: render those words exactly as given, in the place and era's own language and script, and nothing else legible. Invent no other wording a viewer could read. No caption, subtitle, watermark, logo or overlay. Surfaces that would carry writing may still be present — stage any wording they would carry out of legibility: a hand across, an oblique angle, shallow focus.",
  "roll_prompts": {
   "A": "Create ONE FINAL photorealistic live-action film still of the moment below — early 20th-century Korea, Jeolla region; all people are Korean unless explicitly stated otherwise. TIME OF DAY (lock): night.\n\nSHOT TEXT (authoritative, Korean): 텅 빈 집무실 테이블 위, '巨魁 安圭洪' 붓글씨가 짙게 적힌 서류 종이 클로즈업.\n\nLOCATION (lock): The paperwork surface of a table inside the dim police outpost office at night, lit sufficiently to show freshly written brush characters. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: close-up\n- FRAME LAYOUT: Completed document with 巨魁 安圭洪 facing upward in the middle-center of the frame, midground.\n- KEY BACKGROUND ELEMENTS: Completed document (The designation and name have been written in dark brush strokes) — The written face points directly upward toward the camera, with '巨魁 安圭洪' legible in its reading orientation; used as Primary focal detail, isolated from all faces and hands; Table edge and surrounding tabletop (Supporting the completed document) — The camera sees the top and its boundary from directly above; used as Provide a restrained border around the page rather than an oversized furniture foreground; Office floor (Visible beyond the cropped table edge); used as Keep the document grounded in the office and preserve realistic framing proportions.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Continue the restrained nighttime office exposure while preserving clear separation between the dark brush lettering and the paper.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established — the place, its fixed features and wear, each person's clothing and state — persists. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The paper on the table now reads '巨魁 安圭洪' in dark ink; the shroud remains drawn back.\n\nPEOPLE: the SHOT TEXT alone decides who is visible in this shot. People known to appear somewhere in this scene: 배상도 (한국인 남성, 성인 나이대, 자연스러운 검은 머리); 토미이시 (일본인 남성, 성인 나이대, 자연스러운 검은 머리). That list is scene-level, not a cast list for this frame — it may name someone this shot does not show, and it may omit someone this shot does show. If the shot text names a person who is not on the list, draw that person exactly as the shot text describes them; the list does not override the shot text. Never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person.\n\nWORDS TO RENDER (authoritative — the scene itself calls for these; render each as period-real physical lettering in the native script, exactly as written; add no other readable text anywhere):\n- \"巨魁 安圭洪\"\n\nThe WORDS TO RENDER above are the only readable writing in this image: render those words exactly as given, in the place and era's own language and script, and nothing else legible. Invent no other wording a viewer could read. No caption, subtitle, watermark, logo or overlay. Surfaces that would carry writing may still be present — stage any wording they would carry out of legibility: a hand across, an oblique angle, shallow focus.",
   "B": "Create ONE FINAL photorealistic live-action film still of the moment below — early 20th-century Korea, Jeolla region; all people are Korean unless explicitly stated otherwise. TIME OF DAY (lock): night.\n\nSHOT TEXT (authoritative, Korean): 텅 빈 집무실 테이블 위, '巨魁 安圭洪' 붓글씨가 짙게 적힌 서류 종이 클로즈업.\n\nLOCATION (lock): The paperwork surface of a table inside the dim police outpost office at night, lit sufficiently to show freshly written brush characters. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: close-up\n- FRAME LAYOUT: Completed document with 巨魁 安圭洪 facing upward in the middle-center of the frame, midground.\n- KEY BACKGROUND ELEMENTS: Completed document (The designation and name have been written in dark brush strokes) — The written face points directly upward toward the camera, with '巨魁 安圭洪' legible in its reading orientation; used as Primary focal detail, isolated from all faces and hands; Table edge and surrounding tabletop (Supporting the completed document) — The camera sees the top and its boundary from directly above; used as Provide a restrained border around the page rather than an oversized furniture foreground; Office floor (Visible beyond the cropped table edge); used as Keep the document grounded in the office and preserve realistic framing proportions.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Continue the restrained nighttime office exposure while preserving clear separation between the dark brush lettering and the paper.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established — the place, its fixed features and wear, each person's clothing and state — persists. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The paper on the table now reads '巨魁 安圭洪' in dark ink; the shroud remains drawn back.\n\nPEOPLE: the SHOT TEXT alone decides who is visible in this shot. People known to appear somewhere in this scene: 배상도 (한국인 남성, 성인 나이대, 자연스러운 검은 머리); 토미이시 (일본인 남성, 성인 나이대, 자연스러운 검은 머리). That list is scene-level, not a cast list for this frame — it may name someone this shot does not show, and it may omit someone this shot does show. If the shot text names a person who is not on the list, draw that person exactly as the shot text describes them; the list does not override the shot text. Never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person.\n\nWORDS TO RENDER (authoritative — the scene itself calls for these; render each as period-real physical lettering in the native script, exactly as written; add no other readable text anywhere):\n- \"巨魁 安圭洪\"\n\nThe WORDS TO RENDER above are the only readable writing in this image: render those words exactly as given, in the place and era's own language and script, and nothing else legible. Invent no other wording a viewer could read. No caption, subtitle, watermark, logo or overlay. Surfaces that would carry writing may still be present — stage any wording they would carry out of legibility: a hand across, an oblique angle, shallow focus.\n\nCOMPOSITION VARIATION (this candidate only): keep the place, the people and the moment exactly as specified, and keep from the CAMERA & FRAME contract its subject scale, screen placement and key background elements exactly. For this candidate only, its camera angle and height are deliberately left open: choose a camera position distinctly different from the obvious one, as a film director picking a second setup on the same blocking. Stay true to any explicit framing the SHOT TEXT itself dictates."
  },
  "shot_run_spend_attempt_count": 1,
  "readings": [
   {
    "label": "A",
    "direction": "카메라는 책상 위를 수직으로 내려다보며 중앙의 서류를 향하고 있음.",
    "built_space": "레퍼런스와 일치하는 나무 책상 표면이 보이며, 우측에는 나무 바닥이 나타남. 책상 우측 상단에 레퍼런스의 호롱불(램프)이 정확히 배치됨.",
    "entities": "지정된 한자 '巨魁 安圭洪'이 종이 위에 세로 방향 붓글씨로 선명하고 정확하게 적혀 있음.",
    "hard_violations": [],
    "physics": "종이는 책상 표면 위에 평평하게 잘 놓여 있으며, 들뜨거나 어색한 부분 없이 중력이 올바르게 적용됨."
   },
   {
    "label": "B",
    "direction": "카메라는 책상 위를 수직으로 내려다보며 중앙의 서류를 향하고 있음.",
    "built_space": "나무 책상 표면과 그 아래 및 우측의 나무 바닥이 보임. 책상 상단에 붉은빛이 도는 둥근 보관함이 보이나 책상의 폭이 레퍼런스에 비해 좁게 표현됨.",
    "entities": "지정된 한자 '巨魁 安圭洪'이 종이 위에 가로 방향 붓글씨로 정확하게 적혀 있음.",
    "hard_violations": [],
    "physics": "종이는 책상 표면 위에 평평하게 놓여 있으며 접힌 자국 등의 질감이 자연스럽게 표현됨."
   }
  ],
  "cross_model_order": {
   "policy": "cross_model_order_v2_slot0_fwd_slot1_rev_parallel2",
   "slots": [
    {
     "model": "gemini-pro",
     "order": "forward",
     "display_to_canonical": {
      "A": "A",
      "B": "B"
     },
     "raw": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 10,
        "verdict_ko": "지정된 텍스트('巨魁 安圭洪')를 시대적 배경에 알맞게 세로로 정확히 구현했으며, 레퍼런스의 책상 램프와 질감을 훌륭하게 반영한 완벽한 클로즈업 샷입니다."
       },
       {
        "label": "B",
        "score": 8,
        "verdict_ko": "지정된 텍스트를 정확하게 렌더링했으나, 시대적 배경에 비추어 볼 때 가로쓰기 방향이 다소 어색하며 책상의 스케일이 레퍼런스와 다소 차이가 납니다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "카메라는 책상 위를 수직으로 내려다보며 중앙의 서류를 향하고 있음.",
        "built_space": "레퍼런스와 일치하는 나무 책상 표면이 보이며, 우측에는 나무 바닥이 나타남. 책상 우측 상단에 레퍼런스의 호롱불(램프)이 정확히 배치됨.",
        "entities": "지정된 한자 '巨魁 安圭洪'이 종이 위에 세로 방향 붓글씨로 선명하고 정확하게 적혀 있음.",
        "hard_violations": [],
        "physics": "종이는 책상 표면 위에 평평하게 잘 놓여 있으며, 들뜨거나 어색한 부분 없이 중력이 올바르게 적용됨."
       },
       {
        "label": "B",
        "direction": "카메라는 책상 위를 수직으로 내려다보며 중앙의 서류를 향하고 있음.",
        "built_space": "나무 책상 표면과 그 아래 및 우측의 나무 바닥이 보임. 책상 상단에 붉은빛이 도는 둥근 보관함이 보이나 책상의 폭이 레퍼런스에 비해 좁게 표현됨.",
        "entities": "지정된 한자 '巨魁 安圭洪'이 종이 위에 가로 방향 붓글씨로 정확하게 적혀 있음.",
        "hard_violations": [],
        "physics": "종이는 책상 표면 위에 평평하게 놓여 있으며 접힌 자국 등의 질감이 자연스럽게 표현됨."
       }
      ],
      "all_candidates_fail": false
     },
     "normalized": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 10,
        "verdict_ko": "지정된 텍스트('巨魁 安圭洪')를 시대적 배경에 알맞게 세로로 정확히 구현했으며, 레퍼런스의 책상 램프와 질감을 훌륭하게 반영한 완벽한 클로즈업 샷입니다."
       },
       {
        "label": "B",
        "score": 8,
        "verdict_ko": "지정된 텍스트를 정확하게 렌더링했으나, 시대적 배경에 비추어 볼 때 가로쓰기 방향이 다소 어색하며 책상의 스케일이 레퍼런스와 다소 차이가 납니다."
       }
      ],
      "all_candidates_fail": false,
      "readings": [
       {
        "label": "A",
        "direction": "카메라는 책상 위를 수직으로 내려다보며 중앙의 서류를 향하고 있음.",
        "built_space": "레퍼런스와 일치하는 나무 책상 표면이 보이며, 우측에는 나무 바닥이 나타남. 책상 우측 상단에 레퍼런스의 호롱불(램프)이 정확히 배치됨.",
        "entities": "지정된 한자 '巨魁 安圭洪'이 종이 위에 세로 방향 붓글씨로 선명하고 정확하게 적혀 있음.",
        "hard_violations": [],
        "physics": "종이는 책상 표면 위에 평평하게 잘 놓여 있으며, 들뜨거나 어색한 부분 없이 중력이 올바르게 적용됨."
       },
       {
        "label": "B",
        "direction": "카메라는 책상 위를 수직으로 내려다보며 중앙의 서류를 향하고 있음.",
        "built_space": "나무 책상 표면과 그 아래 및 우측의 나무 바닥이 보임. 책상 상단에 붉은빛이 도는 둥근 보관함이 보이나 책상의 폭이 레퍼런스에 비해 좁게 표현됨.",
        "entities": "지정된 한자 '巨魁 安圭洪'이 종이 위에 가로 방향 붓글씨로 정확하게 적혀 있음.",
        "hard_violations": [],
        "physics": "종이는 책상 표면 위에 평평하게 놓여 있으며 접힌 자국 등의 질감이 자연스럽게 표현됨."
       }
      ]
     },
     "ok": true
    },
    {
     "model": "gpt-high",
     "order": "reverse",
     "display_to_canonical": {
      "A": "B",
      "B": "A"
     },
     "raw": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 9,
        "verdict_ko": "정상 방향의 ‘巨魁 安圭洪’을 중앙에 크게 담고, 종이 둘레의 절제된 상판과 가장자리 너머 바닥까지 보여 지정된 클로즈업 구도에 더 충실하다."
       },
       {
        "label": "B",
        "score": 8,
        "verdict_ko": "문구와 수직 부감은 정확하지만, 문서 양옆의 넓은 상판과 눈에 띄는 등잔이 종이 중심의 절제된 테두리 구도를 약화한다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "종이의 글씨 면이 위쪽 카메라를 향하며, ‘巨魁 安圭洪’이 왼쪽에서 오른쪽으로 정상적으로 읽힌다. 시선이나 조준 방향을 확인할 인물·무기·이동 물체는 없다.",
        "built_space": "낡은 목재 탁자 상판 하나를 거의 수직으로 내려다본다. 좌우와 아래쪽 경계 밖으로 판재 바닥이 보이고, 종이 주변 상판은 비교적 좁은 테두리를 이룬다. 위쪽에는 잘린 서류 더미 하나와 둥근 용기 하나, 왼쪽 위에는 인접 목재 가구 일부가 보인다. 참고 이미지의 어두운 목재와 마모된 바닥 분위기는 이어지지만 상판의 테두리 이음새까지 동일한지는 확인하기 어렵다. 인물이나 반사는 없다.",
        "entities": "중앙의 밝은 섬유질 종이에 ‘巨魁 安圭洪’ 다섯 글자가 짙은 붓획으로 쓰여 있으며 추가로 읽히는 문자는 없다. 탁자와 사무실 판재 바닥이 보인다. 위쪽의 서류와 둥근 용기는 참고 이미지에 있는 소품 종류와 부합한다. 얼굴과 손은 완전히 제외되어 있다. 인물의 신원·복장과 수의 상태는 이 크롭에서 확인할 대상이 아니다.",
        "hard_violations": [],
        "physics": "종이는 상판에 평평하게 놓여 지지되고, 접힌 자국과 가장자리의 얕은 그림자가 실제 종이의 두께를 나타낸다. 상단 서류와 둥근 용기도 상판에 닿아 있다. 먹획은 종이 표면에 스며든 물질로 보이며, 지지 없이 떠 있는 물체나 움직이는 신체는 없다."
       },
       {
        "label": "B",
        "direction": "세로로 놓인 문서의 글씨 면이 위쪽 카메라를 향한다. ‘巨魁 安圭洪’이 위에서 아래로 정상적으로 읽히며 뒤집히거나 좌우 반전되지 않았다. 인물의 시선이나 움직임은 없다.",
        "built_space": "목재 상판 하나를 거의 수직으로 내려다보며 오른쪽 테이블 경계와 그 너머 판재 바닥을 보여 준다. 중앙의 세로 문서 양옆, 특히 왼쪽에 넓은 빈 상판이 남는다. 오른쪽 위에는 등잔 하나가 일부 잘려 보인다. 참고 이미지의 목재 재질과 등잔 조명은 이어지지만, 상판이 문서 주변의 절제된 테두리라기보다 넓은 배경을 차지한다. 인물이나 반사는 없다.",
        "entities": "중앙 문서에는 요청한 ‘巨魁 安圭洪’이 짙은 붓글씨로 적혀 있고 다른 판독 가능한 글자는 없다. 누렇게 바랜 종이 아래로 얇은 종이 가장자리가 겹쳐 보인다. 목재 탁자, 판재 바닥, 참고 이미지와 같은 종류의 유리 등잔이 보인다. 얼굴과 손은 없으며 인물의 신원·복장이나 수의 상태는 프레임 밖이다.",
        "hard_violations": [],
        "physics": "문서와 아래쪽 종이는 상판이 받치고 있으며 가장자리 그림자가 접촉을 보여 준다. 등잔은 넓은 밑받침으로 탁자 위에 서 있다. 붓획은 종이의 질감과 조명에 통합되어 보인다. 공중에 뜬 물체나 불가능한 자세는 없다."
       }
      ],
      "all_candidates_fail": false
     },
     "normalized": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 9,
        "verdict_ko": "정상 방향의 ‘巨魁 安圭洪’을 중앙에 크게 담고, 종이 둘레의 절제된 상판과 가장자리 너머 바닥까지 보여 지정된 클로즈업 구도에 더 충실하다."
       },
       {
        "label": "A",
        "score": 8,
        "verdict_ko": "문구와 수직 부감은 정확하지만, 문서 양옆의 넓은 상판과 눈에 띄는 등잔이 종이 중심의 절제된 테두리 구도를 약화한다."
       }
      ],
      "all_candidates_fail": false,
      "readings": [
       {
        "label": "B",
        "direction": "종이의 글씨 면이 위쪽 카메라를 향하며, ‘巨魁 安圭洪’이 왼쪽에서 오른쪽으로 정상적으로 읽힌다. 시선이나 조준 방향을 확인할 인물·무기·이동 물체는 없다.",
        "built_space": "낡은 목재 탁자 상판 하나를 거의 수직으로 내려다본다. 좌우와 아래쪽 경계 밖으로 판재 바닥이 보이고, 종이 주변 상판은 비교적 좁은 테두리를 이룬다. 위쪽에는 잘린 서류 더미 하나와 둥근 용기 하나, 왼쪽 위에는 인접 목재 가구 일부가 보인다. 참고 이미지의 어두운 목재와 마모된 바닥 분위기는 이어지지만 상판의 테두리 이음새까지 동일한지는 확인하기 어렵다. 인물이나 반사는 없다.",
        "entities": "중앙의 밝은 섬유질 종이에 ‘巨魁 安圭洪’ 다섯 글자가 짙은 붓획으로 쓰여 있으며 추가로 읽히는 문자는 없다. 탁자와 사무실 판재 바닥이 보인다. 위쪽의 서류와 둥근 용기는 참고 이미지에 있는 소품 종류와 부합한다. 얼굴과 손은 완전히 제외되어 있다. 인물의 신원·복장과 수의 상태는 이 크롭에서 확인할 대상이 아니다.",
        "hard_violations": [],
        "physics": "종이는 상판에 평평하게 놓여 지지되고, 접힌 자국과 가장자리의 얕은 그림자가 실제 종이의 두께를 나타낸다. 상단 서류와 둥근 용기도 상판에 닿아 있다. 먹획은 종이 표면에 스며든 물질로 보이며, 지지 없이 떠 있는 물체나 움직이는 신체는 없다."
       },
       {
        "label": "A",
        "direction": "세로로 놓인 문서의 글씨 면이 위쪽 카메라를 향한다. ‘巨魁 安圭洪’이 위에서 아래로 정상적으로 읽히며 뒤집히거나 좌우 반전되지 않았다. 인물의 시선이나 움직임은 없다.",
        "built_space": "목재 상판 하나를 거의 수직으로 내려다보며 오른쪽 테이블 경계와 그 너머 판재 바닥을 보여 준다. 중앙의 세로 문서 양옆, 특히 왼쪽에 넓은 빈 상판이 남는다. 오른쪽 위에는 등잔 하나가 일부 잘려 보인다. 참고 이미지의 목재 재질과 등잔 조명은 이어지지만, 상판이 문서 주변의 절제된 테두리라기보다 넓은 배경을 차지한다. 인물이나 반사는 없다.",
        "entities": "중앙 문서에는 요청한 ‘巨魁 安圭洪’이 짙은 붓글씨로 적혀 있고 다른 판독 가능한 글자는 없다. 누렇게 바랜 종이 아래로 얇은 종이 가장자리가 겹쳐 보인다. 목재 탁자, 판재 바닥, 참고 이미지와 같은 종류의 유리 등잔이 보인다. 얼굴과 손은 없으며 인물의 신원·복장이나 수의 상태는 프레임 밖이다.",
        "hard_violations": [],
        "physics": "문서와 아래쪽 종이는 상판이 받치고 있으며 가장자리 그림자가 접촉을 보여 준다. 등잔은 넓은 밑받침으로 탁자 위에 서 있다. 붓획은 종이의 질감과 조명에 통합되어 보인다. 공중에 뜬 물체나 불가능한 자세는 없다."
       }
      ]
     },
     "ok": true
    }
   ],
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "slot_winner_match": false,
   "slot_winner": {
    "gemini-pro": "A",
    "gpt-high": "B"
   },
   "route": "cross_slot_combined"
  },
  "dual": {
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "normalized": {
    "A": 1.889,
    "B": 1.8
   },
   "adjusted": {
    "A": 1.889,
    "B": 1.8
   },
   "violations": {},
   "per_model_winner": {
    "gemini-pro": "A",
    "gpt-high": "B"
   },
   "agreed": false
  },
  "totals": {
   "A": 1889,
   "B": 1800
  },
  "selected": "A",
  "ranking": [
   "A",
   "B"
  ],
  "verdicts": [
   {
    "label": "A",
    "score": 1889,
    "verdict_ko": "지정된 텍스트('巨魁 安圭洪')를 시대적 배경에 알맞게 세로로 정확히 구현했으며, 레퍼런스의 책상 램프와 질감을 훌륭하게 반영한 완벽한 클로즈업 샷입니다."
   },
   {
    "label": "B",
    "score": 1800,
    "verdict_ko": "지정된 텍스트를 정확하게 렌더링했으나, 시대적 배경에 비추어 볼 때 가로쓰기 방향이 다소 어색하며 책상의 스케일이 레퍼런스와 다소 차이가 납니다."
   }
  ],
  "refs": [
   {
    "label": "PREVIOUS SHOT STILL — a visually related earlier shot of this same place: the location's look, materials, fixed features, lighting mood and each person's clothing are LOCKED to this photo; never copy its camera framing. If this photo shows a character who cannot move (dead or unconscious), that character's exact pose, position and orientation are ALSO LOCKED — treat their body as a fixed prop of the set that only the camera moves around.",
    "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/db00517b-165b-4005-a287-f28936b59b95/images/272bd7af-9294-4490-98c6-88a40e9deacc/scene/recipe/S32sh1_sel.png",
    "asset_id": "f8afb6f8-8cdb-4280-b8d1-e2171ae12871",
    "role": "prev_still"
   }
  ],
  "critique_skipped": true,
  "shot_run_produced": false,
  "shot_run_uid": "06aa1428-3f3f-739d-8539-9a3cabc4ee08",
  "ref_mode": "prev만 (배경 전용·공유 계획)",
  "share_plan": {
   "ref_plan": "prev",
   "prev_anchor_tag": "S32sh1"
  },
  "lane_policy": "share_plan_prev_bgonly"
 },
 "S32sh22::cine": {
  "applied": true,
  "attempted_at": "2026-09-09T04:32:11.072655+00:00",
  "fingerprint": "a0f887b085aecd9a2383a16ec6b6fbd6015aa09cb4aa8711e18e10d908d8cb31",
  "fingerprint_version": 2,
  "provider": "mai",
  "endpoint": "openrouter/chat-completions",
  "model": "microsoft/mai-image-2.6",
  "pack": "24.202608252115",
  "source_file": "S32sh22_sel.png",
  "source_sha256": "b29ffc8609292fe866cf52575784f67d256fd465e48a90a2fff4c91dc83de2a6",
  "file": "S32sh22_cine.png",
  "staged_sha256": "3aed828ab22c9e0b00b1a9000fe9bd82fb32dc02b47d47ad5a1fba0e07104c78",
  "latency_ms": 23177
 },
 "S33sh4::signage": {
  "fp": "9c712de1783f976c",
  "inscriptions": [],
  "cues": [],
  "dropped": []
 },
 "S33sh4": {
  "input_fingerprint": "af6d5d4c5e9cd93c",
  "prompt": "Create ONE FINAL photorealistic live-action film still of the moment below — early 20th-century Korea, Jeolla region; all people are Korean unless explicitly stated otherwise. TIME OF DAY (lock): night.\n\nSHOT TEXT (authoritative, Korean): 박제현을 향해 몸을 바짝 기울인 채 은밀하게 속삭이듯 입을 반쯤 벌린 박상현의 측면.\n\nLOCATION (lock): The floor-seating area inside the estate's dark guest-wing room at night, with its door closed. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: close-up\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Keep the nighttime room dark and muted, with enough naturalistic tonal separation to read the lowered, confidential delivery.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The clothing, hair and overall look of 박상현 — who appear both in that photo and in this shot — are LOCKED to that photo. Anyone else visible in that photo is NOT in this shot: never carry their face, body or clothing onto anyone here. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The sarangchae door is closed; the grain-store door beyond the dark yard remains firmly shut with its seizure seal. 박상현: He is inside after closing the door.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 박상현 (한국인 남성, 성인 나이대, 자연스러운 검은 머리) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.",
  "roll_prompts": {
   "A": "Create ONE FINAL photorealistic live-action film still of the moment below — early 20th-century Korea, Jeolla region; all people are Korean unless explicitly stated otherwise. TIME OF DAY (lock): night.\n\nSHOT TEXT (authoritative, Korean): 박제현을 향해 몸을 바짝 기울인 채 은밀하게 속삭이듯 입을 반쯤 벌린 박상현의 측면.\n\nLOCATION (lock): The floor-seating area inside the estate's dark guest-wing room at night, with its door closed. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: close-up\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Keep the nighttime room dark and muted, with enough naturalistic tonal separation to read the lowered, confidential delivery.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The clothing, hair and overall look of 박상현 — who appear both in that photo and in this shot — are LOCKED to that photo. Anyone else visible in that photo is NOT in this shot: never carry their face, body or clothing onto anyone here. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The sarangchae door is closed; the grain-store door beyond the dark yard remains firmly shut with its seizure seal. 박상현: He is inside after closing the door.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 박상현 (한국인 남성, 성인 나이대, 자연스러운 검은 머리) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.",
   "B": "Create ONE FINAL photorealistic live-action film still of the moment below — early 20th-century Korea, Jeolla region; all people are Korean unless explicitly stated otherwise. TIME OF DAY (lock): night.\n\nSHOT TEXT (authoritative, Korean): 박제현을 향해 몸을 바짝 기울인 채 은밀하게 속삭이듯 입을 반쯤 벌린 박상현의 측면.\n\nLOCATION (lock): The floor-seating area inside the estate's dark guest-wing room at night, with its door closed. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: close-up\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Keep the nighttime room dark and muted, with enough naturalistic tonal separation to read the lowered, confidential delivery.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The clothing, hair and overall look of 박상현 — who appear both in that photo and in this shot — are LOCKED to that photo. Anyone else visible in that photo is NOT in this shot: never carry their face, body or clothing onto anyone here. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The sarangchae door is closed; the grain-store door beyond the dark yard remains firmly shut with its seizure seal. 박상현: He is inside after closing the door.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 박상현 (한국인 남성, 성인 나이대, 자연스러운 검은 머리) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.\n\nCOMPOSITION VARIATION (this candidate only): keep the place, the people and the moment exactly as specified, and keep from the CAMERA & FRAME contract its subject scale, screen placement and key background elements exactly. For this candidate only, its camera angle and height are deliberately left open: choose a camera position distinctly different from the obvious one, as a film director picking a second setup on the same blocking. Stay true to any explicit framing the SHOT TEXT itself dictates."
  },
  "shot_run_spend_attempt_count": 1,
  "readings": [
   {
    "label": "A",
    "direction": "박상현의 시선과 상체는 화면 전경에 위치한 다른 인물을 직접 향하고 있음.",
    "built_space": "촛대와 그릇이 놓인 어두운 방 내부이며, 인물들이 바닥에 마주 앉아 있음.",
    "entities": "박상현의 이목구비와 복장은 레퍼런스와 일치하나, 프레임 내에 등장해서는 안 되는 인물(등과 옆모습)이 포함되어 있음.",
    "hard_violations": [
     "[gemini-pro] invented people (프레임 내 허용되지 않은 추가 인물 등장)",
     "[gpt-high] 가시 인물을 박상현으로만 제한했는데 왼쪽에 별도 인물의 얼굴과 몸통을 추가했다."
    ],
    "physics": "박상현이 상대방의 팔을 잡고 있는 손의 형태와 몸을 기울인 자세가 물리적으로 자연스러움."
   },
   {
    "label": "B",
    "direction": "박상현의 시선과 몸은 화면 밖의 대상(박제현)을 향해 비스듬히 앞을 향하고 있음.",
    "built_space": "나무 기둥과 한지 문이 있는 어두운 전통 방 내부로, 이전 샷의 공간적 특징을 잘 유지하고 있음.",
    "entities": "박상현은 레퍼런스와 일치하는 이목구비, 수염, 갓, 회색 도포를 정확히 갖추고 있음.",
    "hard_violations": [],
    "physics": "바닥에 앉아 상체를 앞으로 깊게 숙인 자세가 중력에 맞게 자연스럽게 지지되고 있음."
   }
  ],
  "cross_model_order": {
   "policy": "cross_model_order_v2_slot0_fwd_slot1_rev_parallel2",
   "slots": [
    {
     "model": "gemini-pro",
     "order": "forward",
     "display_to_canonical": {
      "A": "A",
      "B": "B"
     },
     "raw": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 10,
        "verdict_ko": "허용된 인물만 프레임에 유지하면서, 화면 밖의 대상을 향해 몸을 기울이고 입을 반쯤 벌린 박상현의 측면 모습을 샷 텍스트에 맞게 완벽하게 구현했습니다."
       },
       {
        "label": "A",
        "score": 0,
        "verdict_ko": "프레임 내에 허용 목록에 없는 다른 인물을 등장시킨 것은 명백한 하드 위반입니다."
       }
      ],
      "readings": [
       {
        "label": "B",
        "direction": "박상현의 시선과 몸은 화면 밖의 대상(박제현)을 향해 비스듬히 앞을 향하고 있음.",
        "built_space": "나무 기둥과 한지 문이 있는 어두운 전통 방 내부로, 이전 샷의 공간적 특징을 잘 유지하고 있음.",
        "entities": "박상현은 레퍼런스와 일치하는 이목구비, 수염, 갓, 회색 도포를 정확히 갖추고 있음.",
        "hard_violations": [],
        "physics": "바닥에 앉아 상체를 앞으로 깊게 숙인 자세가 중력에 맞게 자연스럽게 지지되고 있음."
       },
       {
        "label": "A",
        "direction": "박상현의 시선과 상체는 화면 전경에 위치한 다른 인물을 직접 향하고 있음.",
        "built_space": "촛대와 그릇이 놓인 어두운 방 내부이며, 인물들이 바닥에 마주 앉아 있음.",
        "entities": "박상현의 이목구비와 복장은 레퍼런스와 일치하나, 프레임 내에 등장해서는 안 되는 인물(등과 옆모습)이 포함되어 있음.",
        "hard_violations": [
         "invented people (프레임 내 허용되지 않은 추가 인물 등장)"
        ],
        "physics": "박상현이 상대방의 팔을 잡고 있는 손의 형태와 몸을 기울인 자세가 물리적으로 자연스러움."
       }
      ],
      "all_candidates_fail": false
     },
     "normalized": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 10,
        "verdict_ko": "허용된 인물만 프레임에 유지하면서, 화면 밖의 대상을 향해 몸을 기울이고 입을 반쯤 벌린 박상현의 측면 모습을 샷 텍스트에 맞게 완벽하게 구현했습니다."
       },
       {
        "label": "A",
        "score": 0,
        "verdict_ko": "프레임 내에 허용 목록에 없는 다른 인물을 등장시킨 것은 명백한 하드 위반입니다."
       }
      ],
      "all_candidates_fail": false,
      "readings": [
       {
        "label": "B",
        "direction": "박상현의 시선과 몸은 화면 밖의 대상(박제현)을 향해 비스듬히 앞을 향하고 있음.",
        "built_space": "나무 기둥과 한지 문이 있는 어두운 전통 방 내부로, 이전 샷의 공간적 특징을 잘 유지하고 있음.",
        "entities": "박상현은 레퍼런스와 일치하는 이목구비, 수염, 갓, 회색 도포를 정확히 갖추고 있음.",
        "hard_violations": [],
        "physics": "바닥에 앉아 상체를 앞으로 깊게 숙인 자세가 중력에 맞게 자연스럽게 지지되고 있음."
       },
       {
        "label": "A",
        "direction": "박상현의 시선과 상체는 화면 전경에 위치한 다른 인물을 직접 향하고 있음.",
        "built_space": "촛대와 그릇이 놓인 어두운 방 내부이며, 인물들이 바닥에 마주 앉아 있음.",
        "entities": "박상현의 이목구비와 복장은 레퍼런스와 일치하나, 프레임 내에 등장해서는 안 되는 인물(등과 옆모습)이 포함되어 있음.",
        "hard_violations": [
         "invented people (프레임 내 허용되지 않은 추가 인물 등장)"
        ],
        "physics": "박상현이 상대방의 팔을 잡고 있는 손의 형태와 몸을 기울인 자세가 물리적으로 자연스러움."
       }
      ]
     },
     "ok": true
    },
    {
     "model": "gpt-high",
     "order": "reverse",
     "display_to_canonical": {
      "A": "B",
      "B": "A"
     },
     "raw": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 7,
        "verdict_ko": "박상현 혼자 몸을 앞으로 깊이 기울이고 입을 벌린 측면 연기는 충실하지만, 얼굴보다 상체와 배경을 넓게 담아 지정된 클로즈업에는 다소 느슨하다."
       },
       {
        "label": "B",
        "score": 3,
        "verdict_ko": "얼굴 중심의 근접 구도와 속삭이는 표정은 맞지만, 박상현만 보여야 하는 화면에 상대 인물을 추가하고 팔을 붙잡는 동작까지 만들어 인물 제한을 위반한다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "박상현의 얼굴과 시선, 앞으로 숙인 상체가 모두 화면 왼쪽 아래의 화면 밖 상대를 향한다. 박제현은 보이지 않아 실제 눈맞춤은 확인할 수 없지만, 화면 밖 박제현에게 낮게 속삭이는 방향으로 읽힌다. 입은 반쯤 열려 있다.",
        "built_space": "배경에는 낡은 목재 기둥과 종이 면, 고리 손잡이 하나가 달린 닫힌 문짝 하나가 보인다. 아래에는 사각 등 하나와 왼쪽 가장자리의 낮은 가구 일부가 있다. 목재와 종이의 재질은 이전 장면과 대체로 연결되지만, 등과 세부 문 구성은 참고 화면에서 확인되지 않는다. 인물은 바닥 가까운 높이에 있으나 좌면과 하체는 잘려 있다. 거울이나 반사는 없다.",
        "entities": "성인 한국인 남성으로 제시된 박상현 한 명만 보인다. 얼굴 윤곽과 콧수염·턱수염, 검은 갓과 망건, 회색 도포와 밝은 깃이 참고 인물에 대체로 부합한다. 머리카락은 대부분 가려져 있고 드러난 부분은 검다. 박제현과 마당 너머 곡간은 프레임 밖이며, 읽을 수 있는 글자는 없다.",
        "hard_violations": [],
        "physics": "목과 어깨, 상체가 함께 앞으로 굽혀져 실제로 가능한 숙임이다. 골반과 발의 지지점은 화면 밖이므로 직접 확인할 수 없지만, 몸이 공중에 떠 있다는 징후는 없다. 갓은 머리에 얹혀 끈으로 고정되어 있고, 옷과 끈은 중력에 따라 내려온다. 하단의 등과 가구는 바닥 쪽에 놓여 있다."
       },
       {
        "label": "B",
        "direction": "박상현은 왼쪽 인물 쪽으로 얼굴을 돌리고 입을 조금 벌렸다. 시선은 상대의 높은 얼굴보다는 턱 아래와 상체 쪽으로 향한다. 손은 그 인물의 위팔을 붙잡고 있어 대화 대상은 분명하지만, 상체를 바짝 기울이는 정도는 A보다 약하다.",
        "built_space": "목재 틀과 종이 벽면 뒤로 낮은 가구가 좌우에 하나씩 보이고, 앞쪽 낮은 상 위에는 놋촛대 하나와 도자기 그릇 하나가 있다. 왼쪽에는 상대에게 가려진 불빛이, 오른쪽 뒤 가구 위에는 작은 용기 하나가 보인다. 참고 장면과 재질은 유사하지만 이 집기 배치는 확인되지 않는다. 닫힌 출입문 전체와 두 사람의 바닥 접촉점은 보이지 않는다. 반사는 없다.",
        "entities": "오른쪽 박상현의 성인 남성 외모, 수염, 검은 갓과 망건, 회색 도포는 참고 이미지와 대체로 맞는다. 왼쪽에는 검은 머리와 짙은 청색 옷을 입은 별도의 성인 인물이 얼굴·몸통 일부까지 등장한다. 상대를 박제현으로 의도했더라도 박상현만 가시 인물로 허용한 명시적 제한과 충돌한다. 읽을 수 있는 글자는 없다.",
        "hard_violations": [
         "가시 인물을 박상현으로만 제한했는데 왼쪽에 별도 인물의 얼굴과 몸통을 추가했다."
        ],
        "physics": "박상현의 손은 상대 위팔 표면에 닿아 실제로 붙잡고 있으며 팔과 소매도 연결되어 있다. 두 사람의 하체 지지점은 프레임 밖이지만 상체가 떠 있거나 불가능하게 꺾인 모습은 아니다. 갓은 머리가 지지하고, 촛대와 그릇은 낮은 상 위에 놓여 있다."
       }
      ],
      "all_candidates_fail": false
     },
     "normalized": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 7,
        "verdict_ko": "박상현 혼자 몸을 앞으로 깊이 기울이고 입을 벌린 측면 연기는 충실하지만, 얼굴보다 상체와 배경을 넓게 담아 지정된 클로즈업에는 다소 느슨하다."
       },
       {
        "label": "A",
        "score": 3,
        "verdict_ko": "얼굴 중심의 근접 구도와 속삭이는 표정은 맞지만, 박상현만 보여야 하는 화면에 상대 인물을 추가하고 팔을 붙잡는 동작까지 만들어 인물 제한을 위반한다."
       }
      ],
      "all_candidates_fail": false,
      "readings": [
       {
        "label": "B",
        "direction": "박상현의 얼굴과 시선, 앞으로 숙인 상체가 모두 화면 왼쪽 아래의 화면 밖 상대를 향한다. 박제현은 보이지 않아 실제 눈맞춤은 확인할 수 없지만, 화면 밖 박제현에게 낮게 속삭이는 방향으로 읽힌다. 입은 반쯤 열려 있다.",
        "built_space": "배경에는 낡은 목재 기둥과 종이 면, 고리 손잡이 하나가 달린 닫힌 문짝 하나가 보인다. 아래에는 사각 등 하나와 왼쪽 가장자리의 낮은 가구 일부가 있다. 목재와 종이의 재질은 이전 장면과 대체로 연결되지만, 등과 세부 문 구성은 참고 화면에서 확인되지 않는다. 인물은 바닥 가까운 높이에 있으나 좌면과 하체는 잘려 있다. 거울이나 반사는 없다.",
        "entities": "성인 한국인 남성으로 제시된 박상현 한 명만 보인다. 얼굴 윤곽과 콧수염·턱수염, 검은 갓과 망건, 회색 도포와 밝은 깃이 참고 인물에 대체로 부합한다. 머리카락은 대부분 가려져 있고 드러난 부분은 검다. 박제현과 마당 너머 곡간은 프레임 밖이며, 읽을 수 있는 글자는 없다.",
        "hard_violations": [],
        "physics": "목과 어깨, 상체가 함께 앞으로 굽혀져 실제로 가능한 숙임이다. 골반과 발의 지지점은 화면 밖이므로 직접 확인할 수 없지만, 몸이 공중에 떠 있다는 징후는 없다. 갓은 머리에 얹혀 끈으로 고정되어 있고, 옷과 끈은 중력에 따라 내려온다. 하단의 등과 가구는 바닥 쪽에 놓여 있다."
       },
       {
        "label": "A",
        "direction": "박상현은 왼쪽 인물 쪽으로 얼굴을 돌리고 입을 조금 벌렸다. 시선은 상대의 높은 얼굴보다는 턱 아래와 상체 쪽으로 향한다. 손은 그 인물의 위팔을 붙잡고 있어 대화 대상은 분명하지만, 상체를 바짝 기울이는 정도는 A보다 약하다.",
        "built_space": "목재 틀과 종이 벽면 뒤로 낮은 가구가 좌우에 하나씩 보이고, 앞쪽 낮은 상 위에는 놋촛대 하나와 도자기 그릇 하나가 있다. 왼쪽에는 상대에게 가려진 불빛이, 오른쪽 뒤 가구 위에는 작은 용기 하나가 보인다. 참고 장면과 재질은 유사하지만 이 집기 배치는 확인되지 않는다. 닫힌 출입문 전체와 두 사람의 바닥 접촉점은 보이지 않는다. 반사는 없다.",
        "entities": "오른쪽 박상현의 성인 남성 외모, 수염, 검은 갓과 망건, 회색 도포는 참고 이미지와 대체로 맞는다. 왼쪽에는 검은 머리와 짙은 청색 옷을 입은 별도의 성인 인물이 얼굴·몸통 일부까지 등장한다. 상대를 박제현으로 의도했더라도 박상현만 가시 인물로 허용한 명시적 제한과 충돌한다. 읽을 수 있는 글자는 없다.",
        "hard_violations": [
         "가시 인물을 박상현으로만 제한했는데 왼쪽에 별도 인물의 얼굴과 몸통을 추가했다."
        ],
        "physics": "박상현의 손은 상대 위팔 표면에 닿아 실제로 붙잡고 있으며 팔과 소매도 연결되어 있다. 두 사람의 하체 지지점은 프레임 밖이지만 상체가 떠 있거나 불가능하게 꺾인 모습은 아니다. 갓은 머리가 지지하고, 촛대와 그릇은 낮은 상 위에 놓여 있다."
       }
      ]
     },
     "ok": true
    }
   ],
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "slot_winner_match": true,
   "slot_winner": {
    "gemini-pro": "B",
    "gpt-high": "B"
   },
   "route": "cross_slot_agree"
  },
  "dual": {
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "normalized": {
    "A": 0.429,
    "B": 2.0
   },
   "adjusted": {
    "A": 0.179,
    "B": 2.0
   },
   "violations": {
    "A": [
     "[gemini-pro] invented people (프레임 내 허용되지 않은 추가 인물 등장)",
     "[gpt-high] 가시 인물을 박상현으로만 제한했는데 왼쪽에 별도 인물의 얼굴과 몸통을 추가했다."
    ]
   },
   "per_model_winner": {
    "gemini-pro": "B",
    "gpt-high": "B"
   },
   "agreed": true
  },
  "totals": {
   "B": 2000,
   "A": 179
  },
  "selected": "B",
  "ranking": [
   "B",
   "A"
  ],
  "verdicts": [
   {
    "label": "B",
    "score": 2000,
    "verdict_ko": "허용된 인물만 프레임에 유지하면서, 화면 밖의 대상을 향해 몸을 기울이고 입을 반쯤 벌린 박상현의 측면 모습을 샷 텍스트에 맞게 완벽하게 구현했습니다."
   },
   {
    "label": "A",
    "score": 179,
    "verdict_ko": "프레임 내에 허용 목록에 없는 다른 인물을 등장시킨 것은 명백한 하드 위반입니다.  ★위반: [gemini-pro] invented people (프레임 내 허용되지 않은 추가 인물 등장) / [gpt-high] 가시 인물을 박상현으로만 제한했는데 왼쪽에 별도 인물의 얼굴과 몸통을 추가했다."
   }
  ],
  "refs": [
   {
    "label": "PREVIOUS SHOT STILL — a visually related earlier shot of this same place: the location's look, materials, fixed features and lighting mood are LOCKED to this photo; never copy its camera framing. The clothing, hair and overall look of 박상현 — who appear both in that photo and in this shot — are LOCKED to that photo. Anyone else visible in that photo is NOT in this shot: never carry their face, body or clothing onto anyone here. If this photo shows a character who cannot move (dead or unconscious), that character's exact pose, position and orientation are ALSO LOCKED — treat their body as a fixed prop of the set that only the camera moves around.",
    "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/db00517b-165b-4005-a287-f28936b59b95/images/272bd7af-9294-4490-98c6-88a40e9deacc/scene/recipe/S20sh6_sel.png",
    "asset_id": "98fed19a-7f4e-4cc6-9cf1-b8f34d6726f8",
    "role": "prev_still"
   },
   {
    "label": "CHARACTER REFERENCE — 박상현: the exact person appearing in this shot; match face, hair and build exactly.",
    "path": "<bytes:840781>",
    "asset_id": "6187751d-5df3-4eb7-9d30-d6fc98a005e6",
    "role": "character_ref"
   }
  ],
  "critique_skipped": true,
  "shot_run_produced": false,
  "shot_run_uid": "06aa1428-40d9-7c27-ba62-76adb37f978b",
  "ref_mode": "prev+엔티티",
  "share_plan": {
   "ref_plan": "prev",
   "prev_anchor_tag": "S20sh6"
  }
 },
 "S33sh4::cine": {
  "applied": true,
  "attempted_at": "2026-09-09T04:33:25.975022+00:00",
  "fingerprint": "c2d965f8ab9853154844e2045c35cff6d25af37fa5b9a016e8b67cb8a73c8b7d",
  "fingerprint_version": 2,
  "provider": "mai",
  "endpoint": "openrouter/chat-completions",
  "model": "microsoft/mai-image-2.6",
  "pack": "24.202608252115",
  "source_file": "S33sh4_sel.png",
  "source_sha256": "64a2d942f90ebc3c403964c9f3bd8fcde28a868a10e587f6d702c496124bb65b",
  "file": "S33sh4_cine.png",
  "staged_sha256": "660b180f8892887731d43e43907d0625911671499dcdc526a4df3cfaab743740",
  "latency_ms": 24322
 },
 "S33sh8::signage": {
  "fp": "41f664469da55adb",
  "inscriptions": [],
  "cues": [],
  "dropped": []
 },
 "S33sh8": {
  "input_fingerprint": "f83d26e252d795f4",
  "prompt": "Create ONE FINAL photorealistic live-action film still of the moment below — early 20th-century Korea, Jeolla region; all people are Korean unless explicitly stated otherwise. TIME OF DAY (lock): night.\n\nSHOT TEXT (authoritative, Korean): 박상현을 똑바로 응시한 채 단호하게 입을 연 박제현의 굳은 정면.\n\nLOCATION (lock): The quiet seating space inside the dim guest-wing room of a traditional estate at night. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: close-up\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Maintain the same dark nighttime interior and restrained contrast, keeping 박제현's firmness readable without an added dramatic light effect.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The people visible in that photo are NOT in this shot: never carry their faces, bodies or clothing onto anyone here. Each person in THIS shot is defined solely by the PEOPLE list, their own CHARACTER REFERENCE images and their wardrobe notes. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The sarangchae door remains closed; the grain-store door is still shut and sealed across the dark yard. 박제현: He remains seated.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 박제현 (한국인 남성, 중년 이상 나이대, 자연스러운 연령대의 얼굴) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.",
  "roll_prompts": {
   "A": "Create ONE FINAL photorealistic live-action film still of the moment below — early 20th-century Korea, Jeolla region; all people are Korean unless explicitly stated otherwise. TIME OF DAY (lock): night.\n\nSHOT TEXT (authoritative, Korean): 박상현을 똑바로 응시한 채 단호하게 입을 연 박제현의 굳은 정면.\n\nLOCATION (lock): The quiet seating space inside the dim guest-wing room of a traditional estate at night. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: close-up\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Maintain the same dark nighttime interior and restrained contrast, keeping 박제현's firmness readable without an added dramatic light effect.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The people visible in that photo are NOT in this shot: never carry their faces, bodies or clothing onto anyone here. Each person in THIS shot is defined solely by the PEOPLE list, their own CHARACTER REFERENCE images and their wardrobe notes. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The sarangchae door remains closed; the grain-store door is still shut and sealed across the dark yard. 박제현: He remains seated.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 박제현 (한국인 남성, 중년 이상 나이대, 자연스러운 연령대의 얼굴) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.",
   "B": "Create ONE FINAL photorealistic live-action film still of the moment below — early 20th-century Korea, Jeolla region; all people are Korean unless explicitly stated otherwise. TIME OF DAY (lock): night.\n\nSHOT TEXT (authoritative, Korean): 박상현을 똑바로 응시한 채 단호하게 입을 연 박제현의 굳은 정면.\n\nLOCATION (lock): The quiet seating space inside the dim guest-wing room of a traditional estate at night. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: close-up\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Maintain the same dark nighttime interior and restrained contrast, keeping 박제현's firmness readable without an added dramatic light effect.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The people visible in that photo are NOT in this shot: never carry their faces, bodies or clothing onto anyone here. Each person in THIS shot is defined solely by the PEOPLE list, their own CHARACTER REFERENCE images and their wardrobe notes. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The sarangchae door remains closed; the grain-store door is still shut and sealed across the dark yard. 박제현: He remains seated.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 박제현 (한국인 남성, 중년 이상 나이대, 자연스러운 연령대의 얼굴) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.\n\nCOMPOSITION VARIATION (this candidate only): keep the place, the people and the moment exactly as specified, and keep from the CAMERA & FRAME contract its subject scale, screen placement and key background elements exactly. For this candidate only, its camera angle and height are deliberately left open: choose a camera position distinctly different from the obvious one, as a film director picking a second setup on the same blocking. Stay true to any explicit framing the SHOT TEXT itself dictates."
  },
  "shot_run_spend_attempt_count": 1,
  "readings": [
   {
    "label": "A",
    "direction": "정면의 카메라 너머(박상현의 위치)를 똑바로 응시하고 있으며, 말을 하기 위해 입을 살짝 벌리고 있음.",
    "built_space": "이전 샷의 분위기를 유지하는 어두운 전통 가옥 실내로, 배경에 격자무늬 창문이 배치되어 있음.",
    "entities": "캐릭터 레퍼런스와 일치하는 외모와 복장(갓, 옅은 색 한복)을 한 박제현이 등장함.",
    "hard_violations": [],
    "physics": "어깨와 목의 자세로 보아 지시된 대로 안정적으로 앉아있는 상태를 유지하고 있음."
   },
   {
    "label": "B",
    "direction": "정면을 향해 시선을 고정하고 있으나, 지시문과 달리 입을 굳게 다물고 있음.",
    "built_space": "전통 가옥의 방 내부가 넓게 보이며, 장식장과 바닥에 놓인 조명 등 방의 구조가 드러남.",
    "entities": "캐릭터 레퍼런스와 일치하는 박제현이 갓과 한복을 착용하고 있음.",
    "hard_violations": [],
    "physics": "방바닥에 무릎을 꿇고 앉아 체중이 안정적으로 실려 있음."
   }
  ],
  "cross_model_order": {
   "policy": "cross_model_order_v2_slot0_fwd_slot1_rev_parallel2",
   "slots": [
    {
     "model": "gemini-pro",
     "order": "forward",
     "display_to_canonical": {
      "A": "A",
      "B": "B"
     },
     "raw": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 9,
        "verdict_ko": "클로즈업 프레이밍을 정확히 준수하였으며, 지시문(입을 연 박제현)의 행동을 훌륭하게 반영하였습니다."
       },
       {
        "label": "B",
        "score": 4,
        "verdict_ko": "클로즈업이라는 프레이밍 지시를 어기고 미디엄 샷으로 렌더링되었으며, 인물이 입을 다물고 있어 지시문과 어긋납니다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "정면의 카메라 너머(박상현의 위치)를 똑바로 응시하고 있으며, 말을 하기 위해 입을 살짝 벌리고 있음.",
        "built_space": "이전 샷의 분위기를 유지하는 어두운 전통 가옥 실내로, 배경에 격자무늬 창문이 배치되어 있음.",
        "entities": "캐릭터 레퍼런스와 일치하는 외모와 복장(갓, 옅은 색 한복)을 한 박제현이 등장함.",
        "hard_violations": [],
        "physics": "어깨와 목의 자세로 보아 지시된 대로 안정적으로 앉아있는 상태를 유지하고 있음."
       },
       {
        "label": "B",
        "direction": "정면을 향해 시선을 고정하고 있으나, 지시문과 달리 입을 굳게 다물고 있음.",
        "built_space": "전통 가옥의 방 내부가 넓게 보이며, 장식장과 바닥에 놓인 조명 등 방의 구조가 드러남.",
        "entities": "캐릭터 레퍼런스와 일치하는 박제현이 갓과 한복을 착용하고 있음.",
        "hard_violations": [],
        "physics": "방바닥에 무릎을 꿇고 앉아 체중이 안정적으로 실려 있음."
       }
      ],
      "all_candidates_fail": false
     },
     "normalized": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 9,
        "verdict_ko": "클로즈업 프레이밍을 정확히 준수하였으며, 지시문(입을 연 박제현)의 행동을 훌륭하게 반영하였습니다."
       },
       {
        "label": "B",
        "score": 4,
        "verdict_ko": "클로즈업이라는 프레이밍 지시를 어기고 미디엄 샷으로 렌더링되었으며, 인물이 입을 다물고 있어 지시문과 어긋납니다."
       }
      ],
      "all_candidates_fail": false,
      "readings": [
       {
        "label": "A",
        "direction": "정면의 카메라 너머(박상현의 위치)를 똑바로 응시하고 있으며, 말을 하기 위해 입을 살짝 벌리고 있음.",
        "built_space": "이전 샷의 분위기를 유지하는 어두운 전통 가옥 실내로, 배경에 격자무늬 창문이 배치되어 있음.",
        "entities": "캐릭터 레퍼런스와 일치하는 외모와 복장(갓, 옅은 색 한복)을 한 박제현이 등장함.",
        "hard_violations": [],
        "physics": "어깨와 목의 자세로 보아 지시된 대로 안정적으로 앉아있는 상태를 유지하고 있음."
       },
       {
        "label": "B",
        "direction": "정면을 향해 시선을 고정하고 있으나, 지시문과 달리 입을 굳게 다물고 있음.",
        "built_space": "전통 가옥의 방 내부가 넓게 보이며, 장식장과 바닥에 놓인 조명 등 방의 구조가 드러남.",
        "entities": "캐릭터 레퍼런스와 일치하는 박제현이 갓과 한복을 착용하고 있음.",
        "hard_violations": [],
        "physics": "방바닥에 무릎을 꿇고 앉아 체중이 안정적으로 실려 있음."
       }
      ]
     },
     "ok": true
    },
    {
     "model": "gpt-high",
     "order": "reverse",
     "display_to_canonical": {
      "A": "B",
      "B": "A"
     },
     "raw": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 5,
        "verdict_ko": "정면의 단호한 인상과 어두운 실내는 맞지만, 허리까지 보이는 넓은 구도와 닫힌 입이 ‘입을 연 굳은 정면 클로즈업’이라는 핵심 지시를 놓친다."
       },
       {
        "label": "B",
        "score": 8,
        "verdict_ko": "정면 응시와 실제로 말을 시작한 입 모양을 얼굴 중심 클로즈업으로 구현해 우세하나, 배경 창호와 얼굴 조명은 요구된 어두운 밤 분위기보다 밝다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "얼굴과 두 눈은 거의 렌즈 정면을 향한다. 화면 밖 카메라 쪽 박상현을 응시하는 배치로 해석할 수 있지만 상대가 보이지 않아 목표 인물을 직접 확인할 수는 없다. 입은 닫혀 있어 발화를 시작한 순간은 보이지 않는다.",
        "built_space": "왼쪽에 닫힌 창호문 한 구역, 인물 뒤와 오른쪽에 격자 창호 구역이 보인다. 바닥 등 하나가 왼쪽 뒤에 있고, 오른쪽에는 크기가 다른 선반 두 개, 왼쪽 아래에는 낮은 수납장이 있다. 낡은 목재와 종이 재질은 이전 장면과 유사하지만, 넓어진 실내와 선반 배치는 참고 사진으로 확인되지 않는다. 인물은 중앙에 있으며 허리까지 드러나 클로즈업보다 넓다. 마당과 곡물창고는 화면 밖이다.",
        "entities": "중년 한국인 남성으로 설정된 박제현 한 명만 보이며, 얼굴의 연령감과 윤곽은 인물 참고와 대체로 일치한다. 검은 갓과 머리띠, 갓끈, 미색 도포, 붉은 허리띠가 보인다. 이전 장면의 회색 옷 인물은 복제되지 않았다. 읽을 수 있는 글자는 없다.",
        "hard_violations": [],
        "physics": "머리는 목과 몸통에 자연스럽게 연결되고, 갓은 머리에 얹혀 있으며 끈은 아래로 늘어진다. 팔과 옷자락에도 중력에 맞는 처짐이 있다. 엉덩이와 좌면은 프레임 밖이므로 앉은 상태의 접촉점은 확인할 수 없지만, 부유나 불가능한 자세의 증거는 없다."
       },
       {
        "label": "B",
        "direction": "얼굴은 정면이고 두 눈은 카메라 바로 근처의 전방을 향한다. 화면 밖 박상현을 마주 보는 대화축으로 성립한다. 눈썹과 입 주변에 긴장이 있고 입술이 벌어져 치아가 드러나므로 단호하게 말을 시작하는 순간이 보인다.",
        "built_space": "배경 중앙에 닫힌 격자 창호 두 짝이 인물 양옆으로 드러나며, 양쪽 목재 기둥과 낡은 종이 벽면이 보인다. 왼쪽 아래에는 어두운 가구 일부만 걸린다. 얼굴과 어깨 중심의 가까운 구도이며 배경 구조물의 크기도 물리적으로 자연스럽다. 목재와 벽 재질은 이전 장면에 부합하지만 창호를 통과하는 밝은 빛은 밤의 어두움을 약화한다. 문이 열린 모습이나 불가능한 반사는 없다.",
        "entities": "박제현에 해당하는 중년 남성 한 명만 등장한다. 참고 인물의 얼굴 윤곽, 눈매와 연령감이 대체로 유지되고 검은 갓, 머리띠와 갓끈, 미색 도포가 보인다. 눈은 정상적인 홍채와 동공을 가진 자연스러운 형태다. 허리띠와 하체는 클로즈업 밖이므로 평가 대상이 아니며 읽을 수 있는 글자나 추가 인물은 없다.",
        "hard_violations": [],
        "physics": "목과 어깨가 머리를 자연스럽게 지지하고 갓은 머리에 안정적으로 놓여 있다. 갓끈과 도포 주름은 아래로 처진다. 발화 중 벌어진 턱과 입도 해부학적으로 자연스럽다. 좌면과 하체는 화면 밖이라 착석 접촉점은 확인되지 않지만, 보이는 자세에 착석과 충돌하는 요소나 지지 없이 떠 있는 물체는 없다."
       }
      ],
      "all_candidates_fail": false
     },
     "normalized": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 5,
        "verdict_ko": "정면의 단호한 인상과 어두운 실내는 맞지만, 허리까지 보이는 넓은 구도와 닫힌 입이 ‘입을 연 굳은 정면 클로즈업’이라는 핵심 지시를 놓친다."
       },
       {
        "label": "A",
        "score": 8,
        "verdict_ko": "정면 응시와 실제로 말을 시작한 입 모양을 얼굴 중심 클로즈업으로 구현해 우세하나, 배경 창호와 얼굴 조명은 요구된 어두운 밤 분위기보다 밝다."
       }
      ],
      "all_candidates_fail": false,
      "readings": [
       {
        "label": "B",
        "direction": "얼굴과 두 눈은 거의 렌즈 정면을 향한다. 화면 밖 카메라 쪽 박상현을 응시하는 배치로 해석할 수 있지만 상대가 보이지 않아 목표 인물을 직접 확인할 수는 없다. 입은 닫혀 있어 발화를 시작한 순간은 보이지 않는다.",
        "built_space": "왼쪽에 닫힌 창호문 한 구역, 인물 뒤와 오른쪽에 격자 창호 구역이 보인다. 바닥 등 하나가 왼쪽 뒤에 있고, 오른쪽에는 크기가 다른 선반 두 개, 왼쪽 아래에는 낮은 수납장이 있다. 낡은 목재와 종이 재질은 이전 장면과 유사하지만, 넓어진 실내와 선반 배치는 참고 사진으로 확인되지 않는다. 인물은 중앙에 있으며 허리까지 드러나 클로즈업보다 넓다. 마당과 곡물창고는 화면 밖이다.",
        "entities": "중년 한국인 남성으로 설정된 박제현 한 명만 보이며, 얼굴의 연령감과 윤곽은 인물 참고와 대체로 일치한다. 검은 갓과 머리띠, 갓끈, 미색 도포, 붉은 허리띠가 보인다. 이전 장면의 회색 옷 인물은 복제되지 않았다. 읽을 수 있는 글자는 없다.",
        "hard_violations": [],
        "physics": "머리는 목과 몸통에 자연스럽게 연결되고, 갓은 머리에 얹혀 있으며 끈은 아래로 늘어진다. 팔과 옷자락에도 중력에 맞는 처짐이 있다. 엉덩이와 좌면은 프레임 밖이므로 앉은 상태의 접촉점은 확인할 수 없지만, 부유나 불가능한 자세의 증거는 없다."
       },
       {
        "label": "A",
        "direction": "얼굴은 정면이고 두 눈은 카메라 바로 근처의 전방을 향한다. 화면 밖 박상현을 마주 보는 대화축으로 성립한다. 눈썹과 입 주변에 긴장이 있고 입술이 벌어져 치아가 드러나므로 단호하게 말을 시작하는 순간이 보인다.",
        "built_space": "배경 중앙에 닫힌 격자 창호 두 짝이 인물 양옆으로 드러나며, 양쪽 목재 기둥과 낡은 종이 벽면이 보인다. 왼쪽 아래에는 어두운 가구 일부만 걸린다. 얼굴과 어깨 중심의 가까운 구도이며 배경 구조물의 크기도 물리적으로 자연스럽다. 목재와 벽 재질은 이전 장면에 부합하지만 창호를 통과하는 밝은 빛은 밤의 어두움을 약화한다. 문이 열린 모습이나 불가능한 반사는 없다.",
        "entities": "박제현에 해당하는 중년 남성 한 명만 등장한다. 참고 인물의 얼굴 윤곽, 눈매와 연령감이 대체로 유지되고 검은 갓, 머리띠와 갓끈, 미색 도포가 보인다. 눈은 정상적인 홍채와 동공을 가진 자연스러운 형태다. 허리띠와 하체는 클로즈업 밖이므로 평가 대상이 아니며 읽을 수 있는 글자나 추가 인물은 없다.",
        "hard_violations": [],
        "physics": "목과 어깨가 머리를 자연스럽게 지지하고 갓은 머리에 안정적으로 놓여 있다. 갓끈과 도포 주름은 아래로 처진다. 발화 중 벌어진 턱과 입도 해부학적으로 자연스럽다. 좌면과 하체는 화면 밖이라 착석 접촉점은 확인되지 않지만, 보이는 자세에 착석과 충돌하는 요소나 지지 없이 떠 있는 물체는 없다."
       }
      ]
     },
     "ok": true
    }
   ],
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "slot_winner_match": true,
   "slot_winner": {
    "gemini-pro": "A",
    "gpt-high": "A"
   },
   "route": "cross_slot_agree"
  },
  "dual": {
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "normalized": {
    "A": 2.0,
    "B": 1.069
   },
   "adjusted": {
    "A": 2.0,
    "B": 1.069
   },
   "violations": {},
   "per_model_winner": {
    "gemini-pro": "A",
    "gpt-high": "A"
   },
   "agreed": true
  },
  "totals": {
   "A": 2000,
   "B": 1069
  },
  "selected": "A",
  "ranking": [
   "A",
   "B"
  ],
  "verdicts": [
   {
    "label": "A",
    "score": 2000,
    "verdict_ko": "클로즈업 프레이밍을 정확히 준수하였으며, 지시문(입을 연 박제현)의 행동을 훌륭하게 반영하였습니다."
   },
   {
    "label": "B",
    "score": 1069,
    "verdict_ko": "클로즈업이라는 프레이밍 지시를 어기고 미디엄 샷으로 렌더링되었으며, 인물이 입을 다물고 있어 지시문과 어긋납니다."
   }
  ],
  "refs": [
   {
    "label": "PREVIOUS SHOT STILL — a visually related earlier shot of this same place: the location's look, materials, fixed features and lighting mood are LOCKED to this photo; never copy its camera framing. The people visible in that photo are NOT in this shot: never carry their faces, bodies or clothing onto anyone here. Each person in THIS shot is defined solely by the PEOPLE list, their own CHARACTER REFERENCE images and their wardrobe notes. If this photo shows a character who cannot move (dead or unconscious), that character's exact pose, position and orientation are ALSO LOCKED — treat their body as a fixed prop of the set that only the camera moves around.",
    "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/db00517b-165b-4005-a287-f28936b59b95/images/272bd7af-9294-4490-98c6-88a40e9deacc/scene/recipe/S33sh4_sel.png",
    "asset_id": "41e4b3f0-cac6-4f4a-886d-36bcab35e9df",
    "role": "prev_still"
   },
   {
    "label": "CHARACTER REFERENCE — 박제현: the exact person appearing in this shot; match face, hair and build exactly.",
    "path": "<bytes:870304>",
    "asset_id": "02a56dab-2641-421d-89df-af22b5de63df",
    "role": "character_ref"
   }
  ],
  "critique_skipped": true,
  "shot_run_produced": false,
  "shot_run_uid": "06aa1428-4293-7c25-9fe6-dddf304c5c5a",
  "ref_mode": "prev+엔티티",
  "share_plan": {
   "ref_plan": "prev",
   "prev_anchor_tag": "S33sh4"
  }
 },
 "S33sh8::cine": {
  "applied": true,
  "attempted_at": "2026-09-09T04:34:35.074208+00:00",
  "fingerprint": "2351817df2e921d7cd3d78043321e795f2960d5ad10b63f4dad8980639de2e67",
  "fingerprint_version": 2,
  "provider": "mai",
  "endpoint": "openrouter/chat-completions",
  "model": "microsoft/mai-image-2.6",
  "pack": "24.202608252115",
  "source_file": "S33sh8_sel.png",
  "source_sha256": "de325b0cd43e75d2f5719db9f50aefc0d7eb9547ecd15c9f13bd996050d0344b",
  "file": "S33sh8_cine.png",
  "staged_sha256": "18cb2108711e87fa694a6d04313d841a78cc8633e37622900882cb92a106a59a",
  "latency_ms": 26351
 },
 "S33sh11::signage": {
  "fp": "347f3ca3b2cd3df7",
  "inscriptions": [],
  "cues": [],
  "dropped": []
 },
 "S33sh11": {
  "input_fingerprint": "52e3294efdaea9d7",
  "prompt": "Create ONE FINAL photorealistic live-action film still of the moment below — early 20th-century Korea, Jeolla region; all people are Korean unless explicitly stated otherwise. TIME OF DAY (lock): night.\n\nSHOT TEXT (authoritative, Korean): 박제현의 시선 너머, 마당 구석의 굳게 닫힌 곳간 문 클로즈업.\n\nLOCATION (lock): The exterior face of the tightly closed granary door across the estate's dark courtyard. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: close-up\n- KEY BACKGROUND ELEMENTS: Storehouse door (Firmly closed) — The courtyard-facing side is visible at a shallow oblique angle from the side nearest the room; used as Central focal anchor for 박제현's preceding look; Storehouse frontage (Visible immediately beside the closed door) — Seen obliquely along the same plane as the entrance; used as Provide context without introducing unsupported construction details; Courtyard (Separating the room from the storehouse); used as A lower margin grounds the door in the intervening space.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Use subdued ambient exposure appropriate to the nighttime courtyard, without extending an unestablished interior light source outdoors.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The people visible in that photo are NOT in this shot: never carry their faces, bodies or clothing onto anyone here. Each person in THIS shot is defined solely by the PEOPLE list, their own CHARACTER REFERENCE images and their wardrobe notes. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The sarangchae door is closed again after the exit; the grain-store door across the yard remains firmly shut with its red seal.\n\nNO PEOPLE IN THIS SHOT: the shot text shows only the place and its state — no living person appears in frame, and no face, and no figure standing, sitting or passing through. Do not invent someone to occupy the space.\n\nOne thing this rule does NOT forbid: the hand or arm of whoever is doing what the shot text describes. If the moment is an object being held, carried, operated, read, aimed, worn or handed over, the hand doing it belongs in frame, resting on or gripping that object exactly as its real use requires. An object photographed floating unsupported, with the hand that must be holding it erased, is a worse failure than the person you were avoiding — render the hand.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.",
  "roll_prompts": {
   "A": "Create ONE FINAL photorealistic live-action film still of the moment below — early 20th-century Korea, Jeolla region; all people are Korean unless explicitly stated otherwise. TIME OF DAY (lock): night.\n\nSHOT TEXT (authoritative, Korean): 박제현의 시선 너머, 마당 구석의 굳게 닫힌 곳간 문 클로즈업.\n\nLOCATION (lock): The exterior face of the tightly closed granary door across the estate's dark courtyard. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: close-up\n- KEY BACKGROUND ELEMENTS: Storehouse door (Firmly closed) — The courtyard-facing side is visible at a shallow oblique angle from the side nearest the room; used as Central focal anchor for 박제현's preceding look; Storehouse frontage (Visible immediately beside the closed door) — Seen obliquely along the same plane as the entrance; used as Provide context without introducing unsupported construction details; Courtyard (Separating the room from the storehouse); used as A lower margin grounds the door in the intervening space.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Use subdued ambient exposure appropriate to the nighttime courtyard, without extending an unestablished interior light source outdoors.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The people visible in that photo are NOT in this shot: never carry their faces, bodies or clothing onto anyone here. Each person in THIS shot is defined solely by the PEOPLE list, their own CHARACTER REFERENCE images and their wardrobe notes. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The sarangchae door is closed again after the exit; the grain-store door across the yard remains firmly shut with its red seal.\n\nNO PEOPLE IN THIS SHOT: the shot text shows only the place and its state — no living person appears in frame, and no face, and no figure standing, sitting or passing through. Do not invent someone to occupy the space.\n\nOne thing this rule does NOT forbid: the hand or arm of whoever is doing what the shot text describes. If the moment is an object being held, carried, operated, read, aimed, worn or handed over, the hand doing it belongs in frame, resting on or gripping that object exactly as its real use requires. An object photographed floating unsupported, with the hand that must be holding it erased, is a worse failure than the person you were avoiding — render the hand.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.",
   "B": "Create ONE FINAL photorealistic live-action film still of the moment below — early 20th-century Korea, Jeolla region; all people are Korean unless explicitly stated otherwise. TIME OF DAY (lock): night.\n\nSHOT TEXT (authoritative, Korean): 박제현의 시선 너머, 마당 구석의 굳게 닫힌 곳간 문 클로즈업.\n\nLOCATION (lock): The exterior face of the tightly closed granary door across the estate's dark courtyard. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: close-up\n- KEY BACKGROUND ELEMENTS: Storehouse door (Firmly closed) — The courtyard-facing side is visible at a shallow oblique angle from the side nearest the room; used as Central focal anchor for 박제현's preceding look; Storehouse frontage (Visible immediately beside the closed door) — Seen obliquely along the same plane as the entrance; used as Provide context without introducing unsupported construction details; Courtyard (Separating the room from the storehouse); used as A lower margin grounds the door in the intervening space.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Use subdued ambient exposure appropriate to the nighttime courtyard, without extending an unestablished interior light source outdoors.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The people visible in that photo are NOT in this shot: never carry their faces, bodies or clothing onto anyone here. Each person in THIS shot is defined solely by the PEOPLE list, their own CHARACTER REFERENCE images and their wardrobe notes. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The sarangchae door is closed again after the exit; the grain-store door across the yard remains firmly shut with its red seal.\n\nNO PEOPLE IN THIS SHOT: the shot text shows only the place and its state — no living person appears in frame, and no face, and no figure standing, sitting or passing through. Do not invent someone to occupy the space.\n\nOne thing this rule does NOT forbid: the hand or arm of whoever is doing what the shot text describes. If the moment is an object being held, carried, operated, read, aimed, worn or handed over, the hand doing it belongs in frame, resting on or gripping that object exactly as its real use requires. An object photographed floating unsupported, with the hand that must be holding it erased, is a worse failure than the person you were avoiding — render the hand.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.\n\nCOMPOSITION VARIATION (this candidate only): keep the place, the people and the moment exactly as specified, and keep from the CAMERA & FRAME contract its subject scale, screen placement and key background elements exactly. For this candidate only, its camera angle and height are deliberately left open: choose a camera position distinctly different from the obvious one, as a film director picking a second setup on the same blocking. Stay true to any explicit framing the SHOT TEXT itself dictates."
  },
  "shot_run_spend_attempt_count": 1,
  "readings": [
   {
    "label": "A",
    "direction": "카메라는 곳간 문을 향하고 있음.",
    "built_space": "나무 문과 붉은 종이 봉인이 보이나, 화면 좌측에 마당 바닥과 기둥 같은 형상이 반투명한 유령처럼 문 위에 겹쳐서 나타남.",
    "entities": "나무 문, 붉은색 봉인 종이.",
    "hard_violations": [
     "[gemini-pro] 물리적으로 불가능한 화면 구성 (단단한 문 표면 위에 마당 배경이 반투명하게 겹쳐진 이중 노출 오류)",
     "[gpt-high] 문과 봉인 위에 별도의 마당 영상이 겹쳐 보이는 합성형 공간으로, 하나의 물리적 장소를 촬영한 실사 프레임이 아니다.",
     "[gpt-high] 봉인의 노출된 부분에 판독 가능한 글자가 있어 무문자 조건을 위반한다."
    ],
    "physics": "좌측 절반의 배경 이미지가 전경의 사물 위에 겹쳐 떠 있는 상태로, 물리적 재질감과 광학을 완전히 위반함."
   },
   {
    "label": "B",
    "direction": "카메라는 비스듬한 각도로 곳간 문을 향하고 있음.",
    "built_space": "나무로 된 곳간 문, 붉은 봉인, 쇠로 된 빗장, 우측의 돌과 흙으로 된 벽면, 하단의 마당 흙바닥이 지시된 대로 공간적으로 올바르게 배치됨.",
    "entities": "굳게 닫힌 나무 문, 붉은색 봉인 종이, 금속 빗장, 돌담.",
    "hard_violations": [
     "[gpt-high] 참조에서 확인되는 봉인 아래 문 표면에 없던 금속 빗장과 걸쇠를 새로 추가했다.",
     "[gpt-high] 붉은 봉인에 판독 가능한 글자가 노출되어 읽을 수 있는 문자를 금지한 조건을 위반한다."
    ],
    "physics": "모든 사물과 구조물이 정상적인 중력과 물리적 형태를 유지하고 있음."
   }
  ],
  "cross_model_order": {
   "policy": "cross_model_order_v2_slot0_fwd_slot1_rev_parallel2",
   "slots": [
    {
     "model": "gemini-pro",
     "order": "forward",
     "display_to_canonical": {
      "A": "A",
      "B": "B"
     },
     "raw": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 9,
        "verdict_ko": "프롬프트가 요구한 비스듬한 앵글의 곳간 문, 붉은 봉인, 주변 벽면과 마당의 바닥 여백을 물리적으로 자연스럽게 잘 구현했습니다."
       },
       {
        "label": "A",
        "score": 0,
        "verdict_ko": "마당의 배경 이미지가 단단한 나무 문과 붉은 부적 위에 반투명하게 겹쳐서 렌더링된 이중 노출 오류가 발생하여 실사 촬영 규칙을 크게 위반했습니다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "카메라는 곳간 문을 향하고 있음.",
        "built_space": "나무 문과 붉은 종이 봉인이 보이나, 화면 좌측에 마당 바닥과 기둥 같은 형상이 반투명한 유령처럼 문 위에 겹쳐서 나타남.",
        "entities": "나무 문, 붉은색 봉인 종이.",
        "hard_violations": [
         "물리적으로 불가능한 화면 구성 (단단한 문 표면 위에 마당 배경이 반투명하게 겹쳐진 이중 노출 오류)"
        ],
        "physics": "좌측 절반의 배경 이미지가 전경의 사물 위에 겹쳐 떠 있는 상태로, 물리적 재질감과 광학을 완전히 위반함."
       },
       {
        "label": "B",
        "direction": "카메라는 비스듬한 각도로 곳간 문을 향하고 있음.",
        "built_space": "나무로 된 곳간 문, 붉은 봉인, 쇠로 된 빗장, 우측의 돌과 흙으로 된 벽면, 하단의 마당 흙바닥이 지시된 대로 공간적으로 올바르게 배치됨.",
        "entities": "굳게 닫힌 나무 문, 붉은색 봉인 종이, 금속 빗장, 돌담.",
        "hard_violations": [],
        "physics": "모든 사물과 구조물이 정상적인 중력과 물리적 형태를 유지하고 있음."
       }
      ],
      "all_candidates_fail": false
     },
     "normalized": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 9,
        "verdict_ko": "프롬프트가 요구한 비스듬한 앵글의 곳간 문, 붉은 봉인, 주변 벽면과 마당의 바닥 여백을 물리적으로 자연스럽게 잘 구현했습니다."
       },
       {
        "label": "A",
        "score": 0,
        "verdict_ko": "마당의 배경 이미지가 단단한 나무 문과 붉은 부적 위에 반투명하게 겹쳐서 렌더링된 이중 노출 오류가 발생하여 실사 촬영 규칙을 크게 위반했습니다."
       }
      ],
      "all_candidates_fail": false,
      "readings": [
       {
        "label": "A",
        "direction": "카메라는 곳간 문을 향하고 있음.",
        "built_space": "나무 문과 붉은 종이 봉인이 보이나, 화면 좌측에 마당 바닥과 기둥 같은 형상이 반투명한 유령처럼 문 위에 겹쳐서 나타남.",
        "entities": "나무 문, 붉은색 봉인 종이.",
        "hard_violations": [
         "물리적으로 불가능한 화면 구성 (단단한 문 표면 위에 마당 배경이 반투명하게 겹쳐진 이중 노출 오류)"
        ],
        "physics": "좌측 절반의 배경 이미지가 전경의 사물 위에 겹쳐 떠 있는 상태로, 물리적 재질감과 광학을 완전히 위반함."
       },
       {
        "label": "B",
        "direction": "카메라는 비스듬한 각도로 곳간 문을 향하고 있음.",
        "built_space": "나무로 된 곳간 문, 붉은 봉인, 쇠로 된 빗장, 우측의 돌과 흙으로 된 벽면, 하단의 마당 흙바닥이 지시된 대로 공간적으로 올바르게 배치됨.",
        "entities": "굳게 닫힌 나무 문, 붉은색 봉인 종이, 금속 빗장, 돌담.",
        "hard_violations": [],
        "physics": "모든 사물과 구조물이 정상적인 중력과 물리적 형태를 유지하고 있음."
       }
      ]
     },
     "ok": true
    },
    {
     "model": "gpt-high",
     "order": "reverse",
     "display_to_canonical": {
      "A": "B",
      "B": "A"
     },
     "raw": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 4,
        "verdict_ko": "닫힌 문을 비스듬히 잡고 옆 벽과 마당 하단을 포함해 구도는 더 충실하지만, 참조에 없는 쇠빗장과 판독 가능한 봉인 글자가 연속성 및 무문자 조건을 위반한다."
       },
       {
        "label": "B",
        "score": 2,
        "verdict_ko": "참조의 목재와 봉인은 유지하지만, 문 위로 마당 영상이 겹치는 합성형 공간과 읽히는 글자 때문에 단일 실사 장면으로 사용할 수 없다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "인물이나 시선, 무기, 이동체는 없다. 카메라는 닫힌 두 짝 문의 바깥 면을 비스듬히 향하며, 중앙 이음매를 가로지르는 붉은 봉인이 주목점이다. 방이 있는 쪽에서 보는지는 화면만으로 확인할 수 없다.",
        "built_space": "닫힌 두 짝 나무문 한 벌, 중앙 봉인 한 장, 그 아래 금속 빗장 한 벌과 매달린 걸쇠가 보인다. 오른쪽에는 문틀과 돌·흙 벽이 이어지고 오른쪽 하단에는 흙마당이 좁게 들어온다. 문과 전면 벽의 공간 관계는 자연스럽지만, 참조에서 봉인 아래로 보이던 목재 면에는 이 금속 잠금장치가 없다.",
        "entities": "사람은 없으며 오래된 세로 판재 곳간 문과 붉은 종이 봉인은 요청에 맞는다. 봉인의 검은 원형 인장과 글자도 남아 있으나 일부 낱글자가 구별되어 무문자 조건에 맞지 않는다. 어두운 청회색 야간 노출과 거친 목재 질감은 참조에 가깝다. 사랑채 문은 화면 밖이므로 상태를 판단할 수 없다.",
        "hard_violations": [
         "참조에서 확인되는 봉인 아래 문 표면에 없던 금속 빗장과 걸쇠를 새로 추가했다.",
         "붉은 봉인에 판독 가능한 글자가 노출되어 읽을 수 있는 문자를 금지한 조건을 위반한다."
        ],
        "physics": "봉인은 양쪽 문짝 표면에 붙어 있고 주름진 부분도 부착면에서 이어진다. 빗장은 문에 고정된 고리들이 받치고 걸쇠는 아래로 매달려 있다. 문과 벽은 지면에 세워져 있으며, 지지 없이 떠 있는 물체나 신체는 없다."
       },
       {
        "label": "B",
        "direction": "인물이나 실제 시선은 없다. 카메라는 문의 중앙 틈과 붉은 봉인을 거의 참조와 같은 근접 각도로 향한다. 왼쪽에 겹친 흐릿한 마당 영상 때문에 닫힌 문을 향한 단일 시점이 흐트러진다.",
        "built_space": "주된 영상에는 두 짝 나무문의 중앙 부분과 봉인 한 장, 하단 가로 보강목 일부가 있다. 왼쪽 아래에는 마당 바닥과 건물 같은 다른 깊이의 영상이 문 표면을 덮으며 봉인 왼쪽까지 흐리게 가린다. 옆 전면 벽과 마당이 문에 물리적으로 이어지는 구도가 아니라 이중노출처럼 겹쳐 보인다.",
        "entities": "확인되는 사람은 없다. 풍화된 목재 문과 붉은 봉인의 모양·배치는 참조에 매우 가깝고 야간 색조도 유지한다. 그러나 봉인 오른쪽과 중앙에 구별 가능한 글자가 남아 있다. 사랑채 문은 식별되지 않는다.",
        "hard_violations": [
         "문과 봉인 위에 별도의 마당 영상이 겹쳐 보이는 합성형 공간으로, 하나의 물리적 장소를 촬영한 실사 프레임이 아니다.",
         "봉인의 노출된 부분에 판독 가능한 글자가 있어 무문자 조건을 위반한다."
        ],
        "physics": "주된 문의 봉인은 목재에 부착되어 있으며 공중에 뜬 물체는 없다. 다만 불투명한 나무문 표면에서 다른 공간의 바닥과 건물이 겹쳐 나타나는 모습은 이 재질의 반사나 정상적인 가림 관계로 설명되지 않는다."
       }
      ],
      "all_candidates_fail": true
     },
     "normalized": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 4,
        "verdict_ko": "닫힌 문을 비스듬히 잡고 옆 벽과 마당 하단을 포함해 구도는 더 충실하지만, 참조에 없는 쇠빗장과 판독 가능한 봉인 글자가 연속성 및 무문자 조건을 위반한다."
       },
       {
        "label": "A",
        "score": 2,
        "verdict_ko": "참조의 목재와 봉인은 유지하지만, 문 위로 마당 영상이 겹치는 합성형 공간과 읽히는 글자 때문에 단일 실사 장면으로 사용할 수 없다."
       }
      ],
      "all_candidates_fail": true,
      "readings": [
       {
        "label": "B",
        "direction": "인물이나 시선, 무기, 이동체는 없다. 카메라는 닫힌 두 짝 문의 바깥 면을 비스듬히 향하며, 중앙 이음매를 가로지르는 붉은 봉인이 주목점이다. 방이 있는 쪽에서 보는지는 화면만으로 확인할 수 없다.",
        "built_space": "닫힌 두 짝 나무문 한 벌, 중앙 봉인 한 장, 그 아래 금속 빗장 한 벌과 매달린 걸쇠가 보인다. 오른쪽에는 문틀과 돌·흙 벽이 이어지고 오른쪽 하단에는 흙마당이 좁게 들어온다. 문과 전면 벽의 공간 관계는 자연스럽지만, 참조에서 봉인 아래로 보이던 목재 면에는 이 금속 잠금장치가 없다.",
        "entities": "사람은 없으며 오래된 세로 판재 곳간 문과 붉은 종이 봉인은 요청에 맞는다. 봉인의 검은 원형 인장과 글자도 남아 있으나 일부 낱글자가 구별되어 무문자 조건에 맞지 않는다. 어두운 청회색 야간 노출과 거친 목재 질감은 참조에 가깝다. 사랑채 문은 화면 밖이므로 상태를 판단할 수 없다.",
        "hard_violations": [
         "참조에서 확인되는 봉인 아래 문 표면에 없던 금속 빗장과 걸쇠를 새로 추가했다.",
         "붉은 봉인에 판독 가능한 글자가 노출되어 읽을 수 있는 문자를 금지한 조건을 위반한다."
        ],
        "physics": "봉인은 양쪽 문짝 표면에 붙어 있고 주름진 부분도 부착면에서 이어진다. 빗장은 문에 고정된 고리들이 받치고 걸쇠는 아래로 매달려 있다. 문과 벽은 지면에 세워져 있으며, 지지 없이 떠 있는 물체나 신체는 없다."
       },
       {
        "label": "A",
        "direction": "인물이나 실제 시선은 없다. 카메라는 문의 중앙 틈과 붉은 봉인을 거의 참조와 같은 근접 각도로 향한다. 왼쪽에 겹친 흐릿한 마당 영상 때문에 닫힌 문을 향한 단일 시점이 흐트러진다.",
        "built_space": "주된 영상에는 두 짝 나무문의 중앙 부분과 봉인 한 장, 하단 가로 보강목 일부가 있다. 왼쪽 아래에는 마당 바닥과 건물 같은 다른 깊이의 영상이 문 표면을 덮으며 봉인 왼쪽까지 흐리게 가린다. 옆 전면 벽과 마당이 문에 물리적으로 이어지는 구도가 아니라 이중노출처럼 겹쳐 보인다.",
        "entities": "확인되는 사람은 없다. 풍화된 목재 문과 붉은 봉인의 모양·배치는 참조에 매우 가깝고 야간 색조도 유지한다. 그러나 봉인 오른쪽과 중앙에 구별 가능한 글자가 남아 있다. 사랑채 문은 식별되지 않는다.",
        "hard_violations": [
         "문과 봉인 위에 별도의 마당 영상이 겹쳐 보이는 합성형 공간으로, 하나의 물리적 장소를 촬영한 실사 프레임이 아니다.",
         "봉인의 노출된 부분에 판독 가능한 글자가 있어 무문자 조건을 위반한다."
        ],
        "physics": "주된 문의 봉인은 목재에 부착되어 있으며 공중에 뜬 물체는 없다. 다만 불투명한 나무문 표면에서 다른 공간의 바닥과 건물이 겹쳐 나타나는 모습은 이 재질의 반사나 정상적인 가림 관계로 설명되지 않는다."
       }
      ]
     },
     "ok": true
    }
   ],
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "slot_winner_match": true,
   "slot_winner": {
    "gemini-pro": "B",
    "gpt-high": "B"
   },
   "route": "cross_slot_agree"
  },
  "dual": {
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "normalized": {
    "A": 0.5,
    "B": 2.0
   },
   "adjusted": {
    "A": 0.25,
    "B": 1.75
   },
   "violations": {
    "A": [
     "[gemini-pro] 물리적으로 불가능한 화면 구성 (단단한 문 표면 위에 마당 배경이 반투명하게 겹쳐진 이중 노출 오류)",
     "[gpt-high] 문과 봉인 위에 별도의 마당 영상이 겹쳐 보이는 합성형 공간으로, 하나의 물리적 장소를 촬영한 실사 프레임이 아니다.",
     "[gpt-high] 봉인의 노출된 부분에 판독 가능한 글자가 있어 무문자 조건을 위반한다."
    ],
    "B": [
     "[gpt-high] 참조에서 확인되는 봉인 아래 문 표면에 없던 금속 빗장과 걸쇠를 새로 추가했다.",
     "[gpt-high] 붉은 봉인에 판독 가능한 글자가 노출되어 읽을 수 있는 문자를 금지한 조건을 위반한다."
    ]
   },
   "per_model_winner": {
    "gemini-pro": "B",
    "gpt-high": "B"
   },
   "agreed": true
  },
  "totals": {
   "B": 1750,
   "A": 250
  },
  "selected": "B",
  "ranking": [
   "B",
   "A"
  ],
  "verdicts": [
   {
    "label": "B",
    "score": 1750,
    "verdict_ko": "프롬프트가 요구한 비스듬한 앵글의 곳간 문, 붉은 봉인, 주변 벽면과 마당의 바닥 여백을 물리적으로 자연스럽게 잘 구현했습니다.  ★위반: [gpt-high] 참조에서 확인되는 봉인 아래 문 표면에 없던 금속 빗장과 걸쇠를 새로 추가했다. / [gpt-high] 붉은 봉인에 판독 가능한 글자가 노출되어 읽을 수 있는 문자를 금지한 조건을 위반한다."
   },
   {
    "label": "A",
    "score": 250,
    "verdict_ko": "마당의 배경 이미지가 단단한 나무 문과 붉은 부적 위에 반투명하게 겹쳐서 렌더링된 이중 노출 오류가 발생하여 실사 촬영 규칙을 크게 위반했습니다.  ★위반: [gemini-pro] 물리적으로 불가능한 화면 구성 (단단한 문 표면 위에 마당 배경이 반투명하게 겹쳐진 이중 노출 오류) / [gpt-high] 문과 봉인 위에 별도의 마당 영상이 겹쳐 보이는 합성형 공간으로, 하나의 물리적 장소를 촬영한 실사 프레임이 아니다. / [gpt-high] 봉인의 노출된 부분에 판독 가능한 글자가 있어 무문자 조건을 위반한다."
   }
  ],
  "refs": [
   {
    "label": "PREVIOUS SHOT STILL — a visually related earlier shot of this same place: the location's look, materials, fixed features and lighting mood are LOCKED to this photo; never copy its camera framing. The people visible in that photo are NOT in this shot: never carry their faces, bodies or clothing onto anyone here. Each person in THIS shot is defined solely by the PEOPLE list, their own CHARACTER REFERENCE images and their wardrobe notes. If this photo shows a character who cannot move (dead or unconscious), that character's exact pose, position and orientation are ALSO LOCKED — treat their body as a fixed prop of the set that only the camera moves around.",
    "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/db00517b-165b-4005-a287-f28936b59b95/images/272bd7af-9294-4490-98c6-88a40e9deacc/scene/recipe/S28sh18_sel.png",
    "asset_id": "ea1e6147-16d0-4438-9390-05c3a08b6e7e",
    "role": "prev_still"
   }
  ],
  "critique_skipped": true,
  "shot_run_produced": false,
  "shot_run_uid": "06aa1428-4422-7174-b1f6-d172ad985f27",
  "ref_mode": "prev만 (배경 전용·공유 계획)",
  "share_plan": {
   "ref_plan": "prev",
   "prev_anchor_tag": "S28sh18"
  },
  "lane_policy": "share_plan_prev_bgonly"
 },
 "S33sh11::cine": {
  "applied": true,
  "attempted_at": "2026-09-09T04:35:51.417723+00:00",
  "fingerprint": "6992344ca7784bbde5b0598e517bee64d74cb97570fcd40b0246a7bb15d131cb",
  "fingerprint_version": 2,
  "provider": "mai",
  "endpoint": "openrouter/chat-completions",
  "model": "microsoft/mai-image-2.6",
  "pack": "24.202608252115",
  "source_file": "S33sh11_sel.png",
  "source_sha256": "748e54487ae17276e6bd29789522e355ebf31cb451eb85471e873158ce18c376",
  "file": "S33sh11_cine.png",
  "staged_sha256": "570f6b98b89972929b1bdb208105c02407cd164e65f93a6f902016090d5ea4a9",
  "latency_ms": 27876
 },
 "S34sh8::signage": {
  "fp": "ad267b2a81054c7e",
  "inscriptions": [],
  "cues": [],
  "dropped": []
 },
 "S34sh8": {
  "input_fingerprint": "80c6c2f438a9a202",
  "prompt": "Create ONE FINAL photorealistic live-action film still of the moment below — early 20th-century Korea, Jeolla region; all people are Korean unless explicitly stated otherwise. TIME OF DAY (lock): night.\n\nSHOT TEXT (authoritative, Korean): 조선인 머슴이 내민 쌀가마니를 얼떨결에 양팔로 꽉 끌어안은 염재보의 멍한 정면.\n\nLOCATION (lock): Immediately outside the estate's opened granary doorway in the nighttime courtyard, where rice sacks are being handed out. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: medium shot\n- KEY BACKGROUND ELEMENTS: Rice sack (Being received and hugged by 염재보) — Held across his lower torso, with its weight supported by both forearms; used as Make the unsolicited assistance readable without obscuring his face; Storehouse door (Opened wide by the servant) — The open door is seen obliquely beside the entrance, leaving the opening unobstructed; used as Retain a background indication that access is being freely provided.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Maintain restrained nighttime ambient light and muted saturation, letting the awkward human exchange supply warmth without a comic lighting shift.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The people visible in that photo are NOT in this shot: never carry their faces, bodies or clothing onto anyone here. Each person in THIS shot is defined solely by the PEOPLE list, their own CHARACTER REFERENCE images and their wardrobe notes. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The main gate is ajar, the grain-store padlock unlocked and the store door wide open; rice bales are being brought toward a cart. 염재보: He holds a rice bale in his arms while retaining his gun. 조선인 머슴(박제현 집): His clothing is dirtied from deliberately rolling on the ground.\n\nTHE HAND THAT IS DOING THIS: the object at the centre of this shot is being held, operated, read, aimed or handed over by 조선인 머슴, 염재보 right now, so 조선인 머슴, 염재보's hand — and as much of the wrist and forearm as the framing reaches — is in the frame, gripping or resting on that object exactly the way its real use requires. Match that hand to 조선인 머슴, 염재보: its size, build, skin, age, grooming, sleeve and anything worn on it belong to that person and to no one else.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 염재보 (한국인 남성, 성인 나이대, 자연스러운 검은 머리); 조선인 머슴(박제현 집) (한국인 남성, 성인 나이대, 자연스러운 검은 머리) — wearing: 대지주 집안의 머슴답게 평범하고 수수한 무명 바지저고리. 다소 낡았으나 정돈되어 있다. — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.",
  "roll_prompts": {
   "A": "Create ONE FINAL photorealistic live-action film still of the moment below — early 20th-century Korea, Jeolla region; all people are Korean unless explicitly stated otherwise. TIME OF DAY (lock): night.\n\nSHOT TEXT (authoritative, Korean): 조선인 머슴이 내민 쌀가마니를 얼떨결에 양팔로 꽉 끌어안은 염재보의 멍한 정면.\n\nLOCATION (lock): Immediately outside the estate's opened granary doorway in the nighttime courtyard, where rice sacks are being handed out. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: medium shot\n- KEY BACKGROUND ELEMENTS: Rice sack (Being received and hugged by 염재보) — Held across his lower torso, with its weight supported by both forearms; used as Make the unsolicited assistance readable without obscuring his face; Storehouse door (Opened wide by the servant) — The open door is seen obliquely beside the entrance, leaving the opening unobstructed; used as Retain a background indication that access is being freely provided.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Maintain restrained nighttime ambient light and muted saturation, letting the awkward human exchange supply warmth without a comic lighting shift.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The people visible in that photo are NOT in this shot: never carry their faces, bodies or clothing onto anyone here. Each person in THIS shot is defined solely by the PEOPLE list, their own CHARACTER REFERENCE images and their wardrobe notes. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The main gate is ajar, the grain-store padlock unlocked and the store door wide open; rice bales are being brought toward a cart. 염재보: He holds a rice bale in his arms while retaining his gun. 조선인 머슴(박제현 집): His clothing is dirtied from deliberately rolling on the ground.\n\nTHE HAND THAT IS DOING THIS: the object at the centre of this shot is being held, operated, read, aimed or handed over by 조선인 머슴, 염재보 right now, so 조선인 머슴, 염재보's hand — and as much of the wrist and forearm as the framing reaches — is in the frame, gripping or resting on that object exactly the way its real use requires. Match that hand to 조선인 머슴, 염재보: its size, build, skin, age, grooming, sleeve and anything worn on it belong to that person and to no one else.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 염재보 (한국인 남성, 성인 나이대, 자연스러운 검은 머리); 조선인 머슴(박제현 집) (한국인 남성, 성인 나이대, 자연스러운 검은 머리) — wearing: 대지주 집안의 머슴답게 평범하고 수수한 무명 바지저고리. 다소 낡았으나 정돈되어 있다. — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.",
   "B": "Create ONE FINAL photorealistic live-action film still of the moment below — early 20th-century Korea, Jeolla region; all people are Korean unless explicitly stated otherwise. TIME OF DAY (lock): night.\n\nSHOT TEXT (authoritative, Korean): 조선인 머슴이 내민 쌀가마니를 얼떨결에 양팔로 꽉 끌어안은 염재보의 멍한 정면.\n\nLOCATION (lock): Immediately outside the estate's opened granary doorway in the nighttime courtyard, where rice sacks are being handed out. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: medium shot\n- KEY BACKGROUND ELEMENTS: Rice sack (Being received and hugged by 염재보) — Held across his lower torso, with its weight supported by both forearms; used as Make the unsolicited assistance readable without obscuring his face; Storehouse door (Opened wide by the servant) — The open door is seen obliquely beside the entrance, leaving the opening unobstructed; used as Retain a background indication that access is being freely provided.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Maintain restrained nighttime ambient light and muted saturation, letting the awkward human exchange supply warmth without a comic lighting shift.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The people visible in that photo are NOT in this shot: never carry their faces, bodies or clothing onto anyone here. Each person in THIS shot is defined solely by the PEOPLE list, their own CHARACTER REFERENCE images and their wardrobe notes. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The main gate is ajar, the grain-store padlock unlocked and the store door wide open; rice bales are being brought toward a cart. 염재보: He holds a rice bale in his arms while retaining his gun. 조선인 머슴(박제현 집): His clothing is dirtied from deliberately rolling on the ground.\n\nTHE HAND THAT IS DOING THIS: the object at the centre of this shot is being held, operated, read, aimed or handed over by 조선인 머슴, 염재보 right now, so 조선인 머슴, 염재보's hand — and as much of the wrist and forearm as the framing reaches — is in the frame, gripping or resting on that object exactly the way its real use requires. Match that hand to 조선인 머슴, 염재보: its size, build, skin, age, grooming, sleeve and anything worn on it belong to that person and to no one else.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 염재보 (한국인 남성, 성인 나이대, 자연스러운 검은 머리); 조선인 머슴(박제현 집) (한국인 남성, 성인 나이대, 자연스러운 검은 머리) — wearing: 대지주 집안의 머슴답게 평범하고 수수한 무명 바지저고리. 다소 낡았으나 정돈되어 있다. — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.\n\nCOMPOSITION VARIATION (this candidate only): keep the place, the people and the moment exactly as specified, and keep from the CAMERA & FRAME contract its subject scale, screen placement and key background elements exactly. For this candidate only, its camera angle and height are deliberately left open: choose a camera position distinctly different from the obvious one, as a film director picking a second setup on the same blocking. Stay true to any explicit framing the SHOT TEXT itself dictates."
  },
  "shot_run_spend_attempt_count": 1,
  "readings": [
   {
    "label": "A",
    "direction": "염재보는 정면을 향해 멍한 시선을 던지고 있으며, 머슴은 염재보와 쌀가마니 쪽을 향하고 있음.",
    "built_space": "야간의 야외 공간. 뒤쪽에 크게 열린 목재 문이 있고, 우측에는 레퍼런스와 일치하는 형태의 흙과 돌로 빚은 벽이 위치함.",
    "entities": "염재보는 레퍼런스와 일치하는 얼굴, 복장, 짙은 색의 원뿔형 모자를 착용함. 머슴은 흙 묻은 무명 옷을 입음. 곡식 자루는 짚으로 엮은 가마니 형태. 총은 염재보의 등에 메어 있음.",
    "hard_violations": [],
    "physics": "염재보는 양팔과 손으로 짚 가마니의 하단과 측면을 받쳐 들고 있으며, 머슴의 왼손은 가마니 앞쪽에 얹혀 있어 물리적 접촉과 무게 지지가 자연스러움. 소총은 어깨 끈에 의해 안정적으로 지탱됨."
   },
   {
    "label": "B",
    "direction": "염재보는 멍한 표정으로 정면을 응시함.",
    "built_space": "야간의 전통 한옥 마당. 불 켜진 등과 마루, 닫힌 창호문 등이 배치되어 있으며, 레퍼런스가 지시한 돌담 벽은 전혀 존재하지 않음.",
    "entities": "염재보의 얼굴과 복장은 맞으나 모자가 레퍼런스와 전혀 다른 둥근 챙의 밀짚모자 형태로 생성됨. 곡식 자루는 황마 자루 형태. 머슴은 우측에 팔과 몸통 일부만 보임.",
    "hard_violations": [
     "[gemini-pro] 장소 레퍼런스(돌담 벽 등 고정된 구조물)를 완전히 무시하고 임의의 전통 가옥 마당을 배경으로 생성함.",
     "[gemini-pro] 캐릭터 레퍼런스의 지정된 모자 형태를 무시하고 다른 종류의 모자를 착용시킴."
    ],
    "physics": "염재보가 황마 자루를 가슴에 밀착해 꽉 껴안고 있으며, 우측에서 뻗어 나온 머슴의 손이 자루 측면에 닿아 있음. 소총은 우측 어깨에 걸쳐 지지됨."
   }
  ],
  "cross_model_order": {
   "policy": "cross_model_order_v2_slot0_fwd_slot1_rev_parallel2",
   "slots": [
    {
     "model": "gemini-pro",
     "order": "forward",
     "display_to_canonical": {
      "A": "A",
      "B": "B"
     },
     "raw": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 9,
        "verdict_ko": "지정된 장소 레퍼런스(돌담)와 캐릭터 레퍼런스(얼굴 및 모자 형태)를 정확히 반영했으며, 멍한 표정으로 가마니를 받아든 프레이밍과 행동 지시를 훌륭하게 구현했습니다."
       },
       {
        "label": "B",
        "score": 3,
        "verdict_ko": "캐릭터의 모자가 레퍼런스와 다르게 변형되었으며, 지정된 돌담 배경 대신 임의의 한옥 마당을 배경으로 생성하여 장소 고정 지시를 크게 위반했습니다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "염재보는 정면을 향해 멍한 시선을 던지고 있으며, 머슴은 염재보와 쌀가마니 쪽을 향하고 있음.",
        "built_space": "야간의 야외 공간. 뒤쪽에 크게 열린 목재 문이 있고, 우측에는 레퍼런스와 일치하는 형태의 흙과 돌로 빚은 벽이 위치함.",
        "entities": "염재보는 레퍼런스와 일치하는 얼굴, 복장, 짙은 색의 원뿔형 모자를 착용함. 머슴은 흙 묻은 무명 옷을 입음. 곡식 자루는 짚으로 엮은 가마니 형태. 총은 염재보의 등에 메어 있음.",
        "hard_violations": [],
        "physics": "염재보는 양팔과 손으로 짚 가마니의 하단과 측면을 받쳐 들고 있으며, 머슴의 왼손은 가마니 앞쪽에 얹혀 있어 물리적 접촉과 무게 지지가 자연스러움. 소총은 어깨 끈에 의해 안정적으로 지탱됨."
       },
       {
        "label": "B",
        "direction": "염재보는 멍한 표정으로 정면을 응시함.",
        "built_space": "야간의 전통 한옥 마당. 불 켜진 등과 마루, 닫힌 창호문 등이 배치되어 있으며, 레퍼런스가 지시한 돌담 벽은 전혀 존재하지 않음.",
        "entities": "염재보의 얼굴과 복장은 맞으나 모자가 레퍼런스와 전혀 다른 둥근 챙의 밀짚모자 형태로 생성됨. 곡식 자루는 황마 자루 형태. 머슴은 우측에 팔과 몸통 일부만 보임.",
        "hard_violations": [
         "장소 레퍼런스(돌담 벽 등 고정된 구조물)를 완전히 무시하고 임의의 전통 가옥 마당을 배경으로 생성함.",
         "캐릭터 레퍼런스의 지정된 모자 형태를 무시하고 다른 종류의 모자를 착용시킴."
        ],
        "physics": "염재보가 황마 자루를 가슴에 밀착해 꽉 껴안고 있으며, 우측에서 뻗어 나온 머슴의 손이 자루 측면에 닿아 있음. 소총은 우측 어깨에 걸쳐 지지됨."
       }
      ],
      "all_candidates_fail": false
     },
     "normalized": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 9,
        "verdict_ko": "지정된 장소 레퍼런스(돌담)와 캐릭터 레퍼런스(얼굴 및 모자 형태)를 정확히 반영했으며, 멍한 표정으로 가마니를 받아든 프레이밍과 행동 지시를 훌륭하게 구현했습니다."
       },
       {
        "label": "B",
        "score": 3,
        "verdict_ko": "캐릭터의 모자가 레퍼런스와 다르게 변형되었으며, 지정된 돌담 배경 대신 임의의 한옥 마당을 배경으로 생성하여 장소 고정 지시를 크게 위반했습니다."
       }
      ],
      "all_candidates_fail": false,
      "readings": [
       {
        "label": "A",
        "direction": "염재보는 정면을 향해 멍한 시선을 던지고 있으며, 머슴은 염재보와 쌀가마니 쪽을 향하고 있음.",
        "built_space": "야간의 야외 공간. 뒤쪽에 크게 열린 목재 문이 있고, 우측에는 레퍼런스와 일치하는 형태의 흙과 돌로 빚은 벽이 위치함.",
        "entities": "염재보는 레퍼런스와 일치하는 얼굴, 복장, 짙은 색의 원뿔형 모자를 착용함. 머슴은 흙 묻은 무명 옷을 입음. 곡식 자루는 짚으로 엮은 가마니 형태. 총은 염재보의 등에 메어 있음.",
        "hard_violations": [],
        "physics": "염재보는 양팔과 손으로 짚 가마니의 하단과 측면을 받쳐 들고 있으며, 머슴의 왼손은 가마니 앞쪽에 얹혀 있어 물리적 접촉과 무게 지지가 자연스러움. 소총은 어깨 끈에 의해 안정적으로 지탱됨."
       },
       {
        "label": "B",
        "direction": "염재보는 멍한 표정으로 정면을 응시함.",
        "built_space": "야간의 전통 한옥 마당. 불 켜진 등과 마루, 닫힌 창호문 등이 배치되어 있으며, 레퍼런스가 지시한 돌담 벽은 전혀 존재하지 않음.",
        "entities": "염재보의 얼굴과 복장은 맞으나 모자가 레퍼런스와 전혀 다른 둥근 챙의 밀짚모자 형태로 생성됨. 곡식 자루는 황마 자루 형태. 머슴은 우측에 팔과 몸통 일부만 보임.",
        "hard_violations": [
         "장소 레퍼런스(돌담 벽 등 고정된 구조물)를 완전히 무시하고 임의의 전통 가옥 마당을 배경으로 생성함.",
         "캐릭터 레퍼런스의 지정된 모자 형태를 무시하고 다른 종류의 모자를 착용시킴."
        ],
        "physics": "염재보가 황마 자루를 가슴에 밀착해 꽉 껴안고 있으며, 우측에서 뻗어 나온 머슴의 손이 자루 측면에 닿아 있음. 소총은 우측 어깨에 걸쳐 지지됨."
       }
      ]
     },
     "ok": true
    },
    {
     "model": "gpt-high",
     "order": "reverse",
     "display_to_canonical": {
      "A": "B",
      "B": "A"
     },
     "raw": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 6,
        "verdict_ko": "멍한 정면과 야간 중간숏은 맞지만, 자루가 가슴까지 세로로 올라와 하복부에 가로로 받치는 지시와 다르고 머슴의 손도 자루보다 염재보의 팔에 닿아 있다."
       },
       {
        "label": "B",
        "score": 9,
        "verdict_ko": "정면의 멍한 표정, 하복부에 가로로 받은 가마니, 두 팔의 하중 지지와 머슴의 인계 동작을 가장 충실히 구현하며 인물·의상·창고 재질도 참조에 가깝다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "염재보는 얼굴과 시선을 카메라 정면으로 향하고 입을 조금 벌려 당황한 기색을 보인다. 오른쪽에서 들어온 머슴의 손은 자루를 직접 잡기보다 염재보의 화면 오른쪽 위팔에 닿아 보인다. 자루는 하복부를 가로지르지 않고 가슴부터 골반까지 세로로 놓였다. 소총 총구는 어깨 뒤에서 위를 향하며 누구도 겨냥하지 않는다.",
        "built_space": "염재보는 창고 문턱 밖 흙마당에 서 있고, 오른쪽 뒤에 열린 출입구 하나와 양옆의 나무 문짝 두 장이 보인다. 문짝은 비스듬히 열려 통로를 막지 않는다. 왼쪽에는 기둥, 격자 창호, 돌기단과 등불 하나가 보인다. 낡은 목재와 돌·흙 재질은 참조의 계열을 유지하지만 참조보다 주변 건물을 많이 드러낸다. 자물쇠의 해제 상태와 대문은 이 구도에서 확인되지 않는다.",
        "entities": "염재보는 참조와 대체로 닮은 성인 한국인 남성으로 보이며 검은 머리, 짙은 누비 저고리와 바지를 갖췄다. 다만 모자는 참조의 뾰족한 원뿔형보다 둥글고 높다. 머슴은 오른쪽 몸통 일부와 성인 남성의 손만 보이고, 밝은 무명 소매에는 흙 얼룩이 있다. 안고 있는 것은 끈으로 입구를 묶은 거친 직물 자루로 소품 참조의 자루 유형과 맞는다. 소총 한 정과 어깨끈이 보인다. 읽을 수 있는 글자나 추가 인물은 없다.",
        "hard_violations": [],
        "physics": "자루는 염재보의 몸통에 밀착되어 두 팔과 손의 압박으로 지지된다. 밑면을 두 전완으로 떠받치는 자세는 아니지만, 팔 사이에 끼워 안는 동작 자체는 가능하며 부유하지 않는다. 소총은 어깨끈으로 지지된다. 인물의 발은 프레임 밖이지만 상체에 공중 동작이나 불가능한 균형은 보이지 않는다."
       },
       {
        "label": "B",
        "direction": "염재보의 얼굴과 시선은 정면을 향하고, 조금 벌어진 입과 멈춘 눈빛이 얼떨결에 받은 순간으로 읽힌다. 왼쪽 머슴은 가마니와 염재보의 팔 쪽을 내려다보며 한 손을 가마니에, 다른 손을 염재보의 위팔에 댄다. 가마니의 긴 축은 하복부를 비스듬히 가로지른다. 등에 멘 소총 총구는 화면 오른쪽 위를 향하고 사람을 겨냥하지 않는다.",
        "built_space": "출입구 하나를 중심으로 문짝 두 장이 좌우로 활짝 열려 비스듬히 보이며, 출입구 자체를 가로막지 않는다. 염재보는 문턱 바로 바깥 마당에, 머슴은 그의 왼쪽에 서 있다. 참조와 같은 계열의 풍화된 판재 문, 목재 문틀, 돌이 섞인 흙벽이 이어진다. 왼쪽에 등불 하나, 오른쪽 벽 아래에 가마니 하나와 기대 놓은 농기구들이 보인다. 대문과 자물쇠의 세부 상태는 확인할 수 없고 불가능한 반사나 중복 출입구는 없다.",
        "entities": "염재보는 성인 한국인 남성으로 보이며 얼굴형, 검은 머리, 원뿔형 삿갓과 턱끈, 짙은 누비옷이 인물 참조에 가깝다. 머슴도 검은 머리의 성인 한국인 남성으로 보이고 수수한 무명 바지저고리에 뚜렷한 흙 얼룩이 있다. 중심 소품은 새끼줄로 묶은 짚 가마니로 참조의 재질과 형태에 부합한다. 소총 한 정을 끈으로 멘 상태가 보이며, 추가 인물이나 읽을 수 있는 글자는 없다.",
        "hard_violations": [],
        "physics": "가마니의 양쪽 아래를 염재보의 손과 굽힌 팔이 받치고 몸통이 뒤에서 지지한다. 머슴의 손도 가마니 왼쪽 면에 닿아 있어 건네는 동작이 연결된다. 세게 끌어안는 압박감은 다소 약하지만 무게를 받아 든 자세는 물리적으로 가능하다. 소총은 몸을 가로지르는 멜빵에 매달려 있다. 발은 잘렸으나 두 인물 모두 서 있는 자세로 읽히며 지지 없이 떠 있는 물체는 없다."
       }
      ],
      "all_candidates_fail": false
     },
     "normalized": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 6,
        "verdict_ko": "멍한 정면과 야간 중간숏은 맞지만, 자루가 가슴까지 세로로 올라와 하복부에 가로로 받치는 지시와 다르고 머슴의 손도 자루보다 염재보의 팔에 닿아 있다."
       },
       {
        "label": "A",
        "score": 9,
        "verdict_ko": "정면의 멍한 표정, 하복부에 가로로 받은 가마니, 두 팔의 하중 지지와 머슴의 인계 동작을 가장 충실히 구현하며 인물·의상·창고 재질도 참조에 가깝다."
       }
      ],
      "all_candidates_fail": false,
      "readings": [
       {
        "label": "B",
        "direction": "염재보는 얼굴과 시선을 카메라 정면으로 향하고 입을 조금 벌려 당황한 기색을 보인다. 오른쪽에서 들어온 머슴의 손은 자루를 직접 잡기보다 염재보의 화면 오른쪽 위팔에 닿아 보인다. 자루는 하복부를 가로지르지 않고 가슴부터 골반까지 세로로 놓였다. 소총 총구는 어깨 뒤에서 위를 향하며 누구도 겨냥하지 않는다.",
        "built_space": "염재보는 창고 문턱 밖 흙마당에 서 있고, 오른쪽 뒤에 열린 출입구 하나와 양옆의 나무 문짝 두 장이 보인다. 문짝은 비스듬히 열려 통로를 막지 않는다. 왼쪽에는 기둥, 격자 창호, 돌기단과 등불 하나가 보인다. 낡은 목재와 돌·흙 재질은 참조의 계열을 유지하지만 참조보다 주변 건물을 많이 드러낸다. 자물쇠의 해제 상태와 대문은 이 구도에서 확인되지 않는다.",
        "entities": "염재보는 참조와 대체로 닮은 성인 한국인 남성으로 보이며 검은 머리, 짙은 누비 저고리와 바지를 갖췄다. 다만 모자는 참조의 뾰족한 원뿔형보다 둥글고 높다. 머슴은 오른쪽 몸통 일부와 성인 남성의 손만 보이고, 밝은 무명 소매에는 흙 얼룩이 있다. 안고 있는 것은 끈으로 입구를 묶은 거친 직물 자루로 소품 참조의 자루 유형과 맞는다. 소총 한 정과 어깨끈이 보인다. 읽을 수 있는 글자나 추가 인물은 없다.",
        "hard_violations": [],
        "physics": "자루는 염재보의 몸통에 밀착되어 두 팔과 손의 압박으로 지지된다. 밑면을 두 전완으로 떠받치는 자세는 아니지만, 팔 사이에 끼워 안는 동작 자체는 가능하며 부유하지 않는다. 소총은 어깨끈으로 지지된다. 인물의 발은 프레임 밖이지만 상체에 공중 동작이나 불가능한 균형은 보이지 않는다."
       },
       {
        "label": "A",
        "direction": "염재보의 얼굴과 시선은 정면을 향하고, 조금 벌어진 입과 멈춘 눈빛이 얼떨결에 받은 순간으로 읽힌다. 왼쪽 머슴은 가마니와 염재보의 팔 쪽을 내려다보며 한 손을 가마니에, 다른 손을 염재보의 위팔에 댄다. 가마니의 긴 축은 하복부를 비스듬히 가로지른다. 등에 멘 소총 총구는 화면 오른쪽 위를 향하고 사람을 겨냥하지 않는다.",
        "built_space": "출입구 하나를 중심으로 문짝 두 장이 좌우로 활짝 열려 비스듬히 보이며, 출입구 자체를 가로막지 않는다. 염재보는 문턱 바로 바깥 마당에, 머슴은 그의 왼쪽에 서 있다. 참조와 같은 계열의 풍화된 판재 문, 목재 문틀, 돌이 섞인 흙벽이 이어진다. 왼쪽에 등불 하나, 오른쪽 벽 아래에 가마니 하나와 기대 놓은 농기구들이 보인다. 대문과 자물쇠의 세부 상태는 확인할 수 없고 불가능한 반사나 중복 출입구는 없다.",
        "entities": "염재보는 성인 한국인 남성으로 보이며 얼굴형, 검은 머리, 원뿔형 삿갓과 턱끈, 짙은 누비옷이 인물 참조에 가깝다. 머슴도 검은 머리의 성인 한국인 남성으로 보이고 수수한 무명 바지저고리에 뚜렷한 흙 얼룩이 있다. 중심 소품은 새끼줄로 묶은 짚 가마니로 참조의 재질과 형태에 부합한다. 소총 한 정을 끈으로 멘 상태가 보이며, 추가 인물이나 읽을 수 있는 글자는 없다.",
        "hard_violations": [],
        "physics": "가마니의 양쪽 아래를 염재보의 손과 굽힌 팔이 받치고 몸통이 뒤에서 지지한다. 머슴의 손도 가마니 왼쪽 면에 닿아 있어 건네는 동작이 연결된다. 세게 끌어안는 압박감은 다소 약하지만 무게를 받아 든 자세는 물리적으로 가능하다. 소총은 몸을 가로지르는 멜빵에 매달려 있다. 발은 잘렸으나 두 인물 모두 서 있는 자세로 읽히며 지지 없이 떠 있는 물체는 없다."
       }
      ]
     },
     "ok": true
    }
   ],
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "slot_winner_match": true,
   "slot_winner": {
    "gemini-pro": "A",
    "gpt-high": "A"
   },
   "route": "cross_slot_agree"
  },
  "dual": {
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "normalized": {
    "A": 2.0,
    "B": 1.0
   },
   "adjusted": {
    "A": 2.0,
    "B": 0.75
   },
   "violations": {
    "B": [
     "[gemini-pro] 장소 레퍼런스(돌담 벽 등 고정된 구조물)를 완전히 무시하고 임의의 전통 가옥 마당을 배경으로 생성함.",
     "[gemini-pro] 캐릭터 레퍼런스의 지정된 모자 형태를 무시하고 다른 종류의 모자를 착용시킴."
    ]
   },
   "per_model_winner": {
    "gemini-pro": "A",
    "gpt-high": "A"
   },
   "agreed": true
  },
  "totals": {
   "A": 2000,
   "B": 750
  },
  "selected": "A",
  "ranking": [
   "A",
   "B"
  ],
  "verdicts": [
   {
    "label": "A",
    "score": 2000,
    "verdict_ko": "지정된 장소 레퍼런스(돌담)와 캐릭터 레퍼런스(얼굴 및 모자 형태)를 정확히 반영했으며, 멍한 표정으로 가마니를 받아든 프레이밍과 행동 지시를 훌륭하게 구현했습니다."
   },
   {
    "label": "B",
    "score": 750,
    "verdict_ko": "캐릭터의 모자가 레퍼런스와 다르게 변형되었으며, 지정된 돌담 배경 대신 임의의 한옥 마당을 배경으로 생성하여 장소 고정 지시를 크게 위반했습니다.  ★위반: [gemini-pro] 장소 레퍼런스(돌담 벽 등 고정된 구조물)를 완전히 무시하고 임의의 전통 가옥 마당을 배경으로 생성함. / [gemini-pro] 캐릭터 레퍼런스의 지정된 모자 형태를 무시하고 다른 종류의 모자를 착용시킴."
   }
  ],
  "refs": [
   {
    "label": "PREVIOUS SHOT STILL — a visually related earlier shot of this same place: the location's look, materials, fixed features and lighting mood are LOCKED to this photo; never copy its camera framing. The people visible in that photo are NOT in this shot: never carry their faces, bodies or clothing onto anyone here. Each person in THIS shot is defined solely by the PEOPLE list, their own CHARACTER REFERENCE images and their wardrobe notes. If this photo shows a character who cannot move (dead or unconscious), that character's exact pose, position and orientation are ALSO LOCKED — treat their body as a fixed prop of the set that only the camera moves around.",
    "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/db00517b-165b-4005-a287-f28936b59b95/images/272bd7af-9294-4490-98c6-88a40e9deacc/scene/recipe/S33sh11_sel.png",
    "asset_id": "4471a836-fa89-4828-86b4-f9e2c0bd5293",
    "role": "prev_still"
   },
   {
    "label": "CHARACTER REFERENCE — 염재보: the exact person appearing in this shot; match face, hair and build exactly.",
    "path": "<bytes:964351>",
    "asset_id": "ab287289-d0b7-436d-8f70-68452f898f73",
    "role": "character_ref"
   },
   {
    "label": "PROP REFERENCE — 곡식 자루·가마니: the exact object appearing in this shot; match its look, material and wear exactly.",
    "path": "<bytes:1807398>",
    "asset_id": "e30c319e-fdc7-4c01-a301-290ddd45d27a",
    "role": "prop_ref"
   }
  ],
  "critique_skipped": true,
  "shot_run_produced": false,
  "shot_run_uid": "06aa1428-45b1-7a19-a6ef-6cf109f9f220",
  "ref_mode": "prev+엔티티",
  "share_plan": {
   "ref_plan": "prev",
   "prev_anchor_tag": "S33sh11"
  }
 },
 "S34sh8::cine": {
  "applied": true,
  "attempted_at": "2026-09-09T04:37:46.958993+00:00",
  "fingerprint": "026cc04f9ccecba3f78649b77b0fa3bee2de570c3c88a2fa15ef82d6d56e3292",
  "fingerprint_version": 2,
  "provider": "mai",
  "endpoint": "openrouter/chat-completions",
  "model": "microsoft/mai-image-2.6",
  "pack": "24.202608252115",
  "source_file": "S34sh8_sel.png",
  "source_sha256": "c5a2a43d29fd0e2ed2de5cbea4a4c81d0f8bddcdfc94b9e6952a79004f08cc82",
  "file": "S34sh8_cine.png",
  "staged_sha256": "b3fca89f593ed194c8145bd57caeb3b9f9a031bf6f219e0901caab3ca0787c99",
  "latency_ms": 27298
 },
 "S34sh16::signage": {
  "fp": "bea5d03b82335623",
  "inscriptions": [],
  "cues": [],
  "dropped": []
 },
 "S34sh16": {
  "input_fingerprint": "6ba316d59dcd49d9",
  "prompt": "Create ONE FINAL photorealistic live-action film still of the moment below — early 20th-century Korea, Jeolla region; all people are Korean unless explicitly stated otherwise. TIME OF DAY (lock): night.\n\nSHOT TEXT (authoritative, Korean): 안규홍이 쥔 환도 손잡이가 박제현의 뒷목에 강하게 맞닿아 충격을 가하는 타격 순간의 찰나.\n\nLOCATION (lock): The outdoor loading space before the estate's granary, beside the rice cart in the nighttime courtyard. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: close-up\n- KEY BACKGROUND ELEMENTS: Sword hilt (In 안규홍's hand, making contact with 박제현's nape) — Seen from the side so the gripping hand and contact end remain distinguishable; used as Small, sharply readable action detail between the two bodies; Courtyard ground (Beneath the two men); used as A limited background area establishes the direction of the impending fall.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Preserve the established muted nighttime exposure and naturalistic contrast, with no stylized flash or added impact illumination.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The people visible in that photo are NOT in this shot: never carry their faces, bodies or clothing onto anyone here. Each person in THIS shot is defined solely by the PEOPLE list, their own CHARACTER REFERENCE images and their wardrobe notes. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The store door and main gate remain open; rice bales are loaded on the cart. 안규홍: He holds the sword by its hilt for the blow, with the jade ring concealed in his clothing. 박제현: He has exposed the back of his neck and is losing consciousness from the blow.\n\nTHE HAND THAT IS DOING THIS: the object at the centre of this shot is being held, operated, read, aimed or handed over by 안규홍 right now, so 안규홍's hand — and as much of the wrist and forearm as the framing reaches — is in the frame, gripping or resting on that object exactly the way its real use requires. Match that hand to 안규홍: its size, build, skin, age, grooming, sleeve and anything worn on it belong to that person and to no one else.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 안규홍 (한국인 남성, 청년 나이대, 검은 머리, 단단히 묶은 머리); 박제현 (한국인 남성, 중년 이상 나이대, 자연스러운 연령대의 얼굴) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.",
  "roll_prompts": {
   "A": "Create ONE FINAL photorealistic live-action film still of the moment below — early 20th-century Korea, Jeolla region; all people are Korean unless explicitly stated otherwise. TIME OF DAY (lock): night.\n\nSHOT TEXT (authoritative, Korean): 안규홍이 쥔 환도 손잡이가 박제현의 뒷목에 강하게 맞닿아 충격을 가하는 타격 순간의 찰나.\n\nLOCATION (lock): The outdoor loading space before the estate's granary, beside the rice cart in the nighttime courtyard. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: close-up\n- KEY BACKGROUND ELEMENTS: Sword hilt (In 안규홍's hand, making contact with 박제현's nape) — Seen from the side so the gripping hand and contact end remain distinguishable; used as Small, sharply readable action detail between the two bodies; Courtyard ground (Beneath the two men); used as A limited background area establishes the direction of the impending fall.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Preserve the established muted nighttime exposure and naturalistic contrast, with no stylized flash or added impact illumination.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The people visible in that photo are NOT in this shot: never carry their faces, bodies or clothing onto anyone here. Each person in THIS shot is defined solely by the PEOPLE list, their own CHARACTER REFERENCE images and their wardrobe notes. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The store door and main gate remain open; rice bales are loaded on the cart. 안규홍: He holds the sword by its hilt for the blow, with the jade ring concealed in his clothing. 박제현: He has exposed the back of his neck and is losing consciousness from the blow.\n\nTHE HAND THAT IS DOING THIS: the object at the centre of this shot is being held, operated, read, aimed or handed over by 안규홍 right now, so 안규홍's hand — and as much of the wrist and forearm as the framing reaches — is in the frame, gripping or resting on that object exactly the way its real use requires. Match that hand to 안규홍: its size, build, skin, age, grooming, sleeve and anything worn on it belong to that person and to no one else.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 안규홍 (한국인 남성, 청년 나이대, 검은 머리, 단단히 묶은 머리); 박제현 (한국인 남성, 중년 이상 나이대, 자연스러운 연령대의 얼굴) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.",
   "B": "Create ONE FINAL photorealistic live-action film still of the moment below — early 20th-century Korea, Jeolla region; all people are Korean unless explicitly stated otherwise. TIME OF DAY (lock): night.\n\nSHOT TEXT (authoritative, Korean): 안규홍이 쥔 환도 손잡이가 박제현의 뒷목에 강하게 맞닿아 충격을 가하는 타격 순간의 찰나.\n\nLOCATION (lock): The outdoor loading space before the estate's granary, beside the rice cart in the nighttime courtyard. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: close-up\n- KEY BACKGROUND ELEMENTS: Sword hilt (In 안규홍's hand, making contact with 박제현's nape) — Seen from the side so the gripping hand and contact end remain distinguishable; used as Small, sharply readable action detail between the two bodies; Courtyard ground (Beneath the two men); used as A limited background area establishes the direction of the impending fall.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Preserve the established muted nighttime exposure and naturalistic contrast, with no stylized flash or added impact illumination.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The people visible in that photo are NOT in this shot: never carry their faces, bodies or clothing onto anyone here. Each person in THIS shot is defined solely by the PEOPLE list, their own CHARACTER REFERENCE images and their wardrobe notes. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The store door and main gate remain open; rice bales are loaded on the cart. 안규홍: He holds the sword by its hilt for the blow, with the jade ring concealed in his clothing. 박제현: He has exposed the back of his neck and is losing consciousness from the blow.\n\nTHE HAND THAT IS DOING THIS: the object at the centre of this shot is being held, operated, read, aimed or handed over by 안규홍 right now, so 안규홍's hand — and as much of the wrist and forearm as the framing reaches — is in the frame, gripping or resting on that object exactly the way its real use requires. Match that hand to 안규홍: its size, build, skin, age, grooming, sleeve and anything worn on it belong to that person and to no one else.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 안규홍 (한국인 남성, 청년 나이대, 검은 머리, 단단히 묶은 머리); 박제현 (한국인 남성, 중년 이상 나이대, 자연스러운 연령대의 얼굴) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.\n\nCOMPOSITION VARIATION (this candidate only): keep the place, the people and the moment exactly as specified, and keep from the CAMERA & FRAME contract its subject scale, screen placement and key background elements exactly. For this candidate only, its camera angle and height are deliberately left open: choose a camera position distinctly different from the obvious one, as a film director picking a second setup on the same blocking. Stay true to any explicit framing the SHOT TEXT itself dictates."
  },
  "shot_run_spend_attempt_count": 1,
  "initial_roll_all_fail": true,
  "readings": [
   {
    "label": "A",
    "direction": "안규홍의 시선은 상대의 머리와 목을 향하고 있으나, 타격하는 손의 방향이 뒷목이 아닌 상대의 뺨과 옆목에 맞닿아 있음.",
    "built_space": "야간의 흙바닥 마당이며, 배경에 프롬프트가 요구한 수레가 배치되어 있음.",
    "entities": "안규홍은 캐릭터 레퍼런스의 밝은 의상을 입고 있으나, 가격당하는 박제현은 얼굴이 가려져 신원을 확인하기 어려움. 손에 쥔 환도는 칼날이 전혀 없는 손잡이 부분만 존재함.",
    "hard_violations": [
     "[gemini-pro] 칼날이 없는 물리적으로 불가능한 형태의 검 (발명된 사물)",
     "[gemini-pro] 프롬프트에 명시된 '뒷목 타격'이 아닌 얼굴 측면/뺨을 찌르듯 누르고 있는 잘못된 액션 연출"
    ],
    "physics": "안규홍이 환도 손잡이를 쥐고 있으나, 코등이(가드) 너머로 이어져야 할 칼날이 아예 존재하지 않아 허공에서 잘려나간 형태임."
   },
   {
    "label": "B",
    "direction": "안규홍이 박제현을 바라보고 있으며, 박제현은 눈이 뒤집히며 의식을 잃어가는 시선 처리를 보여줌.",
    "built_space": "야간의 흙바닥과 배경 구조물이 보임.",
    "entities": "박제현의 얼굴과 표정은 묘사되었으나, 안규홍은 캐릭터 레퍼런스의 의상이 아닌 이전 샷(Previous Shot Still)의 어두운 누비옷을 그대로 입고 있어 캐릭터 지침을 위반함.",
    "hard_violations": [
     "[gemini-pro] 환도의 칼날이 박제현의 목을 관통하여 반대편으로 튀어나온 물리적으로 불가능한 인체/사물 묘사",
     "[gemini-pro] 안규홍에게 이전 샷의 의상(어두운 누비옷)을 입혀, 레퍼런스 의상만 사용하라는 명시적 금지 지침 위반"
    ],
    "physics": "안규홍이 손잡이를 쥐고 있으나, 칼날이 박제현의 뒷목 겉을 타격하는 것이 아니라 목 안쪽을 완전히 관통해 물리적으로 불가능한 상태를 연출함."
   }
  ],
  "cross_model_order": {
   "policy": "cross_model_order_v2_slot0_fwd_slot1_rev_parallel2",
   "slots": [
    {
     "model": "gemini-pro",
     "order": "forward",
     "display_to_canonical": {
      "A": "A",
      "B": "B"
     },
     "raw": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 2,
        "verdict_ko": "검의 칼날이 아예 생략된 채 손잡이만 들고 있으며, 뒷목이 아닌 얼굴 측면을 타격하여 핵심 액션과 사물 묘사에 실패했습니다."
       },
       {
        "label": "B",
        "score": 2,
        "verdict_ko": "칼날이 인체의 목을 관통하는 물리적 오류가 발생했으며, 안규홍이 이전 샷의 의상을 그대로 입고 등장해 지침을 심각하게 위반했습니다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "안규홍의 시선은 상대의 머리와 목을 향하고 있으나, 타격하는 손의 방향이 뒷목이 아닌 상대의 뺨과 옆목에 맞닿아 있음.",
        "built_space": "야간의 흙바닥 마당이며, 배경에 프롬프트가 요구한 수레가 배치되어 있음.",
        "entities": "안규홍은 캐릭터 레퍼런스의 밝은 의상을 입고 있으나, 가격당하는 박제현은 얼굴이 가려져 신원을 확인하기 어려움. 손에 쥔 환도는 칼날이 전혀 없는 손잡이 부분만 존재함.",
        "hard_violations": [
         "칼날이 없는 물리적으로 불가능한 형태의 검 (발명된 사물)",
         "프롬프트에 명시된 '뒷목 타격'이 아닌 얼굴 측면/뺨을 찌르듯 누르고 있는 잘못된 액션 연출"
        ],
        "physics": "안규홍이 환도 손잡이를 쥐고 있으나, 코등이(가드) 너머로 이어져야 할 칼날이 아예 존재하지 않아 허공에서 잘려나간 형태임."
       },
       {
        "label": "B",
        "direction": "안규홍이 박제현을 바라보고 있으며, 박제현은 눈이 뒤집히며 의식을 잃어가는 시선 처리를 보여줌.",
        "built_space": "야간의 흙바닥과 배경 구조물이 보임.",
        "entities": "박제현의 얼굴과 표정은 묘사되었으나, 안규홍은 캐릭터 레퍼런스의 의상이 아닌 이전 샷(Previous Shot Still)의 어두운 누비옷을 그대로 입고 있어 캐릭터 지침을 위반함.",
        "hard_violations": [
         "환도의 칼날이 박제현의 목을 관통하여 반대편으로 튀어나온 물리적으로 불가능한 인체/사물 묘사",
         "안규홍에게 이전 샷의 의상(어두운 누비옷)을 입혀, 레퍼런스 의상만 사용하라는 명시적 금지 지침 위반"
        ],
        "physics": "안규홍이 손잡이를 쥐고 있으나, 칼날이 박제현의 뒷목 겉을 타격하는 것이 아니라 목 안쪽을 완전히 관통해 물리적으로 불가능한 상태를 연출함."
       }
      ],
      "all_candidates_fail": true
     },
     "normalized": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 2,
        "verdict_ko": "검의 칼날이 아예 생략된 채 손잡이만 들고 있으며, 뒷목이 아닌 얼굴 측면을 타격하여 핵심 액션과 사물 묘사에 실패했습니다."
       },
       {
        "label": "B",
        "score": 2,
        "verdict_ko": "칼날이 인체의 목을 관통하는 물리적 오류가 발생했으며, 안규홍이 이전 샷의 의상을 그대로 입고 등장해 지침을 심각하게 위반했습니다."
       }
      ],
      "all_candidates_fail": true,
      "readings": [
       {
        "label": "A",
        "direction": "안규홍의 시선은 상대의 머리와 목을 향하고 있으나, 타격하는 손의 방향이 뒷목이 아닌 상대의 뺨과 옆목에 맞닿아 있음.",
        "built_space": "야간의 흙바닥 마당이며, 배경에 프롬프트가 요구한 수레가 배치되어 있음.",
        "entities": "안규홍은 캐릭터 레퍼런스의 밝은 의상을 입고 있으나, 가격당하는 박제현은 얼굴이 가려져 신원을 확인하기 어려움. 손에 쥔 환도는 칼날이 전혀 없는 손잡이 부분만 존재함.",
        "hard_violations": [
         "칼날이 없는 물리적으로 불가능한 형태의 검 (발명된 사물)",
         "프롬프트에 명시된 '뒷목 타격'이 아닌 얼굴 측면/뺨을 찌르듯 누르고 있는 잘못된 액션 연출"
        ],
        "physics": "안규홍이 환도 손잡이를 쥐고 있으나, 코등이(가드) 너머로 이어져야 할 칼날이 아예 존재하지 않아 허공에서 잘려나간 형태임."
       },
       {
        "label": "B",
        "direction": "안규홍이 박제현을 바라보고 있으며, 박제현은 눈이 뒤집히며 의식을 잃어가는 시선 처리를 보여줌.",
        "built_space": "야간의 흙바닥과 배경 구조물이 보임.",
        "entities": "박제현의 얼굴과 표정은 묘사되었으나, 안규홍은 캐릭터 레퍼런스의 의상이 아닌 이전 샷(Previous Shot Still)의 어두운 누비옷을 그대로 입고 있어 캐릭터 지침을 위반함.",
        "hard_violations": [
         "환도의 칼날이 박제현의 목을 관통하여 반대편으로 튀어나온 물리적으로 불가능한 인체/사물 묘사",
         "안규홍에게 이전 샷의 의상(어두운 누비옷)을 입혀, 레퍼런스 의상만 사용하라는 명시적 금지 지침 위반"
        ],
        "physics": "안규홍이 손잡이를 쥐고 있으나, 칼날이 박제현의 뒷목 겉을 타격하는 것이 아니라 목 안쪽을 완전히 관통해 물리적으로 불가능한 상태를 연출함."
       }
      ]
     },
     "ok": true
    },
    {
     "model": "gpt-high",
     "order": "reverse",
     "display_to_canonical": {
      "A": "B",
      "B": "A"
     },
     "raw": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 2,
        "verdict_ko": "클로즈업은 맞지만 손잡이가 아니라 칼날이 목 뒤를 지나며, 두 사람의 의상과 박제현의 외모가 지정 인물보다 이전 장면 인물을 따른다."
       },
       {
        "label": "B",
        "score": 5,
        "verdict_ko": "안규홍의 의상과 움켜쥔 손, 앞으로 꺾이는 타격 반응은 더 충실하지만 접촉부가 뒷목을 치는 손잡이 끝으로 명확하지 않고 박제현의 신원·복장이 틀린다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "왼쪽 남자는 오른쪽 남자의 목 부근을 바라본다. 칼은 거의 수평으로 오른쪽을 향하며 손잡이 끝은 표적 반대쪽인 왼쪽에 있다. 목 뒤에 이르는 부분은 칼날이고, 손잡이가 뒷목에 맞닿는 타격은 보이지 않는다. 오른쪽 남자는 오른쪽 아래를 곁눈질하며 몸을 앞으로 기울인다.",
        "built_space": "상반신 클로즈업 뒤로 흙바닥과 상단의 창고 문턱 하나, 오른쪽 문짝 일부와 벽 일부가 보인다. 이전 장소의 흙바닥·목재·돌벽 계열은 이어지지만 수레와 대문은 화면 밖이라 상태를 확인할 수 없다. 중복 설비나 반사는 보이지 않는다.",
        "entities": "두 사람만 보인다. 왼쪽은 검은 머리를 뒤로 묶은 젊은 동아시아계 남성이지만, 지정된 안규홍의 낡은 밝은 저고리 대신 이전 장면의 짙은 누비옷을 입었다. 오른쪽은 중년 동아시아계 남성이지만 흐트러진 머리, 수염과 거친 흰옷이 박제현 참고의 얼굴·갓·연한 황색 포와 다르고 이전 장면의 왼쪽 인물을 닮았다. 칼의 마름모 감개와 장식 금속은 참고와 유사하다. 옥가락지는 드러나지 않으며 읽을 수 있는 글자는 없다.",
        "hard_violations": [],
        "physics": "칼은 왼쪽 남자의 손가락과 손바닥에 잡혀 있고 팔은 굽힌 팔꿈치로 이어진다. 오른쪽 남자의 상체는 화면 아래로 이어져 공중에 뜬 모습은 아니다. 발은 잘려 있어 접지 상태는 확인할 수 없다. 자세 자체는 가능하지만 손잡이와 뒷목의 충돌 접점이 없어 요구된 충격의 전달은 성립해 보이지 않는다."
       },
       {
        "label": "B",
        "direction": "안규홍은 상대의 목과 접촉 부위를 내려다보며 팔을 오른쪽으로 뻗는다. 박제현 역할의 남자는 고개를 아래로 숙인다. 손잡이 끝은 왼쪽 아래를 향하고, 목에 닿은 부분은 귀 아래 옆목과 뒷목 경계에 놓인 둥근 코등이 부근이다. A보다 타격 접촉은 읽히지만 노출된 뒷목을 손잡이 끝으로 강타하는 방향은 명확하지 않다.",
        "built_space": "두 상체 사이와 뒤에 제한된 흙바닥이 보이고, 왼쪽 뒤에는 열린 창고 문짝 하나와 어두운 출입구가 있다. 오른쪽 뒤에는 수레 한 대, 보이는 바퀴 하나와 적재된 짚 포장 쌀가마니들이 있다. 가마니는 흐려 정확한 개수는 판독하기 어렵다. 수레가 인물보다 뒤에 작게 놓여 장소와 규모 관계가 자연스럽고, 중복 설비나 반사는 없다. 대문은 화면 밖이다.",
        "entities": "두 사람만 보인다. 왼쪽의 젊은 동아시아계 남성은 묶은 검은 머리와 낡은 밝은 저고리로 안규홍 참고에 A보다 가깝고, 드러난 손과 팔도 그 인물에 연결된다. 오른쪽 중년 남성은 얼굴 대부분이 가려졌지만 풀어진 짧은 머리와 거친 흰옷으로, 박제현 참고의 갓·정돈된 머리·연한 황색 포와 다르다. 칼 손잡이의 감개와 장식 끝은 참고와 유사하나 칼날은 보이지 않아 전체 형태를 확인할 수 없다. 옥가락지는 노출되지 않았고 읽을 수 있는 글자는 없다.",
        "hard_violations": [],
        "physics": "칼 손잡이는 안규홍의 손에 단단히 잡혀 있고 손목과 전완이 타격 부위까지 이어진다. 상대는 목과 상체를 앞으로 굽혀 충격에 무너지는 반응을 보인다. 두 몸은 화면 아래로 이어지며 공중 부양의 증거는 없고, 발은 클로즈업 밖이다. 접촉과 체중 이동은 물리적으로 가능하지만, 주먹과 코등이가 접점을 가려 손잡이 타격인지 목을 밀어 누르는 동작인지 모호하다."
       }
      ],
      "all_candidates_fail": true
     },
     "normalized": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 2,
        "verdict_ko": "클로즈업은 맞지만 손잡이가 아니라 칼날이 목 뒤를 지나며, 두 사람의 의상과 박제현의 외모가 지정 인물보다 이전 장면 인물을 따른다."
       },
       {
        "label": "A",
        "score": 5,
        "verdict_ko": "안규홍의 의상과 움켜쥔 손, 앞으로 꺾이는 타격 반응은 더 충실하지만 접촉부가 뒷목을 치는 손잡이 끝으로 명확하지 않고 박제현의 신원·복장이 틀린다."
       }
      ],
      "all_candidates_fail": true,
      "readings": [
       {
        "label": "B",
        "direction": "왼쪽 남자는 오른쪽 남자의 목 부근을 바라본다. 칼은 거의 수평으로 오른쪽을 향하며 손잡이 끝은 표적 반대쪽인 왼쪽에 있다. 목 뒤에 이르는 부분은 칼날이고, 손잡이가 뒷목에 맞닿는 타격은 보이지 않는다. 오른쪽 남자는 오른쪽 아래를 곁눈질하며 몸을 앞으로 기울인다.",
        "built_space": "상반신 클로즈업 뒤로 흙바닥과 상단의 창고 문턱 하나, 오른쪽 문짝 일부와 벽 일부가 보인다. 이전 장소의 흙바닥·목재·돌벽 계열은 이어지지만 수레와 대문은 화면 밖이라 상태를 확인할 수 없다. 중복 설비나 반사는 보이지 않는다.",
        "entities": "두 사람만 보인다. 왼쪽은 검은 머리를 뒤로 묶은 젊은 동아시아계 남성이지만, 지정된 안규홍의 낡은 밝은 저고리 대신 이전 장면의 짙은 누비옷을 입었다. 오른쪽은 중년 동아시아계 남성이지만 흐트러진 머리, 수염과 거친 흰옷이 박제현 참고의 얼굴·갓·연한 황색 포와 다르고 이전 장면의 왼쪽 인물을 닮았다. 칼의 마름모 감개와 장식 금속은 참고와 유사하다. 옥가락지는 드러나지 않으며 읽을 수 있는 글자는 없다.",
        "hard_violations": [],
        "physics": "칼은 왼쪽 남자의 손가락과 손바닥에 잡혀 있고 팔은 굽힌 팔꿈치로 이어진다. 오른쪽 남자의 상체는 화면 아래로 이어져 공중에 뜬 모습은 아니다. 발은 잘려 있어 접지 상태는 확인할 수 없다. 자세 자체는 가능하지만 손잡이와 뒷목의 충돌 접점이 없어 요구된 충격의 전달은 성립해 보이지 않는다."
       },
       {
        "label": "A",
        "direction": "안규홍은 상대의 목과 접촉 부위를 내려다보며 팔을 오른쪽으로 뻗는다. 박제현 역할의 남자는 고개를 아래로 숙인다. 손잡이 끝은 왼쪽 아래를 향하고, 목에 닿은 부분은 귀 아래 옆목과 뒷목 경계에 놓인 둥근 코등이 부근이다. A보다 타격 접촉은 읽히지만 노출된 뒷목을 손잡이 끝으로 강타하는 방향은 명확하지 않다.",
        "built_space": "두 상체 사이와 뒤에 제한된 흙바닥이 보이고, 왼쪽 뒤에는 열린 창고 문짝 하나와 어두운 출입구가 있다. 오른쪽 뒤에는 수레 한 대, 보이는 바퀴 하나와 적재된 짚 포장 쌀가마니들이 있다. 가마니는 흐려 정확한 개수는 판독하기 어렵다. 수레가 인물보다 뒤에 작게 놓여 장소와 규모 관계가 자연스럽고, 중복 설비나 반사는 없다. 대문은 화면 밖이다.",
        "entities": "두 사람만 보인다. 왼쪽의 젊은 동아시아계 남성은 묶은 검은 머리와 낡은 밝은 저고리로 안규홍 참고에 A보다 가깝고, 드러난 손과 팔도 그 인물에 연결된다. 오른쪽 중년 남성은 얼굴 대부분이 가려졌지만 풀어진 짧은 머리와 거친 흰옷으로, 박제현 참고의 갓·정돈된 머리·연한 황색 포와 다르다. 칼 손잡이의 감개와 장식 끝은 참고와 유사하나 칼날은 보이지 않아 전체 형태를 확인할 수 없다. 옥가락지는 노출되지 않았고 읽을 수 있는 글자는 없다.",
        "hard_violations": [],
        "physics": "칼 손잡이는 안규홍의 손에 단단히 잡혀 있고 손목과 전완이 타격 부위까지 이어진다. 상대는 목과 상체를 앞으로 굽혀 충격에 무너지는 반응을 보인다. 두 몸은 화면 아래로 이어지며 공중 부양의 증거는 없고, 발은 클로즈업 밖이다. 접촉과 체중 이동은 물리적으로 가능하지만, 주먹과 코등이가 접점을 가려 손잡이 타격인지 목을 밀어 누르는 동작인지 모호하다."
       }
      ]
     },
     "ok": true
    }
   ],
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "slot_winner_match": true,
   "slot_winner": {
    "gemini-pro": "A",
    "gpt-high": "A"
   },
   "route": "cross_slot_agree"
  },
  "dual": {
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "normalized": {
    "A": 2.0,
    "B": 1.4
   },
   "adjusted": {
    "A": 1.75,
    "B": 1.15
   },
   "violations": {
    "A": [
     "[gemini-pro] 칼날이 없는 물리적으로 불가능한 형태의 검 (발명된 사물)",
     "[gemini-pro] 프롬프트에 명시된 '뒷목 타격'이 아닌 얼굴 측면/뺨을 찌르듯 누르고 있는 잘못된 액션 연출"
    ],
    "B": [
     "[gemini-pro] 환도의 칼날이 박제현의 목을 관통하여 반대편으로 튀어나온 물리적으로 불가능한 인체/사물 묘사",
     "[gemini-pro] 안규홍에게 이전 샷의 의상(어두운 누비옷)을 입혀, 레퍼런스 의상만 사용하라는 명시적 금지 지침 위반"
    ]
   },
   "per_model_winner": {
    "gemini-pro": "A",
    "gpt-high": "A"
   },
   "agreed": true
  },
  "totals": {
   "A": 1750,
   "B": 1150
  },
  "selected": "A",
  "ranking": [
   "A",
   "B"
  ],
  "verdicts": [
   {
    "label": "A",
    "score": 1750,
    "verdict_ko": "검의 칼날이 아예 생략된 채 손잡이만 들고 있으며, 뒷목이 아닌 얼굴 측면을 타격하여 핵심 액션과 사물 묘사에 실패했습니다.  ★위반: [gemini-pro] 칼날이 없는 물리적으로 불가능한 형태의 검 (발명된 사물) / [gemini-pro] 프롬프트에 명시된 '뒷목 타격'이 아닌 얼굴 측면/뺨을 찌르듯 누르고 있는 잘못된 액션 연출"
   },
   {
    "label": "B",
    "score": 1150,
    "verdict_ko": "칼날이 인체의 목을 관통하는 물리적 오류가 발생했으며, 안규홍이 이전 샷의 의상을 그대로 입고 등장해 지침을 심각하게 위반했습니다.  ★위반: [gemini-pro] 환도의 칼날이 박제현의 목을 관통하여 반대편으로 튀어나온 물리적으로 불가능한 인체/사물 묘사 / [gemini-pro] 안규홍에게 이전 샷의 의상(어두운 누비옷)을 입혀, 레퍼런스 의상만 사용하라는 명시적 금지 지침 위반"
   }
  ],
  "refs": [
   {
    "label": "PREVIOUS SHOT STILL — a visually related earlier shot of this same place: the location's look, materials, fixed features and lighting mood are LOCKED to this photo; never copy its camera framing. The people visible in that photo are NOT in this shot: never carry their faces, bodies or clothing onto anyone here. Each person in THIS shot is defined solely by the PEOPLE list, their own CHARACTER REFERENCE images and their wardrobe notes. If this photo shows a character who cannot move (dead or unconscious), that character's exact pose, position and orientation are ALSO LOCKED — treat their body as a fixed prop of the set that only the camera moves around.",
    "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/db00517b-165b-4005-a287-f28936b59b95/images/272bd7af-9294-4490-98c6-88a40e9deacc/scene/recipe/S34sh8_sel.png",
    "asset_id": "966124bb-5e97-4b1f-ad49-aa414513e4eb",
    "role": "prev_still"
   },
   {
    "label": "CHARACTER REFERENCE — 안규홍: the exact person appearing in this shot; match face, hair and build exactly.",
    "path": "<bytes:1028384>",
    "asset_id": "3b5894db-ce55-4f34-b966-5e725c3aa952",
    "role": "character_ref"
   },
   {
    "label": "CHARACTER REFERENCE — 박제현: the exact person appearing in this shot; match face, hair and build exactly.",
    "path": "<bytes:870304>",
    "asset_id": "02a56dab-2641-421d-89df-af22b5de63df",
    "role": "character_ref"
   },
   {
    "label": "PROP REFERENCE — 칼: the exact object appearing in this shot; match its look, material and wear exactly.",
    "path": "<bytes:780122>",
    "asset_id": "82614014-6091-423c-89ca-1e15d209de72",
    "role": "prop_ref"
   }
  ],
  "critique_skipped": true,
  "needs_reshoot": true,
  "shot_run_produced": false,
  "shot_run_uid": "06aa1428-474d-77cf-92b8-54967e1dd5d9",
  "ref_mode": "prev+엔티티",
  "share_plan": {
   "ref_plan": "prev",
   "prev_anchor_tag": "S34sh8"
  }
 },
 "S34sh16::cine": {
  "applied": true,
  "attempted_at": "2026-09-09T04:39:29.071527+00:00",
  "fingerprint": "0b2cce7dcd7263ce168a79cad9d1e0e21c12f218c288f3470b85d9b2fa22f69f",
  "fingerprint_version": 2,
  "provider": "mai",
  "endpoint": "openrouter/chat-completions",
  "model": "microsoft/mai-image-2.6",
  "pack": "24.202608252115",
  "source_file": "S34sh16_sel.png",
  "source_sha256": "e39aa8a3311877ee8585062935627a571777ffe78eaa5784f997867c1e96b2c6",
  "file": "S34sh16_cine.png",
  "staged_sha256": "c0126d2668229be3eccaf7d1355a1e105db6b4d1f9e16dc5cf868ea8c2d29bf6",
  "latency_ms": 24610
 },
 "S34sh24::signage": {
  "fp": "d9a7e353900eb479",
  "inscriptions": [],
  "cues": [],
  "dropped": []
 },
 "S34sh24": {
  "input_fingerprint": "5ffae50f28606202",
  "prompt": "Create ONE FINAL photorealistic live-action film still of the moment below — early 20th-century Korea, Jeolla region; all people are Korean unless explicitly stated otherwise. TIME OF DAY (lock): night.\n\nSHOT TEXT (authoritative, Korean): 수레가 사라진 텅 빈 대문 밖 어둠을 가만히 내려다보는 박제현의 안도하는 표정.\n\nLOCATION (lock): The estate courtyard beside the granary-loading area, facing the empty main gateway and the darkness beyond. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: close-up\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Keep the courtyard's subdued nighttime ambience unchanged, allowing relief to emerge through expression rather than increased brightness.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The people visible in that photo are NOT in this shot: never carry their faces, bodies or clothing onto anyone here. Each person in THIS shot is defined solely by the PEOPLE list, their own CHARACTER REFERENCE images and their wardrobe notes. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The loaded cart has departed through the still-open main gate; the grain store remains emptied of the donated bales with its door open. 박제현: He has regained consciousness and is upright, clutching the sore back of his neck.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 박제현 (한국인 남성, 중년 이상 나이대, 자연스러운 연령대의 얼굴) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.",
  "roll_prompts": {
   "A": "Create ONE FINAL photorealistic live-action film still of the moment below — early 20th-century Korea, Jeolla region; all people are Korean unless explicitly stated otherwise. TIME OF DAY (lock): night.\n\nSHOT TEXT (authoritative, Korean): 수레가 사라진 텅 빈 대문 밖 어둠을 가만히 내려다보는 박제현의 안도하는 표정.\n\nLOCATION (lock): The estate courtyard beside the granary-loading area, facing the empty main gateway and the darkness beyond. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: close-up\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Keep the courtyard's subdued nighttime ambience unchanged, allowing relief to emerge through expression rather than increased brightness.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The people visible in that photo are NOT in this shot: never carry their faces, bodies or clothing onto anyone here. Each person in THIS shot is defined solely by the PEOPLE list, their own CHARACTER REFERENCE images and their wardrobe notes. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The loaded cart has departed through the still-open main gate; the grain store remains emptied of the donated bales with its door open. 박제현: He has regained consciousness and is upright, clutching the sore back of his neck.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 박제현 (한국인 남성, 중년 이상 나이대, 자연스러운 연령대의 얼굴) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.",
   "B": "Create ONE FINAL photorealistic live-action film still of the moment below — early 20th-century Korea, Jeolla region; all people are Korean unless explicitly stated otherwise. TIME OF DAY (lock): night.\n\nSHOT TEXT (authoritative, Korean): 수레가 사라진 텅 빈 대문 밖 어둠을 가만히 내려다보는 박제현의 안도하는 표정.\n\nLOCATION (lock): The estate courtyard beside the granary-loading area, facing the empty main gateway and the darkness beyond. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: close-up\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Keep the courtyard's subdued nighttime ambience unchanged, allowing relief to emerge through expression rather than increased brightness.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The people visible in that photo are NOT in this shot: never carry their faces, bodies or clothing onto anyone here. Each person in THIS shot is defined solely by the PEOPLE list, their own CHARACTER REFERENCE images and their wardrobe notes. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The loaded cart has departed through the still-open main gate; the grain store remains emptied of the donated bales with its door open. 박제현: He has regained consciousness and is upright, clutching the sore back of his neck.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 박제현 (한국인 남성, 중년 이상 나이대, 자연스러운 연령대의 얼굴) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.\n\nCOMPOSITION VARIATION (this candidate only): keep the place, the people and the moment exactly as specified, and keep from the CAMERA & FRAME contract its subject scale, screen placement and key background elements exactly. For this candidate only, its camera angle and height are deliberately left open: choose a camera position distinctly different from the obvious one, as a film director picking a second setup on the same blocking. Stay true to any explicit framing the SHOT TEXT itself dictates."
  },
  "shot_run_spend_attempt_count": 1,
  "readings": [
   {
    "label": "A",
    "direction": "인물은 우측 배경의 열린 대문 밖 어둠을 향해 시선을 두고 있음.",
    "built_space": "이전 샷의 마당 구조를 바탕으로 하나, 지시와 달리 좌측 곡물창고 문이 닫혀 있고 우측 대문은 문짝이 없이 프레임만 열려 있음.",
    "entities": "박제현의 얼굴과 체형, 의상(갓 포함)이 캐릭터 참고 이미지와 일치함.",
    "hard_violations": [
     "[gpt-high] 참조의 기존 대문을 오른쪽에 그대로 남긴 채 그 옆 담장 자리에 별도의 대형 출입구를 추가하여, 고정된 장소 구조와 출입구 수를 변경했습니다."
    ],
    "physics": "바닥에 서서 오른손을 들어 뒷목을 잡고 있는 자세가 자연스럽게 지탱됨."
   },
   {
    "label": "B",
    "direction": "인물은 우측 배경의 열린 대문 밖 어둠을 향해 시선을 두고 있음.",
    "built_space": "마당 구조가 이전 샷과 일치하며, 지시대로 좌측 곡물창고 문이 활짝 열려 빈 내부가 드러나고 우측 대문 역시 양쪽 문짝이 열린 상태로 잘 묘사됨.",
    "entities": "박제현의 얼굴과 체형, 의상이 참고 이미지와 일치함(갓은 벗은 상태).",
    "hard_violations": [],
    "physics": "바닥에 서서 왼손으로 뒷목을 감싸 쥐는 자세가 안정적으로 유지됨."
   }
  ],
  "cross_model_order": {
   "policy": "cross_model_order_v2_slot0_fwd_slot1_rev_parallel2",
   "slots": [
    {
     "model": "gemini-pro",
     "order": "forward",
     "display_to_canonical": {
      "A": "A",
      "B": "B"
     },
     "raw": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 8,
        "verdict_ko": "곡물창고 문이 열려 텅 빈 내부가 보인다는 프롬프트의 지시를 정확히 구현했으며, 이전 샷의 대문 문짝 형태를 유지하고 클로즈업에 더 가까운 프레이밍을 보여주어 가장 우수합니다."
       },
       {
        "label": "A",
        "score": 4,
        "verdict_ko": "클로즈업 지시를 어기고 인물을 너무 넓게 잡았으며, 곡물창고 문이 닫혀 있고 대문의 문짝이 아예 사라지는 등 배경 세부 묘사 지시를 놓쳤습니다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "인물은 우측 배경의 열린 대문 밖 어둠을 향해 시선을 두고 있음.",
        "built_space": "이전 샷의 마당 구조를 바탕으로 하나, 지시와 달리 좌측 곡물창고 문이 닫혀 있고 우측 대문은 문짝이 없이 프레임만 열려 있음.",
        "entities": "박제현의 얼굴과 체형, 의상(갓 포함)이 캐릭터 참고 이미지와 일치함.",
        "hard_violations": [],
        "physics": "바닥에 서서 오른손을 들어 뒷목을 잡고 있는 자세가 자연스럽게 지탱됨."
       },
       {
        "label": "B",
        "direction": "인물은 우측 배경의 열린 대문 밖 어둠을 향해 시선을 두고 있음.",
        "built_space": "마당 구조가 이전 샷과 일치하며, 지시대로 좌측 곡물창고 문이 활짝 열려 빈 내부가 드러나고 우측 대문 역시 양쪽 문짝이 열린 상태로 잘 묘사됨.",
        "entities": "박제현의 얼굴과 체형, 의상이 참고 이미지와 일치함(갓은 벗은 상태).",
        "hard_violations": [],
        "physics": "바닥에 서서 왼손으로 뒷목을 감싸 쥐는 자세가 안정적으로 유지됨."
       }
      ],
      "all_candidates_fail": false
     },
     "normalized": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 8,
        "verdict_ko": "곡물창고 문이 열려 텅 빈 내부가 보인다는 프롬프트의 지시를 정확히 구현했으며, 이전 샷의 대문 문짝 형태를 유지하고 클로즈업에 더 가까운 프레이밍을 보여주어 가장 우수합니다."
       },
       {
        "label": "A",
        "score": 4,
        "verdict_ko": "클로즈업 지시를 어기고 인물을 너무 넓게 잡았으며, 곡물창고 문이 닫혀 있고 대문의 문짝이 아예 사라지는 등 배경 세부 묘사 지시를 놓쳤습니다."
       }
      ],
      "all_candidates_fail": false,
      "readings": [
       {
        "label": "A",
        "direction": "인물은 우측 배경의 열린 대문 밖 어둠을 향해 시선을 두고 있음.",
        "built_space": "이전 샷의 마당 구조를 바탕으로 하나, 지시와 달리 좌측 곡물창고 문이 닫혀 있고 우측 대문은 문짝이 없이 프레임만 열려 있음.",
        "entities": "박제현의 얼굴과 체형, 의상(갓 포함)이 캐릭터 참고 이미지와 일치함.",
        "hard_violations": [],
        "physics": "바닥에 서서 오른손을 들어 뒷목을 잡고 있는 자세가 자연스럽게 지탱됨."
       },
       {
        "label": "B",
        "direction": "인물은 우측 배경의 열린 대문 밖 어둠을 향해 시선을 두고 있음.",
        "built_space": "마당 구조가 이전 샷과 일치하며, 지시대로 좌측 곡물창고 문이 활짝 열려 빈 내부가 드러나고 우측 대문 역시 양쪽 문짝이 열린 상태로 잘 묘사됨.",
        "entities": "박제현의 얼굴과 체형, 의상이 참고 이미지와 일치함(갓은 벗은 상태).",
        "hard_violations": [],
        "physics": "바닥에 서서 왼손으로 뒷목을 감싸 쥐는 자세가 안정적으로 유지됨."
       }
      ]
     },
     "ok": true
    },
    {
     "model": "gpt-high",
     "order": "reverse",
     "display_to_canonical": {
      "A": "B",
      "B": "A"
     },
     "raw": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 4,
        "verdict_ko": "열린 대문과 빈 창고, 목덜미를 짚는 상태는 맞지만 클로즈업이 아니며, 시선이 대문 밖에 닿지 않고 갓과 머리 모양도 참조와 다릅니다."
       },
       {
        "label": "B",
        "score": 2,
        "verdict_ko": "갓과 의복은 비교적 충실하지만 기존 대문 옆에 별도 출입구를 만들어 장소를 바꾸었고, 클로즈업과 대문 밖을 내려다보는 시선도 구현하지 못했습니다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "얼굴과 눈은 화면 오른쪽 앞쪽의 프레임 밖을 향합니다. 화면 오른쪽 뒤에 보이는 열린 대문은 인물의 시선 뒤편에 있어 실제 응시 대상이 아닙니다. 아래로 내려다보는 눈길도 뚜렷하지 않습니다. 무기나 방향성 있는 휴대 소품은 없습니다.",
        "built_space": "왼쪽에 목조·회벽 건물과 돌기단, 열린 창고 입구 하나 및 벽걸이 등 하나가 보입니다. 오른쪽 뒤에는 기와지붕 대문 하나가 열려 있고 그 너머는 어둡습니다. 창고의 보이는 내부와 마당에는 곡물 짐이나 수레가 없습니다. 참조의 주요 공간 관계는 대체로 유지하지만, 인물을 허벅지 부근까지 담아 얼굴 클로즈업보다 훨씬 넓습니다.",
        "entities": "중년 한국인 남성으로 보이는 인물 한 명만 있습니다. 크림색 포와 붉은 허리띠, 검은 끈은 참조와 대체로 맞지만 검은 갓이 없고 풀린 짧은 머리가 드러납니다. 얼굴도 참조보다 젊고 다른 인상입니다. 벌어진 입과 긴장된 눈에서는 안도보다 경계나 당혹감이 강하게 읽힙니다. 수레와 추가 인물, 읽을 수 있는 문자는 없습니다.",
        "hard_violations": [],
        "physics": "상체는 곧게 서 있고 굽힌 팔 끝의 손이 목덜미에 실제로 닿아 있습니다. 팔과 손의 연결 및 옷의 처짐은 자연스럽습니다. 발은 프레임 밖이어서 지면 접촉을 확인할 수 없지만, 몸이 떠 있거나 지지 없이 매달린 모습은 아닙니다. 문짝은 문틀에 연결되어 있고 등은 건물에 부착되어 있습니다."
       },
       {
        "label": "B",
        "direction": "인물은 카메라 쪽에서 약간 화면 오른쪽을 바라봅니다. 뒤쪽 중앙의 열린 통로와 오른쪽의 닫힌 대문 어느 쪽으로도 눈길이 향하지 않습니다. 대문 밖 어둠을 내려다보는 행동 대신 앞쪽을 멍하니 보는 모습입니다. 무기나 겨누는 소품은 없습니다.",
        "built_space": "왼쪽 목조 건물과 돌기단, 벽걸이 등 하나와 바닥 등 하나가 보입니다. 오른쪽에는 참조의 작은 기와 대문이 닫힌 채 남아 있고, 그 왼쪽에 별도의 큰 지붕 달린 열린 출입구가 추가되었습니다. 참조에서 담장이 이어지던 부분을 새 통로로 바꾼 구조입니다. 열린 빈 창고 내부는 확인되지 않습니다. 인물은 무릎 부근까지 보여 요구된 클로즈업보다 넓습니다.",
        "entities": "중년 이상 한국인 남성으로 보이는 인물 한 명입니다. 검은 갓과 턱끈, 크림색 포, 붉은 허리띠 및 검은 장식 끈은 참조에 비교적 가깝습니다. 다만 얼굴의 수염과 노화 표현은 참조보다 강합니다. 표정은 안도보다는 피로와 걱정에 가깝습니다. 수레나 곡물 짐, 추가 인물은 없으며 대문의 붉은 표식은 읽을 수 있는 글자로 식별되지 않습니다.",
        "hard_violations": [
         "참조의 기존 대문을 오른쪽에 그대로 남긴 채 그 옆 담장 자리에 별도의 대형 출입구를 추가하여, 고정된 장소 구조와 출입구 수를 변경했습니다."
        ],
        "physics": "인물은 상체를 세우고 한 손으로 목덜미를 짚으며 다른 팔은 아래로 내립니다. 목과 손의 접촉, 팔의 굽힘은 가능한 자세입니다. 하체의 지면 접촉은 잘려 보이지 않지만 공중에 떠 있다는 징후는 없습니다. 갓은 머리에 얹혀 있고 장식 끈은 허리에서 늘어지며, 바닥 등은 기단 위에 놓여 있습니다."
       }
      ],
      "all_candidates_fail": true
     },
     "normalized": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 4,
        "verdict_ko": "열린 대문과 빈 창고, 목덜미를 짚는 상태는 맞지만 클로즈업이 아니며, 시선이 대문 밖에 닿지 않고 갓과 머리 모양도 참조와 다릅니다."
       },
       {
        "label": "A",
        "score": 2,
        "verdict_ko": "갓과 의복은 비교적 충실하지만 기존 대문 옆에 별도 출입구를 만들어 장소를 바꾸었고, 클로즈업과 대문 밖을 내려다보는 시선도 구현하지 못했습니다."
       }
      ],
      "all_candidates_fail": true,
      "readings": [
       {
        "label": "B",
        "direction": "얼굴과 눈은 화면 오른쪽 앞쪽의 프레임 밖을 향합니다. 화면 오른쪽 뒤에 보이는 열린 대문은 인물의 시선 뒤편에 있어 실제 응시 대상이 아닙니다. 아래로 내려다보는 눈길도 뚜렷하지 않습니다. 무기나 방향성 있는 휴대 소품은 없습니다.",
        "built_space": "왼쪽에 목조·회벽 건물과 돌기단, 열린 창고 입구 하나 및 벽걸이 등 하나가 보입니다. 오른쪽 뒤에는 기와지붕 대문 하나가 열려 있고 그 너머는 어둡습니다. 창고의 보이는 내부와 마당에는 곡물 짐이나 수레가 없습니다. 참조의 주요 공간 관계는 대체로 유지하지만, 인물을 허벅지 부근까지 담아 얼굴 클로즈업보다 훨씬 넓습니다.",
        "entities": "중년 한국인 남성으로 보이는 인물 한 명만 있습니다. 크림색 포와 붉은 허리띠, 검은 끈은 참조와 대체로 맞지만 검은 갓이 없고 풀린 짧은 머리가 드러납니다. 얼굴도 참조보다 젊고 다른 인상입니다. 벌어진 입과 긴장된 눈에서는 안도보다 경계나 당혹감이 강하게 읽힙니다. 수레와 추가 인물, 읽을 수 있는 문자는 없습니다.",
        "hard_violations": [],
        "physics": "상체는 곧게 서 있고 굽힌 팔 끝의 손이 목덜미에 실제로 닿아 있습니다. 팔과 손의 연결 및 옷의 처짐은 자연스럽습니다. 발은 프레임 밖이어서 지면 접촉을 확인할 수 없지만, 몸이 떠 있거나 지지 없이 매달린 모습은 아닙니다. 문짝은 문틀에 연결되어 있고 등은 건물에 부착되어 있습니다."
       },
       {
        "label": "A",
        "direction": "인물은 카메라 쪽에서 약간 화면 오른쪽을 바라봅니다. 뒤쪽 중앙의 열린 통로와 오른쪽의 닫힌 대문 어느 쪽으로도 눈길이 향하지 않습니다. 대문 밖 어둠을 내려다보는 행동 대신 앞쪽을 멍하니 보는 모습입니다. 무기나 겨누는 소품은 없습니다.",
        "built_space": "왼쪽 목조 건물과 돌기단, 벽걸이 등 하나와 바닥 등 하나가 보입니다. 오른쪽에는 참조의 작은 기와 대문이 닫힌 채 남아 있고, 그 왼쪽에 별도의 큰 지붕 달린 열린 출입구가 추가되었습니다. 참조에서 담장이 이어지던 부분을 새 통로로 바꾼 구조입니다. 열린 빈 창고 내부는 확인되지 않습니다. 인물은 무릎 부근까지 보여 요구된 클로즈업보다 넓습니다.",
        "entities": "중년 이상 한국인 남성으로 보이는 인물 한 명입니다. 검은 갓과 턱끈, 크림색 포, 붉은 허리띠 및 검은 장식 끈은 참조에 비교적 가깝습니다. 다만 얼굴의 수염과 노화 표현은 참조보다 강합니다. 표정은 안도보다는 피로와 걱정에 가깝습니다. 수레나 곡물 짐, 추가 인물은 없으며 대문의 붉은 표식은 읽을 수 있는 글자로 식별되지 않습니다.",
        "hard_violations": [
         "참조의 기존 대문을 오른쪽에 그대로 남긴 채 그 옆 담장 자리에 별도의 대형 출입구를 추가하여, 고정된 장소 구조와 출입구 수를 변경했습니다."
        ],
        "physics": "인물은 상체를 세우고 한 손으로 목덜미를 짚으며 다른 팔은 아래로 내립니다. 목과 손의 접촉, 팔의 굽힘은 가능한 자세입니다. 하체의 지면 접촉은 잘려 보이지 않지만 공중에 떠 있다는 징후는 없습니다. 갓은 머리에 얹혀 있고 장식 끈은 허리에서 늘어지며, 바닥 등은 기단 위에 놓여 있습니다."
       }
      ]
     },
     "ok": true
    }
   ],
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "slot_winner_match": true,
   "slot_winner": {
    "gemini-pro": "B",
    "gpt-high": "B"
   },
   "route": "cross_slot_agree"
  },
  "dual": {
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "normalized": {
    "A": 1.0,
    "B": 2.0
   },
   "adjusted": {
    "A": 0.75,
    "B": 2.0
   },
   "violations": {
    "A": [
     "[gpt-high] 참조의 기존 대문을 오른쪽에 그대로 남긴 채 그 옆 담장 자리에 별도의 대형 출입구를 추가하여, 고정된 장소 구조와 출입구 수를 변경했습니다."
    ]
   },
   "per_model_winner": {
    "gemini-pro": "B",
    "gpt-high": "B"
   },
   "agreed": true
  },
  "totals": {
   "B": 2000,
   "A": 750
  },
  "selected": "B",
  "ranking": [
   "B",
   "A"
  ],
  "verdicts": [
   {
    "label": "B",
    "score": 2000,
    "verdict_ko": "곡물창고 문이 열려 텅 빈 내부가 보인다는 프롬프트의 지시를 정확히 구현했으며, 이전 샷의 대문 문짝 형태를 유지하고 클로즈업에 더 가까운 프레이밍을 보여주어 가장 우수합니다."
   },
   {
    "label": "A",
    "score": 750,
    "verdict_ko": "클로즈업 지시를 어기고 인물을 너무 넓게 잡았으며, 곡물창고 문이 닫혀 있고 대문의 문짝이 아예 사라지는 등 배경 세부 묘사 지시를 놓쳤습니다.  ★위반: [gpt-high] 참조의 기존 대문을 오른쪽에 그대로 남긴 채 그 옆 담장 자리에 별도의 대형 출입구를 추가하여, 고정된 장소 구조와 출입구 수를 변경했습니다."
   }
  ],
  "refs": [
   {
    "label": "PREVIOUS SHOT STILL — a visually related earlier shot of this same place: the location's look, materials, fixed features and lighting mood are LOCKED to this photo; never copy its camera framing. The people visible in that photo are NOT in this shot: never carry their faces, bodies or clothing onto anyone here. Each person in THIS shot is defined solely by the PEOPLE list, their own CHARACTER REFERENCE images and their wardrobe notes. If this photo shows a character who cannot move (dead or unconscious), that character's exact pose, position and orientation are ALSO LOCKED — treat their body as a fixed prop of the set that only the camera moves around.",
    "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/db00517b-165b-4005-a287-f28936b59b95/images/272bd7af-9294-4490-98c6-88a40e9deacc/scene/recipe/S28sh21_sel.png",
    "asset_id": "2e453814-27ba-4736-bcdb-668f831d2650",
    "role": "prev_still"
   },
   {
    "label": "CHARACTER REFERENCE — 박제현: the exact person appearing in this shot; match face, hair and build exactly.",
    "path": "<bytes:870304>",
    "asset_id": "02a56dab-2641-421d-89df-af22b5de63df",
    "role": "character_ref"
   }
  ],
  "critique_skipped": true,
  "shot_run_produced": false,
  "shot_run_uid": "06aa1428-48f3-7604-8dfd-d58683764af8",
  "ref_mode": "prev+엔티티",
  "share_plan": {
   "ref_plan": "prev",
   "prev_anchor_tag": "S28sh21"
  }
 },
 "S34sh24::cine": {
  "applied": true,
  "attempted_at": "2026-09-09T04:40:46.745157+00:00",
  "fingerprint": "c5bd4d613529b72c755d8050d6e0e2f0bc37fd51a8778256f103e8d93fd02cd7",
  "fingerprint_version": 2,
  "provider": "mai",
  "endpoint": "openrouter/chat-completions",
  "model": "microsoft/mai-image-2.6",
  "pack": "24.202608252115",
  "source_file": "S34sh24_sel.png",
  "source_sha256": "4da8b267799a3854b797e96a944cc84ab01db2ae47cee7c74ac687698fb4a1e8",
  "file": "S34sh24_cine.png",
  "staged_sha256": "1c26f5d5885e7c1954022759012a910a8e790399d38b73c12319aff6b0608759",
  "latency_ms": 26761
 },
 "S35sh8::signage": {
  "fp": "ed481983807f7e8d",
  "inscriptions": [],
  "cues": [],
  "dropped": []
 },
 "S35sh8::bgfirst_bg": {
  "input_fingerprint": "01b2f68370120858",
  "prompt": "Create the EMPTY BACKGROUND PLATE for one film shot — NO PEOPLE, no figures, no body parts, no sketch lines, no arrows anywhere.\n\"Empty\" means no people only: KEEP the location's inherent occupants and stock that define the place — animals in an animal shelter, pen or farm, goods in a market, moored boats in a harbour — unless the shot text explicitly removes them.\nThe FIRST attached image is a thin-line storyboard sketch: use ONLY its camera angle, horizon, perspective and the placement/size of buildings and set masses — ignore the sketched people and arrows entirely. The SECOND attached image (LOCATION PHOTOGRAPH) is the real place: take its architecture, materials, signage and fixed features, and RE-PROJECT them into the sketch's camera. If the photograph's camera differs from the sketch's, the sketch's camera wins.\nHUMAN-SCALE CALIBRATION: derive every structure's true size from human-scale elements — a door ≈ 2m, a window ≈ 1–1.5m wide, one storey ≈ 2.5–3m; never inflate a small structure or shrink a large one.\n\nSHOT TEXT this background must serve (Korean): 어둠 속 열린 문틈들 사이로 짚신과 옥수수를 쥔 채 문밖을 향해 불쑥 뻗어 나와 정지된 여러 조선인 손들의 겹친 모습.\n\nLOCATION (lock): The outdoor edge of a village lane lined with opening household gates, where supplies are passed toward a cart at night.\n\nTIME OF DAY (lock): night.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: close-up\n- KEY BACKGROUND ELEMENTS: Successive village doorways (Open as offerings are extended outside) — Seen obliquely along their exterior edges, with interiors largely excluded; used as Create successive narrow frames and depth separations around the hands; Straw shoes and corn (Held out in several villagers' hands) — Presented at differing angles along the line of extended arms; used as Provide distinct offering details without allowing any single object to exceed a small portion of the frame.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Restrained nighttime exposure preserves the hands against dark openings, with muted saturation and naturalistic contrast rather than theatrical highlights.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nRender ONE photorealistic empty location photograph, 16:9, that this shot can be staged inside later. No readable writing anywhere: surfaces that would carry writing may be present, but stage any wording out of legibility — an oblique angle, distance, shallow focus. No captions, watermarks or overlay text.",
  "effective_prompt": "Create the EMPTY BACKGROUND PLATE for one film shot — NO PEOPLE, no figures, no body parts, no sketch lines, no arrows anywhere.\n\"Empty\" means no people only: KEEP the location's inherent occupants and stock that define the place — animals in an animal shelter, pen or farm, goods in a market, moored boats in a harbour — unless the shot text explicitly removes them.\nThe FIRST attached image is a thin-line storyboard sketch: use ONLY its camera angle, horizon, perspective and the placement/size of buildings and set masses — ignore the sketched people and arrows entirely. The SECOND attached image (LOCATION PHOTOGRAPH) is the real place: take its architecture, materials, signage and fixed features, and RE-PROJECT them into the sketch's camera. If the photograph's camera differs from the sketch's, the sketch's camera wins.\nHUMAN-SCALE CALIBRATION: derive every structure's true size from human-scale elements — a door ≈ 2m, a window ≈ 1–1.5m wide, one storey ≈ 2.5–3m; never inflate a small structure or shrink a large one.\n\nSHOT TEXT this background must serve (Korean): 어둠 속 열린 문틈들 사이로 짚신과 옥수수를 쥔 채 문밖을 향해 불쑥 뻗어 나와 정지된 여러 조선인 손들의 겹친 모습.\n\nLOCATION (lock): The outdoor edge of a village lane lined with opening household gates, where supplies are passed toward a cart at night.\n\nTIME OF DAY (lock): night.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: close-up\n- KEY BACKGROUND ELEMENTS: Successive village doorways (Open as offerings are extended outside) — Seen obliquely along their exterior edges, with interiors largely excluded; used as Create successive narrow frames and depth separations around the hands; Straw shoes and corn (Held out in several villagers' hands) — Presented at differing angles along the line of extended arms; used as Provide distinct offering details without allowing any single object to exceed a small portion of the frame.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Restrained nighttime exposure preserves the hands against dark openings, with muted saturation and naturalistic contrast rather than theatrical highlights.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nRender ONE photorealistic empty location photograph, 16:9, that this shot can be staged inside later. No readable writing anywhere: surfaces that would carry writing may be present, but stage any wording out of legibility — an oblique angle, distance, shallow focus. No captions, watermarks or overlay text.",
  "bg_path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/db00517b-165b-4005-a287-f28936b59b95/images/272bd7af-9294-4490-98c6-88a40e9deacc/scene/recipe/S35sh8__bgfirst_bg.png",
  "asset_id": "1b47337e-6382-4b94-a173-30a81265ae3b",
  "input_asset_ids": [
   "55aaf43a-c73c-4965-89dd-f589f5a02422",
   "aeaa845f-49c4-45b0-8a1f-1c8f50c6daed"
  ]
 },
 "S35sh8": {
  "input_fingerprint": "8881b98b1745e2be",
  "prompt": "Create ONE FINAL photorealistic live-action film still of the moment below — early 20th-century Korea, Jeolla region; all people are Korean unless explicitly stated otherwise. TIME OF DAY (lock): night.\n\nSHOT TEXT (authoritative, Korean): 어둠 속 열린 문틈들 사이로 짚신과 옥수수를 쥔 채 문밖을 향해 불쑥 뻗어 나와 정지된 여러 조선인 손들의 겹친 모습.\n\nLOCATION (lock): The outdoor edge of a village lane lined with opening household gates, where supplies are passed toward a cart at night. The shot takes place here — the attached LOCATION STRUCTURE PHOTOGRAPH is the single authority for this exact place — its fixed structure and permanent site details are LOCKED to it. No separate location photograph exists for this place. Build everything else strictly from the location text above and the shot text; the layout sketch (when attached) governs framing and placement only, and the shot text governs time of day, lighting and action.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: close-up\n- KEY BACKGROUND ELEMENTS: Successive village doorways (Open as offerings are extended outside) — Seen obliquely along their exterior edges, with interiors largely excluded; used as Create successive narrow frames and depth separations around the hands; Straw shoes and corn (Held out in several villagers' hands) — Presented at differing angles along the line of extended arms; used as Provide distinct offering details without allowing any single object to exceed a small portion of the frame.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Restrained nighttime exposure preserves the hands against dark openings, with muted saturation and naturalistic contrast rather than theatrical highlights.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The cart holds the donated rice bales plus a barley sack and potato basket; village gates are opening as straw shoes and corn are added.\n\nTHE HAND THAT IS DOING THIS: the object at the centre of this shot is being held, operated, read, aimed or handed over by 여러 조선인 손들 right now, so 여러 조선인 손들's hand — and as much of the wrist and forearm as the framing reaches — is in the frame, gripping or resting on that object exactly the way its real use requires. Match that hand to 여러 조선인 손들: its size, build, skin, age, grooming, sleeve and anything worn on it belong to that person and to no one else.\n\nNo people appear unless the shot text itself says so.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.",
  "roll_prompts": {
   "A": "Stage the shot. The FIRST attached image (SHOT BACKGROUND) is the finished empty background of this shot — keep it EXACTLY: its camera, perspective, architecture, lighting and every fixed feature stay untouched. The SECOND attached image (LAYOUT SKETCH) tells you ONLY where the people go: each sketched person's position, screen size, pose and the gaze/motion arrows. Ignore the sketch's background lines. The CHARACTER REFERENCE photographs show the real people.\nPlace the real people into the background at exactly the sketched positions, sizes and poses, following the arrow directions. No sketch lines or arrows may remain.\n\nCreate ONE FINAL photorealistic live-action film still of the moment below — early 20th-century Korea, Jeolla region; all people are Korean unless explicitly stated otherwise. TIME OF DAY (lock): night.\n\nSHOT TEXT (authoritative, Korean): 어둠 속 열린 문틈들 사이로 짚신과 옥수수를 쥔 채 문밖을 향해 불쑥 뻗어 나와 정지된 여러 조선인 손들의 겹친 모습.\n\nLOCATION (lock): The outdoor edge of a village lane lined with opening household gates, where supplies are passed toward a cart at night. The shot takes place here — the FIRST attached image (SHOT BACKGROUND) is this exact place, already built: its ground, structures, horizon, materials and lighting are the finished truth of this location and must not be redesigned or replaced. No location photograph is attached — read the place from that image alone, and add no scenery, structure, vehicle or fixture that it does not already show. This lock governs the place only; the figures in the shot follow the staging and pose instructions.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: close-up\n- KEY BACKGROUND ELEMENTS: Successive village doorways (Open as offerings are extended outside) — Seen obliquely along their exterior edges, with interiors largely excluded; used as Create successive narrow frames and depth separations around the hands; Straw shoes and corn (Held out in several villagers' hands) — Presented at differing angles along the line of extended arms; used as Provide distinct offering details without allowing any single object to exceed a small portion of the frame.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Restrained nighttime exposure preserves the hands against dark openings, with muted saturation and naturalistic contrast rather than theatrical highlights.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The cart holds the donated rice bales plus a barley sack and potato basket; village gates are opening as straw shoes and corn are added.\n\nTHE HAND THAT IS DOING THIS: the object at the centre of this shot is being held, operated, read, aimed or handed over by 여러 조선인 손들 right now, so 여러 조선인 손들's hand — and as much of the wrist and forearm as the framing reaches — is in the frame, gripping or resting on that object exactly the way its real use requires. Match that hand to 여러 조선인 손들: its size, build, skin, age, grooming, sleeve and anything worn on it belong to that person and to no one else.\n\nNo people appear unless the shot text itself says so.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.",
   "B": "Create ONE FINAL photorealistic live-action film still of the moment below — early 20th-century Korea, Jeolla region; all people are Korean unless explicitly stated otherwise. TIME OF DAY (lock): night.\n\nSHOT TEXT (authoritative, Korean): 어둠 속 열린 문틈들 사이로 짚신과 옥수수를 쥔 채 문밖을 향해 불쑥 뻗어 나와 정지된 여러 조선인 손들의 겹친 모습.\n\nLOCATION (lock): The outdoor edge of a village lane lined with opening household gates, where supplies are passed toward a cart at night. The shot takes place here — the attached LOCATION STRUCTURE PHOTOGRAPH is the single authority for this exact place — its fixed structure and permanent site details are LOCKED to it. No separate location photograph exists for this place. Build everything else strictly from the location text above and the shot text; the layout sketch (when attached) governs framing and placement only, and the shot text governs time of day, lighting and action.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: close-up\n- KEY BACKGROUND ELEMENTS: Successive village doorways (Open as offerings are extended outside) — Seen obliquely along their exterior edges, with interiors largely excluded; used as Create successive narrow frames and depth separations around the hands; Straw shoes and corn (Held out in several villagers' hands) — Presented at differing angles along the line of extended arms; used as Provide distinct offering details without allowing any single object to exceed a small portion of the frame.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Restrained nighttime exposure preserves the hands against dark openings, with muted saturation and naturalistic contrast rather than theatrical highlights.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The cart holds the donated rice bales plus a barley sack and potato basket; village gates are opening as straw shoes and corn are added.\n\nTHE HAND THAT IS DOING THIS: the object at the centre of this shot is being held, operated, read, aimed or handed over by 여러 조선인 손들 right now, so 여러 조선인 손들's hand — and as much of the wrist and forearm as the framing reaches — is in the frame, gripping or resting on that object exactly the way its real use requires. Match that hand to 여러 조선인 손들: its size, build, skin, age, grooming, sleeve and anything worn on it belong to that person and to no one else.\n\nNo people appear unless the shot text itself says so.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.\n\nCOMPOSITION VARIATION (this candidate only): keep the place, the people and the moment exactly as specified, and keep from the CAMERA & FRAME contract its subject scale, screen placement and key background elements exactly. For this candidate only, its camera angle and height are deliberately left open: choose a camera position distinctly different from the obvious one, as a film director picking a second setup on the same blocking. Stay true to any explicit framing the SHOT TEXT itself dictates."
  },
  "roll_refs": {
   "A": [
    {
     "label": "SHOT BACKGROUND",
     "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/db00517b-165b-4005-a287-f28936b59b95/images/272bd7af-9294-4490-98c6-88a40e9deacc/scene/recipe/S35sh8__bgfirst_bg.png",
     "asset_id": "1b47337e-6382-4b94-a173-30a81265ae3b",
     "role": "bgfirst_bg"
    },
    {
     "label": "LAYOUT SKETCH (people placement only)",
     "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/db00517b-165b-4005-a287-f28936b59b95/images/272bd7af-9294-4490-98c6-88a40e9deacc/conti/conti_S35sh8.png",
     "asset_id": "55aaf43a-c73c-4965-89dd-f589f5a02422",
     "role": "conti_light"
    }
   ],
   "B": [
    {
     "label": "LOCATION STRUCTURE PHOTOGRAPH — the confirmed photograph of this exact place and its fixed structure: it is the SINGLE authority for the location, the structure's shape, proportions, materials, colors, openings and every permanent site detail. Never copy its camera framing, time of day or lighting — the shot text is the authority for those.",
     "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/db00517b-165b-4005-a287-f28936b59b95/images/background_chain/seed_bg_village_lanes_sel.png",
     "asset_id": "aeaa845f-49c4-45b0-8a1f-1c8f50c6daed",
     "role": "location_seed_bg"
    }
   ]
  },
  "shot_run_spend_attempt_count": 1,
  "judge_flip_skipped": "judge_fn_owns_order",
  "readings": [
   {
    "label": "A",
    "direction": "뻗은 팔과 손들이 우측의 수레 방향을 향하고 있음.",
    "built_space": "골목 구조는 참조와 유사하나, 문이 없는 흙벽과 돌담 부분에서 팔들이 나옴.",
    "entities": "짚신과 옥수수를 쥔 손들, 수레(볏짚과 감자 바구니 포함), 프롬프트에 없는 인물 1명.",
    "hard_violations": [
     "[gemini-pro] 단단한 돌담과 흙벽을 뚫고 나온 팔의 불가능한 물리적 배치",
     "[gemini-pro] 지시되지 않은 전체 인물 등장",
     "[gpt-high] 손을 내미는 주민들과 별개로 수레 옆에 서 있는 인물을 추가했다. 지문이 명시한 사람 외에는 등장시키지 말라는 조건을 위반한다."
    ],
    "physics": "우측의 팔들이 개구부가 없는 단단한 벽을 물리적으로 투과하여 뻗어 나옴."
   },
   {
    "label": "B",
    "direction": "손들이 전경에 위치한 빈 수레를 향해 뻗어 있음.",
    "built_space": "골목과 열린 대문 구조가 참조 이미지의 공간과 부합함.",
    "entities": "짚신과 옥수수를 쥔 손들, 내용물이 누락된 빈 수레.",
    "hard_violations": [
     "[gemini-pro] 팔이 없는 손이 다른 팔뚝에서 돋아나는 등 물리적으로 불가능한 해부학적 구조"
    ],
    "physics": "왼쪽 대문 사이로 나온 손과 팔들이 뼈대가 연결되지 않은 채 서로 기괴하게 융합됨."
   }
  ],
  "cross_model_order": {
   "policy": "cross_model_order_v2_slot0_fwd_slot1_rev_parallel2",
   "slots": [
    {
     "model": "gemini-pro",
     "order": "forward",
     "display_to_canonical": {
      "A": "A",
      "B": "B"
     },
     "raw": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 3,
        "verdict_ko": "수레의 지정된 물품 묘사는 좋으나, 벽을 뚫고 나온 팔과 지시되지 않은 인물의 등장으로 인해 실격입니다."
       },
       {
        "label": "B",
        "score": 3,
        "verdict_ko": "수레의 필수 내용물이 누락되었으며, 엉킨 팔들의 불가능한 해부학적 구조로 인해 실격입니다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "뻗은 팔과 손들이 우측의 수레 방향을 향하고 있음.",
        "built_space": "골목 구조는 참조와 유사하나, 문이 없는 흙벽과 돌담 부분에서 팔들이 나옴.",
        "entities": "짚신과 옥수수를 쥔 손들, 수레(볏짚과 감자 바구니 포함), 프롬프트에 없는 인물 1명.",
        "hard_violations": [
         "단단한 돌담과 흙벽을 뚫고 나온 팔의 불가능한 물리적 배치",
         "지시되지 않은 전체 인물 등장"
        ],
        "physics": "우측의 팔들이 개구부가 없는 단단한 벽을 물리적으로 투과하여 뻗어 나옴."
       },
       {
        "label": "B",
        "direction": "손들이 전경에 위치한 빈 수레를 향해 뻗어 있음.",
        "built_space": "골목과 열린 대문 구조가 참조 이미지의 공간과 부합함.",
        "entities": "짚신과 옥수수를 쥔 손들, 내용물이 누락된 빈 수레.",
        "hard_violations": [
         "팔이 없는 손이 다른 팔뚝에서 돋아나는 등 물리적으로 불가능한 해부학적 구조"
        ],
        "physics": "왼쪽 대문 사이로 나온 손과 팔들이 뼈대가 연결되지 않은 채 서로 기괴하게 융합됨."
       }
      ],
      "all_candidates_fail": true
     },
     "normalized": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 3,
        "verdict_ko": "수레의 지정된 물품 묘사는 좋으나, 벽을 뚫고 나온 팔과 지시되지 않은 인물의 등장으로 인해 실격입니다."
       },
       {
        "label": "B",
        "score": 3,
        "verdict_ko": "수레의 필수 내용물이 누락되었으며, 엉킨 팔들의 불가능한 해부학적 구조로 인해 실격입니다."
       }
      ],
      "all_candidates_fail": true,
      "readings": [
       {
        "label": "A",
        "direction": "뻗은 팔과 손들이 우측의 수레 방향을 향하고 있음.",
        "built_space": "골목 구조는 참조와 유사하나, 문이 없는 흙벽과 돌담 부분에서 팔들이 나옴.",
        "entities": "짚신과 옥수수를 쥔 손들, 수레(볏짚과 감자 바구니 포함), 프롬프트에 없는 인물 1명.",
        "hard_violations": [
         "단단한 돌담과 흙벽을 뚫고 나온 팔의 불가능한 물리적 배치",
         "지시되지 않은 전체 인물 등장"
        ],
        "physics": "우측의 팔들이 개구부가 없는 단단한 벽을 물리적으로 투과하여 뻗어 나옴."
       },
       {
        "label": "B",
        "direction": "손들이 전경에 위치한 빈 수레를 향해 뻗어 있음.",
        "built_space": "골목과 열린 대문 구조가 참조 이미지의 공간과 부합함.",
        "entities": "짚신과 옥수수를 쥔 손들, 내용물이 누락된 빈 수레.",
        "hard_violations": [
         "팔이 없는 손이 다른 팔뚝에서 돋아나는 등 물리적으로 불가능한 해부학적 구조"
        ],
        "physics": "왼쪽 대문 사이로 나온 손과 팔들이 뼈대가 연결되지 않은 채 서로 기괴하게 융합됨."
       }
      ]
     },
     "ok": true
    },
    {
     "model": "gpt-high",
     "order": "reverse",
     "display_to_canonical": {
      "A": "B",
      "B": "A"
     },
     "raw": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 7,
        "verdict_ko": "문밖으로 겹쳐 뻗은 손과 서로 다른 각도의 짚신·옥수수를 더 충실하게 담았지만, 길과 수레가 차지하는 여백은 클로즈업 요구보다 넓다."
       },
       {
        "label": "B",
        "score": 3,
        "verdict_ko": "손보다 골목과 수레까지 보여 주는 넓은 구도로 이탈했고, 지문에 없는 수레 옆 인물까지 추가했다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "왼쪽 문틈에서 나온 손들이 화면 오른쪽의 골목과 아래쪽 수레 쪽으로 물품을 내민다. 짚신은 정면과 아래쪽으로, 옥수수는 오른쪽 위와 수평 방향으로 서로 다르게 기울어져 있다. 얼굴과 시선은 보이지 않으며, 문밖으로 공물을 내미는 방향은 맞는다.",
        "built_space": "가까운 갈대 엮음 대문과 그 뒤의 좁은 출입구에서 팔들이 나오며, 더 뒤에는 툇마루가 있는 집과 닫힌 목제 대문이 보인다. 흙·돌담, 담장 위 기와, 목재 기둥과 초가지붕은 참조 장소의 재료와 대체로 일치한다. 외벽을 비스듬히 보는 시점이고 실내는 어둠에 가려져 있지만, 오른쪽 길과 하단 수레가 비교적 넓게 포함된다.",
        "entities": "약 여덟 개의 손이 겹쳐 보이고, 세 손이 짚신을, 여러 다른 손이 옥수수를 쥐고 있으며 맨손 하나도 보인다. 짚신은 실제 짚을 엮은 신발로, 옥수수는 껍질이 일부 벗겨진 이삭으로 읽힌다. 손에는 성인의 주름과 힘줄이 있고 소매는 청색·회백색의 거친 천이다. 얼굴이 없어 한국인 여부와 정확한 성별·나이는 판별할 수 없다. 수레는 일부만 보여 쌀가마·보리 자루·감자 바구니의 적재 상태를 확인할 수 없으며, 읽을 수 있는 글씨는 없다.",
        "hard_violations": [],
        "physics": "짚신은 손가락이 가장자리나 끈을 잡고 있고 옥수수는 이삭 또는 껍질 밑부분을 손으로 움켜쥐어 지탱한다. 손목과 팔뚝이 소매를 거쳐 문 안으로 이어지므로 공중에 분리된 손으로 보이지 않는다. 팔들이 겹치지만 불가능한 관절이나 지지 없는 물품은 확인되지 않는다. 수레 하부는 화면 밖에 있어 바닥 접촉은 보이지 않는다."
       },
       {
        "label": "B",
        "direction": "왼쪽의 여러 출입구에서 손들이 오른쪽 골목과 수레 방향으로 물품을 내민다. 앞쪽 짚신은 펼친 손바닥 위에서 카메라 쪽으로도 제시되고, 옥수수들은 대체로 오른쪽 아래를 향한다. 뒤쪽 수레 옆 사람의 얼굴은 가려져 시선의 목표를 확인할 수 없다.",
        "built_space": "왼쪽에는 적어도 세 개의 열린 출입구가 연속되고, 오른쪽에는 바퀴와 적재함까지 보이는 큰 수레가 놓여 있다. 흙·돌담과 초가지붕은 참조와 유사하지만, 가까운 문들은 참조의 갈대 엮음 대문보다 목제 문틀이 강조된다. 골목 바닥, 하늘, 지붕과 수레까지 넓게 드러나 손들의 클로즈업보다는 골목 전체를 설명하는 구도다.",
        "entities": "약 일곱 개의 손이 보이며 앞쪽과 중간의 손에 짚신, 나머지 여러 손에 옥수수가 있다. 거친 회백색 소매와 성인 손은 시대 설정에 어울리지만 얼굴이 없어 손 주인의 한국인 정체성·성별·정확한 나이는 확인할 수 없다. 수레에는 묶은 짚 포장 짐과 감자 바구니가 보이나 보리 자루는 명확하지 않다. 수레 옆에는 회색 계열 옷과 검은 머리쓰개를 착용한 별도의 성인 인물이 추가되어 있다. 읽을 수 있는 글씨는 없다.",
        "hard_violations": [
         "손을 내미는 주민들과 별개로 수레 옆에 서 있는 인물을 추가했다. 지문이 명시한 사람 외에는 등장시키지 말라는 조건을 위반한다."
        ],
        "physics": "앞쪽 짚신은 손바닥과 구부린 손가락 위에 받쳐져 있고, 중간 짚신과 각 옥수수도 손으로 잡혀 있다. 팔들은 소매와 출입구 안쪽으로 이어져 지지 없는 부유물로 보이지 않는다. 수레는 보이는 바퀴로 길 위에 지지되고 짐은 적재함에 놓여 있다. 수레 옆 인물도 발이 바닥에 닿아 있으며, 명백한 부유나 불가능한 동작은 없다."
       }
      ],
      "all_candidates_fail": false
     },
     "normalized": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 7,
        "verdict_ko": "문밖으로 겹쳐 뻗은 손과 서로 다른 각도의 짚신·옥수수를 더 충실하게 담았지만, 길과 수레가 차지하는 여백은 클로즈업 요구보다 넓다."
       },
       {
        "label": "A",
        "score": 3,
        "verdict_ko": "손보다 골목과 수레까지 보여 주는 넓은 구도로 이탈했고, 지문에 없는 수레 옆 인물까지 추가했다."
       }
      ],
      "all_candidates_fail": false,
      "readings": [
       {
        "label": "B",
        "direction": "왼쪽 문틈에서 나온 손들이 화면 오른쪽의 골목과 아래쪽 수레 쪽으로 물품을 내민다. 짚신은 정면과 아래쪽으로, 옥수수는 오른쪽 위와 수평 방향으로 서로 다르게 기울어져 있다. 얼굴과 시선은 보이지 않으며, 문밖으로 공물을 내미는 방향은 맞는다.",
        "built_space": "가까운 갈대 엮음 대문과 그 뒤의 좁은 출입구에서 팔들이 나오며, 더 뒤에는 툇마루가 있는 집과 닫힌 목제 대문이 보인다. 흙·돌담, 담장 위 기와, 목재 기둥과 초가지붕은 참조 장소의 재료와 대체로 일치한다. 외벽을 비스듬히 보는 시점이고 실내는 어둠에 가려져 있지만, 오른쪽 길과 하단 수레가 비교적 넓게 포함된다.",
        "entities": "약 여덟 개의 손이 겹쳐 보이고, 세 손이 짚신을, 여러 다른 손이 옥수수를 쥐고 있으며 맨손 하나도 보인다. 짚신은 실제 짚을 엮은 신발로, 옥수수는 껍질이 일부 벗겨진 이삭으로 읽힌다. 손에는 성인의 주름과 힘줄이 있고 소매는 청색·회백색의 거친 천이다. 얼굴이 없어 한국인 여부와 정확한 성별·나이는 판별할 수 없다. 수레는 일부만 보여 쌀가마·보리 자루·감자 바구니의 적재 상태를 확인할 수 없으며, 읽을 수 있는 글씨는 없다.",
        "hard_violations": [],
        "physics": "짚신은 손가락이 가장자리나 끈을 잡고 있고 옥수수는 이삭 또는 껍질 밑부분을 손으로 움켜쥐어 지탱한다. 손목과 팔뚝이 소매를 거쳐 문 안으로 이어지므로 공중에 분리된 손으로 보이지 않는다. 팔들이 겹치지만 불가능한 관절이나 지지 없는 물품은 확인되지 않는다. 수레 하부는 화면 밖에 있어 바닥 접촉은 보이지 않는다."
       },
       {
        "label": "A",
        "direction": "왼쪽의 여러 출입구에서 손들이 오른쪽 골목과 수레 방향으로 물품을 내민다. 앞쪽 짚신은 펼친 손바닥 위에서 카메라 쪽으로도 제시되고, 옥수수들은 대체로 오른쪽 아래를 향한다. 뒤쪽 수레 옆 사람의 얼굴은 가려져 시선의 목표를 확인할 수 없다.",
        "built_space": "왼쪽에는 적어도 세 개의 열린 출입구가 연속되고, 오른쪽에는 바퀴와 적재함까지 보이는 큰 수레가 놓여 있다. 흙·돌담과 초가지붕은 참조와 유사하지만, 가까운 문들은 참조의 갈대 엮음 대문보다 목제 문틀이 강조된다. 골목 바닥, 하늘, 지붕과 수레까지 넓게 드러나 손들의 클로즈업보다는 골목 전체를 설명하는 구도다.",
        "entities": "약 일곱 개의 손이 보이며 앞쪽과 중간의 손에 짚신, 나머지 여러 손에 옥수수가 있다. 거친 회백색 소매와 성인 손은 시대 설정에 어울리지만 얼굴이 없어 손 주인의 한국인 정체성·성별·정확한 나이는 확인할 수 없다. 수레에는 묶은 짚 포장 짐과 감자 바구니가 보이나 보리 자루는 명확하지 않다. 수레 옆에는 회색 계열 옷과 검은 머리쓰개를 착용한 별도의 성인 인물이 추가되어 있다. 읽을 수 있는 글씨는 없다.",
        "hard_violations": [
         "손을 내미는 주민들과 별개로 수레 옆에 서 있는 인물을 추가했다. 지문이 명시한 사람 외에는 등장시키지 말라는 조건을 위반한다."
        ],
        "physics": "앞쪽 짚신은 손바닥과 구부린 손가락 위에 받쳐져 있고, 중간 짚신과 각 옥수수도 손으로 잡혀 있다. 팔들은 소매와 출입구 안쪽으로 이어져 지지 없는 부유물로 보이지 않는다. 수레는 보이는 바퀴로 길 위에 지지되고 짐은 적재함에 놓여 있다. 수레 옆 인물도 발이 바닥에 닿아 있으며, 명백한 부유나 불가능한 동작은 없다."
       }
      ]
     },
     "ok": true
    }
   ],
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "slot_winner_match": false,
   "slot_winner": {
    "gemini-pro": "A",
    "gpt-high": "B"
   },
   "route": "cross_slot_combined"
  },
  "dual": {
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "normalized": {
    "A": 1.429,
    "B": 2.0
   },
   "adjusted": {
    "A": 1.179,
    "B": 1.75
   },
   "violations": {
    "A": [
     "[gemini-pro] 단단한 돌담과 흙벽을 뚫고 나온 팔의 불가능한 물리적 배치",
     "[gemini-pro] 지시되지 않은 전체 인물 등장",
     "[gpt-high] 손을 내미는 주민들과 별개로 수레 옆에 서 있는 인물을 추가했다. 지문이 명시한 사람 외에는 등장시키지 말라는 조건을 위반한다."
    ],
    "B": [
     "[gemini-pro] 팔이 없는 손이 다른 팔뚝에서 돋아나는 등 물리적으로 불가능한 해부학적 구조"
    ]
   },
   "per_model_winner": {
    "gemini-pro": "A",
    "gpt-high": "B"
   },
   "agreed": false
  },
  "totals": {
   "A": 1179,
   "B": 1750
  },
  "selected": "B",
  "ranking": [
   "B",
   "A"
  ],
  "verdicts": [
   {
    "label": "A",
    "score": 1179,
    "verdict_ko": "수레의 지정된 물품 묘사는 좋으나, 벽을 뚫고 나온 팔과 지시되지 않은 인물의 등장으로 인해 실격입니다.  ★위반: [gemini-pro] 단단한 돌담과 흙벽을 뚫고 나온 팔의 불가능한 물리적 배치 / [gemini-pro] 지시되지 않은 전체 인물 등장 / [gpt-high] 손을 내미는 주민들과 별개로 수레 옆에 서 있는 인물을 추가했다. 지문이 명시한 사람 외에는 등장시키지 말라는 조건을 위반한다."
   },
   {
    "label": "B",
    "score": 1750,
    "verdict_ko": "수레의 필수 내용물이 누락되었으며, 엉킨 팔들의 불가능한 해부학적 구조로 인해 실격입니다.  ★위반: [gemini-pro] 팔이 없는 손이 다른 팔뚝에서 돋아나는 등 물리적으로 불가능한 해부학적 구조"
   }
  ],
  "refs": [
   {
    "label": "LOCATION STRUCTURE PHOTOGRAPH — the confirmed photograph of this exact place and its fixed structure: it is the SINGLE authority for the location, the structure's shape, proportions, materials, colors, openings and every permanent site detail. Never copy its camera framing, time of day or lighting — the shot text is the authority for those.",
    "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/db00517b-165b-4005-a287-f28936b59b95/images/background_chain/seed_bg_village_lanes_sel.png",
    "asset_id": "aeaa845f-49c4-45b0-8a1f-1c8f50c6daed",
    "role": "location_seed_bg"
   }
  ],
  "critique_skipped": true,
  "shot_run_produced": false,
  "shot_run_uid": "06aa1428-4a90-7225-8df9-7b9893c1e2a2",
  "bgfirst": {
   "bg_path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/db00517b-165b-4005-a287-f28936b59b95/images/272bd7af-9294-4490-98c6-88a40e9deacc/scene/recipe/S35sh8__bgfirst_bg.png",
   "bg_asset_id": "1b47337e-6382-4b94-a173-30a81265ae3b",
   "bg_record_key": "S35sh8::bgfirst_bg",
   "chain_winner": false,
   "authority": "seed_bg"
  },
  "ref_mode": "seed-bg+엔티티 (2택1: 무콘티 승)",
  "share_plan": {
   "ref_plan": "background"
  },
  "lane_policy": "ab_select_ready"
 },
 "S35sh8::cine": {
  "applied": true,
  "attempted_at": "2026-09-09T04:43:49.436839+00:00",
  "fingerprint": "734f9181a17f4e7237d6269c7d31a07b6dd3884e56d6553ea13104389f853c4e",
  "fingerprint_version": 2,
  "provider": "mai",
  "endpoint": "openrouter/chat-completions",
  "model": "microsoft/mai-image-2.6",
  "pack": "24.202608252115",
  "source_file": "S35sh8_sel.png",
  "source_sha256": "c9501e26ab2b8e11931c8bcc1f1e75582ceec18273568781f7aedace05f9f73f",
  "file": "S35sh8_cine.png",
  "staged_sha256": "ff113cc3da9ea158ce956a9fa3c18073797fdb5d4548d4ca413f4dd937245a0b",
  "latency_ms": 26843
 },
 "S35sh12::signage": {
  "fp": "852386f048d0b40f",
  "inscriptions": [],
  "cues": [],
  "dropped": []
 },
 "S35sh12": {
  "input_fingerprint": "7b0b222923971a49",
  "prompt": "Create ONE FINAL photorealistic live-action film still of the moment below — early 20th-century Korea, Jeolla region; all people are Korean unless explicitly stated otherwise. TIME OF DAY (lock): night.\n\nSHOT TEXT (authoritative, Korean): 다시 끈을 꽉 쥐고 묵직한 수레를 끌기 위해 어둠 속을 향해 체중을 앞으로 기울이며 한 발을 크게 내디딘 안규홍, 이관회, 염재보의 역동적인 뒷모습.\n\nLOCATION (lock): The dark village dirt lane beyond the donating households, along the heavily loaded cart's departure route. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: wide shot\n- KEY BACKGROUND ELEMENTS: Loaded rice cart (Heavier with the accumulated village offerings) — Rear and near side visible behind the pulling men; used as Anchor the lower-right frame and make the resistance to their first step legible; Village road (The three men resume their departure along it) — Recedes beyond the men in their established direction of travel; used as Leave forward space beyond their bodies without introducing a new travel axis; Pulling ropes (Gripped tightly as the men restart) — Run from their hands back toward the cart; used as Connect the forward body lean to the weight behind them.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Maintain the preceding nighttime darkness and subdued saturation, retaining enough tonal separation to read the three men's combined effort.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The people visible in that photo are NOT in this shot: never carry their faces, bodies or clothing onto anyone here. Each person in THIS shot is defined solely by the PEOPLE list, their own CHARACTER REFERENCE images and their wardrobe notes. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The cart is piled higher with rice, barley, potatoes, straw shoes and corn; opened village gates line the dark lane. 안규홍: He leans into pulling the heavier cart, retaining his sword and concealed jade ring. 이관회: He leans into pulling the heavier cart. 염재보: He pulls the heavier cart while retaining his gun.\n\nTHE HAND THAT IS DOING THIS: the object at the centre of this shot is being held, operated, read, aimed or handed over by 안규홍, 이관회, 염재보 right now, so 안규홍, 이관회, 염재보's hand — and as much of the wrist and forearm as the framing reaches — is in the frame, gripping or resting on that object exactly the way its real use requires. Match that hand to 안규홍, 이관회, 염재보: its size, build, skin, age, grooming, sleeve and anything worn on it belong to that person and to no one else.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 안규홍 (한국인 남성, 청년 나이대, 검은 머리, 단단히 묶은 머리); 이관회 (한국인 남성, 청년 나이대, 자연스러운 검은 머리); 염재보 (한국인 남성, 성인 나이대, 자연스러운 검은 머리) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.",
  "roll_prompts": {
   "A": "Create ONE FINAL photorealistic live-action film still of the moment below — early 20th-century Korea, Jeolla region; all people are Korean unless explicitly stated otherwise. TIME OF DAY (lock): night.\n\nSHOT TEXT (authoritative, Korean): 다시 끈을 꽉 쥐고 묵직한 수레를 끌기 위해 어둠 속을 향해 체중을 앞으로 기울이며 한 발을 크게 내디딘 안규홍, 이관회, 염재보의 역동적인 뒷모습.\n\nLOCATION (lock): The dark village dirt lane beyond the donating households, along the heavily loaded cart's departure route. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: wide shot\n- KEY BACKGROUND ELEMENTS: Loaded rice cart (Heavier with the accumulated village offerings) — Rear and near side visible behind the pulling men; used as Anchor the lower-right frame and make the resistance to their first step legible; Village road (The three men resume their departure along it) — Recedes beyond the men in their established direction of travel; used as Leave forward space beyond their bodies without introducing a new travel axis; Pulling ropes (Gripped tightly as the men restart) — Run from their hands back toward the cart; used as Connect the forward body lean to the weight behind them.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Maintain the preceding nighttime darkness and subdued saturation, retaining enough tonal separation to read the three men's combined effort.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The people visible in that photo are NOT in this shot: never carry their faces, bodies or clothing onto anyone here. Each person in THIS shot is defined solely by the PEOPLE list, their own CHARACTER REFERENCE images and their wardrobe notes. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The cart is piled higher with rice, barley, potatoes, straw shoes and corn; opened village gates line the dark lane. 안규홍: He leans into pulling the heavier cart, retaining his sword and concealed jade ring. 이관회: He leans into pulling the heavier cart. 염재보: He pulls the heavier cart while retaining his gun.\n\nTHE HAND THAT IS DOING THIS: the object at the centre of this shot is being held, operated, read, aimed or handed over by 안규홍, 이관회, 염재보 right now, so 안규홍, 이관회, 염재보's hand — and as much of the wrist and forearm as the framing reaches — is in the frame, gripping or resting on that object exactly the way its real use requires. Match that hand to 안규홍, 이관회, 염재보: its size, build, skin, age, grooming, sleeve and anything worn on it belong to that person and to no one else.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 안규홍 (한국인 남성, 청년 나이대, 검은 머리, 단단히 묶은 머리); 이관회 (한국인 남성, 청년 나이대, 자연스러운 검은 머리); 염재보 (한국인 남성, 성인 나이대, 자연스러운 검은 머리) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.",
   "B": "Create ONE FINAL photorealistic live-action film still of the moment below — early 20th-century Korea, Jeolla region; all people are Korean unless explicitly stated otherwise. TIME OF DAY (lock): night.\n\nSHOT TEXT (authoritative, Korean): 다시 끈을 꽉 쥐고 묵직한 수레를 끌기 위해 어둠 속을 향해 체중을 앞으로 기울이며 한 발을 크게 내디딘 안규홍, 이관회, 염재보의 역동적인 뒷모습.\n\nLOCATION (lock): The dark village dirt lane beyond the donating households, along the heavily loaded cart's departure route. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: wide shot\n- KEY BACKGROUND ELEMENTS: Loaded rice cart (Heavier with the accumulated village offerings) — Rear and near side visible behind the pulling men; used as Anchor the lower-right frame and make the resistance to their first step legible; Village road (The three men resume their departure along it) — Recedes beyond the men in their established direction of travel; used as Leave forward space beyond their bodies without introducing a new travel axis; Pulling ropes (Gripped tightly as the men restart) — Run from their hands back toward the cart; used as Connect the forward body lean to the weight behind them.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Maintain the preceding nighttime darkness and subdued saturation, retaining enough tonal separation to read the three men's combined effort.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The people visible in that photo are NOT in this shot: never carry their faces, bodies or clothing onto anyone here. Each person in THIS shot is defined solely by the PEOPLE list, their own CHARACTER REFERENCE images and their wardrobe notes. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The cart is piled higher with rice, barley, potatoes, straw shoes and corn; opened village gates line the dark lane. 안규홍: He leans into pulling the heavier cart, retaining his sword and concealed jade ring. 이관회: He leans into pulling the heavier cart. 염재보: He pulls the heavier cart while retaining his gun.\n\nTHE HAND THAT IS DOING THIS: the object at the centre of this shot is being held, operated, read, aimed or handed over by 안규홍, 이관회, 염재보 right now, so 안규홍, 이관회, 염재보's hand — and as much of the wrist and forearm as the framing reaches — is in the frame, gripping or resting on that object exactly the way its real use requires. Match that hand to 안규홍, 이관회, 염재보: its size, build, skin, age, grooming, sleeve and anything worn on it belong to that person and to no one else.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 안규홍 (한국인 남성, 청년 나이대, 검은 머리, 단단히 묶은 머리); 이관회 (한국인 남성, 청년 나이대, 자연스러운 검은 머리); 염재보 (한국인 남성, 성인 나이대, 자연스러운 검은 머리) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.\n\nCOMPOSITION VARIATION (this candidate only): keep the place, the people and the moment exactly as specified, and keep from the CAMERA & FRAME contract its subject scale, screen placement and key background elements exactly. For this candidate only, its camera angle and height are deliberately left open: choose a camera position distinctly different from the obvious one, as a film director picking a second setup on the same blocking. Stay true to any explicit framing the SHOT TEXT itself dictates."
  },
  "shot_run_spend_attempt_count": 1,
  "readings": [
   {
    "label": "A",
    "direction": "세 남자가 전방의 어두운 길을 향해 카메라를 등지고 서서 수레를 끄는 방향으로 시선과 체중을 이동하고 있음.",
    "built_space": "초가 지붕과 흙담으로 이루어진 전통 마을의 어두운 흙길이며, 인물들의 진행 방향인 앞쪽으로 뻗어 있음.",
    "entities": "가운데 인물(안규홍)은 흰 의상, 상투 머리, 검, 오른손의 옥가락지를 정확히 갖춤. 왼쪽 인물(이관회)은 갈색 상의. 오른쪽 인물(염재보)은 어두운 의상과 등에 멘 총을 갖춤. 수레에는 짚신, 감자, 곡식 자루가 일치하게 실려 있음.",
    "hard_violations": [],
    "physics": "세 사람 모두 지면에 발을 딛고 몸을 앞으로 기울이는 자세가 밧줄의 장력과 잘 맞물려 물리적으로 자연스럽게 지지되고 있으며, 수레와 짐, 무기 등도 올바르게 위치함."
   },
   {
    "label": "B",
    "direction": "세 남자가 화면 깊숙한 전방으로 체중을 실어 수레를 끄는 자세를 취하고 있음.",
    "built_space": "양옆에 전통 가옥들이 늘어선 야간의 흙길이 앞쪽으로 이어져 있음.",
    "entities": "왼쪽 인물(안규홍)은 흰 의상과 검, 가운데 인물(이관회)은 갈색 상의를 입었으나, 오른쪽 인물(염재보)은 지정된 어두운 누비옷이 아닌 엉뚱한 밝은 상의를 입고 있음. 수레에는 옥수수, 감자, 가마니가 있음.",
    "hard_violations": [],
    "physics": "지면에 닿은 발과 끄는 자세는 지지되고 있으나, 왼쪽 인물이 잡은 밧줄이 가운데 인물의 손으로 이어지며 물리적 형태가 무너지고 수레의 앞쪽 나무 막대가 구조적으로 끊겨 보임."
   }
  ],
  "cross_model_order": {
   "policy": "cross_model_order_v2_slot0_fwd_slot1_rev_parallel2",
   "slots": [
    {
     "model": "gemini-pro",
     "order": "forward",
     "display_to_canonical": {
      "A": "A",
      "B": "B"
     },
     "raw": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 10,
        "verdict_ko": "세 인물의 의상, 헤어스타일 및 소품(안규홍의 검과 옥가락지, 염재보의 총)이 프롬프트와 레퍼런스에 완벽하게 부합하며, 수레를 끄는 역동적인 동작이 매우 사실적으로 묘사되었습니다."
       },
       {
        "label": "B",
        "score": 5,
        "verdict_ko": "오른쪽 인물(염재보)이 레퍼런스의 어두운 누비옷이 아닌 밝은 상의를 입고 있어 캐릭터 의상 일관성에서 크게 벗어났습니다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "세 남자가 전방의 어두운 길을 향해 카메라를 등지고 서서 수레를 끄는 방향으로 시선과 체중을 이동하고 있음.",
        "built_space": "초가 지붕과 흙담으로 이루어진 전통 마을의 어두운 흙길이며, 인물들의 진행 방향인 앞쪽으로 뻗어 있음.",
        "entities": "가운데 인물(안규홍)은 흰 의상, 상투 머리, 검, 오른손의 옥가락지를 정확히 갖춤. 왼쪽 인물(이관회)은 갈색 상의. 오른쪽 인물(염재보)은 어두운 의상과 등에 멘 총을 갖춤. 수레에는 짚신, 감자, 곡식 자루가 일치하게 실려 있음.",
        "hard_violations": [],
        "physics": "세 사람 모두 지면에 발을 딛고 몸을 앞으로 기울이는 자세가 밧줄의 장력과 잘 맞물려 물리적으로 자연스럽게 지지되고 있으며, 수레와 짐, 무기 등도 올바르게 위치함."
       },
       {
        "label": "B",
        "direction": "세 남자가 화면 깊숙한 전방으로 체중을 실어 수레를 끄는 자세를 취하고 있음.",
        "built_space": "양옆에 전통 가옥들이 늘어선 야간의 흙길이 앞쪽으로 이어져 있음.",
        "entities": "왼쪽 인물(안규홍)은 흰 의상과 검, 가운데 인물(이관회)은 갈색 상의를 입었으나, 오른쪽 인물(염재보)은 지정된 어두운 누비옷이 아닌 엉뚱한 밝은 상의를 입고 있음. 수레에는 옥수수, 감자, 가마니가 있음.",
        "hard_violations": [],
        "physics": "지면에 닿은 발과 끄는 자세는 지지되고 있으나, 왼쪽 인물이 잡은 밧줄이 가운데 인물의 손으로 이어지며 물리적 형태가 무너지고 수레의 앞쪽 나무 막대가 구조적으로 끊겨 보임."
       }
      ],
      "all_candidates_fail": false
     },
     "normalized": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 10,
        "verdict_ko": "세 인물의 의상, 헤어스타일 및 소품(안규홍의 검과 옥가락지, 염재보의 총)이 프롬프트와 레퍼런스에 완벽하게 부합하며, 수레를 끄는 역동적인 동작이 매우 사실적으로 묘사되었습니다."
       },
       {
        "label": "B",
        "score": 5,
        "verdict_ko": "오른쪽 인물(염재보)이 레퍼런스의 어두운 누비옷이 아닌 밝은 상의를 입고 있어 캐릭터 의상 일관성에서 크게 벗어났습니다."
       }
      ],
      "all_candidates_fail": false,
      "readings": [
       {
        "label": "A",
        "direction": "세 남자가 전방의 어두운 길을 향해 카메라를 등지고 서서 수레를 끄는 방향으로 시선과 체중을 이동하고 있음.",
        "built_space": "초가 지붕과 흙담으로 이루어진 전통 마을의 어두운 흙길이며, 인물들의 진행 방향인 앞쪽으로 뻗어 있음.",
        "entities": "가운데 인물(안규홍)은 흰 의상, 상투 머리, 검, 오른손의 옥가락지를 정확히 갖춤. 왼쪽 인물(이관회)은 갈색 상의. 오른쪽 인물(염재보)은 어두운 의상과 등에 멘 총을 갖춤. 수레에는 짚신, 감자, 곡식 자루가 일치하게 실려 있음.",
        "hard_violations": [],
        "physics": "세 사람 모두 지면에 발을 딛고 몸을 앞으로 기울이는 자세가 밧줄의 장력과 잘 맞물려 물리적으로 자연스럽게 지지되고 있으며, 수레와 짐, 무기 등도 올바르게 위치함."
       },
       {
        "label": "B",
        "direction": "세 남자가 화면 깊숙한 전방으로 체중을 실어 수레를 끄는 자세를 취하고 있음.",
        "built_space": "양옆에 전통 가옥들이 늘어선 야간의 흙길이 앞쪽으로 이어져 있음.",
        "entities": "왼쪽 인물(안규홍)은 흰 의상과 검, 가운데 인물(이관회)은 갈색 상의를 입었으나, 오른쪽 인물(염재보)은 지정된 어두운 누비옷이 아닌 엉뚱한 밝은 상의를 입고 있음. 수레에는 옥수수, 감자, 가마니가 있음.",
        "hard_violations": [],
        "physics": "지면에 닿은 발과 끄는 자세는 지지되고 있으나, 왼쪽 인물이 잡은 밧줄이 가운데 인물의 손으로 이어지며 물리적 형태가 무너지고 수레의 앞쪽 나무 막대가 구조적으로 끊겨 보임."
       }
      ]
     },
     "ok": true
    },
    {
     "model": "gpt-high",
     "order": "reverse",
     "display_to_canonical": {
      "A": "B",
      "B": "A"
     },
     "raw": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 6,
        "verdict_ko": "뒷모습과 우하단 수레 배치는 맞지만, 큰 첫걸음과 전방 체중 이동이 약하고 염재보의 밝은 상의가 참조 의상과 다르다."
       },
       {
        "label": "B",
        "score": 8,
        "verdict_ko": "세 사람이 팽팽한 끈을 움켜쥐고 앞으로 기울여 크게 내딛는 순간을 더 충실히 구현하지만, 노출된 옥반지와 빠진 삿갓, 젖은 노면은 지시와 어긋난다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "세 사람 모두 카메라에 등을 보이고 화면 왼쪽 위의 골목 안쪽으로 향한다. 얼굴은 대부분 가려져 정확한 눈길은 확인되지 않지만 머리 방향은 진행로와 일치한다. 손에 잡힌 끈은 뒤쪽 우하단 수레로 이어진다. 염재보의 총은 등에 비스듬히 걸려 있고 총구는 위쪽을 향하며, 사람이나 특정 표적을 겨누지는 않는다. 안규홍의 칼집은 허리에서 아래로 늘어진다.",
        "built_space": "수레 한 대가 우하단을 차지하며 가까운 바퀴 하나와 앞쪽으로 뻗은 끌채 두 개가 보인다. 세 사람은 수레 앞의 흙길에 있고, 그 너머 왼쪽 위로 진행 공간이 남는다. 양옆의 돌·흙 담장, 초가지붕, 목조 출입부는 장소 참조의 재료와 대체로 맞는다. 오른쪽에는 불 켜진 등 하나와 격자창 두 면이 보이며, 먼 정면 대문 한 곳은 닫힌 듯하다. 열린 마을 대문들이 이어진다는 조건은 명확하게 드러나지 않는다. 반사는 없다.",
        "entities": "사람은 세 명뿐이며, 보이는 체격과 머리는 젊거나 성인인 남성 설정에 부합한다. 뒷모습이라 한국인 여부와 정확한 얼굴 동일성은 확인할 수 없다. 왼쪽 안규홍은 묶은 검은 머리, 밝은 저고리와 바지, 허리띠 및 칼을 갖췄다. 가운데 이관회는 짧은 검은 머리, 갈색 저고리와 짧은 밝은 바지가 참조와 맞는다. 오른쪽 염재보는 총을 지녔지만 상의가 밝고 삿갓이 없어 검은 복장의 참조와 다르다. 수레에는 포대, 감자, 짚신, 옥수수가 보인다. 포대 속 쌀과 보리는 외관만으로 구별되지 않는다. 옥반지는 보이지 않아 은닉 조건과 충돌하지 않는다. 읽을 수 있는 글자는 없다.",
        "hard_violations": [],
        "physics": "안규홍과 이관회는 각각 땅을 딛는 발과 뒤꿈치를 드는 발이 보여 보행 중의 지지가 성립한다. 염재보의 하체는 수레에 가려져 접지점을 확인하기 어렵지만, 떠 있는 몸으로 보이지는 않는다. 세 사람의 손이 끈을 잡고 있고 끈은 수레 쪽으로 연결된다. 다만 안규홍은 몸이 비교적 곧고 한 팔을 내려뜨려, 무거운 수레를 재출발시키는 강한 전방 체중 이동이 약하다. 짐은 적재함에 놓이고 결박되어 있으며, 수레는 바퀴로 지지된다. 총은 어깨끈, 칼은 허리 부착부로 지지된다."
       },
       {
        "label": "B",
        "direction": "세 사람의 등과 숙인 머리가 모두 화면 왼쪽 위의 어두운 골목을 향한다. 진행 방향 앞에 길이 남아 있고, 손에서 뒤쪽 우하단 수레로 뻗은 끈이 그 이동축과 일치한다. 염재보의 총은 등에 걸쳐져 총구가 왼쪽 위로 향하지만 조준하는 동작은 아니다. 안규홍의 칼집은 허리에서 뒤쪽 아래로 기울어 있다.",
        "built_space": "우하단에 수레 한 대의 뒤쪽과 가까운 옆면, 바퀴 하나가 보이며, 가까운 끌채 하나의 연결이 명확하다. 세 사람은 수레 앞쪽 길에 나란히 배치되어 있고 길은 왼쪽 위로 이어진다. 돌·흙 담장과 초가 처마, 오른쪽 목조 출입부는 참조 장소의 재료 계열을 유지한다. 다만 어두운 출입부들의 문짝 개수와 개방 상태는 판별하기 어려워, 열린 대문들이 줄지어 있다는 조건은 뚜렷하지 않다. 노면은 참조보다 훨씬 젖어 있다. 물기에서 생기는 약한 반사광은 가능한 표현이며 인물의 불가능한 거울상은 없다.",
        "entities": "등장인물은 정확히 세 명이다. 뒷모습이라 얼굴 동일성과 한국인 여부는 직접 검증되지 않지만, 보이는 연령대와 남성 체격은 설정에 부합한다. 왼쪽 이관회의 갈색 저고리와 짧은 밝은 바지, 가운데 안규홍의 묶은 검은 머리와 밝은 복장, 오른쪽 염재보의 짙은 복장과 정강이 감개가 참조에 잘 맞는다. 염재보의 삿갓은 없다. 안규홍은 허리에 칼을, 염재보는 등에 총을 지녔다. 안규홍의 끈 잡은 손에 녹색 반지가 드러나 은닉된 옥반지 조건과 다르다. 수레에는 여러 포대, 감자, 짚신이 보이지만 쌀과 보리의 내용물 및 옥수수는 확실히 식별되지 않는다. 읽을 수 있는 글자는 없다.",
        "hard_violations": [],
        "physics": "세 사람 모두 상체를 앞으로 기울이고 다리를 앞뒤로 벌려 무거운 수레를 당기는 힘의 방향이 읽힌다. 이관회와 안규홍은 지면을 딛는 앞발과 뒤꿈치를 들어 밀어내는 뒷발이 보이고, 염재보도 수레 옆 아래로 신발의 접지가 드러난다. 세 사람의 손이 실제로 끈을 움켜쥐며 팽팽한 끈은 수레 끌채 쪽으로 이어진다. 수레와 짐은 바퀴 및 적재함으로 지지되고 포대는 밧줄로 묶여 있다. 총은 어깨끈으로, 칼은 허리 부착부로 지지되어 떠 있는 물체는 없다."
       }
      ],
      "all_candidates_fail": false
     },
     "normalized": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 6,
        "verdict_ko": "뒷모습과 우하단 수레 배치는 맞지만, 큰 첫걸음과 전방 체중 이동이 약하고 염재보의 밝은 상의가 참조 의상과 다르다."
       },
       {
        "label": "A",
        "score": 8,
        "verdict_ko": "세 사람이 팽팽한 끈을 움켜쥐고 앞으로 기울여 크게 내딛는 순간을 더 충실히 구현하지만, 노출된 옥반지와 빠진 삿갓, 젖은 노면은 지시와 어긋난다."
       }
      ],
      "all_candidates_fail": false,
      "readings": [
       {
        "label": "B",
        "direction": "세 사람 모두 카메라에 등을 보이고 화면 왼쪽 위의 골목 안쪽으로 향한다. 얼굴은 대부분 가려져 정확한 눈길은 확인되지 않지만 머리 방향은 진행로와 일치한다. 손에 잡힌 끈은 뒤쪽 우하단 수레로 이어진다. 염재보의 총은 등에 비스듬히 걸려 있고 총구는 위쪽을 향하며, 사람이나 특정 표적을 겨누지는 않는다. 안규홍의 칼집은 허리에서 아래로 늘어진다.",
        "built_space": "수레 한 대가 우하단을 차지하며 가까운 바퀴 하나와 앞쪽으로 뻗은 끌채 두 개가 보인다. 세 사람은 수레 앞의 흙길에 있고, 그 너머 왼쪽 위로 진행 공간이 남는다. 양옆의 돌·흙 담장, 초가지붕, 목조 출입부는 장소 참조의 재료와 대체로 맞는다. 오른쪽에는 불 켜진 등 하나와 격자창 두 면이 보이며, 먼 정면 대문 한 곳은 닫힌 듯하다. 열린 마을 대문들이 이어진다는 조건은 명확하게 드러나지 않는다. 반사는 없다.",
        "entities": "사람은 세 명뿐이며, 보이는 체격과 머리는 젊거나 성인인 남성 설정에 부합한다. 뒷모습이라 한국인 여부와 정확한 얼굴 동일성은 확인할 수 없다. 왼쪽 안규홍은 묶은 검은 머리, 밝은 저고리와 바지, 허리띠 및 칼을 갖췄다. 가운데 이관회는 짧은 검은 머리, 갈색 저고리와 짧은 밝은 바지가 참조와 맞는다. 오른쪽 염재보는 총을 지녔지만 상의가 밝고 삿갓이 없어 검은 복장의 참조와 다르다. 수레에는 포대, 감자, 짚신, 옥수수가 보인다. 포대 속 쌀과 보리는 외관만으로 구별되지 않는다. 옥반지는 보이지 않아 은닉 조건과 충돌하지 않는다. 읽을 수 있는 글자는 없다.",
        "hard_violations": [],
        "physics": "안규홍과 이관회는 각각 땅을 딛는 발과 뒤꿈치를 드는 발이 보여 보행 중의 지지가 성립한다. 염재보의 하체는 수레에 가려져 접지점을 확인하기 어렵지만, 떠 있는 몸으로 보이지는 않는다. 세 사람의 손이 끈을 잡고 있고 끈은 수레 쪽으로 연결된다. 다만 안규홍은 몸이 비교적 곧고 한 팔을 내려뜨려, 무거운 수레를 재출발시키는 강한 전방 체중 이동이 약하다. 짐은 적재함에 놓이고 결박되어 있으며, 수레는 바퀴로 지지된다. 총은 어깨끈, 칼은 허리 부착부로 지지된다."
       },
       {
        "label": "A",
        "direction": "세 사람의 등과 숙인 머리가 모두 화면 왼쪽 위의 어두운 골목을 향한다. 진행 방향 앞에 길이 남아 있고, 손에서 뒤쪽 우하단 수레로 뻗은 끈이 그 이동축과 일치한다. 염재보의 총은 등에 걸쳐져 총구가 왼쪽 위로 향하지만 조준하는 동작은 아니다. 안규홍의 칼집은 허리에서 뒤쪽 아래로 기울어 있다.",
        "built_space": "우하단에 수레 한 대의 뒤쪽과 가까운 옆면, 바퀴 하나가 보이며, 가까운 끌채 하나의 연결이 명확하다. 세 사람은 수레 앞쪽 길에 나란히 배치되어 있고 길은 왼쪽 위로 이어진다. 돌·흙 담장과 초가 처마, 오른쪽 목조 출입부는 참조 장소의 재료 계열을 유지한다. 다만 어두운 출입부들의 문짝 개수와 개방 상태는 판별하기 어려워, 열린 대문들이 줄지어 있다는 조건은 뚜렷하지 않다. 노면은 참조보다 훨씬 젖어 있다. 물기에서 생기는 약한 반사광은 가능한 표현이며 인물의 불가능한 거울상은 없다.",
        "entities": "등장인물은 정확히 세 명이다. 뒷모습이라 얼굴 동일성과 한국인 여부는 직접 검증되지 않지만, 보이는 연령대와 남성 체격은 설정에 부합한다. 왼쪽 이관회의 갈색 저고리와 짧은 밝은 바지, 가운데 안규홍의 묶은 검은 머리와 밝은 복장, 오른쪽 염재보의 짙은 복장과 정강이 감개가 참조에 잘 맞는다. 염재보의 삿갓은 없다. 안규홍은 허리에 칼을, 염재보는 등에 총을 지녔다. 안규홍의 끈 잡은 손에 녹색 반지가 드러나 은닉된 옥반지 조건과 다르다. 수레에는 여러 포대, 감자, 짚신이 보이지만 쌀과 보리의 내용물 및 옥수수는 확실히 식별되지 않는다. 읽을 수 있는 글자는 없다.",
        "hard_violations": [],
        "physics": "세 사람 모두 상체를 앞으로 기울이고 다리를 앞뒤로 벌려 무거운 수레를 당기는 힘의 방향이 읽힌다. 이관회와 안규홍은 지면을 딛는 앞발과 뒤꿈치를 들어 밀어내는 뒷발이 보이고, 염재보도 수레 옆 아래로 신발의 접지가 드러난다. 세 사람의 손이 실제로 끈을 움켜쥐며 팽팽한 끈은 수레 끌채 쪽으로 이어진다. 수레와 짐은 바퀴 및 적재함으로 지지되고 포대는 밧줄로 묶여 있다. 총은 어깨끈으로, 칼은 허리 부착부로 지지되어 떠 있는 물체는 없다."
       }
      ]
     },
     "ok": true
    }
   ],
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "slot_winner_match": true,
   "slot_winner": {
    "gemini-pro": "A",
    "gpt-high": "A"
   },
   "route": "cross_slot_agree"
  },
  "dual": {
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "normalized": {
    "A": 2.0,
    "B": 1.25
   },
   "adjusted": {
    "A": 2.0,
    "B": 1.25
   },
   "violations": {},
   "per_model_winner": {
    "gemini-pro": "A",
    "gpt-high": "A"
   },
   "agreed": true
  },
  "totals": {
   "A": 2000,
   "B": 1250
  },
  "selected": "A",
  "ranking": [
   "A",
   "B"
  ],
  "verdicts": [
   {
    "label": "A",
    "score": 2000,
    "verdict_ko": "세 인물의 의상, 헤어스타일 및 소품(안규홍의 검과 옥가락지, 염재보의 총)이 프롬프트와 레퍼런스에 완벽하게 부합하며, 수레를 끄는 역동적인 동작이 매우 사실적으로 묘사되었습니다."
   },
   {
    "label": "B",
    "score": 1250,
    "verdict_ko": "오른쪽 인물(염재보)이 레퍼런스의 어두운 누비옷이 아닌 밝은 상의를 입고 있어 캐릭터 의상 일관성에서 크게 벗어났습니다."
   }
  ],
  "refs": [
   {
    "label": "PREVIOUS SHOT STILL — a visually related earlier shot of this same place: the location's look, materials, fixed features and lighting mood are LOCKED to this photo; never copy its camera framing. The people visible in that photo are NOT in this shot: never carry their faces, bodies or clothing onto anyone here. Each person in THIS shot is defined solely by the PEOPLE list, their own CHARACTER REFERENCE images and their wardrobe notes. If this photo shows a character who cannot move (dead or unconscious), that character's exact pose, position and orientation are ALSO LOCKED — treat their body as a fixed prop of the set that only the camera moves around.",
    "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/db00517b-165b-4005-a287-f28936b59b95/images/272bd7af-9294-4490-98c6-88a40e9deacc/scene/recipe/S35sh8_sel.png",
    "asset_id": "f7a890c0-d487-436f-a82d-3054b937f7f9",
    "role": "prev_still"
   },
   {
    "label": "CHARACTER REFERENCE — 안규홍: the exact person appearing in this shot; match face, hair and build exactly.",
    "path": "<bytes:1028384>",
    "asset_id": "3b5894db-ce55-4f34-b966-5e725c3aa952",
    "role": "character_ref"
   },
   {
    "label": "CHARACTER REFERENCE — 이관회: the exact person appearing in this shot; match face, hair and build exactly.",
    "path": "<bytes:933540>",
    "asset_id": "851b450f-ac84-47fe-ac5a-03a56f63fcae",
    "role": "character_ref"
   },
   {
    "label": "CHARACTER REFERENCE — 염재보: the exact person appearing in this shot; match face, hair and build exactly.",
    "path": "<bytes:964351>",
    "asset_id": "ab287289-d0b7-436d-8f70-68452f898f73",
    "role": "character_ref"
   }
  ],
  "critique_skipped": true,
  "shot_run_produced": false,
  "shot_run_uid": "06aa1428-4d7b-7d8b-affc-e5fc8e0c823b",
  "ref_mode": "prev+엔티티",
  "share_plan": {
   "ref_plan": "prev",
   "prev_anchor_tag": "S35sh8"
  }
 },
 "S35sh12::cine": {
  "applied": true,
  "attempted_at": "2026-09-09T04:45:46.226309+00:00",
  "fingerprint": "9e5d5d986e9543743de5b1499aa27e80068325f64a3734f057a2258b69a7a396",
  "fingerprint_version": 2,
  "provider": "mai",
  "endpoint": "openrouter/chat-completions",
  "model": "microsoft/mai-image-2.6",
  "pack": "24.202608252115",
  "source_file": "S35sh12_sel.png",
  "source_sha256": "c515673b4cb8ca2473ce0a7b696be2ea37d5a4339f94e9ae51ad16faa236b431",
  "file": "S35sh12_cine.png",
  "staged_sha256": "08af9961a487aacc021cbba428453ee66b2f5c3942743e6225f6f8d04f0856f6",
  "latency_ms": 34840
 },
 "S35sh13::signage": {
  "fp": "648d99f3fd24f099",
  "inscriptions": [],
  "cues": [],
  "dropped": []
 },
 "S35sh13::bgfirst_bg": {
  "input_fingerprint": "ae12e1eecc61c535",
  "prompt": "Create the EMPTY BACKGROUND PLATE for one film shot — NO PEOPLE, no figures, no body parts, no sketch lines, no arrows anywhere.\n\"Empty\" means no people only: KEEP the location's inherent occupants and stock that define the place — animals in an animal shelter, pen or farm, goods in a market, moored boats in a harbour — unless the shot text explicitly removes them.\nThe FIRST attached image is a thin-line storyboard sketch: use ONLY its camera angle, horizon, perspective and the placement/size of buildings and set masses — ignore the sketched people and arrows entirely. The SECOND attached image (LOCATION PHOTOGRAPH) is the real place: take its architecture, materials, signage and fixed features, and RE-PROJECT them into the sketch's camera. If the photograph's camera differs from the sketch's, the sketch's camera wins.\nHUMAN-SCALE CALIBRATION: derive every structure's true size from human-scale elements — a door ≈ 2m, a window ≈ 1–1.5m wide, one storey ≈ 2.5–3m; never inflate a small structure or shrink a large one.\n\nSHOT TEXT this background must serve (Korean): 골목 끝 어둠 속, 저만치 떨어진 수레를 향해 시선을 고정한 채 뚫어져라 응시하는 박진성과 박상현의 굳은 전신.\n\nLOCATION (lock): The dark end of a village alley, with a view toward the departing cart and the estate's closing main gate.\n\nTIME OF DAY (lock): night.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: wide shot\n- FRAME LAYOUT: Distant departing cart beyond the observers in the upper-right of the frame, background.\n- KEY BACKGROUND ELEMENTS: Alley end (Occupied by the two observers) — Opens onto the departing cart's route beyond their shoulders; used as Establish their separate observation position without reversing the departure direction; Distant loaded cart (Moving away with the accumulated provisions) — Rear load visible beyond the observers; the men pulling it are occluded; used as Supply the small, distant endpoint of both men's attention.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Keep the established nighttime exposure and naturalistic contrast unchanged, allowing the observers' stillness to carry the sober tension.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nRender ONE photorealistic empty location photograph, 16:9, that this shot can be staged inside later. No readable writing anywhere: surfaces that would carry writing may be present, but stage any wording out of legibility — an oblique angle, distance, shallow focus. No captions, watermarks or overlay text.",
  "effective_prompt": "Create the EMPTY BACKGROUND PLATE for one film shot — NO PEOPLE, no figures, no body parts, no sketch lines, no arrows anywhere.\n\"Empty\" means no people only: KEEP the location's inherent occupants and stock that define the place — animals in an animal shelter, pen or farm, goods in a market, moored boats in a harbour — unless the shot text explicitly removes them.\nThe FIRST attached image is a thin-line storyboard sketch: use ONLY its camera angle, horizon, perspective and the placement/size of buildings and set masses — ignore the sketched people and arrows entirely. The SECOND attached image (LOCATION PHOTOGRAPH) is the real place: take its architecture, materials, signage and fixed features, and RE-PROJECT them into the sketch's camera. If the photograph's camera differs from the sketch's, the sketch's camera wins.\nHUMAN-SCALE CALIBRATION: derive every structure's true size from human-scale elements — a door ≈ 2m, a window ≈ 1–1.5m wide, one storey ≈ 2.5–3m; never inflate a small structure or shrink a large one.\n\nSHOT TEXT this background must serve (Korean): 골목 끝 어둠 속, 저만치 떨어진 수레를 향해 시선을 고정한 채 뚫어져라 응시하는 박진성과 박상현의 굳은 전신.\n\nLOCATION (lock): The dark end of a village alley, with a view toward the departing cart and the estate's closing main gate.\n\nTIME OF DAY (lock): night.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: wide shot\n- FRAME LAYOUT: Distant departing cart beyond the observers in the upper-right of the frame, background.\n- KEY BACKGROUND ELEMENTS: Alley end (Occupied by the two observers) — Opens onto the departing cart's route beyond their shoulders; used as Establish their separate observation position without reversing the departure direction; Distant loaded cart (Moving away with the accumulated provisions) — Rear load visible beyond the observers; the men pulling it are occluded; used as Supply the small, distant endpoint of both men's attention.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Keep the established nighttime exposure and naturalistic contrast unchanged, allowing the observers' stillness to carry the sober tension.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nRender ONE photorealistic empty location photograph, 16:9, that this shot can be staged inside later. No readable writing anywhere: surfaces that would carry writing may be present, but stage any wording out of legibility — an oblique angle, distance, shallow focus. No captions, watermarks or overlay text.",
  "bg_path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/db00517b-165b-4005-a287-f28936b59b95/images/272bd7af-9294-4490-98c6-88a40e9deacc/scene/recipe/S35sh13__bgfirst_bg.png",
  "asset_id": "92281e31-07d4-445f-bf06-fefef48361a8",
  "input_asset_ids": [
   "d29e1c04-90d1-4ba7-a4b5-8d8c01f4b6de",
   "aeaa845f-49c4-45b0-8a1f-1c8f50c6daed"
  ]
 },
 "S35sh13": {
  "input_fingerprint": "6a68288c45e8bbd2",
  "prompt": "Create ONE FINAL photorealistic live-action film still of the moment below — early 20th-century Korea, Jeolla region; all people are Korean unless explicitly stated otherwise. TIME OF DAY (lock): night.\n\nSHOT TEXT (authoritative, Korean): 골목 끝 어둠 속, 저만치 떨어진 수레를 향해 시선을 고정한 채 뚫어져라 응시하는 박진성과 박상현의 굳은 전신.\n\nLOCATION (lock): The dark end of a village alley, with a view toward the departing cart and the estate's closing main gate. The shot takes place here — the attached LOCATION STRUCTURE PHOTOGRAPH is the single authority for this exact place — its fixed structure and permanent site details are LOCKED to it. No separate location photograph exists for this place. Build everything else strictly from the location text above and the shot text; the layout sketch (when attached) governs framing and placement only, and the shot text governs time of day, lighting and action.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: wide shot\n- FRAME LAYOUT: Distant departing cart beyond the observers in the upper-right of the frame, background.\n- KEY BACKGROUND ELEMENTS: Alley end (Occupied by the two observers) — Opens onto the departing cart's route beyond their shoulders; used as Establish their separate observation position without reversing the departure direction; Distant loaded cart (Moving away with the accumulated provisions) — Rear load visible beyond the observers; the men pulling it are occluded; used as Supply the small, distant endpoint of both men's attention.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Keep the established nighttime exposure and naturalistic contrast unchanged, allowing the observers' stillness to carry the sober tension.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The cart retains its enlarged load of grain, potatoes, straw shoes and corn; the house's main gate is closing. 박진성: He stands at the dark lane end. 박상현: He stands at the lane end looking toward the closing gate.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 박진성 (한국인 남성, 성인 나이대, 자연스러운 검은 머리); 박상현 (한국인 남성, 성인 나이대, 자연스러운 검은 머리) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.",
  "roll_prompts": {
   "A": "Stage the shot. The FIRST attached image (SHOT BACKGROUND) is the finished empty background of this shot — keep it EXACTLY: its camera, perspective, architecture, lighting and every fixed feature stay untouched. The SECOND attached image (LAYOUT SKETCH) tells you ONLY where the people go: each sketched person's position, screen size, pose and the gaze/motion arrows. Ignore the sketch's background lines. The CHARACTER REFERENCE photographs show the real people.\nPlace the real people into the background at exactly the sketched positions, sizes and poses, following the arrow directions. No sketch lines or arrows may remain.\n\nCreate ONE FINAL photorealistic live-action film still of the moment below — early 20th-century Korea, Jeolla region; all people are Korean unless explicitly stated otherwise. TIME OF DAY (lock): night.\n\nSHOT TEXT (authoritative, Korean): 골목 끝 어둠 속, 저만치 떨어진 수레를 향해 시선을 고정한 채 뚫어져라 응시하는 박진성과 박상현의 굳은 전신.\n\nLOCATION (lock): The dark end of a village alley, with a view toward the departing cart and the estate's closing main gate. The shot takes place here — the FIRST attached image (SHOT BACKGROUND) is this exact place, already built: its ground, structures, horizon, materials and lighting are the finished truth of this location and must not be redesigned or replaced. No location photograph is attached — read the place from that image alone, and add no scenery, structure, vehicle or fixture that it does not already show. This lock governs the place only; the figures in the shot follow the staging and pose instructions.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: wide shot\n- FRAME LAYOUT: Distant departing cart beyond the observers in the upper-right of the frame, background.\n- KEY BACKGROUND ELEMENTS: Alley end (Occupied by the two observers) — Opens onto the departing cart's route beyond their shoulders; used as Establish their separate observation position without reversing the departure direction; Distant loaded cart (Moving away with the accumulated provisions) — Rear load visible beyond the observers; the men pulling it are occluded; used as Supply the small, distant endpoint of both men's attention.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Keep the established nighttime exposure and naturalistic contrast unchanged, allowing the observers' stillness to carry the sober tension.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The cart retains its enlarged load of grain, potatoes, straw shoes and corn; the house's main gate is closing. 박진성: He stands at the dark lane end. 박상현: He stands at the lane end looking toward the closing gate.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 박진성 (한국인 남성, 성인 나이대, 자연스러운 검은 머리); 박상현 (한국인 남성, 성인 나이대, 자연스러운 검은 머리) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.",
   "B": "Create ONE FINAL photorealistic live-action film still of the moment below — early 20th-century Korea, Jeolla region; all people are Korean unless explicitly stated otherwise. TIME OF DAY (lock): night.\n\nSHOT TEXT (authoritative, Korean): 골목 끝 어둠 속, 저만치 떨어진 수레를 향해 시선을 고정한 채 뚫어져라 응시하는 박진성과 박상현의 굳은 전신.\n\nLOCATION (lock): The dark end of a village alley, with a view toward the departing cart and the estate's closing main gate. The shot takes place here — the attached LOCATION STRUCTURE PHOTOGRAPH is the single authority for this exact place — its fixed structure and permanent site details are LOCKED to it. No separate location photograph exists for this place. Build everything else strictly from the location text above and the shot text; the layout sketch (when attached) governs framing and placement only, and the shot text governs time of day, lighting and action.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: wide shot\n- FRAME LAYOUT: Distant departing cart beyond the observers in the upper-right of the frame, background.\n- KEY BACKGROUND ELEMENTS: Alley end (Occupied by the two observers) — Opens onto the departing cart's route beyond their shoulders; used as Establish their separate observation position without reversing the departure direction; Distant loaded cart (Moving away with the accumulated provisions) — Rear load visible beyond the observers; the men pulling it are occluded; used as Supply the small, distant endpoint of both men's attention.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Keep the established nighttime exposure and naturalistic contrast unchanged, allowing the observers' stillness to carry the sober tension.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The cart retains its enlarged load of grain, potatoes, straw shoes and corn; the house's main gate is closing. 박진성: He stands at the dark lane end. 박상현: He stands at the lane end looking toward the closing gate.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 박진성 (한국인 남성, 성인 나이대, 자연스러운 검은 머리); 박상현 (한국인 남성, 성인 나이대, 자연스러운 검은 머리) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.\n\nCOMPOSITION VARIATION (this candidate only): keep the place, the people and the moment exactly as specified, and keep from the CAMERA & FRAME contract its subject scale, screen placement and key background elements exactly. For this candidate only, its camera angle and height are deliberately left open: choose a camera position distinctly different from the obvious one, as a film director picking a second setup on the same blocking. Stay true to any explicit framing the SHOT TEXT itself dictates."
  },
  "roll_refs": {
   "A": [
    {
     "label": "SHOT BACKGROUND",
     "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/db00517b-165b-4005-a287-f28936b59b95/images/272bd7af-9294-4490-98c6-88a40e9deacc/scene/recipe/S35sh13__bgfirst_bg.png",
     "asset_id": "92281e31-07d4-445f-bf06-fefef48361a8",
     "role": "bgfirst_bg"
    },
    {
     "label": "LAYOUT SKETCH (people placement only)",
     "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/db00517b-165b-4005-a287-f28936b59b95/images/272bd7af-9294-4490-98c6-88a40e9deacc/conti/conti_S35sh13.png",
     "asset_id": "d29e1c04-90d1-4ba7-a4b5-8d8c01f4b6de",
     "role": "conti_light"
    },
    {
     "label": "CHARACTER REFERENCE — 박진성: the exact person appearing in this shot; match face, hair and build exactly.",
     "path": "<bytes:912297>",
     "asset_id": "3690b9ae-055b-43ee-9b74-794f83fca3d5",
     "role": "character_ref"
    },
    {
     "label": "CHARACTER REFERENCE — 박상현: the exact person appearing in this shot; match face, hair and build exactly.",
     "path": "<bytes:840781>",
     "asset_id": "6187751d-5df3-4eb7-9d30-d6fc98a005e6",
     "role": "character_ref"
    }
   ],
   "B": [
    {
     "label": "LOCATION STRUCTURE PHOTOGRAPH — the confirmed photograph of this exact place and its fixed structure: it is the SINGLE authority for the location, the structure's shape, proportions, materials, colors, openings and every permanent site detail. Never copy its camera framing, time of day or lighting — the shot text is the authority for those.",
     "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/db00517b-165b-4005-a287-f28936b59b95/images/background_chain/seed_bg_village_lanes_sel.png",
     "asset_id": "aeaa845f-49c4-45b0-8a1f-1c8f50c6daed",
     "role": "location_seed_bg"
    },
    {
     "label": "CHARACTER REFERENCE — 박진성: the exact person appearing in this shot; match face, hair and build exactly.",
     "path": "<bytes:912297>",
     "asset_id": "3690b9ae-055b-43ee-9b74-794f83fca3d5",
     "role": "character_ref"
    },
    {
     "label": "CHARACTER REFERENCE — 박상현: the exact person appearing in this shot; match face, hair and build exactly.",
     "path": "<bytes:840781>",
     "asset_id": "6187751d-5df3-4eb7-9d30-d6fc98a005e6",
     "role": "character_ref"
    }
   ]
  },
  "shot_run_spend_attempt_count": 1,
  "judge_flip_skipped": "judge_fn_owns_order",
  "readings": [
   {
    "label": "A",
    "direction": "두 인물 모두 골목 안쪽 멀리 있는 수레를 향해 시선을 고정하고 응시하고 있음.",
    "built_space": "제공된 로케이션 레퍼런스와 동일한 구조의 골목길. 밤 시간대의 어두운 조명이 잘 반영되었으며, 먼 쪽 대문은 어둠 속에 닫혀 있는 것으로 보임.",
    "entities": "박진성(하늘색 한복)과 박상현(회색 한복)이 올바른 착장으로 등장함. 다만 멀리 있는 수레에는 짚더미만 보일 뿐, 프롬프트에서 명시한 곡식, 짚신 등의 짐이 누락됨.",
    "hard_violations": [
     "[gpt-high] 수레 아래에 제삼자의 다리와 신발이 노출된다. 수레를 끄는 사람의 완전 가림 및 명명된 두 인물 이외의 신체 노출 금지 조건을 위반한다."
    ],
    "physics": "두 사람 모두 땅에 안정적으로 두 발을 디디고 서 있으며, 수레 역시 바퀴가 땅에 정상적으로 닿아 있음."
   },
   {
    "label": "B",
    "direction": "두 인물이 골목 안쪽으로 멀어지는 수레를 향해 시선을 고정하고 있음.",
    "built_space": "레퍼런스와 일치하는 흙벽과 기와가 있는 골목. 야간 배경이며 안쪽 대문에 조명이 있어 문이 닫히는 상황 혹은 동작을 효과적으로 암시함.",
    "entities": "박진성과 박상현이 지정된 의상을 입고 올바르게 위치함. 수레 뒷면에 곡식 포대, 짚더미뿐만 아니라 짚신까지 명확하게 묘사되어 프롬프트의 적재물 디테일을 완벽히 충족함.",
    "hard_violations": [
     "[gpt-high] 수레 오른쪽에 제삼자의 머리와 상체가 보인다. 화면에 허용된 인물은 박진성과 박상현뿐이며 수레를 끄는 사람은 가려져야 한다."
    ],
    "physics": "인물들이 땅에 자연스럽게 서서 체중을 지탱하고 있으며, 수레와 실린 짐들의 배치 및 중력 묘사에도 어긋남이 없음."
   }
  ],
  "cross_model_order": {
   "policy": "cross_model_order_v2_slot0_fwd_slot1_rev_parallel2",
   "slots": [
    {
     "model": "gemini-pro",
     "order": "forward",
     "display_to_canonical": {
      "A": "A",
      "B": "B"
     },
     "raw": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 10,
        "verdict_ko": "프롬프트가 요구한 수레의 구체적인 적재물(짚신, 포대 등)을 매우 정확하게 묘사하였으며, 어두운 골목길과 먼 대문의 묘사까지 세밀하게 살려낸 훌륭한 결과물입니다."
       },
       {
        "label": "A",
        "score": 7,
        "verdict_ko": "지정된 두 인물과 장소의 기본 조건은 충족했으나, 수레에 실린 구체적인 물품(짚신, 곡식 등)의 묘사가 누락되고 짚더미만 표현되어 아쉽습니다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "두 인물 모두 골목 안쪽 멀리 있는 수레를 향해 시선을 고정하고 응시하고 있음.",
        "built_space": "제공된 로케이션 레퍼런스와 동일한 구조의 골목길. 밤 시간대의 어두운 조명이 잘 반영되었으며, 먼 쪽 대문은 어둠 속에 닫혀 있는 것으로 보임.",
        "entities": "박진성(하늘색 한복)과 박상현(회색 한복)이 올바른 착장으로 등장함. 다만 멀리 있는 수레에는 짚더미만 보일 뿐, 프롬프트에서 명시한 곡식, 짚신 등의 짐이 누락됨.",
        "hard_violations": [],
        "physics": "두 사람 모두 땅에 안정적으로 두 발을 디디고 서 있으며, 수레 역시 바퀴가 땅에 정상적으로 닿아 있음."
       },
       {
        "label": "B",
        "direction": "두 인물이 골목 안쪽으로 멀어지는 수레를 향해 시선을 고정하고 있음.",
        "built_space": "레퍼런스와 일치하는 흙벽과 기와가 있는 골목. 야간 배경이며 안쪽 대문에 조명이 있어 문이 닫히는 상황 혹은 동작을 효과적으로 암시함.",
        "entities": "박진성과 박상현이 지정된 의상을 입고 올바르게 위치함. 수레 뒷면에 곡식 포대, 짚더미뿐만 아니라 짚신까지 명확하게 묘사되어 프롬프트의 적재물 디테일을 완벽히 충족함.",
        "hard_violations": [],
        "physics": "인물들이 땅에 자연스럽게 서서 체중을 지탱하고 있으며, 수레와 실린 짐들의 배치 및 중력 묘사에도 어긋남이 없음."
       }
      ],
      "all_candidates_fail": false
     },
     "normalized": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 10,
        "verdict_ko": "프롬프트가 요구한 수레의 구체적인 적재물(짚신, 포대 등)을 매우 정확하게 묘사하였으며, 어두운 골목길과 먼 대문의 묘사까지 세밀하게 살려낸 훌륭한 결과물입니다."
       },
       {
        "label": "A",
        "score": 7,
        "verdict_ko": "지정된 두 인물과 장소의 기본 조건은 충족했으나, 수레에 실린 구체적인 물품(짚신, 곡식 등)의 묘사가 누락되고 짚더미만 표현되어 아쉽습니다."
       }
      ],
      "all_candidates_fail": false,
      "readings": [
       {
        "label": "A",
        "direction": "두 인물 모두 골목 안쪽 멀리 있는 수레를 향해 시선을 고정하고 응시하고 있음.",
        "built_space": "제공된 로케이션 레퍼런스와 동일한 구조의 골목길. 밤 시간대의 어두운 조명이 잘 반영되었으며, 먼 쪽 대문은 어둠 속에 닫혀 있는 것으로 보임.",
        "entities": "박진성(하늘색 한복)과 박상현(회색 한복)이 올바른 착장으로 등장함. 다만 멀리 있는 수레에는 짚더미만 보일 뿐, 프롬프트에서 명시한 곡식, 짚신 등의 짐이 누락됨.",
        "hard_violations": [],
        "physics": "두 사람 모두 땅에 안정적으로 두 발을 디디고 서 있으며, 수레 역시 바퀴가 땅에 정상적으로 닿아 있음."
       },
       {
        "label": "B",
        "direction": "두 인물이 골목 안쪽으로 멀어지는 수레를 향해 시선을 고정하고 있음.",
        "built_space": "레퍼런스와 일치하는 흙벽과 기와가 있는 골목. 야간 배경이며 안쪽 대문에 조명이 있어 문이 닫히는 상황 혹은 동작을 효과적으로 암시함.",
        "entities": "박진성과 박상현이 지정된 의상을 입고 올바르게 위치함. 수레 뒷면에 곡식 포대, 짚더미뿐만 아니라 짚신까지 명확하게 묘사되어 프롬프트의 적재물 디테일을 완벽히 충족함.",
        "hard_violations": [],
        "physics": "인물들이 땅에 자연스럽게 서서 체중을 지탱하고 있으며, 수레와 실린 짐들의 배치 및 중력 묘사에도 어긋남이 없음."
       }
      ]
     },
     "ok": true
    },
    {
     "model": "gpt-high",
     "order": "reverse",
     "display_to_canonical": {
      "A": "B",
      "B": "A"
     },
     "raw": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 3,
        "verdict_ko": "두 사람의 전신과 응시 방향은 맞지만, 수레가 충분히 멀고 작지 않으며 수레 오른쪽에 금지된 제삼자의 머리와 상체가 노출된다."
       },
       {
        "label": "B",
        "score": 4,
        "verdict_ko": "두 관찰자 너머 우측 원경의 작은 수레라는 구도에 더 충실하지만, 수레 아래 제삼자의 다리가 보여 완전 가림 조건을 위반한다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "왼쪽의 청색 도포 남성과 중앙의 회색 도포 남성 모두 고개를 화면 오른쪽 안쪽의 수레와 대문 쪽으로 돌렸다. 수레는 적재함 후면을 카메라에 보이며 골목 안쪽 대문 방향으로 향한다. 관찰 방향과 출발 방향이 뒤집히지는 않았다.",
        "built_space": "왼쪽 전경의 초가지붕 사립문 한 곳, 그 뒤 담 사이의 좁은 출입문 한 곳, 오른쪽 원경의 목조 대문 한 곳이 보인다. 흙과 돌을 쌓고 기와를 얹은 담, 중앙 초가와 툇마루는 장소 참조의 주요 구성을 따른다. 두 사람은 왼쪽 사립문 앞 골목 바닥에 서 있다. 대문은 닫힌 면처럼 보여 닫히는 도중이라는 상태는 분명하지 않다. 수레는 우측 배경에 있지만 요구된 작은 원경 대상보다는 가까운 중경 크기다.",
        "entities": "관찰자는 성인 동아시아계 남성 두 명으로, 청색 도포·검은 갓·흰 바지와 회색 도포·검은 갓·짙은 바지가 각각 인물 참조에 부합한다. 얼굴은 옆면 일부만 보여 정확한 동일인 여부와 검은 머리의 세부는 확인하기 어렵다. 수레 한 대에 자루와 짚 계열 짐, 신발 형태 물건이 실려 있으나 곡물·감자·옥수수를 각각 식별하기는 어렵다. 수레 오른쪽에는 별도 남성의 머리와 푸른 옷의 상체가 노출되어 있다. 밤이며 읽을 수 있는 글자는 없다.",
        "hard_violations": [
         "수레 오른쪽에 제삼자의 머리와 상체가 보인다. 화면에 허용된 인물은 박진성과 박상현뿐이며 수레를 끄는 사람은 가려져야 한다."
        ],
        "physics": "두 관찰자의 신발은 땅에 닿고 몸의 무게도 발 위에 놓여 있다. 수레는 양쪽 바퀴로 지면에 지지되며 자루와 짐은 적재판 위에 얹혀 있다. 수레 옆 인물의 하체와 견인 접점은 가려져 있어 당기는 동작 자체는 확인되지 않지만, 떠 있는 몸이나 지지 없는 짐은 보이지 않는다."
       },
       {
        "label": "B",
        "direction": "두 남성은 등을 카메라 쪽에 두고 고개와 몸을 오른쪽 원경의 수레 및 대문 방향으로 향한다. 눈동자는 보이지 않지만 머리 방향은 요구된 관찰 대상과 일치한다. 수레 후면이 관찰자 쪽으로 보이므로 멀어지는 방향도 맞는다.",
        "built_space": "왼쪽 전경 사립문 한 곳, 중앙 담 사이 좁은 출입문 한 곳, 우측 원경 목조 대문 한 곳이 보이며 초가·툇마루·기와 얹은 돌흙담의 배치가 참조와 대체로 대응한다. 두 남성은 왼쪽 벽과 사립문 곁의 골목 바닥에 서고, 그들 너머로 길이 열린다. 수레가 화면 우측 원경에서 작게 보여 지정된 거리 관계를 A보다 잘 구현한다. 대문은 수레에 일부 가려져 닫히는 순간인지 확인하기 어렵다.",
        "entities": "성인 동아시아계 남성 두 명의 전신이 보인다. 청색 도포와 흰 바지, 회색 도포와 짙은 바지, 각각의 검은 갓과 신발은 참조에 대체로 맞는다. 뒷모습이므로 얼굴 동일성은 판정할 수 없다. 수레 한 대의 적재물은 대부분 큰 짚더미로 읽혀 곡물·감자·짚신·옥수수의 누적 적재는 확인되지 않는다. 수레 적재판 아래 중앙에는 제삼자의 두 다리와 신발이 드러난다. 밤의 노출이며 읽을 수 있는 글자는 없다.",
        "hard_violations": [
         "수레 아래에 제삼자의 다리와 신발이 노출된다. 수레를 끄는 사람의 완전 가림 및 명명된 두 인물 이외의 신체 노출 금지 조건을 위반한다."
        ],
        "physics": "두 관찰자는 양발로 땅을 딛고 있으며 청색 도포 남성의 뒤로 모은 손도 몸에 자연스럽게 연결된다. 수레 바퀴는 지면에 닿고 짚더미는 적재판에 지지된다. 수레 아래 보이는 발도 지면에 놓여 있다. 견인하는 손은 적재물 뒤에 가려져 확인되지 않으며, 명백하게 지지 없이 떠 있는 몸이나 물체는 없다."
       }
      ],
      "all_candidates_fail": true
     },
     "normalized": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 3,
        "verdict_ko": "두 사람의 전신과 응시 방향은 맞지만, 수레가 충분히 멀고 작지 않으며 수레 오른쪽에 금지된 제삼자의 머리와 상체가 노출된다."
       },
       {
        "label": "A",
        "score": 4,
        "verdict_ko": "두 관찰자 너머 우측 원경의 작은 수레라는 구도에 더 충실하지만, 수레 아래 제삼자의 다리가 보여 완전 가림 조건을 위반한다."
       }
      ],
      "all_candidates_fail": true,
      "readings": [
       {
        "label": "B",
        "direction": "왼쪽의 청색 도포 남성과 중앙의 회색 도포 남성 모두 고개를 화면 오른쪽 안쪽의 수레와 대문 쪽으로 돌렸다. 수레는 적재함 후면을 카메라에 보이며 골목 안쪽 대문 방향으로 향한다. 관찰 방향과 출발 방향이 뒤집히지는 않았다.",
        "built_space": "왼쪽 전경의 초가지붕 사립문 한 곳, 그 뒤 담 사이의 좁은 출입문 한 곳, 오른쪽 원경의 목조 대문 한 곳이 보인다. 흙과 돌을 쌓고 기와를 얹은 담, 중앙 초가와 툇마루는 장소 참조의 주요 구성을 따른다. 두 사람은 왼쪽 사립문 앞 골목 바닥에 서 있다. 대문은 닫힌 면처럼 보여 닫히는 도중이라는 상태는 분명하지 않다. 수레는 우측 배경에 있지만 요구된 작은 원경 대상보다는 가까운 중경 크기다.",
        "entities": "관찰자는 성인 동아시아계 남성 두 명으로, 청색 도포·검은 갓·흰 바지와 회색 도포·검은 갓·짙은 바지가 각각 인물 참조에 부합한다. 얼굴은 옆면 일부만 보여 정확한 동일인 여부와 검은 머리의 세부는 확인하기 어렵다. 수레 한 대에 자루와 짚 계열 짐, 신발 형태 물건이 실려 있으나 곡물·감자·옥수수를 각각 식별하기는 어렵다. 수레 오른쪽에는 별도 남성의 머리와 푸른 옷의 상체가 노출되어 있다. 밤이며 읽을 수 있는 글자는 없다.",
        "hard_violations": [
         "수레 오른쪽에 제삼자의 머리와 상체가 보인다. 화면에 허용된 인물은 박진성과 박상현뿐이며 수레를 끄는 사람은 가려져야 한다."
        ],
        "physics": "두 관찰자의 신발은 땅에 닿고 몸의 무게도 발 위에 놓여 있다. 수레는 양쪽 바퀴로 지면에 지지되며 자루와 짐은 적재판 위에 얹혀 있다. 수레 옆 인물의 하체와 견인 접점은 가려져 있어 당기는 동작 자체는 확인되지 않지만, 떠 있는 몸이나 지지 없는 짐은 보이지 않는다."
       },
       {
        "label": "A",
        "direction": "두 남성은 등을 카메라 쪽에 두고 고개와 몸을 오른쪽 원경의 수레 및 대문 방향으로 향한다. 눈동자는 보이지 않지만 머리 방향은 요구된 관찰 대상과 일치한다. 수레 후면이 관찰자 쪽으로 보이므로 멀어지는 방향도 맞는다.",
        "built_space": "왼쪽 전경 사립문 한 곳, 중앙 담 사이 좁은 출입문 한 곳, 우측 원경 목조 대문 한 곳이 보이며 초가·툇마루·기와 얹은 돌흙담의 배치가 참조와 대체로 대응한다. 두 남성은 왼쪽 벽과 사립문 곁의 골목 바닥에 서고, 그들 너머로 길이 열린다. 수레가 화면 우측 원경에서 작게 보여 지정된 거리 관계를 A보다 잘 구현한다. 대문은 수레에 일부 가려져 닫히는 순간인지 확인하기 어렵다.",
        "entities": "성인 동아시아계 남성 두 명의 전신이 보인다. 청색 도포와 흰 바지, 회색 도포와 짙은 바지, 각각의 검은 갓과 신발은 참조에 대체로 맞는다. 뒷모습이므로 얼굴 동일성은 판정할 수 없다. 수레 한 대의 적재물은 대부분 큰 짚더미로 읽혀 곡물·감자·짚신·옥수수의 누적 적재는 확인되지 않는다. 수레 적재판 아래 중앙에는 제삼자의 두 다리와 신발이 드러난다. 밤의 노출이며 읽을 수 있는 글자는 없다.",
        "hard_violations": [
         "수레 아래에 제삼자의 다리와 신발이 노출된다. 수레를 끄는 사람의 완전 가림 및 명명된 두 인물 이외의 신체 노출 금지 조건을 위반한다."
        ],
        "physics": "두 관찰자는 양발로 땅을 딛고 있으며 청색 도포 남성의 뒤로 모은 손도 몸에 자연스럽게 연결된다. 수레 바퀴는 지면에 닿고 짚더미는 적재판에 지지된다. 수레 아래 보이는 발도 지면에 놓여 있다. 견인하는 손은 적재물 뒤에 가려져 확인되지 않으며, 명백하게 지지 없이 떠 있는 몸이나 물체는 없다."
       }
      ]
     },
     "ok": true
    }
   ],
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "slot_winner_match": false,
   "slot_winner": {
    "gemini-pro": "B",
    "gpt-high": "A"
   },
   "route": "cross_slot_combined"
  },
  "dual": {
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "normalized": {
    "A": 1.7,
    "B": 1.75
   },
   "adjusted": {
    "A": 1.45,
    "B": 1.5
   },
   "violations": {
    "B": [
     "[gpt-high] 수레 오른쪽에 제삼자의 머리와 상체가 보인다. 화면에 허용된 인물은 박진성과 박상현뿐이며 수레를 끄는 사람은 가려져야 한다."
    ],
    "A": [
     "[gpt-high] 수레 아래에 제삼자의 다리와 신발이 노출된다. 수레를 끄는 사람의 완전 가림 및 명명된 두 인물 이외의 신체 노출 금지 조건을 위반한다."
    ]
   },
   "per_model_winner": {
    "gemini-pro": "B",
    "gpt-high": "A"
   },
   "agreed": false
  },
  "totals": {
   "B": 1500,
   "A": 1450
  },
  "selected": "B",
  "ranking": [
   "B",
   "A"
  ],
  "verdicts": [
   {
    "label": "B",
    "score": 1500,
    "verdict_ko": "프롬프트가 요구한 수레의 구체적인 적재물(짚신, 포대 등)을 매우 정확하게 묘사하였으며, 어두운 골목길과 먼 대문의 묘사까지 세밀하게 살려낸 훌륭한 결과물입니다.  ★위반: [gpt-high] 수레 오른쪽에 제삼자의 머리와 상체가 보인다. 화면에 허용된 인물은 박진성과 박상현뿐이며 수레를 끄는 사람은 가려져야 한다."
   },
   {
    "label": "A",
    "score": 1450,
    "verdict_ko": "지정된 두 인물과 장소의 기본 조건은 충족했으나, 수레에 실린 구체적인 물품(짚신, 곡식 등)의 묘사가 누락되고 짚더미만 표현되어 아쉽습니다.  ★위반: [gpt-high] 수레 아래에 제삼자의 다리와 신발이 노출된다. 수레를 끄는 사람의 완전 가림 및 명명된 두 인물 이외의 신체 노출 금지 조건을 위반한다."
   }
  ],
  "refs": [
   {
    "label": "LOCATION STRUCTURE PHOTOGRAPH — the confirmed photograph of this exact place and its fixed structure: it is the SINGLE authority for the location, the structure's shape, proportions, materials, colors, openings and every permanent site detail. Never copy its camera framing, time of day or lighting — the shot text is the authority for those.",
    "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/db00517b-165b-4005-a287-f28936b59b95/images/background_chain/seed_bg_village_lanes_sel.png",
    "asset_id": "aeaa845f-49c4-45b0-8a1f-1c8f50c6daed",
    "role": "location_seed_bg"
   },
   {
    "label": "CHARACTER REFERENCE — 박진성: the exact person appearing in this shot; match face, hair and build exactly.",
    "path": "<bytes:912297>",
    "asset_id": "3690b9ae-055b-43ee-9b74-794f83fca3d5",
    "role": "character_ref"
   },
   {
    "label": "CHARACTER REFERENCE — 박상현: the exact person appearing in this shot; match face, hair and build exactly.",
    "path": "<bytes:840781>",
    "asset_id": "6187751d-5df3-4eb7-9d30-d6fc98a005e6",
    "role": "character_ref"
   }
  ],
  "critique_skipped": true,
  "shot_run_produced": false,
  "shot_run_uid": "06aa1428-4f20-7aa6-9cb7-b071a8a75a1a",
  "bgfirst": {
   "bg_path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/db00517b-165b-4005-a287-f28936b59b95/images/272bd7af-9294-4490-98c6-88a40e9deacc/scene/recipe/S35sh13__bgfirst_bg.png",
   "bg_asset_id": "92281e31-07d4-445f-bf06-fefef48361a8",
   "bg_record_key": "S35sh13::bgfirst_bg",
   "chain_winner": false,
   "authority": "seed_bg"
  },
  "ref_mode": "seed-bg+엔티티 (2택1: 무콘티 승)",
  "share_plan": {
   "ref_plan": "background"
  },
  "lane_policy": "ab_select_ready"
 },
 "S35sh13::cine": {
  "applied": true,
  "attempted_at": "2026-09-09T04:48:50.464592+00:00",
  "fingerprint": "2e04b90af316fcc7ffa2abc9ec68251dc3bc16803e8b4ef3586961e5a5f3cd51",
  "fingerprint_version": 2,
  "provider": "mai",
  "endpoint": "openrouter/chat-completions",
  "model": "microsoft/mai-image-2.6",
  "pack": "24.202608252115",
  "source_file": "S35sh13_sel.png",
  "source_sha256": "ed4caf094b23f0661e854de924a4511ce48aaa696922ce7fdb98d6f4fea2dd5f",
  "file": "S35sh13_cine.png",
  "staged_sha256": "3ecea5cbb5348f4884e694ba49ada6e65ac604cc085de02617204736181732ed",
  "latency_ms": 27578
 },
 "S36sh1::signage": {
  "fp": "dac799efbc66cd0e",
  "inscriptions": [],
  "cues": [],
  "dropped": []
 },
 "groupbg::camp_supply_approach": {
  "input_fingerprint": "26d698f5b53bbd14",
  "meta": {
   "model": "gpt-image-2",
   "size": "1536x864",
   "pack": "11.202607220237",
   "contract": "bgfirst_full_v3",
   "group_sig": {
    "key": "camp_supply_approach",
    "tags": [
     "S36sh1",
     "S36sh17",
     "S36sh9",
     "S47sh3"
    ]
   },
   "context_sig": "4da89cac14ae19bb"
  },
  "prompt": "Create ONE empty live-action location background photograph — NO PEOPLE, no figures, no body parts, no silhouettes, no shadows or reflections of people anywhere.\n\"Empty\" means no people only: KEEP the location's inherent occupants and stock that define the place — animals in an animal shelter, pen or farm, goods and produce in a market, moored boats in a harbour — unless the shot text explicitly removes them.\nThe FIRST attached image is a thin-line storyboard sketch of a shot that happens at this location: use it ONLY as spatial evidence — what this place contains, how its ground, structures and landmarks are arranged and proportioned. Ignore the sketched people and arrows entirely, and do NOT copy its line style: render a fully photographic, physically plausible real place that fits THE LOCATION text below.\nHUMAN-SCALE CALIBRATION: derive every structure's true size from human-scale elements — a door ≈ 2m, a window ≈ 1–1.5m wide, one storey ≈ 2.5–3m; never inflate a small structure or shrink a large one.\n\nTHE LOCATION — early 20th-century Korea, Jeolla region; all people are Korean unless explicitly stated otherwise: The lower stretch of the uphill forest approach to the mountain camp, where a loaded cart is being hauled upward at night.\n\nLOCATION DETAIL (descriptive evidence about this place from the production's location records — evidence, not a staging order): use it to understand what kind of place this is and what it permanently contains. Carry over ONLY its enduring physical features — layout, structures, ground and materials, fixed equipment and installations. Do NOT treat any lighting, weather or time-of-day wording, momentary object states, depicted events or subjective impressions in it as fixed facts of the place; the TIME OF DAY lock below is the sole authority for time and lighting.\n파청 산채 공터·훈련장·취사장·막사 주변: 산 중턱에 통나무를 세워 만든 조잡한 목책 안의 넓은 흙바닥 공터. 나무와 거적으로 만든 움막과 막사가 흩어져 있고, 돌과 진흙으로 만든 취사용 아궁이와 무쇠 솥, 모닥불 터가 있다. 전투 후에는 솥이 뒤집히고 막사가 불타거나 무너진다. (특징: 통나무를 박아 세운 목책; 넓게 다져진 흙바닥 (공터/훈련장); 돌과 진흙으로 만든 다수의 아궁이와 무쇠솥; 나무와 거적으로 엮어 만든 움막과 막사; 부서진 토기 밥그릇과 뒤집힌 무쇠솥 (전투 후); 불에 탄 나무 잔해와 흩어진 쌀밥 (전투 후))\n\nSCENE EVIDENCE (verbatim quotes from the screenplay about this place — treat them as evidence of what the location physically contains and looks like; stage the PLACE those moments happen in, but do NOT depict the momentary actions, people or staged props themselves):\n- 산길 아래.\n규홍, 관회, 재보가 무거워진 수레를 끌고 올라온다.\n- 산채의 사내들이 우르르 뛰어나온다.\n- 규홍은 그런 시선을 모른 채 다시 수레로 향한다.\n- 산 아래서 인기척. 나뭇가지 스치는 소리, 발소리.\n- 박상현이 짐꾼 둘과 함께 올라온다.\n\nTIME OF DAY (lock): night.\n\nRender ONE photorealistic empty location photograph, 16:9, neutral enough that every shot of this place can be staged from it later. No readable writing anywhere: surfaces that would carry writing may be present, but stage any wording out of legibility — an oblique angle, distance, shallow focus. No captions, watermarks or overlay text.",
  "effective_prompt": "Create ONE empty live-action location background photograph — NO PEOPLE, no figures, no body parts, no silhouettes, no shadows or reflections of people anywhere.\n\"Empty\" means no people only: KEEP the location's inherent occupants and stock that define the place — animals in an animal shelter, pen or farm, goods and produce in a market, moored boats in a harbour — unless the shot text explicitly removes them.\nThe FIRST attached image is a thin-line storyboard sketch of a shot that happens at this location: use it ONLY as spatial evidence — what this place contains, how its ground, structures and landmarks are arranged and proportioned. Ignore the sketched people and arrows entirely, and do NOT copy its line style: render a fully photographic, physically plausible real place that fits THE LOCATION text below.\nHUMAN-SCALE CALIBRATION: derive every structure's true size from human-scale elements — a door ≈ 2m, a window ≈ 1–1.5m wide, one storey ≈ 2.5–3m; never inflate a small structure or shrink a large one.\n\nTHE LOCATION — early 20th-century Korea, Jeolla region; all people are Korean unless explicitly stated otherwise: The lower stretch of the uphill forest approach to the mountain camp, where a loaded cart is being hauled upward at night.\n\nLOCATION DETAIL (descriptive evidence about this place from the production's location records — evidence, not a staging order): use it to understand what kind of place this is and what it permanently contains. Carry over ONLY its enduring physical features — layout, structures, ground and materials, fixed equipment and installations. Do NOT treat any lighting, weather or time-of-day wording, momentary object states, depicted events or subjective impressions in it as fixed facts of the place; the TIME OF DAY lock below is the sole authority for time and lighting.\n파청 산채 공터·훈련장·취사장·막사 주변: 산 중턱에 통나무를 세워 만든 조잡한 목책 안의 넓은 흙바닥 공터. 나무와 거적으로 만든 움막과 막사가 흩어져 있고, 돌과 진흙으로 만든 취사용 아궁이와 무쇠 솥, 모닥불 터가 있다. 전투 후에는 솥이 뒤집히고 막사가 불타거나 무너진다. (특징: 통나무를 박아 세운 목책; 넓게 다져진 흙바닥 (공터/훈련장); 돌과 진흙으로 만든 다수의 아궁이와 무쇠솥; 나무와 거적으로 엮어 만든 움막과 막사; 부서진 토기 밥그릇과 뒤집힌 무쇠솥 (전투 후); 불에 탄 나무 잔해와 흩어진 쌀밥 (전투 후))\n\nSCENE EVIDENCE (verbatim quotes from the screenplay about this place — treat them as evidence of what the location physically contains and looks like; stage the PLACE those moments happen in, but do NOT depict the momentary actions, people or staged props themselves):\n- 산길 아래.\n규홍, 관회, 재보가 무거워진 수레를 끌고 올라온다.\n- 산채의 사내들이 우르르 뛰어나온다.\n- 규홍은 그런 시선을 모른 채 다시 수레로 향한다.\n- 산 아래서 인기척. 나뭇가지 스치는 소리, 발소리.\n- 박상현이 짐꾼 둘과 함께 올라온다.\n\nTIME OF DAY (lock): night.\n\nRender ONE photorealistic empty location photograph, 16:9, neutral enough that every shot of this place can be staged from it later. No readable writing anywhere: surfaces that would carry writing may be present, but stage any wording out of legibility — an oblique angle, distance, shallow focus. No captions, watermarks or overlay text.",
  "bg_path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/db00517b-165b-4005-a287-f28936b59b95/images/272bd7af-9294-4490-98c6-88a40e9deacc/scene/recipe/groupbg_camp_supply_approach_8601af.png",
  "asset_id": "69bb9661-49a8-4a75-a056-1e48e42c72b5",
  "input_asset_ids": [
   "dd27a023-9a40-4cbd-8596-a4c9b0d59e93"
  ],
  "origin_tag": "S36sh1",
  "place_text": "The lower stretch of the uphill forest approach to the mountain camp, where a loaded cart is being hauled upward at night.",
  "origin_inputs": {
   "place_text": "The lower stretch of the uphill forest approach to the mountain camp, where a loaded cart is being hauled upward at night.",
   "time_of_day_en": "night",
   "conti_asset_id": "dd27a023-9a40-4cbd-8596-a4c9b0d59e93"
  }
 },
 "S36sh1::bgfirst_bg": {
  "input_fingerprint": "21a0c9a25493fcfa",
  "prompt": "Create the EMPTY BACKGROUND PLATE for one film shot — NO PEOPLE, no figures, no body parts, no sketch lines, no arrows anywhere.\n\"Empty\" means no people only: KEEP the location's inherent occupants and stock that define the place — animals in an animal shelter, pen or farm, goods in a market, moored boats in a harbour — unless the shot text explicitly removes them.\nThe FIRST attached image is a thin-line storyboard sketch: use ONLY its camera angle, horizon, perspective and the placement/size of buildings and set masses — ignore the sketched people and arrows entirely. The SECOND attached image (LOCATION PHOTOGRAPH) is the real place: take its architecture, materials, signage and fixed features, and RE-PROJECT them into the sketch's camera. If the photograph's camera differs from the sketch's, the sketch's camera wins.\nHUMAN-SCALE CALIBRATION: derive every structure's true size from human-scale elements — a door ≈ 2m, a window ≈ 1–1.5m wide, one storey ≈ 2.5–3m; never inflate a small structure or shrink a large one.\n\nSHOT TEXT this background must serve (Korean): 산길 아래, 무거운 수레 끈을 어깨에 걸고 한 발을 내디딘 mid-stride 자세의 안규홍, 이관회, 염재보의 땀 맺힌 전신.\n\nLOCATION (lock): The lower stretch of the uphill forest approach to the mountain camp, where a loaded cart is being hauled upward at night.\n\nTIME OF DAY (lock): night.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: wide shot\n- KEY BACKGROUND ELEMENTS: Uphill mountain path (Being climbed by the three men pulling the cart) — Runs from behind the cart toward the camera's uphill position; used as Make the incline readable beneath the full-body action; Loaded cart (Piled with rice sacks, barley, potatoes and straw shoes) — Near side and forward end visible behind the men; used as Provide a proportional background measure of the burden; Cart ropes (Across the men's shoulders during the uphill pull) — Extend backward from the men's shoulders to the cart; used as Visually transmit the load through each man's posture.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Use restrained nighttime exposure with naturalistic contrast, preserving the supplied sweat detail without specifying an unsupported light source.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nRender ONE photorealistic empty location photograph, 16:9, that this shot can be staged inside later. No readable writing anywhere: surfaces that would carry writing may be present, but stage any wording out of legibility — an oblique angle, distance, shallow focus. No captions, watermarks or overlay text.",
  "effective_prompt": "Create the EMPTY BACKGROUND PLATE for one film shot — NO PEOPLE, no figures, no body parts, no sketch lines, no arrows anywhere.\n\"Empty\" means no people only: KEEP the location's inherent occupants and stock that define the place — animals in an animal shelter, pen or farm, goods in a market, moored boats in a harbour — unless the shot text explicitly removes them.\nThe FIRST attached image is a thin-line storyboard sketch: use ONLY its camera angle, horizon, perspective and the placement/size of buildings and set masses — ignore the sketched people and arrows entirely. The SECOND attached image (LOCATION PHOTOGRAPH) is the real place: take its architecture, materials, signage and fixed features, and RE-PROJECT them into the sketch's camera. If the photograph's camera differs from the sketch's, the sketch's camera wins.\nHUMAN-SCALE CALIBRATION: derive every structure's true size from human-scale elements — a door ≈ 2m, a window ≈ 1–1.5m wide, one storey ≈ 2.5–3m; never inflate a small structure or shrink a large one.\n\nSHOT TEXT this background must serve (Korean): 산길 아래, 무거운 수레 끈을 어깨에 걸고 한 발을 내디딘 mid-stride 자세의 안규홍, 이관회, 염재보의 땀 맺힌 전신.\n\nLOCATION (lock): The lower stretch of the uphill forest approach to the mountain camp, where a loaded cart is being hauled upward at night.\n\nTIME OF DAY (lock): night.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: wide shot\n- KEY BACKGROUND ELEMENTS: Uphill mountain path (Being climbed by the three men pulling the cart) — Runs from behind the cart toward the camera's uphill position; used as Make the incline readable beneath the full-body action; Loaded cart (Piled with rice sacks, barley, potatoes and straw shoes) — Near side and forward end visible behind the men; used as Provide a proportional background measure of the burden; Cart ropes (Across the men's shoulders during the uphill pull) — Extend backward from the men's shoulders to the cart; used as Visually transmit the load through each man's posture.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Use restrained nighttime exposure with naturalistic contrast, preserving the supplied sweat detail without specifying an unsupported light source.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nRender ONE photorealistic empty location photograph, 16:9, that this shot can be staged inside later. No readable writing anywhere: surfaces that would carry writing may be present, but stage any wording out of legibility — an oblique angle, distance, shallow focus. No captions, watermarks or overlay text.",
  "bg_path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/db00517b-165b-4005-a287-f28936b59b95/images/272bd7af-9294-4490-98c6-88a40e9deacc/scene/recipe/S36sh1__bgfirst_bg.png",
  "asset_id": "918e7565-903c-4cbe-aa25-7d6133afa09e",
  "input_asset_ids": [
   "dd27a023-9a40-4cbd-8596-a4c9b0d59e93",
   "69bb9661-49a8-4a75-a056-1e48e42c72b5"
  ]
 },
 "S36sh1": {
  "input_fingerprint": "49ee174d60e8c5c7",
  "prompt": "Create ONE FINAL photorealistic live-action film still of the moment below — early 20th-century Korea, Jeolla region; all people are Korean unless explicitly stated otherwise. TIME OF DAY (lock): night.\n\nSHOT TEXT (authoritative, Korean): 산길 아래, 무거운 수레 끈을 어깨에 걸고 한 발을 내디딘 mid-stride 자세의 안규홍, 이관회, 염재보의 땀 맺힌 전신.\n\nLOCATION (lock): The lower stretch of the uphill forest approach to the mountain camp, where a loaded cart is being hauled upward at night. The shot takes place here — the attached LOCATION PHOTOGRAPH shows the exact spot.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: wide shot\n- KEY BACKGROUND ELEMENTS: Uphill mountain path (Being climbed by the three men pulling the cart) — Runs from behind the cart toward the camera's uphill position; used as Make the incline readable beneath the full-body action; Loaded cart (Piled with rice sacks, barley, potatoes and straw shoes) — Near side and forward end visible behind the men; used as Provide a proportional background measure of the burden; Cart ropes (Across the men's shoulders during the uphill pull) — Extend backward from the men's shoulders to the cart; used as Visually transmit the load through each man's posture.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Use restrained nighttime exposure with naturalistic contrast, preserving the supplied sweat detail without specifying an unsupported light source.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The cart arrives loaded with rice bales, barley, potatoes, straw shoes and corn on the dark mountain approach. 안규홍: He strains uphill with the loaded cart, retaining his sword and concealed jade ring. 이관회: He strains uphill with the loaded cart. 염재보: He strains uphill with the loaded cart and retains his gun.\n\nTHE HAND THAT IS DOING THIS: the object at the centre of this shot is being held, operated, read, aimed or handed over by 안규홍, 이관회, 염재보 right now, so 안규홍, 이관회, 염재보's hand — and as much of the wrist and forearm as the framing reaches — is in the frame, gripping or resting on that object exactly the way its real use requires. Match that hand to 안규홍, 이관회, 염재보: its size, build, skin, age, grooming, sleeve and anything worn on it belong to that person and to no one else.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 안규홍 (한국인 남성, 청년 나이대, 검은 머리, 단단히 묶은 머리); 이관회 (한국인 남성, 청년 나이대, 자연스러운 검은 머리); 염재보 (한국인 남성, 성인 나이대, 자연스러운 검은 머리) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.",
  "roll_prompts": {
   "A": "Stage the shot. The FIRST attached image (SHOT BACKGROUND) is the finished empty background of this shot — keep it EXACTLY: its camera, perspective, architecture, lighting and every fixed feature stay untouched. The SECOND attached image (LAYOUT SKETCH) tells you ONLY where the people go: each sketched person's position, screen size, pose and the gaze/motion arrows. Ignore the sketch's background lines. The CHARACTER REFERENCE photographs show the real people.\nPlace the real people into the background at exactly the sketched positions, sizes and poses, following the arrow directions. No sketch lines or arrows may remain.\n\nCreate ONE FINAL photorealistic live-action film still of the moment below — early 20th-century Korea, Jeolla region; all people are Korean unless explicitly stated otherwise. TIME OF DAY (lock): night.\n\nSHOT TEXT (authoritative, Korean): 산길 아래, 무거운 수레 끈을 어깨에 걸고 한 발을 내디딘 mid-stride 자세의 안규홍, 이관회, 염재보의 땀 맺힌 전신.\n\nLOCATION (lock): The lower stretch of the uphill forest approach to the mountain camp, where a loaded cart is being hauled upward at night. The shot takes place here — the FIRST attached image (SHOT BACKGROUND) is this exact place, already built: its ground, structures, horizon, materials and lighting are the finished truth of this location and must not be redesigned or replaced. No location photograph is attached — read the place from that image alone, and add no scenery, structure, vehicle or fixture that it does not already show. This lock governs the place only; the figures in the shot follow the staging and pose instructions.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: wide shot\n- KEY BACKGROUND ELEMENTS: Uphill mountain path (Being climbed by the three men pulling the cart) — Runs from behind the cart toward the camera's uphill position; used as Make the incline readable beneath the full-body action; Loaded cart (Piled with rice sacks, barley, potatoes and straw shoes) — Near side and forward end visible behind the men; used as Provide a proportional background measure of the burden; Cart ropes (Across the men's shoulders during the uphill pull) — Extend backward from the men's shoulders to the cart; used as Visually transmit the load through each man's posture.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Use restrained nighttime exposure with naturalistic contrast, preserving the supplied sweat detail without specifying an unsupported light source.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The cart arrives loaded with rice bales, barley, potatoes, straw shoes and corn on the dark mountain approach. 안규홍: He strains uphill with the loaded cart, retaining his sword and concealed jade ring. 이관회: He strains uphill with the loaded cart. 염재보: He strains uphill with the loaded cart and retains his gun.\n\nTHE HAND THAT IS DOING THIS: the object at the centre of this shot is being held, operated, read, aimed or handed over by 안규홍, 이관회, 염재보 right now, so 안규홍, 이관회, 염재보's hand — and as much of the wrist and forearm as the framing reaches — is in the frame, gripping or resting on that object exactly the way its real use requires. Match that hand to 안규홍, 이관회, 염재보: its size, build, skin, age, grooming, sleeve and anything worn on it belong to that person and to no one else.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 안규홍 (한국인 남성, 청년 나이대, 검은 머리, 단단히 묶은 머리); 이관회 (한국인 남성, 청년 나이대, 자연스러운 검은 머리); 염재보 (한국인 남성, 성인 나이대, 자연스러운 검은 머리) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.",
   "B": "Create ONE FINAL photorealistic live-action film still of the moment below — early 20th-century Korea, Jeolla region; all people are Korean unless explicitly stated otherwise. TIME OF DAY (lock): night.\n\nSHOT TEXT (authoritative, Korean): 산길 아래, 무거운 수레 끈을 어깨에 걸고 한 발을 내디딘 mid-stride 자세의 안규홍, 이관회, 염재보의 땀 맺힌 전신.\n\nLOCATION (lock): The lower stretch of the uphill forest approach to the mountain camp, where a loaded cart is being hauled upward at night. The shot takes place here — the attached LOCATION PHOTOGRAPH shows the exact spot.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: wide shot\n- KEY BACKGROUND ELEMENTS: Uphill mountain path (Being climbed by the three men pulling the cart) — Runs from behind the cart toward the camera's uphill position; used as Make the incline readable beneath the full-body action; Loaded cart (Piled with rice sacks, barley, potatoes and straw shoes) — Near side and forward end visible behind the men; used as Provide a proportional background measure of the burden; Cart ropes (Across the men's shoulders during the uphill pull) — Extend backward from the men's shoulders to the cart; used as Visually transmit the load through each man's posture.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Use restrained nighttime exposure with naturalistic contrast, preserving the supplied sweat detail without specifying an unsupported light source.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The cart arrives loaded with rice bales, barley, potatoes, straw shoes and corn on the dark mountain approach. 안규홍: He strains uphill with the loaded cart, retaining his sword and concealed jade ring. 이관회: He strains uphill with the loaded cart. 염재보: He strains uphill with the loaded cart and retains his gun.\n\nTHE HAND THAT IS DOING THIS: the object at the centre of this shot is being held, operated, read, aimed or handed over by 안규홍, 이관회, 염재보 right now, so 안규홍, 이관회, 염재보's hand — and as much of the wrist and forearm as the framing reaches — is in the frame, gripping or resting on that object exactly the way its real use requires. Match that hand to 안규홍, 이관회, 염재보: its size, build, skin, age, grooming, sleeve and anything worn on it belong to that person and to no one else.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 안규홍 (한국인 남성, 청년 나이대, 검은 머리, 단단히 묶은 머리); 이관회 (한국인 남성, 청년 나이대, 자연스러운 검은 머리); 염재보 (한국인 남성, 성인 나이대, 자연스러운 검은 머리) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.\n\nCOMPOSITION VARIATION (this candidate only): keep the place, the people and the moment exactly as specified, and keep from the CAMERA & FRAME contract its subject scale, screen placement and key background elements exactly. For this candidate only, its camera angle and height are deliberately left open: choose a camera position distinctly different from the obvious one, as a film director picking a second setup on the same blocking. Stay true to any explicit framing the SHOT TEXT itself dictates."
  },
  "roll_refs": {
   "A": [
    {
     "label": "SHOT BACKGROUND",
     "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/db00517b-165b-4005-a287-f28936b59b95/images/272bd7af-9294-4490-98c6-88a40e9deacc/scene/recipe/S36sh1__bgfirst_bg.png",
     "asset_id": "918e7565-903c-4cbe-aa25-7d6133afa09e",
     "role": "bgfirst_bg"
    },
    {
     "label": "LAYOUT SKETCH (people placement only)",
     "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/db00517b-165b-4005-a287-f28936b59b95/images/272bd7af-9294-4490-98c6-88a40e9deacc/conti/conti_S36sh1.png",
     "asset_id": "dd27a023-9a40-4cbd-8596-a4c9b0d59e93",
     "role": "conti_light"
    },
    {
     "label": "CHARACTER REFERENCE — 안규홍: the exact person appearing in this shot; match face, hair and build exactly.",
     "path": "<bytes:1028384>",
     "asset_id": "3b5894db-ce55-4f34-b966-5e725c3aa952",
     "role": "character_ref"
    },
    {
     "label": "CHARACTER REFERENCE — 이관회: the exact person appearing in this shot; match face, hair and build exactly.",
     "path": "<bytes:933540>",
     "asset_id": "851b450f-ac84-47fe-ac5a-03a56f63fcae",
     "role": "character_ref"
    },
    {
     "label": "CHARACTER REFERENCE — 염재보: the exact person appearing in this shot; match face, hair and build exactly.",
     "path": "<bytes:964351>",
     "asset_id": "ab287289-d0b7-436d-8f70-68452f898f73",
     "role": "character_ref"
    }
   ],
   "B": [
    {
     "label": "LOCATION PHOTOGRAPH — the exact place of this shot: its architecture, materials, fixed features and lighting mood are spatial truth; stage the moment inside this place. Never copy its camera framing.",
     "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/db00517b-165b-4005-a287-f28936b59b95/images/272bd7af-9294-4490-98c6-88a40e9deacc/scene/recipe/groupbg_camp_supply_approach_8601af.png",
     "asset_id": "69bb9661-49a8-4a75-a056-1e48e42c72b5",
     "role": "bgfirst_group_bg"
    },
    {
     "label": "CHARACTER REFERENCE — 안규홍: the exact person appearing in this shot; match face, hair and build exactly.",
     "path": "<bytes:1028384>",
     "asset_id": "3b5894db-ce55-4f34-b966-5e725c3aa952",
     "role": "character_ref"
    },
    {
     "label": "CHARACTER REFERENCE — 이관회: the exact person appearing in this shot; match face, hair and build exactly.",
     "path": "<bytes:933540>",
     "asset_id": "851b450f-ac84-47fe-ac5a-03a56f63fcae",
     "role": "character_ref"
    },
    {
     "label": "CHARACTER REFERENCE — 염재보: the exact person appearing in this shot; match face, hair and build exactly.",
     "path": "<bytes:964351>",
     "asset_id": "ab287289-d0b7-436d-8f70-68452f898f73",
     "role": "character_ref"
    }
   ]
  },
  "shot_run_spend_attempt_count": 1,
  "judge_flip_skipped": "judge_fn_owns_order",
  "readings": [
   {
    "label": "A",
    "direction": "세 인물 모두 화면 좌측(오르막 캠프 방향)을 향해 걷고 있으며 시선 역시 진행 방향을 자연스럽게 향하고 있음.",
    "built_space": "야간의 산길 배경으로 좌측 상단에 캠프의 목책과 구조물이 위치함. 인물들이 캠프 방향으로 수레를 당기고 있어 지형과 이동 방향의 호환성이 올바름.",
    "entities": "이관회(좌측), 안규홍(중앙), 염재보(우측) 모두 지정된 레퍼런스의 복장과 외형에 일치함. 염재보는 모자를 착용하고 왼손에 총을 쥐고 있음. 짐이 실린 수레가 인물들 뒤에 위치함.",
    "hard_violations": [
     "[gemini-pro] 객체 중복: 안규홍이 허리에 녹색 술이 달린 검을 차고 있음에도 불구하고, 왼손에 또 다른 검(칼집)을 들고 있어 하나의 검이 중복 생성됨.",
     "[gpt-high] 수레가 있는 높은 뒤쪽에서 인물들이 낮은 전경으로 내려오는 경사 배치로, 명시된 수레 뒤쪽에서 카메라 쪽으로 올라오는 경로와 오르막 견인 방향을 반대로 구현했다."
    ],
    "physics": "세 명 모두 어깨에 밧줄을 걸고 한 발을 내디딘 자세(mid-stride)를 잘 유지하고 있으며 발끝이 지면을 단단히 딛고 있음. 밧줄은 대체로 수레를 향해 뒤쪽으로 당겨지는 장력을 보여줌."
   },
   {
    "label": "B",
    "direction": "세 인물 모두 화면 우측을 향해 이동하고 있으며, 이는 배경 구조상 캠프에서 멀어지는 하산(내리막) 방향임.",
    "built_space": "산길과 좌측 상단의 캠프가 확인되나, 인물들의 위치와 이동 방향이 수레를 이끌고 산 아래로 향하는 형태가 되어 오르막 지시와 모순됨.",
    "entities": "안규홍과 이관회의 복장 및 인상착의는 레퍼런스와 일치하나, 앞장서는 염재보(우측)는 레퍼런스의 모자를 쓰지 않았고 명시된 총도 소지하지 않음.",
    "hard_violations": [
     "[gemini-pro] 구조물 방향 위반: 오르막길(uphill)로 수레를 끌어올려야 한다는 프롬프트 지시와 반대로 우측 내리막 방향으로 이동함.",
     "[gemini-pro] 물리적 지지 위반: 이관회와 염재보의 어깨에 걸린 밧줄이 뒤의 수레와 연결되지 않은 채 허공에서 끊어지거나 아래로 늘어져 아무것도 당기지 않고 있음."
    ],
    "physics": "걷는 자세는 나타나 있으나, 수레를 끄는 핵심 매개체인 밧줄이 물리적으로 연결되지 않아 동작과 하중의 인과관계가 성립하지 않음."
   }
  ],
  "cross_model_order": {
   "policy": "cross_model_order_v2_slot0_fwd_slot1_rev_parallel2",
   "slots": [
    {
     "model": "gemini-pro",
     "order": "forward",
     "display_to_canonical": {
      "A": "A",
      "B": "B"
     },
     "raw": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 5,
        "verdict_ko": "오르막 방향과 인물들의 외형 및 소품(염재보의 총과 모자)은 훌륭하게 구현되었으나, 안규홍의 검이 허리와 손에 두 개로 중복 생성된 물리적 오류로 인해 최종 불합격입니다."
       },
       {
        "label": "B",
        "score": 2,
        "verdict_ko": "명시된 오르막과 반대인 내리막으로 이동하고 있으며, 밧줄이 끊어지고 염재보의 주요 소품이 누락되는 등 핵심 연출 지시를 다수 위반했습니다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "세 인물 모두 화면 좌측(오르막 캠프 방향)을 향해 걷고 있으며 시선 역시 진행 방향을 자연스럽게 향하고 있음.",
        "built_space": "야간의 산길 배경으로 좌측 상단에 캠프의 목책과 구조물이 위치함. 인물들이 캠프 방향으로 수레를 당기고 있어 지형과 이동 방향의 호환성이 올바름.",
        "entities": "이관회(좌측), 안규홍(중앙), 염재보(우측) 모두 지정된 레퍼런스의 복장과 외형에 일치함. 염재보는 모자를 착용하고 왼손에 총을 쥐고 있음. 짐이 실린 수레가 인물들 뒤에 위치함.",
        "hard_violations": [
         "객체 중복: 안규홍이 허리에 녹색 술이 달린 검을 차고 있음에도 불구하고, 왼손에 또 다른 검(칼집)을 들고 있어 하나의 검이 중복 생성됨."
        ],
        "physics": "세 명 모두 어깨에 밧줄을 걸고 한 발을 내디딘 자세(mid-stride)를 잘 유지하고 있으며 발끝이 지면을 단단히 딛고 있음. 밧줄은 대체로 수레를 향해 뒤쪽으로 당겨지는 장력을 보여줌."
       },
       {
        "label": "B",
        "direction": "세 인물 모두 화면 우측을 향해 이동하고 있으며, 이는 배경 구조상 캠프에서 멀어지는 하산(내리막) 방향임.",
        "built_space": "산길과 좌측 상단의 캠프가 확인되나, 인물들의 위치와 이동 방향이 수레를 이끌고 산 아래로 향하는 형태가 되어 오르막 지시와 모순됨.",
        "entities": "안규홍과 이관회의 복장 및 인상착의는 레퍼런스와 일치하나, 앞장서는 염재보(우측)는 레퍼런스의 모자를 쓰지 않았고 명시된 총도 소지하지 않음.",
        "hard_violations": [
         "구조물 방향 위반: 오르막길(uphill)로 수레를 끌어올려야 한다는 프롬프트 지시와 반대로 우측 내리막 방향으로 이동함.",
         "물리적 지지 위반: 이관회와 염재보의 어깨에 걸린 밧줄이 뒤의 수레와 연결되지 않은 채 허공에서 끊어지거나 아래로 늘어져 아무것도 당기지 않고 있음."
        ],
        "physics": "걷는 자세는 나타나 있으나, 수레를 끄는 핵심 매개체인 밧줄이 물리적으로 연결되지 않아 동작과 하중의 인과관계가 성립하지 않음."
       }
      ],
      "all_candidates_fail": true
     },
     "normalized": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 5,
        "verdict_ko": "오르막 방향과 인물들의 외형 및 소품(염재보의 총과 모자)은 훌륭하게 구현되었으나, 안규홍의 검이 허리와 손에 두 개로 중복 생성된 물리적 오류로 인해 최종 불합격입니다."
       },
       {
        "label": "B",
        "score": 2,
        "verdict_ko": "명시된 오르막과 반대인 내리막으로 이동하고 있으며, 밧줄이 끊어지고 염재보의 주요 소품이 누락되는 등 핵심 연출 지시를 다수 위반했습니다."
       }
      ],
      "all_candidates_fail": true,
      "readings": [
       {
        "label": "A",
        "direction": "세 인물 모두 화면 좌측(오르막 캠프 방향)을 향해 걷고 있으며 시선 역시 진행 방향을 자연스럽게 향하고 있음.",
        "built_space": "야간의 산길 배경으로 좌측 상단에 캠프의 목책과 구조물이 위치함. 인물들이 캠프 방향으로 수레를 당기고 있어 지형과 이동 방향의 호환성이 올바름.",
        "entities": "이관회(좌측), 안규홍(중앙), 염재보(우측) 모두 지정된 레퍼런스의 복장과 외형에 일치함. 염재보는 모자를 착용하고 왼손에 총을 쥐고 있음. 짐이 실린 수레가 인물들 뒤에 위치함.",
        "hard_violations": [
         "객체 중복: 안규홍이 허리에 녹색 술이 달린 검을 차고 있음에도 불구하고, 왼손에 또 다른 검(칼집)을 들고 있어 하나의 검이 중복 생성됨."
        ],
        "physics": "세 명 모두 어깨에 밧줄을 걸고 한 발을 내디딘 자세(mid-stride)를 잘 유지하고 있으며 발끝이 지면을 단단히 딛고 있음. 밧줄은 대체로 수레를 향해 뒤쪽으로 당겨지는 장력을 보여줌."
       },
       {
        "label": "B",
        "direction": "세 인물 모두 화면 우측을 향해 이동하고 있으며, 이는 배경 구조상 캠프에서 멀어지는 하산(내리막) 방향임.",
        "built_space": "산길과 좌측 상단의 캠프가 확인되나, 인물들의 위치와 이동 방향이 수레를 이끌고 산 아래로 향하는 형태가 되어 오르막 지시와 모순됨.",
        "entities": "안규홍과 이관회의 복장 및 인상착의는 레퍼런스와 일치하나, 앞장서는 염재보(우측)는 레퍼런스의 모자를 쓰지 않았고 명시된 총도 소지하지 않음.",
        "hard_violations": [
         "구조물 방향 위반: 오르막길(uphill)로 수레를 끌어올려야 한다는 프롬프트 지시와 반대로 우측 내리막 방향으로 이동함.",
         "물리적 지지 위반: 이관회와 염재보의 어깨에 걸린 밧줄이 뒤의 수레와 연결되지 않은 채 허공에서 끊어지거나 아래로 늘어져 아무것도 당기지 않고 있음."
        ],
        "physics": "걷는 자세는 나타나 있으나, 수레를 끄는 핵심 매개체인 밧줄이 물리적으로 연결되지 않아 동작과 하중의 인과관계가 성립하지 않음."
       }
      ]
     },
     "ok": true
    },
    {
     "model": "gpt-high",
     "order": "reverse",
     "display_to_canonical": {
      "A": "B",
      "B": "A"
     },
     "raw": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 6,
        "verdict_ko": "세 사람의 전신과 수레를 담은 야간 와이드 숏에서 오르막 진행은 맞지만, 이관회의 어깨 끈과 손으로 당기는 동작이 빠져 세 사람의 공동 견인이 불완전하다."
       },
       {
        "label": "B",
        "score": 3,
        "verdict_ko": "세 사람의 어깨 끈·손잡이·보행과 무기 소지는 더 명확하지만, 높은 뒤쪽 수레에서 낮은 전경으로 내려오는 배치가 핵심인 오르막 견인 방향을 뒤집는다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "세 사람 모두 화면 오른쪽 앞의 길을 바라보며 그 방향으로 몸을 기울인다. 수레는 왼쪽 뒤의 낮은 지면에 있고 인물들은 오른쪽의 높아지는 지면으로 이동하여 오르막 견인 방향이 읽힌다. 안규홍의 칼집은 허리에서 뒤쪽 아래로 향하며 조준 동작은 없다.",
        "built_space": "수레 한 대가 세 사람 바로 뒤에 있으며 가까운 바퀴 하나와 반대편 바퀴 일부, 적재함 측면과 앞부분이 보인다. 배경에는 목책, 낮은 지붕의 건물과 부속 차양, 오른쪽 굵은 나무, 돌 섞인 흙길이 있어 장소 참조와 대응한다. 돌 화덕 한 곳은 분명하고 나머지 영역은 수레에 가린다. 세 사람 모두 길 위에 발을 놓고 있으며 전신이 프레임 안에 들어온다.",
        "entities": "인물은 세 명뿐이며 모두 지정된 한국인 남성 배역에 부합하는 외관이다. 왼쪽 안규홍은 묶은 검은 머리, 밝은 옷, 허리띠와 칼집이 보인다. 가운데 이관회는 가르마 탄 검은 머리, 갈색 저고리와 짧은 밝은 바지가 참조와 맞지만 어깨의 견인 끈은 확인되지 않는다. 오른쪽 염재보는 검은 옷과 각반을 착용했으나 참조의 삿갓이 없고 총 소지도 명확하지 않다. 수레에는 묶은 포대와 감자로 보이는 둥근 작물이 있지만 쌀·보리·옥수수·짚신을 각각 식별하기는 어렵다. 얼굴에는 땀 같은 광택이 있으며 읽을 수 있는 글자는 없다. 숨긴 옥반지는 확인할 수 없다.",
        "hard_violations": [],
        "physics": "세 사람은 각각 지면에 닿은 발로 체중을 지탱하고, 안규홍은 뒤꿈치를 들고 앞으로 내딛는 순간이다. 안규홍과 염재보는 어깨 부근의 밧줄을 손으로 잡지만 이관회는 양손을 아래로 내려 견인에 참여하는 하중 전달이 보이지 않는다. 염재보 앞의 밧줄 일부는 지면까지 늘어져 팽팽한 견인 상태가 약하다. 수레는 바퀴로 지지되고 짐은 적재함에 놓여 묶여 있다. 지지 없이 떠 있는 사람이나 물체는 없다."
       },
       {
        "label": "B",
        "direction": "세 사람은 화면 왼쪽 앞을 향해 걷고, 이관회는 발밑 길을, 안규홍과 염재보는 진행 방향을 본다. 밧줄은 각자의 어깨에서 오른쪽 뒤 수레로 이어진다. 그러나 수레가 놓인 뒤쪽 지면이 높고 인물들이 향하는 왼쪽 전경이 낮아, 오르기보다 내려오는 방향으로 읽힌다. 염재보가 든 총의 총구는 지면 쪽으로 향하며 사람을 겨누지 않는다.",
        "built_space": "수레 한 대가 오른쪽 뒤에 있고 바퀴 두 개와 적재함 앞면·측면이 보인다. 목책, 낮은 지붕의 건물과 차양, 왼쪽 돌 화덕 구역, 오른쪽 굵은 나무와 바위가 장소 참조에 가깝다. 인물 세 명은 수레보다 앞쪽 길에 나란히 떨어져 있으며 전신을 담은 와이드 숏이다. 다만 길이 카메라 쪽으로 올라오는 것이 아니라 뒤쪽 산영지 쪽으로 높아져, 요구한 카메라의 오르막 쪽 위치와 반대로 보인다.",
        "entities": "지정된 세 남성만 등장한다. 가운데 안규홍은 묶은 검은 머리와 밝은 옷, 왼쪽 이관회는 자연스러운 검은 머리와 갈색 저고리·짧은 밝은 바지, 오른쪽 염재보는 삿갓과 검은 옷·각반으로 참조의 구별되는 특징을 유지한다. 안규홍의 칼과 염재보의 총이 보인다. 수레에는 여러 포대가 있지만 쌀·보리·감자·옥수수·짚신을 개별적으로 확인하기 어렵다. 얼굴의 땀 광택과 야간 노출은 보이며 읽을 수 있는 글자는 없다. 옥반지는 보이지 않아 은닉 여부를 판정할 수 없다.",
        "hard_violations": [
         "수레가 있는 높은 뒤쪽에서 인물들이 낮은 전경으로 내려오는 경사 배치로, 명시된 수레 뒤쪽에서 카메라 쪽으로 올라오는 경로와 오르막 견인 방향을 반대로 구현했다."
        ],
        "physics": "세 사람 모두 한 발로 지면을 지지하면서 다른 발을 옮기는 보행 자세이며, 각자 손으로 어깨의 밧줄을 잡는다. 밧줄은 뒤쪽 수레로 연결되어 A보다 견인 관계가 명확하다. 칼은 안규홍의 허리와 손 부근에서 지지되고 총은 염재보의 손에 잡혀 있다. 수레 바퀴는 지면에 닿고 짐은 적재함에 실려 있어 부유 문제는 없다. 다만 이 물리적으로 가능한 보행은 요구한 오르막이 아닌 내리막 방향으로 배치되어 있다."
       }
      ],
      "all_candidates_fail": false
     },
     "normalized": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 6,
        "verdict_ko": "세 사람의 전신과 수레를 담은 야간 와이드 숏에서 오르막 진행은 맞지만, 이관회의 어깨 끈과 손으로 당기는 동작이 빠져 세 사람의 공동 견인이 불완전하다."
       },
       {
        "label": "A",
        "score": 3,
        "verdict_ko": "세 사람의 어깨 끈·손잡이·보행과 무기 소지는 더 명확하지만, 높은 뒤쪽 수레에서 낮은 전경으로 내려오는 배치가 핵심인 오르막 견인 방향을 뒤집는다."
       }
      ],
      "all_candidates_fail": false,
      "readings": [
       {
        "label": "B",
        "direction": "세 사람 모두 화면 오른쪽 앞의 길을 바라보며 그 방향으로 몸을 기울인다. 수레는 왼쪽 뒤의 낮은 지면에 있고 인물들은 오른쪽의 높아지는 지면으로 이동하여 오르막 견인 방향이 읽힌다. 안규홍의 칼집은 허리에서 뒤쪽 아래로 향하며 조준 동작은 없다.",
        "built_space": "수레 한 대가 세 사람 바로 뒤에 있으며 가까운 바퀴 하나와 반대편 바퀴 일부, 적재함 측면과 앞부분이 보인다. 배경에는 목책, 낮은 지붕의 건물과 부속 차양, 오른쪽 굵은 나무, 돌 섞인 흙길이 있어 장소 참조와 대응한다. 돌 화덕 한 곳은 분명하고 나머지 영역은 수레에 가린다. 세 사람 모두 길 위에 발을 놓고 있으며 전신이 프레임 안에 들어온다.",
        "entities": "인물은 세 명뿐이며 모두 지정된 한국인 남성 배역에 부합하는 외관이다. 왼쪽 안규홍은 묶은 검은 머리, 밝은 옷, 허리띠와 칼집이 보인다. 가운데 이관회는 가르마 탄 검은 머리, 갈색 저고리와 짧은 밝은 바지가 참조와 맞지만 어깨의 견인 끈은 확인되지 않는다. 오른쪽 염재보는 검은 옷과 각반을 착용했으나 참조의 삿갓이 없고 총 소지도 명확하지 않다. 수레에는 묶은 포대와 감자로 보이는 둥근 작물이 있지만 쌀·보리·옥수수·짚신을 각각 식별하기는 어렵다. 얼굴에는 땀 같은 광택이 있으며 읽을 수 있는 글자는 없다. 숨긴 옥반지는 확인할 수 없다.",
        "hard_violations": [],
        "physics": "세 사람은 각각 지면에 닿은 발로 체중을 지탱하고, 안규홍은 뒤꿈치를 들고 앞으로 내딛는 순간이다. 안규홍과 염재보는 어깨 부근의 밧줄을 손으로 잡지만 이관회는 양손을 아래로 내려 견인에 참여하는 하중 전달이 보이지 않는다. 염재보 앞의 밧줄 일부는 지면까지 늘어져 팽팽한 견인 상태가 약하다. 수레는 바퀴로 지지되고 짐은 적재함에 놓여 묶여 있다. 지지 없이 떠 있는 사람이나 물체는 없다."
       },
       {
        "label": "A",
        "direction": "세 사람은 화면 왼쪽 앞을 향해 걷고, 이관회는 발밑 길을, 안규홍과 염재보는 진행 방향을 본다. 밧줄은 각자의 어깨에서 오른쪽 뒤 수레로 이어진다. 그러나 수레가 놓인 뒤쪽 지면이 높고 인물들이 향하는 왼쪽 전경이 낮아, 오르기보다 내려오는 방향으로 읽힌다. 염재보가 든 총의 총구는 지면 쪽으로 향하며 사람을 겨누지 않는다.",
        "built_space": "수레 한 대가 오른쪽 뒤에 있고 바퀴 두 개와 적재함 앞면·측면이 보인다. 목책, 낮은 지붕의 건물과 차양, 왼쪽 돌 화덕 구역, 오른쪽 굵은 나무와 바위가 장소 참조에 가깝다. 인물 세 명은 수레보다 앞쪽 길에 나란히 떨어져 있으며 전신을 담은 와이드 숏이다. 다만 길이 카메라 쪽으로 올라오는 것이 아니라 뒤쪽 산영지 쪽으로 높아져, 요구한 카메라의 오르막 쪽 위치와 반대로 보인다.",
        "entities": "지정된 세 남성만 등장한다. 가운데 안규홍은 묶은 검은 머리와 밝은 옷, 왼쪽 이관회는 자연스러운 검은 머리와 갈색 저고리·짧은 밝은 바지, 오른쪽 염재보는 삿갓과 검은 옷·각반으로 참조의 구별되는 특징을 유지한다. 안규홍의 칼과 염재보의 총이 보인다. 수레에는 여러 포대가 있지만 쌀·보리·감자·옥수수·짚신을 개별적으로 확인하기 어렵다. 얼굴의 땀 광택과 야간 노출은 보이며 읽을 수 있는 글자는 없다. 옥반지는 보이지 않아 은닉 여부를 판정할 수 없다.",
        "hard_violations": [
         "수레가 있는 높은 뒤쪽에서 인물들이 낮은 전경으로 내려오는 경사 배치로, 명시된 수레 뒤쪽에서 카메라 쪽으로 올라오는 경로와 오르막 견인 방향을 반대로 구현했다."
        ],
        "physics": "세 사람 모두 한 발로 지면을 지지하면서 다른 발을 옮기는 보행 자세이며, 각자 손으로 어깨의 밧줄을 잡는다. 밧줄은 뒤쪽 수레로 연결되어 A보다 견인 관계가 명확하다. 칼은 안규홍의 허리와 손 부근에서 지지되고 총은 염재보의 손에 잡혀 있다. 수레 바퀴는 지면에 닿고 짐은 적재함에 실려 있어 부유 문제는 없다. 다만 이 물리적으로 가능한 보행은 요구한 오르막이 아닌 내리막 방향으로 배치되어 있다."
       }
      ]
     },
     "ok": true
    }
   ],
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "slot_winner_match": false,
   "slot_winner": {
    "gemini-pro": "A",
    "gpt-high": "B"
   },
   "route": "cross_slot_combined"
  },
  "dual": {
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "normalized": {
    "A": 1.5,
    "B": 1.4
   },
   "adjusted": {
    "A": 1.25,
    "B": 1.15
   },
   "violations": {
    "A": [
     "[gemini-pro] 객체 중복: 안규홍이 허리에 녹색 술이 달린 검을 차고 있음에도 불구하고, 왼손에 또 다른 검(칼집)을 들고 있어 하나의 검이 중복 생성됨.",
     "[gpt-high] 수레가 있는 높은 뒤쪽에서 인물들이 낮은 전경으로 내려오는 경사 배치로, 명시된 수레 뒤쪽에서 카메라 쪽으로 올라오는 경로와 오르막 견인 방향을 반대로 구현했다."
    ],
    "B": [
     "[gemini-pro] 구조물 방향 위반: 오르막길(uphill)로 수레를 끌어올려야 한다는 프롬프트 지시와 반대로 우측 내리막 방향으로 이동함.",
     "[gemini-pro] 물리적 지지 위반: 이관회와 염재보의 어깨에 걸린 밧줄이 뒤의 수레와 연결되지 않은 채 허공에서 끊어지거나 아래로 늘어져 아무것도 당기지 않고 있음."
    ]
   },
   "per_model_winner": {
    "gemini-pro": "A",
    "gpt-high": "B"
   },
   "agreed": false
  },
  "totals": {
   "A": 1250,
   "B": 1150
  },
  "selected": "A",
  "ranking": [
   "A",
   "B"
  ],
  "verdicts": [
   {
    "label": "A",
    "score": 1250,
    "verdict_ko": "오르막 방향과 인물들의 외형 및 소품(염재보의 총과 모자)은 훌륭하게 구현되었으나, 안규홍의 검이 허리와 손에 두 개로 중복 생성된 물리적 오류로 인해 최종 불합격입니다.  ★위반: [gemini-pro] 객체 중복: 안규홍이 허리에 녹색 술이 달린 검을 차고 있음에도 불구하고, 왼손에 또 다른 검(칼집)을 들고 있어 하나의 검이 중복 생성됨. / [gpt-high] 수레가 있는 높은 뒤쪽에서 인물들이 낮은 전경으로 내려오는 경사 배치로, 명시된 수레 뒤쪽에서 카메라 쪽으로 올라오는 경로와 오르막 견인 방향을 반대로 구현했다."
   },
   {
    "label": "B",
    "score": 1150,
    "verdict_ko": "명시된 오르막과 반대인 내리막으로 이동하고 있으며, 밧줄이 끊어지고 염재보의 주요 소품이 누락되는 등 핵심 연출 지시를 다수 위반했습니다.  ★위반: [gemini-pro] 구조물 방향 위반: 오르막길(uphill)로 수레를 끌어올려야 한다는 프롬프트 지시와 반대로 우측 내리막 방향으로 이동함. / [gemini-pro] 물리적 지지 위반: 이관회와 염재보의 어깨에 걸린 밧줄이 뒤의 수레와 연결되지 않은 채 허공에서 끊어지거나 아래로 늘어져 아무것도 당기지 않고 있음."
   }
  ],
  "refs": [
   {
    "label": "LOCATION PHOTOGRAPH — the exact place of this shot: its architecture, materials, fixed features and lighting mood are spatial truth; stage the moment inside this place. Never copy its camera framing.",
    "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/db00517b-165b-4005-a287-f28936b59b95/images/272bd7af-9294-4490-98c6-88a40e9deacc/scene/recipe/groupbg_camp_supply_approach_8601af.png",
    "asset_id": "69bb9661-49a8-4a75-a056-1e48e42c72b5",
    "role": "bgfirst_group_bg"
   },
   {
    "label": "CHARACTER REFERENCE — 안규홍: the exact person appearing in this shot; match face, hair and build exactly.",
    "path": "<bytes:1028384>",
    "asset_id": "3b5894db-ce55-4f34-b966-5e725c3aa952",
    "role": "character_ref"
   },
   {
    "label": "CHARACTER REFERENCE — 이관회: the exact person appearing in this shot; match face, hair and build exactly.",
    "path": "<bytes:933540>",
    "asset_id": "851b450f-ac84-47fe-ac5a-03a56f63fcae",
    "role": "character_ref"
   },
   {
    "label": "CHARACTER REFERENCE — 염재보: the exact person appearing in this shot; match face, hair and build exactly.",
    "path": "<bytes:964351>",
    "asset_id": "ab287289-d0b7-436d-8f70-68452f898f73",
    "role": "character_ref"
   }
  ],
  "critique_skipped": true,
  "shot_run_produced": false,
  "shot_run_uid": "06aa1428-5213-72f8-86cc-80f874da7d97",
  "bgfirst": {
   "bg_path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/db00517b-165b-4005-a287-f28936b59b95/images/272bd7af-9294-4490-98c6-88a40e9deacc/scene/recipe/S36sh1__bgfirst_bg.png",
   "bg_asset_id": "918e7565-903c-4cbe-aa25-7d6133afa09e",
   "bg_record_key": "S36sh1::bgfirst_bg",
   "chain_winner": true,
   "authority": "groupbg",
   "group_key": "camp_supply_approach",
   "groupbg_asset_id": "69bb9661-49a8-4a75-a056-1e48e42c72b5"
  },
  "ref_mode": "재투영 배경+콘티+엔티티 (2택1: 체인 승)",
  "share_plan": {
   "ref_plan": "background"
  }
 },
 "S36sh1::cine": {
  "applied": true,
  "attempted_at": "2026-09-09T04:54:31.349962+00:00",
  "fingerprint": "1359c6014d6b5cb3cce0f86062cd2725b4b370cc2d8ba88addd7254bcb94b798",
  "fingerprint_version": 2,
  "provider": "mai",
  "endpoint": "openrouter/chat-completions",
  "model": "microsoft/mai-image-2.6",
  "pack": "24.202608252115",
  "source_file": "S36sh1_sel.png",
  "source_sha256": "ed4f25aa121789335a24d416de52cf535617e4b88b666b9b349cd7a369f5c379",
  "file": "S36sh1_cine.png",
  "staged_sha256": "9391c31b9b9d471bf0e7915dd974f4dd96f8b0c6b120a9300de37f3b2fc36256",
  "latency_ms": 27789
 },
 "S36sh9::signage": {
  "fp": "b3eda5d6b535e1f0",
  "inscriptions": [],
  "cues": [],
  "dropped": []
 },
 "S36sh9": {
  "input_fingerprint": "ce5bb023393bed75",
  "prompt": "Create ONE FINAL photorealistic live-action film still of the moment below — early 20th-century Korea, Jeolla region; all people are Korean unless explicitly stated otherwise. TIME OF DAY (lock): night.\n\nSHOT TEXT (authoritative, Korean): 손에 감자를 쥔 채 능청스럽게 눈웃음을 지은 이관회의 펴진 얼굴.\n\nLOCATION (lock): The outdoor unloading spot beside the supply cart at the mountain camp's entrance at night. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: close-up\n- KEY BACKGROUND ELEMENTS: Caught potato (Held in 이관회's hand after the catch); used as Remain a small narrative detail at the lower frame edge beneath his smile; Cart and provisions (Stopped beside the exchange with provisions still aboard) — A partial near-side view remains behind 이관회; used as Provide softly resolved spatial continuity without competing with his face.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Maintain the nighttime tonal continuity while preserving gentle facial detail and understated human warmth in the smile.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The clothing, hair and overall look of 이관회 — who appear both in that photo and in this shot — are LOCKED to that photo. Anyone else visible in that photo is NOT in this shot: never carry their face, body or clothing onto anyone here. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The cart is being unloaded, with rice carried inside and potatoes assigned to shade; barley, shoes and corn remain part of the load. 이관회: He holds the potato he has just caught.\n\nTHE HAND THAT IS DOING THIS: the object at the centre of this shot is being held, operated, read, aimed or handed over by 이관회 right now, so 이관회's hand — and as much of the wrist and forearm as the framing reaches — is in the frame, gripping or resting on that object exactly the way its real use requires. Match that hand to 이관회: its size, build, skin, age, grooming, sleeve and anything worn on it belong to that person and to no one else.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 이관회 (한국인 남성, 청년 나이대, 자연스러운 검은 머리) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.",
  "roll_prompts": {
   "A": "Create ONE FINAL photorealistic live-action film still of the moment below — early 20th-century Korea, Jeolla region; all people are Korean unless explicitly stated otherwise. TIME OF DAY (lock): night.\n\nSHOT TEXT (authoritative, Korean): 손에 감자를 쥔 채 능청스럽게 눈웃음을 지은 이관회의 펴진 얼굴.\n\nLOCATION (lock): The outdoor unloading spot beside the supply cart at the mountain camp's entrance at night. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: close-up\n- KEY BACKGROUND ELEMENTS: Caught potato (Held in 이관회's hand after the catch); used as Remain a small narrative detail at the lower frame edge beneath his smile; Cart and provisions (Stopped beside the exchange with provisions still aboard) — A partial near-side view remains behind 이관회; used as Provide softly resolved spatial continuity without competing with his face.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Maintain the nighttime tonal continuity while preserving gentle facial detail and understated human warmth in the smile.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The clothing, hair and overall look of 이관회 — who appear both in that photo and in this shot — are LOCKED to that photo. Anyone else visible in that photo is NOT in this shot: never carry their face, body or clothing onto anyone here. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The cart is being unloaded, with rice carried inside and potatoes assigned to shade; barley, shoes and corn remain part of the load. 이관회: He holds the potato he has just caught.\n\nTHE HAND THAT IS DOING THIS: the object at the centre of this shot is being held, operated, read, aimed or handed over by 이관회 right now, so 이관회's hand — and as much of the wrist and forearm as the framing reaches — is in the frame, gripping or resting on that object exactly the way its real use requires. Match that hand to 이관회: its size, build, skin, age, grooming, sleeve and anything worn on it belong to that person and to no one else.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 이관회 (한국인 남성, 청년 나이대, 자연스러운 검은 머리) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.",
   "B": "Create ONE FINAL photorealistic live-action film still of the moment below — early 20th-century Korea, Jeolla region; all people are Korean unless explicitly stated otherwise. TIME OF DAY (lock): night.\n\nSHOT TEXT (authoritative, Korean): 손에 감자를 쥔 채 능청스럽게 눈웃음을 지은 이관회의 펴진 얼굴.\n\nLOCATION (lock): The outdoor unloading spot beside the supply cart at the mountain camp's entrance at night. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: close-up\n- KEY BACKGROUND ELEMENTS: Caught potato (Held in 이관회's hand after the catch); used as Remain a small narrative detail at the lower frame edge beneath his smile; Cart and provisions (Stopped beside the exchange with provisions still aboard) — A partial near-side view remains behind 이관회; used as Provide softly resolved spatial continuity without competing with his face.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Maintain the nighttime tonal continuity while preserving gentle facial detail and understated human warmth in the smile.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The clothing, hair and overall look of 이관회 — who appear both in that photo and in this shot — are LOCKED to that photo. Anyone else visible in that photo is NOT in this shot: never carry their face, body or clothing onto anyone here. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The cart is being unloaded, with rice carried inside and potatoes assigned to shade; barley, shoes and corn remain part of the load. 이관회: He holds the potato he has just caught.\n\nTHE HAND THAT IS DOING THIS: the object at the centre of this shot is being held, operated, read, aimed or handed over by 이관회 right now, so 이관회's hand — and as much of the wrist and forearm as the framing reaches — is in the frame, gripping or resting on that object exactly the way its real use requires. Match that hand to 이관회: its size, build, skin, age, grooming, sleeve and anything worn on it belong to that person and to no one else.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 이관회 (한국인 남성, 청년 나이대, 자연스러운 검은 머리) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.\n\nCOMPOSITION VARIATION (this candidate only): keep the place, the people and the moment exactly as specified, and keep from the CAMERA & FRAME contract its subject scale, screen placement and key background elements exactly. For this candidate only, its camera angle and height are deliberately left open: choose a camera position distinctly different from the obvious one, as a film director picking a second setup on the same blocking. Stay true to any explicit framing the SHOT TEXT itself dictates."
  },
  "shot_run_spend_attempt_count": 1,
  "readings": [
   {
    "label": "A",
    "direction": "피사체는 카메라 우측을 응시하며 프롬프트에 명시된 능청스러운 눈웃음을 짓고 있음.",
    "built_space": "어두운 밤의 산채 입구 하역장. 배경에 수레와 짐들이 부드럽게 아웃포커싱되어 인물의 얼굴과 경쟁하지 않고 공간적 연속성을 제공함.",
    "entities": "레퍼런스와 정확히 일치하는 이관회의 얼굴, 헤어스타일, 의상. 프레임 하단 끝에 명시된 대로 감자를 하나 쥐고 있음.",
    "hard_violations": [],
    "physics": "오른손 손가락으로 감자를 안정적으로 쥐고 있으며, 옷의 주름과 인체의 질감이 자연스럽고 물리적으로 타당함."
   },
   {
    "label": "B",
    "direction": "피사체는 카메라 우측을 보며 옅은 미소를 짓고 있음.",
    "built_space": "산채 입구 하역장의 수레 앞. 이전 샷 레퍼런스에는 존재하지 않던 밝은 랜턴이 수레 위에 배치되어 야간의 톤 앤 매너를 깨뜨림.",
    "entities": "이관회의 얼굴 특징이 레퍼런스와 다소 다르며, 요구된 단일 감자가 아닌 두 개의 감자를 양손에 각각 쥐고 있음.",
    "hard_violations": [
     "[gemini-pro] 이전 샷 레퍼런스에 존재하지 않던 임의의 사물(밝게 빛나는 랜턴)이 생성됨",
     "[gemini-pro] 명시된 단일 감자가 아닌 여분의 감자가 추가로 생성되어 양손에 들려 있음",
     "[gpt-high] 막 받은 감자 한 개를 쥔 상태여야 하는데, 반대편 손에도 두 번째 감자를 추가했다.",
     "[gpt-high] 장소와 조명이 고정된 참조에 없는 켜진 등불을 얼굴 뒤에 추가해 새로운 실물 광원을 만들었다."
    ],
    "physics": "양손으로 각각 감자를 들고 무릎과 가슴 쪽에 기대고 있으며, 중력에 따른 물체의 지지나 인체 비례 자체는 정상적임."
   }
  ],
  "cross_model_order": {
   "policy": "cross_model_order_v2_slot0_fwd_slot1_rev_parallel2",
   "slots": [
    {
     "model": "gemini-pro",
     "order": "forward",
     "display_to_canonical": {
      "A": "A",
      "B": "B"
     },
     "raw": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 10,
        "verdict_ko": "프롬프트가 지시한 클로즈업 프레이밍과 이전 샷의 야간 조명 톤을 완벽하게 유지했으며, 프레임 하단에 감자를 쥔 손과 배경의 수레를 정확한 비율과 심도로 배치하여 최적의 결과를 보여줍니다."
       },
       {
        "label": "B",
        "score": 3,
        "verdict_ko": "요구된 클로즈업 앵글을 무시하고 미디엄 샷으로 렌더링했으며, 이전 샷에 없던 랜턴 광원을 추가하고 감자를 두 개나 들고 있어 프롬프트의 핵심 지시사항과 연속성을 크게 위반했습니다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "피사체는 카메라 우측을 응시하며 프롬프트에 명시된 능청스러운 눈웃음을 짓고 있음.",
        "built_space": "어두운 밤의 산채 입구 하역장. 배경에 수레와 짐들이 부드럽게 아웃포커싱되어 인물의 얼굴과 경쟁하지 않고 공간적 연속성을 제공함.",
        "entities": "레퍼런스와 정확히 일치하는 이관회의 얼굴, 헤어스타일, 의상. 프레임 하단 끝에 명시된 대로 감자를 하나 쥐고 있음.",
        "hard_violations": [],
        "physics": "오른손 손가락으로 감자를 안정적으로 쥐고 있으며, 옷의 주름과 인체의 질감이 자연스럽고 물리적으로 타당함."
       },
       {
        "label": "B",
        "direction": "피사체는 카메라 우측을 보며 옅은 미소를 짓고 있음.",
        "built_space": "산채 입구 하역장의 수레 앞. 이전 샷 레퍼런스에는 존재하지 않던 밝은 랜턴이 수레 위에 배치되어 야간의 톤 앤 매너를 깨뜨림.",
        "entities": "이관회의 얼굴 특징이 레퍼런스와 다소 다르며, 요구된 단일 감자가 아닌 두 개의 감자를 양손에 각각 쥐고 있음.",
        "hard_violations": [
         "이전 샷 레퍼런스에 존재하지 않던 임의의 사물(밝게 빛나는 랜턴)이 생성됨",
         "명시된 단일 감자가 아닌 여분의 감자가 추가로 생성되어 양손에 들려 있음"
        ],
        "physics": "양손으로 각각 감자를 들고 무릎과 가슴 쪽에 기대고 있으며, 중력에 따른 물체의 지지나 인체 비례 자체는 정상적임."
       }
      ],
      "all_candidates_fail": false
     },
     "normalized": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 10,
        "verdict_ko": "프롬프트가 지시한 클로즈업 프레이밍과 이전 샷의 야간 조명 톤을 완벽하게 유지했으며, 프레임 하단에 감자를 쥔 손과 배경의 수레를 정확한 비율과 심도로 배치하여 최적의 결과를 보여줍니다."
       },
       {
        "label": "B",
        "score": 3,
        "verdict_ko": "요구된 클로즈업 앵글을 무시하고 미디엄 샷으로 렌더링했으며, 이전 샷에 없던 랜턴 광원을 추가하고 감자를 두 개나 들고 있어 프롬프트의 핵심 지시사항과 연속성을 크게 위반했습니다."
       }
      ],
      "all_candidates_fail": false,
      "readings": [
       {
        "label": "A",
        "direction": "피사체는 카메라 우측을 응시하며 프롬프트에 명시된 능청스러운 눈웃음을 짓고 있음.",
        "built_space": "어두운 밤의 산채 입구 하역장. 배경에 수레와 짐들이 부드럽게 아웃포커싱되어 인물의 얼굴과 경쟁하지 않고 공간적 연속성을 제공함.",
        "entities": "레퍼런스와 정확히 일치하는 이관회의 얼굴, 헤어스타일, 의상. 프레임 하단 끝에 명시된 대로 감자를 하나 쥐고 있음.",
        "hard_violations": [],
        "physics": "오른손 손가락으로 감자를 안정적으로 쥐고 있으며, 옷의 주름과 인체의 질감이 자연스럽고 물리적으로 타당함."
       },
       {
        "label": "B",
        "direction": "피사체는 카메라 우측을 보며 옅은 미소를 짓고 있음.",
        "built_space": "산채 입구 하역장의 수레 앞. 이전 샷 레퍼런스에는 존재하지 않던 밝은 랜턴이 수레 위에 배치되어 야간의 톤 앤 매너를 깨뜨림.",
        "entities": "이관회의 얼굴 특징이 레퍼런스와 다소 다르며, 요구된 단일 감자가 아닌 두 개의 감자를 양손에 각각 쥐고 있음.",
        "hard_violations": [
         "이전 샷 레퍼런스에 존재하지 않던 임의의 사물(밝게 빛나는 랜턴)이 생성됨",
         "명시된 단일 감자가 아닌 여분의 감자가 추가로 생성되어 양손에 들려 있음"
        ],
        "physics": "양손으로 각각 감자를 들고 무릎과 가슴 쪽에 기대고 있으며, 중력에 따른 물체의 지지나 인체 비례 자체는 정상적임."
       }
      ]
     },
     "ok": true
    },
    {
     "model": "gpt-high",
     "order": "reverse",
     "display_to_canonical": {
      "A": "B",
      "B": "A"
     },
     "raw": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 3,
        "verdict_ko": "능청스러운 미소와 갈색 옷은 잘 맞지만, 무릎까지 드러낸 넓은 구도가 얼굴 클로즈업을 벗어나고 양손에 감자를 하나씩 추가해 지정된 소품 상태를 어겼다."
       },
       {
        "label": "B",
        "score": 8,
        "verdict_ko": "얼굴 중심 클로즈업, 눈웃음, 하단의 감자 한 개와 흐릿한 수레가 지시를 충실히 따르지만, 저고리가 참조의 갈색보다 회백색이며 미소는 능청스럽기보다 환하다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "얼굴과 눈은 화면 오른쪽의 프레임 밖 상대를 향한다. 감자를 바라보거나 렌즈를 정면 응시하지는 않는다. 두 손은 각각 감자를 몸 앞에 잡고 있으며 던지는 동작은 아니다.",
        "built_space": "인물 오른쪽 뒤에 목제 수레 한 대가 있고, 가까운 측판과 바퀴 일부 하나가 보인다. 적재함에는 여러 꾸러미, 신발과 옥수수가 남아 있다. 뒤쪽의 나무 울타리와 숲은 참조 장소와 유사하지만, 얼굴 옆에 참조에서 확인되지 않는 켜진 등불 하나가 추가됐다. 인물의 상체뿐 아니라 굽힌 무릎까지 보여 수레와 적재물이 요구보다 크게 드러난다.",
        "entities": "젊은 한국인 남성으로 보이는 인물 한 명이며, 검은 가르마 머리와 얼굴은 이관회 참조에 대체로 부합한다. 해진 갈색 저고리와 밝은 바지도 참조와 가깝다. 눈가를 좁히고 입을 다문 미소는 능청스러운 표정에 맞는다. 그러나 잡은 감자 한 개가 아니라 가슴 앞 손과 화면 오른쪽 하단 손에 감자가 각각 하나씩 있다. 읽을 수 있는 글자는 없다.",
        "hard_violations": [
         "막 받은 감자 한 개를 쥔 상태여야 하는데, 반대편 손에도 두 번째 감자를 추가했다.",
         "장소와 조명이 고정된 참조에 없는 켜진 등불을 얼굴 뒤에 추가해 새로운 실물 광원을 만들었다."
        ],
        "physics": "가슴 앞 감자는 손가락으로 감싸 쥐고, 하단 감자는 다른 손의 손바닥과 손가락이 받친다. 두 감자 모두 지지 접촉이 있다. 인물은 무릎을 굽힌 낮은 자세이며 발과 좌면은 프레임 밖이므로 정확한 하체 지지점은 확인되지 않지만, 공중에 뜬 자세로 보이지는 않는다. 짐은 수레 적재함에 놓여 있다."
       },
       {
        "label": "B",
        "direction": "얼굴은 거의 정면이지만 눈은 화면 왼쪽의 프레임 밖 상대 쪽으로 살짝 향한다. 눈꺼풀을 자연스럽게 좁힌 눈웃음이며, 감자를 내려다보지 않는다. 손에 든 감자는 얼굴 아래에 머물러 있어 이미 받은 뒤의 순간으로 읽힌다.",
        "built_space": "인물 뒤 화면 왼쪽에 수레 한 대의 가까운 측판과 바퀴 하나가 부분적으로 보이고, 적재함 위에는 여러 포대와 꾸러미가 남아 있다. 목제 울타리, 숲과 바위가 어두운 배경으로 이어진다. 수레는 부드럽게 흐려져 얼굴과 경쟁하지 않으며, 얼굴과 어깨를 중심으로 자른 클로즈업이 지정된 공간 관계에 맞는다.",
        "entities": "젊은 한국인 남성으로 보이는 인물 한 명이며 검은 가르마 머리와 얼굴은 이관회 참조와 대체로 일치한다. 다만 보이는 저고리는 참조의 갈색보다 회백색이고 어깨의 해진 형태도 덜 드러난다. 자연스러운 눈웃음과 펴진 얼굴이 보이지만 치아를 드러낸 웃음은 다소 환하다. 감자 한 개와 이를 잡은 본인의 손이 하단에 있다. 배경 짐의 구체적인 곡물 종류는 흐림 때문에 식별할 수 없으며, 추가 인물이나 읽을 수 있는 글자는 없다.",
        "hard_violations": [],
        "physics": "감자는 화면 하단에서 인물의 손가락이 아래와 옆을 감싸 지지한다. 손은 팔과 자연스럽게 이어지며 피부와 크기도 인물에 맞는다. 몸의 하부 지지점은 클로즈업 밖이지만 상체에 부유나 비정상적인 꺾임은 없다. 포대는 수레 위에 실려 있어 지지가 명확하다."
       }
      ],
      "all_candidates_fail": false
     },
     "normalized": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 3,
        "verdict_ko": "능청스러운 미소와 갈색 옷은 잘 맞지만, 무릎까지 드러낸 넓은 구도가 얼굴 클로즈업을 벗어나고 양손에 감자를 하나씩 추가해 지정된 소품 상태를 어겼다."
       },
       {
        "label": "A",
        "score": 8,
        "verdict_ko": "얼굴 중심 클로즈업, 눈웃음, 하단의 감자 한 개와 흐릿한 수레가 지시를 충실히 따르지만, 저고리가 참조의 갈색보다 회백색이며 미소는 능청스럽기보다 환하다."
       }
      ],
      "all_candidates_fail": false,
      "readings": [
       {
        "label": "B",
        "direction": "얼굴과 눈은 화면 오른쪽의 프레임 밖 상대를 향한다. 감자를 바라보거나 렌즈를 정면 응시하지는 않는다. 두 손은 각각 감자를 몸 앞에 잡고 있으며 던지는 동작은 아니다.",
        "built_space": "인물 오른쪽 뒤에 목제 수레 한 대가 있고, 가까운 측판과 바퀴 일부 하나가 보인다. 적재함에는 여러 꾸러미, 신발과 옥수수가 남아 있다. 뒤쪽의 나무 울타리와 숲은 참조 장소와 유사하지만, 얼굴 옆에 참조에서 확인되지 않는 켜진 등불 하나가 추가됐다. 인물의 상체뿐 아니라 굽힌 무릎까지 보여 수레와 적재물이 요구보다 크게 드러난다.",
        "entities": "젊은 한국인 남성으로 보이는 인물 한 명이며, 검은 가르마 머리와 얼굴은 이관회 참조에 대체로 부합한다. 해진 갈색 저고리와 밝은 바지도 참조와 가깝다. 눈가를 좁히고 입을 다문 미소는 능청스러운 표정에 맞는다. 그러나 잡은 감자 한 개가 아니라 가슴 앞 손과 화면 오른쪽 하단 손에 감자가 각각 하나씩 있다. 읽을 수 있는 글자는 없다.",
        "hard_violations": [
         "막 받은 감자 한 개를 쥔 상태여야 하는데, 반대편 손에도 두 번째 감자를 추가했다.",
         "장소와 조명이 고정된 참조에 없는 켜진 등불을 얼굴 뒤에 추가해 새로운 실물 광원을 만들었다."
        ],
        "physics": "가슴 앞 감자는 손가락으로 감싸 쥐고, 하단 감자는 다른 손의 손바닥과 손가락이 받친다. 두 감자 모두 지지 접촉이 있다. 인물은 무릎을 굽힌 낮은 자세이며 발과 좌면은 프레임 밖이므로 정확한 하체 지지점은 확인되지 않지만, 공중에 뜬 자세로 보이지는 않는다. 짐은 수레 적재함에 놓여 있다."
       },
       {
        "label": "A",
        "direction": "얼굴은 거의 정면이지만 눈은 화면 왼쪽의 프레임 밖 상대 쪽으로 살짝 향한다. 눈꺼풀을 자연스럽게 좁힌 눈웃음이며, 감자를 내려다보지 않는다. 손에 든 감자는 얼굴 아래에 머물러 있어 이미 받은 뒤의 순간으로 읽힌다.",
        "built_space": "인물 뒤 화면 왼쪽에 수레 한 대의 가까운 측판과 바퀴 하나가 부분적으로 보이고, 적재함 위에는 여러 포대와 꾸러미가 남아 있다. 목제 울타리, 숲과 바위가 어두운 배경으로 이어진다. 수레는 부드럽게 흐려져 얼굴과 경쟁하지 않으며, 얼굴과 어깨를 중심으로 자른 클로즈업이 지정된 공간 관계에 맞는다.",
        "entities": "젊은 한국인 남성으로 보이는 인물 한 명이며 검은 가르마 머리와 얼굴은 이관회 참조와 대체로 일치한다. 다만 보이는 저고리는 참조의 갈색보다 회백색이고 어깨의 해진 형태도 덜 드러난다. 자연스러운 눈웃음과 펴진 얼굴이 보이지만 치아를 드러낸 웃음은 다소 환하다. 감자 한 개와 이를 잡은 본인의 손이 하단에 있다. 배경 짐의 구체적인 곡물 종류는 흐림 때문에 식별할 수 없으며, 추가 인물이나 읽을 수 있는 글자는 없다.",
        "hard_violations": [],
        "physics": "감자는 화면 하단에서 인물의 손가락이 아래와 옆을 감싸 지지한다. 손은 팔과 자연스럽게 이어지며 피부와 크기도 인물에 맞는다. 몸의 하부 지지점은 클로즈업 밖이지만 상체에 부유나 비정상적인 꺾임은 없다. 포대는 수레 위에 실려 있어 지지가 명확하다."
       }
      ]
     },
     "ok": true
    }
   ],
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "slot_winner_match": true,
   "slot_winner": {
    "gemini-pro": "A",
    "gpt-high": "A"
   },
   "route": "cross_slot_agree"
  },
  "dual": {
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "normalized": {
    "A": 2.0,
    "B": 0.675
   },
   "adjusted": {
    "A": 2.0,
    "B": 0.425
   },
   "violations": {
    "B": [
     "[gemini-pro] 이전 샷 레퍼런스에 존재하지 않던 임의의 사물(밝게 빛나는 랜턴)이 생성됨",
     "[gemini-pro] 명시된 단일 감자가 아닌 여분의 감자가 추가로 생성되어 양손에 들려 있음",
     "[gpt-high] 막 받은 감자 한 개를 쥔 상태여야 하는데, 반대편 손에도 두 번째 감자를 추가했다.",
     "[gpt-high] 장소와 조명이 고정된 참조에 없는 켜진 등불을 얼굴 뒤에 추가해 새로운 실물 광원을 만들었다."
    ]
   },
   "per_model_winner": {
    "gemini-pro": "A",
    "gpt-high": "A"
   },
   "agreed": true
  },
  "totals": {
   "A": 2000,
   "B": 425
  },
  "selected": "A",
  "ranking": [
   "A",
   "B"
  ],
  "verdicts": [
   {
    "label": "A",
    "score": 2000,
    "verdict_ko": "프롬프트가 지시한 클로즈업 프레이밍과 이전 샷의 야간 조명 톤을 완벽하게 유지했으며, 프레임 하단에 감자를 쥔 손과 배경의 수레를 정확한 비율과 심도로 배치하여 최적의 결과를 보여줍니다."
   },
   {
    "label": "B",
    "score": 425,
    "verdict_ko": "요구된 클로즈업 앵글을 무시하고 미디엄 샷으로 렌더링했으며, 이전 샷에 없던 랜턴 광원을 추가하고 감자를 두 개나 들고 있어 프롬프트의 핵심 지시사항과 연속성을 크게 위반했습니다.  ★위반: [gemini-pro] 이전 샷 레퍼런스에 존재하지 않던 임의의 사물(밝게 빛나는 랜턴)이 생성됨 / [gemini-pro] 명시된 단일 감자가 아닌 여분의 감자가 추가로 생성되어 양손에 들려 있음 / [gpt-high] 막 받은 감자 한 개를 쥔 상태여야 하는데, 반대편 손에도 두 번째 감자를 추가했다. / [gpt-high] 장소와 조명이 고정된 참조에 없는 켜진 등불을 얼굴 뒤에 추가해 새로운 실물 광원을 만들었다."
   }
  ],
  "refs": [
   {
    "label": "PREVIOUS SHOT STILL — a visually related earlier shot of this same place: the location's look, materials, fixed features and lighting mood are LOCKED to this photo; never copy its camera framing. The clothing, hair and overall look of 이관회 — who appear both in that photo and in this shot — are LOCKED to that photo. Anyone else visible in that photo is NOT in this shot: never carry their face, body or clothing onto anyone here. If this photo shows a character who cannot move (dead or unconscious), that character's exact pose, position and orientation are ALSO LOCKED — treat their body as a fixed prop of the set that only the camera moves around.",
    "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/db00517b-165b-4005-a287-f28936b59b95/images/272bd7af-9294-4490-98c6-88a40e9deacc/scene/recipe/S36sh1_sel.png",
    "asset_id": "82b6c12a-3238-428f-9aca-67e714a522cd",
    "role": "prev_still"
   },
   {
    "label": "CHARACTER REFERENCE — 이관회: the exact person appearing in this shot; match face, hair and build exactly.",
    "path": "<bytes:933540>",
    "asset_id": "851b450f-ac84-47fe-ac5a-03a56f63fcae",
    "role": "character_ref"
   }
  ],
  "critique_skipped": true,
  "shot_run_produced": false,
  "shot_run_uid": "06aa1428-5633-7d6c-b091-8cd2a0f4de31",
  "ref_mode": "prev+엔티티",
  "share_plan": {
   "ref_plan": "prev",
   "prev_anchor_tag": "S36sh1"
  }
 },
 "S36sh9::cine": {
  "applied": true,
  "attempted_at": "2026-09-09T04:56:25.441145+00:00",
  "fingerprint": "b48e4e9566106d05ce93fb0096f2b042ee1f7154fcc63cd5584e0d1a8ca0c5bd",
  "fingerprint_version": 2,
  "provider": "mai",
  "endpoint": "openrouter/chat-completions",
  "model": "microsoft/mai-image-2.6",
  "pack": "24.202608252115",
  "source_file": "S36sh9_sel.png",
  "source_sha256": "33d02567cdf3d6cae275e44133bdace81e1c20946ac719b633d3b7e31c73cb2a",
  "file": "S36sh9_cine.png",
  "staged_sha256": "ad861509fe77faf54680c961d971ec4ea8293afb8a50ef3109c8013580eecef9",
  "latency_ms": 27808
 },
 "S36sh17::signage": {
  "fp": "adc997d9c65facd9",
  "inscriptions": [],
  "cues": [],
  "dropped": []
 },
 "S36sh17": {
  "input_fingerprint": "defdbb8c46682670",
  "prompt": "Create ONE FINAL photorealistic live-action film still of the moment below — early 20th-century Korea, Jeolla region; all people are Korean unless explicitly stated otherwise. TIME OF DAY (lock): night.\n\nSHOT TEXT (authoritative, Korean): 안규홍의 등을 향해 묘한 시선을 고정한 복만의 굳은 얼굴 클로즈업.\n\nLOCATION (lock): The outdoor sack-staging spot slightly apart from the supply cart in the mountain camp. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: close-up\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Retain the same subdued nighttime illumination and naturalistic facial contrast, allowing the unease to emerge without a new lighting cue.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The people visible in that photo are NOT in this shot: never carry their faces, bodies or clothing onto anyone here. Each person in THIS shot is defined solely by the PEOPLE list, their own CHARACTER REFERENCE images and their wardrobe notes. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): Rice has been moved toward the interior and potatoes toward shade; the partly unloaded cart retains the remaining supplies. 복만: He pauses while lowering a rice bale.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 복만 (한국인 남성, 성인 나이대, 거칠게 자란 수염, 자연스러운 검은 머리) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.",
  "roll_prompts": {
   "A": "Create ONE FINAL photorealistic live-action film still of the moment below — early 20th-century Korea, Jeolla region; all people are Korean unless explicitly stated otherwise. TIME OF DAY (lock): night.\n\nSHOT TEXT (authoritative, Korean): 안규홍의 등을 향해 묘한 시선을 고정한 복만의 굳은 얼굴 클로즈업.\n\nLOCATION (lock): The outdoor sack-staging spot slightly apart from the supply cart in the mountain camp. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: close-up\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Retain the same subdued nighttime illumination and naturalistic facial contrast, allowing the unease to emerge without a new lighting cue.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The people visible in that photo are NOT in this shot: never carry their faces, bodies or clothing onto anyone here. Each person in THIS shot is defined solely by the PEOPLE list, their own CHARACTER REFERENCE images and their wardrobe notes. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): Rice has been moved toward the interior and potatoes toward shade; the partly unloaded cart retains the remaining supplies. 복만: He pauses while lowering a rice bale.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 복만 (한국인 남성, 성인 나이대, 거칠게 자란 수염, 자연스러운 검은 머리) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.",
   "B": "Create ONE FINAL photorealistic live-action film still of the moment below — early 20th-century Korea, Jeolla region; all people are Korean unless explicitly stated otherwise. TIME OF DAY (lock): night.\n\nSHOT TEXT (authoritative, Korean): 안규홍의 등을 향해 묘한 시선을 고정한 복만의 굳은 얼굴 클로즈업.\n\nLOCATION (lock): The outdoor sack-staging spot slightly apart from the supply cart in the mountain camp. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: close-up\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Retain the same subdued nighttime illumination and naturalistic facial contrast, allowing the unease to emerge without a new lighting cue.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The people visible in that photo are NOT in this shot: never carry their faces, bodies or clothing onto anyone here. Each person in THIS shot is defined solely by the PEOPLE list, their own CHARACTER REFERENCE images and their wardrobe notes. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): Rice has been moved toward the interior and potatoes toward shade; the partly unloaded cart retains the remaining supplies. 복만: He pauses while lowering a rice bale.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 복만 (한국인 남성, 성인 나이대, 거칠게 자란 수염, 자연스러운 검은 머리) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.\n\nCOMPOSITION VARIATION (this candidate only): keep the place, the people and the moment exactly as specified, and keep from the CAMERA & FRAME contract its subject scale, screen placement and key background elements exactly. For this candidate only, its camera angle and height are deliberately left open: choose a camera position distinctly different from the obvious one, as a film director picking a second setup on the same blocking. Stay true to any explicit framing the SHOT TEXT itself dictates."
  },
  "shot_run_spend_attempt_count": 1,
  "readings": [
   {
    "label": "A",
    "direction": "복만의 시선이 전경에 등지고 있는 인물(안규홍)의 뒤통수와 목을 향해 정확히 고정되어 있습니다.",
    "built_space": "야외 산채의 흙바닥으로, 배경에 나무 울타리와 보급품이 실린 수레가 배치되어 있습니다. 다만 장소 레퍼런스에는 없던 따뜻한 빛을 내는 호롱불 3개가 배경에 추가되어 있습니다.",
    "entities": "복만은 캐릭터 레퍼런스와 일치하는 갓 형태의 모자와 수염, 어두운 의상을 착용하고 있습니다. 전경에는 텍스트가 요구한 안규홍의 등이 오버더숄더 구도로 걸쳐 있으며, 복만은 짚으로 엮은 쌀가마니를 들고 있습니다.",
    "hard_violations": [
     "[gpt-high] 복만만 화면에 허용한 인물 목록을 어기고 별도의 남성 얼굴 일부와 상반신을 추가했다.",
     "[gpt-high] 장소와 조명을 유지하고 새 조명 신호를 만들지 말라는 지시와 달리 배경에 발광하는 등불 네 곳을 추가했다."
    ],
    "physics": "복만의 두 손이 쌀가마니의 밧줄을 단단히 쥐고 무게를 지탱하고 있으며, 전경의 인물과 복만 모두 자연스러운 자세로 서 있습니다. 물리적으로 불가능한 요소는 없습니다."
   },
   {
    "label": "B",
    "direction": "복만의 시선이 전경에 등지고 있는 인물(안규홍)을 향해 고정되어 있습니다.",
    "built_space": "야외 산채 공간으로 나무 울타리와 수레, 바퀴가 배경에 보이며, 이전 샷의 어두운 조명 분위기를 호롱불 추가 없이 잘 유지하고 있습니다.",
    "entities": "전경에 안규홍의 등이 보이고 복만이 자루를 들고 있으나, 복만의 모자가 캐릭터 레퍼런스(갓 형태)가 아닌 이전 샷 레퍼런스 속 인물의 삿갓을 그대로 복사해 온 형태여서 의상 설정과 금지 지시를 위반했습니다.",
    "hard_violations": [
     "[gpt-high] 복만만 화면에 허용한 인물 목록을 어기고 별도의 남성 상반신을 크게 추가했다."
    ],
    "physics": "복만의 오른손이 자루의 윗부분을 쥐고 무게를 버티고 있으며, 두 인물의 서 있는 자세와 땅과의 접촉 모두 물리적으로 자연스럽습니다."
   }
  ],
  "cross_model_order": {
   "policy": "cross_model_order_v2_slot0_fwd_slot1_rev_parallel2",
   "slots": [
    {
     "model": "gemini-pro",
     "order": "forward",
     "display_to_canonical": {
      "A": "A",
      "B": "B"
     },
     "raw": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 8,
        "verdict_ko": "캐릭터 레퍼런스의 복장(모자 포함)과 인물의 굳은 표정을 클로즈업으로 정확히 구현했으나, 장소 레퍼런스에 없는 호롱불 조명이 배경에 추가되어 조명 락(lock)을 일부 위반했습니다."
       },
       {
        "label": "B",
        "score": 5,
        "verdict_ko": "배경의 조명과 분위기는 이전 샷과 잘 일치하나, 명시적인 금지 지시를 어기고 이전 샷에 등장한 인물의 삿갓을 복만에게 씌워 캐릭터 레퍼런스(의상)를 크게 위반했습니다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "복만의 시선이 전경에 등지고 있는 인물(안규홍)의 뒤통수와 목을 향해 정확히 고정되어 있습니다.",
        "built_space": "야외 산채의 흙바닥으로, 배경에 나무 울타리와 보급품이 실린 수레가 배치되어 있습니다. 다만 장소 레퍼런스에는 없던 따뜻한 빛을 내는 호롱불 3개가 배경에 추가되어 있습니다.",
        "entities": "복만은 캐릭터 레퍼런스와 일치하는 갓 형태의 모자와 수염, 어두운 의상을 착용하고 있습니다. 전경에는 텍스트가 요구한 안규홍의 등이 오버더숄더 구도로 걸쳐 있으며, 복만은 짚으로 엮은 쌀가마니를 들고 있습니다.",
        "hard_violations": [],
        "physics": "복만의 두 손이 쌀가마니의 밧줄을 단단히 쥐고 무게를 지탱하고 있으며, 전경의 인물과 복만 모두 자연스러운 자세로 서 있습니다. 물리적으로 불가능한 요소는 없습니다."
       },
       {
        "label": "B",
        "direction": "복만의 시선이 전경에 등지고 있는 인물(안규홍)을 향해 고정되어 있습니다.",
        "built_space": "야외 산채 공간으로 나무 울타리와 수레, 바퀴가 배경에 보이며, 이전 샷의 어두운 조명 분위기를 호롱불 추가 없이 잘 유지하고 있습니다.",
        "entities": "전경에 안규홍의 등이 보이고 복만이 자루를 들고 있으나, 복만의 모자가 캐릭터 레퍼런스(갓 형태)가 아닌 이전 샷 레퍼런스 속 인물의 삿갓을 그대로 복사해 온 형태여서 의상 설정과 금지 지시를 위반했습니다.",
        "hard_violations": [],
        "physics": "복만의 오른손이 자루의 윗부분을 쥐고 무게를 버티고 있으며, 두 인물의 서 있는 자세와 땅과의 접촉 모두 물리적으로 자연스럽습니다."
       }
      ],
      "all_candidates_fail": false
     },
     "normalized": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 8,
        "verdict_ko": "캐릭터 레퍼런스의 복장(모자 포함)과 인물의 굳은 표정을 클로즈업으로 정확히 구현했으나, 장소 레퍼런스에 없는 호롱불 조명이 배경에 추가되어 조명 락(lock)을 일부 위반했습니다."
       },
       {
        "label": "B",
        "score": 5,
        "verdict_ko": "배경의 조명과 분위기는 이전 샷과 잘 일치하나, 명시적인 금지 지시를 어기고 이전 샷에 등장한 인물의 삿갓을 복만에게 씌워 캐릭터 레퍼런스(의상)를 크게 위반했습니다."
       }
      ],
      "all_candidates_fail": false,
      "readings": [
       {
        "label": "A",
        "direction": "복만의 시선이 전경에 등지고 있는 인물(안규홍)의 뒤통수와 목을 향해 정확히 고정되어 있습니다.",
        "built_space": "야외 산채의 흙바닥으로, 배경에 나무 울타리와 보급품이 실린 수레가 배치되어 있습니다. 다만 장소 레퍼런스에는 없던 따뜻한 빛을 내는 호롱불 3개가 배경에 추가되어 있습니다.",
        "entities": "복만은 캐릭터 레퍼런스와 일치하는 갓 형태의 모자와 수염, 어두운 의상을 착용하고 있습니다. 전경에는 텍스트가 요구한 안규홍의 등이 오버더숄더 구도로 걸쳐 있으며, 복만은 짚으로 엮은 쌀가마니를 들고 있습니다.",
        "hard_violations": [],
        "physics": "복만의 두 손이 쌀가마니의 밧줄을 단단히 쥐고 무게를 지탱하고 있으며, 전경의 인물과 복만 모두 자연스러운 자세로 서 있습니다. 물리적으로 불가능한 요소는 없습니다."
       },
       {
        "label": "B",
        "direction": "복만의 시선이 전경에 등지고 있는 인물(안규홍)을 향해 고정되어 있습니다.",
        "built_space": "야외 산채 공간으로 나무 울타리와 수레, 바퀴가 배경에 보이며, 이전 샷의 어두운 조명 분위기를 호롱불 추가 없이 잘 유지하고 있습니다.",
        "entities": "전경에 안규홍의 등이 보이고 복만이 자루를 들고 있으나, 복만의 모자가 캐릭터 레퍼런스(갓 형태)가 아닌 이전 샷 레퍼런스 속 인물의 삿갓을 그대로 복사해 온 형태여서 의상 설정과 금지 지시를 위반했습니다.",
        "hard_violations": [],
        "physics": "복만의 오른손이 자루의 윗부분을 쥐고 무게를 버티고 있으며, 두 인물의 서 있는 자세와 땅과의 접촉 모두 물리적으로 자연스럽습니다."
       }
      ]
     },
     "ok": true
    },
    {
     "model": "gpt-high",
     "order": "reverse",
     "display_to_canonical": {
      "A": "B",
      "B": "A"
     },
     "raw": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 2,
        "verdict_ko": "복만이 상대의 등이 아니라 얼굴을 바라보는 대면 구도이며, 얼굴 클로즈업 대신 두 사람과 쌀자루를 넓게 보여 주고 허용 인물도 초과한다."
       },
       {
        "label": "B",
        "score": 3,
        "verdict_ko": "돌아선 상대를 뒤에서 응시하는 관계와 복만의 모자는 더 정확하지만, 허용되지 않은 인물과 새 등불을 추가하고 얼굴 클로즈업도 충분히 구현하지 못했다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "복만의 눈은 화면 오른쪽 앞에 있는 남자의 얼굴 쪽을 향한다. 그 남자는 카메라에 등을 보이지만 몸과 고개는 복만을 향하므로, 복만이 안규홍의 등을 응시하는 관계가 아니라 서로 마주 보는 관계다. 무기나 다른 지향성 소품은 보이지 않는다.",
        "built_space": "오른쪽에 짐을 실은 나무 수레 한 대와 바퀴 일부, 뒤쪽에 목책과 숲, 왼쪽 지면에 세워진 자루와 목재가 보인다. 산중 야영지의 재료와 밤 분위기는 대체로 이어진다. 다만 두 사람을 수레 바로 옆에 크게 배치했고, 복만의 얼굴보다 앞사람의 등이 훨씬 넓은 화면을 차지한다. 요구된 얼굴 클로즈업이 아니다.",
        "entities": "수염과 검은 머리를 지닌 한국인 성인 남성으로 설정된 복만 한 명이 보이며, 낡은 회갈색 겹옷은 참고와 유사하다. 얼굴은 대체로 닮았으나 모자는 참고의 높고 평평한 챙이 있는 형태 대신 원뿔형 삿갓이다. 오른쪽에는 밝은 옷을 입은 별도의 성인 남성이 등장한다. 이는 복만만 허용한 인물 목록과 충돌한다. 묶인 포대와 수레의 보급품은 보이지만 내용물이 쌀인지 감자인지는 판별할 수 없다. 읽을 수 있는 글자는 없다.",
        "hard_violations": [
         "복만만 화면에 허용한 인물 목록을 어기고 별도의 남성 상반신을 크게 추가했다."
        ],
        "physics": "복만은 상체를 굽히고 있으며 화면 왼쪽 손이 큰 포대의 묶인 윗부분을 잡고 있다. 포대를 내리다 멈춘 행동으로 가능한 자세이고, 포대 아래쪽과 다른 손은 화면 밖 또는 앞사람에 가려져 있다. 앞사람의 하체도 잘려 있으나 공중에 떠 있다는 징후는 없다. 수레와 뒤쪽 자루는 지면에 놓여 있다."
       },
       {
        "label": "B",
        "direction": "복만은 화면 오른쪽으로 시선을 고정하고 있으며, 앞 남자는 복만에게서 몸을 돌린 채 화면 오른쪽 아래를 보고 있다. 복만의 시선은 그 남자의 뒤통수와 목 뒤쪽을 향해 읽힌다. 등 자체보다 높은 지점이지만, A와 달리 돌아선 상대를 뒤에서 응시하는 관계는 성립한다.",
        "built_space": "왼쪽 뒤에 보급품을 실은 나무 수레 한 대, 그 뒤 지면에 쌓인 자루들, 목책과 숲이 보인다. 수레와 자루는 배경에 놓여 있으나 복만의 몸통과 앞사람의 머리·어깨까지 크게 포함하여 얼굴 중심 클로즈업보다 넓다. 배경에는 따뜻한 등불 네 곳이 새로 보이며, 참고의 억제된 야간 조명에 없던 반복적인 발광 지점을 만든다.",
        "entities": "복만은 한국인 성인 남성으로 읽히며 검은 머리, 거친 수염, 낡은 회갈색 옷을 갖췄다. 높이 솟은 모자와 넓은 챙은 A보다 인물 참고에 가깝다. 오른쪽 전경에는 검은 머리와 밝은 옷을 가진 별도의 성인 남성의 얼굴 일부와 상체가 보이므로 복만만 허용한 목록을 지키지 않았다. 아래에는 끈으로 묶인 포대, 배경에는 수레와 여러 보급품이 보인다. 내용물의 종류는 확인되지 않으며 읽을 수 있는 글자는 없다.",
        "hard_violations": [
         "복만만 화면에 허용한 인물 목록을 어기고 별도의 남성 얼굴 일부와 상반신을 추가했다.",
         "장소와 조명을 유지하고 새 조명 신호를 만들지 말라는 지시와 달리 배경에 발광하는 등불 네 곳을 추가했다."
        ],
        "physics": "복만은 몸을 약간 숙이고 두 팔을 아래 포대 쪽으로 내리고 있다. 포대 윗부분만 보이고 손의 접촉점과 바닥은 전경 인물 및 화면 경계에 가려져 있어, 내려놓다 멈춘 동작을 확증하기는 어렵지만 무지지 부유로 볼 근거도 없다. 앞사람은 자연스럽게 고개를 숙이고 돌아선 자세이며, 수레와 쌓인 자루는 지면에 지지된다."
       }
      ],
      "all_candidates_fail": true
     },
     "normalized": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 2,
        "verdict_ko": "복만이 상대의 등이 아니라 얼굴을 바라보는 대면 구도이며, 얼굴 클로즈업 대신 두 사람과 쌀자루를 넓게 보여 주고 허용 인물도 초과한다."
       },
       {
        "label": "A",
        "score": 3,
        "verdict_ko": "돌아선 상대를 뒤에서 응시하는 관계와 복만의 모자는 더 정확하지만, 허용되지 않은 인물과 새 등불을 추가하고 얼굴 클로즈업도 충분히 구현하지 못했다."
       }
      ],
      "all_candidates_fail": true,
      "readings": [
       {
        "label": "B",
        "direction": "복만의 눈은 화면 오른쪽 앞에 있는 남자의 얼굴 쪽을 향한다. 그 남자는 카메라에 등을 보이지만 몸과 고개는 복만을 향하므로, 복만이 안규홍의 등을 응시하는 관계가 아니라 서로 마주 보는 관계다. 무기나 다른 지향성 소품은 보이지 않는다.",
        "built_space": "오른쪽에 짐을 실은 나무 수레 한 대와 바퀴 일부, 뒤쪽에 목책과 숲, 왼쪽 지면에 세워진 자루와 목재가 보인다. 산중 야영지의 재료와 밤 분위기는 대체로 이어진다. 다만 두 사람을 수레 바로 옆에 크게 배치했고, 복만의 얼굴보다 앞사람의 등이 훨씬 넓은 화면을 차지한다. 요구된 얼굴 클로즈업이 아니다.",
        "entities": "수염과 검은 머리를 지닌 한국인 성인 남성으로 설정된 복만 한 명이 보이며, 낡은 회갈색 겹옷은 참고와 유사하다. 얼굴은 대체로 닮았으나 모자는 참고의 높고 평평한 챙이 있는 형태 대신 원뿔형 삿갓이다. 오른쪽에는 밝은 옷을 입은 별도의 성인 남성이 등장한다. 이는 복만만 허용한 인물 목록과 충돌한다. 묶인 포대와 수레의 보급품은 보이지만 내용물이 쌀인지 감자인지는 판별할 수 없다. 읽을 수 있는 글자는 없다.",
        "hard_violations": [
         "복만만 화면에 허용한 인물 목록을 어기고 별도의 남성 상반신을 크게 추가했다."
        ],
        "physics": "복만은 상체를 굽히고 있으며 화면 왼쪽 손이 큰 포대의 묶인 윗부분을 잡고 있다. 포대를 내리다 멈춘 행동으로 가능한 자세이고, 포대 아래쪽과 다른 손은 화면 밖 또는 앞사람에 가려져 있다. 앞사람의 하체도 잘려 있으나 공중에 떠 있다는 징후는 없다. 수레와 뒤쪽 자루는 지면에 놓여 있다."
       },
       {
        "label": "A",
        "direction": "복만은 화면 오른쪽으로 시선을 고정하고 있으며, 앞 남자는 복만에게서 몸을 돌린 채 화면 오른쪽 아래를 보고 있다. 복만의 시선은 그 남자의 뒤통수와 목 뒤쪽을 향해 읽힌다. 등 자체보다 높은 지점이지만, A와 달리 돌아선 상대를 뒤에서 응시하는 관계는 성립한다.",
        "built_space": "왼쪽 뒤에 보급품을 실은 나무 수레 한 대, 그 뒤 지면에 쌓인 자루들, 목책과 숲이 보인다. 수레와 자루는 배경에 놓여 있으나 복만의 몸통과 앞사람의 머리·어깨까지 크게 포함하여 얼굴 중심 클로즈업보다 넓다. 배경에는 따뜻한 등불 네 곳이 새로 보이며, 참고의 억제된 야간 조명에 없던 반복적인 발광 지점을 만든다.",
        "entities": "복만은 한국인 성인 남성으로 읽히며 검은 머리, 거친 수염, 낡은 회갈색 옷을 갖췄다. 높이 솟은 모자와 넓은 챙은 A보다 인물 참고에 가깝다. 오른쪽 전경에는 검은 머리와 밝은 옷을 가진 별도의 성인 남성의 얼굴 일부와 상체가 보이므로 복만만 허용한 목록을 지키지 않았다. 아래에는 끈으로 묶인 포대, 배경에는 수레와 여러 보급품이 보인다. 내용물의 종류는 확인되지 않으며 읽을 수 있는 글자는 없다.",
        "hard_violations": [
         "복만만 화면에 허용한 인물 목록을 어기고 별도의 남성 얼굴 일부와 상반신을 추가했다.",
         "장소와 조명을 유지하고 새 조명 신호를 만들지 말라는 지시와 달리 배경에 발광하는 등불 네 곳을 추가했다."
        ],
        "physics": "복만은 몸을 약간 숙이고 두 팔을 아래 포대 쪽으로 내리고 있다. 포대 윗부분만 보이고 손의 접촉점과 바닥은 전경 인물 및 화면 경계에 가려져 있어, 내려놓다 멈춘 동작을 확증하기는 어렵지만 무지지 부유로 볼 근거도 없다. 앞사람은 자연스럽게 고개를 숙이고 돌아선 자세이며, 수레와 쌓인 자루는 지면에 지지된다."
       }
      ]
     },
     "ok": true
    }
   ],
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "slot_winner_match": true,
   "slot_winner": {
    "gemini-pro": "A",
    "gpt-high": "A"
   },
   "route": "cross_slot_agree"
  },
  "dual": {
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "normalized": {
    "A": 2.0,
    "B": 1.292
   },
   "adjusted": {
    "A": 1.75,
    "B": 1.042
   },
   "violations": {
    "B": [
     "[gpt-high] 복만만 화면에 허용한 인물 목록을 어기고 별도의 남성 상반신을 크게 추가했다."
    ],
    "A": [
     "[gpt-high] 복만만 화면에 허용한 인물 목록을 어기고 별도의 남성 얼굴 일부와 상반신을 추가했다.",
     "[gpt-high] 장소와 조명을 유지하고 새 조명 신호를 만들지 말라는 지시와 달리 배경에 발광하는 등불 네 곳을 추가했다."
    ]
   },
   "per_model_winner": {
    "gemini-pro": "A",
    "gpt-high": "A"
   },
   "agreed": true
  },
  "totals": {
   "A": 1750,
   "B": 1042
  },
  "selected": "A",
  "ranking": [
   "A",
   "B"
  ],
  "verdicts": [
   {
    "label": "A",
    "score": 1750,
    "verdict_ko": "캐릭터 레퍼런스의 복장(모자 포함)과 인물의 굳은 표정을 클로즈업으로 정확히 구현했으나, 장소 레퍼런스에 없는 호롱불 조명이 배경에 추가되어 조명 락(lock)을 일부 위반했습니다.  ★위반: [gpt-high] 복만만 화면에 허용한 인물 목록을 어기고 별도의 남성 얼굴 일부와 상반신을 추가했다. / [gpt-high] 장소와 조명을 유지하고 새 조명 신호를 만들지 말라는 지시와 달리 배경에 발광하는 등불 네 곳을 추가했다."
   },
   {
    "label": "B",
    "score": 1042,
    "verdict_ko": "배경의 조명과 분위기는 이전 샷과 잘 일치하나, 명시적인 금지 지시를 어기고 이전 샷에 등장한 인물의 삿갓을 복만에게 씌워 캐릭터 레퍼런스(의상)를 크게 위반했습니다.  ★위반: [gpt-high] 복만만 화면에 허용한 인물 목록을 어기고 별도의 남성 상반신을 크게 추가했다."
   }
  ],
  "refs": [
   {
    "label": "PREVIOUS SHOT STILL — a visually related earlier shot of this same place: the location's look, materials, fixed features and lighting mood are LOCKED to this photo; never copy its camera framing. The people visible in that photo are NOT in this shot: never carry their faces, bodies or clothing onto anyone here. Each person in THIS shot is defined solely by the PEOPLE list, their own CHARACTER REFERENCE images and their wardrobe notes. If this photo shows a character who cannot move (dead or unconscious), that character's exact pose, position and orientation are ALSO LOCKED — treat their body as a fixed prop of the set that only the camera moves around.",
    "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/db00517b-165b-4005-a287-f28936b59b95/images/272bd7af-9294-4490-98c6-88a40e9deacc/scene/recipe/S36sh1_sel.png",
    "asset_id": "82b6c12a-3238-428f-9aca-67e714a522cd",
    "role": "prev_still"
   },
   {
    "label": "CHARACTER REFERENCE — 복만: the exact person appearing in this shot; match face, hair and build exactly.",
    "path": "<bytes:1027479>",
    "asset_id": "b7e69570-d15d-4241-b5ea-4cb8d4ece1cf",
    "role": "character_ref"
   }
  ],
  "critique_skipped": true,
  "shot_run_produced": false,
  "shot_run_uid": "06aa1428-57bb-7256-807d-1dcd4eed648e",
  "ref_mode": "prev+엔티티",
  "share_plan": {
   "ref_plan": "prev",
   "prev_anchor_tag": "S36sh1"
  }
 },
 "S36sh17::cine": {
  "applied": true,
  "attempted_at": "2026-09-09T04:58:18.585742+00:00",
  "fingerprint": "b0b7c2777892527442aaeec27a0023844c823f65d66f954122d715edbb824e8b",
  "fingerprint_version": 2,
  "provider": "mai",
  "endpoint": "openrouter/chat-completions",
  "model": "microsoft/mai-image-2.6",
  "pack": "24.202608252115",
  "source_file": "S36sh17_sel.png",
  "source_sha256": "d087436b858452cab69ae72b41961b6ff1c724b4e9b3924f093ec8cb3d2d9b59",
  "file": "S36sh17_cine.png",
  "staged_sha256": "371cca1b5ec01a9e17bebf5464b579a66e523ce5de31f2b4c0ee1dfc7e8672fa",
  "latency_ms": 32847
 },
 "S37sh2::signage": {
  "fp": "4d5948df3e32c8b8",
  "inscriptions": [],
  "cues": [
   {
    "text_native": "",
    "source": "scene_text_implied",
    "source_quote": "벽보"
   }
  ],
  "dropped": []
 },
 "S37sh2::bgfirst_bg": {
  "input_fingerprint": "02d47e2d9df185d2",
  "prompt": "Create the EMPTY BACKGROUND PLATE for one film shot — NO PEOPLE, no figures, no body parts, no sketch lines, no arrows anywhere.\n\"Empty\" means no people only: KEEP the location's inherent occupants and stock that define the place — animals in an animal shelter, pen or farm, goods in a market, moored boats in a harbour — unless the shot text explicitly removes them.\nThe FIRST attached image is a thin-line storyboard sketch: use ONLY its camera angle, horizon, perspective and the placement/size of buildings and set masses — ignore the sketched people and arrows entirely. The SECOND attached image (LOCATION PHOTOGRAPH) is the real place: take its architecture, materials, signage and fixed features, and RE-PROJECT them into the sketch's camera. If the photograph's camera differs from the sketch's, the sketch's camera wins.\nHUMAN-SCALE CALIBRATION: derive every structure's true size from human-scale elements — a door ≈ 2m, a window ≈ 1–1.5m wide, one storey ≈ 2.5–3m; never inflate a small structure or shrink a large one.\n\nSHOT TEXT this background must serve (Korean): 소금과 약재 꾸러미를 손에 쥔 채 벽보를 멍하니 올려다보는 복만의 뒷모습.\n\nLOCATION (lock): Directly before an outdoor market wall bearing a posted bounty notice, beside the passing daytime crowd.\n\nTIME OF DAY (lock): day.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: medium shot\n- FRAME LAYOUT: Written face of the reward notice above and ahead of 복만 in the upper-left of the frame, background.\n- KEY BACKGROUND ELEMENTS: Reward notice (Posted on the wall with 안규홍's name, a reward amount and a red official seal) — The written face is visible obliquely to the camera, including the name and the seal beneath it; used as Occupy less than a third of the upper-left image as the endpoint of 복만's upward attention; Market wall (Bearing the notice) — Seen obliquely beyond 복만's back; used as Form a simple background that separates the notice from his silhouette; Salt and medicinal ingredient bundles (Purchased and held in 복만's hands); used as Keep the ordinary errand visible at the lower edge beneath his arrested attention.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Daytime ambient light gives the scene readable, restrained contrast and muted saturation, with the notice's red official seal retained as written.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nRender ONE photorealistic empty location photograph, 16:9, that this shot can be staged inside later. No readable writing anywhere: surfaces that would carry writing may be present, but stage any wording out of legibility — an oblique angle, distance, shallow focus. No captions, watermarks or overlay text.",
  "effective_prompt": "Create the EMPTY BACKGROUND PLATE for one film shot — NO PEOPLE, no figures, no body parts, no sketch lines, no arrows anywhere.\n\"Empty\" means no people only: KEEP the location's inherent occupants and stock that define the place — animals in an animal shelter, pen or farm, goods in a market, moored boats in a harbour — unless the shot text explicitly removes them.\nThe FIRST attached image is a thin-line storyboard sketch: use ONLY its camera angle, horizon, perspective and the placement/size of buildings and set masses — ignore the sketched people and arrows entirely. The SECOND attached image (LOCATION PHOTOGRAPH) is the real place: take its architecture, materials, signage and fixed features, and RE-PROJECT them into the sketch's camera. If the photograph's camera differs from the sketch's, the sketch's camera wins.\nHUMAN-SCALE CALIBRATION: derive every structure's true size from human-scale elements — a door ≈ 2m, a window ≈ 1–1.5m wide, one storey ≈ 2.5–3m; never inflate a small structure or shrink a large one.\n\nSHOT TEXT this background must serve (Korean): 소금과 약재 꾸러미를 손에 쥔 채 벽보를 멍하니 올려다보는 복만의 뒷모습.\n\nLOCATION (lock): Directly before an outdoor market wall bearing a posted bounty notice, beside the passing daytime crowd.\n\nTIME OF DAY (lock): day.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: medium shot\n- FRAME LAYOUT: Written face of the reward notice above and ahead of 복만 in the upper-left of the frame, background.\n- KEY BACKGROUND ELEMENTS: Reward notice (Posted on the wall with 안규홍's name, a reward amount and a red official seal) — The written face is visible obliquely to the camera, including the name and the seal beneath it; used as Occupy less than a third of the upper-left image as the endpoint of 복만's upward attention; Market wall (Bearing the notice) — Seen obliquely beyond 복만's back; used as Form a simple background that separates the notice from his silhouette; Salt and medicinal ingredient bundles (Purchased and held in 복만's hands); used as Keep the ordinary errand visible at the lower edge beneath his arrested attention.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Daytime ambient light gives the scene readable, restrained contrast and muted saturation, with the notice's red official seal retained as written.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nRender ONE photorealistic empty location photograph, 16:9, that this shot can be staged inside later. No readable writing anywhere: surfaces that would carry writing may be present, but stage any wording out of legibility — an oblique angle, distance, shallow focus. No captions, watermarks or overlay text.",
  "bg_path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/db00517b-165b-4005-a287-f28936b59b95/images/272bd7af-9294-4490-98c6-88a40e9deacc/scene/recipe/S37sh2__bgfirst_bg.png",
  "asset_id": "87472aaf-7f44-408a-b805-dd0474cfa538",
  "input_asset_ids": [
   "e51f1080-c121-43ef-bdb2-c111ecfb3183",
   "32b543ac-1086-4282-b907-3fcd465b580a"
  ]
 },
 "S37sh2": {
  "input_fingerprint": "b1c50cade5e5f313",
  "prompt": "Create ONE FINAL photorealistic live-action film still of the moment below — early 20th-century Korea, Jeolla region; all people are Korean unless explicitly stated otherwise. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 소금과 약재 꾸러미를 손에 쥔 채 벽보를 멍하니 올려다보는 복만의 뒷모습.\n\nLOCATION (lock): Directly before an outdoor market wall bearing a posted bounty notice, beside the passing daytime crowd. The shot takes place here — the attached LOCATION STRUCTURE PHOTOGRAPH is the single authority for this exact place — its fixed structure and permanent site details are LOCKED to it. No separate location photograph exists for this place. Build everything else strictly from the location text above and the shot text; the layout sketch (when attached) governs framing and placement only, and the shot text governs time of day, lighting and action.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: medium shot\n- FRAME LAYOUT: Written face of the reward notice above and ahead of 복만 in the upper-left of the frame, background.\n- KEY BACKGROUND ELEMENTS: Reward notice (Posted on the wall with 안규홍's name, a reward amount and a red official seal) — The written face is visible obliquely to the camera, including the name and the seal beneath it; used as Occupy less than a third of the upper-left image as the endpoint of 복만's upward attention; Market wall (Bearing the notice) — Seen obliquely beyond 복만's back; used as Form a simple background that separates the notice from his silhouette; Salt and medicinal ingredient bundles (Purchased and held in 복만's hands); used as Keep the ordinary errand visible at the lower edge beneath his arrested attention.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Daytime ambient light gives the scene readable, restrained contrast and muted saturation, with the notice's red official seal retained as written.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): A wanted notice on the market wall bears the name '안규홍', a bounty amount and a conspicuous red official seal. 복만: He stands carrying purchased salt and medicinal herbs, in his worn durumagi.\n\nTHE HAND THAT IS DOING THIS: the object at the centre of this shot is being held, operated, read, aimed or handed over by 복만 right now, so 복만's hand — and as much of the wrist and forearm as the framing reaches — is in the frame, gripping or resting on that object exactly the way its real use requires. Match that hand to 복만: its size, build, skin, age, grooming, sleeve and anything worn on it belong to that person and to no one else.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 복만 (한국인 남성, 성인 나이대, 거칠게 자란 수염, 자연스러운 검은 머리) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.",
  "roll_prompts": {
   "A": "Stage the shot. The FIRST attached image (SHOT BACKGROUND) is the finished empty background of this shot — keep it EXACTLY: its camera, perspective, architecture, lighting and every fixed feature stay untouched. The SECOND attached image (LAYOUT SKETCH) tells you ONLY where the people go: each sketched person's position, screen size, pose and the gaze/motion arrows. Ignore the sketch's background lines. The CHARACTER REFERENCE photographs show the real people.\nPlace the real people into the background at exactly the sketched positions, sizes and poses, following the arrow directions. No sketch lines or arrows may remain.\n\nCreate ONE FINAL photorealistic live-action film still of the moment below — early 20th-century Korea, Jeolla region; all people are Korean unless explicitly stated otherwise. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 소금과 약재 꾸러미를 손에 쥔 채 벽보를 멍하니 올려다보는 복만의 뒷모습.\n\nLOCATION (lock): Directly before an outdoor market wall bearing a posted bounty notice, beside the passing daytime crowd. The shot takes place here — the FIRST attached image (SHOT BACKGROUND) is this exact place, already built: its ground, structures, horizon, materials and lighting are the finished truth of this location and must not be redesigned or replaced. No location photograph is attached — read the place from that image alone, and add no scenery, structure, vehicle or fixture that it does not already show. This lock governs the place only; the figures in the shot follow the staging and pose instructions.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: medium shot\n- FRAME LAYOUT: Written face of the reward notice above and ahead of 복만 in the upper-left of the frame, background.\n- KEY BACKGROUND ELEMENTS: Reward notice (Posted on the wall with 안규홍's name, a reward amount and a red official seal) — The written face is visible obliquely to the camera, including the name and the seal beneath it; used as Occupy less than a third of the upper-left image as the endpoint of 복만's upward attention; Market wall (Bearing the notice) — Seen obliquely beyond 복만's back; used as Form a simple background that separates the notice from his silhouette; Salt and medicinal ingredient bundles (Purchased and held in 복만's hands); used as Keep the ordinary errand visible at the lower edge beneath his arrested attention.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Daytime ambient light gives the scene readable, restrained contrast and muted saturation, with the notice's red official seal retained as written.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): A wanted notice on the market wall bears the name '안규홍', a bounty amount and a conspicuous red official seal. 복만: He stands carrying purchased salt and medicinal herbs, in his worn durumagi.\n\nTHE HAND THAT IS DOING THIS: the object at the centre of this shot is being held, operated, read, aimed or handed over by 복만 right now, so 복만's hand — and as much of the wrist and forearm as the framing reaches — is in the frame, gripping or resting on that object exactly the way its real use requires. Match that hand to 복만: its size, build, skin, age, grooming, sleeve and anything worn on it belong to that person and to no one else.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 복만 (한국인 남성, 성인 나이대, 거칠게 자란 수염, 자연스러운 검은 머리) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.",
   "B": "Create ONE FINAL photorealistic live-action film still of the moment below — early 20th-century Korea, Jeolla region; all people are Korean unless explicitly stated otherwise. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 소금과 약재 꾸러미를 손에 쥔 채 벽보를 멍하니 올려다보는 복만의 뒷모습.\n\nLOCATION (lock): Directly before an outdoor market wall bearing a posted bounty notice, beside the passing daytime crowd. The shot takes place here — the attached LOCATION STRUCTURE PHOTOGRAPH is the single authority for this exact place — its fixed structure and permanent site details are LOCKED to it. No separate location photograph exists for this place. Build everything else strictly from the location text above and the shot text; the layout sketch (when attached) governs framing and placement only, and the shot text governs time of day, lighting and action.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: medium shot\n- FRAME LAYOUT: Written face of the reward notice above and ahead of 복만 in the upper-left of the frame, background.\n- KEY BACKGROUND ELEMENTS: Reward notice (Posted on the wall with 안규홍's name, a reward amount and a red official seal) — The written face is visible obliquely to the camera, including the name and the seal beneath it; used as Occupy less than a third of the upper-left image as the endpoint of 복만's upward attention; Market wall (Bearing the notice) — Seen obliquely beyond 복만's back; used as Form a simple background that separates the notice from his silhouette; Salt and medicinal ingredient bundles (Purchased and held in 복만's hands); used as Keep the ordinary errand visible at the lower edge beneath his arrested attention.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Daytime ambient light gives the scene readable, restrained contrast and muted saturation, with the notice's red official seal retained as written.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): A wanted notice on the market wall bears the name '안규홍', a bounty amount and a conspicuous red official seal. 복만: He stands carrying purchased salt and medicinal herbs, in his worn durumagi.\n\nTHE HAND THAT IS DOING THIS: the object at the centre of this shot is being held, operated, read, aimed or handed over by 복만 right now, so 복만's hand — and as much of the wrist and forearm as the framing reaches — is in the frame, gripping or resting on that object exactly the way its real use requires. Match that hand to 복만: its size, build, skin, age, grooming, sleeve and anything worn on it belong to that person and to no one else.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 복만 (한국인 남성, 성인 나이대, 거칠게 자란 수염, 자연스러운 검은 머리) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.\n\nCOMPOSITION VARIATION (this candidate only): keep the place, the people and the moment exactly as specified, and keep from the CAMERA & FRAME contract its subject scale, screen placement and key background elements exactly. For this candidate only, its camera angle and height are deliberately left open: choose a camera position distinctly different from the obvious one, as a film director picking a second setup on the same blocking. Stay true to any explicit framing the SHOT TEXT itself dictates."
  },
  "roll_refs": {
   "A": [
    {
     "label": "SHOT BACKGROUND",
     "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/db00517b-165b-4005-a287-f28936b59b95/images/272bd7af-9294-4490-98c6-88a40e9deacc/scene/recipe/S37sh2__bgfirst_bg.png",
     "asset_id": "87472aaf-7f44-408a-b805-dd0474cfa538",
     "role": "bgfirst_bg"
    },
    {
     "label": "LAYOUT SKETCH (people placement only)",
     "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/db00517b-165b-4005-a287-f28936b59b95/images/272bd7af-9294-4490-98c6-88a40e9deacc/conti/conti_S37sh2.png",
     "asset_id": "e51f1080-c121-43ef-bdb2-c111ecfb3183",
     "role": "conti_light"
    },
    {
     "label": "CHARACTER REFERENCE — 복만: the exact person appearing in this shot; match face, hair and build exactly.",
     "path": "<bytes:1027479>",
     "asset_id": "b7e69570-d15d-4241-b5ea-4cb8d4ece1cf",
     "role": "character_ref"
    }
   ],
   "B": [
    {
     "label": "LOCATION STRUCTURE PHOTOGRAPH — the confirmed photograph of this exact place and its fixed structure: it is the SINGLE authority for the location, the structure's shape, proportions, materials, colors, openings and every permanent site detail. Never copy its camera framing, time of day or lighting — the shot text is the authority for those.",
     "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/db00517b-165b-4005-a287-f28936b59b95/images/background_chain/seed_bg_market_sel.png",
     "asset_id": "32b543ac-1086-4282-b907-3fcd465b580a",
     "role": "location_seed_bg"
    },
    {
     "label": "CHARACTER REFERENCE — 복만: the exact person appearing in this shot; match face, hair and build exactly.",
     "path": "<bytes:1027479>",
     "asset_id": "b7e69570-d15d-4241-b5ea-4cb8d4ece1cf",
     "role": "character_ref"
    }
   ]
  },
  "shot_run_spend_attempt_count": 1,
  "judge_flip_skipped": "judge_fn_owns_order",
  "readings": [
   {
    "label": "A",
    "direction": "복만의 고개와 시선이 화면 왼쪽 상단 벽면에 붙은 벽보를 향해 올려다보고 있습니다.",
    "built_space": "레퍼런스의 건축 양식과 일치하는 초가집들과 흙벽이 늘어선 야외 길거리입니다. 벽면은 화면 왼쪽에 비스듬히 배치되어 있으나, 지시된 '지나가는 군중'은 보이지 않고 거리가 비어 있습니다.",
    "entities": "복만의 뒷모습, 수염, 두루마기 의상이 레퍼런스와 일치합니다. 양손 하단에 하얀 천으로 싸인 꾸러미(소금과 약재로 추정)를 들고 있습니다. 벽보에는 붉은 인장과 함께 읽을 수 없도록 뭉개진 형태의 글귀가 있습니다.",
    "hard_violations": [],
    "physics": "복만은 땅에 안정적으로 서 있으며, 양손의 악력으로 꾸러미의 무게를 쥐고 지탱하고 있습니다."
   },
   {
    "label": "B",
    "direction": "복만의 시선과 얼굴 방향이 화면 왼쪽의 벽보를 향해 고정되어 있습니다.",
    "built_space": "벽면이 화면 왼쪽에 있고 뒤로 초가집 지붕과 시장 거리가 펼쳐져 있습니다. 배경에 한복을 입고 지나다니는 군중들이 잘 배치되어 있습니다.",
    "entities": "복만의 외형과 의상이 레퍼런스와 일치합니다. 손에 하얀 소금이 보이는 짚 묶음과 푸른 잎의 약재 꾸러미를 명확히 들고 있습니다. 그러나 벽보에 '접성간 안규홍...' 이라는 완벽히 읽을 수 있는 이질적인 디지털 폰트가 뚜렷하게 적혀 있습니다.",
    "hard_violations": [
     "[gemini-pro] 벽보 표면의 곡률을 무시한 채 화면에 납작하게 타이핑된 듯한 선명한 한글 텍스트가 존재하여, '읽을 수 있는 글귀 금지(No readable writing)' 및 그래픽 오버레이 금지 제약을 명백히 위반했습니다.",
     "[gpt-high] 복만 외 인물을 보이지 말라는 명시적 제한과 달리 배경에 다수의 행인이 추가되었다.",
     "[gpt-high] 최종 지시가 모든 글을 판독 불가능하게 요구하지만 벽보의 이름·금액·여러 문구가 선명하게 노출되었다."
    ],
    "physics": "복만은 땅을 딛고 서 있으며, 팔과 손으로 소금과 약재 꾸러미를 감싸듯 안아 물리적으로 지탱하고 있습니다."
   }
  ],
  "cross_model_order": {
   "policy": "cross_model_order_v2_slot0_fwd_slot1_rev_parallel2",
   "slots": [
    {
     "model": "gemini-pro",
     "order": "forward",
     "display_to_canonical": {
      "A": "A",
      "B": "B"
     },
     "raw": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 8,
        "verdict_ko": "지시된 카메라 구도와 비스듬한 앵글을 잘 구현했으며 텍스트를 읽을 수 없게 처리하라는 제약을 완벽히 준수했으나, 배경에 명시된 군중이 누락되었고 꾸러미의 내용물(약재) 형태가 불분명한 점이 아쉽습니다."
       },
       {
        "label": "B",
        "score": 3,
        "verdict_ko": "시장 군중과 소금, 약재 꾸러미의 세부 묘사는 훌륭하나, 벽보에 현대적인 디지털 폰트 형태의 선명한 한글이 오버레이처럼 배치되어 '읽을 수 있는 글씨 금지' 및 '질감 사실성' 제약을 심각하게 위반했습니다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "복만의 고개와 시선이 화면 왼쪽 상단 벽면에 붙은 벽보를 향해 올려다보고 있습니다.",
        "built_space": "레퍼런스의 건축 양식과 일치하는 초가집들과 흙벽이 늘어선 야외 길거리입니다. 벽면은 화면 왼쪽에 비스듬히 배치되어 있으나, 지시된 '지나가는 군중'은 보이지 않고 거리가 비어 있습니다.",
        "entities": "복만의 뒷모습, 수염, 두루마기 의상이 레퍼런스와 일치합니다. 양손 하단에 하얀 천으로 싸인 꾸러미(소금과 약재로 추정)를 들고 있습니다. 벽보에는 붉은 인장과 함께 읽을 수 없도록 뭉개진 형태의 글귀가 있습니다.",
        "hard_violations": [],
        "physics": "복만은 땅에 안정적으로 서 있으며, 양손의 악력으로 꾸러미의 무게를 쥐고 지탱하고 있습니다."
       },
       {
        "label": "B",
        "direction": "복만의 시선과 얼굴 방향이 화면 왼쪽의 벽보를 향해 고정되어 있습니다.",
        "built_space": "벽면이 화면 왼쪽에 있고 뒤로 초가집 지붕과 시장 거리가 펼쳐져 있습니다. 배경에 한복을 입고 지나다니는 군중들이 잘 배치되어 있습니다.",
        "entities": "복만의 외형과 의상이 레퍼런스와 일치합니다. 손에 하얀 소금이 보이는 짚 묶음과 푸른 잎의 약재 꾸러미를 명확히 들고 있습니다. 그러나 벽보에 '접성간 안규홍...' 이라는 완벽히 읽을 수 있는 이질적인 디지털 폰트가 뚜렷하게 적혀 있습니다.",
        "hard_violations": [
         "벽보 표면의 곡률을 무시한 채 화면에 납작하게 타이핑된 듯한 선명한 한글 텍스트가 존재하여, '읽을 수 있는 글귀 금지(No readable writing)' 및 그래픽 오버레이 금지 제약을 명백히 위반했습니다."
        ],
        "physics": "복만은 땅을 딛고 서 있으며, 팔과 손으로 소금과 약재 꾸러미를 감싸듯 안아 물리적으로 지탱하고 있습니다."
       }
      ],
      "all_candidates_fail": false
     },
     "normalized": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 8,
        "verdict_ko": "지시된 카메라 구도와 비스듬한 앵글을 잘 구현했으며 텍스트를 읽을 수 없게 처리하라는 제약을 완벽히 준수했으나, 배경에 명시된 군중이 누락되었고 꾸러미의 내용물(약재) 형태가 불분명한 점이 아쉽습니다."
       },
       {
        "label": "B",
        "score": 3,
        "verdict_ko": "시장 군중과 소금, 약재 꾸러미의 세부 묘사는 훌륭하나, 벽보에 현대적인 디지털 폰트 형태의 선명한 한글이 오버레이처럼 배치되어 '읽을 수 있는 글씨 금지' 및 '질감 사실성' 제약을 심각하게 위반했습니다."
       }
      ],
      "all_candidates_fail": false,
      "readings": [
       {
        "label": "A",
        "direction": "복만의 고개와 시선이 화면 왼쪽 상단 벽면에 붙은 벽보를 향해 올려다보고 있습니다.",
        "built_space": "레퍼런스의 건축 양식과 일치하는 초가집들과 흙벽이 늘어선 야외 길거리입니다. 벽면은 화면 왼쪽에 비스듬히 배치되어 있으나, 지시된 '지나가는 군중'은 보이지 않고 거리가 비어 있습니다.",
        "entities": "복만의 뒷모습, 수염, 두루마기 의상이 레퍼런스와 일치합니다. 양손 하단에 하얀 천으로 싸인 꾸러미(소금과 약재로 추정)를 들고 있습니다. 벽보에는 붉은 인장과 함께 읽을 수 없도록 뭉개진 형태의 글귀가 있습니다.",
        "hard_violations": [],
        "physics": "복만은 땅에 안정적으로 서 있으며, 양손의 악력으로 꾸러미의 무게를 쥐고 지탱하고 있습니다."
       },
       {
        "label": "B",
        "direction": "복만의 시선과 얼굴 방향이 화면 왼쪽의 벽보를 향해 고정되어 있습니다.",
        "built_space": "벽면이 화면 왼쪽에 있고 뒤로 초가집 지붕과 시장 거리가 펼쳐져 있습니다. 배경에 한복을 입고 지나다니는 군중들이 잘 배치되어 있습니다.",
        "entities": "복만의 외형과 의상이 레퍼런스와 일치합니다. 손에 하얀 소금이 보이는 짚 묶음과 푸른 잎의 약재 꾸러미를 명확히 들고 있습니다. 그러나 벽보에 '접성간 안규홍...' 이라는 완벽히 읽을 수 있는 이질적인 디지털 폰트가 뚜렷하게 적혀 있습니다.",
        "hard_violations": [
         "벽보 표면의 곡률을 무시한 채 화면에 납작하게 타이핑된 듯한 선명한 한글 텍스트가 존재하여, '읽을 수 있는 글귀 금지(No readable writing)' 및 그래픽 오버레이 금지 제약을 명백히 위반했습니다."
        ],
        "physics": "복만은 땅을 딛고 서 있으며, 팔과 손으로 소금과 약재 꾸러미를 감싸듯 안아 물리적으로 지탱하고 있습니다."
       }
      ]
     },
     "ok": true
    },
    {
     "model": "gpt-high",
     "order": "reverse",
     "display_to_canonical": {
      "A": "B",
      "B": "A"
     },
     "raw": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 2,
        "verdict_ko": "뒷모습 중심의 미디엄 숏과 벽보를 올려다보는 동작은 맞지만, 금지된 추가 인물들과 선명하게 읽히는 벽보 문구 때문에 탈락이다."
       },
       {
        "label": "B",
        "score": 7,
        "verdict_ko": "복만만 등장하고 판독 불가능한 벽보를 올려다보며 양손에 꾸러미를 든 관계는 맞지만, 미디엄 숏보다 넓은 구도와 달라진 담장 구조가 아쉽다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "복만은 등을 보인 채 머리를 왼쪽 위로 들고 있으며, 그 방향에 벽보의 글면이 있다. 눈 자체는 가려져 있지만 머리 방향은 벽보를 향한다. 벽보의 앞면은 복만과 카메라 양쪽에서 볼 수 있는 방향으로 벽에 붙어 있다. 배경 인물들은 골목의 서로 다른 방향으로 이동한다.",
        "built_space": "왼쪽 흙벽에 벽보 한 장이 있고, 복만 앞쪽에 세로 목재 기둥 하나와 위쪽 초가 처마가 보인다. 오른쪽에는 초가 건물과 판매대가 있다. 흙벽·목재·초가라는 재료는 참고와 유사하지만, 참고 장소의 정확한 건물 배치까지 일치한다고 확인하기는 어렵다. 복만은 벽 앞에 서 있고 벽보는 상단 왼쪽의 작은 면적을 차지한다. 반사는 없다.",
        "entities": "복만으로 보이는 성인 남성 한 명은 검은 머리, 옆으로 보이는 거친 수염, 낡은 회색 두루마기를 갖췄다. 참고의 모자는 없고, 뒷모습이라 얼굴의 정확한 동일성은 판단하기 어렵다. 아래쪽에 흰 내용물이 보이는 짚 포장 꾸러미들과 잎·줄기가 있는 약초 묶음이 보인다. 벽보에는 ‘안규홍’, ‘1,000원’과 다른 문구가 선명하게 읽히고 붉은 사각 인장이 있다. 이는 마지막의 판독 가능한 글자 금지 조건에 어긋난다. 배경에는 허용되지 않은 성인 남녀가 최소 일곱 명 더 보인다.",
        "hard_violations": [
         "복만 외 인물을 보이지 말라는 명시적 제한과 달리 배경에 다수의 행인이 추가되었다.",
         "최종 지시가 모든 글을 판독 불가능하게 요구하지만 벽보의 이름·금액·여러 문구가 선명하게 노출되었다."
        ],
        "physics": "복만은 상체를 세운 정지 자세이며, 발은 미디엄 숏 바깥에 있어 지면 접촉을 직접 확인할 수 없다. 왼손이 짚 꾸러미를 받치고 팔이 몸 앞으로 굽어 있어 무게를 지탱하는 관계가 보인다. 오른쪽 약초 묶음은 몸과 팔 가까이에 있으나 손의 접촉부는 가려져 있다. 공중에 떠 있다고 볼 근거는 없다. 벽보는 벽면에 밀착되어 있다."
       },
       {
        "label": "B",
        "direction": "복만은 카메라에 등을 보이고 고개를 왼쪽 위로 젖혀, 앞쪽 벽의 높은 곳에 붙은 벽보를 향한다. 눈은 보이지 않지만 머리와 얼굴 옆선의 방향은 벽보에 맞는다. 벽보의 글면은 복만을 향하면서 카메라에는 비스듬하게 보인다.",
        "built_space": "전경에는 흙벽 한 면, 이를 구획하는 세로 목재 두 개, 돌 기단, 초가 처마와 벽보 한 장이 보인다. 복만은 벽에서 조금 떨어진 길 위에 서 있다. 오른쪽에는 초가 건물들이 이어지고 목재 덮개 형태의 구조물도 보인다. 참고의 흙벽·초가 재료는 따르지만, 전경의 높고 굵은 돌 기단과 일렬로 반복되는 건물 구성은 참고 장소와 차이가 있다. 벽보의 상단 왼쪽 배치는 맞으나 인물을 허벅지 부근까지 포함하고 길을 넓게 보여 미디엄 숏보다 넓다. 반사는 없다.",
        "entities": "복만으로 보이는 성인 남성 한 명만 등장한다. 검은 머리, 옆으로 드러난 수염, 허리띠를 두른 낡은 회색 두루마기는 인물 조건과 대체로 맞지만 참고의 모자는 없다. 양손에는 각각 묶인 꾸러미가 보이며 내용물 대부분은 하단에서 잘려 소금과 약재인지 확정할 수 없다. 벽보 한 장에 흐린 세로 필획과 붉은 원형 인장이 있지만 이름과 금액은 읽히지 않아 최종 비가독성 지시는 충족한다. 인장의 붉은색은 다소 희미하다.",
        "hard_violations": [],
        "physics": "복만의 몸통은 정상적으로 세워져 있고, 양팔이 아래로 내려와 각 손으로 꾸러미의 묶인 부분을 쥔다. 꾸러미는 손 아래로 처져 있어 중력과 지지 관계가 자연스럽다. 발은 화면 밖이므로 접지는 직접 확인할 수 없지만 부유하거나 비정상적으로 기울어진 자세는 아니다. 벽보는 네 모서리의 고정점으로 벽에 붙어 있다."
       }
      ],
      "all_candidates_fail": false
     },
     "normalized": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 2,
        "verdict_ko": "뒷모습 중심의 미디엄 숏과 벽보를 올려다보는 동작은 맞지만, 금지된 추가 인물들과 선명하게 읽히는 벽보 문구 때문에 탈락이다."
       },
       {
        "label": "A",
        "score": 7,
        "verdict_ko": "복만만 등장하고 판독 불가능한 벽보를 올려다보며 양손에 꾸러미를 든 관계는 맞지만, 미디엄 숏보다 넓은 구도와 달라진 담장 구조가 아쉽다."
       }
      ],
      "all_candidates_fail": false,
      "readings": [
       {
        "label": "B",
        "direction": "복만은 등을 보인 채 머리를 왼쪽 위로 들고 있으며, 그 방향에 벽보의 글면이 있다. 눈 자체는 가려져 있지만 머리 방향은 벽보를 향한다. 벽보의 앞면은 복만과 카메라 양쪽에서 볼 수 있는 방향으로 벽에 붙어 있다. 배경 인물들은 골목의 서로 다른 방향으로 이동한다.",
        "built_space": "왼쪽 흙벽에 벽보 한 장이 있고, 복만 앞쪽에 세로 목재 기둥 하나와 위쪽 초가 처마가 보인다. 오른쪽에는 초가 건물과 판매대가 있다. 흙벽·목재·초가라는 재료는 참고와 유사하지만, 참고 장소의 정확한 건물 배치까지 일치한다고 확인하기는 어렵다. 복만은 벽 앞에 서 있고 벽보는 상단 왼쪽의 작은 면적을 차지한다. 반사는 없다.",
        "entities": "복만으로 보이는 성인 남성 한 명은 검은 머리, 옆으로 보이는 거친 수염, 낡은 회색 두루마기를 갖췄다. 참고의 모자는 없고, 뒷모습이라 얼굴의 정확한 동일성은 판단하기 어렵다. 아래쪽에 흰 내용물이 보이는 짚 포장 꾸러미들과 잎·줄기가 있는 약초 묶음이 보인다. 벽보에는 ‘안규홍’, ‘1,000원’과 다른 문구가 선명하게 읽히고 붉은 사각 인장이 있다. 이는 마지막의 판독 가능한 글자 금지 조건에 어긋난다. 배경에는 허용되지 않은 성인 남녀가 최소 일곱 명 더 보인다.",
        "hard_violations": [
         "복만 외 인물을 보이지 말라는 명시적 제한과 달리 배경에 다수의 행인이 추가되었다.",
         "최종 지시가 모든 글을 판독 불가능하게 요구하지만 벽보의 이름·금액·여러 문구가 선명하게 노출되었다."
        ],
        "physics": "복만은 상체를 세운 정지 자세이며, 발은 미디엄 숏 바깥에 있어 지면 접촉을 직접 확인할 수 없다. 왼손이 짚 꾸러미를 받치고 팔이 몸 앞으로 굽어 있어 무게를 지탱하는 관계가 보인다. 오른쪽 약초 묶음은 몸과 팔 가까이에 있으나 손의 접촉부는 가려져 있다. 공중에 떠 있다고 볼 근거는 없다. 벽보는 벽면에 밀착되어 있다."
       },
       {
        "label": "A",
        "direction": "복만은 카메라에 등을 보이고 고개를 왼쪽 위로 젖혀, 앞쪽 벽의 높은 곳에 붙은 벽보를 향한다. 눈은 보이지 않지만 머리와 얼굴 옆선의 방향은 벽보에 맞는다. 벽보의 글면은 복만을 향하면서 카메라에는 비스듬하게 보인다.",
        "built_space": "전경에는 흙벽 한 면, 이를 구획하는 세로 목재 두 개, 돌 기단, 초가 처마와 벽보 한 장이 보인다. 복만은 벽에서 조금 떨어진 길 위에 서 있다. 오른쪽에는 초가 건물들이 이어지고 목재 덮개 형태의 구조물도 보인다. 참고의 흙벽·초가 재료는 따르지만, 전경의 높고 굵은 돌 기단과 일렬로 반복되는 건물 구성은 참고 장소와 차이가 있다. 벽보의 상단 왼쪽 배치는 맞으나 인물을 허벅지 부근까지 포함하고 길을 넓게 보여 미디엄 숏보다 넓다. 반사는 없다.",
        "entities": "복만으로 보이는 성인 남성 한 명만 등장한다. 검은 머리, 옆으로 드러난 수염, 허리띠를 두른 낡은 회색 두루마기는 인물 조건과 대체로 맞지만 참고의 모자는 없다. 양손에는 각각 묶인 꾸러미가 보이며 내용물 대부분은 하단에서 잘려 소금과 약재인지 확정할 수 없다. 벽보 한 장에 흐린 세로 필획과 붉은 원형 인장이 있지만 이름과 금액은 읽히지 않아 최종 비가독성 지시는 충족한다. 인장의 붉은색은 다소 희미하다.",
        "hard_violations": [],
        "physics": "복만의 몸통은 정상적으로 세워져 있고, 양팔이 아래로 내려와 각 손으로 꾸러미의 묶인 부분을 쥔다. 꾸러미는 손 아래로 처져 있어 중력과 지지 관계가 자연스럽다. 발은 화면 밖이므로 접지는 직접 확인할 수 없지만 부유하거나 비정상적으로 기울어진 자세는 아니다. 벽보는 네 모서리의 고정점으로 벽에 붙어 있다."
       }
      ]
     },
     "ok": true
    }
   ],
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "slot_winner_match": true,
   "slot_winner": {
    "gemini-pro": "A",
    "gpt-high": "A"
   },
   "route": "cross_slot_agree"
  },
  "dual": {
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "normalized": {
    "A": 2.0,
    "B": 0.661
   },
   "adjusted": {
    "A": 2.0,
    "B": 0.411
   },
   "violations": {
    "B": [
     "[gemini-pro] 벽보 표면의 곡률을 무시한 채 화면에 납작하게 타이핑된 듯한 선명한 한글 텍스트가 존재하여, '읽을 수 있는 글귀 금지(No readable writing)' 및 그래픽 오버레이 금지 제약을 명백히 위반했습니다.",
     "[gpt-high] 복만 외 인물을 보이지 말라는 명시적 제한과 달리 배경에 다수의 행인이 추가되었다.",
     "[gpt-high] 최종 지시가 모든 글을 판독 불가능하게 요구하지만 벽보의 이름·금액·여러 문구가 선명하게 노출되었다."
    ]
   },
   "per_model_winner": {
    "gemini-pro": "A",
    "gpt-high": "A"
   },
   "agreed": true
  },
  "totals": {
   "A": 2000,
   "B": 411
  },
  "selected": "A",
  "ranking": [
   "A",
   "B"
  ],
  "verdicts": [
   {
    "label": "A",
    "score": 2000,
    "verdict_ko": "지시된 카메라 구도와 비스듬한 앵글을 잘 구현했으며 텍스트를 읽을 수 없게 처리하라는 제약을 완벽히 준수했으나, 배경에 명시된 군중이 누락되었고 꾸러미의 내용물(약재) 형태가 불분명한 점이 아쉽습니다."
   },
   {
    "label": "B",
    "score": 411,
    "verdict_ko": "시장 군중과 소금, 약재 꾸러미의 세부 묘사는 훌륭하나, 벽보에 현대적인 디지털 폰트 형태의 선명한 한글이 오버레이처럼 배치되어 '읽을 수 있는 글씨 금지' 및 '질감 사실성' 제약을 심각하게 위반했습니다.  ★위반: [gemini-pro] 벽보 표면의 곡률을 무시한 채 화면에 납작하게 타이핑된 듯한 선명한 한글 텍스트가 존재하여, '읽을 수 있는 글귀 금지(No readable writing)' 및 그래픽 오버레이 금지 제약을 명백히 위반했습니다. / [gpt-high] 복만 외 인물을 보이지 말라는 명시적 제한과 달리 배경에 다수의 행인이 추가되었다. / [gpt-high] 최종 지시가 모든 글을 판독 불가능하게 요구하지만 벽보의 이름·금액·여러 문구가 선명하게 노출되었다."
   }
  ],
  "refs": [
   {
    "label": "LOCATION STRUCTURE PHOTOGRAPH — the confirmed photograph of this exact place and its fixed structure: it is the SINGLE authority for the location, the structure's shape, proportions, materials, colors, openings and every permanent site detail. Never copy its camera framing, time of day or lighting — the shot text is the authority for those.",
    "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/db00517b-165b-4005-a287-f28936b59b95/images/background_chain/seed_bg_market_sel.png",
    "asset_id": "32b543ac-1086-4282-b907-3fcd465b580a",
    "role": "location_seed_bg"
   },
   {
    "label": "CHARACTER REFERENCE — 복만: the exact person appearing in this shot; match face, hair and build exactly.",
    "path": "<bytes:1027479>",
    "asset_id": "b7e69570-d15d-4241-b5ea-4cb8d4ece1cf",
    "role": "character_ref"
   }
  ],
  "critique_skipped": true,
  "shot_run_produced": false,
  "shot_run_uid": "06aa1428-595e-71d2-8d4b-3640f9f8098a",
  "bgfirst": {
   "bg_path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/db00517b-165b-4005-a287-f28936b59b95/images/272bd7af-9294-4490-98c6-88a40e9deacc/scene/recipe/S37sh2__bgfirst_bg.png",
   "bg_asset_id": "87472aaf-7f44-408a-b805-dd0474cfa538",
   "bg_record_key": "S37sh2::bgfirst_bg",
   "chain_winner": true,
   "authority": "seed_bg"
  },
  "ref_mode": "재투영 배경+콘티+엔티티 (2택1: 체인 승)",
  "share_plan": {
   "ref_plan": "background"
  },
  "lane_policy": "ab_select_ready"
 },
 "S37sh2::cine": {
  "applied": true,
  "attempted_at": "2026-09-09T05:01:33.497362+00:00",
  "fingerprint": "0872124dada6a53f2b1f3544e025f85b627bb6653a73fdd5f6169162900db783",
  "fingerprint_version": 2,
  "provider": "mai",
  "endpoint": "openrouter/chat-completions",
  "model": "microsoft/mai-image-2.6",
  "pack": "24.202608252115",
  "source_file": "S37sh2_sel.png",
  "source_sha256": "59e850e2584d9d4ecaaa9fd9486dac92872dd8566c7ccb5d1b6462f87303e230",
  "file": "S37sh2_cine.png",
  "staged_sha256": "67c967285f95f4ed9ea2aba32646d55a5408739e086821717e604631bdbdffb0",
  "latency_ms": 28227
 },
 "S37sh13::signage": {
  "fp": "53b23fba18d0cb62",
  "inscriptions": [],
  "cues": [],
  "dropped": []
 },
 "S37sh13": {
  "input_fingerprint": "3d37c5a7712509fe",
  "prompt": "Create ONE FINAL photorealistic live-action film still of the moment below — early 20th-century Korea, Jeolla region; all people are Korean unless explicitly stated otherwise. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 덫에 걸린 복만을 보며 입가에 희미한 웃음이 반쯤 번진 배상도의 얼굴 클로즈업.\n\nLOCATION (lock): The market-side standing space beside the wall-mounted bounty notice in daylight. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: close-up\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Preserve the established daytime ambient light and muted contrast, making the smile unsettling through performance rather than sinister lighting.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The people visible in that photo are NOT in this shot: never carry their faces, bodies or clothing onto anyone here. Each person in THIS shot is defined solely by the PEOPLE list, their own CHARACTER REFERENCE images and their wardrobe notes. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nPREVIOUS STILL USAGE (follow exactly — what to take from the attached still and what to exclude; it governs place and objects only, never who is in this shot or how the camera sees them): Take the plaster or masonry wall as visible, the posted notice, its red official seal, and daylight. Exclude the salt and medicine bundles as background fixtures; they are carried objects.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The wanted notice remains on the wall with the name, bounty and red official seal unchanged. 배상도: He has paused mid-departure and turned his head back.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 배상도 (한국인 남성, 성인 나이대, 자연스러운 검은 머리) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.",
  "roll_prompts": {
   "A": "Create ONE FINAL photorealistic live-action film still of the moment below — early 20th-century Korea, Jeolla region; all people are Korean unless explicitly stated otherwise. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 덫에 걸린 복만을 보며 입가에 희미한 웃음이 반쯤 번진 배상도의 얼굴 클로즈업.\n\nLOCATION (lock): The market-side standing space beside the wall-mounted bounty notice in daylight. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: close-up\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Preserve the established daytime ambient light and muted contrast, making the smile unsettling through performance rather than sinister lighting.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The people visible in that photo are NOT in this shot: never carry their faces, bodies or clothing onto anyone here. Each person in THIS shot is defined solely by the PEOPLE list, their own CHARACTER REFERENCE images and their wardrobe notes. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nPREVIOUS STILL USAGE (follow exactly — what to take from the attached still and what to exclude; it governs place and objects only, never who is in this shot or how the camera sees them): Take the plaster or masonry wall as visible, the posted notice, its red official seal, and daylight. Exclude the salt and medicine bundles as background fixtures; they are carried objects.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The wanted notice remains on the wall with the name, bounty and red official seal unchanged. 배상도: He has paused mid-departure and turned his head back.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 배상도 (한국인 남성, 성인 나이대, 자연스러운 검은 머리) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.",
   "B": "Create ONE FINAL photorealistic live-action film still of the moment below — early 20th-century Korea, Jeolla region; all people are Korean unless explicitly stated otherwise. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 덫에 걸린 복만을 보며 입가에 희미한 웃음이 반쯤 번진 배상도의 얼굴 클로즈업.\n\nLOCATION (lock): The market-side standing space beside the wall-mounted bounty notice in daylight. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: close-up\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Preserve the established daytime ambient light and muted contrast, making the smile unsettling through performance rather than sinister lighting.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The people visible in that photo are NOT in this shot: never carry their faces, bodies or clothing onto anyone here. Each person in THIS shot is defined solely by the PEOPLE list, their own CHARACTER REFERENCE images and their wardrobe notes. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nPREVIOUS STILL USAGE (follow exactly — what to take from the attached still and what to exclude; it governs place and objects only, never who is in this shot or how the camera sees them): Take the plaster or masonry wall as visible, the posted notice, its red official seal, and daylight. Exclude the salt and medicine bundles as background fixtures; they are carried objects.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The wanted notice remains on the wall with the name, bounty and red official seal unchanged. 배상도: He has paused mid-departure and turned his head back.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 배상도 (한국인 남성, 성인 나이대, 자연스러운 검은 머리) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.\n\nCOMPOSITION VARIATION (this candidate only): keep the place, the people and the moment exactly as specified, and keep from the CAMERA & FRAME contract its subject scale, screen placement and key background elements exactly. For this candidate only, its camera angle and height are deliberately left open: choose a camera position distinctly different from the obvious one, as a film director picking a second setup on the same blocking. Stay true to any explicit framing the SHOT TEXT itself dictates."
  },
  "shot_run_spend_attempt_count": 1,
  "readings": [
   {
    "label": "A",
    "direction": "고개를 돌리고 있으나 시선이 프레임 밖의 대상이 아닌 카메라 렌즈 부근을 향하고 있어 연출 의도와 다름.",
    "built_space": "흙벽과 방이 뒤편에 위치해 있으며, 이전 컷의 공간적 특징을 기본적인 수준에서 유지함.",
    "entities": "배상도의 얼굴 형태는 일치하지만, 캐릭터 참고 이미지에 있는 검은 모자와 조끼가 완전히 누락되어 흰옷만 입고 있음.",
    "hard_violations": [],
    "physics": "뒤를 돌아보는 자세 자체는 자연스럽고 안정적으로 서 있음."
   },
   {
    "label": "B",
    "direction": "고개를 돌려 프레임 밖의 특정 대상(덫에 걸린 복만)을 응시하는 시선이 정확히 연출됨.",
    "built_space": "이전 컷과 동일한 흙벽과 붉은 관인이 찍힌 방이 보이며, 배경에 아웃포커싱된 초가집들이 공간의 연속성을 더해줌.",
    "entities": "배상도의 얼굴이 참고 이미지와 일치하며, 특히 명시된 의상인 검은 중절모와 검은 조끼를 정확하게 착용하고 있음.",
    "hard_violations": [],
    "physics": "가던 길을 멈추고 어깨 너머로 뒤를 돌아보는 몸의 회전과 무게 중심이 매우 자연스럽게 지탱됨."
   }
  ],
  "cross_model_order": {
   "policy": "cross_model_order_v2_slot0_fwd_slot1_rev_parallel2",
   "slots": [
    {
     "model": "gemini-pro",
     "order": "forward",
     "display_to_canonical": {
      "A": "A",
      "B": "B"
     },
     "raw": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 10,
        "verdict_ko": "캐릭터 참고 이미지의 의상(검은 모자와 조끼)을 완벽하게 재현했으며, 프레임 밖을 향하는 시선과 미소가 프롬프트의 상황을 훌륭하게 표현했습니다."
       },
       {
        "label": "A",
        "score": 5,
        "verdict_ko": "얼굴의 클로즈업과 미소는 표현되었으나, 참고 이미지에 명시된 핵심 의상(모자와 검은 조끼)이 완전히 누락되었고 시선이 카메라 렌즈를 향해 부자연스럽습니다."
       }
      ],
      "readings": [
       {
        "label": "B",
        "direction": "고개를 돌려 프레임 밖의 특정 대상(덫에 걸린 복만)을 응시하는 시선이 정확히 연출됨.",
        "built_space": "이전 컷과 동일한 흙벽과 붉은 관인이 찍힌 방이 보이며, 배경에 아웃포커싱된 초가집들이 공간의 연속성을 더해줌.",
        "entities": "배상도의 얼굴이 참고 이미지와 일치하며, 특히 명시된 의상인 검은 중절모와 검은 조끼를 정확하게 착용하고 있음.",
        "hard_violations": [],
        "physics": "가던 길을 멈추고 어깨 너머로 뒤를 돌아보는 몸의 회전과 무게 중심이 매우 자연스럽게 지탱됨."
       },
       {
        "label": "A",
        "direction": "고개를 돌리고 있으나 시선이 프레임 밖의 대상이 아닌 카메라 렌즈 부근을 향하고 있어 연출 의도와 다름.",
        "built_space": "흙벽과 방이 뒤편에 위치해 있으며, 이전 컷의 공간적 특징을 기본적인 수준에서 유지함.",
        "entities": "배상도의 얼굴 형태는 일치하지만, 캐릭터 참고 이미지에 있는 검은 모자와 조끼가 완전히 누락되어 흰옷만 입고 있음.",
        "hard_violations": [],
        "physics": "뒤를 돌아보는 자세 자체는 자연스럽고 안정적으로 서 있음."
       }
      ],
      "all_candidates_fail": false
     },
     "normalized": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 10,
        "verdict_ko": "캐릭터 참고 이미지의 의상(검은 모자와 조끼)을 완벽하게 재현했으며, 프레임 밖을 향하는 시선과 미소가 프롬프트의 상황을 훌륭하게 표현했습니다."
       },
       {
        "label": "A",
        "score": 5,
        "verdict_ko": "얼굴의 클로즈업과 미소는 표현되었으나, 참고 이미지에 명시된 핵심 의상(모자와 검은 조끼)이 완전히 누락되었고 시선이 카메라 렌즈를 향해 부자연스럽습니다."
       }
      ],
      "all_candidates_fail": false,
      "readings": [
       {
        "label": "B",
        "direction": "고개를 돌려 프레임 밖의 특정 대상(덫에 걸린 복만)을 응시하는 시선이 정확히 연출됨.",
        "built_space": "이전 컷과 동일한 흙벽과 붉은 관인이 찍힌 방이 보이며, 배경에 아웃포커싱된 초가집들이 공간의 연속성을 더해줌.",
        "entities": "배상도의 얼굴이 참고 이미지와 일치하며, 특히 명시된 의상인 검은 중절모와 검은 조끼를 정확하게 착용하고 있음.",
        "hard_violations": [],
        "physics": "가던 길을 멈추고 어깨 너머로 뒤를 돌아보는 몸의 회전과 무게 중심이 매우 자연스럽게 지탱됨."
       },
       {
        "label": "A",
        "direction": "고개를 돌리고 있으나 시선이 프레임 밖의 대상이 아닌 카메라 렌즈 부근을 향하고 있어 연출 의도와 다름.",
        "built_space": "흙벽과 방이 뒤편에 위치해 있으며, 이전 컷의 공간적 특징을 기본적인 수준에서 유지함.",
        "entities": "배상도의 얼굴 형태는 일치하지만, 캐릭터 참고 이미지에 있는 검은 모자와 조끼가 완전히 누락되어 흰옷만 입고 있음.",
        "hard_violations": [],
        "physics": "뒤를 돌아보는 자세 자체는 자연스럽고 안정적으로 서 있음."
       }
      ]
     },
     "ok": true
    },
    {
     "model": "gpt-high",
     "order": "reverse",
     "display_to_canonical": {
      "A": "B",
      "B": "A"
     },
     "raw": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 8,
        "verdict_ko": "얼굴 중심의 타이트한 클로즈업, 돌아본 자세와 희미한 웃음, 검은 모자·조끼가 지시에 충실하지만 참조보다 얼굴이 다소 나이 들어 보인다."
       },
       {
        "label": "B",
        "score": 5,
        "verdict_ko": "돌아보며 짓는 옅은 웃음은 맞지만 얼굴보다 어깨까지 넓게 담았고, 참조의 검은 모자와 조끼가 빠져 인물 연속성이 크게 떨어진다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "몸통과 어깨는 화면 오른쪽으로 향하고 얼굴은 뒤돌아 카메라 쪽을 향한다. 눈은 렌즈 가까운 화면 왼쪽 바깥을 보며, 현상수배 고지문을 보는 방향은 아니다. 복만은 프레임 밖이므로 실제 시선 도착점은 확인할 수 없지만, 뒤쪽의 복만을 돌아본다는 설정과 양립한다. 무기나 방향성 소품은 없다.",
        "built_space": "왼쪽 회벽에 고지문 한 장과 그 아래쪽 붉은 원형 관인 하나가 보인다. 오른쪽 배경에는 목재 기둥과 처마, 멀리 초가 건물이 보여 이전 장면의 시장 벽 옆 공간을 이어간다. 배상도는 벽 앞에 서 있으며 고정 시설과 겹치거나 충돌하지 않는다. 고지문은 배경에 흐리게 놓여 있고 반사면이나 중복 시설은 없다.",
        "entities": "성인 한국인 남성으로 설정된 배상도 한 명만 보인다. 자연스러운 검은 머리, 검은 띠 달린 모자, 검은 조끼와 흰 전통 의복의 깃이 참조와 부합한다. 얼굴은 참조보다 주름과 수염 자국이 두드러져 다소 연상으로 보인다. 입가에 비대칭으로 옅은 웃음이 번져 있다. 고지문과 붉은 관인은 있으나 글자는 흐려 이름과 현상금의 동일 여부를 판독할 수 없다. 복만이나 이전 장면 인물, 운반 꾸러미는 추가되지 않았다.",
        "hard_violations": [],
        "physics": "머리는 목과 어깨에 자연스럽게 연결되어 있고, 몸통 방향에서 고개를 돌린 정도도 가능한 범위다. 모자는 머리에 얹혀 있고 조끼와 흰옷은 어깨에 걸쳐 자연스럽게 내려온다. 발과 지면은 클로즈업 밖이어서 접지를 직접 확인할 수 없지만 공중 부유나 지지 없는 물체는 보이지 않는다."
       },
       {
        "label": "B",
        "direction": "몸통은 화면 오른쪽으로 향하고 얼굴은 뒤로 돌아 카메라 쪽을 향한다. 시선은 렌즈에 매우 가까운 화면 왼쪽 바깥으로 향해 직접 카메라를 보는 인상도 있다. 복만은 보이지 않아 시선이 그에게 닿는지는 확인할 수 없다. 고지문을 읽는 시선은 아니며 무기나 지시 소품은 없다.",
        "built_space": "거친 회벽과 벽에 붙은 고지문 한 장, 붉은 원형 관인 하나가 보인다. 고지문 아래쪽에는 고정점 두 개가 보이고 윗부분은 화면 밖으로 잘려 있다. 인물은 벽 앞에 있으며 공간 충돌은 없다. 이전 장면의 목재 기둥과 처마는 이 구도에 포함되지 않아 장소 연결은 주로 벽과 고지문에 의존한다. 반사나 시설 중복은 없다.",
        "entities": "검은 머리의 젊은 성인 남성 한 명이 보이며, 흰 전통 의복을 입고 옅게 웃는다. 참조에 있는 검은 모자와 검은 조끼가 해당 부위가 보이는 구도인데도 없다. 노출된 가르마 머리와 얼굴 인상도 참조 인물과 차이가 난다. 고지문과 붉은 관인은 있지만 이름과 현상금은 판독되지 않아 내용 보존을 확인할 수 없다. 다른 사람이나 꾸러미는 없다.",
        "hard_violations": [],
        "physics": "목과 어깨가 자연스럽게 연결되어 뒤돌아보는 자세를 지탱하며, 옷은 어깨와 몸통에 걸쳐 주름진다. 정지한 상체 자세는 물리적으로 가능하다. 하체와 발은 화면 밖이므로 떠나는 걸음을 멈춘 접지 상태는 확인할 수 없지만, 지지 없이 떠 있는 신체나 물체는 없다."
       }
      ],
      "all_candidates_fail": false
     },
     "normalized": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 8,
        "verdict_ko": "얼굴 중심의 타이트한 클로즈업, 돌아본 자세와 희미한 웃음, 검은 모자·조끼가 지시에 충실하지만 참조보다 얼굴이 다소 나이 들어 보인다."
       },
       {
        "label": "A",
        "score": 5,
        "verdict_ko": "돌아보며 짓는 옅은 웃음은 맞지만 얼굴보다 어깨까지 넓게 담았고, 참조의 검은 모자와 조끼가 빠져 인물 연속성이 크게 떨어진다."
       }
      ],
      "all_candidates_fail": false,
      "readings": [
       {
        "label": "B",
        "direction": "몸통과 어깨는 화면 오른쪽으로 향하고 얼굴은 뒤돌아 카메라 쪽을 향한다. 눈은 렌즈 가까운 화면 왼쪽 바깥을 보며, 현상수배 고지문을 보는 방향은 아니다. 복만은 프레임 밖이므로 실제 시선 도착점은 확인할 수 없지만, 뒤쪽의 복만을 돌아본다는 설정과 양립한다. 무기나 방향성 소품은 없다.",
        "built_space": "왼쪽 회벽에 고지문 한 장과 그 아래쪽 붉은 원형 관인 하나가 보인다. 오른쪽 배경에는 목재 기둥과 처마, 멀리 초가 건물이 보여 이전 장면의 시장 벽 옆 공간을 이어간다. 배상도는 벽 앞에 서 있으며 고정 시설과 겹치거나 충돌하지 않는다. 고지문은 배경에 흐리게 놓여 있고 반사면이나 중복 시설은 없다.",
        "entities": "성인 한국인 남성으로 설정된 배상도 한 명만 보인다. 자연스러운 검은 머리, 검은 띠 달린 모자, 검은 조끼와 흰 전통 의복의 깃이 참조와 부합한다. 얼굴은 참조보다 주름과 수염 자국이 두드러져 다소 연상으로 보인다. 입가에 비대칭으로 옅은 웃음이 번져 있다. 고지문과 붉은 관인은 있으나 글자는 흐려 이름과 현상금의 동일 여부를 판독할 수 없다. 복만이나 이전 장면 인물, 운반 꾸러미는 추가되지 않았다.",
        "hard_violations": [],
        "physics": "머리는 목과 어깨에 자연스럽게 연결되어 있고, 몸통 방향에서 고개를 돌린 정도도 가능한 범위다. 모자는 머리에 얹혀 있고 조끼와 흰옷은 어깨에 걸쳐 자연스럽게 내려온다. 발과 지면은 클로즈업 밖이어서 접지를 직접 확인할 수 없지만 공중 부유나 지지 없는 물체는 보이지 않는다."
       },
       {
        "label": "A",
        "direction": "몸통은 화면 오른쪽으로 향하고 얼굴은 뒤로 돌아 카메라 쪽을 향한다. 시선은 렌즈에 매우 가까운 화면 왼쪽 바깥으로 향해 직접 카메라를 보는 인상도 있다. 복만은 보이지 않아 시선이 그에게 닿는지는 확인할 수 없다. 고지문을 읽는 시선은 아니며 무기나 지시 소품은 없다.",
        "built_space": "거친 회벽과 벽에 붙은 고지문 한 장, 붉은 원형 관인 하나가 보인다. 고지문 아래쪽에는 고정점 두 개가 보이고 윗부분은 화면 밖으로 잘려 있다. 인물은 벽 앞에 있으며 공간 충돌은 없다. 이전 장면의 목재 기둥과 처마는 이 구도에 포함되지 않아 장소 연결은 주로 벽과 고지문에 의존한다. 반사나 시설 중복은 없다.",
        "entities": "검은 머리의 젊은 성인 남성 한 명이 보이며, 흰 전통 의복을 입고 옅게 웃는다. 참조에 있는 검은 모자와 검은 조끼가 해당 부위가 보이는 구도인데도 없다. 노출된 가르마 머리와 얼굴 인상도 참조 인물과 차이가 난다. 고지문과 붉은 관인은 있지만 이름과 현상금은 판독되지 않아 내용 보존을 확인할 수 없다. 다른 사람이나 꾸러미는 없다.",
        "hard_violations": [],
        "physics": "목과 어깨가 자연스럽게 연결되어 뒤돌아보는 자세를 지탱하며, 옷은 어깨와 몸통에 걸쳐 주름진다. 정지한 상체 자세는 물리적으로 가능하다. 하체와 발은 화면 밖이므로 떠나는 걸음을 멈춘 접지 상태는 확인할 수 없지만, 지지 없이 떠 있는 신체나 물체는 없다."
       }
      ]
     },
     "ok": true
    }
   ],
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "slot_winner_match": true,
   "slot_winner": {
    "gemini-pro": "B",
    "gpt-high": "B"
   },
   "route": "cross_slot_agree"
  },
  "dual": {
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "normalized": {
    "A": 1.125,
    "B": 2.0
   },
   "adjusted": {
    "A": 1.125,
    "B": 2.0
   },
   "violations": {},
   "per_model_winner": {
    "gemini-pro": "B",
    "gpt-high": "B"
   },
   "agreed": true
  },
  "totals": {
   "B": 2000,
   "A": 1125
  },
  "selected": "B",
  "ranking": [
   "B",
   "A"
  ],
  "verdicts": [
   {
    "label": "B",
    "score": 2000,
    "verdict_ko": "캐릭터 참고 이미지의 의상(검은 모자와 조끼)을 완벽하게 재현했으며, 프레임 밖을 향하는 시선과 미소가 프롬프트의 상황을 훌륭하게 표현했습니다."
   },
   {
    "label": "A",
    "score": 1125,
    "verdict_ko": "얼굴의 클로즈업과 미소는 표현되었으나, 참고 이미지에 명시된 핵심 의상(모자와 검은 조끼)이 완전히 누락되었고 시선이 카메라 렌즈를 향해 부자연스럽습니다."
   }
  ],
  "refs": [
   {
    "label": "PREVIOUS SHOT STILL — a visually related earlier shot of this same place: the location's look, materials, fixed features and lighting mood are LOCKED to this photo; never copy its camera framing. The people visible in that photo are NOT in this shot: never carry their faces, bodies or clothing onto anyone here. Each person in THIS shot is defined solely by the PEOPLE list, their own CHARACTER REFERENCE images and their wardrobe notes. If this photo shows a character who cannot move (dead or unconscious), that character's exact pose, position and orientation are ALSO LOCKED — treat their body as a fixed prop of the set that only the camera moves around.",
    "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/db00517b-165b-4005-a287-f28936b59b95/images/272bd7af-9294-4490-98c6-88a40e9deacc/scene/recipe/S37sh2_sel.png",
    "asset_id": "120d7022-792e-4f91-a991-6beb79815f8b",
    "role": "prev_still"
   },
   {
    "label": "CHARACTER REFERENCE — 배상도: the exact person appearing in this shot; match face, hair and build exactly.",
    "path": "<bytes:851718>",
    "asset_id": "034f5325-975e-4474-854b-e48359710bf8",
    "role": "character_ref"
   }
  ],
  "critique_skipped": true,
  "shot_run_produced": false,
  "shot_run_uid": "06aa1428-5c5f-70e4-ba2e-0870b3eb0403",
  "ref_mode": "prev+엔티티",
  "share_plan": {
   "ref_plan": "prev",
   "prev_anchor_tag": "S37sh2"
  }
 },
 "S37sh13::cine": {
  "applied": true,
  "attempted_at": "2026-09-09T05:03:04.406633+00:00",
  "fingerprint": "bc710d91c101f1d6c85190e69a15e0b64f6dbbea6126756f59a2671aa3ca24be",
  "fingerprint_version": 2,
  "provider": "mai",
  "endpoint": "openrouter/chat-completions",
  "model": "microsoft/mai-image-2.6",
  "pack": "24.202608252115",
  "source_file": "S37sh13_sel.png",
  "source_sha256": "91bf700cacfc729c319dcca1e5901d9962365a4a4093552c45efb399ad6c711e",
  "file": "S37sh13_cine.png",
  "staged_sha256": "49a6607fd93a5067b9f3abe7d75b084d5dd584336a75a14b4067d47ce9ed7f7f",
  "latency_ms": 27213
 },
 "S37sh17::signage": {
  "fp": "48e57eee601263b1",
  "inscriptions": [],
  "cues": [
   {
    "text_native": "",
    "source": "scene_text_implied",
    "source_quote": "벽보"
   }
  ],
  "dropped": []
 },
 "S37sh17": {
  "input_fingerprint": "da3e7ebd1a0ff7ae",
  "prompt": "Create ONE FINAL photorealistic live-action film still of the moment below — early 20th-century Korea, Jeolla region; all people are Korean unless explicitly stated otherwise. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 벽보의 이름을 다시 올려다보며 미간이 깊게 팬 복만의 짙게 가라앉은 표정.\n\nLOCATION (lock): The outdoor frontage of the market wall with the bounty notice and red official seal. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: close-up\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Keep the same restrained daylight and naturalistic contrast, preserving the weight of the expression without darkening the setting for effect.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The clothing, hair and overall look of 복만 — who appear both in that photo and in this shot — are LOCKED to that photo. Anyone else visible in that photo is NOT in this shot: never carry their face, body or clothing onto anyone here. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nPREVIOUS STILL USAGE (follow exactly — what to take from the attached still and what to exclude; it governs place and objects only, never who is in this shot or how the camera sees them): Take the notice-bearing wall, paper texture, red official seal, and daytime light. Exclude the departing informant from the notice area.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The wanted notice still displays '안규홍' above the sharply visible red official seal. 복만: He remains by the notice carrying salt and medicinal herbs.\n\nTHE HAND THAT IS DOING THIS: the object at the centre of this shot is being held, operated, read, aimed or handed over by 복만 right now, so 복만's hand — and as much of the wrist and forearm as the framing reaches — is in the frame, gripping or resting on that object exactly the way its real use requires. Match that hand to 복만: its size, build, skin, age, grooming, sleeve and anything worn on it belong to that person and to no one else.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 복만 (한국인 남성, 성인 나이대, 거칠게 자란 수염, 자연스러운 검은 머리) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.",
  "roll_prompts": {
   "A": "Create ONE FINAL photorealistic live-action film still of the moment below — early 20th-century Korea, Jeolla region; all people are Korean unless explicitly stated otherwise. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 벽보의 이름을 다시 올려다보며 미간이 깊게 팬 복만의 짙게 가라앉은 표정.\n\nLOCATION (lock): The outdoor frontage of the market wall with the bounty notice and red official seal. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: close-up\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Keep the same restrained daylight and naturalistic contrast, preserving the weight of the expression without darkening the setting for effect.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The clothing, hair and overall look of 복만 — who appear both in that photo and in this shot — are LOCKED to that photo. Anyone else visible in that photo is NOT in this shot: never carry their face, body or clothing onto anyone here. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nPREVIOUS STILL USAGE (follow exactly — what to take from the attached still and what to exclude; it governs place and objects only, never who is in this shot or how the camera sees them): Take the notice-bearing wall, paper texture, red official seal, and daytime light. Exclude the departing informant from the notice area.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The wanted notice still displays '안규홍' above the sharply visible red official seal. 복만: He remains by the notice carrying salt and medicinal herbs.\n\nTHE HAND THAT IS DOING THIS: the object at the centre of this shot is being held, operated, read, aimed or handed over by 복만 right now, so 복만's hand — and as much of the wrist and forearm as the framing reaches — is in the frame, gripping or resting on that object exactly the way its real use requires. Match that hand to 복만: its size, build, skin, age, grooming, sleeve and anything worn on it belong to that person and to no one else.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 복만 (한국인 남성, 성인 나이대, 거칠게 자란 수염, 자연스러운 검은 머리) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.",
   "B": "Create ONE FINAL photorealistic live-action film still of the moment below — early 20th-century Korea, Jeolla region; all people are Korean unless explicitly stated otherwise. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 벽보의 이름을 다시 올려다보며 미간이 깊게 팬 복만의 짙게 가라앉은 표정.\n\nLOCATION (lock): The outdoor frontage of the market wall with the bounty notice and red official seal. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: close-up\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Keep the same restrained daylight and naturalistic contrast, preserving the weight of the expression without darkening the setting for effect.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The clothing, hair and overall look of 복만 — who appear both in that photo and in this shot — are LOCKED to that photo. Anyone else visible in that photo is NOT in this shot: never carry their face, body or clothing onto anyone here. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nPREVIOUS STILL USAGE (follow exactly — what to take from the attached still and what to exclude; it governs place and objects only, never who is in this shot or how the camera sees them): Take the notice-bearing wall, paper texture, red official seal, and daytime light. Exclude the departing informant from the notice area.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The wanted notice still displays '안규홍' above the sharply visible red official seal. 복만: He remains by the notice carrying salt and medicinal herbs.\n\nTHE HAND THAT IS DOING THIS: the object at the centre of this shot is being held, operated, read, aimed or handed over by 복만 right now, so 복만's hand — and as much of the wrist and forearm as the framing reaches — is in the frame, gripping or resting on that object exactly the way its real use requires. Match that hand to 복만: its size, build, skin, age, grooming, sleeve and anything worn on it belong to that person and to no one else.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 복만 (한국인 남성, 성인 나이대, 거칠게 자란 수염, 자연스러운 검은 머리) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.\n\nCOMPOSITION VARIATION (this candidate only): keep the place, the people and the moment exactly as specified, and keep from the CAMERA & FRAME contract its subject scale, screen placement and key background elements exactly. For this candidate only, its camera angle and height are deliberately left open: choose a camera position distinctly different from the obvious one, as a film director picking a second setup on the same blocking. Stay true to any explicit framing the SHOT TEXT itself dictates."
  },
  "shot_run_spend_attempt_count": 1,
  "readings": [
   {
    "label": "A",
    "direction": "인물은 화면 좌측 상단을 바라보지만, 벽보는 인물의 우측 뒤편 흙벽에 위치하여 시선이 대상을 향하지 않습니다.",
    "built_space": "흙벽과 붉은 관인이 찍힌 벽보가 인물의 배경으로 존재합니다.",
    "entities": "복만의 얼굴 묘사와 손에 든 짐자루는 나타나 있으나, 이전 숏 스틸에서 착용하지 않았던 모자를 쓰고 있습니다.",
    "hard_violations": [
     "[gemini-pro] 이전 숏 스틸로 고정(LOCKED)된 복장 지시를 위반하여 모자를 착용함"
    ],
    "physics": "두 손으로 짐자루의 윗부분을 단단히 움켜쥐고 있습니다."
   },
   {
    "label": "B",
    "direction": "인물은 화면 좌측 상단 벽면에 붙은 벽보를 향해 정확히 시선을 올려다보고 있습니다.",
    "built_space": "참조 이미지와 일치하는 질감의 흙벽, 처마, 그리고 붉은 관인이 선명한 벽보가 올바른 위치에 구성되어 있습니다.",
    "entities": "복만이 이전 숏 스틸과 동일하게 모자 없이 헝클어진 머리를 하고 있으며, 미간을 찌푸린 표정으로 짐자루를 안고 있습니다.",
    "hard_violations": [],
    "physics": "손과 팔을 이용해 두 개의 짐자루 하단을 자연스럽게 받치고 있습니다."
   }
  ],
  "cross_model_order": {
   "policy": "cross_model_order_v2_slot0_fwd_slot1_rev_parallel2",
   "slots": [
    {
     "model": "gemini-pro",
     "order": "forward",
     "display_to_canonical": {
      "A": "A",
      "B": "B"
     },
     "raw": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 3,
        "verdict_ko": "이전 장면 스틸에서 없었던 모자를 쓰고 있어 엄격한 복장 연속성 조건을 위반했으며, 시선이 등 뒤의 벽보를 향하지 않고 엉뚱한 곳을 바라봅니다."
       },
       {
        "label": "B",
        "score": 9,
        "verdict_ko": "이전 장면 스틸의 머리 모양과 복장을 정확히 유지하고 있으며, 벽보를 올려다보는 시선 방향과 미간이 깊게 팬 표정을 훌륭하게 구현했습니다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "인물은 화면 좌측 상단을 바라보지만, 벽보는 인물의 우측 뒤편 흙벽에 위치하여 시선이 대상을 향하지 않습니다.",
        "built_space": "흙벽과 붉은 관인이 찍힌 벽보가 인물의 배경으로 존재합니다.",
        "entities": "복만의 얼굴 묘사와 손에 든 짐자루는 나타나 있으나, 이전 숏 스틸에서 착용하지 않았던 모자를 쓰고 있습니다.",
        "hard_violations": [
         "이전 숏 스틸로 고정(LOCKED)된 복장 지시를 위반하여 모자를 착용함"
        ],
        "physics": "두 손으로 짐자루의 윗부분을 단단히 움켜쥐고 있습니다."
       },
       {
        "label": "B",
        "direction": "인물은 화면 좌측 상단 벽면에 붙은 벽보를 향해 정확히 시선을 올려다보고 있습니다.",
        "built_space": "참조 이미지와 일치하는 질감의 흙벽, 처마, 그리고 붉은 관인이 선명한 벽보가 올바른 위치에 구성되어 있습니다.",
        "entities": "복만이 이전 숏 스틸과 동일하게 모자 없이 헝클어진 머리를 하고 있으며, 미간을 찌푸린 표정으로 짐자루를 안고 있습니다.",
        "hard_violations": [],
        "physics": "손과 팔을 이용해 두 개의 짐자루 하단을 자연스럽게 받치고 있습니다."
       }
      ],
      "all_candidates_fail": false
     },
     "normalized": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 3,
        "verdict_ko": "이전 장면 스틸에서 없었던 모자를 쓰고 있어 엄격한 복장 연속성 조건을 위반했으며, 시선이 등 뒤의 벽보를 향하지 않고 엉뚱한 곳을 바라봅니다."
       },
       {
        "label": "B",
        "score": 9,
        "verdict_ko": "이전 장면 스틸의 머리 모양과 복장을 정확히 유지하고 있으며, 벽보를 올려다보는 시선 방향과 미간이 깊게 팬 표정을 훌륭하게 구현했습니다."
       }
      ],
      "all_candidates_fail": false,
      "readings": [
       {
        "label": "A",
        "direction": "인물은 화면 좌측 상단을 바라보지만, 벽보는 인물의 우측 뒤편 흙벽에 위치하여 시선이 대상을 향하지 않습니다.",
        "built_space": "흙벽과 붉은 관인이 찍힌 벽보가 인물의 배경으로 존재합니다.",
        "entities": "복만의 얼굴 묘사와 손에 든 짐자루는 나타나 있으나, 이전 숏 스틸에서 착용하지 않았던 모자를 쓰고 있습니다.",
        "hard_violations": [
         "이전 숏 스틸로 고정(LOCKED)된 복장 지시를 위반하여 모자를 착용함"
        ],
        "physics": "두 손으로 짐자루의 윗부분을 단단히 움켜쥐고 있습니다."
       },
       {
        "label": "B",
        "direction": "인물은 화면 좌측 상단 벽면에 붙은 벽보를 향해 정확히 시선을 올려다보고 있습니다.",
        "built_space": "참조 이미지와 일치하는 질감의 흙벽, 처마, 그리고 붉은 관인이 선명한 벽보가 올바른 위치에 구성되어 있습니다.",
        "entities": "복만이 이전 숏 스틸과 동일하게 모자 없이 헝클어진 머리를 하고 있으며, 미간을 찌푸린 표정으로 짐자루를 안고 있습니다.",
        "hard_violations": [],
        "physics": "손과 팔을 이용해 두 개의 짐자루 하단을 자연스럽게 받치고 있습니다."
       }
      ]
     },
     "ok": true
    },
    {
     "model": "gpt-high",
     "order": "reverse",
     "display_to_canonical": {
      "A": "B",
      "B": "A"
     },
     "raw": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 7,
        "verdict_ko": "벽보 쪽으로 올려다보는 시선, 깊게 팬 미간, 이전 컷의 맨머리를 유지하지만, 얼굴 클로즈업보다 넓고 벽보에 닿는 손이 없다."
       },
       {
        "label": "B",
        "score": 4,
        "verdict_ko": "가라앉은 표정은 있으나 벽보가 시선 뒤쪽에 있어 이름을 다시 읽는 동작이 성립하지 않으며, 이전 컷에 없던 모자를 착용했다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "얼굴과 눈이 화면 왼쪽 위를 향하며, 그쪽 벽에 벽보가 있다. 이름 자체는 흐려 확인할 수 없지만 벽보 상단을 올려다보는 방향 관계는 대체로 맞는다.",
        "built_space": "거친 흙벽에 벽보 한 장과 붉은 원형 관인 하나가 보이고, 오른쪽 위에는 목재 서까래와 처마가 이어진다. 복만은 벽 바로 옆에 서 있으며 이전 장소의 재료와 낮빛을 대체로 유지한다. 다만 얼굴뿐 아니라 가슴과 꾸러미까지 포함한 상반신 구도로, 요구한 클로즈업보다 넓다.",
        "entities": "성인 한국인 남성으로 보이는 인물 한 명만 있다. 검은 머리, 거친 수염, 낡은 회갈색 옷과 모자 없는 모습은 이전 컷에 부합한다. 벽보의 종이와 붉은 관인이 있으며 글은 판독하기 어렵다. 천 꾸러미 두 개는 보이지만 내용물이 소금과 약초인지는 확인할 수 없다. 손은 하단 꾸러미에 있고, 읽는 대상인 벽보에 손을 대라는 별도 지시는 충족하지 않는다.",
        "hard_violations": [],
        "physics": "머리와 상체는 자연스럽게 연결되어 있고, 서 있는 자세로 보인다. 발은 프레임 밖이므로 접지는 확인할 수 없다. 하단의 손과 팔이 꾸러미를 몸 앞에서 받치며, 지지 없이 떠 있는 물체나 불가능한 관절은 보이지 않는다."
       },
       {
        "label": "B",
        "direction": "눈과 코는 화면 오른쪽 위의 화면 밖 공간을 향한다. 벽보는 머리 뒤쪽 벽에 붙어 있어 현재 시선이 그 표면의 이름에 닿는 배치가 아니다. 위를 보기는 하지만 벽보를 다시 읽는 시선은 아니다.",
        "built_space": "흙벽에 벽보 한 장과 붉은 원형 관인 하나가 있으며, 왼쪽에는 세로 목재와 흐릿한 초가 건물이 보인다. 복만은 벽보를 뒤에 둔 채 바깥쪽을 향한다. 재료와 낮빛은 유사하지만 읽기 동작에 필요한 벽과 인물의 방향 관계가 맞지 않는다. 양손과 가슴까지 포함해 얼굴 클로즈업보다 넓다.",
        "entities": "성인 한국인 남성 한 명이며 수염, 얼굴과 낡은 옷은 인물 참고와 대체로 맞는다. 그러나 갓 형태의 모자와 머리띠는 모자 없이 검은 머리를 드러낸 이전 컷의 외형 잠금을 따르지 않는다. 벽보와 붉은 관인이 있고 글은 흐려 판독하기 어렵다. 두 손으로 천 꾸러미 두 개를 잡고 있으나 내용물은 확인되지 않으며, 어느 손도 벽보에 닿지 않는다.",
        "hard_violations": [],
        "physics": "두 손이 각각 꾸러미의 윗부분을 움켜쥐고 있어 무게를 지탱하는 접촉이 명확하다. 모자는 머리에 얹혀 있고 끈이 턱 아래로 내려온다. 상체와 팔의 자세는 가능하며, 발이 잘린 구도 자체를 부유로 볼 근거는 없다."
       }
      ],
      "all_candidates_fail": true
     },
     "normalized": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 7,
        "verdict_ko": "벽보 쪽으로 올려다보는 시선, 깊게 팬 미간, 이전 컷의 맨머리를 유지하지만, 얼굴 클로즈업보다 넓고 벽보에 닿는 손이 없다."
       },
       {
        "label": "A",
        "score": 4,
        "verdict_ko": "가라앉은 표정은 있으나 벽보가 시선 뒤쪽에 있어 이름을 다시 읽는 동작이 성립하지 않으며, 이전 컷에 없던 모자를 착용했다."
       }
      ],
      "all_candidates_fail": true,
      "readings": [
       {
        "label": "B",
        "direction": "얼굴과 눈이 화면 왼쪽 위를 향하며, 그쪽 벽에 벽보가 있다. 이름 자체는 흐려 확인할 수 없지만 벽보 상단을 올려다보는 방향 관계는 대체로 맞는다.",
        "built_space": "거친 흙벽에 벽보 한 장과 붉은 원형 관인 하나가 보이고, 오른쪽 위에는 목재 서까래와 처마가 이어진다. 복만은 벽 바로 옆에 서 있으며 이전 장소의 재료와 낮빛을 대체로 유지한다. 다만 얼굴뿐 아니라 가슴과 꾸러미까지 포함한 상반신 구도로, 요구한 클로즈업보다 넓다.",
        "entities": "성인 한국인 남성으로 보이는 인물 한 명만 있다. 검은 머리, 거친 수염, 낡은 회갈색 옷과 모자 없는 모습은 이전 컷에 부합한다. 벽보의 종이와 붉은 관인이 있으며 글은 판독하기 어렵다. 천 꾸러미 두 개는 보이지만 내용물이 소금과 약초인지는 확인할 수 없다. 손은 하단 꾸러미에 있고, 읽는 대상인 벽보에 손을 대라는 별도 지시는 충족하지 않는다.",
        "hard_violations": [],
        "physics": "머리와 상체는 자연스럽게 연결되어 있고, 서 있는 자세로 보인다. 발은 프레임 밖이므로 접지는 확인할 수 없다. 하단의 손과 팔이 꾸러미를 몸 앞에서 받치며, 지지 없이 떠 있는 물체나 불가능한 관절은 보이지 않는다."
       },
       {
        "label": "A",
        "direction": "눈과 코는 화면 오른쪽 위의 화면 밖 공간을 향한다. 벽보는 머리 뒤쪽 벽에 붙어 있어 현재 시선이 그 표면의 이름에 닿는 배치가 아니다. 위를 보기는 하지만 벽보를 다시 읽는 시선은 아니다.",
        "built_space": "흙벽에 벽보 한 장과 붉은 원형 관인 하나가 있으며, 왼쪽에는 세로 목재와 흐릿한 초가 건물이 보인다. 복만은 벽보를 뒤에 둔 채 바깥쪽을 향한다. 재료와 낮빛은 유사하지만 읽기 동작에 필요한 벽과 인물의 방향 관계가 맞지 않는다. 양손과 가슴까지 포함해 얼굴 클로즈업보다 넓다.",
        "entities": "성인 한국인 남성 한 명이며 수염, 얼굴과 낡은 옷은 인물 참고와 대체로 맞는다. 그러나 갓 형태의 모자와 머리띠는 모자 없이 검은 머리를 드러낸 이전 컷의 외형 잠금을 따르지 않는다. 벽보와 붉은 관인이 있고 글은 흐려 판독하기 어렵다. 두 손으로 천 꾸러미 두 개를 잡고 있으나 내용물은 확인되지 않으며, 어느 손도 벽보에 닿지 않는다.",
        "hard_violations": [],
        "physics": "두 손이 각각 꾸러미의 윗부분을 움켜쥐고 있어 무게를 지탱하는 접촉이 명확하다. 모자는 머리에 얹혀 있고 끈이 턱 아래로 내려온다. 상체와 팔의 자세는 가능하며, 발이 잘린 구도 자체를 부유로 볼 근거는 없다."
       }
      ]
     },
     "ok": true
    }
   ],
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "slot_winner_match": true,
   "slot_winner": {
    "gemini-pro": "B",
    "gpt-high": "B"
   },
   "route": "cross_slot_agree"
  },
  "dual": {
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "normalized": {
    "A": 0.905,
    "B": 2.0
   },
   "adjusted": {
    "A": 0.655,
    "B": 2.0
   },
   "violations": {
    "A": [
     "[gemini-pro] 이전 숏 스틸로 고정(LOCKED)된 복장 지시를 위반하여 모자를 착용함"
    ]
   },
   "per_model_winner": {
    "gemini-pro": "B",
    "gpt-high": "B"
   },
   "agreed": true
  },
  "totals": {
   "A": 655,
   "B": 2000
  },
  "selected": "B",
  "ranking": [
   "B",
   "A"
  ],
  "verdicts": [
   {
    "label": "A",
    "score": 655,
    "verdict_ko": "이전 장면 스틸에서 없었던 모자를 쓰고 있어 엄격한 복장 연속성 조건을 위반했으며, 시선이 등 뒤의 벽보를 향하지 않고 엉뚱한 곳을 바라봅니다.  ★위반: [gemini-pro] 이전 숏 스틸로 고정(LOCKED)된 복장 지시를 위반하여 모자를 착용함"
   },
   {
    "label": "B",
    "score": 2000,
    "verdict_ko": "이전 장면 스틸의 머리 모양과 복장을 정확히 유지하고 있으며, 벽보를 올려다보는 시선 방향과 미간이 깊게 팬 표정을 훌륭하게 구현했습니다."
   }
  ],
  "refs": [
   {
    "label": "PREVIOUS SHOT STILL — a visually related earlier shot of this same place: the location's look, materials, fixed features and lighting mood are LOCKED to this photo; never copy its camera framing. The clothing, hair and overall look of 복만 — who appear both in that photo and in this shot — are LOCKED to that photo. Anyone else visible in that photo is NOT in this shot: never carry their face, body or clothing onto anyone here. If this photo shows a character who cannot move (dead or unconscious), that character's exact pose, position and orientation are ALSO LOCKED — treat their body as a fixed prop of the set that only the camera moves around.",
    "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/db00517b-165b-4005-a287-f28936b59b95/images/272bd7af-9294-4490-98c6-88a40e9deacc/scene/recipe/S37sh2_sel.png",
    "asset_id": "120d7022-792e-4f91-a991-6beb79815f8b",
    "role": "prev_still"
   },
   {
    "label": "CHARACTER REFERENCE — 복만: the exact person appearing in this shot; match face, hair and build exactly.",
    "path": "<bytes:1027479>",
    "asset_id": "b7e69570-d15d-4241-b5ea-4cb8d4ece1cf",
    "role": "character_ref"
   }
  ],
  "critique_skipped": true,
  "shot_run_produced": false,
  "shot_run_uid": "06aa1428-5df9-712d-a84b-977f96db18fc",
  "ref_mode": "prev+엔티티",
  "share_plan": {
   "ref_plan": "prev",
   "prev_anchor_tag": "S37sh2"
  }
 },
 "S37sh17::cine": {
  "applied": true,
  "attempted_at": "2026-09-09T05:04:25.199037+00:00",
  "fingerprint": "d08097e2a1d8a240deb7ff6f528caf11690a7b6b1dcfa2e2c326a1ef84418bbe",
  "fingerprint_version": 2,
  "provider": "mai",
  "endpoint": "openrouter/chat-completions",
  "model": "microsoft/mai-image-2.6",
  "pack": "24.202608252115",
  "source_file": "S37sh17_sel.png",
  "source_sha256": "f88beab89284dcc9441bc0bb5fcebeb1631c108499a3b3010c309bbaf28564cd",
  "file": "S37sh17_cine.png",
  "staged_sha256": "9000c94f01fdf99e675618f1a2af29c25f0ea250457bbe0d23f56d3220dc5092",
  "latency_ms": 32024
 },
 "S38sh3::signage": {
  "fp": "c6ce632c5ddf0fac",
  "inscriptions": [],
  "cues": [],
  "dropped": []
 },
 "groupbg::mountain_camp_clearing": {
  "input_fingerprint": "fb2e082099ce2f00",
  "meta": {
   "model": "gpt-image-2",
   "size": "1536x864",
   "pack": "11.202607220237",
   "contract": "bgfirst_full_v3",
   "group_sig": {
    "key": "mountain_camp_clearing",
    "tags": [
     "S26sh10",
     "S26sh13",
     "S26sh2",
     "S27sh3",
     "S29sh12",
     "S29sh2",
     "S29sh6",
     "S38sh13",
     "S38sh17",
     "S38sh3",
     "S45sh20",
     "S45sh33",
     "S45sh5",
     "S46sh16",
     "S46sh3",
     "S46sh7",
     "S47sh13",
     "S47sh18"
    ]
   },
   "context_sig": "70a7bc99b0c01248"
  },
  "prompt": "Create ONE empty live-action location background photograph — NO PEOPLE, no figures, no body parts, no silhouettes, no shadows or reflections of people anywhere.\n\"Empty\" means no people only: KEEP the location's inherent occupants and stock that define the place — animals in an animal shelter, pen or farm, goods and produce in a market, moored boats in a harbour — unless the shot text explicitly removes them.\nThe FIRST attached image is a thin-line storyboard sketch of a shot that happens at this location: use it ONLY as spatial evidence — what this place contains, how its ground, structures and landmarks are arranged and proportioned. Ignore the sketched people and arrows entirely, and do NOT copy its line style: render a fully photographic, physically plausible real place that fits THE LOCATION text below.\nHUMAN-SCALE CALIBRATION: derive every structure's true size from human-scale elements — a door ≈ 2m, a window ≈ 1–1.5m wide, one storey ≈ 2.5–3m; never inflate a small structure or shrink a large one.\n\nTHE LOCATION — early 20th-century Korea, Jeolla region; all people are Korean unless explicitly stated otherwise: The entrance edge of the mountain camp's dirt meeting clearing, crossed by red sunset light through the palisade.\n\nLOCATION DETAIL (descriptive evidence about this place from the production's location records — evidence, not a staging order): use it to understand what kind of place this is and what it permanently contains. Carry over ONLY its enduring physical features — layout, structures, ground and materials, fixed equipment and installations. Do NOT treat any lighting, weather or time-of-day wording, momentary object states, depicted events or subjective impressions in it as fixed facts of the place; the TIME OF DAY lock below is the sole authority for time and lighting.\n파청 산채 공터·훈련장·취사장·막사 주변: 산 중턱에 통나무를 세워 만든 조잡한 목책 안의 넓은 흙바닥 공터. 나무와 거적으로 만든 움막과 막사가 흩어져 있고, 돌과 진흙으로 만든 취사용 아궁이와 무쇠 솥, 모닥불 터가 있다. 전투 후에는 솥이 뒤집히고 막사가 불타거나 무너진다. (특징: 통나무를 박아 세운 목책; 넓게 다져진 흙바닥 (공터/훈련장); 돌과 진흙으로 만든 다수의 아궁이와 무쇠솥; 나무와 거적으로 엮어 만든 움막과 막사; 부서진 토기 밥그릇과 뒤집힌 무쇠솥 (전투 후); 불에 탄 나무 잔해와 흩어진 쌀밥 (전투 후))\n\nSCENE EVIDENCE (verbatim quotes from the screenplay about this place — treat them as evidence of what the location physically contains and looks like; stage the PLACE those moments happen in, but do NOT depict the momentary actions, people or staged props themselves):\n- 산채 공터.\n- 1. 오주일이 나뭇가지를 지휘봉처럼 들고 사내들 앞에 선다.\n- 훈련하던 규홍과 사내들이 돌아본다.\n- 해가 산등성이 너머로 기운다. 붉은 빛이 목책 사이로 길게 드리운다.\n- S#45. 파청 산채 / 밤\n- 달래의 시신이 짚멍석 위에 뉘어져 있다.\n- 산채 한가운데. 의병들이 둘러앉아 있다.\n- 옆에서 듣고 있던 임정현. 쌀자루를 세다 말고 손이 멈춘다.\n- 규홍과 관회가 박상현을 따라 산길을 내려간다.\n임정현은 그 뒷모습을 잠시 바라본다.\n\nTIME OF DAY (lock): sunset, low reddish sunlight.\n\nRender ONE photorealistic empty location photograph, 16:9, neutral enough that every shot of this place can be staged from it later. No readable writing anywhere: surfaces that would carry writing may be present, but stage any wording out of legibility — an oblique angle, distance, shallow focus. No captions, watermarks or overlay text.",
  "effective_prompt": "Create ONE empty live-action location background photograph — NO PEOPLE, no figures, no body parts, no silhouettes, no shadows or reflections of people anywhere.\n\"Empty\" means no people only: KEEP the location's inherent occupants and stock that define the place — animals in an animal shelter, pen or farm, goods and produce in a market, moored boats in a harbour — unless the shot text explicitly removes them.\nThe FIRST attached image is a thin-line storyboard sketch of a shot that happens at this location: use it ONLY as spatial evidence — what this place contains, how its ground, structures and landmarks are arranged and proportioned. Ignore the sketched people and arrows entirely, and do NOT copy its line style: render a fully photographic, physically plausible real place that fits THE LOCATION text below.\nHUMAN-SCALE CALIBRATION: derive every structure's true size from human-scale elements — a door ≈ 2m, a window ≈ 1–1.5m wide, one storey ≈ 2.5–3m; never inflate a small structure or shrink a large one.\n\nTHE LOCATION — early 20th-century Korea, Jeolla region; all people are Korean unless explicitly stated otherwise: The entrance edge of the mountain camp's dirt meeting clearing, crossed by red sunset light through the palisade.\n\nLOCATION DETAIL (descriptive evidence about this place from the production's location records — evidence, not a staging order): use it to understand what kind of place this is and what it permanently contains. Carry over ONLY its enduring physical features — layout, structures, ground and materials, fixed equipment and installations. Do NOT treat any lighting, weather or time-of-day wording, momentary object states, depicted events or subjective impressions in it as fixed facts of the place; the TIME OF DAY lock below is the sole authority for time and lighting.\n파청 산채 공터·훈련장·취사장·막사 주변: 산 중턱에 통나무를 세워 만든 조잡한 목책 안의 넓은 흙바닥 공터. 나무와 거적으로 만든 움막과 막사가 흩어져 있고, 돌과 진흙으로 만든 취사용 아궁이와 무쇠 솥, 모닥불 터가 있다. 전투 후에는 솥이 뒤집히고 막사가 불타거나 무너진다. (특징: 통나무를 박아 세운 목책; 넓게 다져진 흙바닥 (공터/훈련장); 돌과 진흙으로 만든 다수의 아궁이와 무쇠솥; 나무와 거적으로 엮어 만든 움막과 막사; 부서진 토기 밥그릇과 뒤집힌 무쇠솥 (전투 후); 불에 탄 나무 잔해와 흩어진 쌀밥 (전투 후))\n\nSCENE EVIDENCE (verbatim quotes from the screenplay about this place — treat them as evidence of what the location physically contains and looks like; stage the PLACE those moments happen in, but do NOT depict the momentary actions, people or staged props themselves):\n- 산채 공터.\n- 1. 오주일이 나뭇가지를 지휘봉처럼 들고 사내들 앞에 선다.\n- 훈련하던 규홍과 사내들이 돌아본다.\n- 해가 산등성이 너머로 기운다. 붉은 빛이 목책 사이로 길게 드리운다.\n- S#45. 파청 산채 / 밤\n- 달래의 시신이 짚멍석 위에 뉘어져 있다.\n- 산채 한가운데. 의병들이 둘러앉아 있다.\n- 옆에서 듣고 있던 임정현. 쌀자루를 세다 말고 손이 멈춘다.\n- 규홍과 관회가 박상현을 따라 산길을 내려간다.\n임정현은 그 뒷모습을 잠시 바라본다.\n\nTIME OF DAY (lock): sunset, low reddish sunlight.\n\nRender ONE photorealistic empty location photograph, 16:9, neutral enough that every shot of this place can be staged from it later. No readable writing anywhere: surfaces that would carry writing may be present, but stage any wording out of legibility — an oblique angle, distance, shallow focus. No captions, watermarks or overlay text.",
  "bg_path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/db00517b-165b-4005-a287-f28936b59b95/images/272bd7af-9294-4490-98c6-88a40e9deacc/scene/recipe/groupbg_mountain_camp_clearing_cb9567.png",
  "asset_id": "c5df4442-a576-4bdd-9e42-441cb4544a5e",
  "input_asset_ids": [
   "925c2002-e6b7-484f-b9c1-767408a0c3fb"
  ],
  "origin_tag": "S38sh3",
  "place_text": "The entrance edge of the mountain camp's dirt meeting clearing, crossed by red sunset light through the palisade.",
  "origin_inputs": {
   "place_text": "The entrance edge of the mountain camp's dirt meeting clearing, crossed by red sunset light through the palisade.",
   "time_of_day_en": "sunset, low reddish sunlight",
   "conti_asset_id": "925c2002-e6b7-484f-b9c1-767408a0c3fb"
  }
 },
 "S38sh3::bgfirst_bg": {
  "input_fingerprint": "5b8b41853e745cbd",
  "prompt": "Create the EMPTY BACKGROUND PLATE for one film shot — NO PEOPLE, no figures, no body parts, no sketch lines, no arrows anywhere.\n\"Empty\" means no people only: KEEP the location's inherent occupants and stock that define the place — animals in an animal shelter, pen or farm, goods in a market, moored boats in a harbour — unless the shot text explicitly removes them.\nThe FIRST attached image is a thin-line storyboard sketch: use ONLY its camera angle, horizon, perspective and the placement/size of buildings and set masses — ignore the sketched people and arrows entirely. The SECOND attached image (LOCATION PHOTOGRAPH) is the real place: take its architecture, materials, signage and fixed features, and RE-PROJECT them into the sketch's camera. If the photograph's camera differs from the sketch's, the sketch's camera wins.\nHUMAN-SCALE CALIBRATION: derive every structure's true size from human-scale elements — a door ≈ 2m, a window ≈ 1–1.5m wide, one storey ≈ 2.5–3m; never inflate a small structure or shrink a large one.\n\nSHOT TEXT this background must serve (Korean): 이마에 땀방울이 맺힌 채 한쪽 발을 공터 흙바닥에 막 디딘 복만의 역동적인 찰나.\n\nLOCATION (lock): The entrance edge of the mountain camp's dirt meeting clearing, crossed by red sunset light through the palisade.\n\nTIME OF DAY (lock): sunset, low reddish sunlight.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: wide shot\n- KEY BACKGROUND ELEMENTS: Clearing's earth ground (복만's forward foot has just made contact); used as Leave clear space beneath the whole figure so the arrival's weight transfer reads in one frame; Camp palisade (Present around the camp with gaps between its uprights) — Seen across the background beyond 복만; used as Provide the camp boundary behind the arrival without crowding the listener-facing space.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: The scripted red sunset light stretches through the palisade, revealing the sweat on 복만's forehead with restrained saturation and sober naturalistic contrast.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nRender ONE photorealistic empty location photograph, 16:9, that this shot can be staged inside later. No readable writing anywhere: surfaces that would carry writing may be present, but stage any wording out of legibility — an oblique angle, distance, shallow focus. No captions, watermarks or overlay text.",
  "effective_prompt": "Create the EMPTY BACKGROUND PLATE for one film shot — NO PEOPLE, no figures, no body parts, no sketch lines, no arrows anywhere.\n\"Empty\" means no people only: KEEP the location's inherent occupants and stock that define the place — animals in an animal shelter, pen or farm, goods in a market, moored boats in a harbour — unless the shot text explicitly removes them.\nThe FIRST attached image is a thin-line storyboard sketch: use ONLY its camera angle, horizon, perspective and the placement/size of buildings and set masses — ignore the sketched people and arrows entirely. The SECOND attached image (LOCATION PHOTOGRAPH) is the real place: take its architecture, materials, signage and fixed features, and RE-PROJECT them into the sketch's camera. If the photograph's camera differs from the sketch's, the sketch's camera wins.\nHUMAN-SCALE CALIBRATION: derive every structure's true size from human-scale elements — a door ≈ 2m, a window ≈ 1–1.5m wide, one storey ≈ 2.5–3m; never inflate a small structure or shrink a large one.\n\nSHOT TEXT this background must serve (Korean): 이마에 땀방울이 맺힌 채 한쪽 발을 공터 흙바닥에 막 디딘 복만의 역동적인 찰나.\n\nLOCATION (lock): The entrance edge of the mountain camp's dirt meeting clearing, crossed by red sunset light through the palisade.\n\nTIME OF DAY (lock): sunset, low reddish sunlight.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: wide shot\n- KEY BACKGROUND ELEMENTS: Clearing's earth ground (복만's forward foot has just made contact); used as Leave clear space beneath the whole figure so the arrival's weight transfer reads in one frame; Camp palisade (Present around the camp with gaps between its uprights) — Seen across the background beyond 복만; used as Provide the camp boundary behind the arrival without crowding the listener-facing space.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: The scripted red sunset light stretches through the palisade, revealing the sweat on 복만's forehead with restrained saturation and sober naturalistic contrast.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nRender ONE photorealistic empty location photograph, 16:9, that this shot can be staged inside later. No readable writing anywhere: surfaces that would carry writing may be present, but stage any wording out of legibility — an oblique angle, distance, shallow focus. No captions, watermarks or overlay text.",
  "bg_path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/db00517b-165b-4005-a287-f28936b59b95/images/272bd7af-9294-4490-98c6-88a40e9deacc/scene/recipe/S38sh3__bgfirst_bg.png",
  "asset_id": "9bb45b30-2a77-4fb4-bcd0-7d92ce98096a",
  "input_asset_ids": [
   "925c2002-e6b7-484f-b9c1-767408a0c3fb",
   "c5df4442-a576-4bdd-9e42-441cb4544a5e"
  ]
 },
 "S38sh3": {
  "input_fingerprint": "32f66dd0ff952bb9",
  "prompt": "Create ONE FINAL photorealistic live-action film still of the moment below — early 20th-century Korea, Jeolla region; all people are Korean unless explicitly stated otherwise. TIME OF DAY (lock): sunset, low reddish sunlight.\n\nSHOT TEXT (authoritative, Korean): 이마에 땀방울이 맺힌 채 한쪽 발을 공터 흙바닥에 막 디딘 복만의 역동적인 찰나.\n\nLOCATION (lock): The entrance edge of the mountain camp's dirt meeting clearing, crossed by red sunset light through the palisade. The shot takes place here — the attached LOCATION PHOTOGRAPH shows the exact spot.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: wide shot\n- KEY BACKGROUND ELEMENTS: Clearing's earth ground (복만's forward foot has just made contact); used as Leave clear space beneath the whole figure so the arrival's weight transfer reads in one frame; Camp palisade (Present around the camp with gaps between its uprights) — Seen across the background beyond 복만; used as Provide the camp boundary behind the arrival without crowding the listener-facing space.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: The scripted red sunset light stretches through the palisade, revealing the sweat on 복만's forehead with restrained saturation and sober naturalistic contrast.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): Long red sunset light falls through the camp palisade. 복만: He arrives breathless with sweat on his forehead, retaining the purchased salt and herbs.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 복만 (한국인 남성, 성인 나이대, 거칠게 자란 수염, 자연스러운 검은 머리) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.",
  "roll_prompts": {
   "A": "Stage the shot. The FIRST attached image (SHOT BACKGROUND) is the finished empty background of this shot — keep it EXACTLY: its camera, perspective, architecture, lighting and every fixed feature stay untouched. The SECOND attached image (LAYOUT SKETCH) tells you ONLY where the people go: each sketched person's position, screen size, pose and the gaze/motion arrows. Ignore the sketch's background lines. The CHARACTER REFERENCE photographs show the real people.\nPlace the real people into the background at exactly the sketched positions, sizes and poses, following the arrow directions. No sketch lines or arrows may remain.\n\nCreate ONE FINAL photorealistic live-action film still of the moment below — early 20th-century Korea, Jeolla region; all people are Korean unless explicitly stated otherwise. TIME OF DAY (lock): sunset, low reddish sunlight.\n\nSHOT TEXT (authoritative, Korean): 이마에 땀방울이 맺힌 채 한쪽 발을 공터 흙바닥에 막 디딘 복만의 역동적인 찰나.\n\nLOCATION (lock): The entrance edge of the mountain camp's dirt meeting clearing, crossed by red sunset light through the palisade. The shot takes place here — the FIRST attached image (SHOT BACKGROUND) is this exact place, already built: its ground, structures, horizon, materials and lighting are the finished truth of this location and must not be redesigned or replaced. No location photograph is attached — read the place from that image alone, and add no scenery, structure, vehicle or fixture that it does not already show. This lock governs the place only; the figures in the shot follow the staging and pose instructions.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: wide shot\n- KEY BACKGROUND ELEMENTS: Clearing's earth ground (복만's forward foot has just made contact); used as Leave clear space beneath the whole figure so the arrival's weight transfer reads in one frame; Camp palisade (Present around the camp with gaps between its uprights) — Seen across the background beyond 복만; used as Provide the camp boundary behind the arrival without crowding the listener-facing space.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: The scripted red sunset light stretches through the palisade, revealing the sweat on 복만's forehead with restrained saturation and sober naturalistic contrast.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): Long red sunset light falls through the camp palisade. 복만: He arrives breathless with sweat on his forehead, retaining the purchased salt and herbs.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 복만 (한국인 남성, 성인 나이대, 거칠게 자란 수염, 자연스러운 검은 머리) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.",
   "B": "Create ONE FINAL photorealistic live-action film still of the moment below — early 20th-century Korea, Jeolla region; all people are Korean unless explicitly stated otherwise. TIME OF DAY (lock): sunset, low reddish sunlight.\n\nSHOT TEXT (authoritative, Korean): 이마에 땀방울이 맺힌 채 한쪽 발을 공터 흙바닥에 막 디딘 복만의 역동적인 찰나.\n\nLOCATION (lock): The entrance edge of the mountain camp's dirt meeting clearing, crossed by red sunset light through the palisade. The shot takes place here — the attached LOCATION PHOTOGRAPH shows the exact spot.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: wide shot\n- KEY BACKGROUND ELEMENTS: Clearing's earth ground (복만's forward foot has just made contact); used as Leave clear space beneath the whole figure so the arrival's weight transfer reads in one frame; Camp palisade (Present around the camp with gaps between its uprights) — Seen across the background beyond 복만; used as Provide the camp boundary behind the arrival without crowding the listener-facing space.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: The scripted red sunset light stretches through the palisade, revealing the sweat on 복만's forehead with restrained saturation and sober naturalistic contrast.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): Long red sunset light falls through the camp palisade. 복만: He arrives breathless with sweat on his forehead, retaining the purchased salt and herbs.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 복만 (한국인 남성, 성인 나이대, 거칠게 자란 수염, 자연스러운 검은 머리) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.\n\nCOMPOSITION VARIATION (this candidate only): keep the place, the people and the moment exactly as specified, and keep from the CAMERA & FRAME contract its subject scale, screen placement and key background elements exactly. For this candidate only, its camera angle and height are deliberately left open: choose a camera position distinctly different from the obvious one, as a film director picking a second setup on the same blocking. Stay true to any explicit framing the SHOT TEXT itself dictates."
  },
  "roll_refs": {
   "A": [
    {
     "label": "SHOT BACKGROUND",
     "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/db00517b-165b-4005-a287-f28936b59b95/images/272bd7af-9294-4490-98c6-88a40e9deacc/scene/recipe/S38sh3__bgfirst_bg.png",
     "asset_id": "9bb45b30-2a77-4fb4-bcd0-7d92ce98096a",
     "role": "bgfirst_bg"
    },
    {
     "label": "LAYOUT SKETCH (people placement only)",
     "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/db00517b-165b-4005-a287-f28936b59b95/images/272bd7af-9294-4490-98c6-88a40e9deacc/conti/conti_S38sh3.png",
     "asset_id": "925c2002-e6b7-484f-b9c1-767408a0c3fb",
     "role": "conti_light"
    },
    {
     "label": "CHARACTER REFERENCE — 복만: the exact person appearing in this shot; match face, hair and build exactly.",
     "path": "<bytes:1027479>",
     "asset_id": "b7e69570-d15d-4241-b5ea-4cb8d4ece1cf",
     "role": "character_ref"
    }
   ],
   "B": [
    {
     "label": "LOCATION PHOTOGRAPH — the exact place of this shot: its architecture, materials, fixed features and lighting mood are spatial truth; stage the moment inside this place. Never copy its camera framing.",
     "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/db00517b-165b-4005-a287-f28936b59b95/images/272bd7af-9294-4490-98c6-88a40e9deacc/scene/recipe/groupbg_mountain_camp_clearing_cb9567.png",
     "asset_id": "c5df4442-a576-4bdd-9e42-441cb4544a5e",
     "role": "bgfirst_group_bg"
    },
    {
     "label": "CHARACTER REFERENCE — 복만: the exact person appearing in this shot; match face, hair and build exactly.",
     "path": "<bytes:1027479>",
     "asset_id": "b7e69570-d15d-4241-b5ea-4cb8d4ece1cf",
     "role": "character_ref"
    }
   ]
  },
  "shot_run_spend_attempt_count": 1,
  "judge_flip_skipped": "judge_fn_owns_order",
  "readings": [
   {
    "label": "A",
    "direction": "복만의 시선과 몸의 방향이 화면 왼쪽 앞을 향하고 있습니다.",
    "built_space": "배경 레퍼런스의 흙바닥, 목책, 가마솥, 오두막 등이 올바른 위치에 배치되어 있습니다.",
    "entities": "복만이 자루와 약초가 든 바구니를 양손에 들고 있으나, 화면 왼쪽 전경에 프롬프트에 없는 또 다른 인물의 뒷모습이 크게 자리 잡고 있습니다.",
    "hard_violations": [
     "[gemini-pro] 프롬프트에 없는 인물 추가 (화면 왼쪽 앞 인물 뒷모습)",
     "[gemini-pro] 물리적 지지 기반 부재 (달리는 자세이나 두 발이 모두 땅에서 떨어져 허공에 떠 있음)",
     "[gpt-high] 샷 텍스트가 등장시키지 않고 인물 제한에서도 허용하지 않은 남자 한 명을 왼쪽 전경에 추가했다."
    ],
    "physics": "달리는 듯한 자세를 취하고 있지만 지면에 닿은 발이 없고 몸을 띄운 도약의 역학도 부족하여 전체적으로 공중에 떠 있는 듯한 물리적 오류를 보입니다."
   },
   {
    "label": "B",
    "direction": "복만의 시선이 화면 왼쪽 바깥을 향하고 있습니다.",
    "built_space": "배경 레퍼런스와 일치하는 흙바닥, 목책, 가마솥 등이 잘 구현되었으며, 목책 틈으로 들어오는 붉은 햇빛이 바닥에 길게 드리워져 있습니다.",
    "entities": "레퍼런스와 복장, 외모가 일치하는 복만이 자루를 메고 단독으로 등장합니다. 약초의 형태는 명확히 보이지 않습니다.",
    "hard_violations": [],
    "physics": "앞발이 흙바닥에 막 닿으면서 먼지가 일고 뒷발은 공중에 떠 있어, 체중이 실리며 착지하는 찰나의 역학이 자연스럽게 표현되었습니다."
   }
  ],
  "cross_model_order": {
   "policy": "cross_model_order_v2_slot0_fwd_slot1_rev_parallel2",
   "slots": [
    {
     "model": "gemini-pro",
     "order": "forward",
     "display_to_canonical": {
      "A": "A",
      "B": "B"
     },
     "raw": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 9,
        "verdict_ko": "프롬프트가 요구한 단독 인물 구성, 붉은 노을빛 조명, 그리고 흙바닥에 발을 막 디딘 역동적인 순간을 매우 정확하고 사실적으로 구현했습니다."
       },
       {
        "label": "A",
        "score": 0,
        "verdict_ko": "프롬프트에 명시되지 않은 정체불명의 인물이 화면 왼쪽 앞에 크게 등장하는 치명적인 위반이 발생했습니다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "복만의 시선과 몸의 방향이 화면 왼쪽 앞을 향하고 있습니다.",
        "built_space": "배경 레퍼런스의 흙바닥, 목책, 가마솥, 오두막 등이 올바른 위치에 배치되어 있습니다.",
        "entities": "복만이 자루와 약초가 든 바구니를 양손에 들고 있으나, 화면 왼쪽 전경에 프롬프트에 없는 또 다른 인물의 뒷모습이 크게 자리 잡고 있습니다.",
        "hard_violations": [
         "프롬프트에 없는 인물 추가 (화면 왼쪽 앞 인물 뒷모습)",
         "물리적 지지 기반 부재 (달리는 자세이나 두 발이 모두 땅에서 떨어져 허공에 떠 있음)"
        ],
        "physics": "달리는 듯한 자세를 취하고 있지만 지면에 닿은 발이 없고 몸을 띄운 도약의 역학도 부족하여 전체적으로 공중에 떠 있는 듯한 물리적 오류를 보입니다."
       },
       {
        "label": "B",
        "direction": "복만의 시선이 화면 왼쪽 바깥을 향하고 있습니다.",
        "built_space": "배경 레퍼런스와 일치하는 흙바닥, 목책, 가마솥 등이 잘 구현되었으며, 목책 틈으로 들어오는 붉은 햇빛이 바닥에 길게 드리워져 있습니다.",
        "entities": "레퍼런스와 복장, 외모가 일치하는 복만이 자루를 메고 단독으로 등장합니다. 약초의 형태는 명확히 보이지 않습니다.",
        "hard_violations": [],
        "physics": "앞발이 흙바닥에 막 닿으면서 먼지가 일고 뒷발은 공중에 떠 있어, 체중이 실리며 착지하는 찰나의 역학이 자연스럽게 표현되었습니다."
       }
      ],
      "all_candidates_fail": false
     },
     "normalized": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 9,
        "verdict_ko": "프롬프트가 요구한 단독 인물 구성, 붉은 노을빛 조명, 그리고 흙바닥에 발을 막 디딘 역동적인 순간을 매우 정확하고 사실적으로 구현했습니다."
       },
       {
        "label": "A",
        "score": 0,
        "verdict_ko": "프롬프트에 명시되지 않은 정체불명의 인물이 화면 왼쪽 앞에 크게 등장하는 치명적인 위반이 발생했습니다."
       }
      ],
      "all_candidates_fail": false,
      "readings": [
       {
        "label": "A",
        "direction": "복만의 시선과 몸의 방향이 화면 왼쪽 앞을 향하고 있습니다.",
        "built_space": "배경 레퍼런스의 흙바닥, 목책, 가마솥, 오두막 등이 올바른 위치에 배치되어 있습니다.",
        "entities": "복만이 자루와 약초가 든 바구니를 양손에 들고 있으나, 화면 왼쪽 전경에 프롬프트에 없는 또 다른 인물의 뒷모습이 크게 자리 잡고 있습니다.",
        "hard_violations": [
         "프롬프트에 없는 인물 추가 (화면 왼쪽 앞 인물 뒷모습)",
         "물리적 지지 기반 부재 (달리는 자세이나 두 발이 모두 땅에서 떨어져 허공에 떠 있음)"
        ],
        "physics": "달리는 듯한 자세를 취하고 있지만 지면에 닿은 발이 없고 몸을 띄운 도약의 역학도 부족하여 전체적으로 공중에 떠 있는 듯한 물리적 오류를 보입니다."
       },
       {
        "label": "B",
        "direction": "복만의 시선이 화면 왼쪽 바깥을 향하고 있습니다.",
        "built_space": "배경 레퍼런스와 일치하는 흙바닥, 목책, 가마솥 등이 잘 구현되었으며, 목책 틈으로 들어오는 붉은 햇빛이 바닥에 길게 드리워져 있습니다.",
        "entities": "레퍼런스와 복장, 외모가 일치하는 복만이 자루를 메고 단독으로 등장합니다. 약초의 형태는 명확히 보이지 않습니다.",
        "hard_violations": [],
        "physics": "앞발이 흙바닥에 막 닿으면서 먼지가 일고 뒷발은 공중에 떠 있어, 체중이 실리며 착지하는 찰나의 역학이 자연스럽게 표현되었습니다."
       }
      ]
     },
     "ok": true
    },
    {
     "model": "gpt-high",
     "order": "reverse",
     "display_to_canonical": {
      "A": "B",
      "B": "A"
     },
     "raw": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 8,
        "verdict_ko": "복만만 등장하는 전신 와이드숏과 앞발 착지, 참조 의상·장소를 충실히 구현했으나 이마의 땀방울은 불분명하고 목책 사이 붉은빛은 다소 과포화되어 있다."
       },
       {
        "label": "B",
        "score": 2,
        "verdict_ko": "허용되지 않은 전경 인물을 추가해 어깨너머 구도로 바꾼 것이 결정적 위반이며, 복만의 모자와 긴 겉옷도 참조에서 벗어난다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "복만의 얼굴과 시선은 화면 왼쪽의 화면 밖 공터를 향하며, 몸도 그 방향으로 전진한다. 특정 청자는 보이지 않는다. 앞발은 왼쪽 앞으로 디뎠고 뒷발은 오른쪽 뒤로 들려 있어 도착 동작의 방향이 일관된다.",
        "built_space": "좌우 초가 구조물 두 동, 뒤쪽을 가로지르는 틈 있는 통나무 목책, 오른쪽 뒤 망루 한 기가 보인다. 왼쪽 전경의 큰 가마솥 아궁이 한 곳과 오른쪽 뒤 작은 취사시설 한 곳도 참조 배치에 대응한다. 복만은 시설과 겹쳐 서 있지 않고 공터 흙바닥 위에 있으며, 전신과 발 아래 지면이 함께 담겼다. 반사는 없다.",
        "entities": "인물은 복만 한 명이다. 성인 한국인 남성이라는 설정에 부합하는 외모, 검은 머리와 거친 수염, 참조와 유사한 얼굴을 보인다. 검은 챙모자, 긴 짙은 겉옷, 회색 속옷, 허리끈과 헐렁한 바지 및 천 신발이 참조에 가깝다. 양옆에 천 주머니 두 개를 지녔으나 내용물이 가려져 소금과 약초인지는 확인할 수 없다. 얼굴에는 윤기가 있지만 이마의 개별 땀방울은 뚜렷하지 않다. 읽을 수 있는 글자는 없다.",
        "hard_violations": [],
        "physics": "앞발의 밑창이 흙바닥에 닿아 체중을 받으며, 상체가 앞으로 기울고 뒷무릎과 발이 들린 자세는 달려오다 한 발을 디딘 순간으로 가능하다. 뒷발 아래 먼지는 직전 지면 차기의 결과로 읽힌다. 천 주머니는 몸에 걸친 끈과 손으로 지지되며, 지지 없이 떠 있는 물체나 신체는 없다."
       },
       {
        "label": "B",
        "direction": "복만은 화면 왼쪽 전경에 있는 남자를 바라보며 그에게 달려온다. 전경 남자도 복만을 향해 고개를 돌렸다. 상호 시선은 맞지만, 그 시선의 표적인 청자는 이 장면에서 허용되지 않은 추가 인물이다.",
        "built_space": "초가 구조물 두 동, 배경 목책, 오른쪽 뒤 망루 한 기가 참조와 유사하게 놓였다. 왼쪽 전경 취사시설과 오른쪽 뒤 아궁이도 보이며, 전경에는 참조에 있는 그릇과 탄 목재가 남아 있다. 복만은 공터 중앙에서 진입하지만 추가 남자가 왼쪽 전경을 크게 차지하여 단독 도착 와이드숏을 어깨너머 구도로 바꾼다. 반사는 없다.",
        "entities": "복만 외에 머리와 상체가 보이는 성인 남자 한 명이 추가되었다. 복만은 검은 머리와 수염을 지닌 성인 한국인 남성 설정에 부합하지만 참조의 챙모자가 없고 겉옷도 훨씬 짧다. 한 손에는 천 자루, 다른 손에는 녹색 식물이 든 바구니가 있어 약초 운반은 비교적 명확하나 자루 속 소금은 확인되지 않는다. 이마의 땀방울은 선명하지 않다. 읽을 수 있는 글자는 없다.",
        "hard_violations": [
         "샷 텍스트가 등장시키지 않고 인물 제한에서도 허용하지 않은 남자 한 명을 왼쪽 전경에 추가했다."
        ],
        "physics": "앞발은 발끝을 든 채 뒤꿈치부터 흙바닥에 닿으려는 낮은 위치에 있고, 먼지가 접촉 경계를 일부 가린다. 뒷다리는 굽혀 들려 있어 달리기의 착지 직전 또는 착지 순간으로 가능한 자세이며, 근거 없는 공중 부양은 아니다. 천 자루는 한 손의 매듭 파지로, 바구니는 다른 손의 손잡이 파지로 지지된다. 전경 인물의 하체는 화면 밖이므로 지지 상태를 판정할 수 없다."
       }
      ],
      "all_candidates_fail": false
     },
     "normalized": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 8,
        "verdict_ko": "복만만 등장하는 전신 와이드숏과 앞발 착지, 참조 의상·장소를 충실히 구현했으나 이마의 땀방울은 불분명하고 목책 사이 붉은빛은 다소 과포화되어 있다."
       },
       {
        "label": "A",
        "score": 2,
        "verdict_ko": "허용되지 않은 전경 인물을 추가해 어깨너머 구도로 바꾼 것이 결정적 위반이며, 복만의 모자와 긴 겉옷도 참조에서 벗어난다."
       }
      ],
      "all_candidates_fail": false,
      "readings": [
       {
        "label": "B",
        "direction": "복만의 얼굴과 시선은 화면 왼쪽의 화면 밖 공터를 향하며, 몸도 그 방향으로 전진한다. 특정 청자는 보이지 않는다. 앞발은 왼쪽 앞으로 디뎠고 뒷발은 오른쪽 뒤로 들려 있어 도착 동작의 방향이 일관된다.",
        "built_space": "좌우 초가 구조물 두 동, 뒤쪽을 가로지르는 틈 있는 통나무 목책, 오른쪽 뒤 망루 한 기가 보인다. 왼쪽 전경의 큰 가마솥 아궁이 한 곳과 오른쪽 뒤 작은 취사시설 한 곳도 참조 배치에 대응한다. 복만은 시설과 겹쳐 서 있지 않고 공터 흙바닥 위에 있으며, 전신과 발 아래 지면이 함께 담겼다. 반사는 없다.",
        "entities": "인물은 복만 한 명이다. 성인 한국인 남성이라는 설정에 부합하는 외모, 검은 머리와 거친 수염, 참조와 유사한 얼굴을 보인다. 검은 챙모자, 긴 짙은 겉옷, 회색 속옷, 허리끈과 헐렁한 바지 및 천 신발이 참조에 가깝다. 양옆에 천 주머니 두 개를 지녔으나 내용물이 가려져 소금과 약초인지는 확인할 수 없다. 얼굴에는 윤기가 있지만 이마의 개별 땀방울은 뚜렷하지 않다. 읽을 수 있는 글자는 없다.",
        "hard_violations": [],
        "physics": "앞발의 밑창이 흙바닥에 닿아 체중을 받으며, 상체가 앞으로 기울고 뒷무릎과 발이 들린 자세는 달려오다 한 발을 디딘 순간으로 가능하다. 뒷발 아래 먼지는 직전 지면 차기의 결과로 읽힌다. 천 주머니는 몸에 걸친 끈과 손으로 지지되며, 지지 없이 떠 있는 물체나 신체는 없다."
       },
       {
        "label": "A",
        "direction": "복만은 화면 왼쪽 전경에 있는 남자를 바라보며 그에게 달려온다. 전경 남자도 복만을 향해 고개를 돌렸다. 상호 시선은 맞지만, 그 시선의 표적인 청자는 이 장면에서 허용되지 않은 추가 인물이다.",
        "built_space": "초가 구조물 두 동, 배경 목책, 오른쪽 뒤 망루 한 기가 참조와 유사하게 놓였다. 왼쪽 전경 취사시설과 오른쪽 뒤 아궁이도 보이며, 전경에는 참조에 있는 그릇과 탄 목재가 남아 있다. 복만은 공터 중앙에서 진입하지만 추가 남자가 왼쪽 전경을 크게 차지하여 단독 도착 와이드숏을 어깨너머 구도로 바꾼다. 반사는 없다.",
        "entities": "복만 외에 머리와 상체가 보이는 성인 남자 한 명이 추가되었다. 복만은 검은 머리와 수염을 지닌 성인 한국인 남성 설정에 부합하지만 참조의 챙모자가 없고 겉옷도 훨씬 짧다. 한 손에는 천 자루, 다른 손에는 녹색 식물이 든 바구니가 있어 약초 운반은 비교적 명확하나 자루 속 소금은 확인되지 않는다. 이마의 땀방울은 선명하지 않다. 읽을 수 있는 글자는 없다.",
        "hard_violations": [
         "샷 텍스트가 등장시키지 않고 인물 제한에서도 허용하지 않은 남자 한 명을 왼쪽 전경에 추가했다."
        ],
        "physics": "앞발은 발끝을 든 채 뒤꿈치부터 흙바닥에 닿으려는 낮은 위치에 있고, 먼지가 접촉 경계를 일부 가린다. 뒷다리는 굽혀 들려 있어 달리기의 착지 직전 또는 착지 순간으로 가능한 자세이며, 근거 없는 공중 부양은 아니다. 천 자루는 한 손의 매듭 파지로, 바구니는 다른 손의 손잡이 파지로 지지된다. 전경 인물의 하체는 화면 밖이므로 지지 상태를 판정할 수 없다."
       }
      ]
     },
     "ok": true
    }
   ],
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "slot_winner_match": true,
   "slot_winner": {
    "gemini-pro": "B",
    "gpt-high": "B"
   },
   "route": "cross_slot_agree"
  },
  "dual": {
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "normalized": {
    "A": 0.25,
    "B": 2.0
   },
   "adjusted": {
    "A": 0.0,
    "B": 2.0
   },
   "violations": {
    "A": [
     "[gemini-pro] 프롬프트에 없는 인물 추가 (화면 왼쪽 앞 인물 뒷모습)",
     "[gemini-pro] 물리적 지지 기반 부재 (달리는 자세이나 두 발이 모두 땅에서 떨어져 허공에 떠 있음)",
     "[gpt-high] 샷 텍스트가 등장시키지 않고 인물 제한에서도 허용하지 않은 남자 한 명을 왼쪽 전경에 추가했다."
    ]
   },
   "per_model_winner": {
    "gemini-pro": "B",
    "gpt-high": "B"
   },
   "agreed": true
  },
  "totals": {
   "B": 2000,
   "A": 0
  },
  "selected": "B",
  "ranking": [
   "B",
   "A"
  ],
  "verdicts": [
   {
    "label": "B",
    "score": 2000,
    "verdict_ko": "프롬프트가 요구한 단독 인물 구성, 붉은 노을빛 조명, 그리고 흙바닥에 발을 막 디딘 역동적인 순간을 매우 정확하고 사실적으로 구현했습니다."
   },
   {
    "label": "A",
    "score": 0,
    "verdict_ko": "프롬프트에 명시되지 않은 정체불명의 인물이 화면 왼쪽 앞에 크게 등장하는 치명적인 위반이 발생했습니다.  ★위반: [gemini-pro] 프롬프트에 없는 인물 추가 (화면 왼쪽 앞 인물 뒷모습) / [gemini-pro] 물리적 지지 기반 부재 (달리는 자세이나 두 발이 모두 땅에서 떨어져 허공에 떠 있음) / [gpt-high] 샷 텍스트가 등장시키지 않고 인물 제한에서도 허용하지 않은 남자 한 명을 왼쪽 전경에 추가했다."
   }
  ],
  "refs": [
   {
    "label": "LOCATION PHOTOGRAPH — the exact place of this shot: its architecture, materials, fixed features and lighting mood are spatial truth; stage the moment inside this place. Never copy its camera framing.",
    "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/db00517b-165b-4005-a287-f28936b59b95/images/272bd7af-9294-4490-98c6-88a40e9deacc/scene/recipe/groupbg_mountain_camp_clearing_cb9567.png",
    "asset_id": "c5df4442-a576-4bdd-9e42-441cb4544a5e",
    "role": "bgfirst_group_bg"
   },
   {
    "label": "CHARACTER REFERENCE — 복만: the exact person appearing in this shot; match face, hair and build exactly.",
    "path": "<bytes:1027479>",
    "asset_id": "b7e69570-d15d-4241-b5ea-4cb8d4ece1cf",
    "role": "character_ref"
   }
  ],
  "critique_skipped": true,
  "shot_run_produced": false,
  "shot_run_uid": "06aa1428-5f8b-7d55-b71d-0114ba7fb9b7",
  "bgfirst": {
   "bg_path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/db00517b-165b-4005-a287-f28936b59b95/images/272bd7af-9294-4490-98c6-88a40e9deacc/scene/recipe/S38sh3__bgfirst_bg.png",
   "bg_asset_id": "9bb45b30-2a77-4fb4-bcd0-7d92ce98096a",
   "bg_record_key": "S38sh3::bgfirst_bg",
   "chain_winner": false,
   "authority": "groupbg",
   "group_key": "mountain_camp_clearing",
   "groupbg_asset_id": "c5df4442-a576-4bdd-9e42-441cb4544a5e"
  },
  "ref_mode": "그룹 배경+엔티티 (2택1: 무콘티 승)",
  "share_plan": {
   "ref_plan": "background"
  }
 },
 "S38sh3::cine": {
  "applied": true,
  "attempted_at": "2026-09-09T05:08:57.652840+00:00",
  "fingerprint": "4857fdeb0e34ca8bf47450e06aa5f57916fa744b11e10af957acbaf8208400cc",
  "fingerprint_version": 2,
  "provider": "mai",
  "endpoint": "openrouter/chat-completions",
  "model": "microsoft/mai-image-2.6",
  "pack": "24.202608252115",
  "source_file": "S38sh3_sel.png",
  "source_sha256": "5608204f530123ce3b7cb3186b911585f3acb1543252abc6f0c2f3c4e7d1d081",
  "file": "S38sh3_cine.png",
  "staged_sha256": "22083d99afe6f81da5b0bccd3062f0cd100c1c7ef8015807d1041e2f903198d9",
  "latency_ms": 27868
 },
 "S38sh13::signage": {
  "fp": "820264363826fb2b",
  "inscriptions": [],
  "cues": [],
  "dropped": []
 },
 "S38sh13": {
  "input_fingerprint": "4f9766aae4857bfb",
  "prompt": "Create ONE FINAL photorealistic live-action film still of the moment below — early 20th-century Korea, Jeolla region; all people are Korean unless explicitly stated otherwise. TIME OF DAY (lock): sunset, low reddish sunlight.\n\nSHOT TEXT (authoritative, Korean): 일행을 향해 결연한 눈빛으로 작전을 알리는 안규홍의 다부진 상체.\n\nLOCATION (lock): The outdoor planning circle on the mountain camp's dirt clearing, near the palisade in fading red daylight. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: medium shot\n- KEY BACKGROUND ELEMENTS: 목책 (Visible behind the seated speaker, with gaps between its upright elements) — Seen obliquely behind 안규홍; used as Provides a subdued background rhythm without competing with his face.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: The scripted red sunset stretches through the palisade with restrained saturation and natural contrast, preserving warmth beneath the sober instructions.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The people visible in that photo are NOT in this shot: never carry their faces, bodies or clothing onto anyone here. Each person in THIS shot is defined solely by the PEOPLE list, their own CHARACTER REFERENCE images and their wardrobe notes. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nPREVIOUS STILL USAGE (follow exactly — what to take from the attached still and what to exclude; it governs place and objects only, never who is in this shot or how the camera sees them): Take the dirt ground, wooden palisade where visible, and reddish evening light. Exclude dust from the messenger's hurried arrival and any duplicated travel supplies.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): Red sunset light continues through the palisade. 안규홍: He remains seated with the jade ring retained in his clothing.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 안규홍 (한국인 남성, 청년 나이대, 검은 머리, 단단히 묶은 머리) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.",
  "roll_prompts": {
   "A": "Create ONE FINAL photorealistic live-action film still of the moment below — early 20th-century Korea, Jeolla region; all people are Korean unless explicitly stated otherwise. TIME OF DAY (lock): sunset, low reddish sunlight.\n\nSHOT TEXT (authoritative, Korean): 일행을 향해 결연한 눈빛으로 작전을 알리는 안규홍의 다부진 상체.\n\nLOCATION (lock): The outdoor planning circle on the mountain camp's dirt clearing, near the palisade in fading red daylight. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: medium shot\n- KEY BACKGROUND ELEMENTS: 목책 (Visible behind the seated speaker, with gaps between its upright elements) — Seen obliquely behind 안규홍; used as Provides a subdued background rhythm without competing with his face.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: The scripted red sunset stretches through the palisade with restrained saturation and natural contrast, preserving warmth beneath the sober instructions.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The people visible in that photo are NOT in this shot: never carry their faces, bodies or clothing onto anyone here. Each person in THIS shot is defined solely by the PEOPLE list, their own CHARACTER REFERENCE images and their wardrobe notes. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nPREVIOUS STILL USAGE (follow exactly — what to take from the attached still and what to exclude; it governs place and objects only, never who is in this shot or how the camera sees them): Take the dirt ground, wooden palisade where visible, and reddish evening light. Exclude dust from the messenger's hurried arrival and any duplicated travel supplies.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): Red sunset light continues through the palisade. 안규홍: He remains seated with the jade ring retained in his clothing.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 안규홍 (한국인 남성, 청년 나이대, 검은 머리, 단단히 묶은 머리) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.",
   "B": "Create ONE FINAL photorealistic live-action film still of the moment below — early 20th-century Korea, Jeolla region; all people are Korean unless explicitly stated otherwise. TIME OF DAY (lock): sunset, low reddish sunlight.\n\nSHOT TEXT (authoritative, Korean): 일행을 향해 결연한 눈빛으로 작전을 알리는 안규홍의 다부진 상체.\n\nLOCATION (lock): The outdoor planning circle on the mountain camp's dirt clearing, near the palisade in fading red daylight. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: medium shot\n- KEY BACKGROUND ELEMENTS: 목책 (Visible behind the seated speaker, with gaps between its upright elements) — Seen obliquely behind 안규홍; used as Provides a subdued background rhythm without competing with his face.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: The scripted red sunset stretches through the palisade with restrained saturation and natural contrast, preserving warmth beneath the sober instructions.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The people visible in that photo are NOT in this shot: never carry their faces, bodies or clothing onto anyone here. Each person in THIS shot is defined solely by the PEOPLE list, their own CHARACTER REFERENCE images and their wardrobe notes. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nPREVIOUS STILL USAGE (follow exactly — what to take from the attached still and what to exclude; it governs place and objects only, never who is in this shot or how the camera sees them): Take the dirt ground, wooden palisade where visible, and reddish evening light. Exclude dust from the messenger's hurried arrival and any duplicated travel supplies.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): Red sunset light continues through the palisade. 안규홍: He remains seated with the jade ring retained in his clothing.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 안규홍 (한국인 남성, 청년 나이대, 검은 머리, 단단히 묶은 머리) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.\n\nCOMPOSITION VARIATION (this candidate only): keep the place, the people and the moment exactly as specified, and keep from the CAMERA & FRAME contract its subject scale, screen placement and key background elements exactly. For this candidate only, its camera angle and height are deliberately left open: choose a camera position distinctly different from the obvious one, as a film director picking a second setup on the same blocking. Stay true to any explicit framing the SHOT TEXT itself dictates."
  },
  "shot_run_spend_attempt_count": 1,
  "readings": [
   {
    "label": "A",
    "direction": "안규홍의 시선은 화면 밖 왼쪽의 보이지 않는 대상을 향해 결연하게 고정되어 있습니다.",
    "built_space": "배경에 나무 목책이 비스듬하게 세워져 있고 그 틈새로 붉은 빛이 스며들며, 바닥은 흙바닥으로 이전 샷의 환경(빛의 그림자 등)이 일관되게 잘 유지되어 있습니다.",
    "entities": "안규홍의 얼굴, 체형, 헤어스타일, 의상은 제공된 레퍼런스와 정확히 일치하며, 옥가락지는 옷 안이 아닌 그의 왼쪽 손에 쥐어져 있습니다.",
    "hard_violations": [],
    "physics": "작은 나무 의자 위에 안정적으로 앉아 있으며 양발은 흙바닥을 단단히 딛고 있어 자세와 무게 중심이 자연스럽게 지탱됩니다."
   },
   {
    "label": "B",
    "direction": "안규홍이 전경에 위치한 두 명의 인물을 바라보며 왼손으로 무언가 설명하듯 제스처를 취하고 있습니다.",
    "built_space": "흙바닥 공터, 뒤편의 목책 및 이전 샷에 있던 가마솥과 초가집 등 지정된 구조물들이 잘 배치되어 있으며 붉은 조명이 적용되었습니다.",
    "entities": "안규홍의 모습은 레퍼런스와 일치하나, 프롬프트가 명시적으로 금지한 추가 인물 2명이 화면 앞쪽 양옆에 배치되어 있습니다.",
    "hard_violations": [
     "[gemini-pro] invented people (프롬프트의 'never anyone else' 지침을 어기고 전경에 2명의 인물을 임의로 추가함)",
     "[gpt-high] 화면 양쪽 전경에 명단에서 허용하지 않은 청자 두 명의 머리와 상체를 추가했다."
    ],
    "physics": "앉은 자세로 하반신이 지지되고 있으며, 허공에 제스처를 취하는 왼손과 허벅지 위에 놓인 오른손의 움직임은 물리적으로 자연스럽습니다."
   }
  ],
  "cross_model_order": {
   "policy": "cross_model_order_v2_slot0_fwd_slot1_rev_parallel2",
   "slots": [
    {
     "model": "gemini-pro",
     "order": "forward",
     "display_to_canonical": {
      "A": "A",
      "B": "B"
     },
     "raw": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 9,
        "verdict_ko": "명시된 인물(안규홍)만을 미디엄 샷으로 담아냈으며, 요구된 붉은 저녁 노을과 목책 배경의 디테일을 완벽하게 준수하여 프롬프트의 의도를 정확히 구현했습니다."
       },
       {
        "label": "B",
        "score": 2,
        "verdict_ko": "프롬프트가 안규홍 외 다른 인물의 등장을 명시적으로 금지했음에도 화면 전경에 두 명의 인물을 임의로 추가하는 치명적인 위반을 범했습니다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "안규홍의 시선은 화면 밖 왼쪽의 보이지 않는 대상을 향해 결연하게 고정되어 있습니다.",
        "built_space": "배경에 나무 목책이 비스듬하게 세워져 있고 그 틈새로 붉은 빛이 스며들며, 바닥은 흙바닥으로 이전 샷의 환경(빛의 그림자 등)이 일관되게 잘 유지되어 있습니다.",
        "entities": "안규홍의 얼굴, 체형, 헤어스타일, 의상은 제공된 레퍼런스와 정확히 일치하며, 옥가락지는 옷 안이 아닌 그의 왼쪽 손에 쥐어져 있습니다.",
        "hard_violations": [],
        "physics": "작은 나무 의자 위에 안정적으로 앉아 있으며 양발은 흙바닥을 단단히 딛고 있어 자세와 무게 중심이 자연스럽게 지탱됩니다."
       },
       {
        "label": "B",
        "direction": "안규홍이 전경에 위치한 두 명의 인물을 바라보며 왼손으로 무언가 설명하듯 제스처를 취하고 있습니다.",
        "built_space": "흙바닥 공터, 뒤편의 목책 및 이전 샷에 있던 가마솥과 초가집 등 지정된 구조물들이 잘 배치되어 있으며 붉은 조명이 적용되었습니다.",
        "entities": "안규홍의 모습은 레퍼런스와 일치하나, 프롬프트가 명시적으로 금지한 추가 인물 2명이 화면 앞쪽 양옆에 배치되어 있습니다.",
        "hard_violations": [
         "invented people (프롬프트의 'never anyone else' 지침을 어기고 전경에 2명의 인물을 임의로 추가함)"
        ],
        "physics": "앉은 자세로 하반신이 지지되고 있으며, 허공에 제스처를 취하는 왼손과 허벅지 위에 놓인 오른손의 움직임은 물리적으로 자연스럽습니다."
       }
      ],
      "all_candidates_fail": false
     },
     "normalized": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 9,
        "verdict_ko": "명시된 인물(안규홍)만을 미디엄 샷으로 담아냈으며, 요구된 붉은 저녁 노을과 목책 배경의 디테일을 완벽하게 준수하여 프롬프트의 의도를 정확히 구현했습니다."
       },
       {
        "label": "B",
        "score": 2,
        "verdict_ko": "프롬프트가 안규홍 외 다른 인물의 등장을 명시적으로 금지했음에도 화면 전경에 두 명의 인물을 임의로 추가하는 치명적인 위반을 범했습니다."
       }
      ],
      "all_candidates_fail": false,
      "readings": [
       {
        "label": "A",
        "direction": "안규홍의 시선은 화면 밖 왼쪽의 보이지 않는 대상을 향해 결연하게 고정되어 있습니다.",
        "built_space": "배경에 나무 목책이 비스듬하게 세워져 있고 그 틈새로 붉은 빛이 스며들며, 바닥은 흙바닥으로 이전 샷의 환경(빛의 그림자 등)이 일관되게 잘 유지되어 있습니다.",
        "entities": "안규홍의 얼굴, 체형, 헤어스타일, 의상은 제공된 레퍼런스와 정확히 일치하며, 옥가락지는 옷 안이 아닌 그의 왼쪽 손에 쥐어져 있습니다.",
        "hard_violations": [],
        "physics": "작은 나무 의자 위에 안정적으로 앉아 있으며 양발은 흙바닥을 단단히 딛고 있어 자세와 무게 중심이 자연스럽게 지탱됩니다."
       },
       {
        "label": "B",
        "direction": "안규홍이 전경에 위치한 두 명의 인물을 바라보며 왼손으로 무언가 설명하듯 제스처를 취하고 있습니다.",
        "built_space": "흙바닥 공터, 뒤편의 목책 및 이전 샷에 있던 가마솥과 초가집 등 지정된 구조물들이 잘 배치되어 있으며 붉은 조명이 적용되었습니다.",
        "entities": "안규홍의 모습은 레퍼런스와 일치하나, 프롬프트가 명시적으로 금지한 추가 인물 2명이 화면 앞쪽 양옆에 배치되어 있습니다.",
        "hard_violations": [
         "invented people (프롬프트의 'never anyone else' 지침을 어기고 전경에 2명의 인물을 임의로 추가함)"
        ],
        "physics": "앉은 자세로 하반신이 지지되고 있으며, 허공에 제스처를 취하는 왼손과 허벅지 위에 놓인 오른손의 움직임은 물리적으로 자연스럽습니다."
       }
      ]
     },
     "ok": true
    },
    {
     "model": "gpt-high",
     "order": "reverse",
     "display_to_canonical": {
      "A": "B",
      "B": "A"
     },
     "raw": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 2,
        "verdict_ko": "결연하게 작전을 설명하는 표정과 손짓은 맞지만, 허용되지 않은 청자 두 명을 전경에 추가하여 단독 인물 조건을 위반했다."
       },
       {
        "label": "B",
        "score": 8,
        "verdict_ko": "안규홍만 앉힌 채 화면 밖 일행에게 말하는 순간과 붉은 목책 배경을 충실히 구현했지만, 옥가락지를 옷 속이 아닌 손가락에 드러냈다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "안규홍의 얼굴과 시선은 화면 왼쪽 전경의 청자를 향하고, 입을 벌린 채 가슴 앞에 손을 들어 설명한다. 전경 양쪽 인물도 안규홍을 향한다. 작전을 전달하는 방향은 명확하지만, 그 대상을 실제 인물로 추가한 것은 허용되지 않는다. 무기나 별도의 지시 도구는 없다.",
        "built_space": "흙 공터 뒤로 수십 개의 간격 있는 세로 통나무와 가로 연결대가 하나의 목책을 이루며 비스듬히 이어진다. 왼쪽에는 초가 구조물 일부, 솥이 놓인 화덕 하나와 별도 용기들이 보여 이전 장소와 대체로 연결된다. 안규홍은 중앙 오른쪽에 앉아 있고, 추가된 두 사람의 머리와 어깨가 양쪽 전경을 크게 차지한다. 무릎까지 포함하여 요청한 상체 중심 화면보다 넓다. 반사면은 없다.",
        "entities": "중앙 인물은 검은 머리를 뒤로 단단히 묶은 청년 남성으로, 참고의 한국인 안규홍과 얼굴·체격이 대체로 닮았다. 때 묻은 미색 저고리와 바지, 천 허리띠도 부합한다. 옥가락지는 보이지 않아 옷 속 보관 여부는 확인할 수 없다. 양쪽 전경에는 명단에 없는 검은 머리의 남성 두 명이 부분적으로 등장한다. 읽을 수 있는 글자나 도식은 없다.",
        "hard_violations": [
         "화면 양쪽 전경에 명단에서 허용하지 않은 청자 두 명의 머리와 상체를 추가했다."
        ],
        "physics": "벌어진 무릎과 접힌 다리, 낮은 골반은 앉은 자세로 자연스럽게 이어지지만 좌면과 엉덩이 접촉점은 가려져 있다. 내린 손은 무릎에 걸쳐 있고, 든 손은 굽힌 팔로 지지되는 설명 동작이다. 전경 인물들의 하체와 지지면은 화면 밖이다. 배경 솥과 용기는 화덕 또는 지면 위에 놓여 있으며 공중에 뜬 물체는 없다."
       },
       {
        "label": "B",
        "direction": "안규홍은 고개와 눈을 화면 왼쪽 바깥으로 돌리고 입을 조금 벌려 말한다. 보이는 청자는 없지만 화면 밖 일행에게 작전을 전달하는 시선으로 읽히며, 렌즈를 응시하지 않는다. 두 손은 다리 위에 있고 무기나 가리키는 도구는 없다.",
        "built_space": "흙 공터 뒤에 하나의 연속된 목책이 비스듬히 놓여 있다. 수십 개의 불규칙한 세로 말뚝 사이로 붉은 빛이 들어오고, 가로 연결대와 왼쪽 사선 버팀목이 보인다. 안규홍은 목책 앞의 낮은 목제 좌석 하나에 앉아 있으며 다른 사람은 없다. 머리부터 무릎 부근까지 담아 상체 중심 미디엄 숏보다 하체를 다소 많이 포함한다. 목책의 재료와 흙바닥, 붉은 저녁빛은 장소 참고와 부합하며 반사면은 없다.",
        "entities": "검은 머리를 뒤로 묶은 청년 남성 한 명만 있으며, 참고 안규홍의 얼굴과 체격에 대체로 부합한다. 낡고 오염된 미색 저고리·바지와 천 허리띠도 일치한다. 다만 초록색 고리 형태의 옥가락지가 화면 오른쪽 손의 손가락에 드러나 있어, 옷 속에 보관한다는 상태 지시와 다르다. 추가 인물이나 읽을 수 있는 글자는 없다.",
        "hard_violations": [],
        "physics": "엉덩이 아래 목제 좌면과 그 아래 지지 다리가 보여 앉은 몸의 하중을 받는다. 다리는 좌면에서 자연스럽게 내려가며 발은 화면 밖이다. 양손은 각각 허벅지와 무릎에 닿아 있고 옥가락지는 손가락에 끼워져 지지된다. 앞으로 약간 기울인 상체도 앉아서 말하는 동작으로 가능하며, 지지 없이 떠 있는 신체나 물체는 없다."
       }
      ],
      "all_candidates_fail": false
     },
     "normalized": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 2,
        "verdict_ko": "결연하게 작전을 설명하는 표정과 손짓은 맞지만, 허용되지 않은 청자 두 명을 전경에 추가하여 단독 인물 조건을 위반했다."
       },
       {
        "label": "A",
        "score": 8,
        "verdict_ko": "안규홍만 앉힌 채 화면 밖 일행에게 말하는 순간과 붉은 목책 배경을 충실히 구현했지만, 옥가락지를 옷 속이 아닌 손가락에 드러냈다."
       }
      ],
      "all_candidates_fail": false,
      "readings": [
       {
        "label": "B",
        "direction": "안규홍의 얼굴과 시선은 화면 왼쪽 전경의 청자를 향하고, 입을 벌린 채 가슴 앞에 손을 들어 설명한다. 전경 양쪽 인물도 안규홍을 향한다. 작전을 전달하는 방향은 명확하지만, 그 대상을 실제 인물로 추가한 것은 허용되지 않는다. 무기나 별도의 지시 도구는 없다.",
        "built_space": "흙 공터 뒤로 수십 개의 간격 있는 세로 통나무와 가로 연결대가 하나의 목책을 이루며 비스듬히 이어진다. 왼쪽에는 초가 구조물 일부, 솥이 놓인 화덕 하나와 별도 용기들이 보여 이전 장소와 대체로 연결된다. 안규홍은 중앙 오른쪽에 앉아 있고, 추가된 두 사람의 머리와 어깨가 양쪽 전경을 크게 차지한다. 무릎까지 포함하여 요청한 상체 중심 화면보다 넓다. 반사면은 없다.",
        "entities": "중앙 인물은 검은 머리를 뒤로 단단히 묶은 청년 남성으로, 참고의 한국인 안규홍과 얼굴·체격이 대체로 닮았다. 때 묻은 미색 저고리와 바지, 천 허리띠도 부합한다. 옥가락지는 보이지 않아 옷 속 보관 여부는 확인할 수 없다. 양쪽 전경에는 명단에 없는 검은 머리의 남성 두 명이 부분적으로 등장한다. 읽을 수 있는 글자나 도식은 없다.",
        "hard_violations": [
         "화면 양쪽 전경에 명단에서 허용하지 않은 청자 두 명의 머리와 상체를 추가했다."
        ],
        "physics": "벌어진 무릎과 접힌 다리, 낮은 골반은 앉은 자세로 자연스럽게 이어지지만 좌면과 엉덩이 접촉점은 가려져 있다. 내린 손은 무릎에 걸쳐 있고, 든 손은 굽힌 팔로 지지되는 설명 동작이다. 전경 인물들의 하체와 지지면은 화면 밖이다. 배경 솥과 용기는 화덕 또는 지면 위에 놓여 있으며 공중에 뜬 물체는 없다."
       },
       {
        "label": "A",
        "direction": "안규홍은 고개와 눈을 화면 왼쪽 바깥으로 돌리고 입을 조금 벌려 말한다. 보이는 청자는 없지만 화면 밖 일행에게 작전을 전달하는 시선으로 읽히며, 렌즈를 응시하지 않는다. 두 손은 다리 위에 있고 무기나 가리키는 도구는 없다.",
        "built_space": "흙 공터 뒤에 하나의 연속된 목책이 비스듬히 놓여 있다. 수십 개의 불규칙한 세로 말뚝 사이로 붉은 빛이 들어오고, 가로 연결대와 왼쪽 사선 버팀목이 보인다. 안규홍은 목책 앞의 낮은 목제 좌석 하나에 앉아 있으며 다른 사람은 없다. 머리부터 무릎 부근까지 담아 상체 중심 미디엄 숏보다 하체를 다소 많이 포함한다. 목책의 재료와 흙바닥, 붉은 저녁빛은 장소 참고와 부합하며 반사면은 없다.",
        "entities": "검은 머리를 뒤로 묶은 청년 남성 한 명만 있으며, 참고 안규홍의 얼굴과 체격에 대체로 부합한다. 낡고 오염된 미색 저고리·바지와 천 허리띠도 일치한다. 다만 초록색 고리 형태의 옥가락지가 화면 오른쪽 손의 손가락에 드러나 있어, 옷 속에 보관한다는 상태 지시와 다르다. 추가 인물이나 읽을 수 있는 글자는 없다.",
        "hard_violations": [],
        "physics": "엉덩이 아래 목제 좌면과 그 아래 지지 다리가 보여 앉은 몸의 하중을 받는다. 다리는 좌면에서 자연스럽게 내려가며 발은 화면 밖이다. 양손은 각각 허벅지와 무릎에 닿아 있고 옥가락지는 손가락에 끼워져 지지된다. 앞으로 약간 기울인 상체도 앉아서 말하는 동작으로 가능하며, 지지 없이 떠 있는 신체나 물체는 없다."
       }
      ]
     },
     "ok": true
    }
   ],
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "slot_winner_match": true,
   "slot_winner": {
    "gemini-pro": "A",
    "gpt-high": "A"
   },
   "route": "cross_slot_agree"
  },
  "dual": {
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "normalized": {
    "A": 2.0,
    "B": 0.472
   },
   "adjusted": {
    "A": 2.0,
    "B": 0.222
   },
   "violations": {
    "B": [
     "[gemini-pro] invented people (프롬프트의 'never anyone else' 지침을 어기고 전경에 2명의 인물을 임의로 추가함)",
     "[gpt-high] 화면 양쪽 전경에 명단에서 허용하지 않은 청자 두 명의 머리와 상체를 추가했다."
    ]
   },
   "per_model_winner": {
    "gemini-pro": "A",
    "gpt-high": "A"
   },
   "agreed": true
  },
  "totals": {
   "A": 2000,
   "B": 222
  },
  "selected": "A",
  "ranking": [
   "A",
   "B"
  ],
  "verdicts": [
   {
    "label": "A",
    "score": 2000,
    "verdict_ko": "명시된 인물(안규홍)만을 미디엄 샷으로 담아냈으며, 요구된 붉은 저녁 노을과 목책 배경의 디테일을 완벽하게 준수하여 프롬프트의 의도를 정확히 구현했습니다."
   },
   {
    "label": "B",
    "score": 222,
    "verdict_ko": "프롬프트가 안규홍 외 다른 인물의 등장을 명시적으로 금지했음에도 화면 전경에 두 명의 인물을 임의로 추가하는 치명적인 위반을 범했습니다.  ★위반: [gemini-pro] invented people (프롬프트의 'never anyone else' 지침을 어기고 전경에 2명의 인물을 임의로 추가함) / [gpt-high] 화면 양쪽 전경에 명단에서 허용하지 않은 청자 두 명의 머리와 상체를 추가했다."
   }
  ],
  "refs": [
   {
    "label": "PREVIOUS SHOT STILL — a visually related earlier shot of this same place: the location's look, materials, fixed features and lighting mood are LOCKED to this photo; never copy its camera framing. The people visible in that photo are NOT in this shot: never carry their faces, bodies or clothing onto anyone here. Each person in THIS shot is defined solely by the PEOPLE list, their own CHARACTER REFERENCE images and their wardrobe notes. If this photo shows a character who cannot move (dead or unconscious), that character's exact pose, position and orientation are ALSO LOCKED — treat their body as a fixed prop of the set that only the camera moves around.",
    "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/db00517b-165b-4005-a287-f28936b59b95/images/272bd7af-9294-4490-98c6-88a40e9deacc/scene/recipe/S38sh3_sel.png",
    "asset_id": "f44d1ca8-ad89-451e-9ef0-e9b6acaf69cf",
    "role": "prev_still"
   },
   {
    "label": "CHARACTER REFERENCE — 안규홍: the exact person appearing in this shot; match face, hair and build exactly.",
    "path": "<bytes:1028384>",
    "asset_id": "3b5894db-ce55-4f34-b966-5e725c3aa952",
    "role": "character_ref"
   }
  ],
  "critique_skipped": true,
  "shot_run_produced": false,
  "shot_run_uid": "06aa1428-63bc-703b-91c3-7d14af903a60",
  "ref_mode": "prev+엔티티",
  "share_plan": {
   "ref_plan": "prev",
   "prev_anchor_tag": "S38sh3"
  }
 },
 "S38sh13::cine": {
  "applied": true,
  "attempted_at": "2026-09-09T05:10:22.931118+00:00",
  "fingerprint": "5d4aa80fb986390b3d8e73b99bf54d4e8a78ed44d2c891876bdddf99b6b30469",
  "fingerprint_version": 2,
  "provider": "mai",
  "endpoint": "openrouter/chat-completions",
  "model": "microsoft/mai-image-2.6",
  "pack": "24.202608252115",
  "source_file": "S38sh13_sel.png",
  "source_sha256": "be5a42ee6f5ea061bef4f7123c64c1b1acadc3ad596c87f3c016ac824c5df8f1",
  "file": "S38sh13_cine.png",
  "staged_sha256": "46f58ccffc3e5cb6f4ea9f5b5f5fbe043b8ed069f77ea2c61bc5fa4628a5900a",
  "latency_ms": 25123
 },
 "S38sh17::signage": {
  "fp": "ef3ac0215ea8fc72",
  "inscriptions": [],
  "cues": [],
  "dropped": []
 },
 "S38sh17": {
  "input_fingerprint": "a3b369622480b935",
  "prompt": "Create ONE FINAL photorealistic live-action film still of the moment below — early 20th-century Korea, Jeolla region; all people are Korean unless explicitly stated otherwise. TIME OF DAY (lock): sunset, low reddish sunlight.\n\nSHOT TEXT (authoritative, Korean): 바쁘게 움직이는 의병들 사이에서 홀로 우두커니 멈춰 선 복만의 고립된 전신.\n\nLOCATION (lock): A spot slightly apart from the planning circle in the mountain camp's outdoor clearing, among men preparing to depart. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: wide shot\n- FRAME LAYOUT: Unoccupied opening revealing 복만's entire figure in the middle-center of the frame, midground.\n- KEY BACKGROUND ELEMENTS: Former seated circle (The group has risen and dispersed to prepare, leaving an opening toward 복만); used as The unoccupied opening separates his still figure from the nearer activity; 흙바닥 (Visible between the dispersed figures); used as Keeps 복만's feet and the separation around him legible; 총과 짚신 (Being checked and adjusted by different members of the group) — Held or worn close to the working bodies, without enlarged foreground presentation; used as Makes the surrounding preparation readable through small, asynchronous actions.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Maintain the same restrained red sunset through the palisade, allowing the unchanged light to contrast with 복만's growing isolation.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The clothing, hair and overall look of 복만 — who appear both in that photo and in this shot — are LOCKED to that photo. Anyone else visible in that photo is NOT in this shot: never carry their face, body or clothing onto anyone here. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nPREVIOUS STILL USAGE (follow exactly — what to take from the attached still and what to exclude; it governs place and objects only, never who is in this shot or how the camera sees them): Take the clearing's dirt surface, visible palisade materials, and sunset colors. Exclude the arrival's transient dust and do not retain a static arrangement of equipment now being prepared.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): Sunset light stretches through the palisade as weapons and footwear are prepared. 복만: He alone remains standing still with sweat on his forehead and his purchased supplies retained.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 복만 (한국인 남성, 성인 나이대, 거칠게 자란 수염, 자연스러운 검은 머리) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.",
  "roll_prompts": {
   "A": "Create ONE FINAL photorealistic live-action film still of the moment below — early 20th-century Korea, Jeolla region; all people are Korean unless explicitly stated otherwise. TIME OF DAY (lock): sunset, low reddish sunlight.\n\nSHOT TEXT (authoritative, Korean): 바쁘게 움직이는 의병들 사이에서 홀로 우두커니 멈춰 선 복만의 고립된 전신.\n\nLOCATION (lock): A spot slightly apart from the planning circle in the mountain camp's outdoor clearing, among men preparing to depart. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: wide shot\n- FRAME LAYOUT: Unoccupied opening revealing 복만's entire figure in the middle-center of the frame, midground.\n- KEY BACKGROUND ELEMENTS: Former seated circle (The group has risen and dispersed to prepare, leaving an opening toward 복만); used as The unoccupied opening separates his still figure from the nearer activity; 흙바닥 (Visible between the dispersed figures); used as Keeps 복만's feet and the separation around him legible; 총과 짚신 (Being checked and adjusted by different members of the group) — Held or worn close to the working bodies, without enlarged foreground presentation; used as Makes the surrounding preparation readable through small, asynchronous actions.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Maintain the same restrained red sunset through the palisade, allowing the unchanged light to contrast with 복만's growing isolation.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The clothing, hair and overall look of 복만 — who appear both in that photo and in this shot — are LOCKED to that photo. Anyone else visible in that photo is NOT in this shot: never carry their face, body or clothing onto anyone here. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nPREVIOUS STILL USAGE (follow exactly — what to take from the attached still and what to exclude; it governs place and objects only, never who is in this shot or how the camera sees them): Take the clearing's dirt surface, visible palisade materials, and sunset colors. Exclude the arrival's transient dust and do not retain a static arrangement of equipment now being prepared.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): Sunset light stretches through the palisade as weapons and footwear are prepared. 복만: He alone remains standing still with sweat on his forehead and his purchased supplies retained.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 복만 (한국인 남성, 성인 나이대, 거칠게 자란 수염, 자연스러운 검은 머리) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.",
   "B": "Create ONE FINAL photorealistic live-action film still of the moment below — early 20th-century Korea, Jeolla region; all people are Korean unless explicitly stated otherwise. TIME OF DAY (lock): sunset, low reddish sunlight.\n\nSHOT TEXT (authoritative, Korean): 바쁘게 움직이는 의병들 사이에서 홀로 우두커니 멈춰 선 복만의 고립된 전신.\n\nLOCATION (lock): A spot slightly apart from the planning circle in the mountain camp's outdoor clearing, among men preparing to depart. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: wide shot\n- FRAME LAYOUT: Unoccupied opening revealing 복만's entire figure in the middle-center of the frame, midground.\n- KEY BACKGROUND ELEMENTS: Former seated circle (The group has risen and dispersed to prepare, leaving an opening toward 복만); used as The unoccupied opening separates his still figure from the nearer activity; 흙바닥 (Visible between the dispersed figures); used as Keeps 복만's feet and the separation around him legible; 총과 짚신 (Being checked and adjusted by different members of the group) — Held or worn close to the working bodies, without enlarged foreground presentation; used as Makes the surrounding preparation readable through small, asynchronous actions.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Maintain the same restrained red sunset through the palisade, allowing the unchanged light to contrast with 복만's growing isolation.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The clothing, hair and overall look of 복만 — who appear both in that photo and in this shot — are LOCKED to that photo. Anyone else visible in that photo is NOT in this shot: never carry their face, body or clothing onto anyone here. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nPREVIOUS STILL USAGE (follow exactly — what to take from the attached still and what to exclude; it governs place and objects only, never who is in this shot or how the camera sees them): Take the clearing's dirt surface, visible palisade materials, and sunset colors. Exclude the arrival's transient dust and do not retain a static arrangement of equipment now being prepared.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): Sunset light stretches through the palisade as weapons and footwear are prepared. 복만: He alone remains standing still with sweat on his forehead and his purchased supplies retained.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 복만 (한국인 남성, 성인 나이대, 거칠게 자란 수염, 자연스러운 검은 머리) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.\n\nCOMPOSITION VARIATION (this candidate only): keep the place, the people and the moment exactly as specified, and keep from the CAMERA & FRAME contract its subject scale, screen placement and key background elements exactly. For this candidate only, its camera angle and height are deliberately left open: choose a camera position distinctly different from the obvious one, as a film director picking a second setup on the same blocking. Stay true to any explicit framing the SHOT TEXT itself dictates."
  },
  "shot_run_spend_attempt_count": 1,
  "readings": [
   {
    "label": "A",
    "direction": "주변의 의병들은 각자 정비 중인 총과 짚신 등 자신의 장비로 시선을 향하고 있으며, 중앙의 복만은 홀로 고립된 채 정면 카메라 방향을 응시하고 있습니다.",
    "built_space": "이전 샷의 레퍼런스를 충실히 반영하여 좌측의 아궁이와 가마솥, 후경의 빽빽한 나무 목책, 양옆의 초가집 형태 구조물들이 올바른 위치에 자리 잡고 있습니다.",
    "entities": "복만의 얼굴, 체형, 복장은 캐릭터 레퍼런스와 잘 일치하며 화면 중앙에 정확한 크기로 배치되었습니다. 다만 이전 샷에서 쥐고 있던 흰색 보따리(물품)가 누락되었습니다. 주변 인물들은 조선시대 의병의 모습에 부합합니다.",
    "hard_violations": [],
    "physics": "모든 인물이 흙바닥 위에 안정적으로 서 있거나 웅크리고 있으며, 각자 들고 있는 총기와 짐 등은 손에 의해 물리적으로 올바르게 지지되고 있습니다."
   },
   {
    "label": "B",
    "direction": "주변 인물들은 각자의 총기와 장비로 시선을 향하고 있고, 중앙의 복만은 정면을 무표정하게 응시하고 있습니다.",
    "built_space": "이전 샷에서는 일렬로 막혀 있던 후경의 나무 목책 한가운데에 거대한 문이나 뚫린 공간이 임의로 창조되어 구조적으로 심각하게 달라졌습니다.",
    "entities": "중앙에 복만이 서 있으나, 프롬프트의 지시와 달리 전경에 사람들을 매우 크게 배치하여 화면을 가리고 있습니다. 복만 역시 이전 샷의 짐을 들고 있지 않습니다.",
    "hard_violations": [
     "[gemini-pro] 이전 샷 레퍼런스에는 존재하지 않는 목책의 거대한 개방구(문)가 임의로 생성됨 (Invented object/structure)",
     "[gemini-pro] 우측에서 총을 쥐고 있는 인물의 손이 총기 부품과 살점으로 기괴하게 융합되어 물리적으로 불가능한 해부학적 구조를 보임 (Physically impossible anatomy)"
    ],
    "physics": "우측 인물의 손과 총기가 하나로 뭉개져 있어 정상적인 파지 상태를 유지하지 못하고 있습니다. 다른 인물들은 땅에 서 있습니다."
   }
  ],
  "cross_model_order": {
   "policy": "cross_model_order_v2_slot0_fwd_slot1_rev_parallel2",
   "slots": [
    {
     "model": "gemini-pro",
     "order": "forward",
     "display_to_canonical": {
      "A": "A",
      "B": "B"
     },
     "raw": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 8,
        "verdict_ko": "프롬프트가 요구한 와이드 샷 프레이밍과 전경의 빈 공간을 완벽하게 구현하여 복만의 고립감을 잘 살렸고 배경 디테일도 훌륭히 유지했으나, 이전 샷에서 들고 있던 구입한 물품(보따리)이 누락된 점이 아쉽습니다."
       },
       {
        "label": "B",
        "score": 3,
        "verdict_ko": "프롬프트에서 명시적으로 금지한 '확대된 전경 피사체'가 화면을 가득 채워 구도 지시를 어겼으며, 레퍼런스에 없는 거대한 목책 문을 임의로 추가하여 공간의 연속성을 훼손했습니다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "주변의 의병들은 각자 정비 중인 총과 짚신 등 자신의 장비로 시선을 향하고 있으며, 중앙의 복만은 홀로 고립된 채 정면 카메라 방향을 응시하고 있습니다.",
        "built_space": "이전 샷의 레퍼런스를 충실히 반영하여 좌측의 아궁이와 가마솥, 후경의 빽빽한 나무 목책, 양옆의 초가집 형태 구조물들이 올바른 위치에 자리 잡고 있습니다.",
        "entities": "복만의 얼굴, 체형, 복장은 캐릭터 레퍼런스와 잘 일치하며 화면 중앙에 정확한 크기로 배치되었습니다. 다만 이전 샷에서 쥐고 있던 흰색 보따리(물품)가 누락되었습니다. 주변 인물들은 조선시대 의병의 모습에 부합합니다.",
        "hard_violations": [],
        "physics": "모든 인물이 흙바닥 위에 안정적으로 서 있거나 웅크리고 있으며, 각자 들고 있는 총기와 짐 등은 손에 의해 물리적으로 올바르게 지지되고 있습니다."
       },
       {
        "label": "B",
        "direction": "주변 인물들은 각자의 총기와 장비로 시선을 향하고 있고, 중앙의 복만은 정면을 무표정하게 응시하고 있습니다.",
        "built_space": "이전 샷에서는 일렬로 막혀 있던 후경의 나무 목책 한가운데에 거대한 문이나 뚫린 공간이 임의로 창조되어 구조적으로 심각하게 달라졌습니다.",
        "entities": "중앙에 복만이 서 있으나, 프롬프트의 지시와 달리 전경에 사람들을 매우 크게 배치하여 화면을 가리고 있습니다. 복만 역시 이전 샷의 짐을 들고 있지 않습니다.",
        "hard_violations": [
         "이전 샷 레퍼런스에는 존재하지 않는 목책의 거대한 개방구(문)가 임의로 생성됨 (Invented object/structure)",
         "우측에서 총을 쥐고 있는 인물의 손이 총기 부품과 살점으로 기괴하게 융합되어 물리적으로 불가능한 해부학적 구조를 보임 (Physically impossible anatomy)"
        ],
        "physics": "우측 인물의 손과 총기가 하나로 뭉개져 있어 정상적인 파지 상태를 유지하지 못하고 있습니다. 다른 인물들은 땅에 서 있습니다."
       }
      ],
      "all_candidates_fail": false
     },
     "normalized": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 8,
        "verdict_ko": "프롬프트가 요구한 와이드 샷 프레이밍과 전경의 빈 공간을 완벽하게 구현하여 복만의 고립감을 잘 살렸고 배경 디테일도 훌륭히 유지했으나, 이전 샷에서 들고 있던 구입한 물품(보따리)이 누락된 점이 아쉽습니다."
       },
       {
        "label": "B",
        "score": 3,
        "verdict_ko": "프롬프트에서 명시적으로 금지한 '확대된 전경 피사체'가 화면을 가득 채워 구도 지시를 어겼으며, 레퍼런스에 없는 거대한 목책 문을 임의로 추가하여 공간의 연속성을 훼손했습니다."
       }
      ],
      "all_candidates_fail": false,
      "readings": [
       {
        "label": "A",
        "direction": "주변의 의병들은 각자 정비 중인 총과 짚신 등 자신의 장비로 시선을 향하고 있으며, 중앙의 복만은 홀로 고립된 채 정면 카메라 방향을 응시하고 있습니다.",
        "built_space": "이전 샷의 레퍼런스를 충실히 반영하여 좌측의 아궁이와 가마솥, 후경의 빽빽한 나무 목책, 양옆의 초가집 형태 구조물들이 올바른 위치에 자리 잡고 있습니다.",
        "entities": "복만의 얼굴, 체형, 복장은 캐릭터 레퍼런스와 잘 일치하며 화면 중앙에 정확한 크기로 배치되었습니다. 다만 이전 샷에서 쥐고 있던 흰색 보따리(물품)가 누락되었습니다. 주변 인물들은 조선시대 의병의 모습에 부합합니다.",
        "hard_violations": [],
        "physics": "모든 인물이 흙바닥 위에 안정적으로 서 있거나 웅크리고 있으며, 각자 들고 있는 총기와 짐 등은 손에 의해 물리적으로 올바르게 지지되고 있습니다."
       },
       {
        "label": "B",
        "direction": "주변 인물들은 각자의 총기와 장비로 시선을 향하고 있고, 중앙의 복만은 정면을 무표정하게 응시하고 있습니다.",
        "built_space": "이전 샷에서는 일렬로 막혀 있던 후경의 나무 목책 한가운데에 거대한 문이나 뚫린 공간이 임의로 창조되어 구조적으로 심각하게 달라졌습니다.",
        "entities": "중앙에 복만이 서 있으나, 프롬프트의 지시와 달리 전경에 사람들을 매우 크게 배치하여 화면을 가리고 있습니다. 복만 역시 이전 샷의 짐을 들고 있지 않습니다.",
        "hard_violations": [
         "이전 샷 레퍼런스에는 존재하지 않는 목책의 거대한 개방구(문)가 임의로 생성됨 (Invented object/structure)",
         "우측에서 총을 쥐고 있는 인물의 손이 총기 부품과 살점으로 기괴하게 융합되어 물리적으로 불가능한 해부학적 구조를 보임 (Physically impossible anatomy)"
        ],
        "physics": "우측 인물의 손과 총기가 하나로 뭉개져 있어 정상적인 파지 상태를 유지하지 못하고 있습니다. 다른 인물들은 땅에 서 있습니다."
       }
      ]
     },
     "ok": true
    },
    {
     "model": "gpt-high",
     "order": "reverse",
     "display_to_canonical": {
      "A": "B",
      "B": "A"
     },
     "raw": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 5,
        "verdict_ko": "중앙 중경의 고립된 전신과 주변 활동은 구현했지만, 짚신을 손보는 행동이 불분명하고 목책 중앙 공간이 달라졌으며 복만의 구입 물품 꾸러미가 빠졌다."
       },
       {
        "label": "B",
        "score": 7,
        "verdict_ko": "중앙의 빈 흙바닥에 선 복만과 총·짚신을 각기 준비하는 의병들, 기존 야영지 배치를 더 충실히 구현했지만 복만의 꾸러미와 동작 도중 멈춘 자세는 놓쳤다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "복만은 거의 정면으로 서서 카메라 쪽을 바라본다. 왼쪽 의병들은 등을 보이거나 아래를 보고, 오른쪽 의병들은 손에 든 총과 작업 부위를 내려다본다. 오른쪽 중앙 총구는 왼쪽 아래의 빈 흙바닥을 향하며, 다른 총들은 몸 가까이 비스듬히 놓여 있다. 복만이나 다른 사람을 겨누는 행동은 보이지 않는다.",
        "built_space": "양옆에 초가 시설이 하나씩, 오른쪽 뒤에 지붕 있는 망루 하나, 뒤쪽에 통나무 목책이 보인다. 복만은 중앙 중경의 넓은 흙바닥에 전신이 드러나고, 가까운 의병들이 양쪽 가장자리를 차지한다. 다만 참조에서 목책이 이어져 보이던 중앙에 큰 통로가 생겨 공간의 연속성이 약해졌다. 반사면이나 중복된 망루는 보이지 않는다.",
        "entities": "복만은 검은 머리와 수염을 가진 한국인 성인 남성으로 보이며, 모자와 낡은 짙은 겉옷, 허리띠, 바지와 신발은 참조의 전반적 외형에 가깝다. 그러나 참조에서 양옆으로 매달렸던 구입 물품 주머니는 보이지 않는다. 주변에는 시대 의복을 입은 성인 남성 의병들과 장총이 있지만, 서로 다른 의병이 짚신을 점검하거나 조이는 행동은 확인하기 어렵다. 이마의 땀은 이 크기에서 판별하기 어렵고, 읽을 수 있는 글자는 없다.",
        "hard_violations": [],
        "physics": "복만의 두 발은 흙바닥에 닿아 있고, 의병들은 보이는 발이나 화면 밖으로 이어지는 다리로 체중을 지탱한다. 총은 손으로 잡거나 어깨끈에 걸고 있으며, 공중에 지지 없이 뜬 몸이나 물건은 보이지 않는다. 다만 복만은 양팔을 내리고 양발을 벌린 거의 대칭적인 자세여서, 한쪽에 체중을 실은 채 하던 동작이 멎었다기보다 정면으로 서도록 배치된 모습이다."
       },
       {
        "label": "B",
        "direction": "복만은 중앙에서 정면 가까운 방향을 응시한다. 왼쪽의 세 의병은 화면 왼쪽으로 걸어가고, 오른쪽의 의병들은 각자 총의 작동 부위나 발의 짚신을 내려다본다. 오른쪽 끝 장총은 총구가 위로 향하고, 그 옆 장총도 오른쪽 위를 향한다. 총구가 복만을 겨누지 않으며, 손과 시선이 장비 점검 대상에 모인다.",
        "built_space": "좌우 초가 시설 두 곳, 오른쪽 뒤 망루 하나, 뒤를 막는 통나무 목책, 왼쪽 앞의 솥이 놓인 흙·돌 아궁이 하나가 참조와 유사한 위치 관계로 유지된다. 복만은 중앙 중경의 빈 흙바닥에 서 있고 발 주변까지 가려지지 않는다. 왼쪽 이동 인원과 오른쪽 준비 인원 사이의 빈 공간이 고립을 만든다. 오른쪽 앞 인물과 총이 비교적 크게 보이지만 실제 거리 차이로 설명되며, 불가능한 반사나 중복 고정 시설은 없다.",
        "entities": "복만의 성인 남성 외형, 검은 머리와 수염, 모자, 짙고 해진 겉옷과 바지는 참조에 대체로 맞는다. 다만 전신이 보이는데도 참조의 양옆 구입 물품 주머니는 없고, 오른쪽 바닥의 짐은 복만이 지닌 물건으로 읽히지 않는다. 주변 한국인 성인 남성 의병들은 장총을 점검하고 짚신 끈을 조이며 떠날 준비를 한다. 흙바닥과 낮은 붉은 석양도 보인다. 이마의 땀은 확실하게 식별되지 않으며 읽을 수 있는 글자는 없다.",
        "hard_violations": [],
        "physics": "복만은 두 발로 땅을 딛고 있고, 왼쪽의 보행 인물들은 지지하는 발과 앞으로 옮기는 발이 구분된다. 오른쪽 앞에서 짚신을 조이는 사람은 굽힌 다리와 바닥에 닿은 발로 몸을 지탱하며 양손이 신발에 닿는다. 세운 총은 손과 바닥의 개머리판 접촉으로, 비스듬한 총은 양손으로 지지된다. 바닥 꾸러미도 흙 위에 놓여 있다. 물리적으로 불가능한 부유는 없지만 복만의 정지는 거의 대칭적인 정자세라 요구된 중단 동작의 체중 편향이 부족하다."
       }
      ],
      "all_candidates_fail": true
     },
     "normalized": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 5,
        "verdict_ko": "중앙 중경의 고립된 전신과 주변 활동은 구현했지만, 짚신을 손보는 행동이 불분명하고 목책 중앙 공간이 달라졌으며 복만의 구입 물품 꾸러미가 빠졌다."
       },
       {
        "label": "A",
        "score": 7,
        "verdict_ko": "중앙의 빈 흙바닥에 선 복만과 총·짚신을 각기 준비하는 의병들, 기존 야영지 배치를 더 충실히 구현했지만 복만의 꾸러미와 동작 도중 멈춘 자세는 놓쳤다."
       }
      ],
      "all_candidates_fail": true,
      "readings": [
       {
        "label": "B",
        "direction": "복만은 거의 정면으로 서서 카메라 쪽을 바라본다. 왼쪽 의병들은 등을 보이거나 아래를 보고, 오른쪽 의병들은 손에 든 총과 작업 부위를 내려다본다. 오른쪽 중앙 총구는 왼쪽 아래의 빈 흙바닥을 향하며, 다른 총들은 몸 가까이 비스듬히 놓여 있다. 복만이나 다른 사람을 겨누는 행동은 보이지 않는다.",
        "built_space": "양옆에 초가 시설이 하나씩, 오른쪽 뒤에 지붕 있는 망루 하나, 뒤쪽에 통나무 목책이 보인다. 복만은 중앙 중경의 넓은 흙바닥에 전신이 드러나고, 가까운 의병들이 양쪽 가장자리를 차지한다. 다만 참조에서 목책이 이어져 보이던 중앙에 큰 통로가 생겨 공간의 연속성이 약해졌다. 반사면이나 중복된 망루는 보이지 않는다.",
        "entities": "복만은 검은 머리와 수염을 가진 한국인 성인 남성으로 보이며, 모자와 낡은 짙은 겉옷, 허리띠, 바지와 신발은 참조의 전반적 외형에 가깝다. 그러나 참조에서 양옆으로 매달렸던 구입 물품 주머니는 보이지 않는다. 주변에는 시대 의복을 입은 성인 남성 의병들과 장총이 있지만, 서로 다른 의병이 짚신을 점검하거나 조이는 행동은 확인하기 어렵다. 이마의 땀은 이 크기에서 판별하기 어렵고, 읽을 수 있는 글자는 없다.",
        "hard_violations": [],
        "physics": "복만의 두 발은 흙바닥에 닿아 있고, 의병들은 보이는 발이나 화면 밖으로 이어지는 다리로 체중을 지탱한다. 총은 손으로 잡거나 어깨끈에 걸고 있으며, 공중에 지지 없이 뜬 몸이나 물건은 보이지 않는다. 다만 복만은 양팔을 내리고 양발을 벌린 거의 대칭적인 자세여서, 한쪽에 체중을 실은 채 하던 동작이 멎었다기보다 정면으로 서도록 배치된 모습이다."
       },
       {
        "label": "A",
        "direction": "복만은 중앙에서 정면 가까운 방향을 응시한다. 왼쪽의 세 의병은 화면 왼쪽으로 걸어가고, 오른쪽의 의병들은 각자 총의 작동 부위나 발의 짚신을 내려다본다. 오른쪽 끝 장총은 총구가 위로 향하고, 그 옆 장총도 오른쪽 위를 향한다. 총구가 복만을 겨누지 않으며, 손과 시선이 장비 점검 대상에 모인다.",
        "built_space": "좌우 초가 시설 두 곳, 오른쪽 뒤 망루 하나, 뒤를 막는 통나무 목책, 왼쪽 앞의 솥이 놓인 흙·돌 아궁이 하나가 참조와 유사한 위치 관계로 유지된다. 복만은 중앙 중경의 빈 흙바닥에 서 있고 발 주변까지 가려지지 않는다. 왼쪽 이동 인원과 오른쪽 준비 인원 사이의 빈 공간이 고립을 만든다. 오른쪽 앞 인물과 총이 비교적 크게 보이지만 실제 거리 차이로 설명되며, 불가능한 반사나 중복 고정 시설은 없다.",
        "entities": "복만의 성인 남성 외형, 검은 머리와 수염, 모자, 짙고 해진 겉옷과 바지는 참조에 대체로 맞는다. 다만 전신이 보이는데도 참조의 양옆 구입 물품 주머니는 없고, 오른쪽 바닥의 짐은 복만이 지닌 물건으로 읽히지 않는다. 주변 한국인 성인 남성 의병들은 장총을 점검하고 짚신 끈을 조이며 떠날 준비를 한다. 흙바닥과 낮은 붉은 석양도 보인다. 이마의 땀은 확실하게 식별되지 않으며 읽을 수 있는 글자는 없다.",
        "hard_violations": [],
        "physics": "복만은 두 발로 땅을 딛고 있고, 왼쪽의 보행 인물들은 지지하는 발과 앞으로 옮기는 발이 구분된다. 오른쪽 앞에서 짚신을 조이는 사람은 굽힌 다리와 바닥에 닿은 발로 몸을 지탱하며 양손이 신발에 닿는다. 세운 총은 손과 바닥의 개머리판 접촉으로, 비스듬한 총은 양손으로 지지된다. 바닥 꾸러미도 흙 위에 놓여 있다. 물리적으로 불가능한 부유는 없지만 복만의 정지는 거의 대칭적인 정자세라 요구된 중단 동작의 체중 편향이 부족하다."
       }
      ]
     },
     "ok": true
    }
   ],
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "slot_winner_match": true,
   "slot_winner": {
    "gemini-pro": "A",
    "gpt-high": "A"
   },
   "route": "cross_slot_agree"
  },
  "dual": {
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "normalized": {
    "A": 2.0,
    "B": 1.089
   },
   "adjusted": {
    "A": 2.0,
    "B": 0.839
   },
   "violations": {
    "B": [
     "[gemini-pro] 이전 샷 레퍼런스에는 존재하지 않는 목책의 거대한 개방구(문)가 임의로 생성됨 (Invented object/structure)",
     "[gemini-pro] 우측에서 총을 쥐고 있는 인물의 손이 총기 부품과 살점으로 기괴하게 융합되어 물리적으로 불가능한 해부학적 구조를 보임 (Physically impossible anatomy)"
    ]
   },
   "per_model_winner": {
    "gemini-pro": "A",
    "gpt-high": "A"
   },
   "agreed": true
  },
  "totals": {
   "A": 2000,
   "B": 839
  },
  "selected": "A",
  "ranking": [
   "A",
   "B"
  ],
  "verdicts": [
   {
    "label": "A",
    "score": 2000,
    "verdict_ko": "프롬프트가 요구한 와이드 샷 프레이밍과 전경의 빈 공간을 완벽하게 구현하여 복만의 고립감을 잘 살렸고 배경 디테일도 훌륭히 유지했으나, 이전 샷에서 들고 있던 구입한 물품(보따리)이 누락된 점이 아쉽습니다."
   },
   {
    "label": "B",
    "score": 839,
    "verdict_ko": "프롬프트에서 명시적으로 금지한 '확대된 전경 피사체'가 화면을 가득 채워 구도 지시를 어겼으며, 레퍼런스에 없는 거대한 목책 문을 임의로 추가하여 공간의 연속성을 훼손했습니다.  ★위반: [gemini-pro] 이전 샷 레퍼런스에는 존재하지 않는 목책의 거대한 개방구(문)가 임의로 생성됨 (Invented object/structure) / [gemini-pro] 우측에서 총을 쥐고 있는 인물의 손이 총기 부품과 살점으로 기괴하게 융합되어 물리적으로 불가능한 해부학적 구조를 보임 (Physically impossible anatomy)"
   }
  ],
  "refs": [
   {
    "label": "PREVIOUS SHOT STILL — a visually related earlier shot of this same place: the location's look, materials, fixed features and lighting mood are LOCKED to this photo; never copy its camera framing. The clothing, hair and overall look of 복만 — who appear both in that photo and in this shot — are LOCKED to that photo. Anyone else visible in that photo is NOT in this shot: never carry their face, body or clothing onto anyone here. If this photo shows a character who cannot move (dead or unconscious), that character's exact pose, position and orientation are ALSO LOCKED — treat their body as a fixed prop of the set that only the camera moves around.",
    "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/db00517b-165b-4005-a287-f28936b59b95/images/272bd7af-9294-4490-98c6-88a40e9deacc/scene/recipe/S38sh3_sel.png",
    "asset_id": "f44d1ca8-ad89-451e-9ef0-e9b6acaf69cf",
    "role": "prev_still"
   },
   {
    "label": "CHARACTER REFERENCE — 복만: the exact person appearing in this shot; match face, hair and build exactly.",
    "path": "<bytes:1027479>",
    "asset_id": "b7e69570-d15d-4241-b5ea-4cb8d4ece1cf",
    "role": "character_ref"
   }
  ],
  "critique_skipped": true,
  "shot_run_produced": false,
  "shot_run_uid": "06aa1428-654d-70aa-8130-832caa7affce",
  "ref_mode": "prev+엔티티",
  "share_plan": {
   "ref_plan": "prev",
   "prev_anchor_tag": "S38sh3"
  }
 },
 "S38sh17::cine": {
  "applied": true,
  "attempted_at": "2026-09-09T05:12:02.876413+00:00",
  "fingerprint": "f85fbe8dbf936ce24f99a13ccde85628326aab5da8c47901cf2d787662dd9c99",
  "fingerprint_version": 2,
  "provider": "mai",
  "endpoint": "openrouter/chat-completions",
  "model": "microsoft/mai-image-2.6",
  "pack": "24.202608252115",
  "source_file": "S38sh17_sel.png",
  "source_sha256": "c16e5fb8cef1385d6f02e14b5202a7d43de35009544a32525f817a9abfe1a329",
  "file": "S38sh17_cine.png",
  "staged_sha256": "0824ff21144417772fc518bf50cb04c7669003d7266b8715ea918f59c7e7241c",
  "latency_ms": 28867
 },
 "S39sh14::signage": {
  "fp": "e48f6f30d42cdd37",
  "inscriptions": [],
  "cues": [],
  "dropped": []
 },
 "era_assess::124832509e650275": {
  "subjects": [],
  "subject_text": "새재 산길·수레 습격 지점·길가 숲\n양쪽 능선 사이로 이어지는 산중 흙길. 길가에 나무가 빽빽하고, 굽은 길과 숲 가장자리 사이로 좁은 빈틈이 난다.",
  "identity": "canonical",
  "scope_id": "L29",
  "scope_role": "location_exterior",
  "scope_sha": "c62d508f63228f19"
 },
 "groupbg::ambush_upper_trees": {
  "input_fingerprint": "361bb153b366f9e5",
  "meta": {
   "model": "gpt-image-2",
   "size": "1536x864",
   "pack": "11.202607220237",
   "contract": "bgfirst_full_v3",
   "group_sig": {
    "key": "ambush_upper_trees",
    "tags": [
     "S39sh14"
    ]
   },
   "context_sig": "5b1e8771561d8a6a"
  },
  "prompt": "Create ONE empty live-action location background photograph — NO PEOPLE, no figures, no body parts, no silhouettes, no shadows or reflections of people anywhere.\n\"Empty\" means no people only: KEEP the location's inherent occupants and stock that define the place — animals in an animal shelter, pen or farm, goods and produce in a market, moored boats in a harbour — unless the shot text explicitly removes them.\nThe FIRST attached image is a thin-line storyboard sketch of a shot that happens at this location: use it ONLY as spatial evidence — what this place contains, how its ground, structures and landmarks are arranged and proportioned. Ignore the sketched people and arrows entirely, and do NOT copy its line style: render a fully photographic, physically plausible real place that fits THE LOCATION text below.\nHUMAN-SCALE CALIBRATION: derive every structure's true size from human-scale elements — a door ≈ 2m, a window ≈ 1–1.5m wide, one storey ≈ 2.5–3m; never inflate a small structure or shrink a large one.\n\nTHE LOCATION — early 20th-century Korea, Jeolla region; all people are Korean unless explicitly stated otherwise: The wooded ridges on both sides of a mountain-pass cart track, overlooking the exposed ambush site.\n\nLOCATION DETAIL (descriptive evidence about this place from the production's location records — evidence, not a staging order): use it to understand what kind of place this is and what it permanently contains. Carry over ONLY its enduring physical features — layout, structures, ground and materials, fixed equipment and installations. Do NOT treat any lighting, weather or time-of-day wording, momentary object states, depicted events or subjective impressions in it as fixed facts of the place; the TIME OF DAY lock below is the sole authority for time and lighting.\n새재 산길·수레 습격 지점·길가 숲: 수레 한 대가 지나갈 수 있을 정도의 너비를 가진 산속 흙길. 길 양옆으로 울창한 잡목 숲이 맞닿아 있어 기습에 용이한 지형이다. (특징: 수레가 이동할 수 있는 너비의 다져진 흙길; 수레 타이어 자국이 남은 흙바닥; 길과 맞닿은 빽빽한 잡목림과 수풀)\n\nSCENE EVIDENCE (verbatim quotes from the screenplay about this place — treat them as evidence of what the location physically contains and looks like; stage the PLACE those moments happen in, but do NOT depict the momentary actions, people or staged props themselves):\n- 양쪽 능선에서 총성이 쏟아진다.\n- 숲 사이로 모습을 드러내는 일본군들.\n\nTIME OF DAY (lock): day.\n\nRender ONE photorealistic empty location photograph, 16:9, neutral enough that every shot of this place can be staged from it later. No readable writing anywhere: surfaces that would carry writing may be present, but stage any wording out of legibility — an oblique angle, distance, shallow focus. No captions, watermarks or overlay text.",
  "effective_prompt": "Create ONE empty live-action location background photograph — NO PEOPLE, no figures, no body parts, no silhouettes, no shadows or reflections of people anywhere.\n\"Empty\" means no people only: KEEP the location's inherent occupants and stock that define the place — animals in an animal shelter, pen or farm, goods and produce in a market, moored boats in a harbour — unless the shot text explicitly removes them.\nThe FIRST attached image is a thin-line storyboard sketch of a shot that happens at this location: use it ONLY as spatial evidence — what this place contains, how its ground, structures and landmarks are arranged and proportioned. Ignore the sketched people and arrows entirely, and do NOT copy its line style: render a fully photographic, physically plausible real place that fits THE LOCATION text below.\nHUMAN-SCALE CALIBRATION: derive every structure's true size from human-scale elements — a door ≈ 2m, a window ≈ 1–1.5m wide, one storey ≈ 2.5–3m; never inflate a small structure or shrink a large one.\n\nTHE LOCATION — early 20th-century Korea, Jeolla region; all people are Korean unless explicitly stated otherwise: The wooded ridges on both sides of a mountain-pass cart track, overlooking the exposed ambush site.\n\nLOCATION DETAIL (descriptive evidence about this place from the production's location records — evidence, not a staging order): use it to understand what kind of place this is and what it permanently contains. Carry over ONLY its enduring physical features — layout, structures, ground and materials, fixed equipment and installations. Do NOT treat any lighting, weather or time-of-day wording, momentary object states, depicted events or subjective impressions in it as fixed facts of the place; the TIME OF DAY lock below is the sole authority for time and lighting.\n새재 산길·수레 습격 지점·길가 숲: 수레 한 대가 지나갈 수 있을 정도의 너비를 가진 산속 흙길. 길 양옆으로 울창한 잡목 숲이 맞닿아 있어 기습에 용이한 지형이다. (특징: 수레가 이동할 수 있는 너비의 다져진 흙길; 수레 타이어 자국이 남은 흙바닥; 길과 맞닿은 빽빽한 잡목림과 수풀)\n\nSCENE EVIDENCE (verbatim quotes from the screenplay about this place — treat them as evidence of what the location physically contains and looks like; stage the PLACE those moments happen in, but do NOT depict the momentary actions, people or staged props themselves):\n- 양쪽 능선에서 총성이 쏟아진다.\n- 숲 사이로 모습을 드러내는 일본군들.\n\nTIME OF DAY (lock): day.\n\nRender ONE photorealistic empty location photograph, 16:9, neutral enough that every shot of this place can be staged from it later. No readable writing anywhere: surfaces that would carry writing may be present, but stage any wording out of legibility — an oblique angle, distance, shallow focus. No captions, watermarks or overlay text.",
  "bg_path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/db00517b-165b-4005-a287-f28936b59b95/images/272bd7af-9294-4490-98c6-88a40e9deacc/scene/recipe/groupbg_ambush_upper_trees_269326.png",
  "asset_id": "df94889d-3b54-4560-8b7b-cb09c93785df",
  "input_asset_ids": [
   "464a7862-b636-4a03-82eb-b0aabb506b77"
  ],
  "origin_tag": "S39sh14",
  "place_text": "The wooded ridges on both sides of a mountain-pass cart track, overlooking the exposed ambush site.",
  "origin_inputs": {
   "place_text": "The wooded ridges on both sides of a mountain-pass cart track, overlooking the exposed ambush site.",
   "time_of_day_en": "day",
   "conti_asset_id": "464a7862-b636-4a03-82eb-b0aabb506b77"
  }
 },
 "S39sh14::bgfirst_bg": {
  "input_fingerprint": "16dc3cea5c68401a",
  "prompt": "Create the EMPTY BACKGROUND PLATE for one film shot — NO PEOPLE, no figures, no body parts, no sketch lines, no arrows anywhere.\n\"Empty\" means no people only: KEEP the location's inherent occupants and stock that define the place — animals in an animal shelter, pen or farm, goods in a market, moored boats in a harbour — unless the shot text explicitly removes them.\nThe FIRST attached image is a thin-line storyboard sketch: use ONLY its camera angle, horizon, perspective and the placement/size of buildings and set masses — ignore the sketched people and arrows entirely. The SECOND attached image (LOCATION PHOTOGRAPH) is the real place: take its architecture, materials, signage and fixed features, and RE-PROJECT them into the sketch's camera. If the photograph's camera differs from the sketch's, the sketch's camera wins.\nHUMAN-SCALE CALIBRATION: derive every structure's true size from human-scale elements — a door ≈ 2m, a window ≈ 1–1.5m wide, one storey ≈ 2.5–3m; never inflate a small structure or shrink a large one.\n\nSHOT TEXT this background must serve (Korean): 나무 틈새로 일제히 모습을 드러낸 채 총을 겨누고 선 수많은 일본인 군인들의 전신.\n\nLOCATION (lock): The wooded ridges on both sides of a mountain-pass cart track, overlooking the exposed ambush site.\n\nTIME OF DAY (lock): day.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: wide shot\n- KEY BACKGROUND ELEMENTS: 능선의 나무 (Japanese soldiers have appeared in the gaps between them) — Trunks cross the oblique uphill view at different depths; used as Divide the force into irregular visible intervals while preserving full-body readings through selected gaps; 산길과 능선 (The soldiers occupy the higher ground above the path) — The path edge lies low in the image and the slope recedes upward; used as Establishes the attackers' height advantage without adopting a combatant's subjective viewpoint; 일본군의 총 (Raised and aimed toward the militia below) — Seen from the side at oblique angles, with barrels directed away from the camera's axis; used as Connects the dispersed figures through a common downhill threat direction.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Use natural daytime illumination with muted saturation and grounded contrast, keeping the uphill threat legible without theatrical lighting changes.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nRender ONE photorealistic empty location photograph, 16:9, that this shot can be staged inside later. No readable writing anywhere: surfaces that would carry writing may be present, but stage any wording out of legibility — an oblique angle, distance, shallow focus. No captions, watermarks or overlay text.",
  "effective_prompt": "Create the EMPTY BACKGROUND PLATE for one film shot — NO PEOPLE, no figures, no body parts, no sketch lines, no arrows anywhere.\n\"Empty\" means no people only: KEEP the location's inherent occupants and stock that define the place — animals in an animal shelter, pen or farm, goods in a market, moored boats in a harbour — unless the shot text explicitly removes them.\nThe FIRST attached image is a thin-line storyboard sketch: use ONLY its camera angle, horizon, perspective and the placement/size of buildings and set masses — ignore the sketched people and arrows entirely. The SECOND attached image (LOCATION PHOTOGRAPH) is the real place: take its architecture, materials, signage and fixed features, and RE-PROJECT them into the sketch's camera. If the photograph's camera differs from the sketch's, the sketch's camera wins.\nHUMAN-SCALE CALIBRATION: derive every structure's true size from human-scale elements — a door ≈ 2m, a window ≈ 1–1.5m wide, one storey ≈ 2.5–3m; never inflate a small structure or shrink a large one.\n\nSHOT TEXT this background must serve (Korean): 나무 틈새로 일제히 모습을 드러낸 채 총을 겨누고 선 수많은 일본인 군인들의 전신.\n\nLOCATION (lock): The wooded ridges on both sides of a mountain-pass cart track, overlooking the exposed ambush site.\n\nTIME OF DAY (lock): day.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: wide shot\n- KEY BACKGROUND ELEMENTS: 능선의 나무 (Japanese soldiers have appeared in the gaps between them) — Trunks cross the oblique uphill view at different depths; used as Divide the force into irregular visible intervals while preserving full-body readings through selected gaps; 산길과 능선 (The soldiers occupy the higher ground above the path) — The path edge lies low in the image and the slope recedes upward; used as Establishes the attackers' height advantage without adopting a combatant's subjective viewpoint; 일본군의 총 (Raised and aimed toward the militia below) — Seen from the side at oblique angles, with barrels directed away from the camera's axis; used as Connects the dispersed figures through a common downhill threat direction.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Use natural daytime illumination with muted saturation and grounded contrast, keeping the uphill threat legible without theatrical lighting changes.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nRender ONE photorealistic empty location photograph, 16:9, that this shot can be staged inside later. No readable writing anywhere: surfaces that would carry writing may be present, but stage any wording out of legibility — an oblique angle, distance, shallow focus. No captions, watermarks or overlay text.",
  "bg_path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/db00517b-165b-4005-a287-f28936b59b95/images/272bd7af-9294-4490-98c6-88a40e9deacc/scene/recipe/S39sh14__bgfirst_bg.png",
  "asset_id": "bd7bba71-1819-42da-aace-6a405084bcaa",
  "input_asset_ids": [
   "464a7862-b636-4a03-82eb-b0aabb506b77",
   "df94889d-3b54-4560-8b7b-cb09c93785df"
  ]
 },
 "S39sh14": {
  "input_fingerprint": "3e5660a2eb864207",
  "prompt": "Create ONE FINAL photorealistic live-action film still of the moment below — early 20th-century Korea, Jeolla region; all people are Korean unless explicitly stated otherwise. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 나무 틈새로 일제히 모습을 드러낸 채 총을 겨누고 선 수많은 일본인 군인들의 전신.\n\nLOCATION (lock): The wooded ridges on both sides of a mountain-pass cart track, overlooking the exposed ambush site. The shot takes place here — the attached LOCATION PHOTOGRAPH shows the exact spot.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: wide shot\n- KEY BACKGROUND ELEMENTS: 능선의 나무 (Japanese soldiers have appeared in the gaps between them) — Trunks cross the oblique uphill view at different depths; used as Divide the force into irregular visible intervals while preserving full-body readings through selected gaps; 산길과 능선 (The soldiers occupy the higher ground above the path) — The path edge lies low in the image and the slope recedes upward; used as Establishes the attackers' height advantage without adopting a combatant's subjective viewpoint; 일본군의 총 (Raised and aimed toward the militia below) — Seen from the side at oblique angles, with barrels directed away from the camera's axis; used as Connects the dispersed figures through a common downhill threat direction.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Use natural daytime illumination with muted saturation and grounded contrast, keeping the uphill threat legible without theatrical lighting changes.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The bait supply cart is stopped on the mountain path; gunfire comes from both wooded ridges.\n\nTHE HAND THAT IS DOING THIS: the object at the centre of this shot is being held, operated, read, aimed or handed over by 수많은 일본인 군인들 right now, so 수많은 일본인 군인들's hand — and as much of the wrist and forearm as the framing reaches — is in the frame, gripping or resting on that object exactly the way its real use requires. Match that hand to 수많은 일본인 군인들: its size, build, skin, age, grooming, sleeve and anything worn on it belong to that person and to no one else.\n\nNo people appear unless the shot text itself says so.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.",
  "roll_prompts": {
   "A": "Stage the shot. The FIRST attached image (SHOT BACKGROUND) is the finished empty background of this shot — keep it EXACTLY: its camera, perspective, architecture, lighting and every fixed feature stay untouched. The SECOND attached image (LAYOUT SKETCH) tells you ONLY where the people go: each sketched person's position, screen size, pose and the gaze/motion arrows. Ignore the sketch's background lines. The CHARACTER REFERENCE photographs show the real people.\nPlace the real people into the background at exactly the sketched positions, sizes and poses, following the arrow directions. No sketch lines or arrows may remain.\n\nCreate ONE FINAL photorealistic live-action film still of the moment below — early 20th-century Korea, Jeolla region; all people are Korean unless explicitly stated otherwise. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 나무 틈새로 일제히 모습을 드러낸 채 총을 겨누고 선 수많은 일본인 군인들의 전신.\n\nLOCATION (lock): The wooded ridges on both sides of a mountain-pass cart track, overlooking the exposed ambush site. The shot takes place here — the FIRST attached image (SHOT BACKGROUND) is this exact place, already built: its ground, structures, horizon, materials and lighting are the finished truth of this location and must not be redesigned or replaced. No location photograph is attached — read the place from that image alone, and add no scenery, structure, vehicle or fixture that it does not already show. This lock governs the place only; the figures in the shot follow the staging and pose instructions.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: wide shot\n- KEY BACKGROUND ELEMENTS: 능선의 나무 (Japanese soldiers have appeared in the gaps between them) — Trunks cross the oblique uphill view at different depths; used as Divide the force into irregular visible intervals while preserving full-body readings through selected gaps; 산길과 능선 (The soldiers occupy the higher ground above the path) — The path edge lies low in the image and the slope recedes upward; used as Establishes the attackers' height advantage without adopting a combatant's subjective viewpoint; 일본군의 총 (Raised and aimed toward the militia below) — Seen from the side at oblique angles, with barrels directed away from the camera's axis; used as Connects the dispersed figures through a common downhill threat direction.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Use natural daytime illumination with muted saturation and grounded contrast, keeping the uphill threat legible without theatrical lighting changes.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The bait supply cart is stopped on the mountain path; gunfire comes from both wooded ridges.\n\nTHE HAND THAT IS DOING THIS: the object at the centre of this shot is being held, operated, read, aimed or handed over by 수많은 일본인 군인들 right now, so 수많은 일본인 군인들's hand — and as much of the wrist and forearm as the framing reaches — is in the frame, gripping or resting on that object exactly the way its real use requires. Match that hand to 수많은 일본인 군인들: its size, build, skin, age, grooming, sleeve and anything worn on it belong to that person and to no one else.\n\nNo people appear unless the shot text itself says so.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.",
   "B": "Create ONE FINAL photorealistic live-action film still of the moment below — early 20th-century Korea, Jeolla region; all people are Korean unless explicitly stated otherwise. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 나무 틈새로 일제히 모습을 드러낸 채 총을 겨누고 선 수많은 일본인 군인들의 전신.\n\nLOCATION (lock): The wooded ridges on both sides of a mountain-pass cart track, overlooking the exposed ambush site. The shot takes place here — the attached LOCATION PHOTOGRAPH shows the exact spot.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: wide shot\n- KEY BACKGROUND ELEMENTS: 능선의 나무 (Japanese soldiers have appeared in the gaps between them) — Trunks cross the oblique uphill view at different depths; used as Divide the force into irregular visible intervals while preserving full-body readings through selected gaps; 산길과 능선 (The soldiers occupy the higher ground above the path) — The path edge lies low in the image and the slope recedes upward; used as Establishes the attackers' height advantage without adopting a combatant's subjective viewpoint; 일본군의 총 (Raised and aimed toward the militia below) — Seen from the side at oblique angles, with barrels directed away from the camera's axis; used as Connects the dispersed figures through a common downhill threat direction.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Use natural daytime illumination with muted saturation and grounded contrast, keeping the uphill threat legible without theatrical lighting changes.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The bait supply cart is stopped on the mountain path; gunfire comes from both wooded ridges.\n\nTHE HAND THAT IS DOING THIS: the object at the centre of this shot is being held, operated, read, aimed or handed over by 수많은 일본인 군인들 right now, so 수많은 일본인 군인들's hand — and as much of the wrist and forearm as the framing reaches — is in the frame, gripping or resting on that object exactly the way its real use requires. Match that hand to 수많은 일본인 군인들: its size, build, skin, age, grooming, sleeve and anything worn on it belong to that person and to no one else.\n\nNo people appear unless the shot text itself says so.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.\n\nCOMPOSITION VARIATION (this candidate only): keep the place, the people and the moment exactly as specified, and keep from the CAMERA & FRAME contract its subject scale, screen placement and key background elements exactly. For this candidate only, its camera angle and height are deliberately left open: choose a camera position distinctly different from the obvious one, as a film director picking a second setup on the same blocking. Stay true to any explicit framing the SHOT TEXT itself dictates."
  },
  "roll_refs": {
   "A": [
    {
     "label": "SHOT BACKGROUND",
     "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/db00517b-165b-4005-a287-f28936b59b95/images/272bd7af-9294-4490-98c6-88a40e9deacc/scene/recipe/S39sh14__bgfirst_bg.png",
     "asset_id": "bd7bba71-1819-42da-aace-6a405084bcaa",
     "role": "bgfirst_bg"
    },
    {
     "label": "LAYOUT SKETCH (people placement only)",
     "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/db00517b-165b-4005-a287-f28936b59b95/images/272bd7af-9294-4490-98c6-88a40e9deacc/conti/conti_S39sh14.png",
     "asset_id": "464a7862-b636-4a03-82eb-b0aabb506b77",
     "role": "conti_light"
    }
   ],
   "B": [
    {
     "label": "LOCATION PHOTOGRAPH — the exact place of this shot: its architecture, materials, fixed features and lighting mood are spatial truth; stage the moment inside this place. Never copy its camera framing.",
     "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/db00517b-165b-4005-a287-f28936b59b95/images/272bd7af-9294-4490-98c6-88a40e9deacc/scene/recipe/groupbg_ambush_upper_trees_269326.png",
     "asset_id": "df94889d-3b54-4560-8b7b-cb09c93785df",
     "role": "bgfirst_group_bg"
    }
   ]
  },
  "shot_run_spend_attempt_count": 1,
  "judge_flip_skipped": "judge_fn_owns_order",
  "readings": [
   {
    "label": "A",
    "direction": "우측 능선에 있는 군인들이 좌측 아래의 산길을 향해 일제히 총구를 겨누고 있다.",
    "built_space": "좌측에 흙으로 된 산길이 있고, 우측에는 나무가 우거진 가파른 능선이 있어 레퍼런스의 자연 환경과 잘 일치한다.",
    "entities": "일본군 군복을 입은 수많은 군인들이 묘사되었으나, 프롬프트에 명시된 미끼 보급 마차(bait supply cart)는 화면에 존재하지 않는다.",
    "hard_violations": [
     "[gemini-pro] 물리적으로 불가능한 사물 형태 (여러 군인들의 총열이 몇 미터에 달할 정도로 비정상적으로 길게 늘어나 허공과 나무 사이를 가로지름)"
    ],
    "physics": "군인들은 가파른 경사면에 서 있으나, 여러 군인들의 총열이 비정상적으로 길게 늘어나 있으며 형태가 붕괴되어 물리 법칙을 위반하고 있다."
   },
   {
    "label": "B",
    "direction": "좌측 능선에 매복한 군인들이 우측 아래 산길에 정지해 있는 마차를 향해 총을 겨누고 발사하고 있다.",
    "built_space": "좌측에 가파른 나무 능선이 있고 우측 아래에 흙길이 위치해 있어 레퍼런스의 지형적 특징을 훌륭히 살렸다.",
    "entities": "일본군 군복을 입은 군인들과 우측 하단 산길에 정지해 있는 미끼 보급 마차가 모두 정확히 묘사되었다.",
    "hard_violations": [
     "[gpt-high] 지시된 병사와 보급 수레 외에 말 한 마리를 추가했다."
    ],
    "physics": "군인들은 경사진 지면에 딛고 서서 총을 안정적으로 파지 및 발사하고 있으며, 산길 위에 놓인 마차 역시 바닥에 정상적으로 지지되어 있어 물리적으로 자연스럽다."
   }
  ],
  "cross_model_order": {
   "policy": "cross_model_order_v2_slot0_fwd_slot1_rev_parallel2",
   "slots": [
    {
     "model": "gemini-pro",
     "order": "forward",
     "display_to_canonical": {
      "A": "A",
      "B": "B"
     },
     "raw": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 9,
        "verdict_ko": "프롬프트가 요구한 지형적 특징과 일본군의 매복 자세를 훌륭하게 구현했으며, 특히 명시된 미끼 보급 마차를 산길에 정확히 배치하고 사격 상황을 자연스럽게 연출하여 높은 완성도를 보여줍니다."
       },
       {
        "label": "A",
        "score": 2,
        "verdict_ko": "매복한 군인들의 구도는 잡았으나 필수 요소인 보급 마차가 누락되었고, 무엇보다 여러 군인들의 총열이 화면을 가로지를 정도로 기괴하게 늘어나는 치명적인 물리적 오류가 발생하여 활용이 불가능합니다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "우측 능선에 있는 군인들이 좌측 아래의 산길을 향해 일제히 총구를 겨누고 있다.",
        "built_space": "좌측에 흙으로 된 산길이 있고, 우측에는 나무가 우거진 가파른 능선이 있어 레퍼런스의 자연 환경과 잘 일치한다.",
        "entities": "일본군 군복을 입은 수많은 군인들이 묘사되었으나, 프롬프트에 명시된 미끼 보급 마차(bait supply cart)는 화면에 존재하지 않는다.",
        "hard_violations": [
         "물리적으로 불가능한 사물 형태 (여러 군인들의 총열이 몇 미터에 달할 정도로 비정상적으로 길게 늘어나 허공과 나무 사이를 가로지름)"
        ],
        "physics": "군인들은 가파른 경사면에 서 있으나, 여러 군인들의 총열이 비정상적으로 길게 늘어나 있으며 형태가 붕괴되어 물리 법칙을 위반하고 있다."
       },
       {
        "label": "B",
        "direction": "좌측 능선에 매복한 군인들이 우측 아래 산길에 정지해 있는 마차를 향해 총을 겨누고 발사하고 있다.",
        "built_space": "좌측에 가파른 나무 능선이 있고 우측 아래에 흙길이 위치해 있어 레퍼런스의 지형적 특징을 훌륭히 살렸다.",
        "entities": "일본군 군복을 입은 군인들과 우측 하단 산길에 정지해 있는 미끼 보급 마차가 모두 정확히 묘사되었다.",
        "hard_violations": [],
        "physics": "군인들은 경사진 지면에 딛고 서서 총을 안정적으로 파지 및 발사하고 있으며, 산길 위에 놓인 마차 역시 바닥에 정상적으로 지지되어 있어 물리적으로 자연스럽다."
       }
      ],
      "all_candidates_fail": false
     },
     "normalized": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 9,
        "verdict_ko": "프롬프트가 요구한 지형적 특징과 일본군의 매복 자세를 훌륭하게 구현했으며, 특히 명시된 미끼 보급 마차를 산길에 정확히 배치하고 사격 상황을 자연스럽게 연출하여 높은 완성도를 보여줍니다."
       },
       {
        "label": "A",
        "score": 2,
        "verdict_ko": "매복한 군인들의 구도는 잡았으나 필수 요소인 보급 마차가 누락되었고, 무엇보다 여러 군인들의 총열이 화면을 가로지를 정도로 기괴하게 늘어나는 치명적인 물리적 오류가 발생하여 활용이 불가능합니다."
       }
      ],
      "all_candidates_fail": false,
      "readings": [
       {
        "label": "A",
        "direction": "우측 능선에 있는 군인들이 좌측 아래의 산길을 향해 일제히 총구를 겨누고 있다.",
        "built_space": "좌측에 흙으로 된 산길이 있고, 우측에는 나무가 우거진 가파른 능선이 있어 레퍼런스의 자연 환경과 잘 일치한다.",
        "entities": "일본군 군복을 입은 수많은 군인들이 묘사되었으나, 프롬프트에 명시된 미끼 보급 마차(bait supply cart)는 화면에 존재하지 않는다.",
        "hard_violations": [
         "물리적으로 불가능한 사물 형태 (여러 군인들의 총열이 몇 미터에 달할 정도로 비정상적으로 길게 늘어나 허공과 나무 사이를 가로지름)"
        ],
        "physics": "군인들은 가파른 경사면에 서 있으나, 여러 군인들의 총열이 비정상적으로 길게 늘어나 있으며 형태가 붕괴되어 물리 법칙을 위반하고 있다."
       },
       {
        "label": "B",
        "direction": "좌측 능선에 매복한 군인들이 우측 아래 산길에 정지해 있는 마차를 향해 총을 겨누고 발사하고 있다.",
        "built_space": "좌측에 가파른 나무 능선이 있고 우측 아래에 흙길이 위치해 있어 레퍼런스의 지형적 특징을 훌륭히 살렸다.",
        "entities": "일본군 군복을 입은 군인들과 우측 하단 산길에 정지해 있는 미끼 보급 마차가 모두 정확히 묘사되었다.",
        "hard_violations": [],
        "physics": "군인들은 경사진 지면에 딛고 서서 총을 안정적으로 파지 및 발사하고 있으며, 산길 위에 놓인 마차 역시 바닥에 정상적으로 지지되어 있어 물리적으로 자연스럽다."
       }
      ]
     },
     "ok": true
    },
    {
     "model": "gpt-high",
     "order": "reverse",
     "display_to_canonical": {
      "A": "B",
      "B": "A"
     },
     "raw": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 3,
        "verdict_ko": "나무 사이에 선 병사들의 전신과 측면 조준은 잘 보이지만, 지시되지 않은 말을 추가했고 조준 순간을 총구 화염이 보이는 발포 순간으로 바꾸었다."
       },
       {
        "label": "B",
        "score": 7,
        "verdict_ko": "길 아래쪽에서 비스듬히 올려다본 능선과 나무 사이 병력의 하향 조준이 충실하지만, 여러 병사가 서 있지 않고 쪼그리거나 무릎을 꿇고 있다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "병사들의 얼굴과 소총 총구는 대체로 화면 오른쪽, 비탈 아래 산길 방향을 향한다. 총열은 카메라 정면이 아니라 측면으로 길게 보인다. 앞쪽 병사들의 조준선은 비교적 수평이고, 오른쪽 병사의 총열은 더 뚜렷하게 아래를 향한다. 화면에 의병은 없어 실제 표적에 조준선이 도달하는지는 확인할 수 없다.",
        "built_space": "화면 왼쪽 위로 솟는 흙·바위 비탈에 병사들이 여러 높이로 서 있고, 굵은 나무들이 전경과 중경에서 이들을 불규칙하게 나눈다. 산길은 오른쪽 하단에 있으며 그 위에 보급 수레 한 대가 있다. 인공 건물이나 고정 설비는 없다. 참조의 숲길, 암반과 수풀 재질은 유사하지만 앞쪽 병사들이 상당히 크게 배치된다.",
        "entities": "동아시아계 성인 남성으로 보이는 병사 십여 명 이상이 군모, 갈색 군복, 각반과 장총을 갖췄다. 일본군 역할과는 부합하지만 얼굴만으로 국적을 확인할 수는 없다. 여러 병사의 머리부터 발까지 보이며 일부는 나무와 수풀에 가린다. 보급품을 실은 수레 한 대와 그 앞의 검은 말 한 마리가 보인다. 말은 지시된 등장 대상이 아니다. 여러 총구에 연기와 화염이 있어 단순 조준보다 발포 중인 순간이다. 읽을 수 있는 글자는 없다.",
        "hard_violations": [
         "지시된 병사와 보급 수레 외에 말 한 마리를 추가했다."
        ],
        "physics": "전경 병사들은 양발을 비탈의 서로 다른 높이에 디뎌 체중을 지탱한다. 소총은 어깨에 개머리판을 대고 양손으로 잡고 있어 지지가 분명하다. 뒤쪽 인물의 발 일부는 수풀에 가리지만 공중에 떠 있다고 볼 근거는 없다. 수레는 바퀴로 길에 놓여 있고 말도 다리로 지면을 딛는다. 발포 연기와 화염은 총구 부근에서 발생한다."
       },
       {
        "label": "B",
        "direction": "병사들의 시선과 총열은 공통으로 화면 왼쪽 아래의 산길 또는 그 너머 화면 밖을 향한다. 소총은 대부분 비스듬한 측면으로 보이며 카메라를 직접 겨누지 않는다. 높은 비탈에서 아래쪽을 위협한다는 방향 관계가 명확하다. 의병은 화면 밖이므로 개별 표적까지의 정확한 조준은 확인할 수 없다.",
        "built_space": "돌이 섞인 산길이 하단에서 왼쪽 중경으로 이어지고, 오른쪽 비탈이 화면 위쪽으로 올라간다. 병사들은 길보다 높은 곳에 여러 층으로 분산되어 있다. 깊이가 다른 나무줄기들이 병력 사이를 나누며 선택된 틈에서 전신이 드러난다. 건물이나 고정 설비는 없다. 참조와 같은 흙길·암반·울창한 숲의 성격은 유지하지만 길이 전경에서 차지하는 면적은 다소 크다.",
        "entities": "동아시아계 성인 남성으로 보이는 병사 약 스무 명이 갈색 군복과 군모를 쓰고 장총을 들고 있다. 일본군이라는 배역에 맞는 외형이며 다른 민간인은 보이지 않는다. 일부는 똑바로 서 있지만 여러 명은 무릎을 꿇거나 쪼그려 앉아 있어 '겨누고 선'이라는 지시와 다르다. 여러 인물의 전신이 보이고 일부는 나무에 가린다. 수레와 반대편 능선은 이 구도 밖이므로 누락으로 판단하지 않는다. 읽을 수 있는 글자는 없다.",
        "hard_violations": [],
        "physics": "서 있는 병사들은 비탈의 흙과 돌에 발을 딛고, 낮게 자세를 취한 병사들은 굽힌 다리와 발 또는 무릎으로 몸을 받친다. 가려진 발이 있지만 지지 없이 떠 있는 인물은 확인되지 않는다. 소총은 양손과 어깨로 지탱되며, 경사면에서 조준하는 자세 자체는 물리적으로 가능하다."
       }
      ],
      "all_candidates_fail": false
     },
     "normalized": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 3,
        "verdict_ko": "나무 사이에 선 병사들의 전신과 측면 조준은 잘 보이지만, 지시되지 않은 말을 추가했고 조준 순간을 총구 화염이 보이는 발포 순간으로 바꾸었다."
       },
       {
        "label": "A",
        "score": 7,
        "verdict_ko": "길 아래쪽에서 비스듬히 올려다본 능선과 나무 사이 병력의 하향 조준이 충실하지만, 여러 병사가 서 있지 않고 쪼그리거나 무릎을 꿇고 있다."
       }
      ],
      "all_candidates_fail": false,
      "readings": [
       {
        "label": "B",
        "direction": "병사들의 얼굴과 소총 총구는 대체로 화면 오른쪽, 비탈 아래 산길 방향을 향한다. 총열은 카메라 정면이 아니라 측면으로 길게 보인다. 앞쪽 병사들의 조준선은 비교적 수평이고, 오른쪽 병사의 총열은 더 뚜렷하게 아래를 향한다. 화면에 의병은 없어 실제 표적에 조준선이 도달하는지는 확인할 수 없다.",
        "built_space": "화면 왼쪽 위로 솟는 흙·바위 비탈에 병사들이 여러 높이로 서 있고, 굵은 나무들이 전경과 중경에서 이들을 불규칙하게 나눈다. 산길은 오른쪽 하단에 있으며 그 위에 보급 수레 한 대가 있다. 인공 건물이나 고정 설비는 없다. 참조의 숲길, 암반과 수풀 재질은 유사하지만 앞쪽 병사들이 상당히 크게 배치된다.",
        "entities": "동아시아계 성인 남성으로 보이는 병사 십여 명 이상이 군모, 갈색 군복, 각반과 장총을 갖췄다. 일본군 역할과는 부합하지만 얼굴만으로 국적을 확인할 수는 없다. 여러 병사의 머리부터 발까지 보이며 일부는 나무와 수풀에 가린다. 보급품을 실은 수레 한 대와 그 앞의 검은 말 한 마리가 보인다. 말은 지시된 등장 대상이 아니다. 여러 총구에 연기와 화염이 있어 단순 조준보다 발포 중인 순간이다. 읽을 수 있는 글자는 없다.",
        "hard_violations": [
         "지시된 병사와 보급 수레 외에 말 한 마리를 추가했다."
        ],
        "physics": "전경 병사들은 양발을 비탈의 서로 다른 높이에 디뎌 체중을 지탱한다. 소총은 어깨에 개머리판을 대고 양손으로 잡고 있어 지지가 분명하다. 뒤쪽 인물의 발 일부는 수풀에 가리지만 공중에 떠 있다고 볼 근거는 없다. 수레는 바퀴로 길에 놓여 있고 말도 다리로 지면을 딛는다. 발포 연기와 화염은 총구 부근에서 발생한다."
       },
       {
        "label": "A",
        "direction": "병사들의 시선과 총열은 공통으로 화면 왼쪽 아래의 산길 또는 그 너머 화면 밖을 향한다. 소총은 대부분 비스듬한 측면으로 보이며 카메라를 직접 겨누지 않는다. 높은 비탈에서 아래쪽을 위협한다는 방향 관계가 명확하다. 의병은 화면 밖이므로 개별 표적까지의 정확한 조준은 확인할 수 없다.",
        "built_space": "돌이 섞인 산길이 하단에서 왼쪽 중경으로 이어지고, 오른쪽 비탈이 화면 위쪽으로 올라간다. 병사들은 길보다 높은 곳에 여러 층으로 분산되어 있다. 깊이가 다른 나무줄기들이 병력 사이를 나누며 선택된 틈에서 전신이 드러난다. 건물이나 고정 설비는 없다. 참조와 같은 흙길·암반·울창한 숲의 성격은 유지하지만 길이 전경에서 차지하는 면적은 다소 크다.",
        "entities": "동아시아계 성인 남성으로 보이는 병사 약 스무 명이 갈색 군복과 군모를 쓰고 장총을 들고 있다. 일본군이라는 배역에 맞는 외형이며 다른 민간인은 보이지 않는다. 일부는 똑바로 서 있지만 여러 명은 무릎을 꿇거나 쪼그려 앉아 있어 '겨누고 선'이라는 지시와 다르다. 여러 인물의 전신이 보이고 일부는 나무에 가린다. 수레와 반대편 능선은 이 구도 밖이므로 누락으로 판단하지 않는다. 읽을 수 있는 글자는 없다.",
        "hard_violations": [],
        "physics": "서 있는 병사들은 비탈의 흙과 돌에 발을 딛고, 낮게 자세를 취한 병사들은 굽힌 다리와 발 또는 무릎으로 몸을 받친다. 가려진 발이 있지만 지지 없이 떠 있는 인물은 확인되지 않는다. 소총은 양손과 어깨로 지탱되며, 경사면에서 조준하는 자세 자체는 물리적으로 가능하다."
       }
      ]
     },
     "ok": true
    }
   ],
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "slot_winner_match": false,
   "slot_winner": {
    "gemini-pro": "B",
    "gpt-high": "A"
   },
   "route": "cross_slot_combined"
  },
  "dual": {
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "normalized": {
    "A": 1.222,
    "B": 1.429
   },
   "adjusted": {
    "A": 0.972,
    "B": 1.179
   },
   "violations": {
    "A": [
     "[gemini-pro] 물리적으로 불가능한 사물 형태 (여러 군인들의 총열이 몇 미터에 달할 정도로 비정상적으로 길게 늘어나 허공과 나무 사이를 가로지름)"
    ],
    "B": [
     "[gpt-high] 지시된 병사와 보급 수레 외에 말 한 마리를 추가했다."
    ]
   },
   "per_model_winner": {
    "gemini-pro": "B",
    "gpt-high": "A"
   },
   "agreed": false
  },
  "totals": {
   "B": 1179,
   "A": 972
  },
  "selected": "B",
  "ranking": [
   "B",
   "A"
  ],
  "verdicts": [
   {
    "label": "B",
    "score": 1179,
    "verdict_ko": "프롬프트가 요구한 지형적 특징과 일본군의 매복 자세를 훌륭하게 구현했으며, 특히 명시된 미끼 보급 마차를 산길에 정확히 배치하고 사격 상황을 자연스럽게 연출하여 높은 완성도를 보여줍니다.  ★위반: [gpt-high] 지시된 병사와 보급 수레 외에 말 한 마리를 추가했다."
   },
   {
    "label": "A",
    "score": 972,
    "verdict_ko": "매복한 군인들의 구도는 잡았으나 필수 요소인 보급 마차가 누락되었고, 무엇보다 여러 군인들의 총열이 화면을 가로지를 정도로 기괴하게 늘어나는 치명적인 물리적 오류가 발생하여 활용이 불가능합니다.  ★위반: [gemini-pro] 물리적으로 불가능한 사물 형태 (여러 군인들의 총열이 몇 미터에 달할 정도로 비정상적으로 길게 늘어나 허공과 나무 사이를 가로지름)"
   }
  ],
  "refs": [
   {
    "label": "LOCATION PHOTOGRAPH — the exact place of this shot: its architecture, materials, fixed features and lighting mood are spatial truth; stage the moment inside this place. Never copy its camera framing.",
    "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/db00517b-165b-4005-a287-f28936b59b95/images/272bd7af-9294-4490-98c6-88a40e9deacc/scene/recipe/groupbg_ambush_upper_trees_269326.png",
    "asset_id": "df94889d-3b54-4560-8b7b-cb09c93785df",
    "role": "bgfirst_group_bg"
   }
  ],
  "critique_skipped": true,
  "shot_run_produced": false,
  "shot_run_uid": "06aa1428-66dc-7fdd-b69d-9fe9f7f15d2d",
  "bgfirst": {
   "bg_path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/db00517b-165b-4005-a287-f28936b59b95/images/272bd7af-9294-4490-98c6-88a40e9deacc/scene/recipe/S39sh14__bgfirst_bg.png",
   "bg_asset_id": "bd7bba71-1819-42da-aace-6a405084bcaa",
   "bg_record_key": "S39sh14::bgfirst_bg",
   "chain_winner": false,
   "authority": "groupbg",
   "group_key": "ambush_upper_trees",
   "groupbg_asset_id": "df94889d-3b54-4560-8b7b-cb09c93785df"
  },
  "ref_mode": "그룹 배경+엔티티 (2택1: 무콘티 승)",
  "share_plan": {
   "ref_plan": "background"
  }
 },
 "S39sh14::cine": {
  "applied": true,
  "attempted_at": "2026-09-09T05:17:26.020475+00:00",
  "fingerprint": "bbce09821a4e017b0f6c1964055768eb4b5e8933735df80b6cd6b688926780c3",
  "fingerprint_version": 2,
  "provider": "mai",
  "endpoint": "openrouter/chat-completions",
  "model": "microsoft/mai-image-2.6",
  "pack": "24.202608252115",
  "source_file": "S39sh14_sel.png",
  "source_sha256": "eb71635b8981e8b578744d1c882d1e53ec453a0b63bc8cf85389d7242c5db07b",
  "file": "S39sh14_cine.png",
  "staged_sha256": "e0f0e4a2acb61d8eb6580d6b67d2842f86eeb28563f8acf6327354df90882e55",
  "latency_ms": 31679
 },
 "S39sh18::signage": {
  "fp": "6b03bbbab33467dc",
  "inscriptions": [],
  "cues": [],
  "dropped": []
 },
 "groupbg::ambush_cart_path": {
  "input_fingerprint": "39578afd540e6a44",
  "meta": {
   "model": "gpt-image-2",
   "size": "1536x864",
   "pack": "11.202607220237",
   "contract": "bgfirst_full_v3",
   "group_sig": {
    "key": "ambush_cart_path",
    "tags": [
     "S39sh18",
     "S39sh25"
    ]
   },
   "context_sig": "39d2b930665a57e1"
  },
  "prompt": "Create ONE empty live-action location background photograph — NO PEOPLE, no figures, no body parts, no silhouettes, no shadows or reflections of people anywhere.\n\"Empty\" means no people only: KEEP the location's inherent occupants and stock that define the place — animals in an animal shelter, pen or farm, goods and produce in a market, moored boats in a harbour — unless the shot text explicitly removes them.\nThe FIRST attached image is a thin-line storyboard sketch of a shot that happens at this location: use it ONLY as spatial evidence — what this place contains, how its ground, structures and landmarks are arranged and proportioned. Ignore the sketched people and arrows entirely, and do NOT copy its line style: render a fully photographic, physically plausible real place that fits THE LOCATION text below.\nHUMAN-SCALE CALIBRATION: derive every structure's true size from human-scale elements — a door ≈ 2m, a window ≈ 1–1.5m wide, one storey ≈ 2.5–3m; never inflate a small structure or shrink a large one.\n\nTHE LOCATION — early 20th-century Korea, Jeolla region; all people are Korean unless explicitly stated otherwise: The exposed mountain track beside the seized supply cart, between the wooded slopes firing down on it.\n\nLOCATION DETAIL (descriptive evidence about this place from the production's location records — evidence, not a staging order): use it to understand what kind of place this is and what it permanently contains. Carry over ONLY its enduring physical features — layout, structures, ground and materials, fixed equipment and installations. Do NOT treat any lighting, weather or time-of-day wording, momentary object states, depicted events or subjective impressions in it as fixed facts of the place; the TIME OF DAY lock below is the sole authority for time and lighting.\n새재 산길·수레 습격 지점·길가 숲: 수레 한 대가 지나갈 수 있을 정도의 너비를 가진 산속 흙길. 길 양옆으로 울창한 잡목 숲이 맞닿아 있어 기습에 용이한 지형이다. (특징: 수레가 이동할 수 있는 너비의 다져진 흙길; 수레 타이어 자국이 남은 흙바닥; 길과 맞닿은 빽빽한 잡목림과 수풀)\n\nSCENE EVIDENCE (verbatim quotes from the screenplay about this place — treat them as evidence of what the location physically contains and looks like; stage the PLACE those moments happen in, but do NOT depict the momentary actions, people or staged props themselves):\n- 의병들이 수레에 달라붙는 순간!\n- 그때 총에 맞은 의병 하나가 쓰러진다. 대장장이가 달려가 부축한다.\n- 규홍이 쓰러진 대장장이를 본다. 이를 악문다.\n\nTIME OF DAY (lock): day.\n\nRender ONE photorealistic empty location photograph, 16:9, neutral enough that every shot of this place can be staged from it later. No readable writing anywhere: surfaces that would carry writing may be present, but stage any wording out of legibility — an oblique angle, distance, shallow focus. No captions, watermarks or overlay text.",
  "effective_prompt": "Create ONE empty live-action location background photograph — NO PEOPLE, no figures, no body parts, no silhouettes, no shadows or reflections of people anywhere.\n\"Empty\" means no people only: KEEP the location's inherent occupants and stock that define the place — animals in an animal shelter, pen or farm, goods and produce in a market, moored boats in a harbour — unless the shot text explicitly removes them.\nThe FIRST attached image is a thin-line storyboard sketch of a shot that happens at this location: use it ONLY as spatial evidence — what this place contains, how its ground, structures and landmarks are arranged and proportioned. Ignore the sketched people and arrows entirely, and do NOT copy its line style: render a fully photographic, physically plausible real place that fits THE LOCATION text below.\nHUMAN-SCALE CALIBRATION: derive every structure's true size from human-scale elements — a door ≈ 2m, a window ≈ 1–1.5m wide, one storey ≈ 2.5–3m; never inflate a small structure or shrink a large one.\n\nTHE LOCATION — early 20th-century Korea, Jeolla region; all people are Korean unless explicitly stated otherwise: The exposed mountain track beside the seized supply cart, between the wooded slopes firing down on it.\n\nLOCATION DETAIL (descriptive evidence about this place from the production's location records — evidence, not a staging order): use it to understand what kind of place this is and what it permanently contains. Carry over ONLY its enduring physical features — layout, structures, ground and materials, fixed equipment and installations. Do NOT treat any lighting, weather or time-of-day wording, momentary object states, depicted events or subjective impressions in it as fixed facts of the place; the TIME OF DAY lock below is the sole authority for time and lighting.\n새재 산길·수레 습격 지점·길가 숲: 수레 한 대가 지나갈 수 있을 정도의 너비를 가진 산속 흙길. 길 양옆으로 울창한 잡목 숲이 맞닿아 있어 기습에 용이한 지형이다. (특징: 수레가 이동할 수 있는 너비의 다져진 흙길; 수레 타이어 자국이 남은 흙바닥; 길과 맞닿은 빽빽한 잡목림과 수풀)\n\nSCENE EVIDENCE (verbatim quotes from the screenplay about this place — treat them as evidence of what the location physically contains and looks like; stage the PLACE those moments happen in, but do NOT depict the momentary actions, people or staged props themselves):\n- 의병들이 수레에 달라붙는 순간!\n- 그때 총에 맞은 의병 하나가 쓰러진다. 대장장이가 달려가 부축한다.\n- 규홍이 쓰러진 대장장이를 본다. 이를 악문다.\n\nTIME OF DAY (lock): day.\n\nRender ONE photorealistic empty location photograph, 16:9, neutral enough that every shot of this place can be staged from it later. No readable writing anywhere: surfaces that would carry writing may be present, but stage any wording out of legibility — an oblique angle, distance, shallow focus. No captions, watermarks or overlay text.",
  "bg_path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/db00517b-165b-4005-a287-f28936b59b95/images/272bd7af-9294-4490-98c6-88a40e9deacc/scene/recipe/groupbg_ambush_cart_path_502072.png",
  "asset_id": "0f5668b1-c00a-4ca6-b8eb-c9dd1de1ec36",
  "input_asset_ids": [
   "674fc9be-f179-475f-aebc-486d89cffdc6"
  ],
  "origin_tag": "S39sh18",
  "place_text": "The exposed mountain track beside the seized supply cart, between the wooded slopes firing down on it.",
  "origin_inputs": {
   "place_text": "The exposed mountain track beside the seized supply cart, between the wooded slopes firing down on it.",
   "time_of_day_en": "day",
   "conti_asset_id": "674fc9be-f179-475f-aebc-486d89cffdc6"
  }
 },
 "S39sh18::bgfirst_bg": {
  "input_fingerprint": "996f28c8e59b58fe",
  "prompt": "Create the EMPTY BACKGROUND PLATE for one film shot — NO PEOPLE, no figures, no body parts, no sketch lines, no arrows anywhere.\n\"Empty\" means no people only: KEEP the location's inherent occupants and stock that define the place — animals in an animal shelter, pen or farm, goods in a market, moored boats in a harbour — unless the shot text explicitly removes them.\nThe FIRST attached image is a thin-line storyboard sketch: use ONLY its camera angle, horizon, perspective and the placement/size of buildings and set masses — ignore the sketched people and arrows entirely. The SECOND attached image (LOCATION PHOTOGRAPH) is the real place: take its architecture, materials, signage and fixed features, and RE-PROJECT them into the sketch's camera. If the photograph's camera differs from the sketch's, the sketch's camera wins.\nHUMAN-SCALE CALIBRATION: derive every structure's true size from human-scale elements — a door ≈ 2m, a window ≈ 1–1.5m wide, one storey ≈ 2.5–3m; never inflate a small structure or shrink a large one.\n\nSHOT TEXT this background must serve (Korean): 대장장이의 등 한가운데 찢겨진 옷감 사이로 붉은 피가 확 번져나가는 찰나.\n\nLOCATION (lock): The exposed mountain track beside the seized supply cart, between the wooded slopes firing down on it.\n\nTIME OF DAY (lock): day.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: close-up\n- KEY BACKGROUND ELEMENTS: 숲의 나무 (Partially visible beyond the wounded figure) — Only narrow, oblique background portions remain visible; used as Provides a soft spatial reference around the bodily detail without distracting from the interrupted rescue.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Maintain natural daytime contrast and subdued surrounding color so the spreading red blood registers plainly rather than as a stylized impact effect.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nRender ONE photorealistic empty location photograph, 16:9, that this shot can be staged inside later. No readable writing anywhere: surfaces that would carry writing may be present, but stage any wording out of legibility — an oblique angle, distance, shallow focus. No captions, watermarks or overlay text.",
  "effective_prompt": "Create the EMPTY BACKGROUND PLATE for one film shot — NO PEOPLE, no figures, no body parts, no sketch lines, no arrows anywhere.\n\"Empty\" means no people only: KEEP the location's inherent occupants and stock that define the place — animals in an animal shelter, pen or farm, goods in a market, moored boats in a harbour — unless the shot text explicitly removes them.\nThe FIRST attached image is a thin-line storyboard sketch: use ONLY its camera angle, horizon, perspective and the placement/size of buildings and set masses — ignore the sketched people and arrows entirely. The SECOND attached image (LOCATION PHOTOGRAPH) is the real place: take its architecture, materials, signage and fixed features, and RE-PROJECT them into the sketch's camera. If the photograph's camera differs from the sketch's, the sketch's camera wins.\nHUMAN-SCALE CALIBRATION: derive every structure's true size from human-scale elements — a door ≈ 2m, a window ≈ 1–1.5m wide, one storey ≈ 2.5–3m; never inflate a small structure or shrink a large one.\n\nSHOT TEXT this background must serve (Korean): 대장장이의 등 한가운데 찢겨진 옷감 사이로 붉은 피가 확 번져나가는 찰나.\n\nLOCATION (lock): The exposed mountain track beside the seized supply cart, between the wooded slopes firing down on it.\n\nTIME OF DAY (lock): day.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: close-up\n- KEY BACKGROUND ELEMENTS: 숲의 나무 (Partially visible beyond the wounded figure) — Only narrow, oblique background portions remain visible; used as Provides a soft spatial reference around the bodily detail without distracting from the interrupted rescue.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Maintain natural daytime contrast and subdued surrounding color so the spreading red blood registers plainly rather than as a stylized impact effect.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nRender ONE photorealistic empty location photograph, 16:9, that this shot can be staged inside later. No readable writing anywhere: surfaces that would carry writing may be present, but stage any wording out of legibility — an oblique angle, distance, shallow focus. No captions, watermarks or overlay text.",
  "bg_path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/db00517b-165b-4005-a287-f28936b59b95/images/272bd7af-9294-4490-98c6-88a40e9deacc/scene/recipe/S39sh18__bgfirst_bg.png",
  "asset_id": "28707102-b98d-4a26-9ac3-1ba28087e8b5",
  "input_asset_ids": [
   "674fc9be-f179-475f-aebc-486d89cffdc6",
   "0f5668b1-c00a-4ca6-b8eb-c9dd1de1ec36"
  ]
 },
 "S39sh18": {
  "input_fingerprint": "c98740cf5b2f15b9",
  "prompt": "Create ONE FINAL photorealistic live-action film still of the moment below — early 20th-century Korea, Jeolla region; all people are Korean unless explicitly stated otherwise. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 대장장이의 등 한가운데 찢겨진 옷감 사이로 붉은 피가 확 번져나가는 찰나.\n\nLOCATION (lock): The exposed mountain track beside the seized supply cart, between the wooded slopes firing down on it. The shot takes place here — the attached LOCATION PHOTOGRAPH shows the exact spot.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: close-up\n- KEY BACKGROUND ELEMENTS: 숲의 나무 (Partially visible beyond the wounded figure) — Only narrow, oblique background portions remain visible; used as Provides a soft spatial reference around the bodily detail without distracting from the interrupted rescue.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Maintain natural daytime contrast and subdued surrounding color so the spreading red blood registers plainly rather than as a stylized impact effect.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The supply cart remains abandoned under fire on the wooded path. 대장장이: He has just been shot in the back, with blood spreading through his clothing.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 대장장이 (한국인 남성, 성인 나이대, 자연스러운 검은 머리) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.",
  "roll_prompts": {
   "A": "Stage the shot. The FIRST attached image (SHOT BACKGROUND) is the finished empty background of this shot — keep it EXACTLY: its camera, perspective, architecture, lighting and every fixed feature stay untouched. The SECOND attached image (LAYOUT SKETCH) tells you ONLY where the people go: each sketched person's position, screen size, pose and the gaze/motion arrows. Ignore the sketch's background lines. The CHARACTER REFERENCE photographs show the real people.\nPlace the real people into the background at exactly the sketched positions, sizes and poses, following the arrow directions. No sketch lines or arrows may remain.\n\nCreate ONE FINAL photorealistic live-action film still of the moment below — early 20th-century Korea, Jeolla region; all people are Korean unless explicitly stated otherwise. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 대장장이의 등 한가운데 찢겨진 옷감 사이로 붉은 피가 확 번져나가는 찰나.\n\nLOCATION (lock): The exposed mountain track beside the seized supply cart, between the wooded slopes firing down on it. The shot takes place here — the FIRST attached image (SHOT BACKGROUND) is this exact place, already built: its ground, structures, horizon, materials and lighting are the finished truth of this location and must not be redesigned or replaced. No location photograph is attached — read the place from that image alone, and add no scenery, structure, vehicle or fixture that it does not already show. This lock governs the place only; the figures in the shot follow the staging and pose instructions.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: close-up\n- KEY BACKGROUND ELEMENTS: 숲의 나무 (Partially visible beyond the wounded figure) — Only narrow, oblique background portions remain visible; used as Provides a soft spatial reference around the bodily detail without distracting from the interrupted rescue.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Maintain natural daytime contrast and subdued surrounding color so the spreading red blood registers plainly rather than as a stylized impact effect.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The supply cart remains abandoned under fire on the wooded path. 대장장이: He has just been shot in the back, with blood spreading through his clothing.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 대장장이 (한국인 남성, 성인 나이대, 자연스러운 검은 머리) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.",
   "B": "Create ONE FINAL photorealistic live-action film still of the moment below — early 20th-century Korea, Jeolla region; all people are Korean unless explicitly stated otherwise. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 대장장이의 등 한가운데 찢겨진 옷감 사이로 붉은 피가 확 번져나가는 찰나.\n\nLOCATION (lock): The exposed mountain track beside the seized supply cart, between the wooded slopes firing down on it. The shot takes place here — the attached LOCATION PHOTOGRAPH shows the exact spot.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: close-up\n- KEY BACKGROUND ELEMENTS: 숲의 나무 (Partially visible beyond the wounded figure) — Only narrow, oblique background portions remain visible; used as Provides a soft spatial reference around the bodily detail without distracting from the interrupted rescue.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Maintain natural daytime contrast and subdued surrounding color so the spreading red blood registers plainly rather than as a stylized impact effect.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The supply cart remains abandoned under fire on the wooded path. 대장장이: He has just been shot in the back, with blood spreading through his clothing.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 대장장이 (한국인 남성, 성인 나이대, 자연스러운 검은 머리) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.\n\nCOMPOSITION VARIATION (this candidate only): keep the place, the people and the moment exactly as specified, and keep from the CAMERA & FRAME contract its subject scale, screen placement and key background elements exactly. For this candidate only, its camera angle and height are deliberately left open: choose a camera position distinctly different from the obvious one, as a film director picking a second setup on the same blocking. Stay true to any explicit framing the SHOT TEXT itself dictates."
  },
  "roll_refs": {
   "A": [
    {
     "label": "SHOT BACKGROUND",
     "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/db00517b-165b-4005-a287-f28936b59b95/images/272bd7af-9294-4490-98c6-88a40e9deacc/scene/recipe/S39sh18__bgfirst_bg.png",
     "asset_id": "28707102-b98d-4a26-9ac3-1ba28087e8b5",
     "role": "bgfirst_bg"
    },
    {
     "label": "LAYOUT SKETCH (people placement only)",
     "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/db00517b-165b-4005-a287-f28936b59b95/images/272bd7af-9294-4490-98c6-88a40e9deacc/conti/conti_S39sh18.png",
     "asset_id": "674fc9be-f179-475f-aebc-486d89cffdc6",
     "role": "conti_light"
    },
    {
     "label": "CHARACTER REFERENCE — 대장장이: the exact person appearing in this shot; match face, hair and build exactly.",
     "path": "<bytes:1287019>",
     "asset_id": "d8b98efb-5427-4326-9eea-47f9a34e754a",
     "role": "character_ref"
    }
   ],
   "B": [
    {
     "label": "LOCATION PHOTOGRAPH — the exact place of this shot: its architecture, materials, fixed features and lighting mood are spatial truth; stage the moment inside this place. Never copy its camera framing.",
     "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/db00517b-165b-4005-a287-f28936b59b95/images/272bd7af-9294-4490-98c6-88a40e9deacc/scene/recipe/groupbg_ambush_cart_path_502072.png",
     "asset_id": "0f5668b1-c00a-4ca6-b8eb-c9dd1de1ec36",
     "role": "bgfirst_group_bg"
    },
    {
     "label": "CHARACTER REFERENCE — 대장장이: the exact person appearing in this shot; match face, hair and build exactly.",
     "path": "<bytes:1287019>",
     "asset_id": "d8b98efb-5427-4326-9eea-47f9a34e754a",
     "role": "character_ref"
    }
   ]
  },
  "shot_run_spend_attempt_count": 1,
  "judge_flip_skipped": "judge_fn_owns_order",
  "readings": [
   {
    "label": "A",
    "direction": "카메라를 등지고 상체를 숙인 상태로 앞을 향함.",
    "built_space": "숲이 우거진 산길, 배경 왼쪽에 수레의 일부가 보임.",
    "entities": "대장장이(검은 머리, 푸른색 한복)와 프레임 우측에 대장장이의 팔을 붙잡고 있는 정체불명의 추가 인물.",
    "hard_violations": [
     "[gemini-pro] 프롬프트에 명시되지 않은 추가 인물 등장 (Invented people)",
     "[gpt-high] 대장장이 외에 별도의 인물을 오른쪽 전경에 추가하여, 이 숏에 다른 사람을 등장시키지 말라는 명시적 제한을 위반했다."
    ],
    "physics": "대장장이는 상체를 앞으로 숙이고 있으며, 우측에 있는 다른 인물이 그의 팔을 잡아 지탱해주고 있음."
   },
   {
    "label": "B",
    "direction": "카메라를 등지고 상체를 약간 비틀며 앞을 향함.",
    "built_space": "좁고 노출된 산길과 숲 배경. 프레임이 클로즈업되어 수레는 보이지 않음.",
    "entities": "대장장이(검은 머리, 푸른색 한복). 지시된 대로 단독으로 등장함.",
    "hard_violations": [],
    "physics": "총에 맞아 비틀거리는 자세로 서 있으며, 양팔은 자연스럽게 아래로 향해 있음."
   }
  ],
  "cross_model_order": {
   "policy": "cross_model_order_v2_slot0_fwd_slot1_rev_parallel2",
   "slots": [
    {
     "model": "gemini-pro",
     "order": "forward",
     "display_to_canonical": {
      "A": "A",
      "B": "B"
     },
     "raw": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 9,
        "verdict_ko": "지시된 클로즈업 앵글을 잘 따르며, 추가 인물 없이 대장장이의 등에서 피가 번지는 순간을 사실적으로 묘사하여 프롬프트를 충실히 이행했습니다."
       },
       {
        "label": "A",
        "score": 0,
        "verdict_ko": "대장장이 외에 프롬프트에 명시되지 않은 제3의 인물이 프레임 우측에 등장하여 심각한 규정 위반이 발생했습니다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "카메라를 등지고 상체를 숙인 상태로 앞을 향함.",
        "built_space": "숲이 우거진 산길, 배경 왼쪽에 수레의 일부가 보임.",
        "entities": "대장장이(검은 머리, 푸른색 한복)와 프레임 우측에 대장장이의 팔을 붙잡고 있는 정체불명의 추가 인물.",
        "hard_violations": [
         "프롬프트에 명시되지 않은 추가 인물 등장 (Invented people)"
        ],
        "physics": "대장장이는 상체를 앞으로 숙이고 있으며, 우측에 있는 다른 인물이 그의 팔을 잡아 지탱해주고 있음."
       },
       {
        "label": "B",
        "direction": "카메라를 등지고 상체를 약간 비틀며 앞을 향함.",
        "built_space": "좁고 노출된 산길과 숲 배경. 프레임이 클로즈업되어 수레는 보이지 않음.",
        "entities": "대장장이(검은 머리, 푸른색 한복). 지시된 대로 단독으로 등장함.",
        "hard_violations": [],
        "physics": "총에 맞아 비틀거리는 자세로 서 있으며, 양팔은 자연스럽게 아래로 향해 있음."
       }
      ],
      "all_candidates_fail": false
     },
     "normalized": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 9,
        "verdict_ko": "지시된 클로즈업 앵글을 잘 따르며, 추가 인물 없이 대장장이의 등에서 피가 번지는 순간을 사실적으로 묘사하여 프롬프트를 충실히 이행했습니다."
       },
       {
        "label": "A",
        "score": 0,
        "verdict_ko": "대장장이 외에 프롬프트에 명시되지 않은 제3의 인물이 프레임 우측에 등장하여 심각한 규정 위반이 발생했습니다."
       }
      ],
      "all_candidates_fail": false,
      "readings": [
       {
        "label": "A",
        "direction": "카메라를 등지고 상체를 숙인 상태로 앞을 향함.",
        "built_space": "숲이 우거진 산길, 배경 왼쪽에 수레의 일부가 보임.",
        "entities": "대장장이(검은 머리, 푸른색 한복)와 프레임 우측에 대장장이의 팔을 붙잡고 있는 정체불명의 추가 인물.",
        "hard_violations": [
         "프롬프트에 명시되지 않은 추가 인물 등장 (Invented people)"
        ],
        "physics": "대장장이는 상체를 앞으로 숙이고 있으며, 우측에 있는 다른 인물이 그의 팔을 잡아 지탱해주고 있음."
       },
       {
        "label": "B",
        "direction": "카메라를 등지고 상체를 약간 비틀며 앞을 향함.",
        "built_space": "좁고 노출된 산길과 숲 배경. 프레임이 클로즈업되어 수레는 보이지 않음.",
        "entities": "대장장이(검은 머리, 푸른색 한복). 지시된 대로 단독으로 등장함.",
        "hard_violations": [],
        "physics": "총에 맞아 비틀거리는 자세로 서 있으며, 양팔은 자연스럽게 아래로 향해 있음."
       }
      ]
     },
     "ok": true
    },
    {
     "model": "gpt-high",
     "order": "reverse",
     "display_to_canonical": {
      "A": "B",
      "B": "A"
     },
     "raw": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 8,
        "verdict_ko": "대장장이 한 명의 등 중앙에서 찢긴 옷감에 피가 스미는 순간을 충실히 보여주지만, 요구한 신체 디테일 클로즈업보다 다소 넓고 혈색이 어둡다."
       },
       {
        "label": "B",
        "score": 2,
        "verdict_ko": "등 중앙의 출혈은 보이지만, 허용되지 않은 두 번째 인물을 추가했으며 넓은 배경과 방사형 혈흔이 요구한 밀착 구도와 자연스러운 출혈 표현에서 벗어난다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "대장장이는 카메라에 등을 보이고 머리를 화면 왼쪽 아래로 숙였다. 눈은 보이지 않아 시선의 정확한 대상은 확인할 수 없다. 몸통은 앞으로 기울고 오른팔은 바깥쪽 아래로 벌어져 있다. 등 중앙의 파열부와 출혈이 카메라를 향하며, 총기나 조준 대상은 화면에 없다.",
        "built_space": "인물 양옆으로 흙길과 나무, 수풀이 보인다. 인공 구조물이나 수레는 보이지 않으며 중복된 시설도 없다. 숲속 산길의 재료와 낮의 환경은 참조와 부합하지만, 고유 지형까지 확인하기에는 배경이 흐리다. 배경과 양팔이 요구한 좁은 주변부보다 다소 많이 포함된다.",
        "entities": "인물은 한 명뿐이다. 검은 머리와 드러난 귀·목·손은 성인 남성의 외형에 부합하며, 얼굴이 가려져 정확한 신원과 한국인 여부는 독립적으로 확인하기 어렵다. 청회색의 낡은 전통 상의, 밝은 깃과 접힌 소매는 인물 참조와 잘 맞는다. 등 한가운데 찢긴 천 사이로 피부와 젖은 붉은 피가 보이고 주변 천에는 짙은 혈액이 스며 있다. 수레는 프레임 밖이며 읽을 수 있는 글자나 다른 소품은 없다.",
        "hard_violations": [],
        "physics": "머리와 팔이 몸통에 자연스럽게 연결되고, 옷은 어깨와 몸통에 걸쳐 주름진다. 하체와 발은 잘려 지면 접촉은 확인할 수 없지만, 보이는 자세는 총격 직후 앞으로 휘청이는 동작으로 가능하며 공중에 떠 있다는 증거는 없다. 혈액은 찢긴 천과 피부에 붙어 광택을 띠고 일부 아래로 흐르므로 물리적인 지지가 있다."
       },
       {
        "label": "B",
        "direction": "대장장이는 카메라에 등을 보인 채 화면 오른쪽 앞의 다른 인물 쪽으로 몸과 머리를 숙이고 오른팔을 뻗는다. 상대 인물의 얼굴은 대부분 잘려 정확한 시선은 확인되지 않는다. 등 중앙의 상처는 카메라에 드러나지만, 자세가 단독 피격 순간보다 다른 사람에게 기대는 장면으로 읽힌다. 무기나 총구는 없다.",
        "built_space": "왼쪽에 나무와 바위 사면, 아래와 오른쪽에 흙길이 보이며 왼쪽 중경에는 목재 수레의 적재함 일부가 한 대 보인다. 숲과 돌, 길의 재료는 장소 참조와 대체로 맞는다. 다만 배경이 넓고 비교적 선명하며 두 번째 인물이 오른쪽을 크게 차지해, 상처 주위의 좁은 배경만 남기는 구도와 다르다.",
        "entities": "검은 머리와 청회색 전통 상의를 입은 대장장이의 후면 외형은 참조에 대체로 부합한다. 얼굴은 보이지 않아 정확한 신원은 확인할 수 없다. 등 중앙의 긴 파열부에는 밝은 붉은 피와 주변 혈흔이 있다. 오른쪽에는 회갈색 전통 옷을 입은 별도의 성인 남성으로 보이는 인물이 추가되어 인물 제한을 위반한다. 목재 수레 일부는 보이고 읽을 수 있는 글자는 없다.",
        "hard_violations": [
         "대장장이 외에 별도의 인물을 오른쪽 전경에 추가하여, 이 숏에 다른 사람을 등장시키지 말라는 명시적 제한을 위반했다."
        ],
        "physics": "대장장이의 상체가 앞으로 기울고 오른팔이 옆 인물의 몸과 겹쳐 접촉하는 듯 보이나, 실제로 붙잡는 손은 프레임 밖이라 확정할 수 없다. 하체와 지면 접촉도 잘려 있지만, 몸이 공중에 떠 있는 것으로 보이지는 않는다. 옷과 혈액은 몸 표면에 지지되어 있다. 다만 혈흔이 뚜렷한 방사형으로 퍼져 있어 천을 통해 번지는 피보다 연출된 충격 효과에 가깝다."
       }
      ],
      "all_candidates_fail": false
     },
     "normalized": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 8,
        "verdict_ko": "대장장이 한 명의 등 중앙에서 찢긴 옷감에 피가 스미는 순간을 충실히 보여주지만, 요구한 신체 디테일 클로즈업보다 다소 넓고 혈색이 어둡다."
       },
       {
        "label": "A",
        "score": 2,
        "verdict_ko": "등 중앙의 출혈은 보이지만, 허용되지 않은 두 번째 인물을 추가했으며 넓은 배경과 방사형 혈흔이 요구한 밀착 구도와 자연스러운 출혈 표현에서 벗어난다."
       }
      ],
      "all_candidates_fail": false,
      "readings": [
       {
        "label": "B",
        "direction": "대장장이는 카메라에 등을 보이고 머리를 화면 왼쪽 아래로 숙였다. 눈은 보이지 않아 시선의 정확한 대상은 확인할 수 없다. 몸통은 앞으로 기울고 오른팔은 바깥쪽 아래로 벌어져 있다. 등 중앙의 파열부와 출혈이 카메라를 향하며, 총기나 조준 대상은 화면에 없다.",
        "built_space": "인물 양옆으로 흙길과 나무, 수풀이 보인다. 인공 구조물이나 수레는 보이지 않으며 중복된 시설도 없다. 숲속 산길의 재료와 낮의 환경은 참조와 부합하지만, 고유 지형까지 확인하기에는 배경이 흐리다. 배경과 양팔이 요구한 좁은 주변부보다 다소 많이 포함된다.",
        "entities": "인물은 한 명뿐이다. 검은 머리와 드러난 귀·목·손은 성인 남성의 외형에 부합하며, 얼굴이 가려져 정확한 신원과 한국인 여부는 독립적으로 확인하기 어렵다. 청회색의 낡은 전통 상의, 밝은 깃과 접힌 소매는 인물 참조와 잘 맞는다. 등 한가운데 찢긴 천 사이로 피부와 젖은 붉은 피가 보이고 주변 천에는 짙은 혈액이 스며 있다. 수레는 프레임 밖이며 읽을 수 있는 글자나 다른 소품은 없다.",
        "hard_violations": [],
        "physics": "머리와 팔이 몸통에 자연스럽게 연결되고, 옷은 어깨와 몸통에 걸쳐 주름진다. 하체와 발은 잘려 지면 접촉은 확인할 수 없지만, 보이는 자세는 총격 직후 앞으로 휘청이는 동작으로 가능하며 공중에 떠 있다는 증거는 없다. 혈액은 찢긴 천과 피부에 붙어 광택을 띠고 일부 아래로 흐르므로 물리적인 지지가 있다."
       },
       {
        "label": "A",
        "direction": "대장장이는 카메라에 등을 보인 채 화면 오른쪽 앞의 다른 인물 쪽으로 몸과 머리를 숙이고 오른팔을 뻗는다. 상대 인물의 얼굴은 대부분 잘려 정확한 시선은 확인되지 않는다. 등 중앙의 상처는 카메라에 드러나지만, 자세가 단독 피격 순간보다 다른 사람에게 기대는 장면으로 읽힌다. 무기나 총구는 없다.",
        "built_space": "왼쪽에 나무와 바위 사면, 아래와 오른쪽에 흙길이 보이며 왼쪽 중경에는 목재 수레의 적재함 일부가 한 대 보인다. 숲과 돌, 길의 재료는 장소 참조와 대체로 맞는다. 다만 배경이 넓고 비교적 선명하며 두 번째 인물이 오른쪽을 크게 차지해, 상처 주위의 좁은 배경만 남기는 구도와 다르다.",
        "entities": "검은 머리와 청회색 전통 상의를 입은 대장장이의 후면 외형은 참조에 대체로 부합한다. 얼굴은 보이지 않아 정확한 신원은 확인할 수 없다. 등 중앙의 긴 파열부에는 밝은 붉은 피와 주변 혈흔이 있다. 오른쪽에는 회갈색 전통 옷을 입은 별도의 성인 남성으로 보이는 인물이 추가되어 인물 제한을 위반한다. 목재 수레 일부는 보이고 읽을 수 있는 글자는 없다.",
        "hard_violations": [
         "대장장이 외에 별도의 인물을 오른쪽 전경에 추가하여, 이 숏에 다른 사람을 등장시키지 말라는 명시적 제한을 위반했다."
        ],
        "physics": "대장장이의 상체가 앞으로 기울고 오른팔이 옆 인물의 몸과 겹쳐 접촉하는 듯 보이나, 실제로 붙잡는 손은 프레임 밖이라 확정할 수 없다. 하체와 지면 접촉도 잘려 있지만, 몸이 공중에 떠 있는 것으로 보이지는 않는다. 옷과 혈액은 몸 표면에 지지되어 있다. 다만 혈흔이 뚜렷한 방사형으로 퍼져 있어 천을 통해 번지는 피보다 연출된 충격 효과에 가깝다."
       }
      ]
     },
     "ok": true
    }
   ],
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "slot_winner_match": true,
   "slot_winner": {
    "gemini-pro": "B",
    "gpt-high": "B"
   },
   "route": "cross_slot_agree"
  },
  "dual": {
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "normalized": {
    "A": 0.25,
    "B": 2.0
   },
   "adjusted": {
    "A": 0.0,
    "B": 2.0
   },
   "violations": {
    "A": [
     "[gemini-pro] 프롬프트에 명시되지 않은 추가 인물 등장 (Invented people)",
     "[gpt-high] 대장장이 외에 별도의 인물을 오른쪽 전경에 추가하여, 이 숏에 다른 사람을 등장시키지 말라는 명시적 제한을 위반했다."
    ]
   },
   "per_model_winner": {
    "gemini-pro": "B",
    "gpt-high": "B"
   },
   "agreed": true
  },
  "totals": {
   "B": 2000,
   "A": 0
  },
  "selected": "B",
  "ranking": [
   "B",
   "A"
  ],
  "verdicts": [
   {
    "label": "B",
    "score": 2000,
    "verdict_ko": "지시된 클로즈업 앵글을 잘 따르며, 추가 인물 없이 대장장이의 등에서 피가 번지는 순간을 사실적으로 묘사하여 프롬프트를 충실히 이행했습니다."
   },
   {
    "label": "A",
    "score": 0,
    "verdict_ko": "대장장이 외에 프롬프트에 명시되지 않은 제3의 인물이 프레임 우측에 등장하여 심각한 규정 위반이 발생했습니다.  ★위반: [gemini-pro] 프롬프트에 명시되지 않은 추가 인물 등장 (Invented people) / [gpt-high] 대장장이 외에 별도의 인물을 오른쪽 전경에 추가하여, 이 숏에 다른 사람을 등장시키지 말라는 명시적 제한을 위반했다."
   }
  ],
  "refs": [
   {
    "label": "LOCATION PHOTOGRAPH — the exact place of this shot: its architecture, materials, fixed features and lighting mood are spatial truth; stage the moment inside this place. Never copy its camera framing.",
    "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/db00517b-165b-4005-a287-f28936b59b95/images/272bd7af-9294-4490-98c6-88a40e9deacc/scene/recipe/groupbg_ambush_cart_path_502072.png",
    "asset_id": "0f5668b1-c00a-4ca6-b8eb-c9dd1de1ec36",
    "role": "bgfirst_group_bg"
   },
   {
    "label": "CHARACTER REFERENCE — 대장장이: the exact person appearing in this shot; match face, hair and build exactly.",
    "path": "<bytes:1287019>",
    "asset_id": "d8b98efb-5427-4326-9eea-47f9a34e754a",
    "role": "character_ref"
   }
  ],
  "critique_skipped": true,
  "shot_run_produced": false,
  "shot_run_uid": "06aa1428-6b1a-738f-ad3c-49967bfdead7",
  "bgfirst": {
   "bg_path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/db00517b-165b-4005-a287-f28936b59b95/images/272bd7af-9294-4490-98c6-88a40e9deacc/scene/recipe/S39sh18__bgfirst_bg.png",
   "bg_asset_id": "28707102-b98d-4a26-9ac3-1ba28087e8b5",
   "bg_record_key": "S39sh18::bgfirst_bg",
   "chain_winner": false,
   "authority": "groupbg",
   "group_key": "ambush_cart_path",
   "groupbg_asset_id": "0f5668b1-c00a-4ca6-b8eb-c9dd1de1ec36"
  },
  "ref_mode": "그룹 배경+엔티티 (2택1: 무콘티 승)",
  "share_plan": {
   "ref_plan": "background"
  }
 },
 "S39sh18::cine": {
  "applied": false,
  "attempted_at": "2026-09-09T09:08:33.684640+00:00",
  "fingerprint": "1fc2b88e7c291eb0b95388133e1692c76a2033ac791473637f4a630557b88c3f",
  "fingerprint_version": 2,
  "provider": "mai",
  "endpoint": "openrouter/chat-completions",
  "model": "microsoft/mai-image-2.6",
  "pack": "24.202608252115",
  "source_file": "S39sh18_sel.png",
  "source_sha256": "95c32f1981917fd5e86e59622b646f3f493ceba2df74bfea3e1748a7853fc884",
  "error": "RuntimeError: moderation blocked: {\"message\": \"Response content blocked by label 'MultiSeverity_ViolenceScore'.\", \"code\": 400}",
  "moderation_refusals": 2,
  "declined": true,
  "declined_reason": "moderation"
 },
 "S39sh25::signage": {
  "fp": "6728677b96cbc9e5",
  "inscriptions": [],
  "cues": [],
  "dropped": []
 },
 "S39sh25": {
  "input_fingerprint": "01410c81c6788fbb",
  "prompt": "Create ONE FINAL photorealistic live-action film still of the moment below — early 20th-century Korea, Jeolla region; all people are Korean unless explicitly stated otherwise. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 바닥에 미동 없이 엎어진 대장장이를 내려다보며 이를 꽉 악문 안규홍의 참담한 표정.\n\nLOCATION (lock): The tree-lined edge of the mountain-pass ambush site, beside the route back into forest cover. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: close-up\n- KEY BACKGROUND ELEMENTS: 숲 (Visible only as a limited background behind 안규홍); used as Maintains location continuity while leaving the dead man outside the close framing.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Retain sober natural daylight and restrained saturation, allowing readable facial tension without adding a dramatic lighting cue.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The people visible in that photo are NOT in this shot: never carry their faces, bodies or clothing onto anyone here. Each person in THIS shot is defined solely by the PEOPLE list, their own CHARACTER REFERENCE images and their wardrobe notes. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nIMMOBILE CHARACTER POSE — CANONICAL (identical wherever this character appears in ANY panel; on any conflict THIS POSE WINS): His dead body is collapsed motionless on the mountain-path ground, with blood spreading across his back. The screenplay does not establish which side of his torso touches the ground, his head direction, or the exact positions of his arms and legs.\n\nIMMOBILE BODIES OBEY GRAVITY: a person who is dead or unconscious\nexerts NO muscular effort. Every part of their body — head, torso,\narms, hands, fingers, legs — rests fully on whatever supports it\n(floor, wall, furniture, their own lap) and hangs or slumps with\ngravity. NEVER show any part of an immobile person's body lifted,\nraised, held up in the air, or posed as if presenting something:\nan object in their grip stays clenched in a hand that itself lies\nfallen on a support — the hand does not hold the object up. If the\ncanonical pose leaves a body part unspecified, resolve it as the\nmost gravity-compliant, fully-supported position.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The rice cart remains behind; a fresh bullet impact marks the nearby tree. 안규홍: He pauses before retreat, jaw clenched, with his waist sickle and jade ring retained.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 안규홍 (한국인 남성, 청년 나이대, 검은 머리, 단단히 묶은 머리) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.",
  "roll_prompts": {
   "A": "Create ONE FINAL photorealistic live-action film still of the moment below — early 20th-century Korea, Jeolla region; all people are Korean unless explicitly stated otherwise. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 바닥에 미동 없이 엎어진 대장장이를 내려다보며 이를 꽉 악문 안규홍의 참담한 표정.\n\nLOCATION (lock): The tree-lined edge of the mountain-pass ambush site, beside the route back into forest cover. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: close-up\n- KEY BACKGROUND ELEMENTS: 숲 (Visible only as a limited background behind 안규홍); used as Maintains location continuity while leaving the dead man outside the close framing.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Retain sober natural daylight and restrained saturation, allowing readable facial tension without adding a dramatic lighting cue.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The people visible in that photo are NOT in this shot: never carry their faces, bodies or clothing onto anyone here. Each person in THIS shot is defined solely by the PEOPLE list, their own CHARACTER REFERENCE images and their wardrobe notes. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nIMMOBILE CHARACTER POSE — CANONICAL (identical wherever this character appears in ANY panel; on any conflict THIS POSE WINS): His dead body is collapsed motionless on the mountain-path ground, with blood spreading across his back. The screenplay does not establish which side of his torso touches the ground, his head direction, or the exact positions of his arms and legs.\n\nIMMOBILE BODIES OBEY GRAVITY: a person who is dead or unconscious\nexerts NO muscular effort. Every part of their body — head, torso,\narms, hands, fingers, legs — rests fully on whatever supports it\n(floor, wall, furniture, their own lap) and hangs or slumps with\ngravity. NEVER show any part of an immobile person's body lifted,\nraised, held up in the air, or posed as if presenting something:\nan object in their grip stays clenched in a hand that itself lies\nfallen on a support — the hand does not hold the object up. If the\ncanonical pose leaves a body part unspecified, resolve it as the\nmost gravity-compliant, fully-supported position.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The rice cart remains behind; a fresh bullet impact marks the nearby tree. 안규홍: He pauses before retreat, jaw clenched, with his waist sickle and jade ring retained.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 안규홍 (한국인 남성, 청년 나이대, 검은 머리, 단단히 묶은 머리) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.",
   "B": "Create ONE FINAL photorealistic live-action film still of the moment below — early 20th-century Korea, Jeolla region; all people are Korean unless explicitly stated otherwise. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 바닥에 미동 없이 엎어진 대장장이를 내려다보며 이를 꽉 악문 안규홍의 참담한 표정.\n\nLOCATION (lock): The tree-lined edge of the mountain-pass ambush site, beside the route back into forest cover. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: close-up\n- KEY BACKGROUND ELEMENTS: 숲 (Visible only as a limited background behind 안규홍); used as Maintains location continuity while leaving the dead man outside the close framing.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Retain sober natural daylight and restrained saturation, allowing readable facial tension without adding a dramatic lighting cue.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The people visible in that photo are NOT in this shot: never carry their faces, bodies or clothing onto anyone here. Each person in THIS shot is defined solely by the PEOPLE list, their own CHARACTER REFERENCE images and their wardrobe notes. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nIMMOBILE CHARACTER POSE — CANONICAL (identical wherever this character appears in ANY panel; on any conflict THIS POSE WINS): His dead body is collapsed motionless on the mountain-path ground, with blood spreading across his back. The screenplay does not establish which side of his torso touches the ground, his head direction, or the exact positions of his arms and legs.\n\nIMMOBILE BODIES OBEY GRAVITY: a person who is dead or unconscious\nexerts NO muscular effort. Every part of their body — head, torso,\narms, hands, fingers, legs — rests fully on whatever supports it\n(floor, wall, furniture, their own lap) and hangs or slumps with\ngravity. NEVER show any part of an immobile person's body lifted,\nraised, held up in the air, or posed as if presenting something:\nan object in their grip stays clenched in a hand that itself lies\nfallen on a support — the hand does not hold the object up. If the\ncanonical pose leaves a body part unspecified, resolve it as the\nmost gravity-compliant, fully-supported position.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The rice cart remains behind; a fresh bullet impact marks the nearby tree. 안규홍: He pauses before retreat, jaw clenched, with his waist sickle and jade ring retained.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 안규홍 (한국인 남성, 청년 나이대, 검은 머리, 단단히 묶은 머리) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.\n\nCOMPOSITION VARIATION (this candidate only): keep the place, the people and the moment exactly as specified, and keep from the CAMERA & FRAME contract its subject scale, screen placement and key background elements exactly. For this candidate only, its camera angle and height are deliberately left open: choose a camera position distinctly different from the obvious one, as a film director picking a second setup on the same blocking. Stay true to any explicit framing the SHOT TEXT itself dictates."
  },
  "shot_run_spend_attempt_count": 1,
  "readings": [
   {
    "label": "A",
    "direction": "안규홍의 시선이 화면 밖 아래쪽 바닥을 뚜렷하게 향하고 있다.",
    "built_space": "흙과 돌이 깔린 산길이며, 안규홍의 등 뒤로 숲의 나무들이 연속성을 띠고 배치되어 있다.",
    "entities": "안규홍의 얼굴과 의상이 레퍼런스와 정확히 일치한다. 허리띠에 낫이 꽂혀 있고 오른손 손가락에 옥가락지를 끼고 있다. 지시대로 대장장이의 시체는 화면에 나타나지 않는다.",
    "hard_violations": [],
    "physics": "안규홍은 땅에 발을 디디고 서 있는 자세로 체중을 지탱하고 있으며, 허리에 꽂힌 낫은 천 허리띠에 의해 물리적으로 자연스럽게 고정되어 있다."
   },
   {
    "label": "B",
    "direction": "안규홍의 시선이 화면 우측 하단에 쓰러진 대장장이를 향하고 있다.",
    "built_space": "산길 바닥과 흙길, 배경의 나무들이 보이나 카메라 구도가 낮아져 있다.",
    "entities": "안규홍의 얼굴과 의상은 일치하나, 화면에서 배제되어야 할 대장장이 시체가 프레임 안에 포함되었다. 대장장이는 등에 피가 퍼져야 한다는 설정과 달리 가슴 쪽에 피가 묻은 채 정면을 향해 누워 있다. 안규홍의 옥가락지와 허리 낫이 보이지 않는다.",
    "hard_violations": [
     "[gemini-pro] 화면 왼쪽 아래에 안규홍의 자세와 물리적으로 연결될 수 없는 위치에 짚신을 신은 발이 잘린 채 그려져 있어 해부학적으로 불가능함 (physically impossible anatomy)",
     "[gemini-pro] 프레임 밖으로 배제되어야 할 인물(대장장이 시체)이 화면에 포함됨 (person placed where the staging does not put them)",
     "[gpt-high] 명시적으로 클로즈업 밖에 남겨야 하는 대장장이의 얼굴과 몸을 화면 안에 배치했다."
    ],
    "physics": "안규홍이 웅크리거나 주저앉은 자세를 취하고 있으나, 좌측 하단의 짚신 신은 발은 몸의 중심이나 각도와 맞지 않게 허공이나 땅바닥에 기형적으로 놓여 있다. 대장장이 시체는 바닥에 누워 중력의 영향을 받고 있다."
   }
  ],
  "cross_model_order": {
   "policy": "cross_model_order_v2_slot0_fwd_slot1_rev_parallel2",
   "slots": [
    {
     "model": "gemini-pro",
     "order": "forward",
     "display_to_canonical": {
      "A": "A",
      "B": "B"
     },
     "raw": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 10,
        "verdict_ko": "지시문이 요구한 클로즈업 앵글을 유지하여 시체를 프레임 밖으로 올바르게 배제했고, 요구된 소품(옥가락지, 허리 낫)과 이를 꽉 악문 참담한 표정을 매우 충실하게 구현했습니다."
       },
       {
        "label": "B",
        "score": 2,
        "verdict_ko": "시체를 프레임에서 배제하라는 명시적인 카메라 프레이밍 지시를 위반했으며, 요구된 소품이 누락되었고 해부학적으로 불가능한 신체 부위가 묘사되었습니다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "안규홍의 시선이 화면 밖 아래쪽 바닥을 뚜렷하게 향하고 있다.",
        "built_space": "흙과 돌이 깔린 산길이며, 안규홍의 등 뒤로 숲의 나무들이 연속성을 띠고 배치되어 있다.",
        "entities": "안규홍의 얼굴과 의상이 레퍼런스와 정확히 일치한다. 허리띠에 낫이 꽂혀 있고 오른손 손가락에 옥가락지를 끼고 있다. 지시대로 대장장이의 시체는 화면에 나타나지 않는다.",
        "hard_violations": [],
        "physics": "안규홍은 땅에 발을 디디고 서 있는 자세로 체중을 지탱하고 있으며, 허리에 꽂힌 낫은 천 허리띠에 의해 물리적으로 자연스럽게 고정되어 있다."
       },
       {
        "label": "B",
        "direction": "안규홍의 시선이 화면 우측 하단에 쓰러진 대장장이를 향하고 있다.",
        "built_space": "산길 바닥과 흙길, 배경의 나무들이 보이나 카메라 구도가 낮아져 있다.",
        "entities": "안규홍의 얼굴과 의상은 일치하나, 화면에서 배제되어야 할 대장장이 시체가 프레임 안에 포함되었다. 대장장이는 등에 피가 퍼져야 한다는 설정과 달리 가슴 쪽에 피가 묻은 채 정면을 향해 누워 있다. 안규홍의 옥가락지와 허리 낫이 보이지 않는다.",
        "hard_violations": [
         "화면 왼쪽 아래에 안규홍의 자세와 물리적으로 연결될 수 없는 위치에 짚신을 신은 발이 잘린 채 그려져 있어 해부학적으로 불가능함 (physically impossible anatomy)",
         "프레임 밖으로 배제되어야 할 인물(대장장이 시체)이 화면에 포함됨 (person placed where the staging does not put them)"
        ],
        "physics": "안규홍이 웅크리거나 주저앉은 자세를 취하고 있으나, 좌측 하단의 짚신 신은 발은 몸의 중심이나 각도와 맞지 않게 허공이나 땅바닥에 기형적으로 놓여 있다. 대장장이 시체는 바닥에 누워 중력의 영향을 받고 있다."
       }
      ],
      "all_candidates_fail": false
     },
     "normalized": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 10,
        "verdict_ko": "지시문이 요구한 클로즈업 앵글을 유지하여 시체를 프레임 밖으로 올바르게 배제했고, 요구된 소품(옥가락지, 허리 낫)과 이를 꽉 악문 참담한 표정을 매우 충실하게 구현했습니다."
       },
       {
        "label": "B",
        "score": 2,
        "verdict_ko": "시체를 프레임에서 배제하라는 명시적인 카메라 프레이밍 지시를 위반했으며, 요구된 소품이 누락되었고 해부학적으로 불가능한 신체 부위가 묘사되었습니다."
       }
      ],
      "all_candidates_fail": false,
      "readings": [
       {
        "label": "A",
        "direction": "안규홍의 시선이 화면 밖 아래쪽 바닥을 뚜렷하게 향하고 있다.",
        "built_space": "흙과 돌이 깔린 산길이며, 안규홍의 등 뒤로 숲의 나무들이 연속성을 띠고 배치되어 있다.",
        "entities": "안규홍의 얼굴과 의상이 레퍼런스와 정확히 일치한다. 허리띠에 낫이 꽂혀 있고 오른손 손가락에 옥가락지를 끼고 있다. 지시대로 대장장이의 시체는 화면에 나타나지 않는다.",
        "hard_violations": [],
        "physics": "안규홍은 땅에 발을 디디고 서 있는 자세로 체중을 지탱하고 있으며, 허리에 꽂힌 낫은 천 허리띠에 의해 물리적으로 자연스럽게 고정되어 있다."
       },
       {
        "label": "B",
        "direction": "안규홍의 시선이 화면 우측 하단에 쓰러진 대장장이를 향하고 있다.",
        "built_space": "산길 바닥과 흙길, 배경의 나무들이 보이나 카메라 구도가 낮아져 있다.",
        "entities": "안규홍의 얼굴과 의상은 일치하나, 화면에서 배제되어야 할 대장장이 시체가 프레임 안에 포함되었다. 대장장이는 등에 피가 퍼져야 한다는 설정과 달리 가슴 쪽에 피가 묻은 채 정면을 향해 누워 있다. 안규홍의 옥가락지와 허리 낫이 보이지 않는다.",
        "hard_violations": [
         "화면 왼쪽 아래에 안규홍의 자세와 물리적으로 연결될 수 없는 위치에 짚신을 신은 발이 잘린 채 그려져 있어 해부학적으로 불가능함 (physically impossible anatomy)",
         "프레임 밖으로 배제되어야 할 인물(대장장이 시체)이 화면에 포함됨 (person placed where the staging does not put them)"
        ],
        "physics": "안규홍이 웅크리거나 주저앉은 자세를 취하고 있으나, 좌측 하단의 짚신 신은 발은 몸의 중심이나 각도와 맞지 않게 허공이나 땅바닥에 기형적으로 놓여 있다. 대장장이 시체는 바닥에 누워 중력의 영향을 받고 있다."
       }
      ]
     },
     "ok": true
    },
    {
     "model": "gpt-high",
     "order": "reverse",
     "display_to_canonical": {
      "A": "B",
      "B": "A"
     },
     "raw": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 2,
        "verdict_ko": "얼굴의 크기와 악문 이는 적절하지만, 제외해야 할 대장장이를 뒤쪽에 노출하고 시선은 그와 반대인 앞쪽 아래로 향해 핵심 배치를 위반한다."
       },
       {
        "label": "B",
        "score": 6,
        "verdict_ko": "화면 밖 아래를 보는 시선과 단독 인물 배치는 맞지만, 허리까지 담은 중간 크기 구도여서 얼굴 클로즈업과 제한된 숲 배경이라는 지시를 충족하지 못한다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "안규홍은 고개를 숙여 자신의 앞쪽 아래를 본다. 대장장이는 안규홍의 뒤쪽인 화면 오른쪽 아래에 누워 있어, 보이는 시선이 그에게 닿지 않는다. 무기나 이동 동작은 보이지 않는다.",
        "built_space": "흙길과 수목, 오른쪽의 굵은 나무줄기가 보이며 인공 구조물이나 고정 설비는 없다. 안규홍의 얼굴과 상체가 전경을 차지하고 대장장이가 바로 뒤 지면에 놓여 있다. 숲길의 재질과 낮 조명은 참고와 유사하지만, 시신을 프레임 밖에 두라는 공간 배치를 어겼다.",
        "entities": "안규홍은 참고와 유사한 얼굴의 젊은 동아시아계 남성으로, 뒤로 모은 검은 머리와 때 묻은 회백색 저고리가 맞는다. 찌푸린 미간과 드러난 맞물린 이가 고통과 분노를 나타낸다. 청색 옷의 대장장이까지 보여 총 두 사람이 등장하며, 그의 상처와 혈흔은 등이 아니라 위로 드러난 가슴 쪽에 보인다. 낫, 옥가락지, 쌀수레와 탄흔은 확인되지 않으며 읽을 수 있는 문자는 없다.",
        "hard_violations": [
         "명시적으로 클로즈업 밖에 남겨야 하는 대장장이의 얼굴과 몸을 화면 안에 배치했다."
        ],
        "physics": "안규홍의 머리는 목과 몸통으로 지지되고 상체는 앞으로 숙여져 있다. 하체와 접지점은 프레임 밖이므로 지지 불능이라고 볼 근거는 없다. 대장장이의 머리와 몸통은 지면에 누워 있고, 왼쪽 아래에 보이는 손과 팔도 지면 가까이 놓여 있다. 떠 있는 신체나 물체는 보이지 않는다."
       },
       {
        "label": "B",
        "direction": "안규홍의 눈과 고개가 정면 아래의 화면 밖을 향한다. 대장장이는 보이지 않아 시선의 실제 도착점은 확인할 수 없지만, 바닥의 시신을 내려다보는 지시와 양립한다. 허리의 낫은 위로 굽은 채 수납되어 있으며 누구를 겨누거나 휘두르지 않는다.",
        "built_space": "안규홍이 중앙에 있고 왼쪽 뒤로 흙과 돌이 섞인 산길이 이어진다. 양옆에 여러 나무줄기와 관목이 있으며 인공 구조물이나 고정 설비는 없다. 숲길과 자연광은 장소 설정에 부합하지만, 허리까지 보여 배경 숲의 면적이 크고 지정된 클로즈업보다 넓다.",
        "entities": "한 명의 젊은 동아시아계 남성만 보이며, 얼굴과 뒤로 단단히 모은 검은 머리, 오염된 회백색 저고리가 안규홍 참고와 대체로 일치한다. 미간을 찌푸리고 이를 드러내 악문 표정이다. 허리띠에 낫이 있고 손가락에 녹색 가락지가 보인다. 대장장이와 쌀수레는 화면 밖이며 탄흔은 식별되지 않는다. 읽을 수 있는 문자는 없다.",
        "hard_violations": [],
        "physics": "머리와 상체는 자연스럽게 연결되어 있고 한 손은 허리 앞에서 주먹을 쥔 상태다. 하체의 접지는 프레임 밖이지만 부유를 시사하지 않는다. 낫은 허리띠에 끼워 지지되고 가락지는 손가락에 착용되어 있다. 지지 없이 떠 있는 물체나 불가능한 관절 자세는 보이지 않는다."
       }
      ],
      "all_candidates_fail": true
     },
     "normalized": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 2,
        "verdict_ko": "얼굴의 크기와 악문 이는 적절하지만, 제외해야 할 대장장이를 뒤쪽에 노출하고 시선은 그와 반대인 앞쪽 아래로 향해 핵심 배치를 위반한다."
       },
       {
        "label": "A",
        "score": 6,
        "verdict_ko": "화면 밖 아래를 보는 시선과 단독 인물 배치는 맞지만, 허리까지 담은 중간 크기 구도여서 얼굴 클로즈업과 제한된 숲 배경이라는 지시를 충족하지 못한다."
       }
      ],
      "all_candidates_fail": true,
      "readings": [
       {
        "label": "B",
        "direction": "안규홍은 고개를 숙여 자신의 앞쪽 아래를 본다. 대장장이는 안규홍의 뒤쪽인 화면 오른쪽 아래에 누워 있어, 보이는 시선이 그에게 닿지 않는다. 무기나 이동 동작은 보이지 않는다.",
        "built_space": "흙길과 수목, 오른쪽의 굵은 나무줄기가 보이며 인공 구조물이나 고정 설비는 없다. 안규홍의 얼굴과 상체가 전경을 차지하고 대장장이가 바로 뒤 지면에 놓여 있다. 숲길의 재질과 낮 조명은 참고와 유사하지만, 시신을 프레임 밖에 두라는 공간 배치를 어겼다.",
        "entities": "안규홍은 참고와 유사한 얼굴의 젊은 동아시아계 남성으로, 뒤로 모은 검은 머리와 때 묻은 회백색 저고리가 맞는다. 찌푸린 미간과 드러난 맞물린 이가 고통과 분노를 나타낸다. 청색 옷의 대장장이까지 보여 총 두 사람이 등장하며, 그의 상처와 혈흔은 등이 아니라 위로 드러난 가슴 쪽에 보인다. 낫, 옥가락지, 쌀수레와 탄흔은 확인되지 않으며 읽을 수 있는 문자는 없다.",
        "hard_violations": [
         "명시적으로 클로즈업 밖에 남겨야 하는 대장장이의 얼굴과 몸을 화면 안에 배치했다."
        ],
        "physics": "안규홍의 머리는 목과 몸통으로 지지되고 상체는 앞으로 숙여져 있다. 하체와 접지점은 프레임 밖이므로 지지 불능이라고 볼 근거는 없다. 대장장이의 머리와 몸통은 지면에 누워 있고, 왼쪽 아래에 보이는 손과 팔도 지면 가까이 놓여 있다. 떠 있는 신체나 물체는 보이지 않는다."
       },
       {
        "label": "A",
        "direction": "안규홍의 눈과 고개가 정면 아래의 화면 밖을 향한다. 대장장이는 보이지 않아 시선의 실제 도착점은 확인할 수 없지만, 바닥의 시신을 내려다보는 지시와 양립한다. 허리의 낫은 위로 굽은 채 수납되어 있으며 누구를 겨누거나 휘두르지 않는다.",
        "built_space": "안규홍이 중앙에 있고 왼쪽 뒤로 흙과 돌이 섞인 산길이 이어진다. 양옆에 여러 나무줄기와 관목이 있으며 인공 구조물이나 고정 설비는 없다. 숲길과 자연광은 장소 설정에 부합하지만, 허리까지 보여 배경 숲의 면적이 크고 지정된 클로즈업보다 넓다.",
        "entities": "한 명의 젊은 동아시아계 남성만 보이며, 얼굴과 뒤로 단단히 모은 검은 머리, 오염된 회백색 저고리가 안규홍 참고와 대체로 일치한다. 미간을 찌푸리고 이를 드러내 악문 표정이다. 허리띠에 낫이 있고 손가락에 녹색 가락지가 보인다. 대장장이와 쌀수레는 화면 밖이며 탄흔은 식별되지 않는다. 읽을 수 있는 문자는 없다.",
        "hard_violations": [],
        "physics": "머리와 상체는 자연스럽게 연결되어 있고 한 손은 허리 앞에서 주먹을 쥔 상태다. 하체의 접지는 프레임 밖이지만 부유를 시사하지 않는다. 낫은 허리띠에 끼워 지지되고 가락지는 손가락에 착용되어 있다. 지지 없이 떠 있는 물체나 불가능한 관절 자세는 보이지 않는다."
       }
      ]
     },
     "ok": true
    }
   ],
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "slot_winner_match": true,
   "slot_winner": {
    "gemini-pro": "A",
    "gpt-high": "A"
   },
   "route": "cross_slot_agree"
  },
  "dual": {
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "normalized": {
    "A": 2.0,
    "B": 0.533
   },
   "adjusted": {
    "A": 2.0,
    "B": 0.283
   },
   "violations": {
    "B": [
     "[gemini-pro] 화면 왼쪽 아래에 안규홍의 자세와 물리적으로 연결될 수 없는 위치에 짚신을 신은 발이 잘린 채 그려져 있어 해부학적으로 불가능함 (physically impossible anatomy)",
     "[gemini-pro] 프레임 밖으로 배제되어야 할 인물(대장장이 시체)이 화면에 포함됨 (person placed where the staging does not put them)",
     "[gpt-high] 명시적으로 클로즈업 밖에 남겨야 하는 대장장이의 얼굴과 몸을 화면 안에 배치했다."
    ]
   },
   "per_model_winner": {
    "gemini-pro": "A",
    "gpt-high": "A"
   },
   "agreed": true
  },
  "totals": {
   "A": 2000,
   "B": 283
  },
  "selected": "A",
  "ranking": [
   "A",
   "B"
  ],
  "verdicts": [
   {
    "label": "A",
    "score": 2000,
    "verdict_ko": "지시문이 요구한 클로즈업 앵글을 유지하여 시체를 프레임 밖으로 올바르게 배제했고, 요구된 소품(옥가락지, 허리 낫)과 이를 꽉 악문 참담한 표정을 매우 충실하게 구현했습니다."
   },
   {
    "label": "B",
    "score": 283,
    "verdict_ko": "시체를 프레임에서 배제하라는 명시적인 카메라 프레이밍 지시를 위반했으며, 요구된 소품이 누락되었고 해부학적으로 불가능한 신체 부위가 묘사되었습니다.  ★위반: [gemini-pro] 화면 왼쪽 아래에 안규홍의 자세와 물리적으로 연결될 수 없는 위치에 짚신을 신은 발이 잘린 채 그려져 있어 해부학적으로 불가능함 (physically impossible anatomy) / [gemini-pro] 프레임 밖으로 배제되어야 할 인물(대장장이 시체)이 화면에 포함됨 (person placed where the staging does not put them) / [gpt-high] 명시적으로 클로즈업 밖에 남겨야 하는 대장장이의 얼굴과 몸을 화면 안에 배치했다."
   }
  ],
  "refs": [
   {
    "label": "PREVIOUS SHOT STILL — a visually related earlier shot of this same place: the location's look, materials, fixed features and lighting mood are LOCKED to this photo; never copy its camera framing. The people visible in that photo are NOT in this shot: never carry their faces, bodies or clothing onto anyone here. Each person in THIS shot is defined solely by the PEOPLE list, their own CHARACTER REFERENCE images and their wardrobe notes. If this photo shows a character who cannot move (dead or unconscious), that character's exact pose, position and orientation are ALSO LOCKED — treat their body as a fixed prop of the set that only the camera moves around.",
    "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/db00517b-165b-4005-a287-f28936b59b95/images/272bd7af-9294-4490-98c6-88a40e9deacc/scene/recipe/S39sh18_sel.png",
    "asset_id": "e893c628-4b5e-4c17-b33a-3381a06b66c9",
    "role": "prev_still"
   },
   {
    "label": "CHARACTER REFERENCE — 안규홍: the exact person appearing in this shot; match face, hair and build exactly.",
    "path": "<bytes:1028384>",
    "asset_id": "3b5894db-ce55-4f34-b966-5e725c3aa952",
    "role": "character_ref"
   }
  ],
  "critique_skipped": true,
  "shot_run_produced": false,
  "shot_run_uid": "06aa1429-8a7f-724d-b653-03cc3da82bd0",
  "ref_mode": "prev+엔티티",
  "share_plan": {
   "ref_plan": "prev",
   "prev_anchor_tag": "S39sh18"
  }
 },
 "S39sh25::cine": {
  "applied": true,
  "attempted_at": "2026-09-09T05:24:07.509783+00:00",
  "fingerprint": "c1c82bbd90ef5ebe6568d240139079ab697c19da25fd79c6d95860f7dd97f5a8",
  "fingerprint_version": 2,
  "provider": "mai",
  "endpoint": "openrouter/chat-completions",
  "model": "microsoft/mai-image-2.6",
  "pack": "24.202608252115",
  "source_file": "S39sh25_sel.png",
  "source_sha256": "d8db69a5052b526022b14c77bcef90b2c400c8ff77b3fd5a626b2eaa490608e8",
  "file": "S39sh25_cine.png",
  "staged_sha256": "9fb56c793e34ddd48a1621c801c99ac562a44da7e4030302fa82889612318767",
  "latency_ms": 25042
 },
 "S40sh4::signage": {
  "fp": "97408393f45d58e8",
  "inscriptions": [],
  "cues": [],
  "dropped": []
 },
 "S40sh4::bgfirst_bg": {
  "input_fingerprint": "cae49bfeb4ce83e9",
  "prompt": "Create the EMPTY BACKGROUND PLATE for one film shot — NO PEOPLE, no figures, no body parts, no sketch lines, no arrows anywhere.\n\"Empty\" means no people only: KEEP the location's inherent occupants and stock that define the place — animals in an animal shelter, pen or farm, goods in a market, moored boats in a harbour — unless the shot text explicitly removes them.\nThe FIRST attached image is a thin-line storyboard sketch: use ONLY its camera angle, horizon, perspective and the placement/size of buildings and set masses — ignore the sketched people and arrows entirely. The SECOND attached image (LOCATION PHOTOGRAPH) is the real place: take its architecture, materials, signage and fixed features, and RE-PROJECT them into the sketch's camera. If the photograph's camera differs from the sketch's, the sketch's camera wins.\nHUMAN-SCALE CALIBRATION: derive every structure's true size from human-scale elements — a door ≈ 2m, a window ≈ 1–1.5m wide, one storey ≈ 2.5–3m; never inflate a small structure or shrink a large one.\n\nSHOT TEXT this background must serve (Korean): 박제현을 향해 턱을 치켜들고 거만한 눈빛으로 내려다보는 배상도의 측면.\n\nLOCATION (lock): The outdoor courtyard of a traditional estate, between the opened main gate and the guest wing in daylight.\n\nTIME OF DAY (lock): day.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: medium shot\n- KEY BACKGROUND ELEMENTS: 사랑채 (박제현 has just emerged from it) — An oblique portion of its courtyard-facing side remains behind 박제현; used as Anchors his position within his own household during the confrontation.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Use natural courtyard daylight with restrained saturation and unembellished contrast, keeping 배상도's fleeting discomfort visible beneath his asserted authority.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nRender ONE photorealistic empty location photograph, 16:9, that this shot can be staged inside later. No readable writing anywhere: surfaces that would carry writing may be present, but stage any wording out of legibility — an oblique angle, distance, shallow focus. No captions, watermarks or overlay text.",
  "effective_prompt": "Create the EMPTY BACKGROUND PLATE for one film shot — NO PEOPLE, no figures, no body parts, no sketch lines, no arrows anywhere.\n\"Empty\" means no people only: KEEP the location's inherent occupants and stock that define the place — animals in an animal shelter, pen or farm, goods in a market, moored boats in a harbour — unless the shot text explicitly removes them.\nThe FIRST attached image is a thin-line storyboard sketch: use ONLY its camera angle, horizon, perspective and the placement/size of buildings and set masses — ignore the sketched people and arrows entirely. The SECOND attached image (LOCATION PHOTOGRAPH) is the real place: take its architecture, materials, signage and fixed features, and RE-PROJECT them into the sketch's camera. If the photograph's camera differs from the sketch's, the sketch's camera wins.\nHUMAN-SCALE CALIBRATION: derive every structure's true size from human-scale elements — a door ≈ 2m, a window ≈ 1–1.5m wide, one storey ≈ 2.5–3m; never inflate a small structure or shrink a large one.\n\nSHOT TEXT this background must serve (Korean): 박제현을 향해 턱을 치켜들고 거만한 눈빛으로 내려다보는 배상도의 측면.\n\nLOCATION (lock): The outdoor courtyard of a traditional estate, between the opened main gate and the guest wing in daylight.\n\nTIME OF DAY (lock): day.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: medium shot\n- KEY BACKGROUND ELEMENTS: 사랑채 (박제현 has just emerged from it) — An oblique portion of its courtyard-facing side remains behind 박제현; used as Anchors his position within his own household during the confrontation.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Use natural courtyard daylight with restrained saturation and unembellished contrast, keeping 배상도's fleeting discomfort visible beneath his asserted authority.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nRender ONE photorealistic empty location photograph, 16:9, that this shot can be staged inside later. No readable writing anywhere: surfaces that would carry writing may be present, but stage any wording out of legibility — an oblique angle, distance, shallow focus. No captions, watermarks or overlay text.",
  "bg_path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/db00517b-165b-4005-a287-f28936b59b95/images/272bd7af-9294-4490-98c6-88a40e9deacc/scene/recipe/S40sh4__bgfirst_bg.png",
  "asset_id": "aa20f43d-9e57-410f-b956-dd882862f1bb",
  "input_asset_ids": [
   "0a7af2d1-5155-4265-b928-6ff03a1c5af0",
   "64556ed8-9e6e-4367-ade4-e6cfd6cae678"
  ]
 },
 "S40sh4": {
  "input_fingerprint": "1a6707cf3f6704fc",
  "prompt": "Create ONE FINAL photorealistic live-action film still of the moment below — early 20th-century Korea, Jeolla region; all people are Korean unless explicitly stated otherwise. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 박제현을 향해 턱을 치켜들고 거만한 눈빛으로 내려다보는 배상도의 측면.\n\nLOCATION (lock): The outdoor courtyard of a traditional estate, between the opened main gate and the guest wing in daylight. The shot takes place here — the attached LOCATION PHOTOGRAPH shows the exact spot.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: medium shot\n- KEY BACKGROUND ELEMENTS: 사랑채 (박제현 has just emerged from it) — An oblique portion of its courtyard-facing side remains behind 박제현; used as Anchors his position within his own household during the confrontation.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Use natural courtyard daylight with restrained saturation and unembellished contrast, keeping 배상도's fleeting discomfort visible beneath his asserted authority.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The main gate stands open after being forced wide. 배상도: He stands in the yard with his chin raised.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 배상도 (한국인 남성, 성인 나이대, 자연스러운 검은 머리) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.",
  "roll_prompts": {
   "A": "Stage the shot. The FIRST attached image (SHOT BACKGROUND) is the finished empty background of this shot — keep it EXACTLY: its camera, perspective, architecture, lighting and every fixed feature stay untouched. The SECOND attached image (LAYOUT SKETCH) tells you ONLY where the people go: each sketched person's position, screen size, pose and the gaze/motion arrows. Ignore the sketch's background lines. The CHARACTER REFERENCE photographs show the real people.\nPlace the real people into the background at exactly the sketched positions, sizes and poses, following the arrow directions. No sketch lines or arrows may remain.\n\nCreate ONE FINAL photorealistic live-action film still of the moment below — early 20th-century Korea, Jeolla region; all people are Korean unless explicitly stated otherwise. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 박제현을 향해 턱을 치켜들고 거만한 눈빛으로 내려다보는 배상도의 측면.\n\nLOCATION (lock): The outdoor courtyard of a traditional estate, between the opened main gate and the guest wing in daylight. The shot takes place here — the FIRST attached image (SHOT BACKGROUND) is this exact place, already built: its ground, structures, horizon, materials and lighting are the finished truth of this location and must not be redesigned or replaced. No location photograph is attached — read the place from that image alone, and add no scenery, structure, vehicle or fixture that it does not already show. This lock governs the place only; the figures in the shot follow the staging and pose instructions.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: medium shot\n- KEY BACKGROUND ELEMENTS: 사랑채 (박제현 has just emerged from it) — An oblique portion of its courtyard-facing side remains behind 박제현; used as Anchors his position within his own household during the confrontation.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Use natural courtyard daylight with restrained saturation and unembellished contrast, keeping 배상도's fleeting discomfort visible beneath his asserted authority.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The main gate stands open after being forced wide. 배상도: He stands in the yard with his chin raised.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 배상도 (한국인 남성, 성인 나이대, 자연스러운 검은 머리) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.",
   "B": "Create ONE FINAL photorealistic live-action film still of the moment below — early 20th-century Korea, Jeolla region; all people are Korean unless explicitly stated otherwise. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 박제현을 향해 턱을 치켜들고 거만한 눈빛으로 내려다보는 배상도의 측면.\n\nLOCATION (lock): The outdoor courtyard of a traditional estate, between the opened main gate and the guest wing in daylight. The shot takes place here — the attached LOCATION PHOTOGRAPH shows the exact spot.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: medium shot\n- KEY BACKGROUND ELEMENTS: 사랑채 (박제현 has just emerged from it) — An oblique portion of its courtyard-facing side remains behind 박제현; used as Anchors his position within his own household during the confrontation.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Use natural courtyard daylight with restrained saturation and unembellished contrast, keeping 배상도's fleeting discomfort visible beneath his asserted authority.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The main gate stands open after being forced wide. 배상도: He stands in the yard with his chin raised.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 배상도 (한국인 남성, 성인 나이대, 자연스러운 검은 머리) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.\n\nCOMPOSITION VARIATION (this candidate only): keep the place, the people and the moment exactly as specified, and keep from the CAMERA & FRAME contract its subject scale, screen placement and key background elements exactly. For this candidate only, its camera angle and height are deliberately left open: choose a camera position distinctly different from the obvious one, as a film director picking a second setup on the same blocking. Stay true to any explicit framing the SHOT TEXT itself dictates."
  },
  "roll_refs": {
   "A": [
    {
     "label": "SHOT BACKGROUND",
     "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/db00517b-165b-4005-a287-f28936b59b95/images/272bd7af-9294-4490-98c6-88a40e9deacc/scene/recipe/S40sh4__bgfirst_bg.png",
     "asset_id": "aa20f43d-9e57-410f-b956-dd882862f1bb",
     "role": "bgfirst_bg"
    },
    {
     "label": "LAYOUT SKETCH (people placement only)",
     "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/db00517b-165b-4005-a287-f28936b59b95/images/272bd7af-9294-4490-98c6-88a40e9deacc/conti/conti_S40sh4.png",
     "asset_id": "0a7af2d1-5155-4265-b928-6ff03a1c5af0",
     "role": "conti_light"
    },
    {
     "label": "CHARACTER REFERENCE — 배상도: the exact person appearing in this shot; match face, hair and build exactly.",
     "path": "<bytes:851718>",
     "asset_id": "034f5325-975e-4474-854b-e48359710bf8",
     "role": "character_ref"
    }
   ],
   "B": [
    {
     "label": "LOCATION PHOTOGRAPH — the exact place of this shot: its architecture, materials, fixed features and lighting mood are spatial truth; stage the moment inside this place. Never copy its camera framing.",
     "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/db00517b-165b-4005-a287-f28936b59b95/episodes/272bd7af-9294-4490-98c6-88a40e9deacc/images/background_chain/L08B01.png",
     "asset_id": "64556ed8-9e6e-4367-ade4-e6cfd6cae678",
     "role": "location_plate"
    },
    {
     "label": "CHARACTER REFERENCE — 배상도: the exact person appearing in this shot; match face, hair and build exactly.",
     "path": "<bytes:851718>",
     "asset_id": "034f5325-975e-4474-854b-e48359710bf8",
     "role": "character_ref"
    }
   ]
  },
  "shot_run_spend_attempt_count": 1,
  "judge_flip_skipped": "judge_fn_owns_order",
  "readings": [
   {
    "label": "A",
    "direction": "배상도는 화면 왼쪽의 추가된 인물을 향해 턱을 들고 있으나 눈을 감고 있음. 프레임 안의 다른 인물은 배상도를 마주보고 있음.",
    "built_space": "화면 왼쪽에 사랑채, 오른쪽에 열린 솟을대문이 위치하며, 두 인물이 마당의 올바른 위치에 서 있음.",
    "entities": "배상도의 복장은 참조 이미지와 일치하지만, 프롬프트의 등장 허용 목록에 없는 인물(박제현으로 추정)이 추가로 등장함.",
    "hard_violations": [
     "[gemini-pro] 허용되지 않은 인물 등장 (invented people)",
     "[gpt-high] 배상도만 허용한 등장인물 조건과 달리 왼쪽 전경에 별도의 남성을 추가했다.",
     "[gpt-high] 참조에 하나인 기와 대문을 앞뒤 두 개로 중복 생성하여 장소의 고정 구조를 바꾸었다.",
     "[gpt-high] 뒤쪽 대문 너머에 현대 차량으로 보이는 미지정 물체가 나타나 시대 설정을 훼손한다."
    ],
    "physics": "두 인물 모두 두 발로 마당을 안정적으로 디디고 서 있음."
   },
   {
    "label": "B",
    "direction": "배상도는 화면 왼쪽(프레임 밖의 대상)을 향해 턱을 치켜들고 거만한 시선으로 아래를 내려다보고 있음.",
    "built_space": "왼쪽에 사랑채, 오른쪽에 열린 대문이 배치되어 참조 사진의 공간 구도와 완벽하게 일치함.",
    "entities": "지정된 인물인 배상도 한 명만 등장하며, 참조 사진의 인상, 검은 모자, 조끼 등의 의상 디테일을 정확히 재현함.",
    "hard_violations": [
     "[gpt-high] 열린 대문 밖에 현대 자동차를 추가하여 20세기 초 조선의 시대 설정과 미지정 소품 금지 조건을 위반했다."
    ],
    "physics": "인물이 마당의 지면 위에 안정적인 자세로 서 있음."
   }
  ],
  "cross_model_order": {
   "policy": "cross_model_order_v2_slot0_fwd_slot1_rev_parallel2",
   "slots": [
    {
     "model": "gemini-pro",
     "order": "forward",
     "display_to_canonical": {
      "A": "A",
      "B": "B"
     },
     "raw": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 3,
        "verdict_ko": "프롬프트의 인물 목록에서 허용하지 않은 인물(박제현)을 프레임 안에 추가한 치명적인 규정 위반이 있으며, 배상도의 눈이 감겨 있어 지시된 거만한 시선이 연출되지 않았습니다."
       },
       {
        "label": "B",
        "score": 9,
        "verdict_ko": "허용된 인물(배상도)만 화면에 배치하여 오프스크린 대상을 향해 턱을 들고 내려다보는 측면 구도를 완벽히 구현했으며, 참조 이미지의 공간 및 의상 지침을 정확하게 준수했습니다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "배상도는 화면 왼쪽의 추가된 인물을 향해 턱을 들고 있으나 눈을 감고 있음. 프레임 안의 다른 인물은 배상도를 마주보고 있음.",
        "built_space": "화면 왼쪽에 사랑채, 오른쪽에 열린 솟을대문이 위치하며, 두 인물이 마당의 올바른 위치에 서 있음.",
        "entities": "배상도의 복장은 참조 이미지와 일치하지만, 프롬프트의 등장 허용 목록에 없는 인물(박제현으로 추정)이 추가로 등장함.",
        "hard_violations": [
         "허용되지 않은 인물 등장 (invented people)"
        ],
        "physics": "두 인물 모두 두 발로 마당을 안정적으로 디디고 서 있음."
       },
       {
        "label": "B",
        "direction": "배상도는 화면 왼쪽(프레임 밖의 대상)을 향해 턱을 치켜들고 거만한 시선으로 아래를 내려다보고 있음.",
        "built_space": "왼쪽에 사랑채, 오른쪽에 열린 대문이 배치되어 참조 사진의 공간 구도와 완벽하게 일치함.",
        "entities": "지정된 인물인 배상도 한 명만 등장하며, 참조 사진의 인상, 검은 모자, 조끼 등의 의상 디테일을 정확히 재현함.",
        "hard_violations": [],
        "physics": "인물이 마당의 지면 위에 안정적인 자세로 서 있음."
       }
      ],
      "all_candidates_fail": false
     },
     "normalized": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 3,
        "verdict_ko": "프롬프트의 인물 목록에서 허용하지 않은 인물(박제현)을 프레임 안에 추가한 치명적인 규정 위반이 있으며, 배상도의 눈이 감겨 있어 지시된 거만한 시선이 연출되지 않았습니다."
       },
       {
        "label": "B",
        "score": 9,
        "verdict_ko": "허용된 인물(배상도)만 화면에 배치하여 오프스크린 대상을 향해 턱을 들고 내려다보는 측면 구도를 완벽히 구현했으며, 참조 이미지의 공간 및 의상 지침을 정확하게 준수했습니다."
       }
      ],
      "all_candidates_fail": false,
      "readings": [
       {
        "label": "A",
        "direction": "배상도는 화면 왼쪽의 추가된 인물을 향해 턱을 들고 있으나 눈을 감고 있음. 프레임 안의 다른 인물은 배상도를 마주보고 있음.",
        "built_space": "화면 왼쪽에 사랑채, 오른쪽에 열린 솟을대문이 위치하며, 두 인물이 마당의 올바른 위치에 서 있음.",
        "entities": "배상도의 복장은 참조 이미지와 일치하지만, 프롬프트의 등장 허용 목록에 없는 인물(박제현으로 추정)이 추가로 등장함.",
        "hard_violations": [
         "허용되지 않은 인물 등장 (invented people)"
        ],
        "physics": "두 인물 모두 두 발로 마당을 안정적으로 디디고 서 있음."
       },
       {
        "label": "B",
        "direction": "배상도는 화면 왼쪽(프레임 밖의 대상)을 향해 턱을 치켜들고 거만한 시선으로 아래를 내려다보고 있음.",
        "built_space": "왼쪽에 사랑채, 오른쪽에 열린 대문이 배치되어 참조 사진의 공간 구도와 완벽하게 일치함.",
        "entities": "지정된 인물인 배상도 한 명만 등장하며, 참조 사진의 인상, 검은 모자, 조끼 등의 의상 디테일을 정확히 재현함.",
        "hard_violations": [],
        "physics": "인물이 마당의 지면 위에 안정적인 자세로 서 있음."
       }
      ]
     },
     "ok": true
    },
    {
     "model": "gpt-high",
     "order": "reverse",
     "display_to_canonical": {
      "A": "B",
      "B": "A"
     },
     "raw": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 3,
        "verdict_ko": "배상도 한 명의 측면 미디엄 숏과 장소 재현은 더 충실하지만, 대문 밖 현대 자동차가 시대·소품 조건을 깨고 턱을 치켜든 거만한 연기도 약하다."
       },
       {
        "label": "B",
        "score": 1,
        "verdict_ko": "턱을 드는 동작은 뚜렷하지만 허용되지 않은 두 번째 인물과 중복 대문을 추가했고, 시선도 상대의 얼굴보다 위로 향한다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "배상도는 왼쪽 사랑채 방향으로 얼굴과 눈을 향한다. 턱은 약간 들렸지만 눈길은 거의 수평이어서 거만하게 내려다보는 동작이 선명하지 않다. 박제현은 화면 밖에 있으므로 실제로 그 얼굴을 겨냥하는지는 확인할 수 없다. 무기나 이동 중인 인물은 없다.",
        "built_space": "왼쪽에 사랑채 한 동의 비스듬한 마당 쪽 면, 열린 출입구 두 곳, 격자창 한 곳, 길게 이어진 툇마루가 보인다. 오른쪽에는 열린 양짝 대문 한 곳과 돌담이 있고, 사랑채 앞과 대문 옆에 돌 디딤돌이 각각 보인다. 왼쪽 아래에는 평상 일부가 걸린다. 배상도는 사랑채와 대문 사이 마당에 서 있어 기본 배치와 목재·회벽·기와 재료는 참조에 가깝다. 불가능한 반사는 없다.",
        "entities": "보이는 사람은 배상도 한 명이며, 성인 한국인 남성으로 설정된 참조의 외형과 대체로 부합한다. 검은 머리, 검은 중절모와 조끼, 흰 전통 의복이 일치한다. 측면이라 얼굴의 정확한 동일성 확인에는 한계가 있다. 허리 장식과 신발은 크롭 때문에 판정하지 않는다. 대문 밖에는 참조와 지문에 없는 현대 자동차가 보인다. 읽을 수 있는 글자는 보이지 않는다.",
        "hard_violations": [
         "열린 대문 밖에 현대 자동차를 추가하여 20세기 초 조선의 시대 설정과 미지정 소품 금지 조건을 위반했다."
        ],
        "physics": "상체는 곧게 서 있고 목과 턱의 자세는 신체적으로 가능하다. 발은 화면 밖이지만 마당에 서 있는 상체 배치에 부유를 의심할 근거는 없다. 모자는 머리에 얹혀 있고 옷은 어깨와 몸에 지지되어 자연스럽게 드리워진다. 툇마루와 평상에는 받침이 보인다."
       },
       {
        "label": "B",
        "direction": "오른쪽 배상도는 턱을 확실히 치켜들고 왼쪽을 향하지만, 눈길은 왼쪽 남자의 얼굴보다 위쪽으로 지나가는 듯하다. 왼쪽 남자는 배상도 쪽으로 고개를 돌린다. 서로 대치하는 배치는 있으나 배상도가 상대를 내려다보는 정확한 시선 연결은 약하다. 무기나 이동 동작은 없다.",
        "built_space": "왼쪽 사랑채에는 열린 출입구 두 곳과 가장자리 격자창, 툇마루가 보이고, 전경에 평상 한 개가 있다. 중앙에는 가까운 기와 대문 뒤로 또 다른 기와 대문이 겹쳐 나타나며 돌담도 두 겹으로 나뉜다. 참조의 단일 대문 구조와 다르다. 사랑채 앞에는 툇마루 아래 별도의 긴 목제 단도 보인다. 배상도는 오른쪽 마당에, 다른 남자는 왼쪽 사랑채 앞에 서 있다. 불가능한 반사는 보이지 않는다.",
        "entities": "배상도는 검은 중절모·조끼와 흰 의복을 입은 검은 머리의 성인 남성으로, 참조 인물의 외형에 대체로 가깝다. 왼쪽에는 상투와 머리띠, 밝은색 전통 의복을 갖춘 별도의 성인 남성이 추가되었다. 박제현을 나타내려는 배치로 보이지만, 등장 허용 인물을 배상도로 제한한 조건에는 맞지 않는다. 뒤쪽 대문 너머에는 현대 차량의 일부로 보이는 물체도 있다. 판독 가능한 글자는 없다.",
        "hard_violations": [
         "배상도만 허용한 등장인물 조건과 달리 왼쪽 전경에 별도의 남성을 추가했다.",
         "참조에 하나인 기와 대문을 앞뒤 두 개로 중복 생성하여 장소의 고정 구조를 바꾸었다.",
         "뒤쪽 대문 너머에 현대 차량으로 보이는 미지정 물체가 나타나 시대 설정을 훼손한다."
        ],
        "physics": "두 사람 모두 발은 크롭 밖이지만 몸통은 마당에 서 있는 자세로 이어지며, 공중에 뜬 징후는 없다. 배상도의 젖힌 목과 뒤로 둔 팔은 가능한 자세다. 모자는 머리가 지지하고 옷은 몸을 따라 늘어진다. 평상과 목제 단에는 다리가 있으며, 대문 지붕도 기둥 위에 놓여 있다. 구조 중복은 장소 위반이지 무지지 부유 현상은 아니다."
       }
      ],
      "all_candidates_fail": true
     },
     "normalized": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 3,
        "verdict_ko": "배상도 한 명의 측면 미디엄 숏과 장소 재현은 더 충실하지만, 대문 밖 현대 자동차가 시대·소품 조건을 깨고 턱을 치켜든 거만한 연기도 약하다."
       },
       {
        "label": "A",
        "score": 1,
        "verdict_ko": "턱을 드는 동작은 뚜렷하지만 허용되지 않은 두 번째 인물과 중복 대문을 추가했고, 시선도 상대의 얼굴보다 위로 향한다."
       }
      ],
      "all_candidates_fail": true,
      "readings": [
       {
        "label": "B",
        "direction": "배상도는 왼쪽 사랑채 방향으로 얼굴과 눈을 향한다. 턱은 약간 들렸지만 눈길은 거의 수평이어서 거만하게 내려다보는 동작이 선명하지 않다. 박제현은 화면 밖에 있으므로 실제로 그 얼굴을 겨냥하는지는 확인할 수 없다. 무기나 이동 중인 인물은 없다.",
        "built_space": "왼쪽에 사랑채 한 동의 비스듬한 마당 쪽 면, 열린 출입구 두 곳, 격자창 한 곳, 길게 이어진 툇마루가 보인다. 오른쪽에는 열린 양짝 대문 한 곳과 돌담이 있고, 사랑채 앞과 대문 옆에 돌 디딤돌이 각각 보인다. 왼쪽 아래에는 평상 일부가 걸린다. 배상도는 사랑채와 대문 사이 마당에 서 있어 기본 배치와 목재·회벽·기와 재료는 참조에 가깝다. 불가능한 반사는 없다.",
        "entities": "보이는 사람은 배상도 한 명이며, 성인 한국인 남성으로 설정된 참조의 외형과 대체로 부합한다. 검은 머리, 검은 중절모와 조끼, 흰 전통 의복이 일치한다. 측면이라 얼굴의 정확한 동일성 확인에는 한계가 있다. 허리 장식과 신발은 크롭 때문에 판정하지 않는다. 대문 밖에는 참조와 지문에 없는 현대 자동차가 보인다. 읽을 수 있는 글자는 보이지 않는다.",
        "hard_violations": [
         "열린 대문 밖에 현대 자동차를 추가하여 20세기 초 조선의 시대 설정과 미지정 소품 금지 조건을 위반했다."
        ],
        "physics": "상체는 곧게 서 있고 목과 턱의 자세는 신체적으로 가능하다. 발은 화면 밖이지만 마당에 서 있는 상체 배치에 부유를 의심할 근거는 없다. 모자는 머리에 얹혀 있고 옷은 어깨와 몸에 지지되어 자연스럽게 드리워진다. 툇마루와 평상에는 받침이 보인다."
       },
       {
        "label": "A",
        "direction": "오른쪽 배상도는 턱을 확실히 치켜들고 왼쪽을 향하지만, 눈길은 왼쪽 남자의 얼굴보다 위쪽으로 지나가는 듯하다. 왼쪽 남자는 배상도 쪽으로 고개를 돌린다. 서로 대치하는 배치는 있으나 배상도가 상대를 내려다보는 정확한 시선 연결은 약하다. 무기나 이동 동작은 없다.",
        "built_space": "왼쪽 사랑채에는 열린 출입구 두 곳과 가장자리 격자창, 툇마루가 보이고, 전경에 평상 한 개가 있다. 중앙에는 가까운 기와 대문 뒤로 또 다른 기와 대문이 겹쳐 나타나며 돌담도 두 겹으로 나뉜다. 참조의 단일 대문 구조와 다르다. 사랑채 앞에는 툇마루 아래 별도의 긴 목제 단도 보인다. 배상도는 오른쪽 마당에, 다른 남자는 왼쪽 사랑채 앞에 서 있다. 불가능한 반사는 보이지 않는다.",
        "entities": "배상도는 검은 중절모·조끼와 흰 의복을 입은 검은 머리의 성인 남성으로, 참조 인물의 외형에 대체로 가깝다. 왼쪽에는 상투와 머리띠, 밝은색 전통 의복을 갖춘 별도의 성인 남성이 추가되었다. 박제현을 나타내려는 배치로 보이지만, 등장 허용 인물을 배상도로 제한한 조건에는 맞지 않는다. 뒤쪽 대문 너머에는 현대 차량의 일부로 보이는 물체도 있다. 판독 가능한 글자는 없다.",
        "hard_violations": [
         "배상도만 허용한 등장인물 조건과 달리 왼쪽 전경에 별도의 남성을 추가했다.",
         "참조에 하나인 기와 대문을 앞뒤 두 개로 중복 생성하여 장소의 고정 구조를 바꾸었다.",
         "뒤쪽 대문 너머에 현대 차량으로 보이는 미지정 물체가 나타나 시대 설정을 훼손한다."
        ],
        "physics": "두 사람 모두 발은 크롭 밖이지만 몸통은 마당에 서 있는 자세로 이어지며, 공중에 뜬 징후는 없다. 배상도의 젖힌 목과 뒤로 둔 팔은 가능한 자세다. 모자는 머리가 지지하고 옷은 몸을 따라 늘어진다. 평상과 목제 단에는 다리가 있으며, 대문 지붕도 기둥 위에 놓여 있다. 구조 중복은 장소 위반이지 무지지 부유 현상은 아니다."
       }
      ]
     },
     "ok": true
    }
   ],
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "slot_winner_match": true,
   "slot_winner": {
    "gemini-pro": "B",
    "gpt-high": "B"
   },
   "route": "cross_slot_agree"
  },
  "dual": {
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "normalized": {
    "A": 0.667,
    "B": 2.0
   },
   "adjusted": {
    "A": 0.417,
    "B": 1.75
   },
   "violations": {
    "A": [
     "[gemini-pro] 허용되지 않은 인물 등장 (invented people)",
     "[gpt-high] 배상도만 허용한 등장인물 조건과 달리 왼쪽 전경에 별도의 남성을 추가했다.",
     "[gpt-high] 참조에 하나인 기와 대문을 앞뒤 두 개로 중복 생성하여 장소의 고정 구조를 바꾸었다.",
     "[gpt-high] 뒤쪽 대문 너머에 현대 차량으로 보이는 미지정 물체가 나타나 시대 설정을 훼손한다."
    ],
    "B": [
     "[gpt-high] 열린 대문 밖에 현대 자동차를 추가하여 20세기 초 조선의 시대 설정과 미지정 소품 금지 조건을 위반했다."
    ]
   },
   "per_model_winner": {
    "gemini-pro": "B",
    "gpt-high": "B"
   },
   "agreed": true
  },
  "totals": {
   "A": 417,
   "B": 1750
  },
  "selected": "B",
  "ranking": [
   "B",
   "A"
  ],
  "verdicts": [
   {
    "label": "A",
    "score": 417,
    "verdict_ko": "프롬프트의 인물 목록에서 허용하지 않은 인물(박제현)을 프레임 안에 추가한 치명적인 규정 위반이 있으며, 배상도의 눈이 감겨 있어 지시된 거만한 시선이 연출되지 않았습니다.  ★위반: [gemini-pro] 허용되지 않은 인물 등장 (invented people) / [gpt-high] 배상도만 허용한 등장인물 조건과 달리 왼쪽 전경에 별도의 남성을 추가했다. / [gpt-high] 참조에 하나인 기와 대문을 앞뒤 두 개로 중복 생성하여 장소의 고정 구조를 바꾸었다. / [gpt-high] 뒤쪽 대문 너머에 현대 차량으로 보이는 미지정 물체가 나타나 시대 설정을 훼손한다."
   },
   {
    "label": "B",
    "score": 1750,
    "verdict_ko": "허용된 인물(배상도)만 화면에 배치하여 오프스크린 대상을 향해 턱을 들고 내려다보는 측면 구도를 완벽히 구현했으며, 참조 이미지의 공간 및 의상 지침을 정확하게 준수했습니다.  ★위반: [gpt-high] 열린 대문 밖에 현대 자동차를 추가하여 20세기 초 조선의 시대 설정과 미지정 소품 금지 조건을 위반했다."
   }
  ],
  "refs": [
   {
    "label": "LOCATION PHOTOGRAPH — the exact place of this shot: its architecture, materials, fixed features and lighting mood are spatial truth; stage the moment inside this place. Never copy its camera framing.",
    "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/db00517b-165b-4005-a287-f28936b59b95/episodes/272bd7af-9294-4490-98c6-88a40e9deacc/images/background_chain/L08B01.png",
    "asset_id": "64556ed8-9e6e-4367-ade4-e6cfd6cae678",
    "role": "location_plate"
   },
   {
    "label": "CHARACTER REFERENCE — 배상도: the exact person appearing in this shot; match face, hair and build exactly.",
    "path": "<bytes:851718>",
    "asset_id": "034f5325-975e-4474-854b-e48359710bf8",
    "role": "character_ref"
   }
  ],
  "critique_skipped": true,
  "shot_run_produced": false,
  "shot_run_uid": "06aa1429-8c23-7a7d-b54a-333a5f945287",
  "bgfirst": {
   "bg_path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/db00517b-165b-4005-a287-f28936b59b95/images/272bd7af-9294-4490-98c6-88a40e9deacc/scene/recipe/S40sh4__bgfirst_bg.png",
   "bg_asset_id": "aa20f43d-9e57-410f-b956-dd882862f1bb",
   "bg_record_key": "S40sh4::bgfirst_bg",
   "chain_winner": false,
   "authority": "plate"
  },
  "ref_mode": "플레이트+엔티티 (2택1: 무콘티 승)",
  "share_plan": {
   "ref_plan": "background"
  }
 },
 "S40sh4::cine": {
  "applied": false,
  "attempted_at": "2026-09-09T09:08:50.694387+00:00",
  "fingerprint": "7569d36cca204d45e251052442d5e349ed2ca396ff217d13b2dfcc2373c53f00",
  "fingerprint_version": 2,
  "provider": "mai",
  "endpoint": "openrouter/chat-completions",
  "model": "microsoft/mai-image-2.6",
  "pack": "24.202608252115",
  "source_file": "S40sh4_sel.png",
  "source_sha256": "4a1b454883eeb00dee59e0aac98e014d7b2ed63d7ca9f4710fac759499d5f679",
  "error": "RuntimeError: moderation blocked: {\"message\": \"Response content blocked by label 'DallEBlockList'.\", \"code\": 400}",
  "moderation_refusals": 2,
  "declined": true,
  "declined_reason": "moderation"
 },
 "S40sh12::signage": {
  "fp": "8968c03fac29269f",
  "inscriptions": [],
  "cues": [],
  "dropped": []
 },
 "S40sh12": {
  "input_fingerprint": "018cb68263e1db15",
  "prompt": "Create ONE FINAL photorealistic live-action film still of the moment below — early 20th-century Korea, Jeolla region; all people are Korean unless explicitly stated otherwise. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 입술을 꾹 깨문 채 박진성의 팔을 꽉 부여잡고 시선을 다른 곳으로 향한 박상현의 굳은 정면.\n\nLOCATION (lock): The inner main-gate area of the estate's courtyard, beside the arrest party's exit route. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: medium shot\n- KEY BACKGROUND ELEMENTS: 마당 (Visible in a narrow interval around the two men); used as Leaves the restraining contact unobstructed and maintains the shared courtyard setting.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Maintain the courtyard's natural daytime illumination and muted color, with sufficient tonal separation to read both the bitten lip and the gripping hands.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The people visible in that photo are NOT in this shot: never carry their faces, bodies or clothing onto anyone here. Each person in THIS shot is defined solely by the PEOPLE list, their own CHARACTER REFERENCE images and their wardrobe notes. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The main gate remains wide open during the arrest. 박상현: He stands with one hand locked in a firm restraining grip, fingers tightening. 박진성: He stands checked mid-advance with his arm held.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 박진성 (한국인 남성, 성인 나이대, 자연스러운 검은 머리); 박상현 (한국인 남성, 성인 나이대, 자연스러운 검은 머리) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.",
  "roll_prompts": {
   "A": "Create ONE FINAL photorealistic live-action film still of the moment below — early 20th-century Korea, Jeolla region; all people are Korean unless explicitly stated otherwise. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 입술을 꾹 깨문 채 박진성의 팔을 꽉 부여잡고 시선을 다른 곳으로 향한 박상현의 굳은 정면.\n\nLOCATION (lock): The inner main-gate area of the estate's courtyard, beside the arrest party's exit route. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: medium shot\n- KEY BACKGROUND ELEMENTS: 마당 (Visible in a narrow interval around the two men); used as Leaves the restraining contact unobstructed and maintains the shared courtyard setting.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Maintain the courtyard's natural daytime illumination and muted color, with sufficient tonal separation to read both the bitten lip and the gripping hands.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The people visible in that photo are NOT in this shot: never carry their faces, bodies or clothing onto anyone here. Each person in THIS shot is defined solely by the PEOPLE list, their own CHARACTER REFERENCE images and their wardrobe notes. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The main gate remains wide open during the arrest. 박상현: He stands with one hand locked in a firm restraining grip, fingers tightening. 박진성: He stands checked mid-advance with his arm held.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 박진성 (한국인 남성, 성인 나이대, 자연스러운 검은 머리); 박상현 (한국인 남성, 성인 나이대, 자연스러운 검은 머리) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.",
   "B": "Create ONE FINAL photorealistic live-action film still of the moment below — early 20th-century Korea, Jeolla region; all people are Korean unless explicitly stated otherwise. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 입술을 꾹 깨문 채 박진성의 팔을 꽉 부여잡고 시선을 다른 곳으로 향한 박상현의 굳은 정면.\n\nLOCATION (lock): The inner main-gate area of the estate's courtyard, beside the arrest party's exit route. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: medium shot\n- KEY BACKGROUND ELEMENTS: 마당 (Visible in a narrow interval around the two men); used as Leaves the restraining contact unobstructed and maintains the shared courtyard setting.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Maintain the courtyard's natural daytime illumination and muted color, with sufficient tonal separation to read both the bitten lip and the gripping hands.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The people visible in that photo are NOT in this shot: never carry their faces, bodies or clothing onto anyone here. Each person in THIS shot is defined solely by the PEOPLE list, their own CHARACTER REFERENCE images and their wardrobe notes. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The main gate remains wide open during the arrest. 박상현: He stands with one hand locked in a firm restraining grip, fingers tightening. 박진성: He stands checked mid-advance with his arm held.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 박진성 (한국인 남성, 성인 나이대, 자연스러운 검은 머리); 박상현 (한국인 남성, 성인 나이대, 자연스러운 검은 머리) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.\n\nCOMPOSITION VARIATION (this candidate only): keep the place, the people and the moment exactly as specified, and keep from the CAMERA & FRAME contract its subject scale, screen placement and key background elements exactly. For this candidate only, its camera angle and height are deliberately left open: choose a camera position distinctly different from the obvious one, as a film director picking a second setup on the same blocking. Stay true to any explicit framing the SHOT TEXT itself dictates."
  },
  "shot_run_spend_attempt_count": 1,
  "readings": [
   {
    "label": "A",
    "direction": "박상현(회색)은 화면 오른쪽 바깥을 향해 시선을 두고 있어 '다른 곳으로 향한' 지문을 정확히 충족합니다.",
    "built_space": "왼쪽의 대청마루가 있는 한옥과 오른쪽의 열린 대문 등 기준 사진의 마당 배경을 올바르게 배치했습니다.",
    "entities": "박상현과 박진성의 복장 색상은 맞으나, 기준 이미지와 달리 갓을 쓰지 않고 현대적인 머리 모양을 하고 있습니다.",
    "hard_violations": [
     "[gemini-pro] 해부학적으로 불가능한 신체 구조 (박상현의 오른쪽 소매 하나에서 두 개의 손이 나와 푸른 소매를 잡고 있으며, 소매 안쪽에 또 다른 손이 존재해 한쪽 팔에 3개의 손이 엉켜 있음)"
    ],
    "physics": "두 인물이 서 있는 기본 자세는 지면의 지지를 받으나, 팔을 잡고 있는 손들의 구조가 물리적으로 성립하지 않습니다."
   },
   {
    "label": "B",
    "direction": "박상현(회색)은 다른 곳이 아닌 박진성의 얼굴 쪽을 바라보고 있어 시선 처리 지문과 어긋납니다.",
    "built_space": "왼쪽의 한옥과 오른쪽의 대문 등 주어진 마당 공간의 구조와 위치를 기준 사진과 동일하게 구현했습니다.",
    "entities": "두 인물 모두 기준 사진의 인상착의를 따르고 있으나, 갓을 벗고 탕건만 착용한 상태입니다.",
    "hard_violations": [],
    "physics": "박상현이 두 발로 서서 왼손으로 박진성의 팔을 꽉 잡고 있는 자세가 물리적으로 자연스럽게 지탱되고 있습니다."
   }
  ],
  "cross_model_order": {
   "policy": "cross_model_order_v2_slot0_fwd_slot1_rev_parallel2",
   "slots": [
    {
     "model": "gemini-pro",
     "order": "forward",
     "display_to_canonical": {
      "A": "A",
      "B": "B"
     },
     "raw": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 0,
        "verdict_ko": "정면 구도와 시선 처리 등 지문 연출은 훌륭하나, 박상현의 한 소매에서 3개의 손이 나오는 치명적인 해부학적 오류가 발생해 탈락입니다."
       },
       {
        "label": "B",
        "score": 5,
        "verdict_ko": "해부학적 오류 없이 배경과 인물 간의 신체 접촉을 안정적으로 구현했으나, '정면 구도'와 '다른 곳을 향한 시선'이라는 핵심 연출을 놓쳤습니다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "박상현(회색)은 화면 오른쪽 바깥을 향해 시선을 두고 있어 '다른 곳으로 향한' 지문을 정확히 충족합니다.",
        "built_space": "왼쪽의 대청마루가 있는 한옥과 오른쪽의 열린 대문 등 기준 사진의 마당 배경을 올바르게 배치했습니다.",
        "entities": "박상현과 박진성의 복장 색상은 맞으나, 기준 이미지와 달리 갓을 쓰지 않고 현대적인 머리 모양을 하고 있습니다.",
        "hard_violations": [
         "해부학적으로 불가능한 신체 구조 (박상현의 오른쪽 소매 하나에서 두 개의 손이 나와 푸른 소매를 잡고 있으며, 소매 안쪽에 또 다른 손이 존재해 한쪽 팔에 3개의 손이 엉켜 있음)"
        ],
        "physics": "두 인물이 서 있는 기본 자세는 지면의 지지를 받으나, 팔을 잡고 있는 손들의 구조가 물리적으로 성립하지 않습니다."
       },
       {
        "label": "B",
        "direction": "박상현(회색)은 다른 곳이 아닌 박진성의 얼굴 쪽을 바라보고 있어 시선 처리 지문과 어긋납니다.",
        "built_space": "왼쪽의 한옥과 오른쪽의 대문 등 주어진 마당 공간의 구조와 위치를 기준 사진과 동일하게 구현했습니다.",
        "entities": "두 인물 모두 기준 사진의 인상착의를 따르고 있으나, 갓을 벗고 탕건만 착용한 상태입니다.",
        "hard_violations": [],
        "physics": "박상현이 두 발로 서서 왼손으로 박진성의 팔을 꽉 잡고 있는 자세가 물리적으로 자연스럽게 지탱되고 있습니다."
       }
      ],
      "all_candidates_fail": false
     },
     "normalized": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 0,
        "verdict_ko": "정면 구도와 시선 처리 등 지문 연출은 훌륭하나, 박상현의 한 소매에서 3개의 손이 나오는 치명적인 해부학적 오류가 발생해 탈락입니다."
       },
       {
        "label": "B",
        "score": 5,
        "verdict_ko": "해부학적 오류 없이 배경과 인물 간의 신체 접촉을 안정적으로 구현했으나, '정면 구도'와 '다른 곳을 향한 시선'이라는 핵심 연출을 놓쳤습니다."
       }
      ],
      "all_candidates_fail": false,
      "readings": [
       {
        "label": "A",
        "direction": "박상현(회색)은 화면 오른쪽 바깥을 향해 시선을 두고 있어 '다른 곳으로 향한' 지문을 정확히 충족합니다.",
        "built_space": "왼쪽의 대청마루가 있는 한옥과 오른쪽의 열린 대문 등 기준 사진의 마당 배경을 올바르게 배치했습니다.",
        "entities": "박상현과 박진성의 복장 색상은 맞으나, 기준 이미지와 달리 갓을 쓰지 않고 현대적인 머리 모양을 하고 있습니다.",
        "hard_violations": [
         "해부학적으로 불가능한 신체 구조 (박상현의 오른쪽 소매 하나에서 두 개의 손이 나와 푸른 소매를 잡고 있으며, 소매 안쪽에 또 다른 손이 존재해 한쪽 팔에 3개의 손이 엉켜 있음)"
        ],
        "physics": "두 인물이 서 있는 기본 자세는 지면의 지지를 받으나, 팔을 잡고 있는 손들의 구조가 물리적으로 성립하지 않습니다."
       },
       {
        "label": "B",
        "direction": "박상현(회색)은 다른 곳이 아닌 박진성의 얼굴 쪽을 바라보고 있어 시선 처리 지문과 어긋납니다.",
        "built_space": "왼쪽의 한옥과 오른쪽의 대문 등 주어진 마당 공간의 구조와 위치를 기준 사진과 동일하게 구현했습니다.",
        "entities": "두 인물 모두 기준 사진의 인상착의를 따르고 있으나, 갓을 벗고 탕건만 착용한 상태입니다.",
        "hard_violations": [],
        "physics": "박상현이 두 발로 서서 왼손으로 박진성의 팔을 꽉 잡고 있는 자세가 물리적으로 자연스럽게 지탱되고 있습니다."
       }
      ]
     },
     "ok": true
    },
    {
     "model": "gpt-high",
     "order": "reverse",
     "display_to_canonical": {
      "A": "B",
      "B": "A"
     },
     "raw": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 5,
        "verdict_ko": "입술을 깨물고 팔을 붙잡는 동작은 보이지만, 박상현이 다른 곳이 아니라 박진성을 바라보고 두 손으로 제지하며 회색 의상도 참조와 다릅니다."
       },
       {
        "label": "B",
        "score": 8,
        "verdict_ko": "박상현의 굳은 정면에 가까운 얼굴, 옆으로 돌린 시선, 깨문 입술과 한 손의 제지 접촉을 가장 잘 구현했으나 두 사람의 갓과 머리 모양은 참조에서 벗어납니다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "왼쪽 박상현은 고개와 눈을 오른쪽 박진성의 얼굴 쪽으로 향한다. 박진성도 뒤돌아 박상현을 바라보므로 서로 마주 보는 관계이며, 박상현이 다른 곳으로 시선을 피한다는 지시와 다르다. 박진성의 몸통은 대문 쪽을 향하지만 실제 전진 움직임은 뚜렷하지 않다.",
        "built_space": "왼쪽에 목조 건물 한 동의 툇마루, 어두운 출입구 하나와 격자창 하나가 보이고, 아래에 돌 디딤돌 하나가 있다. 오른쪽 뒤에는 기와지붕 대문 하나와 이어진 담장이 있으며 문짝 사이로 바깥이 보여 개방 상태가 읽힌다. 두 사람은 건물과 대문 사이 흙마당에 서 있다. 참조의 재료와 배치는 대체로 이어지지만 인물 주변 마당의 노출이 비교적 넓다. 반사면이나 중복된 대문은 없다.",
        "entities": "등장인물은 성인 한국인 남성으로 보이는 두 명뿐이다. 왼쪽의 짧은 수염과 얼굴은 박상현, 오른쪽의 수염 없는 얼굴과 연한 청색 포는 박진성에 대응한다. 두 사람 모두 검은 머리이지만 참조의 갓을 쓰지 않고 상투와 망건만 보인다. 박상현은 참조의 회색 긴소매 포 대신 흰 소매 위에 회색 겉옷을 입었다. 박진성의 청색 허리끈은 유지되지만 참조의 갓끈 장식은 없다. 읽을 수 있는 글자는 없다.",
        "hard_violations": [],
        "physics": "박상현의 두 손이 박진성의 위팔과 팔꿈치 부근 소매를 각각 잡고 있어 접촉과 당김은 물리적으로 가능하다. 다만 한 손을 단단히 잠가 제지한다는 지정과 달리 두 손으로 붙잡는다. 박진성은 몸통을 돌리고 잡힌 팔을 아래로 내린 자세로, 제지당한 상태로 가능하다. 발은 화면 밖이지만 두 몸은 정상적인 직립 자세이며 공중 부양이나 지지 없는 물체는 보이지 않는다."
       },
       {
        "label": "B",
        "direction": "오른쪽 박상현은 얼굴을 거의 정면으로 드러낸 채 눈을 화면 오른쪽 바깥으로 돌려 왼쪽 박진성과 시선을 맞추지 않는다. 화면 밖의 구체적인 대상은 보이지 않는다. 박진성도 오른쪽을 바라보며 몸과 잡힌 팔을 그쪽으로 내밀어, 나아가려다 제지된 관계가 읽힌다.",
        "built_space": "왼쪽 가장자리에 목조 건물의 격자창 하나와 툇마루 일부, 뒤에는 기와지붕 대문 하나와 양옆 돌담이 보인다. 두 사람은 대문 안쪽 흙마당에 나란히 가까이 서 있고, 몸 사이와 가장자리로 마당이 드러난다. 참조의 목재·기와·돌담과 낮의 조명은 대체로 유지된다. 대문 통로 대부분이 박상현에게 가려져 활짝 열린 상태인지는 확정하기 어렵다. 중복된 시설이나 불가능한 반사는 없다.",
        "entities": "성인 한국인 남성으로 보이는 두 명만 등장한다. 오른쪽 박상현의 짧은 수염, 얼굴과 회색 포, 왼쪽 박진성의 수염 없는 얼굴과 청색 포·허리끈은 참조의 구분에 부합한다. 그러나 둘 다 갓이 없고 짧게 정돈한 검은 머리를 드러내며, 박진성의 장식 갓끈도 없다. 박상현의 입술은 안쪽으로 눌러 깨문 듯 보이고 미간과 턱에 긴장이 있다. 읽을 수 있는 글자는 없다.",
        "hard_violations": [],
        "physics": "박상현의 한 손이 박진성의 팔꿈치 가까운 아래팔 소매를 감싸 쥐며 천을 눌러, 단단한 제지 접촉이 보인다. 잡힌 팔은 팔꿈치에서 굽어 몸 앞을 가로지르고 손은 주먹에 가깝게 오므려져 있다. 손과 팔의 연결 및 제지 자세는 가능하다. 두 사람의 하체와 발은 잘렸지만 몸통은 서 있는 자세와 일관되며, 떠 있거나 지지 없이 매달린 신체·물체는 없다."
       }
      ],
      "all_candidates_fail": false
     },
     "normalized": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 5,
        "verdict_ko": "입술을 깨물고 팔을 붙잡는 동작은 보이지만, 박상현이 다른 곳이 아니라 박진성을 바라보고 두 손으로 제지하며 회색 의상도 참조와 다릅니다."
       },
       {
        "label": "A",
        "score": 8,
        "verdict_ko": "박상현의 굳은 정면에 가까운 얼굴, 옆으로 돌린 시선, 깨문 입술과 한 손의 제지 접촉을 가장 잘 구현했으나 두 사람의 갓과 머리 모양은 참조에서 벗어납니다."
       }
      ],
      "all_candidates_fail": false,
      "readings": [
       {
        "label": "B",
        "direction": "왼쪽 박상현은 고개와 눈을 오른쪽 박진성의 얼굴 쪽으로 향한다. 박진성도 뒤돌아 박상현을 바라보므로 서로 마주 보는 관계이며, 박상현이 다른 곳으로 시선을 피한다는 지시와 다르다. 박진성의 몸통은 대문 쪽을 향하지만 실제 전진 움직임은 뚜렷하지 않다.",
        "built_space": "왼쪽에 목조 건물 한 동의 툇마루, 어두운 출입구 하나와 격자창 하나가 보이고, 아래에 돌 디딤돌 하나가 있다. 오른쪽 뒤에는 기와지붕 대문 하나와 이어진 담장이 있으며 문짝 사이로 바깥이 보여 개방 상태가 읽힌다. 두 사람은 건물과 대문 사이 흙마당에 서 있다. 참조의 재료와 배치는 대체로 이어지지만 인물 주변 마당의 노출이 비교적 넓다. 반사면이나 중복된 대문은 없다.",
        "entities": "등장인물은 성인 한국인 남성으로 보이는 두 명뿐이다. 왼쪽의 짧은 수염과 얼굴은 박상현, 오른쪽의 수염 없는 얼굴과 연한 청색 포는 박진성에 대응한다. 두 사람 모두 검은 머리이지만 참조의 갓을 쓰지 않고 상투와 망건만 보인다. 박상현은 참조의 회색 긴소매 포 대신 흰 소매 위에 회색 겉옷을 입었다. 박진성의 청색 허리끈은 유지되지만 참조의 갓끈 장식은 없다. 읽을 수 있는 글자는 없다.",
        "hard_violations": [],
        "physics": "박상현의 두 손이 박진성의 위팔과 팔꿈치 부근 소매를 각각 잡고 있어 접촉과 당김은 물리적으로 가능하다. 다만 한 손을 단단히 잠가 제지한다는 지정과 달리 두 손으로 붙잡는다. 박진성은 몸통을 돌리고 잡힌 팔을 아래로 내린 자세로, 제지당한 상태로 가능하다. 발은 화면 밖이지만 두 몸은 정상적인 직립 자세이며 공중 부양이나 지지 없는 물체는 보이지 않는다."
       },
       {
        "label": "A",
        "direction": "오른쪽 박상현은 얼굴을 거의 정면으로 드러낸 채 눈을 화면 오른쪽 바깥으로 돌려 왼쪽 박진성과 시선을 맞추지 않는다. 화면 밖의 구체적인 대상은 보이지 않는다. 박진성도 오른쪽을 바라보며 몸과 잡힌 팔을 그쪽으로 내밀어, 나아가려다 제지된 관계가 읽힌다.",
        "built_space": "왼쪽 가장자리에 목조 건물의 격자창 하나와 툇마루 일부, 뒤에는 기와지붕 대문 하나와 양옆 돌담이 보인다. 두 사람은 대문 안쪽 흙마당에 나란히 가까이 서 있고, 몸 사이와 가장자리로 마당이 드러난다. 참조의 목재·기와·돌담과 낮의 조명은 대체로 유지된다. 대문 통로 대부분이 박상현에게 가려져 활짝 열린 상태인지는 확정하기 어렵다. 중복된 시설이나 불가능한 반사는 없다.",
        "entities": "성인 한국인 남성으로 보이는 두 명만 등장한다. 오른쪽 박상현의 짧은 수염, 얼굴과 회색 포, 왼쪽 박진성의 수염 없는 얼굴과 청색 포·허리끈은 참조의 구분에 부합한다. 그러나 둘 다 갓이 없고 짧게 정돈한 검은 머리를 드러내며, 박진성의 장식 갓끈도 없다. 박상현의 입술은 안쪽으로 눌러 깨문 듯 보이고 미간과 턱에 긴장이 있다. 읽을 수 있는 글자는 없다.",
        "hard_violations": [],
        "physics": "박상현의 한 손이 박진성의 팔꿈치 가까운 아래팔 소매를 감싸 쥐며 천을 눌러, 단단한 제지 접촉이 보인다. 잡힌 팔은 팔꿈치에서 굽어 몸 앞을 가로지르고 손은 주먹에 가깝게 오므려져 있다. 손과 팔의 연결 및 제지 자세는 가능하다. 두 사람의 하체와 발은 잘렸지만 몸통은 서 있는 자세와 일관되며, 떠 있거나 지지 없이 매달린 신체·물체는 없다."
       }
      ]
     },
     "ok": true
    }
   ],
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "slot_winner_match": false,
   "slot_winner": {
    "gemini-pro": "B",
    "gpt-high": "A"
   },
   "route": "cross_slot_combined"
  },
  "dual": {
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "normalized": {
    "A": 1.0,
    "B": 1.625
   },
   "adjusted": {
    "A": 0.75,
    "B": 1.625
   },
   "violations": {
    "A": [
     "[gemini-pro] 해부학적으로 불가능한 신체 구조 (박상현의 오른쪽 소매 하나에서 두 개의 손이 나와 푸른 소매를 잡고 있으며, 소매 안쪽에 또 다른 손이 존재해 한쪽 팔에 3개의 손이 엉켜 있음)"
    ]
   },
   "per_model_winner": {
    "gemini-pro": "B",
    "gpt-high": "A"
   },
   "agreed": false
  },
  "totals": {
   "A": 750,
   "B": 1625
  },
  "selected": "B",
  "ranking": [
   "B",
   "A"
  ],
  "verdicts": [
   {
    "label": "A",
    "score": 750,
    "verdict_ko": "정면 구도와 시선 처리 등 지문 연출은 훌륭하나, 박상현의 한 소매에서 3개의 손이 나오는 치명적인 해부학적 오류가 발생해 탈락입니다.  ★위반: [gemini-pro] 해부학적으로 불가능한 신체 구조 (박상현의 오른쪽 소매 하나에서 두 개의 손이 나와 푸른 소매를 잡고 있으며, 소매 안쪽에 또 다른 손이 존재해 한쪽 팔에 3개의 손이 엉켜 있음)"
   },
   {
    "label": "B",
    "score": 1625,
    "verdict_ko": "해부학적 오류 없이 배경과 인물 간의 신체 접촉을 안정적으로 구현했으나, '정면 구도'와 '다른 곳을 향한 시선'이라는 핵심 연출을 놓쳤습니다."
   }
  ],
  "refs": [
   {
    "label": "PREVIOUS SHOT STILL — a visually related earlier shot of this same place: the location's look, materials, fixed features and lighting mood are LOCKED to this photo; never copy its camera framing. The people visible in that photo are NOT in this shot: never carry their faces, bodies or clothing onto anyone here. Each person in THIS shot is defined solely by the PEOPLE list, their own CHARACTER REFERENCE images and their wardrobe notes. If this photo shows a character who cannot move (dead or unconscious), that character's exact pose, position and orientation are ALSO LOCKED — treat their body as a fixed prop of the set that only the camera moves around.",
    "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/db00517b-165b-4005-a287-f28936b59b95/images/272bd7af-9294-4490-98c6-88a40e9deacc/scene/recipe/S40sh4_sel.png",
    "asset_id": "836eec0a-8646-4f80-9c8e-ec07622b6f0b",
    "role": "prev_still"
   },
   {
    "label": "CHARACTER REFERENCE — 박진성: the exact person appearing in this shot; match face, hair and build exactly.",
    "path": "<bytes:912297>",
    "asset_id": "3690b9ae-055b-43ee-9b74-794f83fca3d5",
    "role": "character_ref"
   },
   {
    "label": "CHARACTER REFERENCE — 박상현: the exact person appearing in this shot; match face, hair and build exactly.",
    "path": "<bytes:840781>",
    "asset_id": "6187751d-5df3-4eb7-9d30-d6fc98a005e6",
    "role": "character_ref"
   }
  ],
  "critique_skipped": true,
  "shot_run_produced": false,
  "shot_run_uid": "06aa142a-6975-7914-b72e-3d74ee62aa6f",
  "ref_mode": "prev+엔티티",
  "share_plan": {
   "ref_plan": "prev",
   "prev_anchor_tag": "S40sh4"
  }
 },
 "S40sh12::cine": {
  "applied": true,
  "attempted_at": "2026-09-09T05:29:45.592493+00:00",
  "fingerprint": "3d8224edd841f94565aa3ad302936bdef4f437023c57099685b4d31451634f52",
  "fingerprint_version": 2,
  "provider": "mai",
  "endpoint": "openrouter/chat-completions",
  "model": "microsoft/mai-image-2.6",
  "pack": "24.202608252115",
  "source_file": "S40sh12_sel.png",
  "source_sha256": "79265fe43e6dad26d62be0f6ad36a9be2038b9e8d3f0549f8f2e098c2ace1b71",
  "file": "S40sh12_cine.png",
  "staged_sha256": "48a6b5cf3b4337980d91f4b90c69128f25cec028c0588eaa17d78962df5bb0ce",
  "latency_ms": 23692
 },
 "S40sh13::signage": {
  "fp": "a1563eb67a3aa9a1",
  "inscriptions": [],
  "cues": [],
  "dropped": []
 },
 "S40sh13": {
  "input_fingerprint": "551b761662eb8937",
  "prompt": "Create ONE FINAL photorealistic live-action film still of the moment below — early 20th-century Korea, Jeolla region; all people are Korean unless explicitly stated otherwise. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 일본인 헌병들에게 양팔이 붙들린 채 대문 밖을 향해 한쪽 발을 내딛고 억지로 걷는 mid-action 중인 박제현의 뒷모습.\n\nLOCATION (lock): The estate's outdoor main-gate threshold, leading from the courtyard to the lane outside. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: wide shot\n- FRAME LAYOUT: Open gate ahead of the departing escort in the upper-center of the frame, background.\n- KEY BACKGROUND ELEMENTS: 대문 (Open as 박제현 is escorted out) — Its courtyard-facing side and the passage through it are seen ahead of the departing figures; used as Provides the destination for the forced step without obscuring the restrained arms; 마당 바닥 (Visible beneath the three departing figures); used as Makes the advancing foot and unequal stride phases legible.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Continue the same natural courtyard daylight and restrained contrast through the forced departure, avoiding a new brightness cue at the gate.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The people visible in that photo are NOT in this shot: never carry their faces, bodies or clothing onto anyone here. Each person in THIS shot is defined solely by the PEOPLE list, their own CHARACTER REFERENCE images and their wardrobe notes. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The main gate remains open. 박제현: He is being taken through the gate with his arms held and twisted.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 박제현 (한국인 남성, 중년 이상 나이대, 자연스러운 연령대의 얼굴) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.",
  "roll_prompts": {
   "A": "Create ONE FINAL photorealistic live-action film still of the moment below — early 20th-century Korea, Jeolla region; all people are Korean unless explicitly stated otherwise. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 일본인 헌병들에게 양팔이 붙들린 채 대문 밖을 향해 한쪽 발을 내딛고 억지로 걷는 mid-action 중인 박제현의 뒷모습.\n\nLOCATION (lock): The estate's outdoor main-gate threshold, leading from the courtyard to the lane outside. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: wide shot\n- FRAME LAYOUT: Open gate ahead of the departing escort in the upper-center of the frame, background.\n- KEY BACKGROUND ELEMENTS: 대문 (Open as 박제현 is escorted out) — Its courtyard-facing side and the passage through it are seen ahead of the departing figures; used as Provides the destination for the forced step without obscuring the restrained arms; 마당 바닥 (Visible beneath the three departing figures); used as Makes the advancing foot and unequal stride phases legible.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Continue the same natural courtyard daylight and restrained contrast through the forced departure, avoiding a new brightness cue at the gate.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The people visible in that photo are NOT in this shot: never carry their faces, bodies or clothing onto anyone here. Each person in THIS shot is defined solely by the PEOPLE list, their own CHARACTER REFERENCE images and their wardrobe notes. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The main gate remains open. 박제현: He is being taken through the gate with his arms held and twisted.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 박제현 (한국인 남성, 중년 이상 나이대, 자연스러운 연령대의 얼굴) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.",
   "B": "Create ONE FINAL photorealistic live-action film still of the moment below — early 20th-century Korea, Jeolla region; all people are Korean unless explicitly stated otherwise. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 일본인 헌병들에게 양팔이 붙들린 채 대문 밖을 향해 한쪽 발을 내딛고 억지로 걷는 mid-action 중인 박제현의 뒷모습.\n\nLOCATION (lock): The estate's outdoor main-gate threshold, leading from the courtyard to the lane outside. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: wide shot\n- FRAME LAYOUT: Open gate ahead of the departing escort in the upper-center of the frame, background.\n- KEY BACKGROUND ELEMENTS: 대문 (Open as 박제현 is escorted out) — Its courtyard-facing side and the passage through it are seen ahead of the departing figures; used as Provides the destination for the forced step without obscuring the restrained arms; 마당 바닥 (Visible beneath the three departing figures); used as Makes the advancing foot and unequal stride phases legible.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Continue the same natural courtyard daylight and restrained contrast through the forced departure, avoiding a new brightness cue at the gate.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The people visible in that photo are NOT in this shot: never carry their faces, bodies or clothing onto anyone here. Each person in THIS shot is defined solely by the PEOPLE list, their own CHARACTER REFERENCE images and their wardrobe notes. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The main gate remains open. 박제현: He is being taken through the gate with his arms held and twisted.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 박제현 (한국인 남성, 중년 이상 나이대, 자연스러운 연령대의 얼굴) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.\n\nCOMPOSITION VARIATION (this candidate only): keep the place, the people and the moment exactly as specified, and keep from the CAMERA & FRAME contract its subject scale, screen placement and key background elements exactly. For this candidate only, its camera angle and height are deliberately left open: choose a camera position distinctly different from the obvious one, as a film director picking a second setup on the same blocking. Stay true to any explicit framing the SHOT TEXT itself dictates."
  },
  "shot_run_spend_attempt_count": 1,
  "readings": [
   {
    "label": "A",
    "direction": "세 인물 모두 프레임 중앙 상단의 열린 대문을 향해 나아가고 있다.",
    "built_space": "레퍼런스 사진의 마당 구조를 정확히 따름. 왼쪽에 마루가 있는 한옥 건물, 정면에 기와가 얹힌 대문과 돌담이 올바른 방향과 위치에 배치되어 있음.",
    "entities": "박제현은 레퍼런스와 동일하게 갓을 쓰고 흰색 두루마기를 입은 뒷모습으로 등장함. 두 명의 일본인 헌병이 그의 양팔을 붙들고 있음.",
    "hard_violations": [],
    "physics": "박제현의 왼발이 바닥을 딛고 체중을 지탱하며, 오른발은 앞으로 내딛기 위해 자연스럽게 들려 있음. 헌병들의 발도 지면에 안정적으로 닿아 정상적인 보행을 보여줌."
   },
   {
    "label": "B",
    "direction": "인물들이 정면에 있는 열린 대문을 향해 나아가고 있다.",
    "built_space": "정면에 대문이 있으나, 프레임 양옆이 어두운 실내 기둥과 벽으로 둘러싸여 있어 프롬프트가 지시한 야외 마당 시점이 아님.",
    "entities": "박제현이 갓을 쓰지 않았고 두루마기 대신 짧은 저고리와 바지만 입고 있어 레퍼런스와 완전히 불일치함. 두 명의 헌병이 양팔을 잡고 있음.",
    "hard_violations": [
     "[gemini-pro] 아무것도 지탱하지 않는 공중에 뜬 몸 (박제현의 두 발이 모두 지면에서 떨어져 있음)",
     "[gemini-pro] 신체 구조 불가능 (오른쪽 헌병의 다리 주변에 세 번째 발이 존재함)",
     "[gemini-pro] 지정된 카메라 위치 위반 (야외 마당이 아닌 실내 기둥 사이에서 바라보는 구도)"
    ],
    "physics": "박제현의 두 발이 모두 지면에서 떠 있어 몸을 지지하는 곳이 전혀 없는 허공에 뜬 상태임. 우측 헌병의 다리 뒤로 여분의 신발이 보여 해부학적으로 불가능한 모습임."
   }
  ],
  "cross_model_order": {
   "policy": "cross_model_order_v2_slot0_fwd_slot1_rev_parallel2",
   "slots": [
    {
     "model": "gemini-pro",
     "order": "forward",
     "display_to_canonical": {
      "A": "A",
      "B": "B"
     },
     "raw": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 9,
        "verdict_ko": "레퍼런스에 맞춰 캐릭터의 외형과 배경 공간을 정확하게 구현했으며, 양팔이 붙들린 채 걷는 뒷모습의 물리적 묘사도 안정적입니다."
       },
       {
        "label": "B",
        "score": 0,
        "verdict_ko": "주인공이 공중에 떠 있고, 갓과 두루마기가 누락되어 캐릭터 외형이 불일치하며, 실내에서 밖을 보는 잘못된 구도를 취한 데다 헌병의 신체 구조까지 붕괴된 치명적인 오류가 있습니다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "세 인물 모두 프레임 중앙 상단의 열린 대문을 향해 나아가고 있다.",
        "built_space": "레퍼런스 사진의 마당 구조를 정확히 따름. 왼쪽에 마루가 있는 한옥 건물, 정면에 기와가 얹힌 대문과 돌담이 올바른 방향과 위치에 배치되어 있음.",
        "entities": "박제현은 레퍼런스와 동일하게 갓을 쓰고 흰색 두루마기를 입은 뒷모습으로 등장함. 두 명의 일본인 헌병이 그의 양팔을 붙들고 있음.",
        "hard_violations": [],
        "physics": "박제현의 왼발이 바닥을 딛고 체중을 지탱하며, 오른발은 앞으로 내딛기 위해 자연스럽게 들려 있음. 헌병들의 발도 지면에 안정적으로 닿아 정상적인 보행을 보여줌."
       },
       {
        "label": "B",
        "direction": "인물들이 정면에 있는 열린 대문을 향해 나아가고 있다.",
        "built_space": "정면에 대문이 있으나, 프레임 양옆이 어두운 실내 기둥과 벽으로 둘러싸여 있어 프롬프트가 지시한 야외 마당 시점이 아님.",
        "entities": "박제현이 갓을 쓰지 않았고 두루마기 대신 짧은 저고리와 바지만 입고 있어 레퍼런스와 완전히 불일치함. 두 명의 헌병이 양팔을 잡고 있음.",
        "hard_violations": [
         "아무것도 지탱하지 않는 공중에 뜬 몸 (박제현의 두 발이 모두 지면에서 떨어져 있음)",
         "신체 구조 불가능 (오른쪽 헌병의 다리 주변에 세 번째 발이 존재함)",
         "지정된 카메라 위치 위반 (야외 마당이 아닌 실내 기둥 사이에서 바라보는 구도)"
        ],
        "physics": "박제현의 두 발이 모두 지면에서 떠 있어 몸을 지지하는 곳이 전혀 없는 허공에 뜬 상태임. 우측 헌병의 다리 뒤로 여분의 신발이 보여 해부학적으로 불가능한 모습임."
       }
      ],
      "all_candidates_fail": false
     },
     "normalized": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 9,
        "verdict_ko": "레퍼런스에 맞춰 캐릭터의 외형과 배경 공간을 정확하게 구현했으며, 양팔이 붙들린 채 걷는 뒷모습의 물리적 묘사도 안정적입니다."
       },
       {
        "label": "B",
        "score": 0,
        "verdict_ko": "주인공이 공중에 떠 있고, 갓과 두루마기가 누락되어 캐릭터 외형이 불일치하며, 실내에서 밖을 보는 잘못된 구도를 취한 데다 헌병의 신체 구조까지 붕괴된 치명적인 오류가 있습니다."
       }
      ],
      "all_candidates_fail": false,
      "readings": [
       {
        "label": "A",
        "direction": "세 인물 모두 프레임 중앙 상단의 열린 대문을 향해 나아가고 있다.",
        "built_space": "레퍼런스 사진의 마당 구조를 정확히 따름. 왼쪽에 마루가 있는 한옥 건물, 정면에 기와가 얹힌 대문과 돌담이 올바른 방향과 위치에 배치되어 있음.",
        "entities": "박제현은 레퍼런스와 동일하게 갓을 쓰고 흰색 두루마기를 입은 뒷모습으로 등장함. 두 명의 일본인 헌병이 그의 양팔을 붙들고 있음.",
        "hard_violations": [],
        "physics": "박제현의 왼발이 바닥을 딛고 체중을 지탱하며, 오른발은 앞으로 내딛기 위해 자연스럽게 들려 있음. 헌병들의 발도 지면에 안정적으로 닿아 정상적인 보행을 보여줌."
       },
       {
        "label": "B",
        "direction": "인물들이 정면에 있는 열린 대문을 향해 나아가고 있다.",
        "built_space": "정면에 대문이 있으나, 프레임 양옆이 어두운 실내 기둥과 벽으로 둘러싸여 있어 프롬프트가 지시한 야외 마당 시점이 아님.",
        "entities": "박제현이 갓을 쓰지 않았고 두루마기 대신 짧은 저고리와 바지만 입고 있어 레퍼런스와 완전히 불일치함. 두 명의 헌병이 양팔을 잡고 있음.",
        "hard_violations": [
         "아무것도 지탱하지 않는 공중에 뜬 몸 (박제현의 두 발이 모두 지면에서 떨어져 있음)",
         "신체 구조 불가능 (오른쪽 헌병의 다리 주변에 세 번째 발이 존재함)",
         "지정된 카메라 위치 위반 (야외 마당이 아닌 실내 기둥 사이에서 바라보는 구도)"
        ],
        "physics": "박제현의 두 발이 모두 지면에서 떠 있어 몸을 지지하는 곳이 전혀 없는 허공에 뜬 상태임. 우측 헌병의 다리 뒤로 여분의 신발이 보여 해부학적으로 불가능한 모습임."
       }
      ]
     },
     "ok": true
    },
    {
     "model": "gpt-high",
     "order": "reverse",
     "display_to_canonical": {
      "A": "B",
      "B": "A"
     },
     "raw": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 4,
        "verdict_ko": "양팔을 붙잡힌 강제 보행은 선명하지만, 대문을 크게 당긴 구도와 갓·긴 도포가 사라진 복장이 지정된 와이드 배치 및 인물 참조와 크게 어긋납니다."
       },
       {
        "label": "B",
        "score": 9,
        "verdict_ko": "상단 중앙의 열린 대문, 세 사람의 뒷모습과 엇갈린 걸음, 참조 복장을 충실히 구현했으며, 다만 팔을 강제로 비트는 긴장감은 다소 약합니다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "중앙 남성은 카메라에 등을 보이고 앞의 열린 대문과 바깥길을 향한다. 좌우 군인은 몸을 중앙 남성 쪽으로 기울이고 그의 팔을 바라보며 붙잡는다. 중앙 남성의 왼발은 뒤로 들려 있고 오른발이 앞쪽에서 체중을 받는 보행 순간이다. 조준하는 무기는 보이지 않는다.",
        "built_space": "앞쪽에 기와지붕 대문 한 곳, 좌우 기둥 두 개와 열린 목제 문짝 두 개, 돌 문턱과 양옆 담장이 보인다. 세 사람은 문턱 앞 흙마당에 있다. 화면 양 가장자리에는 가까운 어두운 목재 문틀이 들어온다. 대문 지붕이 화면 상단에서 잘리고 인물도 크게 차지해, 대문을 상단 중앙 배경으로 두는 지정 와이드 배치보다 가까운 구도다. 목재·기와·돌담 재질은 참조와 유사하지만 참조의 왼쪽 툇마루와 마당 관계는 확인하기 어렵다. 반사는 없다.",
        "entities": "인물은 중앙 남성 한 명과 군인 두 명으로 총 세 명이며, 이전 장면의 검은 조끼 남성은 없다. 중앙 인물은 중년으로 보이는 동아시아계 남성이지만 뒷모습만으로 한국인 여부나 정확한 얼굴 일치는 확인할 수 없다. 검은 머리는 보이나 참조의 갓이 없고, 크림색 긴 도포와 붉은 허리띠 대신 얼룩진 짧은 상의와 바지를 입었다. 좌우 남성은 녹색 군복과 군모 차림이며 일본 헌병 역할로 읽히지만 국적은 외형만으로 확정할 수 없다. 열린 대문과 마당은 있으며 읽을 수 있는 글자는 없다.",
        "hard_violations": [],
        "physics": "중앙 남성의 오른발 앞부분이 지면을 받치고 왼발은 뒤로 들려 있다. 양옆 군인은 발을 마당에 딛고 손으로 각각 중앙 남성의 상완과 팔 부근을 잡아 지지한다. 팔이 벌어지고 몸이 앞으로 끌리는 자세는 강제 보행으로 가능하다. 공중의 왼발에는 반대쪽 지지발과 양쪽의 붙잡는 손이 있으므로 무지지 부유는 아니다."
       },
       {
        "label": "B",
        "direction": "세 사람 모두 등을 보이며 상단 중앙의 열린 대문과 그 너머 길을 향한다. 중앙 남성의 고개도 대문 방향이다. 양옆 헌병의 안쪽 팔은 중앙 남성의 양팔로 향해 접촉한다. 중앙 남성은 왼발로 버티면서 오른발을 들고 이동 중이며, 헌병들의 발도 서로 다른 보행 단계에 있다. 조준하는 무기는 보이지 않는다.",
        "built_space": "기와지붕 대문 한 곳과 양쪽 기둥, 열린 문짝 및 중앙 통로가 세 사람 앞의 상단 중앙에 놓인다. 문짝 일부는 인물에 가리지만 통로는 열려 있다. 왼쪽에는 목조 건물의 툇마루와 격자창, 대문 양옆에는 기와를 얹은 돌담, 아래에는 돌 문턱과 흙마당이 이어져 참조 장소의 주요 관계를 유지한다. 세 사람은 마당에서 문턱으로 접근하며 발밑 공간이 충분히 보인다. 오른쪽 담 너머 건물은 참조에서 확인되지 않는 부분이지만 고정 시설의 중복은 보이지 않는다. 반사는 없다.",
        "entities": "인물은 박제현 역할의 중앙 남성과 헌병 두 명뿐이다. 중앙 남성의 검은 갓, 크림색 긴 도포, 붉은 허리띠, 흰 버선과 검은 신발은 인물 참조에 부합한다. 체격도 유사하며 얼굴은 지정된 뒷모습 때문에 확인되지 않는다. 헌병들은 검은 계열 군복, 군모, 가죽 허리띠와 장화를 착용한 성인 남성으로 보이며 국적은 외형만으로 확인할 수 없다. 대문 너머에는 이전 장소 참조에도 있던 차량 일부가 보인다. 읽을 수 있는 글자는 없다.",
        "hard_violations": [],
        "physics": "중앙 남성은 왼발을 지면에 딛고 오른발을 들어 보행 중이다. 양옆 헌병도 각각 지지발을 마당에 둔 채 걷고 있으며 안쪽 손으로 그의 양팔을 잡는다. 중앙 남성의 두 손은 허리 뒤에서 모여 있어 뒤로 제한된 팔 상태는 읽히지만, 강하게 비틀리는 모습보다는 손을 모은 자세에 가깝다. 몸과 의복에는 발·손의 지지와 중력에 맞는 접촉 및 처짐이 있으며 떠 있는 물체는 없다."
       }
      ],
      "all_candidates_fail": false
     },
     "normalized": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 4,
        "verdict_ko": "양팔을 붙잡힌 강제 보행은 선명하지만, 대문을 크게 당긴 구도와 갓·긴 도포가 사라진 복장이 지정된 와이드 배치 및 인물 참조와 크게 어긋납니다."
       },
       {
        "label": "A",
        "score": 9,
        "verdict_ko": "상단 중앙의 열린 대문, 세 사람의 뒷모습과 엇갈린 걸음, 참조 복장을 충실히 구현했으며, 다만 팔을 강제로 비트는 긴장감은 다소 약합니다."
       }
      ],
      "all_candidates_fail": false,
      "readings": [
       {
        "label": "B",
        "direction": "중앙 남성은 카메라에 등을 보이고 앞의 열린 대문과 바깥길을 향한다. 좌우 군인은 몸을 중앙 남성 쪽으로 기울이고 그의 팔을 바라보며 붙잡는다. 중앙 남성의 왼발은 뒤로 들려 있고 오른발이 앞쪽에서 체중을 받는 보행 순간이다. 조준하는 무기는 보이지 않는다.",
        "built_space": "앞쪽에 기와지붕 대문 한 곳, 좌우 기둥 두 개와 열린 목제 문짝 두 개, 돌 문턱과 양옆 담장이 보인다. 세 사람은 문턱 앞 흙마당에 있다. 화면 양 가장자리에는 가까운 어두운 목재 문틀이 들어온다. 대문 지붕이 화면 상단에서 잘리고 인물도 크게 차지해, 대문을 상단 중앙 배경으로 두는 지정 와이드 배치보다 가까운 구도다. 목재·기와·돌담 재질은 참조와 유사하지만 참조의 왼쪽 툇마루와 마당 관계는 확인하기 어렵다. 반사는 없다.",
        "entities": "인물은 중앙 남성 한 명과 군인 두 명으로 총 세 명이며, 이전 장면의 검은 조끼 남성은 없다. 중앙 인물은 중년으로 보이는 동아시아계 남성이지만 뒷모습만으로 한국인 여부나 정확한 얼굴 일치는 확인할 수 없다. 검은 머리는 보이나 참조의 갓이 없고, 크림색 긴 도포와 붉은 허리띠 대신 얼룩진 짧은 상의와 바지를 입었다. 좌우 남성은 녹색 군복과 군모 차림이며 일본 헌병 역할로 읽히지만 국적은 외형만으로 확정할 수 없다. 열린 대문과 마당은 있으며 읽을 수 있는 글자는 없다.",
        "hard_violations": [],
        "physics": "중앙 남성의 오른발 앞부분이 지면을 받치고 왼발은 뒤로 들려 있다. 양옆 군인은 발을 마당에 딛고 손으로 각각 중앙 남성의 상완과 팔 부근을 잡아 지지한다. 팔이 벌어지고 몸이 앞으로 끌리는 자세는 강제 보행으로 가능하다. 공중의 왼발에는 반대쪽 지지발과 양쪽의 붙잡는 손이 있으므로 무지지 부유는 아니다."
       },
       {
        "label": "A",
        "direction": "세 사람 모두 등을 보이며 상단 중앙의 열린 대문과 그 너머 길을 향한다. 중앙 남성의 고개도 대문 방향이다. 양옆 헌병의 안쪽 팔은 중앙 남성의 양팔로 향해 접촉한다. 중앙 남성은 왼발로 버티면서 오른발을 들고 이동 중이며, 헌병들의 발도 서로 다른 보행 단계에 있다. 조준하는 무기는 보이지 않는다.",
        "built_space": "기와지붕 대문 한 곳과 양쪽 기둥, 열린 문짝 및 중앙 통로가 세 사람 앞의 상단 중앙에 놓인다. 문짝 일부는 인물에 가리지만 통로는 열려 있다. 왼쪽에는 목조 건물의 툇마루와 격자창, 대문 양옆에는 기와를 얹은 돌담, 아래에는 돌 문턱과 흙마당이 이어져 참조 장소의 주요 관계를 유지한다. 세 사람은 마당에서 문턱으로 접근하며 발밑 공간이 충분히 보인다. 오른쪽 담 너머 건물은 참조에서 확인되지 않는 부분이지만 고정 시설의 중복은 보이지 않는다. 반사는 없다.",
        "entities": "인물은 박제현 역할의 중앙 남성과 헌병 두 명뿐이다. 중앙 남성의 검은 갓, 크림색 긴 도포, 붉은 허리띠, 흰 버선과 검은 신발은 인물 참조에 부합한다. 체격도 유사하며 얼굴은 지정된 뒷모습 때문에 확인되지 않는다. 헌병들은 검은 계열 군복, 군모, 가죽 허리띠와 장화를 착용한 성인 남성으로 보이며 국적은 외형만으로 확인할 수 없다. 대문 너머에는 이전 장소 참조에도 있던 차량 일부가 보인다. 읽을 수 있는 글자는 없다.",
        "hard_violations": [],
        "physics": "중앙 남성은 왼발을 지면에 딛고 오른발을 들어 보행 중이다. 양옆 헌병도 각각 지지발을 마당에 둔 채 걷고 있으며 안쪽 손으로 그의 양팔을 잡는다. 중앙 남성의 두 손은 허리 뒤에서 모여 있어 뒤로 제한된 팔 상태는 읽히지만, 강하게 비틀리는 모습보다는 손을 모은 자세에 가깝다. 몸과 의복에는 발·손의 지지와 중력에 맞는 접촉 및 처짐이 있으며 떠 있는 물체는 없다."
       }
      ]
     },
     "ok": true
    }
   ],
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "slot_winner_match": true,
   "slot_winner": {
    "gemini-pro": "A",
    "gpt-high": "A"
   },
   "route": "cross_slot_agree"
  },
  "dual": {
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "normalized": {
    "A": 2.0,
    "B": 0.444
   },
   "adjusted": {
    "A": 2.0,
    "B": 0.194
   },
   "violations": {
    "B": [
     "[gemini-pro] 아무것도 지탱하지 않는 공중에 뜬 몸 (박제현의 두 발이 모두 지면에서 떨어져 있음)",
     "[gemini-pro] 신체 구조 불가능 (오른쪽 헌병의 다리 주변에 세 번째 발이 존재함)",
     "[gemini-pro] 지정된 카메라 위치 위반 (야외 마당이 아닌 실내 기둥 사이에서 바라보는 구도)"
    ]
   },
   "per_model_winner": {
    "gemini-pro": "A",
    "gpt-high": "A"
   },
   "agreed": true
  },
  "totals": {
   "A": 2000,
   "B": 194
  },
  "selected": "A",
  "ranking": [
   "A",
   "B"
  ],
  "verdicts": [
   {
    "label": "A",
    "score": 2000,
    "verdict_ko": "레퍼런스에 맞춰 캐릭터의 외형과 배경 공간을 정확하게 구현했으며, 양팔이 붙들린 채 걷는 뒷모습의 물리적 묘사도 안정적입니다."
   },
   {
    "label": "B",
    "score": 194,
    "verdict_ko": "주인공이 공중에 떠 있고, 갓과 두루마기가 누락되어 캐릭터 외형이 불일치하며, 실내에서 밖을 보는 잘못된 구도를 취한 데다 헌병의 신체 구조까지 붕괴된 치명적인 오류가 있습니다.  ★위반: [gemini-pro] 아무것도 지탱하지 않는 공중에 뜬 몸 (박제현의 두 발이 모두 지면에서 떨어져 있음) / [gemini-pro] 신체 구조 불가능 (오른쪽 헌병의 다리 주변에 세 번째 발이 존재함) / [gemini-pro] 지정된 카메라 위치 위반 (야외 마당이 아닌 실내 기둥 사이에서 바라보는 구도)"
   }
  ],
  "refs": [
   {
    "label": "PREVIOUS SHOT STILL — a visually related earlier shot of this same place: the location's look, materials, fixed features and lighting mood are LOCKED to this photo; never copy its camera framing. The people visible in that photo are NOT in this shot: never carry their faces, bodies or clothing onto anyone here. Each person in THIS shot is defined solely by the PEOPLE list, their own CHARACTER REFERENCE images and their wardrobe notes. If this photo shows a character who cannot move (dead or unconscious), that character's exact pose, position and orientation are ALSO LOCKED — treat their body as a fixed prop of the set that only the camera moves around.",
    "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/db00517b-165b-4005-a287-f28936b59b95/images/272bd7af-9294-4490-98c6-88a40e9deacc/scene/recipe/S40sh4_sel.png",
    "asset_id": "836eec0a-8646-4f80-9c8e-ec07622b6f0b",
    "role": "prev_still"
   },
   {
    "label": "CHARACTER REFERENCE — 박제현: the exact person appearing in this shot; match face, hair and build exactly.",
    "path": "<bytes:870304>",
    "asset_id": "02a56dab-2641-421d-89df-af22b5de63df",
    "role": "character_ref"
   }
  ],
  "critique_skipped": true,
  "shot_run_produced": false,
  "shot_run_uid": "06aa142a-6b1d-71b2-a5e1-03cd78604ee1",
  "ref_mode": "prev+엔티티",
  "share_plan": {
   "ref_plan": "prev",
   "prev_anchor_tag": "S40sh4"
  },
  "lane_policy": "ab_select_bypass:prev"
 },
 "S40sh13::cine": {
  "applied": true,
  "attempted_at": "2026-09-09T05:31:07.256544+00:00",
  "fingerprint": "6e704b885cad6d97d0a0f7c7adb77e8d46e459339f040536c0ce1e0ffcda9fd4",
  "fingerprint_version": 2,
  "provider": "mai",
  "endpoint": "openrouter/chat-completions",
  "model": "microsoft/mai-image-2.6",
  "pack": "24.202608252115",
  "source_file": "S40sh13_sel.png",
  "source_sha256": "50429daad68bf401f11d2964cf5a1c34078e252acf5c2e6e92a97ce918541620",
  "file": "S40sh13_cine.png",
  "staged_sha256": "7e49a2bd14c41b23155f09704467cfc1d3e138999108133fbe37a08cfb29ef56",
  "latency_ms": 27840
 },
 "S41sh1::signage": {
  "fp": "5642193da98e4395",
  "inscriptions": [],
  "cues": [],
  "dropped": []
 },
 "S41sh1::bgfirst_bg": {
  "input_fingerprint": "9ad0527cd9089901",
  "prompt": "Create the EMPTY BACKGROUND PLATE for one film shot — NO PEOPLE, no figures, no body parts, no sketch lines, no arrows anywhere.\n\"Empty\" means no people only: KEEP the location's inherent occupants and stock that define the place — animals in an animal shelter, pen or farm, goods in a market, moored boats in a harbour — unless the shot text explicitly removes them.\nThe FIRST attached image is a thin-line storyboard sketch: use ONLY its camera angle, horizon, perspective and the placement/size of buildings and set masses — ignore the sketched people and arrows entirely. The SECOND attached image (LOCATION PHOTOGRAPH) is the real place: take its architecture, materials, signage and fixed features, and RE-PROJECT them into the sketch's camera. If the photograph's camera differs from the sketch's, the sketch's camera wins.\nHUMAN-SCALE CALIBRATION: derive every structure's true size from human-scale elements — a door ≈ 2m, a window ≈ 1–1.5m wide, one storey ≈ 2.5–3m; never inflate a small structure or shrink a large one.\n\nSHOT TEXT this background must serve (Korean): 죽어가는 모닥불 주변에 무거운 표정으로 흩어져 앉은 조선인 의병들 전경.\n\nLOCATION (lock): The outdoor gathering circle around a low, dying campfire in the mountain camp at night.\n\nTIME OF DAY (lock): night.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: wide shot\n- FRAME LAYOUT: Unoccupied place beside the blacksmith's hammer in the lower-right of the frame, midground.\n- KEY BACKGROUND ELEMENTS: 모닥불 (A single fire burning unusually low); used as Serves as a small spatial center around which the seated intervals are readable; 대장장이의 빈 자리와 망치 (The place is unoccupied; his hammer lies there and nobody sits beside it) — The hammer lies obliquely within the empty place, at a modest scale relative to the seated people; used as Creates a visible break in the group that gives the loss a physical location.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: The single low-burning campfire provides subdued nighttime illumination with restrained saturation, leaving the surrounding darkness intact without losing the seated figures.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nRender ONE photorealistic empty location photograph, 16:9, that this shot can be staged inside later. No readable writing anywhere: surfaces that would carry writing may be present, but stage any wording out of legibility — an oblique angle, distance, shallow focus. No captions, watermarks or overlay text.",
  "effective_prompt": "Create the EMPTY BACKGROUND PLATE for one film shot — NO PEOPLE, no figures, no body parts, no sketch lines, no arrows anywhere.\n\"Empty\" means no people only: KEEP the location's inherent occupants and stock that define the place — animals in an animal shelter, pen or farm, goods in a market, moored boats in a harbour — unless the shot text explicitly removes them.\nThe FIRST attached image is a thin-line storyboard sketch: use ONLY its camera angle, horizon, perspective and the placement/size of buildings and set masses — ignore the sketched people and arrows entirely. The SECOND attached image (LOCATION PHOTOGRAPH) is the real place: take its architecture, materials, signage and fixed features, and RE-PROJECT them into the sketch's camera. If the photograph's camera differs from the sketch's, the sketch's camera wins.\nHUMAN-SCALE CALIBRATION: derive every structure's true size from human-scale elements — a door ≈ 2m, a window ≈ 1–1.5m wide, one storey ≈ 2.5–3m; never inflate a small structure or shrink a large one.\n\nSHOT TEXT this background must serve (Korean): 죽어가는 모닥불 주변에 무거운 표정으로 흩어져 앉은 조선인 의병들 전경.\n\nLOCATION (lock): The outdoor gathering circle around a low, dying campfire in the mountain camp at night.\n\nTIME OF DAY (lock): night.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: wide shot\n- FRAME LAYOUT: Unoccupied place beside the blacksmith's hammer in the lower-right of the frame, midground.\n- KEY BACKGROUND ELEMENTS: 모닥불 (A single fire burning unusually low); used as Serves as a small spatial center around which the seated intervals are readable; 대장장이의 빈 자리와 망치 (The place is unoccupied; his hammer lies there and nobody sits beside it) — The hammer lies obliquely within the empty place, at a modest scale relative to the seated people; used as Creates a visible break in the group that gives the loss a physical location.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: The single low-burning campfire provides subdued nighttime illumination with restrained saturation, leaving the surrounding darkness intact without losing the seated figures.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nRender ONE photorealistic empty location photograph, 16:9, that this shot can be staged inside later. No readable writing anywhere: surfaces that would carry writing may be present, but stage any wording out of legibility — an oblique angle, distance, shallow focus. No captions, watermarks or overlay text.",
  "bg_path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/db00517b-165b-4005-a287-f28936b59b95/images/272bd7af-9294-4490-98c6-88a40e9deacc/scene/recipe/S41sh1__bgfirst_bg.png",
  "asset_id": "293d627c-0109-4923-b81d-e1da0fcbec5c",
  "input_asset_ids": [
   "ba41a792-d2b2-4f36-98c5-72e5dbd65f1a",
   "87bcdd2b-607c-40d6-8d7f-f4dfbf3ed054"
  ]
 },
 "S41sh1": {
  "input_fingerprint": "780c7aa912a6c006",
  "prompt": "Create ONE FINAL photorealistic live-action film still of the moment below — early 20th-century Korea, Jeolla region; all people are Korean unless explicitly stated otherwise. TIME OF DAY (lock): night.\n\nSHOT TEXT (authoritative, Korean): 죽어가는 모닥불 주변에 무거운 표정으로 흩어져 앉은 조선인 의병들 전경.\n\nLOCATION (lock): The outdoor gathering circle around a low, dying campfire in the mountain camp at night. The shot takes place here — the attached LOCATION PHOTOGRAPH shows the exact spot.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: wide shot\n- FRAME LAYOUT: Unoccupied place beside the blacksmith's hammer in the lower-right of the frame, midground.\n- KEY BACKGROUND ELEMENTS: 모닥불 (A single fire burning unusually low); used as Serves as a small spatial center around which the seated intervals are readable; 대장장이의 빈 자리와 망치 (The place is unoccupied; his hammer lies there and nobody sits beside it) — The hammer lies obliquely within the empty place, at a modest scale relative to the seated people; used as Creates a visible break in the group that gives the loss a physical location.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: The single low-burning campfire provides subdued nighttime illumination with restrained saturation, leaving the surrounding darkness intact without losing the seated figures.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): A low campfire burns beside an unoccupied place marked only by a hammer.\n\nNo people appear unless the shot text itself says so.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.",
  "roll_prompts": {
   "A": "Stage the shot. The FIRST attached image (SHOT BACKGROUND) is the finished empty background of this shot — keep it EXACTLY: its camera, perspective, architecture, lighting and every fixed feature stay untouched. The SECOND attached image (LAYOUT SKETCH) tells you ONLY where the people go: each sketched person's position, screen size, pose and the gaze/motion arrows. Ignore the sketch's background lines. The CHARACTER REFERENCE photographs show the real people.\nPlace the real people into the background at exactly the sketched positions, sizes and poses, following the arrow directions. No sketch lines or arrows may remain.\n\nCreate ONE FINAL photorealistic live-action film still of the moment below — early 20th-century Korea, Jeolla region; all people are Korean unless explicitly stated otherwise. TIME OF DAY (lock): night.\n\nSHOT TEXT (authoritative, Korean): 죽어가는 모닥불 주변에 무거운 표정으로 흩어져 앉은 조선인 의병들 전경.\n\nLOCATION (lock): The outdoor gathering circle around a low, dying campfire in the mountain camp at night. The shot takes place here — the FIRST attached image (SHOT BACKGROUND) is this exact place, already built: its ground, structures, horizon, materials and lighting are the finished truth of this location and must not be redesigned or replaced. No location photograph is attached — read the place from that image alone, and add no scenery, structure, vehicle or fixture that it does not already show. This lock governs the place only; the figures in the shot follow the staging and pose instructions.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: wide shot\n- FRAME LAYOUT: Unoccupied place beside the blacksmith's hammer in the lower-right of the frame, midground.\n- KEY BACKGROUND ELEMENTS: 모닥불 (A single fire burning unusually low); used as Serves as a small spatial center around which the seated intervals are readable; 대장장이의 빈 자리와 망치 (The place is unoccupied; his hammer lies there and nobody sits beside it) — The hammer lies obliquely within the empty place, at a modest scale relative to the seated people; used as Creates a visible break in the group that gives the loss a physical location.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: The single low-burning campfire provides subdued nighttime illumination with restrained saturation, leaving the surrounding darkness intact without losing the seated figures.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): A low campfire burns beside an unoccupied place marked only by a hammer.\n\nNo people appear unless the shot text itself says so.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.",
   "B": "Create ONE FINAL photorealistic live-action film still of the moment below — early 20th-century Korea, Jeolla region; all people are Korean unless explicitly stated otherwise. TIME OF DAY (lock): night.\n\nSHOT TEXT (authoritative, Korean): 죽어가는 모닥불 주변에 무거운 표정으로 흩어져 앉은 조선인 의병들 전경.\n\nLOCATION (lock): The outdoor gathering circle around a low, dying campfire in the mountain camp at night. The shot takes place here — the attached LOCATION PHOTOGRAPH shows the exact spot.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: wide shot\n- FRAME LAYOUT: Unoccupied place beside the blacksmith's hammer in the lower-right of the frame, midground.\n- KEY BACKGROUND ELEMENTS: 모닥불 (A single fire burning unusually low); used as Serves as a small spatial center around which the seated intervals are readable; 대장장이의 빈 자리와 망치 (The place is unoccupied; his hammer lies there and nobody sits beside it) — The hammer lies obliquely within the empty place, at a modest scale relative to the seated people; used as Creates a visible break in the group that gives the loss a physical location.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: The single low-burning campfire provides subdued nighttime illumination with restrained saturation, leaving the surrounding darkness intact without losing the seated figures.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): A low campfire burns beside an unoccupied place marked only by a hammer.\n\nNo people appear unless the shot text itself says so.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.\n\nCOMPOSITION VARIATION (this candidate only): keep the place, the people and the moment exactly as specified, and keep from the CAMERA & FRAME contract its subject scale, screen placement and key background elements exactly. For this candidate only, its camera angle and height are deliberately left open: choose a camera position distinctly different from the obvious one, as a film director picking a second setup on the same blocking. Stay true to any explicit framing the SHOT TEXT itself dictates."
  },
  "roll_refs": {
   "A": [
    {
     "label": "SHOT BACKGROUND",
     "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/db00517b-165b-4005-a287-f28936b59b95/images/272bd7af-9294-4490-98c6-88a40e9deacc/scene/recipe/S41sh1__bgfirst_bg.png",
     "asset_id": "293d627c-0109-4923-b81d-e1da0fcbec5c",
     "role": "bgfirst_bg"
    },
    {
     "label": "LAYOUT SKETCH (people placement only)",
     "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/db00517b-165b-4005-a287-f28936b59b95/images/272bd7af-9294-4490-98c6-88a40e9deacc/conti/conti_S41sh1.png",
     "asset_id": "ba41a792-d2b2-4f36-98c5-72e5dbd65f1a",
     "role": "conti_light"
    }
   ],
   "B": [
    {
     "label": "LOCATION PHOTOGRAPH — the exact place of this shot: its architecture, materials, fixed features and lighting mood are spatial truth; stage the moment inside this place. Never copy its camera framing.",
     "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/db00517b-165b-4005-a287-f28936b59b95/images/272bd7af-9294-4490-98c6-88a40e9deacc/scene/recipe/groupbg_mountain_camp_hearth_81c95f.png",
     "asset_id": "87bcdd2b-607c-40d6-8d7f-f4dfbf3ed054",
     "role": "bgfirst_group_bg"
    }
   ]
  },
  "shot_run_spend_attempt_count": 1,
  "judge_flip_skipped": "judge_fn_owns_order",
  "readings": [
   {
    "label": "A",
    "direction": "인물들은 고개를 숙이거나 모닥불 쪽으로 시선을 두고 있으며, 우측에서 두 번째 인물은 무릎 위의 책을 보며 무언가를 적는 자세를 취하고 있다.",
    "built_space": "배경에 목책과 구조물이 보이지만 기준 이미지의 거대한 가마솥이 누락되었다. 또한 전경 좌측과 중앙에 각각 활활 타오르는 모닥불이 두 개 존재하여 '단 하나의 모닥불'이라는 공간적 지시와 충돌한다.",
    "entities": "조선인 의병들이 묘사되었으나, 우측 하단의 망치가 의미를 알 수 없는 바닥의 둥근 선 안에 놓여 있어 지시문의 의도와 다르게 표현되었다.",
    "hard_violations": [
     "[gemini-pro] 지시문에 명시된 '단 하나의 모닥불(A single fire)' 규정을 위반하고 화면에 두 개의 모닥불을 중복하여 배치함.",
     "[gpt-high] 단 하나로 지정된 모닥불을 두 개로 중복 배치했습니다.",
     "[gpt-high] 모여 앉은 의병 장면에 지시되지 않은 펼친 책과 독서 행동을 추가했습니다.",
     "[gpt-high] 망치만으로 표시되어야 할 빈자리에 별도의 원형 테두리 표식을 추가했습니다."
    ],
    "physics": "인물들의 앉은 자세와 지면에 닿은 형태는 물리적으로 무리가 없으나, 바닥에 그려진 원형 선은 자연스러운 질감이 아닌 그래픽적 요소처럼 보인다."
   },
   {
    "label": "B",
    "direction": "수많은 의병들이 원형으로 둘러앉아 있으며, 대부분 고개를 푹 숙인 채 중앙의 잦아드는 불씨나 바닥을 무거운 표정으로 응시하고 있다.",
    "built_space": "기준 이미지의 장소(목책, 초가집, 진흙 화덕, 거대한 검은 가마솥 등)가 정확한 위치에 구현되었다. 공간 중앙의 돌무더기 화로 안에 단 하나의 모닥불이 배치되어 있다.",
    "entities": "지시된 조선인 의병들이 묘사되었으며, 화면 우측 하단 전경에는 대장장이의 망치가 바위 위에 비스듬히 놓여 있어 의병들 사이의 시각적 단절과 빈자리를 정확하게 형성하고 있다.",
    "hard_violations": [
     "[gpt-high] 명시적으로 아무도 앉지 않아야 하는 망치 옆에 오른쪽 앞 남성이 앉아 있어, 핵심 빈자리의 인물 배치를 위반합니다."
    ],
    "physics": "인물들은 땅바닥에 자연스럽게 무게를 두고 앉아 있으며, 우측 하단의 망치 역시 돌 표면에 안정적으로 지탱되어 있다."
   }
  ],
  "cross_model_order": {
   "policy": "cross_model_order_v2_slot0_fwd_slot1_rev_parallel2",
   "slots": [
    {
     "model": "gemini-pro",
     "order": "forward",
     "display_to_canonical": {
      "A": "A",
      "B": "B"
     },
     "raw": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 10,
        "verdict_ko": "지시문이 요구한 꺼져가는 단 하나의 모닥불, 우측 하단의 망치가 놓인 빈자리, 기준 이미지의 핵심 지형물(가마솥 등)을 완벽하게 재현했습니다."
       },
       {
        "label": "A",
        "score": 3,
        "verdict_ko": "'단 하나의 모닥불'이라는 명시적 지시를 어기고 두 개의 불을 배치했으며, 화면 우측 하단의 망치가 작위적인 원형 선 안에 놓여 있어 부자연스럽습니다."
       }
      ],
      "readings": [
       {
        "label": "B",
        "direction": "수많은 의병들이 원형으로 둘러앉아 있으며, 대부분 고개를 푹 숙인 채 중앙의 잦아드는 불씨나 바닥을 무거운 표정으로 응시하고 있다.",
        "built_space": "기준 이미지의 장소(목책, 초가집, 진흙 화덕, 거대한 검은 가마솥 등)가 정확한 위치에 구현되었다. 공간 중앙의 돌무더기 화로 안에 단 하나의 모닥불이 배치되어 있다.",
        "entities": "지시된 조선인 의병들이 묘사되었으며, 화면 우측 하단 전경에는 대장장이의 망치가 바위 위에 비스듬히 놓여 있어 의병들 사이의 시각적 단절과 빈자리를 정확하게 형성하고 있다.",
        "hard_violations": [],
        "physics": "인물들은 땅바닥에 자연스럽게 무게를 두고 앉아 있으며, 우측 하단의 망치 역시 돌 표면에 안정적으로 지탱되어 있다."
       },
       {
        "label": "A",
        "direction": "인물들은 고개를 숙이거나 모닥불 쪽으로 시선을 두고 있으며, 우측에서 두 번째 인물은 무릎 위의 책을 보며 무언가를 적는 자세를 취하고 있다.",
        "built_space": "배경에 목책과 구조물이 보이지만 기준 이미지의 거대한 가마솥이 누락되었다. 또한 전경 좌측과 중앙에 각각 활활 타오르는 모닥불이 두 개 존재하여 '단 하나의 모닥불'이라는 공간적 지시와 충돌한다.",
        "entities": "조선인 의병들이 묘사되었으나, 우측 하단의 망치가 의미를 알 수 없는 바닥의 둥근 선 안에 놓여 있어 지시문의 의도와 다르게 표현되었다.",
        "hard_violations": [
         "지시문에 명시된 '단 하나의 모닥불(A single fire)' 규정을 위반하고 화면에 두 개의 모닥불을 중복하여 배치함."
        ],
        "physics": "인물들의 앉은 자세와 지면에 닿은 형태는 물리적으로 무리가 없으나, 바닥에 그려진 원형 선은 자연스러운 질감이 아닌 그래픽적 요소처럼 보인다."
       }
      ],
      "all_candidates_fail": false
     },
     "normalized": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 10,
        "verdict_ko": "지시문이 요구한 꺼져가는 단 하나의 모닥불, 우측 하단의 망치가 놓인 빈자리, 기준 이미지의 핵심 지형물(가마솥 등)을 완벽하게 재현했습니다."
       },
       {
        "label": "A",
        "score": 3,
        "verdict_ko": "'단 하나의 모닥불'이라는 명시적 지시를 어기고 두 개의 불을 배치했으며, 화면 우측 하단의 망치가 작위적인 원형 선 안에 놓여 있어 부자연스럽습니다."
       }
      ],
      "all_candidates_fail": false,
      "readings": [
       {
        "label": "B",
        "direction": "수많은 의병들이 원형으로 둘러앉아 있으며, 대부분 고개를 푹 숙인 채 중앙의 잦아드는 불씨나 바닥을 무거운 표정으로 응시하고 있다.",
        "built_space": "기준 이미지의 장소(목책, 초가집, 진흙 화덕, 거대한 검은 가마솥 등)가 정확한 위치에 구현되었다. 공간 중앙의 돌무더기 화로 안에 단 하나의 모닥불이 배치되어 있다.",
        "entities": "지시된 조선인 의병들이 묘사되었으며, 화면 우측 하단 전경에는 대장장이의 망치가 바위 위에 비스듬히 놓여 있어 의병들 사이의 시각적 단절과 빈자리를 정확하게 형성하고 있다.",
        "hard_violations": [],
        "physics": "인물들은 땅바닥에 자연스럽게 무게를 두고 앉아 있으며, 우측 하단의 망치 역시 돌 표면에 안정적으로 지탱되어 있다."
       },
       {
        "label": "A",
        "direction": "인물들은 고개를 숙이거나 모닥불 쪽으로 시선을 두고 있으며, 우측에서 두 번째 인물은 무릎 위의 책을 보며 무언가를 적는 자세를 취하고 있다.",
        "built_space": "배경에 목책과 구조물이 보이지만 기준 이미지의 거대한 가마솥이 누락되었다. 또한 전경 좌측과 중앙에 각각 활활 타오르는 모닥불이 두 개 존재하여 '단 하나의 모닥불'이라는 공간적 지시와 충돌한다.",
        "entities": "조선인 의병들이 묘사되었으나, 우측 하단의 망치가 의미를 알 수 없는 바닥의 둥근 선 안에 놓여 있어 지시문의 의도와 다르게 표현되었다.",
        "hard_violations": [
         "지시문에 명시된 '단 하나의 모닥불(A single fire)' 규정을 위반하고 화면에 두 개의 모닥불을 중복하여 배치함."
        ],
        "physics": "인물들의 앉은 자세와 지면에 닿은 형태는 물리적으로 무리가 없으나, 바닥에 그려진 원형 선은 자연스러운 질감이 아닌 그래픽적 요소처럼 보인다."
       }
      ]
     },
     "ok": true
    },
    {
     "model": "gpt-high",
     "order": "reverse",
     "display_to_canonical": {
      "A": "B",
      "B": "A"
     },
     "raw": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 4,
        "verdict_ko": "밤의 와이드 숏과 하나의 죽어가는 모닥불은 충실하지만, 망치 바로 옆에 사람이 앉아 있어 핵심인 대장장이의 빈자리 배치를 위반합니다."
       },
       {
        "label": "B",
        "score": 2,
        "verdict_ko": "여섯 인물 사이의 간격은 읽히지만, 모닥불이 두 개이고 독서 행동과 빈자리의 원형 표식까지 추가되어 명시된 장면에서 크게 벗어납니다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "대부분의 인물은 고개를 숙여 자기 무릎이나 모닥불 쪽 바닥을 바라봅니다. 앞쪽의 등을 보인 인물들도 몸을 불 쪽으로 향하고 있습니다. 망치는 오른쪽 아래에서 쇠머리가 왼쪽 아래, 자루가 오른쪽 위를 향하도록 비스듬히 놓여 있습니다. 누군가를 겨냥하는 무기는 보이지 않습니다.",
        "built_space": "왼쪽의 일부 잘린 움막, 중앙의 작은 천막형 작업 공간, 오른쪽의 긴 초가 구조물 등 세 구조물과 뒤쪽 목책이 보여 참조 장소의 배치를 상당히 유지합니다. 중앙 뒤에는 낮은 흙 화덕 두 기와 오른쪽에 세운 큰 금속 솥뚜껑 하나가 보입니다. 모닥불은 하나입니다. 많은 사람이 불 주위에 여러 겹으로 밀집해 앉았고 앞쪽은 넓게 열려 있습니다. 그러나 오른쪽 아래 망치와 돌 바로 옆에 남성이 앉아 있어, 아무도 옆에 앉지 않는 빈자리라는 조건을 충족하지 못합니다.",
        "entities": "인물들은 성인에서 중년의 한국인 남성으로 설정에 부합하는 외양이며, 상투·머리띠와 낡은 저고리·바지를 착용했습니다. 정확한 국적은 외양만으로 확정할 수 없지만 시대 배경에 어긋나는 인물은 보이지 않습니다. 표정과 숙인 자세는 무겁고 침울합니다. 불은 장작 사이의 붉은 잔불과 극히 작은 불꽃으로 표현되어 죽어가는 모닥불에 부합합니다. 나무 자루와 금속 머리가 있는 대장장이용 망치 한 개가 보입니다. 읽을 수 있는 글자는 없습니다.",
        "hard_violations": [
         "명시적으로 아무도 앉지 않아야 하는 망치 옆에 오른쪽 앞 남성이 앉아 있어, 핵심 빈자리의 인물 배치를 위반합니다."
        ],
        "physics": "인물들의 엉덩이와 접은 다리는 땅에 놓여 있고 손은 무릎이나 다리에 얹혀 있습니다. 망치 머리는 흙바닥에 닿고 자루는 돌에 기대어 지지됩니다. 장작과 불씨는 돌로 둘러싸인 지면 위에 놓여 있습니다. 공중에 떠 있거나 지지 없이 매달린 몸과 물체는 보이지 않습니다."
       },
       {
        "label": "B",
        "direction": "왼쪽 세 남성과 오른쪽 끝 남성은 주로 무릎이나 앞쪽 바닥으로 시선을 낮춥니다. 오른쪽 중앙 남성은 왼쪽의 동료들을 바라보고, 그 왼쪽 남성은 손에 든 펼친 책을 내려다봅니다. 책의 펼친 면은 독자의 눈을 향해 있어 사용 방향 자체는 자연스럽습니다. 오른쪽 아래 망치는 머리가 왼쪽 위, 자루가 오른쪽 아래를 향합니다. 왼쪽 인물들의 등 뒤 긴 무기는 위쪽을 향하며 특정 사람을 겨냥하지 않습니다.",
        "built_space": "참조와 대응하는 왼쪽 움막, 중앙 작업 천막, 오른쪽 긴 초가 구조물 세 곳과 목책, 산비탈이 보입니다. 낮은 흙 화덕들도 중앙 뒤에 남아 있습니다. 여섯 남성은 좌우 세 명씩 떨어져 앉아 있지만, 그 사이의 불과 화면 아래 돌테두리 안의 불이 따로 있어 모닥불이 두 개입니다. 망치는 오른쪽 아래 열린 바닥에 놓였고 바로 붙어 앉은 사람은 없습니다. 다만 그 자리를 둘러싼 원형 테두리가 추가되어 있습니다. 오른쪽 건물 안에도 별도의 따뜻한 불빛이 보입니다.",
        "entities": "여섯 인물은 성인·중년의 한국인 남성 설정과 양립하는 외양이고, 머리띠와 낡은 전통 복식을 착용했습니다. 표정은 대체로 무겁습니다. 망치 한 개는 적당한 크기로 보이지만, 모닥불 두 개 모두 뚜렷한 불꽃이 솟아 유난히 낮게 타는 단 하나의 불이라는 조건과 다릅니다. 한 남성에게 펼친 책과 독서 행동이 추가되었습니다. 빈자리에는 망치뿐 아니라 원형 표식도 있습니다. 읽을 수 있는 글자는 보이지 않습니다.",
        "hard_violations": [
         "단 하나로 지정된 모닥불을 두 개로 중복 배치했습니다.",
         "모여 앉은 의병 장면에 지시되지 않은 펼친 책과 독서 행동을 추가했습니다.",
         "망치만으로 표시되어야 할 빈자리에 별도의 원형 테두리 표식을 추가했습니다."
        ],
        "physics": "여섯 인물 모두 엉덩이와 접은 다리가 지면에 지지되고, 손은 다리 위에 놓이거나 책을 받칩니다. 책은 양손으로 지지되며 망치도 바닥에 놓여 있습니다. 두 불의 장작은 각각 지면에 닿아 있습니다. 지지 없는 부유나 명백히 불가능한 자세는 보이지 않습니다."
       }
      ],
      "all_candidates_fail": true
     },
     "normalized": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 4,
        "verdict_ko": "밤의 와이드 숏과 하나의 죽어가는 모닥불은 충실하지만, 망치 바로 옆에 사람이 앉아 있어 핵심인 대장장이의 빈자리 배치를 위반합니다."
       },
       {
        "label": "A",
        "score": 2,
        "verdict_ko": "여섯 인물 사이의 간격은 읽히지만, 모닥불이 두 개이고 독서 행동과 빈자리의 원형 표식까지 추가되어 명시된 장면에서 크게 벗어납니다."
       }
      ],
      "all_candidates_fail": true,
      "readings": [
       {
        "label": "B",
        "direction": "대부분의 인물은 고개를 숙여 자기 무릎이나 모닥불 쪽 바닥을 바라봅니다. 앞쪽의 등을 보인 인물들도 몸을 불 쪽으로 향하고 있습니다. 망치는 오른쪽 아래에서 쇠머리가 왼쪽 아래, 자루가 오른쪽 위를 향하도록 비스듬히 놓여 있습니다. 누군가를 겨냥하는 무기는 보이지 않습니다.",
        "built_space": "왼쪽의 일부 잘린 움막, 중앙의 작은 천막형 작업 공간, 오른쪽의 긴 초가 구조물 등 세 구조물과 뒤쪽 목책이 보여 참조 장소의 배치를 상당히 유지합니다. 중앙 뒤에는 낮은 흙 화덕 두 기와 오른쪽에 세운 큰 금속 솥뚜껑 하나가 보입니다. 모닥불은 하나입니다. 많은 사람이 불 주위에 여러 겹으로 밀집해 앉았고 앞쪽은 넓게 열려 있습니다. 그러나 오른쪽 아래 망치와 돌 바로 옆에 남성이 앉아 있어, 아무도 옆에 앉지 않는 빈자리라는 조건을 충족하지 못합니다.",
        "entities": "인물들은 성인에서 중년의 한국인 남성으로 설정에 부합하는 외양이며, 상투·머리띠와 낡은 저고리·바지를 착용했습니다. 정확한 국적은 외양만으로 확정할 수 없지만 시대 배경에 어긋나는 인물은 보이지 않습니다. 표정과 숙인 자세는 무겁고 침울합니다. 불은 장작 사이의 붉은 잔불과 극히 작은 불꽃으로 표현되어 죽어가는 모닥불에 부합합니다. 나무 자루와 금속 머리가 있는 대장장이용 망치 한 개가 보입니다. 읽을 수 있는 글자는 없습니다.",
        "hard_violations": [
         "명시적으로 아무도 앉지 않아야 하는 망치 옆에 오른쪽 앞 남성이 앉아 있어, 핵심 빈자리의 인물 배치를 위반합니다."
        ],
        "physics": "인물들의 엉덩이와 접은 다리는 땅에 놓여 있고 손은 무릎이나 다리에 얹혀 있습니다. 망치 머리는 흙바닥에 닿고 자루는 돌에 기대어 지지됩니다. 장작과 불씨는 돌로 둘러싸인 지면 위에 놓여 있습니다. 공중에 떠 있거나 지지 없이 매달린 몸과 물체는 보이지 않습니다."
       },
       {
        "label": "A",
        "direction": "왼쪽 세 남성과 오른쪽 끝 남성은 주로 무릎이나 앞쪽 바닥으로 시선을 낮춥니다. 오른쪽 중앙 남성은 왼쪽의 동료들을 바라보고, 그 왼쪽 남성은 손에 든 펼친 책을 내려다봅니다. 책의 펼친 면은 독자의 눈을 향해 있어 사용 방향 자체는 자연스럽습니다. 오른쪽 아래 망치는 머리가 왼쪽 위, 자루가 오른쪽 아래를 향합니다. 왼쪽 인물들의 등 뒤 긴 무기는 위쪽을 향하며 특정 사람을 겨냥하지 않습니다.",
        "built_space": "참조와 대응하는 왼쪽 움막, 중앙 작업 천막, 오른쪽 긴 초가 구조물 세 곳과 목책, 산비탈이 보입니다. 낮은 흙 화덕들도 중앙 뒤에 남아 있습니다. 여섯 남성은 좌우 세 명씩 떨어져 앉아 있지만, 그 사이의 불과 화면 아래 돌테두리 안의 불이 따로 있어 모닥불이 두 개입니다. 망치는 오른쪽 아래 열린 바닥에 놓였고 바로 붙어 앉은 사람은 없습니다. 다만 그 자리를 둘러싼 원형 테두리가 추가되어 있습니다. 오른쪽 건물 안에도 별도의 따뜻한 불빛이 보입니다.",
        "entities": "여섯 인물은 성인·중년의 한국인 남성 설정과 양립하는 외양이고, 머리띠와 낡은 전통 복식을 착용했습니다. 표정은 대체로 무겁습니다. 망치 한 개는 적당한 크기로 보이지만, 모닥불 두 개 모두 뚜렷한 불꽃이 솟아 유난히 낮게 타는 단 하나의 불이라는 조건과 다릅니다. 한 남성에게 펼친 책과 독서 행동이 추가되었습니다. 빈자리에는 망치뿐 아니라 원형 표식도 있습니다. 읽을 수 있는 글자는 보이지 않습니다.",
        "hard_violations": [
         "단 하나로 지정된 모닥불을 두 개로 중복 배치했습니다.",
         "모여 앉은 의병 장면에 지시되지 않은 펼친 책과 독서 행동을 추가했습니다.",
         "망치만으로 표시되어야 할 빈자리에 별도의 원형 테두리 표식을 추가했습니다."
        ],
        "physics": "여섯 인물 모두 엉덩이와 접은 다리가 지면에 지지되고, 손은 다리 위에 놓이거나 책을 받칩니다. 책은 양손으로 지지되며 망치도 바닥에 놓여 있습니다. 두 불의 장작은 각각 지면에 닿아 있습니다. 지지 없는 부유나 명백히 불가능한 자세는 보이지 않습니다."
       }
      ]
     },
     "ok": true
    }
   ],
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "slot_winner_match": true,
   "slot_winner": {
    "gemini-pro": "B",
    "gpt-high": "B"
   },
   "route": "cross_slot_agree"
  },
  "dual": {
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "normalized": {
    "A": 0.8,
    "B": 2.0
   },
   "adjusted": {
    "A": 0.55,
    "B": 1.75
   },
   "violations": {
    "A": [
     "[gemini-pro] 지시문에 명시된 '단 하나의 모닥불(A single fire)' 규정을 위반하고 화면에 두 개의 모닥불을 중복하여 배치함.",
     "[gpt-high] 단 하나로 지정된 모닥불을 두 개로 중복 배치했습니다.",
     "[gpt-high] 모여 앉은 의병 장면에 지시되지 않은 펼친 책과 독서 행동을 추가했습니다.",
     "[gpt-high] 망치만으로 표시되어야 할 빈자리에 별도의 원형 테두리 표식을 추가했습니다."
    ],
    "B": [
     "[gpt-high] 명시적으로 아무도 앉지 않아야 하는 망치 옆에 오른쪽 앞 남성이 앉아 있어, 핵심 빈자리의 인물 배치를 위반합니다."
    ]
   },
   "per_model_winner": {
    "gemini-pro": "B",
    "gpt-high": "B"
   },
   "agreed": true
  },
  "totals": {
   "B": 1750,
   "A": 550
  },
  "selected": "B",
  "ranking": [
   "B",
   "A"
  ],
  "verdicts": [
   {
    "label": "B",
    "score": 1750,
    "verdict_ko": "지시문이 요구한 꺼져가는 단 하나의 모닥불, 우측 하단의 망치가 놓인 빈자리, 기준 이미지의 핵심 지형물(가마솥 등)을 완벽하게 재현했습니다.  ★위반: [gpt-high] 명시적으로 아무도 앉지 않아야 하는 망치 옆에 오른쪽 앞 남성이 앉아 있어, 핵심 빈자리의 인물 배치를 위반합니다."
   },
   {
    "label": "A",
    "score": 550,
    "verdict_ko": "'단 하나의 모닥불'이라는 명시적 지시를 어기고 두 개의 불을 배치했으며, 화면 우측 하단의 망치가 작위적인 원형 선 안에 놓여 있어 부자연스럽습니다.  ★위반: [gemini-pro] 지시문에 명시된 '단 하나의 모닥불(A single fire)' 규정을 위반하고 화면에 두 개의 모닥불을 중복하여 배치함. / [gpt-high] 단 하나로 지정된 모닥불을 두 개로 중복 배치했습니다. / [gpt-high] 모여 앉은 의병 장면에 지시되지 않은 펼친 책과 독서 행동을 추가했습니다. / [gpt-high] 망치만으로 표시되어야 할 빈자리에 별도의 원형 테두리 표식을 추가했습니다."
   }
  ],
  "refs": [
   {
    "label": "LOCATION PHOTOGRAPH — the exact place of this shot: its architecture, materials, fixed features and lighting mood are spatial truth; stage the moment inside this place. Never copy its camera framing.",
    "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/db00517b-165b-4005-a287-f28936b59b95/images/272bd7af-9294-4490-98c6-88a40e9deacc/scene/recipe/groupbg_mountain_camp_hearth_81c95f.png",
    "asset_id": "87bcdd2b-607c-40d6-8d7f-f4dfbf3ed054",
    "role": "bgfirst_group_bg"
   }
  ],
  "critique_skipped": true,
  "shot_run_produced": false,
  "shot_run_uid": "06aa142a-6cb1-7fcd-8da0-41081d11b158",
  "bgfirst": {
   "bg_path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/db00517b-165b-4005-a287-f28936b59b95/images/272bd7af-9294-4490-98c6-88a40e9deacc/scene/recipe/S41sh1__bgfirst_bg.png",
   "bg_asset_id": "293d627c-0109-4923-b81d-e1da0fcbec5c",
   "bg_record_key": "S41sh1::bgfirst_bg",
   "chain_winner": false,
   "authority": "groupbg",
   "group_key": "mountain_camp_hearth",
   "groupbg_asset_id": "87bcdd2b-607c-40d6-8d7f-f4dfbf3ed054"
  },
  "ref_mode": "그룹 배경+엔티티 (2택1: 무콘티 승)",
  "share_plan": {
   "ref_plan": "background"
  }
 },
 "S41sh1::cine": {
  "applied": true,
  "attempted_at": "2026-09-09T05:34:07.622384+00:00",
  "fingerprint": "e100e71960a857262e45fbdaa80ed84a922fb38fbd41103c7542d386d7f415ff",
  "fingerprint_version": 2,
  "provider": "mai",
  "endpoint": "openrouter/chat-completions",
  "model": "microsoft/mai-image-2.6",
  "pack": "24.202608252115",
  "source_file": "S41sh1_sel.png",
  "source_sha256": "7121a3546fe3625f3e3f9656882f3e200aa5b0f52423f7c38af8291cbdc77d80",
  "file": "S41sh1_cine.png",
  "staged_sha256": "271aa778554236feef4e713b5545eb196d793e1dee0dfa2fb4a262362b2af095",
  "latency_ms": 26814
 },
 "S41sh8::signage": {
  "fp": "cba052a72a185b4a",
  "inscriptions": [],
  "cues": [],
  "dropped": []
 },
 "S41sh8": {
  "input_fingerprint": "2e20b99ebe5e37d4",
  "prompt": "Create ONE FINAL photorealistic live-action film still of the moment below — early 20th-century Korea, Jeolla region; all people are Korean unless explicitly stated otherwise. TIME OF DAY (lock): night.\n\nSHOT TEXT (authoritative, Korean): 모닥불 너머로 붓글씨를 쓰는 자세로 멈춘 임정현을 뚫어져라 바라보는 안규홍의 침통한 정면.\n\nLOCATION (lock): The far side of the mountain camp's low campfire, facing the seated scribe and an unoccupied place beside a hammer. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: medium shot\n- KEY BACKGROUND ELEMENTS: 모닥불 (Still burning low between the seated men); used as Occupies a small lower-frame interval below the line connecting the writing hand and 안규홍's face; 임정현의 붓 (Paused in his writing hand) — Seen obliquely beside his foreground shoulder, small in relation to his hand; used as Provides the concrete object of 안규홍's attention without becoming an isolated insert.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Preserve the subdued illumination of the low-burning campfire and sober nighttime contrast, keeping the face and paused hand readable without adding another source.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The people visible in that photo are NOT in this shot: never carry their faces, bodies or clothing onto anyone here. Each person in THIS shot is defined solely by the PEOPLE list, their own CHARACTER REFERENCE images and their wardrobe notes. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nPREVIOUS STILL USAGE (follow exactly — what to take from the attached still and what to exclude; it governs place and objects only, never who is in this shot or how the camera sees them): Take the low campfire, surrounding dirt, and subdued nighttime color. Exclude the grain cart and food-unloading props from the earlier supply arrival.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The hammer remains alone at the empty place; the low campfire burns while a name is added to the written record. 안규홍: He remains at the gathering, retaining his sickle and jade ring.\n\nTHE HAND THAT IS DOING THIS: the object at the centre of this shot is being held, operated, read, aimed or handed over by 임정현 right now, so 임정현's hand — and as much of the wrist and forearm as the framing reaches — is in the frame, gripping or resting on that object exactly the way its real use requires. Match that hand to 임정현: its size, build, skin, age, grooming, sleeve and anything worn on it belong to that person and to no one else.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 안규홍 (한국인 남성, 청년 나이대, 검은 머리, 단단히 묶은 머리) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.",
  "roll_prompts": {
   "A": "Create ONE FINAL photorealistic live-action film still of the moment below — early 20th-century Korea, Jeolla region; all people are Korean unless explicitly stated otherwise. TIME OF DAY (lock): night.\n\nSHOT TEXT (authoritative, Korean): 모닥불 너머로 붓글씨를 쓰는 자세로 멈춘 임정현을 뚫어져라 바라보는 안규홍의 침통한 정면.\n\nLOCATION (lock): The far side of the mountain camp's low campfire, facing the seated scribe and an unoccupied place beside a hammer. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: medium shot\n- KEY BACKGROUND ELEMENTS: 모닥불 (Still burning low between the seated men); used as Occupies a small lower-frame interval below the line connecting the writing hand and 안규홍's face; 임정현의 붓 (Paused in his writing hand) — Seen obliquely beside his foreground shoulder, small in relation to his hand; used as Provides the concrete object of 안규홍's attention without becoming an isolated insert.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Preserve the subdued illumination of the low-burning campfire and sober nighttime contrast, keeping the face and paused hand readable without adding another source.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The people visible in that photo are NOT in this shot: never carry their faces, bodies or clothing onto anyone here. Each person in THIS shot is defined solely by the PEOPLE list, their own CHARACTER REFERENCE images and their wardrobe notes. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nPREVIOUS STILL USAGE (follow exactly — what to take from the attached still and what to exclude; it governs place and objects only, never who is in this shot or how the camera sees them): Take the low campfire, surrounding dirt, and subdued nighttime color. Exclude the grain cart and food-unloading props from the earlier supply arrival.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The hammer remains alone at the empty place; the low campfire burns while a name is added to the written record. 안규홍: He remains at the gathering, retaining his sickle and jade ring.\n\nTHE HAND THAT IS DOING THIS: the object at the centre of this shot is being held, operated, read, aimed or handed over by 임정현 right now, so 임정현's hand — and as much of the wrist and forearm as the framing reaches — is in the frame, gripping or resting on that object exactly the way its real use requires. Match that hand to 임정현: its size, build, skin, age, grooming, sleeve and anything worn on it belong to that person and to no one else.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 안규홍 (한국인 남성, 청년 나이대, 검은 머리, 단단히 묶은 머리) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.",
   "B": "Create ONE FINAL photorealistic live-action film still of the moment below — early 20th-century Korea, Jeolla region; all people are Korean unless explicitly stated otherwise. TIME OF DAY (lock): night.\n\nSHOT TEXT (authoritative, Korean): 모닥불 너머로 붓글씨를 쓰는 자세로 멈춘 임정현을 뚫어져라 바라보는 안규홍의 침통한 정면.\n\nLOCATION (lock): The far side of the mountain camp's low campfire, facing the seated scribe and an unoccupied place beside a hammer. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: medium shot\n- KEY BACKGROUND ELEMENTS: 모닥불 (Still burning low between the seated men); used as Occupies a small lower-frame interval below the line connecting the writing hand and 안규홍's face; 임정현의 붓 (Paused in his writing hand) — Seen obliquely beside his foreground shoulder, small in relation to his hand; used as Provides the concrete object of 안규홍's attention without becoming an isolated insert.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Preserve the subdued illumination of the low-burning campfire and sober nighttime contrast, keeping the face and paused hand readable without adding another source.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The people visible in that photo are NOT in this shot: never carry their faces, bodies or clothing onto anyone here. Each person in THIS shot is defined solely by the PEOPLE list, their own CHARACTER REFERENCE images and their wardrobe notes. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nPREVIOUS STILL USAGE (follow exactly — what to take from the attached still and what to exclude; it governs place and objects only, never who is in this shot or how the camera sees them): Take the low campfire, surrounding dirt, and subdued nighttime color. Exclude the grain cart and food-unloading props from the earlier supply arrival.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The hammer remains alone at the empty place; the low campfire burns while a name is added to the written record. 안규홍: He remains at the gathering, retaining his sickle and jade ring.\n\nTHE HAND THAT IS DOING THIS: the object at the centre of this shot is being held, operated, read, aimed or handed over by 임정현 right now, so 임정현's hand — and as much of the wrist and forearm as the framing reaches — is in the frame, gripping or resting on that object exactly the way its real use requires. Match that hand to 임정현: its size, build, skin, age, grooming, sleeve and anything worn on it belong to that person and to no one else.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 안규홍 (한국인 남성, 청년 나이대, 검은 머리, 단단히 묶은 머리) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.\n\nCOMPOSITION VARIATION (this candidate only): keep the place, the people and the moment exactly as specified, and keep from the CAMERA & FRAME contract its subject scale, screen placement and key background elements exactly. For this candidate only, its camera angle and height are deliberately left open: choose a camera position distinctly different from the obvious one, as a film director picking a second setup on the same blocking. Stay true to any explicit framing the SHOT TEXT itself dictates."
  },
  "shot_run_spend_attempt_count": 1,
  "readings": [
   {
    "label": "A",
    "direction": "안규홍의 뚫어지게 보는 시선은 전경 우측에 있는 임정현의 손과 붓을 향해 정확히 고정되어 있습니다. 붓은 허공에 멈춘 채 대기하고 있습니다.",
    "built_space": "화면 하단 중앙에 모닥불이 위치하며, 배경의 초가집과 구조물은 레퍼런스 사진과 잘 일치합니다. 다만 지시된 바위 대신 짚으로 엮은 멍석 위에 망치가 놓여 있습니다.",
    "entities": "안규홍의 얼굴과 묶은 머리, 흙 묻은 복장은 레퍼런스와 완벽히 일치하지만, 프롬프트에서 요구한 낫과 옥반지를 소지하지 않고 빈손을 무릎에 두고 있습니다. 전경의 임정현은 어깨와 팔, 손이 지시된 비례에 맞게 등장합니다.",
    "hard_violations": [],
    "physics": "안규홍은 맨땅에 안정적인 자세로 앉아 있으며, 임정현의 손은 붓을 자연스럽게 쥐고 허공에서 버티고 있습니다. 떠 있거나 지지되지 않은 물체는 없습니다."
   },
   {
    "label": "B",
    "direction": "안규홍의 시선이 화면 우측의 임정현을 향하고 있습니다. 임정현의 붓은 멈춰 있지 않고 종이에 닿아 선을 긋고 있습니다.",
    "built_space": "모닥불이 두 인물 사이 하단에 놓여 있고, 배경 구조물은 레퍼런스를 잘 따랐습니다. 좌측 바닥에는 바위 대신 멍석이 깔려 있고 그 위에 망치가 있습니다.",
    "entities": "안규홍은 레퍼런스와 동일한 외형을 지니고 있으며 오른손에 낫을 쥐고 있으나, 앉아있어야 할 상황에서 일어서 있습니다. 전경의 임정현은 뒷모습과 손이 노출됩니다.",
    "hard_violations": [
     "[gpt-high] 모닥불을 사이에 둔 착석 배치와 달리 안규홍을 일으켜 세워, 지정된 인물의 자세와 공간 배치를 바꾸었다."
    ],
    "physics": "안규홍이 땅을 딛고 똑바로 서 있으며, 임정현은 손에 붓을 쥐고 판자 위에 대고 있습니다. 물리적으로 불가능한 요소는 발견되지 않습니다."
   }
  ],
  "cross_model_order": {
   "policy": "cross_model_order_v2_slot0_fwd_slot1_rev_parallel2",
   "slots": [
    {
     "model": "gemini-pro",
     "order": "forward",
     "display_to_canonical": {
      "A": "A",
      "B": "B"
     },
     "raw": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 8,
        "verdict_ko": "지정된 미디엄 샷 프레이밍과 전경에서 허공에 멈춘 붓의 구도를 정확히 구현했으나, 안규홍의 낫과 옥반지가 누락되었고 읽을 수 없는 글씨 지시에도 불구하고 종이에 획이 너무 선명하게 드러난 점이 아쉽습니다."
       },
       {
        "label": "B",
        "score": 4,
        "verdict_ko": "안규홍이 낫을 들고 있는 점은 맞췄으나, 미디엄 샷 지시를 어기고 풀 샷으로 처리했으며 인물이 서 있고 붓이 종이에 닿아 있어 핵심적인 구도와 상황 연출(Priority 2)에서 크게 실패했습니다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "안규홍의 뚫어지게 보는 시선은 전경 우측에 있는 임정현의 손과 붓을 향해 정확히 고정되어 있습니다. 붓은 허공에 멈춘 채 대기하고 있습니다.",
        "built_space": "화면 하단 중앙에 모닥불이 위치하며, 배경의 초가집과 구조물은 레퍼런스 사진과 잘 일치합니다. 다만 지시된 바위 대신 짚으로 엮은 멍석 위에 망치가 놓여 있습니다.",
        "entities": "안규홍의 얼굴과 묶은 머리, 흙 묻은 복장은 레퍼런스와 완벽히 일치하지만, 프롬프트에서 요구한 낫과 옥반지를 소지하지 않고 빈손을 무릎에 두고 있습니다. 전경의 임정현은 어깨와 팔, 손이 지시된 비례에 맞게 등장합니다.",
        "hard_violations": [],
        "physics": "안규홍은 맨땅에 안정적인 자세로 앉아 있으며, 임정현의 손은 붓을 자연스럽게 쥐고 허공에서 버티고 있습니다. 떠 있거나 지지되지 않은 물체는 없습니다."
       },
       {
        "label": "B",
        "direction": "안규홍의 시선이 화면 우측의 임정현을 향하고 있습니다. 임정현의 붓은 멈춰 있지 않고 종이에 닿아 선을 긋고 있습니다.",
        "built_space": "모닥불이 두 인물 사이 하단에 놓여 있고, 배경 구조물은 레퍼런스를 잘 따랐습니다. 좌측 바닥에는 바위 대신 멍석이 깔려 있고 그 위에 망치가 있습니다.",
        "entities": "안규홍은 레퍼런스와 동일한 외형을 지니고 있으며 오른손에 낫을 쥐고 있으나, 앉아있어야 할 상황에서 일어서 있습니다. 전경의 임정현은 뒷모습과 손이 노출됩니다.",
        "hard_violations": [],
        "physics": "안규홍이 땅을 딛고 똑바로 서 있으며, 임정현은 손에 붓을 쥐고 판자 위에 대고 있습니다. 물리적으로 불가능한 요소는 발견되지 않습니다."
       }
      ],
      "all_candidates_fail": false
     },
     "normalized": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 8,
        "verdict_ko": "지정된 미디엄 샷 프레이밍과 전경에서 허공에 멈춘 붓의 구도를 정확히 구현했으나, 안규홍의 낫과 옥반지가 누락되었고 읽을 수 없는 글씨 지시에도 불구하고 종이에 획이 너무 선명하게 드러난 점이 아쉽습니다."
       },
       {
        "label": "B",
        "score": 4,
        "verdict_ko": "안규홍이 낫을 들고 있는 점은 맞췄으나, 미디엄 샷 지시를 어기고 풀 샷으로 처리했으며 인물이 서 있고 붓이 종이에 닿아 있어 핵심적인 구도와 상황 연출(Priority 2)에서 크게 실패했습니다."
       }
      ],
      "all_candidates_fail": false,
      "readings": [
       {
        "label": "A",
        "direction": "안규홍의 뚫어지게 보는 시선은 전경 우측에 있는 임정현의 손과 붓을 향해 정확히 고정되어 있습니다. 붓은 허공에 멈춘 채 대기하고 있습니다.",
        "built_space": "화면 하단 중앙에 모닥불이 위치하며, 배경의 초가집과 구조물은 레퍼런스 사진과 잘 일치합니다. 다만 지시된 바위 대신 짚으로 엮은 멍석 위에 망치가 놓여 있습니다.",
        "entities": "안규홍의 얼굴과 묶은 머리, 흙 묻은 복장은 레퍼런스와 완벽히 일치하지만, 프롬프트에서 요구한 낫과 옥반지를 소지하지 않고 빈손을 무릎에 두고 있습니다. 전경의 임정현은 어깨와 팔, 손이 지시된 비례에 맞게 등장합니다.",
        "hard_violations": [],
        "physics": "안규홍은 맨땅에 안정적인 자세로 앉아 있으며, 임정현의 손은 붓을 자연스럽게 쥐고 허공에서 버티고 있습니다. 떠 있거나 지지되지 않은 물체는 없습니다."
       },
       {
        "label": "B",
        "direction": "안규홍의 시선이 화면 우측의 임정현을 향하고 있습니다. 임정현의 붓은 멈춰 있지 않고 종이에 닿아 선을 긋고 있습니다.",
        "built_space": "모닥불이 두 인물 사이 하단에 놓여 있고, 배경 구조물은 레퍼런스를 잘 따랐습니다. 좌측 바닥에는 바위 대신 멍석이 깔려 있고 그 위에 망치가 있습니다.",
        "entities": "안규홍은 레퍼런스와 동일한 외형을 지니고 있으며 오른손에 낫을 쥐고 있으나, 앉아있어야 할 상황에서 일어서 있습니다. 전경의 임정현은 뒷모습과 손이 노출됩니다.",
        "hard_violations": [],
        "physics": "안규홍이 땅을 딛고 똑바로 서 있으며, 임정현은 손에 붓을 쥐고 판자 위에 대고 있습니다. 물리적으로 불가능한 요소는 발견되지 않습니다."
       }
      ]
     },
     "ok": true
    },
    {
     "model": "gpt-high",
     "order": "reverse",
     "display_to_canonical": {
      "A": "B",
      "B": "A"
     },
     "raw": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 3,
        "verdict_ko": "안규홍을 앉은 인물 대신 전신으로 세워 배치했고, 시선도 임정현보다 모닥불 쪽으로 내려가 있어 지정된 침통한 정면 미디엄 숏을 구현하지 못했다."
       },
       {
        "label": "B",
        "score": 7,
        "verdict_ko": "앉은 안규홍의 침통한 정면과 임정현을 향한 응시가 더 정확하지만, 전경의 붓과 기록판이 지나치게 커서 작게 배치하라는 구도 지시는 약해졌다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "안규홍의 얼굴은 정면에 가깝지만 눈은 아래쪽 모닥불과 지면 방향으로 내려가 있으며, 오른쪽 전경의 임정현을 뚫어져라 보는 시선은 아니다. 임정현은 고개를 숙이고 종이 쪽을 향하며, 오른손의 붓끝도 종이를 향한다. 안규홍이 든 낫은 아래로 늘어져 있고 누구를 겨누지 않는다.",
        "built_space": "흙바닥에 돌로 둘러싼 모닥불 하나가 있고, 뒤에는 목책과 좌우 움막 및 중앙 천막이 보인다. 중앙 뒤쪽에는 화덕 형태의 시설 두 개와 오른쪽의 세워진 큰 솥뚜껑이 보여 이전 장소의 주요 요소를 이어간다. 망치 하나는 뒤쪽 돗자리 위에 놓였다. 그러나 안규홍은 모닥불 건너편에 서 있어 두 사람이 앉아 불을 사이에 둔 배치와 다르며, 카메라도 그의 전신과 넓은 마당을 담는다.",
        "entities": "두 인물만 보이며 추가 군중은 없다. 안규홍은 참고와 유사한 청년 한국인 남성 외형, 뒤로 묶은 검은 머리, 낡은 밝은색 저고리와 바지, 허리띠를 갖췄다. 오른손에는 낫이 보이나 옥반지는 확인하기 어렵다. 전경 임정현의 머리·어깨·손과 짙은 소매가 보이며, 붓은 그의 손에 들려 있다. 망치와 낮은 장작불, 종이와 먹 도구가 있고 읽을 수 있는 글자는 보이지 않는다.",
        "hard_violations": [
         "모닥불을 사이에 둔 착석 배치와 달리 안규홍을 일으켜 세워, 지정된 인물의 자세와 공간 배치를 바꾸었다."
        ],
        "physics": "안규홍의 두 발은 지면에 닿고 오른손은 낫을 잡고 있어 몸과 도구의 지지가 분명하다. 임정현의 손가락은 붓대를 잡고 붓끝을 종이 위에 멈춰 두어 중단된 필기 동작이 가능하다. 기록판의 받침은 하단 크롭 밖이라 전체 구조는 확인되지 않지만 공중에 떠 있다고 볼 근거는 없다. 망치는 돗자리 위에, 장작과 돌은 지면에 놓여 있다."
       },
       {
        "label": "B",
        "direction": "안규홍은 정면에 가까운 얼굴로 눈을 약간 오른쪽 전경의 임정현 쪽에 고정하고 있다. 낮게 겨눈 눈과 굳은 입매가 침통한 응시를 만든다. 임정현의 드러난 얼굴 부분은 기록판 쪽으로 숙여져 있고, 손에 잡힌 붓끝은 종이 위를 향한 채 멈춰 있다. 붓을 무기처럼 다른 사람에게 겨누는 오류는 없다.",
        "built_space": "모닥불 하나를 사이에 두고 안규홍이 흙바닥에 가부좌로 앉아 있으며, 임정현의 어깨와 필기하는 손이 오른쪽 전경을 차지한다. 왼쪽 빈 돗자리에는 망치 하나가 홀로 놓였다. 목책, 좌우 움막, 중앙 천막과 뒤쪽 화덕 두 개 및 솥뚜껑이 보여 장소의 연속성은 유지된다. 안규홍이 화면에서 충분히 커져 A보다 미디엄 숏에 가깝다. 불은 얼굴과 필기 손을 잇는 선 아래에 있지만, 큰 기록판과 붓이 하단 및 전경을 상당히 점유한다.",
        "entities": "안규홍과 부분적으로 보이는 임정현 외의 사람은 없다. 안규홍의 청년 한국인 남성 외형, 묶은 검은 머리, 얼굴과 체격, 낡은 밝은색 옷과 허리띠가 참고에 대체로 부합한다. 드러난 손에서는 옥반지가 확인되지 않으며 낫도 화면에서 확인할 수 없다. 임정현의 손과 짙은 소매는 전경의 몸에 연결되고 붓을 잡고 있다. 붓, 종이, 먹 도구, 망치가 실제 물체로 표현되지만 붓은 손에 비해 작게 보이지 않는다. 종이에 먹자국이 선명하게 보이지만 확실히 판독되는 문구는 없다.",
        "hard_violations": [],
        "physics": "안규홍의 엉덩이와 접은 다리는 지면에 놓이고 양손은 무릎 부근에 얹혀 있어 착석 자세가 성립한다. 임정현의 오른손은 붓을 잡고 다른 손은 기록판 가장자리를 받친다. 기록판 아래에는 목재 받침도 일부 보여 지지 없는 부유물은 아니다. 붓끝을 종이 바로 위에 멈춘 자세도 가능하다. 망치는 돗자리에 놓이고 불붙은 장작은 돌 테두리 안에서 서로 기대어 지지된다."
       }
      ],
      "all_candidates_fail": false
     },
     "normalized": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 3,
        "verdict_ko": "안규홍을 앉은 인물 대신 전신으로 세워 배치했고, 시선도 임정현보다 모닥불 쪽으로 내려가 있어 지정된 침통한 정면 미디엄 숏을 구현하지 못했다."
       },
       {
        "label": "A",
        "score": 7,
        "verdict_ko": "앉은 안규홍의 침통한 정면과 임정현을 향한 응시가 더 정확하지만, 전경의 붓과 기록판이 지나치게 커서 작게 배치하라는 구도 지시는 약해졌다."
       }
      ],
      "all_candidates_fail": false,
      "readings": [
       {
        "label": "B",
        "direction": "안규홍의 얼굴은 정면에 가깝지만 눈은 아래쪽 모닥불과 지면 방향으로 내려가 있으며, 오른쪽 전경의 임정현을 뚫어져라 보는 시선은 아니다. 임정현은 고개를 숙이고 종이 쪽을 향하며, 오른손의 붓끝도 종이를 향한다. 안규홍이 든 낫은 아래로 늘어져 있고 누구를 겨누지 않는다.",
        "built_space": "흙바닥에 돌로 둘러싼 모닥불 하나가 있고, 뒤에는 목책과 좌우 움막 및 중앙 천막이 보인다. 중앙 뒤쪽에는 화덕 형태의 시설 두 개와 오른쪽의 세워진 큰 솥뚜껑이 보여 이전 장소의 주요 요소를 이어간다. 망치 하나는 뒤쪽 돗자리 위에 놓였다. 그러나 안규홍은 모닥불 건너편에 서 있어 두 사람이 앉아 불을 사이에 둔 배치와 다르며, 카메라도 그의 전신과 넓은 마당을 담는다.",
        "entities": "두 인물만 보이며 추가 군중은 없다. 안규홍은 참고와 유사한 청년 한국인 남성 외형, 뒤로 묶은 검은 머리, 낡은 밝은색 저고리와 바지, 허리띠를 갖췄다. 오른손에는 낫이 보이나 옥반지는 확인하기 어렵다. 전경 임정현의 머리·어깨·손과 짙은 소매가 보이며, 붓은 그의 손에 들려 있다. 망치와 낮은 장작불, 종이와 먹 도구가 있고 읽을 수 있는 글자는 보이지 않는다.",
        "hard_violations": [
         "모닥불을 사이에 둔 착석 배치와 달리 안규홍을 일으켜 세워, 지정된 인물의 자세와 공간 배치를 바꾸었다."
        ],
        "physics": "안규홍의 두 발은 지면에 닿고 오른손은 낫을 잡고 있어 몸과 도구의 지지가 분명하다. 임정현의 손가락은 붓대를 잡고 붓끝을 종이 위에 멈춰 두어 중단된 필기 동작이 가능하다. 기록판의 받침은 하단 크롭 밖이라 전체 구조는 확인되지 않지만 공중에 떠 있다고 볼 근거는 없다. 망치는 돗자리 위에, 장작과 돌은 지면에 놓여 있다."
       },
       {
        "label": "A",
        "direction": "안규홍은 정면에 가까운 얼굴로 눈을 약간 오른쪽 전경의 임정현 쪽에 고정하고 있다. 낮게 겨눈 눈과 굳은 입매가 침통한 응시를 만든다. 임정현의 드러난 얼굴 부분은 기록판 쪽으로 숙여져 있고, 손에 잡힌 붓끝은 종이 위를 향한 채 멈춰 있다. 붓을 무기처럼 다른 사람에게 겨누는 오류는 없다.",
        "built_space": "모닥불 하나를 사이에 두고 안규홍이 흙바닥에 가부좌로 앉아 있으며, 임정현의 어깨와 필기하는 손이 오른쪽 전경을 차지한다. 왼쪽 빈 돗자리에는 망치 하나가 홀로 놓였다. 목책, 좌우 움막, 중앙 천막과 뒤쪽 화덕 두 개 및 솥뚜껑이 보여 장소의 연속성은 유지된다. 안규홍이 화면에서 충분히 커져 A보다 미디엄 숏에 가깝다. 불은 얼굴과 필기 손을 잇는 선 아래에 있지만, 큰 기록판과 붓이 하단 및 전경을 상당히 점유한다.",
        "entities": "안규홍과 부분적으로 보이는 임정현 외의 사람은 없다. 안규홍의 청년 한국인 남성 외형, 묶은 검은 머리, 얼굴과 체격, 낡은 밝은색 옷과 허리띠가 참고에 대체로 부합한다. 드러난 손에서는 옥반지가 확인되지 않으며 낫도 화면에서 확인할 수 없다. 임정현의 손과 짙은 소매는 전경의 몸에 연결되고 붓을 잡고 있다. 붓, 종이, 먹 도구, 망치가 실제 물체로 표현되지만 붓은 손에 비해 작게 보이지 않는다. 종이에 먹자국이 선명하게 보이지만 확실히 판독되는 문구는 없다.",
        "hard_violations": [],
        "physics": "안규홍의 엉덩이와 접은 다리는 지면에 놓이고 양손은 무릎 부근에 얹혀 있어 착석 자세가 성립한다. 임정현의 오른손은 붓을 잡고 다른 손은 기록판 가장자리를 받친다. 기록판 아래에는 목재 받침도 일부 보여 지지 없는 부유물은 아니다. 붓끝을 종이 바로 위에 멈춘 자세도 가능하다. 망치는 돗자리에 놓이고 불붙은 장작은 돌 테두리 안에서 서로 기대어 지지된다."
       }
      ]
     },
     "ok": true
    }
   ],
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "slot_winner_match": true,
   "slot_winner": {
    "gemini-pro": "A",
    "gpt-high": "A"
   },
   "route": "cross_slot_agree"
  },
  "dual": {
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "normalized": {
    "A": 2.0,
    "B": 0.929
   },
   "adjusted": {
    "A": 2.0,
    "B": 0.679
   },
   "violations": {
    "B": [
     "[gpt-high] 모닥불을 사이에 둔 착석 배치와 달리 안규홍을 일으켜 세워, 지정된 인물의 자세와 공간 배치를 바꾸었다."
    ]
   },
   "per_model_winner": {
    "gemini-pro": "A",
    "gpt-high": "A"
   },
   "agreed": true
  },
  "totals": {
   "A": 2000,
   "B": 679
  },
  "selected": "A",
  "ranking": [
   "A",
   "B"
  ],
  "verdicts": [
   {
    "label": "A",
    "score": 2000,
    "verdict_ko": "지정된 미디엄 샷 프레이밍과 전경에서 허공에 멈춘 붓의 구도를 정확히 구현했으나, 안규홍의 낫과 옥반지가 누락되었고 읽을 수 없는 글씨 지시에도 불구하고 종이에 획이 너무 선명하게 드러난 점이 아쉽습니다."
   },
   {
    "label": "B",
    "score": 679,
    "verdict_ko": "안규홍이 낫을 들고 있는 점은 맞췄으나, 미디엄 샷 지시를 어기고 풀 샷으로 처리했으며 인물이 서 있고 붓이 종이에 닿아 있어 핵심적인 구도와 상황 연출(Priority 2)에서 크게 실패했습니다.  ★위반: [gpt-high] 모닥불을 사이에 둔 착석 배치와 달리 안규홍을 일으켜 세워, 지정된 인물의 자세와 공간 배치를 바꾸었다."
   }
  ],
  "refs": [
   {
    "label": "PREVIOUS SHOT STILL — a visually related earlier shot of this same place: the location's look, materials, fixed features and lighting mood are LOCKED to this photo; never copy its camera framing. The people visible in that photo are NOT in this shot: never carry their faces, bodies or clothing onto anyone here. Each person in THIS shot is defined solely by the PEOPLE list, their own CHARACTER REFERENCE images and their wardrobe notes. If this photo shows a character who cannot move (dead or unconscious), that character's exact pose, position and orientation are ALSO LOCKED — treat their body as a fixed prop of the set that only the camera moves around.",
    "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/db00517b-165b-4005-a287-f28936b59b95/images/272bd7af-9294-4490-98c6-88a40e9deacc/scene/recipe/S41sh1_sel.png",
    "asset_id": "9d44441b-9724-444d-92a1-4d8d77806821",
    "role": "prev_still"
   },
   {
    "label": "CHARACTER REFERENCE — 안규홍: the exact person appearing in this shot; match face, hair and build exactly.",
    "path": "<bytes:1028384>",
    "asset_id": "3b5894db-ce55-4f34-b966-5e725c3aa952",
    "role": "character_ref"
   }
  ],
  "critique_skipped": true,
  "shot_run_produced": false,
  "shot_run_uid": "06aa142a-70c6-736b-9310-fbfe6b9cf4ac",
  "ref_mode": "prev+엔티티",
  "share_plan": {
   "ref_plan": "prev",
   "prev_anchor_tag": "S41sh1"
  }
 },
 "S41sh8::cine": {
  "applied": true,
  "attempted_at": "2026-09-09T05:36:05.577253+00:00",
  "fingerprint": "8c5fd859913698a5b30c0cc375102bd52f40b587212e0298e532c6a362ab80c0",
  "fingerprint_version": 2,
  "provider": "mai",
  "endpoint": "openrouter/chat-completions",
  "model": "microsoft/mai-image-2.6",
  "pack": "24.202608252115",
  "source_file": "S41sh8_sel.png",
  "source_sha256": "1133ffe6c389ee7176fc2884a55962372b4819c80105790790e7571c3374b20d",
  "file": "S41sh8_cine.png",
  "staged_sha256": "68f7e361b04d456ca77ffcc0c9bc5dd2a09087a4e0181001c123d2c799cf0f10",
  "latency_ms": 26262
 },
 "S41sh12::signage": {
  "fp": "9ea1987f74c5cf62",
  "inscriptions": [],
  "cues": [],
  "dropped": []
 },
 "S41sh12": {
  "input_fingerprint": "00ce9e6fc1f5cf88",
  "prompt": "Create ONE FINAL photorealistic live-action film still of the moment below — early 20th-century Korea, Jeolla region; all people are Korean unless explicitly stated otherwise. TIME OF DAY (lock): night.\n\nSHOT TEXT (authoritative, Korean): 자리에서 벌떡 일어선 안규홍의 꼿꼿하고 거대한 전신.\n\nLOCATION (lock): The edge of the nighttime campfire circle in the mountain camp, beside the fallen blacksmith's empty place. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: wide shot\n- KEY BACKGROUND ELEMENTS: Unoccupied blacksmith's place (Empty, with nobody seated beside it); used as A peripheral gap in the camp composition reinforces the absence without competing with the full figure; Blacksmith's hammer (Left at the unoccupied place) — Seen obliquely beside the empty place; used as A small peripheral narrative detail, reserved for the following camera movement.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: The low-burning campfire provides restrained illumination within the night setting, preserving subdued saturation and sober, naturalistic contrast.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The people visible in that photo are NOT in this shot: never carry their faces, bodies or clothing onto anyone here. Each person in THIS shot is defined solely by the PEOPLE list, their own CHARACTER REFERENCE images and their wardrobe notes. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nPREVIOUS STILL USAGE (follow exactly — what to take from the attached still and what to exclude; it governs place and objects only, never who is in this shot or how the camera sees them): Take the low fire, dirt ground, and nearby camp structures visible in the reference. Exclude supplies from the earlier unloading scene and any sparks unique to the reference instant.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The hammer remains at the empty place and the death record now includes a name; the campfire burns low. 안규홍: He has risen, retaining his sickle and jade ring.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 안규홍 (한국인 남성, 청년 나이대, 검은 머리, 단단히 묶은 머리) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.",
  "roll_prompts": {
   "A": "Create ONE FINAL photorealistic live-action film still of the moment below — early 20th-century Korea, Jeolla region; all people are Korean unless explicitly stated otherwise. TIME OF DAY (lock): night.\n\nSHOT TEXT (authoritative, Korean): 자리에서 벌떡 일어선 안규홍의 꼿꼿하고 거대한 전신.\n\nLOCATION (lock): The edge of the nighttime campfire circle in the mountain camp, beside the fallen blacksmith's empty place. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: wide shot\n- KEY BACKGROUND ELEMENTS: Unoccupied blacksmith's place (Empty, with nobody seated beside it); used as A peripheral gap in the camp composition reinforces the absence without competing with the full figure; Blacksmith's hammer (Left at the unoccupied place) — Seen obliquely beside the empty place; used as A small peripheral narrative detail, reserved for the following camera movement.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: The low-burning campfire provides restrained illumination within the night setting, preserving subdued saturation and sober, naturalistic contrast.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The people visible in that photo are NOT in this shot: never carry their faces, bodies or clothing onto anyone here. Each person in THIS shot is defined solely by the PEOPLE list, their own CHARACTER REFERENCE images and their wardrobe notes. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nPREVIOUS STILL USAGE (follow exactly — what to take from the attached still and what to exclude; it governs place and objects only, never who is in this shot or how the camera sees them): Take the low fire, dirt ground, and nearby camp structures visible in the reference. Exclude supplies from the earlier unloading scene and any sparks unique to the reference instant.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The hammer remains at the empty place and the death record now includes a name; the campfire burns low. 안규홍: He has risen, retaining his sickle and jade ring.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 안규홍 (한국인 남성, 청년 나이대, 검은 머리, 단단히 묶은 머리) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.",
   "B": "Create ONE FINAL photorealistic live-action film still of the moment below — early 20th-century Korea, Jeolla region; all people are Korean unless explicitly stated otherwise. TIME OF DAY (lock): night.\n\nSHOT TEXT (authoritative, Korean): 자리에서 벌떡 일어선 안규홍의 꼿꼿하고 거대한 전신.\n\nLOCATION (lock): The edge of the nighttime campfire circle in the mountain camp, beside the fallen blacksmith's empty place. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: wide shot\n- KEY BACKGROUND ELEMENTS: Unoccupied blacksmith's place (Empty, with nobody seated beside it); used as A peripheral gap in the camp composition reinforces the absence without competing with the full figure; Blacksmith's hammer (Left at the unoccupied place) — Seen obliquely beside the empty place; used as A small peripheral narrative detail, reserved for the following camera movement.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: The low-burning campfire provides restrained illumination within the night setting, preserving subdued saturation and sober, naturalistic contrast.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The people visible in that photo are NOT in this shot: never carry their faces, bodies or clothing onto anyone here. Each person in THIS shot is defined solely by the PEOPLE list, their own CHARACTER REFERENCE images and their wardrobe notes. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nPREVIOUS STILL USAGE (follow exactly — what to take from the attached still and what to exclude; it governs place and objects only, never who is in this shot or how the camera sees them): Take the low fire, dirt ground, and nearby camp structures visible in the reference. Exclude supplies from the earlier unloading scene and any sparks unique to the reference instant.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The hammer remains at the empty place and the death record now includes a name; the campfire burns low. 안규홍: He has risen, retaining his sickle and jade ring.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 안규홍 (한국인 남성, 청년 나이대, 검은 머리, 단단히 묶은 머리) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.\n\nCOMPOSITION VARIATION (this candidate only): keep the place, the people and the moment exactly as specified, and keep from the CAMERA & FRAME contract its subject scale, screen placement and key background elements exactly. For this candidate only, its camera angle and height are deliberately left open: choose a camera position distinctly different from the obvious one, as a film director picking a second setup on the same blocking. Stay true to any explicit framing the SHOT TEXT itself dictates."
  },
  "shot_run_spend_attempt_count": 1,
  "readings": [
   {
    "label": "A",
    "direction": "안규홍의 시선은 카메라 우측 상단의 프레임 밖을 향하고 있습니다.",
    "built_space": "참조 이미지와 동일하게 야영지의 구조물과 흙바닥, 모닥불을 둘러싼 형태를 보여줍니다. 화면 좌측 전경의 바닥에 망치가 놓여 있습니다.",
    "entities": "안규홍(참조 이미지의 외형과 일치하며 오른손에 낫을 쥐고 있음)이 중앙에 서 있고, 프롬프트에서 배제하라고 명시한 다수의 야영지 인물들이 모닥불 주변에 앉아 있습니다.",
    "hard_violations": [
     "[gemini-pro] 프롬프트에 명시되지 않은 다수의 추가 인물 등장 (Shot text에 없는 인물 추가 및 이전 장면의 인물 유지 금지 조건 위반)",
     "[gpt-high] 안규홍만 보여야 하는 장면에 다수의 다른 남성과 부분 신체를 추가했다.",
     "[gpt-high] 아무도 곁에 앉지 않아야 할 대장장이의 빈자리와 망치 바로 옆에 다른 사람을 앉혔다."
    ],
    "physics": "안규홍은 바닥에 두 발을 딛고 서 있으며, 오른손이 낫의 손잡이를 정상적으로 쥐고 지탱하고 있습니다. 다른 인물들 역시 바닥에 앉아 무게를 지탱하고 있습니다."
   },
   {
    "label": "B",
    "direction": "안규홍의 시선은 정면의 카메라 렌즈를 향하고 있습니다.",
    "built_space": "야간의 산속 야영지 배경으로 뒤편에 막사와 목책이 보이며, 안규홍의 뒤로 낮게 타오르는 모닥불이 위치합니다. 좌측에는 빈 자리와 함께 바닥(작은 깔개 위)에 놓인 망치가 보입니다.",
    "entities": "안규홍(참조 이미지의 외형과 복장에 일치하며 단독으로 등장함)이 전신으로 잡혀 있습니다. 허리띠에 낫을 차고 있고 왼손 검지에 옥반지가 확인되며, 주변에 다른 인물은 전혀 없습니다.",
    "hard_violations": [],
    "physics": "안규홍은 두 발을 흙바닥에 딛고 서서 체중을 지탱하고 있습니다. 낫은 허리띠에 꽂혀 지지되고 있으며, 망치는 바닥에 놓여 중력을 받고 있습니다."
   }
  ],
  "cross_model_order": {
   "policy": "cross_model_order_v2_slot0_fwd_slot1_rev_parallel2",
   "slots": [
    {
     "model": "gemini-pro",
     "order": "forward",
     "display_to_canonical": {
      "A": "A",
      "B": "B"
     },
     "raw": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 9,
        "verdict_ko": "프롬프트의 지시대로 이전 장면의 다른 인물들을 모두 완벽하게 배제하고 안규홍 단 한 명만을 와이드 샷으로 담아냈으며, 지정된 소품(낫, 옥반지, 빈 자리의 망치)의 상태를 훌륭하게 구현했습니다."
       },
       {
        "label": "A",
        "score": 2,
        "verdict_ko": "주인공의 외형과 자세는 잘 표현했으나, 프롬프트에서 절대 추가하지 말라고 강력하게 명시한 다수의 주변 인물들을 그대로 등장시킨 치명적인 하드 위반을 범했습니다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "안규홍의 시선은 카메라 우측 상단의 프레임 밖을 향하고 있습니다.",
        "built_space": "참조 이미지와 동일하게 야영지의 구조물과 흙바닥, 모닥불을 둘러싼 형태를 보여줍니다. 화면 좌측 전경의 바닥에 망치가 놓여 있습니다.",
        "entities": "안규홍(참조 이미지의 외형과 일치하며 오른손에 낫을 쥐고 있음)이 중앙에 서 있고, 프롬프트에서 배제하라고 명시한 다수의 야영지 인물들이 모닥불 주변에 앉아 있습니다.",
        "hard_violations": [
         "프롬프트에 명시되지 않은 다수의 추가 인물 등장 (Shot text에 없는 인물 추가 및 이전 장면의 인물 유지 금지 조건 위반)"
        ],
        "physics": "안규홍은 바닥에 두 발을 딛고 서 있으며, 오른손이 낫의 손잡이를 정상적으로 쥐고 지탱하고 있습니다. 다른 인물들 역시 바닥에 앉아 무게를 지탱하고 있습니다."
       },
       {
        "label": "B",
        "direction": "안규홍의 시선은 정면의 카메라 렌즈를 향하고 있습니다.",
        "built_space": "야간의 산속 야영지 배경으로 뒤편에 막사와 목책이 보이며, 안규홍의 뒤로 낮게 타오르는 모닥불이 위치합니다. 좌측에는 빈 자리와 함께 바닥(작은 깔개 위)에 놓인 망치가 보입니다.",
        "entities": "안규홍(참조 이미지의 외형과 복장에 일치하며 단독으로 등장함)이 전신으로 잡혀 있습니다. 허리띠에 낫을 차고 있고 왼손 검지에 옥반지가 확인되며, 주변에 다른 인물은 전혀 없습니다.",
        "hard_violations": [],
        "physics": "안규홍은 두 발을 흙바닥에 딛고 서서 체중을 지탱하고 있습니다. 낫은 허리띠에 꽂혀 지지되고 있으며, 망치는 바닥에 놓여 중력을 받고 있습니다."
       }
      ],
      "all_candidates_fail": false
     },
     "normalized": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 9,
        "verdict_ko": "프롬프트의 지시대로 이전 장면의 다른 인물들을 모두 완벽하게 배제하고 안규홍 단 한 명만을 와이드 샷으로 담아냈으며, 지정된 소품(낫, 옥반지, 빈 자리의 망치)의 상태를 훌륭하게 구현했습니다."
       },
       {
        "label": "A",
        "score": 2,
        "verdict_ko": "주인공의 외형과 자세는 잘 표현했으나, 프롬프트에서 절대 추가하지 말라고 강력하게 명시한 다수의 주변 인물들을 그대로 등장시킨 치명적인 하드 위반을 범했습니다."
       }
      ],
      "all_candidates_fail": false,
      "readings": [
       {
        "label": "A",
        "direction": "안규홍의 시선은 카메라 우측 상단의 프레임 밖을 향하고 있습니다.",
        "built_space": "참조 이미지와 동일하게 야영지의 구조물과 흙바닥, 모닥불을 둘러싼 형태를 보여줍니다. 화면 좌측 전경의 바닥에 망치가 놓여 있습니다.",
        "entities": "안규홍(참조 이미지의 외형과 일치하며 오른손에 낫을 쥐고 있음)이 중앙에 서 있고, 프롬프트에서 배제하라고 명시한 다수의 야영지 인물들이 모닥불 주변에 앉아 있습니다.",
        "hard_violations": [
         "프롬프트에 명시되지 않은 다수의 추가 인물 등장 (Shot text에 없는 인물 추가 및 이전 장면의 인물 유지 금지 조건 위반)"
        ],
        "physics": "안규홍은 바닥에 두 발을 딛고 서 있으며, 오른손이 낫의 손잡이를 정상적으로 쥐고 지탱하고 있습니다. 다른 인물들 역시 바닥에 앉아 무게를 지탱하고 있습니다."
       },
       {
        "label": "B",
        "direction": "안규홍의 시선은 정면의 카메라 렌즈를 향하고 있습니다.",
        "built_space": "야간의 산속 야영지 배경으로 뒤편에 막사와 목책이 보이며, 안규홍의 뒤로 낮게 타오르는 모닥불이 위치합니다. 좌측에는 빈 자리와 함께 바닥(작은 깔개 위)에 놓인 망치가 보입니다.",
        "entities": "안규홍(참조 이미지의 외형과 복장에 일치하며 단독으로 등장함)이 전신으로 잡혀 있습니다. 허리띠에 낫을 차고 있고 왼손 검지에 옥반지가 확인되며, 주변에 다른 인물은 전혀 없습니다.",
        "hard_violations": [],
        "physics": "안규홍은 두 발을 흙바닥에 딛고 서서 체중을 지탱하고 있습니다. 낫은 허리띠에 꽂혀 지지되고 있으며, 망치는 바닥에 놓여 중력을 받고 있습니다."
       }
      ]
     },
     "ok": true
    },
    {
     "model": "gpt-high",
     "order": "reverse",
     "display_to_canonical": {
      "A": "B",
      "B": "A"
     },
     "raw": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 7,
        "verdict_ko": "안규홍만의 꼿꼿한 전신과 주변부의 빈자리·망치를 구현했지만, 정면 응시가 경직되고 모닥불이 지정된 낮은 불보다 강하다."
       },
       {
        "label": "B",
        "score": 2,
        "verdict_ko": "전신과 낮은 모닥불은 맞지만, 금지된 군중을 추가하고 빈자리 바로 곁에도 사람을 앉혀 핵심 인물·배치 조건을 위반한다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "안규홍은 몸과 얼굴을 정면으로 두고 카메라 쪽을 응시한다. 특정 응시 대상은 지시되지 않았으나 자연스러운 순간 연기보다는 정면 초상에 가깝다. 낫은 허리띠에 꽂혀 곡선 날이 배 앞에서 위로 향하며, 누구를 겨누지는 않는다. 왼쪽 빈자리의 망치는 비스듬히 보이며 손잡이는 화면 왼쪽으로 놓였다.",
        "built_space": "흙바닥과 목책 뒤로 왼쪽 움막 하나, 중앙 천막 하나, 오른쪽 긴 초가 구조물 하나가 보여 참고 장소의 주요 재료와 구조를 유지한다. 돌로 두른 모닥불은 하나이며 인물 바로 뒤에 있다. 안규홍은 그 앞쪽 가장자리에 전신이 보이도록 서 있다. 왼쪽 주변부에 아무도 앉지 않은 돗자리와 망치 하나가 있지만, 참고의 바위 옆 빈자리와는 세부 구성이 달라졌다. 반사면이나 불가능한 반사는 없다.",
        "entities": "인물은 안규홍 한 명뿐이며, 청년 한국인 남성 설정에 부합하는 외모와 뒤로 단단히 묶은 검은 머리다. 얼굴은 참고와 대체로 유사하고, 때 탄 밝은 저고리·바지·천 허리띠·짚신을 유지한다. 낫은 허리에 있고 망치는 빈자리 쪽에 남아 있다. 옥반지는 손의 작은 크기 때문에 식별하기 어렵다. 사망 기록은 보이지 않아 이름 추가 여부를 확인할 수 없다. 읽을 수 있는 글자는 없다. 밤이지만 달과 별이 강조되고 불꽃은 참고의 잔불보다 높다.",
        "hard_violations": [],
        "physics": "두 짚신이 흙바닥에 닿아 선 몸을 지탱하고 있으며, 다리를 벌리고 곧게 선 자세는 실제로 가능하다. 낫은 허리띠에 끼워져 지지되고 망치는 돗자리 위에 놓여 있다. 모닥불의 장작과 돌도 지면에 놓여 있어 공중에 뜬 물체는 없다. 다만 벌떡 일어난 직후의 체중 이동보다는 이미 자세를 정돈한 모습이다."
       },
       {
        "label": "B",
        "direction": "안규홍은 몸을 약간 틀고 화면 오른쪽 위의 화면 밖을 바라본다. 오른손에 든 낫은 허벅지 옆으로 내려가 있으며 날 끝은 아래쪽을 향하고, 다른 사람을 겨누지 않는다. 주변 남성들은 대체로 아래쪽 땅이나 불 쪽을 본다. 왼쪽 망치의 손잡이는 화면 왼쪽 위로 뻗어 있다.",
        "built_space": "흙바닥, 뒤쪽 목책, 왼쪽 초가 구조물과 오른쪽 천막이 보여 참고 장소의 재료는 잘 이어진다. 모닥불 하나가 오른쪽 아래에 있고 안규홍은 그 왼쪽 가장자리에 서 있다. 그러나 안규홍 외에 약 15명의 앉은 남성과 부분 신체가 둘레를 채운다. 망치 하나는 왼쪽 아래 땅에 있지만 그 바로 곁에도 사람이 앉아 있어, 아무도 곁에 앉지 않은 주변부 빈자리라는 배치가 성립하지 않는다. 반사 관련 문제는 없다.",
        "entities": "중앙 인물은 청년 한국인 남성 설정에 부합하는 외모, 묶은 검은 머리, 밝고 낡은 전통 복식과 짚신을 갖추며 참고 안규홍과 대체로 닮았다. 낫과 망치가 있고 모닥불도 낮게 탄다. 옥반지는 명확히 식별되지 않으며 사망 기록도 보이지 않는다. 하지만 허용 인물 목록에 없는 다수의 남성이 추가되어 이전 장면의 군중을 배제하라는 조건을 어긴다. 읽을 수 있는 글자는 없다.",
        "hard_violations": [
         "안규홍만 보여야 하는 장면에 다수의 다른 남성과 부분 신체를 추가했다.",
         "아무도 곁에 앉지 않아야 할 대장장이의 빈자리와 망치 바로 옆에 다른 사람을 앉혔다."
        ],
        "physics": "안규홍의 두 발은 지면에 닿고, 오른손은 낫 손잡이를 잡아 무게를 지탱한다. 곧게 선 몸과 내려 든 낫의 자세는 물리적으로 가능하다. 주변 인물들도 엉덩이와 접은 다리를 땅에 대고 앉아 있다. 망치와 화덕 돌, 장작은 지면에 놓여 있으며 지지 없이 떠 있는 물체는 보이지 않는다."
       }
      ],
      "all_candidates_fail": false
     },
     "normalized": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 7,
        "verdict_ko": "안규홍만의 꼿꼿한 전신과 주변부의 빈자리·망치를 구현했지만, 정면 응시가 경직되고 모닥불이 지정된 낮은 불보다 강하다."
       },
       {
        "label": "A",
        "score": 2,
        "verdict_ko": "전신과 낮은 모닥불은 맞지만, 금지된 군중을 추가하고 빈자리 바로 곁에도 사람을 앉혀 핵심 인물·배치 조건을 위반한다."
       }
      ],
      "all_candidates_fail": false,
      "readings": [
       {
        "label": "B",
        "direction": "안규홍은 몸과 얼굴을 정면으로 두고 카메라 쪽을 응시한다. 특정 응시 대상은 지시되지 않았으나 자연스러운 순간 연기보다는 정면 초상에 가깝다. 낫은 허리띠에 꽂혀 곡선 날이 배 앞에서 위로 향하며, 누구를 겨누지는 않는다. 왼쪽 빈자리의 망치는 비스듬히 보이며 손잡이는 화면 왼쪽으로 놓였다.",
        "built_space": "흙바닥과 목책 뒤로 왼쪽 움막 하나, 중앙 천막 하나, 오른쪽 긴 초가 구조물 하나가 보여 참고 장소의 주요 재료와 구조를 유지한다. 돌로 두른 모닥불은 하나이며 인물 바로 뒤에 있다. 안규홍은 그 앞쪽 가장자리에 전신이 보이도록 서 있다. 왼쪽 주변부에 아무도 앉지 않은 돗자리와 망치 하나가 있지만, 참고의 바위 옆 빈자리와는 세부 구성이 달라졌다. 반사면이나 불가능한 반사는 없다.",
        "entities": "인물은 안규홍 한 명뿐이며, 청년 한국인 남성 설정에 부합하는 외모와 뒤로 단단히 묶은 검은 머리다. 얼굴은 참고와 대체로 유사하고, 때 탄 밝은 저고리·바지·천 허리띠·짚신을 유지한다. 낫은 허리에 있고 망치는 빈자리 쪽에 남아 있다. 옥반지는 손의 작은 크기 때문에 식별하기 어렵다. 사망 기록은 보이지 않아 이름 추가 여부를 확인할 수 없다. 읽을 수 있는 글자는 없다. 밤이지만 달과 별이 강조되고 불꽃은 참고의 잔불보다 높다.",
        "hard_violations": [],
        "physics": "두 짚신이 흙바닥에 닿아 선 몸을 지탱하고 있으며, 다리를 벌리고 곧게 선 자세는 실제로 가능하다. 낫은 허리띠에 끼워져 지지되고 망치는 돗자리 위에 놓여 있다. 모닥불의 장작과 돌도 지면에 놓여 있어 공중에 뜬 물체는 없다. 다만 벌떡 일어난 직후의 체중 이동보다는 이미 자세를 정돈한 모습이다."
       },
       {
        "label": "A",
        "direction": "안규홍은 몸을 약간 틀고 화면 오른쪽 위의 화면 밖을 바라본다. 오른손에 든 낫은 허벅지 옆으로 내려가 있으며 날 끝은 아래쪽을 향하고, 다른 사람을 겨누지 않는다. 주변 남성들은 대체로 아래쪽 땅이나 불 쪽을 본다. 왼쪽 망치의 손잡이는 화면 왼쪽 위로 뻗어 있다.",
        "built_space": "흙바닥, 뒤쪽 목책, 왼쪽 초가 구조물과 오른쪽 천막이 보여 참고 장소의 재료는 잘 이어진다. 모닥불 하나가 오른쪽 아래에 있고 안규홍은 그 왼쪽 가장자리에 서 있다. 그러나 안규홍 외에 약 15명의 앉은 남성과 부분 신체가 둘레를 채운다. 망치 하나는 왼쪽 아래 땅에 있지만 그 바로 곁에도 사람이 앉아 있어, 아무도 곁에 앉지 않은 주변부 빈자리라는 배치가 성립하지 않는다. 반사 관련 문제는 없다.",
        "entities": "중앙 인물은 청년 한국인 남성 설정에 부합하는 외모, 묶은 검은 머리, 밝고 낡은 전통 복식과 짚신을 갖추며 참고 안규홍과 대체로 닮았다. 낫과 망치가 있고 모닥불도 낮게 탄다. 옥반지는 명확히 식별되지 않으며 사망 기록도 보이지 않는다. 하지만 허용 인물 목록에 없는 다수의 남성이 추가되어 이전 장면의 군중을 배제하라는 조건을 어긴다. 읽을 수 있는 글자는 없다.",
        "hard_violations": [
         "안규홍만 보여야 하는 장면에 다수의 다른 남성과 부분 신체를 추가했다.",
         "아무도 곁에 앉지 않아야 할 대장장이의 빈자리와 망치 바로 옆에 다른 사람을 앉혔다."
        ],
        "physics": "안규홍의 두 발은 지면에 닿고, 오른손은 낫 손잡이를 잡아 무게를 지탱한다. 곧게 선 몸과 내려 든 낫의 자세는 물리적으로 가능하다. 주변 인물들도 엉덩이와 접은 다리를 땅에 대고 앉아 있다. 망치와 화덕 돌, 장작은 지면에 놓여 있으며 지지 없이 떠 있는 물체는 보이지 않는다."
       }
      ]
     },
     "ok": true
    }
   ],
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "slot_winner_match": true,
   "slot_winner": {
    "gemini-pro": "B",
    "gpt-high": "B"
   },
   "route": "cross_slot_agree"
  },
  "dual": {
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "normalized": {
    "A": 0.508,
    "B": 2.0
   },
   "adjusted": {
    "A": 0.258,
    "B": 2.0
   },
   "violations": {
    "A": [
     "[gemini-pro] 프롬프트에 명시되지 않은 다수의 추가 인물 등장 (Shot text에 없는 인물 추가 및 이전 장면의 인물 유지 금지 조건 위반)",
     "[gpt-high] 안규홍만 보여야 하는 장면에 다수의 다른 남성과 부분 신체를 추가했다.",
     "[gpt-high] 아무도 곁에 앉지 않아야 할 대장장이의 빈자리와 망치 바로 옆에 다른 사람을 앉혔다."
    ]
   },
   "per_model_winner": {
    "gemini-pro": "B",
    "gpt-high": "B"
   },
   "agreed": true
  },
  "totals": {
   "B": 2000,
   "A": 258
  },
  "selected": "B",
  "ranking": [
   "B",
   "A"
  ],
  "verdicts": [
   {
    "label": "B",
    "score": 2000,
    "verdict_ko": "프롬프트의 지시대로 이전 장면의 다른 인물들을 모두 완벽하게 배제하고 안규홍 단 한 명만을 와이드 샷으로 담아냈으며, 지정된 소품(낫, 옥반지, 빈 자리의 망치)의 상태를 훌륭하게 구현했습니다."
   },
   {
    "label": "A",
    "score": 258,
    "verdict_ko": "주인공의 외형과 자세는 잘 표현했으나, 프롬프트에서 절대 추가하지 말라고 강력하게 명시한 다수의 주변 인물들을 그대로 등장시킨 치명적인 하드 위반을 범했습니다.  ★위반: [gemini-pro] 프롬프트에 명시되지 않은 다수의 추가 인물 등장 (Shot text에 없는 인물 추가 및 이전 장면의 인물 유지 금지 조건 위반) / [gpt-high] 안규홍만 보여야 하는 장면에 다수의 다른 남성과 부분 신체를 추가했다. / [gpt-high] 아무도 곁에 앉지 않아야 할 대장장이의 빈자리와 망치 바로 옆에 다른 사람을 앉혔다."
   }
  ],
  "refs": [
   {
    "label": "PREVIOUS SHOT STILL — a visually related earlier shot of this same place: the location's look, materials, fixed features and lighting mood are LOCKED to this photo; never copy its camera framing. The people visible in that photo are NOT in this shot: never carry their faces, bodies or clothing onto anyone here. Each person in THIS shot is defined solely by the PEOPLE list, their own CHARACTER REFERENCE images and their wardrobe notes. If this photo shows a character who cannot move (dead or unconscious), that character's exact pose, position and orientation are ALSO LOCKED — treat their body as a fixed prop of the set that only the camera moves around.",
    "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/db00517b-165b-4005-a287-f28936b59b95/images/272bd7af-9294-4490-98c6-88a40e9deacc/scene/recipe/S41sh1_sel.png",
    "asset_id": "9d44441b-9724-444d-92a1-4d8d77806821",
    "role": "prev_still"
   },
   {
    "label": "CHARACTER REFERENCE — 안규홍: the exact person appearing in this shot; match face, hair and build exactly.",
    "path": "<bytes:1028384>",
    "asset_id": "3b5894db-ce55-4f34-b966-5e725c3aa952",
    "role": "character_ref"
   }
  ],
  "critique_skipped": true,
  "shot_run_produced": false,
  "shot_run_uid": "06aa142a-7268-77c2-aeb3-1eacf27c7745",
  "ref_mode": "prev+엔티티",
  "share_plan": {
   "ref_plan": "prev",
   "prev_anchor_tag": "S41sh1"
  }
 },
 "S41sh12::cine": {
  "applied": true,
  "attempted_at": "2026-09-09T05:37:25.489553+00:00",
  "fingerprint": "f8be92fd86ed412fc67ee4958e8956c9cde2d6d2a5b7f8393aad96c53c5ac9d9",
  "fingerprint_version": 2,
  "provider": "mai",
  "endpoint": "openrouter/chat-completions",
  "model": "microsoft/mai-image-2.6",
  "pack": "24.202608252115",
  "source_file": "S41sh12_sel.png",
  "source_sha256": "a91896bc060687189a698a318ff6864011b1d3d6668e1b5de2df825787b0e530",
  "file": "S41sh12_cine.png",
  "staged_sha256": "5ed6b5a83d30137cc9e69dc8e9af74e713308717bd2a7b8311d5baeec316b7bf",
  "latency_ms": 26940
 },
 "S42sh4::signage": {
  "fp": "28389b9e1c7bbaa6",
  "inscriptions": [],
  "cues": [],
  "dropped": []
 },
 "era_assess::7f423dbf7fadfd3d": {
  "subjects": [],
  "subject_text": "관청 뒤편 담장 골목\n관청 뒤편의 높은 담장을 따라 이어지는 좁은 골목. 담 아래에 짙은 그늘이 드리우고, 길이 건물 뒤로 길게 뻗는다.",
  "identity": "canonical",
  "scope_id": "L31",
  "scope_role": "location_exterior",
  "scope_sha": "08a946ec2fb9209d"
 },
 "groupbg::office_rear_alley": {
  "input_fingerprint": "bb7bb34dd265f66a",
  "meta": {
   "model": "gpt-image-2",
   "size": "1536x864",
   "pack": "11.202607220237",
   "contract": "bgfirst_full_v3",
   "group_sig": {
    "key": "office_rear_alley",
    "tags": [
     "S42sh12",
     "S42sh4",
     "S42sh9"
    ]
   },
   "context_sig": "6ecc0c2734a59be3"
  },
  "prompt": "Create ONE empty live-action location background photograph — NO PEOPLE, no figures, no body parts, no silhouettes, no shadows or reflections of people anywhere.\n\"Empty\" means no people only: KEEP the location's inherent occupants and stock that define the place — animals in an animal shelter, pen or farm, goods and produce in a market, moored boats in a harbour — unless the shot text explicitly removes them.\nThe FIRST attached image is a thin-line storyboard sketch of a shot that happens at this location: use it ONLY as spatial evidence — what this place contains, how its ground, structures and landmarks are arranged and proportioned. Ignore the sketched people and arrows entirely, and do NOT copy its line style: render a fully photographic, physically plausible real place that fits THE LOCATION text below.\nHUMAN-SCALE CALIBRATION: derive every structure's true size from human-scale elements — a door ≈ 2m, a window ≈ 1–1.5m wide, one storey ≈ 2.5–3m; never inflate a small structure or shrink a large one.\n\nTHE LOCATION — early 20th-century Korea, Jeolla region; all people are Korean unless explicitly stated otherwise: The dim outdoor alley directly along the rear wall of a government compound at night.\n\nLOCATION DETAIL (descriptive evidence about this place from the production's location records — evidence, not a staging order): use it to understand what kind of place this is and what it permanently contains. Carry over ONLY its enduring physical features — layout, structures, ground and materials, fixed equipment and installations. Do NOT treat any lighting, weather or time-of-day wording, momentary object states, depicted events or subjective impressions in it as fixed facts of the place; the TIME OF DAY lock below is the sole authority for time and lighting.\n관청 뒤편 담장 골목: 높은 관청 담벼락을 따라 나 있는 어둑하고 비좁은 후미진 흙길 골목. (특징: 높게 축조된 관청 외부 담장; 어둠이 깔린 좁은 흙길 바닥)\n\nSCENE EVIDENCE (verbatim quotes from the screenplay about this place — treat them as evidence of what the location physically contains and looks like; stage the PLACE those moments happen in, but do NOT depict the momentary actions, people or staged props themselves):\n- S#42. 관청 뒤편 골목 / 밤\n- 어둑한 관청 담장 아래.\n박상현이 서성인다.\n\nTIME OF DAY (lock): night.\n\nRender ONE photorealistic empty location photograph, 16:9, neutral enough that every shot of this place can be staged from it later. No readable writing anywhere: surfaces that would carry writing may be present, but stage any wording out of legibility — an oblique angle, distance, shallow focus. No captions, watermarks or overlay text.",
  "effective_prompt": "Create ONE empty live-action location background photograph — NO PEOPLE, no figures, no body parts, no silhouettes, no shadows or reflections of people anywhere.\n\"Empty\" means no people only: KEEP the location's inherent occupants and stock that define the place — animals in an animal shelter, pen or farm, goods and produce in a market, moored boats in a harbour — unless the shot text explicitly removes them.\nThe FIRST attached image is a thin-line storyboard sketch of a shot that happens at this location: use it ONLY as spatial evidence — what this place contains, how its ground, structures and landmarks are arranged and proportioned. Ignore the sketched people and arrows entirely, and do NOT copy its line style: render a fully photographic, physically plausible real place that fits THE LOCATION text below.\nHUMAN-SCALE CALIBRATION: derive every structure's true size from human-scale elements — a door ≈ 2m, a window ≈ 1–1.5m wide, one storey ≈ 2.5–3m; never inflate a small structure or shrink a large one.\n\nTHE LOCATION — early 20th-century Korea, Jeolla region; all people are Korean unless explicitly stated otherwise: The dim outdoor alley directly along the rear wall of a government compound at night.\n\nLOCATION DETAIL (descriptive evidence about this place from the production's location records — evidence, not a staging order): use it to understand what kind of place this is and what it permanently contains. Carry over ONLY its enduring physical features — layout, structures, ground and materials, fixed equipment and installations. Do NOT treat any lighting, weather or time-of-day wording, momentary object states, depicted events or subjective impressions in it as fixed facts of the place; the TIME OF DAY lock below is the sole authority for time and lighting.\n관청 뒤편 담장 골목: 높은 관청 담벼락을 따라 나 있는 어둑하고 비좁은 후미진 흙길 골목. (특징: 높게 축조된 관청 외부 담장; 어둠이 깔린 좁은 흙길 바닥)\n\nSCENE EVIDENCE (verbatim quotes from the screenplay about this place — treat them as evidence of what the location physically contains and looks like; stage the PLACE those moments happen in, but do NOT depict the momentary actions, people or staged props themselves):\n- S#42. 관청 뒤편 골목 / 밤\n- 어둑한 관청 담장 아래.\n박상현이 서성인다.\n\nTIME OF DAY (lock): night.\n\nRender ONE photorealistic empty location photograph, 16:9, neutral enough that every shot of this place can be staged from it later. No readable writing anywhere: surfaces that would carry writing may be present, but stage any wording out of legibility — an oblique angle, distance, shallow focus. No captions, watermarks or overlay text.",
  "bg_path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/db00517b-165b-4005-a287-f28936b59b95/images/272bd7af-9294-4490-98c6-88a40e9deacc/scene/recipe/groupbg_office_rear_alley_50e959.png",
  "asset_id": "6ec5270c-a39a-4524-86d4-275694646492",
  "input_asset_ids": [
   "adbe5aa6-eb3b-4267-94b9-472eeb2efec2"
  ],
  "origin_tag": "S42sh4",
  "place_text": "The dim outdoor alley directly along the rear wall of a government compound at night.",
  "origin_inputs": {
   "place_text": "The dim outdoor alley directly along the rear wall of a government compound at night.",
   "time_of_day_en": "night",
   "conti_asset_id": "adbe5aa6-eb3b-4267-94b9-472eeb2efec2"
  }
 },
 "S42sh4::bgfirst_bg": {
  "input_fingerprint": "7663ac7356c1a932",
  "prompt": "Create the EMPTY BACKGROUND PLATE for one film shot — NO PEOPLE, no figures, no body parts, no sketch lines, no arrows anywhere.\n\"Empty\" means no people only: KEEP the location's inherent occupants and stock that define the place — animals in an animal shelter, pen or farm, goods in a market, moored boats in a harbour — unless the shot text explicitly removes them.\nThe FIRST attached image is a thin-line storyboard sketch: use ONLY its camera angle, horizon, perspective and the placement/size of buildings and set masses — ignore the sketched people and arrows entirely. The SECOND attached image (LOCATION PHOTOGRAPH) is the real place: take its architecture, materials, signage and fixed features, and RE-PROJECT them into the sketch's camera. If the photograph's camera differs from the sketch's, the sketch's camera wins.\nHUMAN-SCALE CALIBRATION: derive every structure's true size from human-scale elements — a door ≈ 2m, a window ≈ 1–1.5m wide, one storey ≈ 2.5–3m; never inflate a small structure or shrink a large one.\n\nSHOT TEXT this background must serve (Korean): 박상현을 향해 시선을 고정한 채 비웃듯 한쪽 입꼬리를 비틀어 올린 배상도의 얼굴.\n\nLOCATION (lock): The dim outdoor alley directly along the rear wall of a government compound at night.\n\nTIME OF DAY (lock): night.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: close-up\n- KEY BACKGROUND ELEMENTS: Government-office boundary wall (Present behind the encounter) — Its alley-facing side recedes obliquely behind 배상도; used as A softly resolved background plane maintains the established side of the conversation.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Keep the nighttime alley dim, with restrained saturation and enough naturalistic tonal separation to reveal the asymmetric smile without inventing a visible light source.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nRender ONE photorealistic empty location photograph, 16:9, that this shot can be staged inside later. No readable writing anywhere: surfaces that would carry writing may be present, but stage any wording out of legibility — an oblique angle, distance, shallow focus. No captions, watermarks or overlay text.",
  "effective_prompt": "Create the EMPTY BACKGROUND PLATE for one film shot — NO PEOPLE, no figures, no body parts, no sketch lines, no arrows anywhere.\n\"Empty\" means no people only: KEEP the location's inherent occupants and stock that define the place — animals in an animal shelter, pen or farm, goods in a market, moored boats in a harbour — unless the shot text explicitly removes them.\nThe FIRST attached image is a thin-line storyboard sketch: use ONLY its camera angle, horizon, perspective and the placement/size of buildings and set masses — ignore the sketched people and arrows entirely. The SECOND attached image (LOCATION PHOTOGRAPH) is the real place: take its architecture, materials, signage and fixed features, and RE-PROJECT them into the sketch's camera. If the photograph's camera differs from the sketch's, the sketch's camera wins.\nHUMAN-SCALE CALIBRATION: derive every structure's true size from human-scale elements — a door ≈ 2m, a window ≈ 1–1.5m wide, one storey ≈ 2.5–3m; never inflate a small structure or shrink a large one.\n\nSHOT TEXT this background must serve (Korean): 박상현을 향해 시선을 고정한 채 비웃듯 한쪽 입꼬리를 비틀어 올린 배상도의 얼굴.\n\nLOCATION (lock): The dim outdoor alley directly along the rear wall of a government compound at night.\n\nTIME OF DAY (lock): night.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: close-up\n- KEY BACKGROUND ELEMENTS: Government-office boundary wall (Present behind the encounter) — Its alley-facing side recedes obliquely behind 배상도; used as A softly resolved background plane maintains the established side of the conversation.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Keep the nighttime alley dim, with restrained saturation and enough naturalistic tonal separation to reveal the asymmetric smile without inventing a visible light source.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nRender ONE photorealistic empty location photograph, 16:9, that this shot can be staged inside later. No readable writing anywhere: surfaces that would carry writing may be present, but stage any wording out of legibility — an oblique angle, distance, shallow focus. No captions, watermarks or overlay text.",
  "bg_path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/db00517b-165b-4005-a287-f28936b59b95/images/272bd7af-9294-4490-98c6-88a40e9deacc/scene/recipe/S42sh4__bgfirst_bg.png",
  "asset_id": "493baa60-87b8-4a78-9d1c-aaa2c976e500",
  "input_asset_ids": [
   "adbe5aa6-eb3b-4267-94b9-472eeb2efec2",
   "6ec5270c-a39a-4524-86d4-275694646492"
  ]
 },
 "S42sh4": {
  "input_fingerprint": "936fcc9529a3d341",
  "prompt": "Create ONE FINAL photorealistic live-action film still of the moment below — early 20th-century Korea, Jeolla region; all people are Korean unless explicitly stated otherwise. TIME OF DAY (lock): night.\n\nSHOT TEXT (authoritative, Korean): 박상현을 향해 시선을 고정한 채 비웃듯 한쪽 입꼬리를 비틀어 올린 배상도의 얼굴.\n\nLOCATION (lock): The dim outdoor alley directly along the rear wall of a government compound at night. The shot takes place here — the attached LOCATION PHOTOGRAPH shows the exact spot.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: close-up\n- KEY BACKGROUND ELEMENTS: Government-office boundary wall (Present behind the encounter) — Its alley-facing side recedes obliquely behind 배상도; used as A softly resolved background plane maintains the established side of the conversation.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Keep the nighttime alley dim, with restrained saturation and enough naturalistic tonal separation to reveal the asymmetric smile without inventing a visible light source.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The alley beneath the government-office wall is dim at night. 배상도: He has stopped in the alley.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 배상도 (한국인 남성, 성인 나이대, 자연스러운 검은 머리) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.",
  "roll_prompts": {
   "A": "Stage the shot. The FIRST attached image (SHOT BACKGROUND) is the finished empty background of this shot — keep it EXACTLY: its camera, perspective, architecture, lighting and every fixed feature stay untouched. The SECOND attached image (LAYOUT SKETCH) tells you ONLY where the people go: each sketched person's position, screen size, pose and the gaze/motion arrows. Ignore the sketch's background lines. The CHARACTER REFERENCE photographs show the real people.\nPlace the real people into the background at exactly the sketched positions, sizes and poses, following the arrow directions. No sketch lines or arrows may remain.\n\nCreate ONE FINAL photorealistic live-action film still of the moment below — early 20th-century Korea, Jeolla region; all people are Korean unless explicitly stated otherwise. TIME OF DAY (lock): night.\n\nSHOT TEXT (authoritative, Korean): 박상현을 향해 시선을 고정한 채 비웃듯 한쪽 입꼬리를 비틀어 올린 배상도의 얼굴.\n\nLOCATION (lock): The dim outdoor alley directly along the rear wall of a government compound at night. The shot takes place here — the FIRST attached image (SHOT BACKGROUND) is this exact place, already built: its ground, structures, horizon, materials and lighting are the finished truth of this location and must not be redesigned or replaced. No location photograph is attached — read the place from that image alone, and add no scenery, structure, vehicle or fixture that it does not already show. This lock governs the place only; the figures in the shot follow the staging and pose instructions.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: close-up\n- KEY BACKGROUND ELEMENTS: Government-office boundary wall (Present behind the encounter) — Its alley-facing side recedes obliquely behind 배상도; used as A softly resolved background plane maintains the established side of the conversation.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Keep the nighttime alley dim, with restrained saturation and enough naturalistic tonal separation to reveal the asymmetric smile without inventing a visible light source.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The alley beneath the government-office wall is dim at night. 배상도: He has stopped in the alley.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 배상도 (한국인 남성, 성인 나이대, 자연스러운 검은 머리) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.",
   "B": "Create ONE FINAL photorealistic live-action film still of the moment below — early 20th-century Korea, Jeolla region; all people are Korean unless explicitly stated otherwise. TIME OF DAY (lock): night.\n\nSHOT TEXT (authoritative, Korean): 박상현을 향해 시선을 고정한 채 비웃듯 한쪽 입꼬리를 비틀어 올린 배상도의 얼굴.\n\nLOCATION (lock): The dim outdoor alley directly along the rear wall of a government compound at night. The shot takes place here — the attached LOCATION PHOTOGRAPH shows the exact spot.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: close-up\n- KEY BACKGROUND ELEMENTS: Government-office boundary wall (Present behind the encounter) — Its alley-facing side recedes obliquely behind 배상도; used as A softly resolved background plane maintains the established side of the conversation.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Keep the nighttime alley dim, with restrained saturation and enough naturalistic tonal separation to reveal the asymmetric smile without inventing a visible light source.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The alley beneath the government-office wall is dim at night. 배상도: He has stopped in the alley.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 배상도 (한국인 남성, 성인 나이대, 자연스러운 검은 머리) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.\n\nCOMPOSITION VARIATION (this candidate only): keep the place, the people and the moment exactly as specified, and keep from the CAMERA & FRAME contract its subject scale, screen placement and key background elements exactly. For this candidate only, its camera angle and height are deliberately left open: choose a camera position distinctly different from the obvious one, as a film director picking a second setup on the same blocking. Stay true to any explicit framing the SHOT TEXT itself dictates."
  },
  "roll_refs": {
   "A": [
    {
     "label": "SHOT BACKGROUND",
     "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/db00517b-165b-4005-a287-f28936b59b95/images/272bd7af-9294-4490-98c6-88a40e9deacc/scene/recipe/S42sh4__bgfirst_bg.png",
     "asset_id": "493baa60-87b8-4a78-9d1c-aaa2c976e500",
     "role": "bgfirst_bg"
    },
    {
     "label": "LAYOUT SKETCH (people placement only)",
     "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/db00517b-165b-4005-a287-f28936b59b95/images/272bd7af-9294-4490-98c6-88a40e9deacc/conti/conti_S42sh4.png",
     "asset_id": "adbe5aa6-eb3b-4267-94b9-472eeb2efec2",
     "role": "conti_light"
    },
    {
     "label": "CHARACTER REFERENCE — 배상도: the exact person appearing in this shot; match face, hair and build exactly.",
     "path": "<bytes:851718>",
     "asset_id": "034f5325-975e-4474-854b-e48359710bf8",
     "role": "character_ref"
    }
   ],
   "B": [
    {
     "label": "LOCATION PHOTOGRAPH — the exact place of this shot: its architecture, materials, fixed features and lighting mood are spatial truth; stage the moment inside this place. Never copy its camera framing.",
     "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/db00517b-165b-4005-a287-f28936b59b95/images/272bd7af-9294-4490-98c6-88a40e9deacc/scene/recipe/groupbg_office_rear_alley_50e959.png",
     "asset_id": "6ec5270c-a39a-4524-86d4-275694646492",
     "role": "bgfirst_group_bg"
    },
    {
     "label": "CHARACTER REFERENCE — 배상도: the exact person appearing in this shot; match face, hair and build exactly.",
     "path": "<bytes:851718>",
     "asset_id": "034f5325-975e-4474-854b-e48359710bf8",
     "role": "character_ref"
    }
   ]
  },
  "shot_run_spend_attempt_count": 1,
  "judge_flip_skipped": "judge_fn_owns_order",
  "readings": [
   {
    "label": "A",
    "direction": "배상도의 시선이 프레임 좌측 전경에 위치한 다른 인물의 뒷모습(어깨와 머리)을 향해 고정되어 있습니다.",
    "built_space": "어두운 야외 골목길입니다. 화면 좌측에 돌담이 있고 우측에 건물 구조물이 보이나, 돌담이 배상도의 등 뒤가 아닌 전경의 인물 뒤쪽으로 배치되어 공간적 요구사항과 다소 어긋납니다.",
    "entities": "배상도의 얼굴과 비웃듯 올라간 입꼬리 표정은 잘 묘사되었으나, 레퍼런스의 모자가 누락되었습니다. 또한 허용 인물 목록에 없는 두 번째 인물이 좌측 전경에 크게 등장합니다.",
    "hard_violations": [
     "[gemini-pro] 지침에 허용된 유일한 인물(배상도) 외에 허용되지 않은 추가 인물(프레임 좌측 전경의 뒷모습)이 등장함.",
     "[gpt-high] 배상도 외에는 인물을 보이지 말라는 명시적 제한에도 왼쪽 전경에 별도 인물의 머리와 어깨를 추가했다."
    ],
    "physics": "화면 속 두 인물 모두 땅에 발을 딛고 서 있는 자연스러운 자세를 유지하고 있습니다."
   },
   {
    "label": "B",
    "direction": "배상도의 시선이 화면 우측 밖의 보이지 않는 대상을 향해 고정되어 있습니다.",
    "built_space": "어두운 야외 골목길입니다. 지정된 관아의 돌담이 배상도의 등 뒤에서 비스듬히 멀어지며 배경을 잘 형성하고 있습니다.",
    "entities": "배상도는 모자와 조끼를 포함해 레퍼런스의 외모 및 복장과 완벽히 일치하며, 지문대로 한쪽 입꼬리를 비틀어 올린 표정을 짓고 있습니다.",
    "hard_violations": [],
    "physics": "인물이 지면에 안정적으로 서서 자연스러운 자세를 취하고 있습니다."
   }
  ],
  "cross_model_order": {
   "policy": "cross_model_order_v2_slot0_fwd_slot1_rev_parallel2",
   "slots": [
    {
     "model": "gemini-pro",
     "order": "forward",
     "display_to_canonical": {
      "A": "A",
      "B": "B"
     },
     "raw": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 9,
        "verdict_ko": "레퍼런스의 복장(모자 포함)과 배경 돌담의 배치를 완벽하게 따랐으며, 화면 밖의 대상을 향한 시선과 한쪽 입꼬리가 올라간 표정을 정확히 구현했습니다."
       },
       {
        "label": "A",
        "score": 2,
        "verdict_ko": "비웃는 표정 묘사는 훌륭하나, 프롬프트에서 유일하게 허용한 인물(배상도) 외의 추가 인물을 전경에 배치하는 치명적인 규정 위반을 범했습니다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "배상도의 시선이 프레임 좌측 전경에 위치한 다른 인물의 뒷모습(어깨와 머리)을 향해 고정되어 있습니다.",
        "built_space": "어두운 야외 골목길입니다. 화면 좌측에 돌담이 있고 우측에 건물 구조물이 보이나, 돌담이 배상도의 등 뒤가 아닌 전경의 인물 뒤쪽으로 배치되어 공간적 요구사항과 다소 어긋납니다.",
        "entities": "배상도의 얼굴과 비웃듯 올라간 입꼬리 표정은 잘 묘사되었으나, 레퍼런스의 모자가 누락되었습니다. 또한 허용 인물 목록에 없는 두 번째 인물이 좌측 전경에 크게 등장합니다.",
        "hard_violations": [
         "지침에 허용된 유일한 인물(배상도) 외에 허용되지 않은 추가 인물(프레임 좌측 전경의 뒷모습)이 등장함."
        ],
        "physics": "화면 속 두 인물 모두 땅에 발을 딛고 서 있는 자연스러운 자세를 유지하고 있습니다."
       },
       {
        "label": "B",
        "direction": "배상도의 시선이 화면 우측 밖의 보이지 않는 대상을 향해 고정되어 있습니다.",
        "built_space": "어두운 야외 골목길입니다. 지정된 관아의 돌담이 배상도의 등 뒤에서 비스듬히 멀어지며 배경을 잘 형성하고 있습니다.",
        "entities": "배상도는 모자와 조끼를 포함해 레퍼런스의 외모 및 복장과 완벽히 일치하며, 지문대로 한쪽 입꼬리를 비틀어 올린 표정을 짓고 있습니다.",
        "hard_violations": [],
        "physics": "인물이 지면에 안정적으로 서서 자연스러운 자세를 취하고 있습니다."
       }
      ],
      "all_candidates_fail": false
     },
     "normalized": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 9,
        "verdict_ko": "레퍼런스의 복장(모자 포함)과 배경 돌담의 배치를 완벽하게 따랐으며, 화면 밖의 대상을 향한 시선과 한쪽 입꼬리가 올라간 표정을 정확히 구현했습니다."
       },
       {
        "label": "A",
        "score": 2,
        "verdict_ko": "비웃는 표정 묘사는 훌륭하나, 프롬프트에서 유일하게 허용한 인물(배상도) 외의 추가 인물을 전경에 배치하는 치명적인 규정 위반을 범했습니다."
       }
      ],
      "all_candidates_fail": false,
      "readings": [
       {
        "label": "A",
        "direction": "배상도의 시선이 프레임 좌측 전경에 위치한 다른 인물의 뒷모습(어깨와 머리)을 향해 고정되어 있습니다.",
        "built_space": "어두운 야외 골목길입니다. 화면 좌측에 돌담이 있고 우측에 건물 구조물이 보이나, 돌담이 배상도의 등 뒤가 아닌 전경의 인물 뒤쪽으로 배치되어 공간적 요구사항과 다소 어긋납니다.",
        "entities": "배상도의 얼굴과 비웃듯 올라간 입꼬리 표정은 잘 묘사되었으나, 레퍼런스의 모자가 누락되었습니다. 또한 허용 인물 목록에 없는 두 번째 인물이 좌측 전경에 크게 등장합니다.",
        "hard_violations": [
         "지침에 허용된 유일한 인물(배상도) 외에 허용되지 않은 추가 인물(프레임 좌측 전경의 뒷모습)이 등장함."
        ],
        "physics": "화면 속 두 인물 모두 땅에 발을 딛고 서 있는 자연스러운 자세를 유지하고 있습니다."
       },
       {
        "label": "B",
        "direction": "배상도의 시선이 화면 우측 밖의 보이지 않는 대상을 향해 고정되어 있습니다.",
        "built_space": "어두운 야외 골목길입니다. 지정된 관아의 돌담이 배상도의 등 뒤에서 비스듬히 멀어지며 배경을 잘 형성하고 있습니다.",
        "entities": "배상도는 모자와 조끼를 포함해 레퍼런스의 외모 및 복장과 완벽히 일치하며, 지문대로 한쪽 입꼬리를 비틀어 올린 표정을 짓고 있습니다.",
        "hard_violations": [],
        "physics": "인물이 지면에 안정적으로 서서 자연스러운 자세를 취하고 있습니다."
       }
      ]
     },
     "ok": true
    },
    {
     "model": "gpt-high",
     "order": "reverse",
     "display_to_canonical": {
      "A": "B",
      "B": "A"
     },
     "raw": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 9,
        "verdict_ko": "화면 밖 상대에게 고정한 시선과 비틀어 올린 입꼬리를 얼굴 중심 클로즈업으로 담고, 참조의 모자·복장과 비스듬히 물러나는 밤의 담장을 충실히 유지한다."
       },
       {
        "label": "B",
        "score": 3,
        "verdict_ko": "비웃는 표정과 상대를 향한 시선은 명확하지만, 허용되지 않은 상대 인물을 전경에 추가한 어깨너머 구도이며 얼굴 비중도 작고 참조의 모자가 빠졌다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "얼굴은 화면 오른쪽으로 조금 돌아가 있고 두 눈도 오른쪽 화면 밖 한 지점을 응시한다. 박상현은 보이지 않아 대상의 신원은 확인할 수 없지만, 화면 밖 박상현에게 시선을 고정한 설정과 양립한다. 렌즈를 응시하지 않으며 한쪽 입꼬리가 올라가 있다. 무기나 방향을 확인할 소품은 없다.",
        "built_space": "인물 바로 뒤에 하나의 연속된 관청 경계 담장이 있고 오른쪽 원경으로 비스듬히 물러난다. 하부의 큰 석재, 상부의 작은 사각 구획, 위쪽 기와 덮개가 보이며 참조의 담장 재료와 구성을 따른다. 인물은 담장 바깥 골목 쪽에 있고 배경은 부드럽게 흐려져 있다. 중복 시설이나 반사는 없다.",
        "entities": "성인 한국인 남성으로 설정된 배상도 한 명만 보이며, 검은 머리 일부와 검은 중절모, 흰 교차 깃 의복, 검은 조끼가 참조와 부합한다. 얼굴의 눈매와 코·턱 윤곽도 대체로 유사하다. 정상적인 눈과 자연스러운 피부를 유지한 채 비대칭적인 비웃음을 짓는다. 허리 아래 장신구와 신발은 클로즈업 밖이므로 비교 대상이 아니다. 읽을 수 있는 글자나 가시적인 광원은 없다.",
        "hard_violations": [],
        "physics": "머리는 목과 어깨에 자연스럽게 연결되고 모자는 머리에 얹혀 있다. 조끼와 흰옷은 몸에 걸쳐 정상적으로 드리워진다. 발과 지면 접촉은 프레임 밖이지만 상체에 부유나 비정상적인 기울기는 없고, 골목에서 멈춰 선 순간으로 성립한다."
       },
       {
        "label": "B",
        "direction": "배상도의 눈은 화면 왼쪽 전경에 있는 상대의 얼굴 방향으로 향한다. 상대의 뒤통수와 어깨가 보여 대화 상대를 응시하는 관계는 명확하지만, 그 인물이 박상현인지는 보이는 부분만으로 확인할 수 없다. 한쪽 입꼬리를 올린 비웃음은 뚜렷하다. 전경 인물의 눈은 보이지 않는다.",
        "built_space": "왼쪽에서 중앙 원경으로 이어지는 경계 담장 하나에 큰 석재 기단, 사각 구획 상부, 기와 덮개가 보인다. 오른쪽에는 참조에 있는 낮은 건물과 어두운 출입구 하나가 보이고 가운데 흙길이 이어진다. 장소 구성은 참조와 가깝지만 담장과 골목이 비교적 선명하고 넓게 드러나 얼굴 중심 클로즈업보다 공간 설명 비중이 크다. 배상도는 오른쪽 중경, 추가 인물은 왼쪽 바로 앞에 배치되어 어깨너머 구도를 이룬다.",
        "entities": "중앙 오른쪽에는 검은 머리의 성인 남성 배상도가 흰 교차 깃 의복과 검은 조끼를 입고 있다. 참조의 검은 중절모는 없으며, 드러난 가르마 머리와 얼굴 인상에도 차이가 있다. 왼쪽 전경에는 별도의 남성으로 보이는 뒤통수·귀·목·어깨가 추가되어 있다. 이 부분 인물은 배상도만 보이도록 제한한 인물 조건에 어긋난다. 눈과 피부는 정상적이며 읽을 수 있는 글자나 가시적인 광원은 없다.",
        "hard_violations": [
         "배상도 외에는 인물을 보이지 말라는 명시적 제한에도 왼쪽 전경에 별도 인물의 머리와 어깨를 추가했다."
        ],
        "physics": "배상도의 머리와 상체는 정상적으로 연결되어 있고 의복은 어깨와 몸통에 지지되어 있다. 전경 인물 역시 머리·목·어깨가 이어진 자연스러운 직립 자세다. 두 사람의 발은 프레임 밖이므로 접지는 직접 확인할 수 없지만, 떠 있거나 지지 없이 매달린 신체·물체는 보이지 않는다."
       }
      ],
      "all_candidates_fail": false
     },
     "normalized": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 9,
        "verdict_ko": "화면 밖 상대에게 고정한 시선과 비틀어 올린 입꼬리를 얼굴 중심 클로즈업으로 담고, 참조의 모자·복장과 비스듬히 물러나는 밤의 담장을 충실히 유지한다."
       },
       {
        "label": "A",
        "score": 3,
        "verdict_ko": "비웃는 표정과 상대를 향한 시선은 명확하지만, 허용되지 않은 상대 인물을 전경에 추가한 어깨너머 구도이며 얼굴 비중도 작고 참조의 모자가 빠졌다."
       }
      ],
      "all_candidates_fail": false,
      "readings": [
       {
        "label": "B",
        "direction": "얼굴은 화면 오른쪽으로 조금 돌아가 있고 두 눈도 오른쪽 화면 밖 한 지점을 응시한다. 박상현은 보이지 않아 대상의 신원은 확인할 수 없지만, 화면 밖 박상현에게 시선을 고정한 설정과 양립한다. 렌즈를 응시하지 않으며 한쪽 입꼬리가 올라가 있다. 무기나 방향을 확인할 소품은 없다.",
        "built_space": "인물 바로 뒤에 하나의 연속된 관청 경계 담장이 있고 오른쪽 원경으로 비스듬히 물러난다. 하부의 큰 석재, 상부의 작은 사각 구획, 위쪽 기와 덮개가 보이며 참조의 담장 재료와 구성을 따른다. 인물은 담장 바깥 골목 쪽에 있고 배경은 부드럽게 흐려져 있다. 중복 시설이나 반사는 없다.",
        "entities": "성인 한국인 남성으로 설정된 배상도 한 명만 보이며, 검은 머리 일부와 검은 중절모, 흰 교차 깃 의복, 검은 조끼가 참조와 부합한다. 얼굴의 눈매와 코·턱 윤곽도 대체로 유사하다. 정상적인 눈과 자연스러운 피부를 유지한 채 비대칭적인 비웃음을 짓는다. 허리 아래 장신구와 신발은 클로즈업 밖이므로 비교 대상이 아니다. 읽을 수 있는 글자나 가시적인 광원은 없다.",
        "hard_violations": [],
        "physics": "머리는 목과 어깨에 자연스럽게 연결되고 모자는 머리에 얹혀 있다. 조끼와 흰옷은 몸에 걸쳐 정상적으로 드리워진다. 발과 지면 접촉은 프레임 밖이지만 상체에 부유나 비정상적인 기울기는 없고, 골목에서 멈춰 선 순간으로 성립한다."
       },
       {
        "label": "A",
        "direction": "배상도의 눈은 화면 왼쪽 전경에 있는 상대의 얼굴 방향으로 향한다. 상대의 뒤통수와 어깨가 보여 대화 상대를 응시하는 관계는 명확하지만, 그 인물이 박상현인지는 보이는 부분만으로 확인할 수 없다. 한쪽 입꼬리를 올린 비웃음은 뚜렷하다. 전경 인물의 눈은 보이지 않는다.",
        "built_space": "왼쪽에서 중앙 원경으로 이어지는 경계 담장 하나에 큰 석재 기단, 사각 구획 상부, 기와 덮개가 보인다. 오른쪽에는 참조에 있는 낮은 건물과 어두운 출입구 하나가 보이고 가운데 흙길이 이어진다. 장소 구성은 참조와 가깝지만 담장과 골목이 비교적 선명하고 넓게 드러나 얼굴 중심 클로즈업보다 공간 설명 비중이 크다. 배상도는 오른쪽 중경, 추가 인물은 왼쪽 바로 앞에 배치되어 어깨너머 구도를 이룬다.",
        "entities": "중앙 오른쪽에는 검은 머리의 성인 남성 배상도가 흰 교차 깃 의복과 검은 조끼를 입고 있다. 참조의 검은 중절모는 없으며, 드러난 가르마 머리와 얼굴 인상에도 차이가 있다. 왼쪽 전경에는 별도의 남성으로 보이는 뒤통수·귀·목·어깨가 추가되어 있다. 이 부분 인물은 배상도만 보이도록 제한한 인물 조건에 어긋난다. 눈과 피부는 정상적이며 읽을 수 있는 글자나 가시적인 광원은 없다.",
        "hard_violations": [
         "배상도 외에는 인물을 보이지 말라는 명시적 제한에도 왼쪽 전경에 별도 인물의 머리와 어깨를 추가했다."
        ],
        "physics": "배상도의 머리와 상체는 정상적으로 연결되어 있고 의복은 어깨와 몸통에 지지되어 있다. 전경 인물 역시 머리·목·어깨가 이어진 자연스러운 직립 자세다. 두 사람의 발은 프레임 밖이므로 접지는 직접 확인할 수 없지만, 떠 있거나 지지 없이 매달린 신체·물체는 보이지 않는다."
       }
      ]
     },
     "ok": true
    }
   ],
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "slot_winner_match": true,
   "slot_winner": {
    "gemini-pro": "B",
    "gpt-high": "B"
   },
   "route": "cross_slot_agree"
  },
  "dual": {
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "normalized": {
    "A": 0.556,
    "B": 2.0
   },
   "adjusted": {
    "A": 0.306,
    "B": 2.0
   },
   "violations": {
    "A": [
     "[gemini-pro] 지침에 허용된 유일한 인물(배상도) 외에 허용되지 않은 추가 인물(프레임 좌측 전경의 뒷모습)이 등장함.",
     "[gpt-high] 배상도 외에는 인물을 보이지 말라는 명시적 제한에도 왼쪽 전경에 별도 인물의 머리와 어깨를 추가했다."
    ]
   },
   "per_model_winner": {
    "gemini-pro": "B",
    "gpt-high": "B"
   },
   "agreed": true
  },
  "totals": {
   "B": 2000,
   "A": 306
  },
  "selected": "B",
  "ranking": [
   "B",
   "A"
  ],
  "verdicts": [
   {
    "label": "B",
    "score": 2000,
    "verdict_ko": "레퍼런스의 복장(모자 포함)과 배경 돌담의 배치를 완벽하게 따랐으며, 화면 밖의 대상을 향한 시선과 한쪽 입꼬리가 올라간 표정을 정확히 구현했습니다."
   },
   {
    "label": "A",
    "score": 306,
    "verdict_ko": "비웃는 표정 묘사는 훌륭하나, 프롬프트에서 유일하게 허용한 인물(배상도) 외의 추가 인물을 전경에 배치하는 치명적인 규정 위반을 범했습니다.  ★위반: [gemini-pro] 지침에 허용된 유일한 인물(배상도) 외에 허용되지 않은 추가 인물(프레임 좌측 전경의 뒷모습)이 등장함. / [gpt-high] 배상도 외에는 인물을 보이지 말라는 명시적 제한에도 왼쪽 전경에 별도 인물의 머리와 어깨를 추가했다."
   }
  ],
  "refs": [
   {
    "label": "LOCATION PHOTOGRAPH — the exact place of this shot: its architecture, materials, fixed features and lighting mood are spatial truth; stage the moment inside this place. Never copy its camera framing.",
    "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/db00517b-165b-4005-a287-f28936b59b95/images/272bd7af-9294-4490-98c6-88a40e9deacc/scene/recipe/groupbg_office_rear_alley_50e959.png",
    "asset_id": "6ec5270c-a39a-4524-86d4-275694646492",
    "role": "bgfirst_group_bg"
   },
   {
    "label": "CHARACTER REFERENCE — 배상도: the exact person appearing in this shot; match face, hair and build exactly.",
    "path": "<bytes:851718>",
    "asset_id": "034f5325-975e-4474-854b-e48359710bf8",
    "role": "character_ref"
   }
  ],
  "critique_skipped": true,
  "shot_run_produced": false,
  "shot_run_uid": "06aa142a-73fa-742f-951d-520bd99db4ac",
  "bgfirst": {
   "bg_path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/db00517b-165b-4005-a287-f28936b59b95/images/272bd7af-9294-4490-98c6-88a40e9deacc/scene/recipe/S42sh4__bgfirst_bg.png",
   "bg_asset_id": "493baa60-87b8-4a78-9d1c-aaa2c976e500",
   "bg_record_key": "S42sh4::bgfirst_bg",
   "chain_winner": false,
   "authority": "groupbg",
   "group_key": "office_rear_alley",
   "groupbg_asset_id": "6ec5270c-a39a-4524-86d4-275694646492"
  },
  "ref_mode": "그룹 배경+엔티티 (2택1: 무콘티 승)",
  "share_plan": {
   "ref_plan": "background"
  }
 },
 "S42sh4::cine": {
  "applied": false,
  "attempted_at": "2026-09-09T09:09:14.830965+00:00",
  "fingerprint": "a0114b77715e8793a1752e235ae4fc8d84923dd11f49589a9f024657c1cc8349",
  "fingerprint_version": 2,
  "provider": "mai",
  "endpoint": "openrouter/chat-completions",
  "model": "microsoft/mai-image-2.6",
  "pack": "24.202608252115",
  "source_file": "S42sh4_sel.png",
  "source_sha256": "9256f30bc06f6ed59c964bdf542a19e8beb2aad59b07f0994d558ce3a9459869",
  "error": "RuntimeError: moderation blocked: {\"message\": \"Response content blocked by label 'DallEBlockList'.\", \"code\": 400}",
  "moderation_refusals": 2,
  "declined": true,
  "declined_reason": "moderation"
 },
 "S42sh9::signage": {
  "fp": "8a25b3b5090db129",
  "inscriptions": [],
  "cues": [],
  "dropped": []
 },
 "S42sh9": {
  "input_fingerprint": "5b334fbd3bda523a",
  "prompt": "Create ONE FINAL photorealistic live-action film still of the moment below — early 20th-century Korea, Jeolla region; all people are Korean unless explicitly stated otherwise. TIME OF DAY (lock): night.\n\nSHOT TEXT (authoritative, Korean): 배상도의 귓가에 입을 바짝 댄 채 은밀하게 입술을 벌리고 있는 박상현의 비굴한 옆얼굴.\n\nLOCATION (lock): A secluded standing spot against the government compound's rear boundary wall in the nighttime alley. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: close-up\n- KEY BACKGROUND ELEMENTS: Government-office boundary wall (Behind the two men) — The alley-facing side remains oblique to the lens; used as A narrow, unfocused background margin preserves location without diluting the face-to-ear relationship.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Maintain the same dim nighttime exposure and muted saturation, allowing restrained facial modeling to carry the intimacy and humiliation.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The clothing, hair and overall look of 배상도 — who appear both in that photo and in this shot — are LOCKED to that photo. Anyone else visible in that photo is NOT in this shot: never carry their face, body or clothing onto anyone here. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The government-office alley remains dark. 박상현: He stands in the alley with a tense, subdued expression. 배상도: He has stopped and turned back from leaving.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 박상현 (한국인 남성, 성인 나이대, 자연스러운 검은 머리); 배상도 (한국인 남성, 성인 나이대, 자연스러운 검은 머리) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.",
  "roll_prompts": {
   "A": "Create ONE FINAL photorealistic live-action film still of the moment below — early 20th-century Korea, Jeolla region; all people are Korean unless explicitly stated otherwise. TIME OF DAY (lock): night.\n\nSHOT TEXT (authoritative, Korean): 배상도의 귓가에 입을 바짝 댄 채 은밀하게 입술을 벌리고 있는 박상현의 비굴한 옆얼굴.\n\nLOCATION (lock): A secluded standing spot against the government compound's rear boundary wall in the nighttime alley. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: close-up\n- KEY BACKGROUND ELEMENTS: Government-office boundary wall (Behind the two men) — The alley-facing side remains oblique to the lens; used as A narrow, unfocused background margin preserves location without diluting the face-to-ear relationship.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Maintain the same dim nighttime exposure and muted saturation, allowing restrained facial modeling to carry the intimacy and humiliation.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The clothing, hair and overall look of 배상도 — who appear both in that photo and in this shot — are LOCKED to that photo. Anyone else visible in that photo is NOT in this shot: never carry their face, body or clothing onto anyone here. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The government-office alley remains dark. 박상현: He stands in the alley with a tense, subdued expression. 배상도: He has stopped and turned back from leaving.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 박상현 (한국인 남성, 성인 나이대, 자연스러운 검은 머리); 배상도 (한국인 남성, 성인 나이대, 자연스러운 검은 머리) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.",
   "B": "Create ONE FINAL photorealistic live-action film still of the moment below — early 20th-century Korea, Jeolla region; all people are Korean unless explicitly stated otherwise. TIME OF DAY (lock): night.\n\nSHOT TEXT (authoritative, Korean): 배상도의 귓가에 입을 바짝 댄 채 은밀하게 입술을 벌리고 있는 박상현의 비굴한 옆얼굴.\n\nLOCATION (lock): A secluded standing spot against the government compound's rear boundary wall in the nighttime alley. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: close-up\n- KEY BACKGROUND ELEMENTS: Government-office boundary wall (Behind the two men) — The alley-facing side remains oblique to the lens; used as A narrow, unfocused background margin preserves location without diluting the face-to-ear relationship.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Maintain the same dim nighttime exposure and muted saturation, allowing restrained facial modeling to carry the intimacy and humiliation.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The clothing, hair and overall look of 배상도 — who appear both in that photo and in this shot — are LOCKED to that photo. Anyone else visible in that photo is NOT in this shot: never carry their face, body or clothing onto anyone here. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The government-office alley remains dark. 박상현: He stands in the alley with a tense, subdued expression. 배상도: He has stopped and turned back from leaving.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 박상현 (한국인 남성, 성인 나이대, 자연스러운 검은 머리); 배상도 (한국인 남성, 성인 나이대, 자연스러운 검은 머리) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.\n\nCOMPOSITION VARIATION (this candidate only): keep the place, the people and the moment exactly as specified, and keep from the CAMERA & FRAME contract its subject scale, screen placement and key background elements exactly. For this candidate only, its camera angle and height are deliberately left open: choose a camera position distinctly different from the obvious one, as a film director picking a second setup on the same blocking. Stay true to any explicit framing the SHOT TEXT itself dictates."
  },
  "shot_run_spend_attempt_count": 1,
  "readings": [
   {
    "label": "A",
    "direction": "박상현이 배상도의 귀가 아닌 얼굴 정면(코와 입 부근)을 향해 입을 맞추듯 다가가 있으며, 서로 시선이 교차하고 있음.",
    "built_space": "어두운 밤의 골목길을 배경으로 하고 있으며, 두 사람 뒤로 돌담이 렌즈에 대해 비스듬하게 올바르게 배치됨.",
    "entities": "박상현은 회색 한복과 갓을 착용했고, 배상도는 검은 페도라와 조끼 등 참조 이미지의 의상과 외형을 잘 반영함.",
    "hard_violations": [],
    "physics": "두 인물이 지면에 서서 서로를 향해 상체를 기울인 자세가 무리 없이 지탱되고 있음."
   },
   {
    "label": "B",
    "direction": "박상현이 배상도의 귓가를 정확히 타겟으로 삼아 입을 바짝 대고 속삭이는 방향성이 뚜렷함.",
    "built_space": "밤의 골목길 배경으로, 지시된 돌담이 인물들 뒤로 비스듬한 각도로 적절히 자리잡고 있음.",
    "entities": "배상도는 페도라와 조끼를 착용해 참조와 일치하나, 박상현은 회색 한복만 입고 갓을 착용하지 않은 상태임.",
    "hard_violations": [],
    "physics": "박상현이 배상도의 뒤쪽 측면에서 다가가 귓속말을 하기 위해 상체를 기울인 자세가 물리적으로 자연스럽게 지지됨."
   }
  ],
  "cross_model_order": {
   "policy": "cross_model_order_v2_slot0_fwd_slot1_rev_parallel2",
   "slots": [
    {
     "model": "gemini-pro",
     "order": "forward",
     "display_to_canonical": {
      "A": "A",
      "B": "B"
     },
     "raw": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 8,
        "verdict_ko": "박상현이 갓을 쓰지 않은 의상 누락(우선순위 4)이 있으나, '배상도의 귓가에 입을 바짝 댄' 핵심 동작과 구도(우선순위 2, 3)를 정확하게 묘사하여 더 우수합니다."
       },
       {
        "label": "A",
        "score": 5,
        "verdict_ko": "두 인물의 의상은 참조 이미지와 잘 일치하지만, 박상현이 배상도의 귓가가 아닌 얼굴 정면(코와 입 부근)을 향해 입을 대고 있어 프롬프트가 요구한 은밀한 귓속말 동작을 전혀 살리지 못했습니다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "박상현이 배상도의 귀가 아닌 얼굴 정면(코와 입 부근)을 향해 입을 맞추듯 다가가 있으며, 서로 시선이 교차하고 있음.",
        "built_space": "어두운 밤의 골목길을 배경으로 하고 있으며, 두 사람 뒤로 돌담이 렌즈에 대해 비스듬하게 올바르게 배치됨.",
        "entities": "박상현은 회색 한복과 갓을 착용했고, 배상도는 검은 페도라와 조끼 등 참조 이미지의 의상과 외형을 잘 반영함.",
        "hard_violations": [],
        "physics": "두 인물이 지면에 서서 서로를 향해 상체를 기울인 자세가 무리 없이 지탱되고 있음."
       },
       {
        "label": "B",
        "direction": "박상현이 배상도의 귓가를 정확히 타겟으로 삼아 입을 바짝 대고 속삭이는 방향성이 뚜렷함.",
        "built_space": "밤의 골목길 배경으로, 지시된 돌담이 인물들 뒤로 비스듬한 각도로 적절히 자리잡고 있음.",
        "entities": "배상도는 페도라와 조끼를 착용해 참조와 일치하나, 박상현은 회색 한복만 입고 갓을 착용하지 않은 상태임.",
        "hard_violations": [],
        "physics": "박상현이 배상도의 뒤쪽 측면에서 다가가 귓속말을 하기 위해 상체를 기울인 자세가 물리적으로 자연스럽게 지지됨."
       }
      ],
      "all_candidates_fail": false
     },
     "normalized": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 8,
        "verdict_ko": "박상현이 갓을 쓰지 않은 의상 누락(우선순위 4)이 있으나, '배상도의 귓가에 입을 바짝 댄' 핵심 동작과 구도(우선순위 2, 3)를 정확하게 묘사하여 더 우수합니다."
       },
       {
        "label": "A",
        "score": 5,
        "verdict_ko": "두 인물의 의상은 참조 이미지와 잘 일치하지만, 박상현이 배상도의 귓가가 아닌 얼굴 정면(코와 입 부근)을 향해 입을 대고 있어 프롬프트가 요구한 은밀한 귓속말 동작을 전혀 살리지 못했습니다."
       }
      ],
      "all_candidates_fail": false,
      "readings": [
       {
        "label": "A",
        "direction": "박상현이 배상도의 귀가 아닌 얼굴 정면(코와 입 부근)을 향해 입을 맞추듯 다가가 있으며, 서로 시선이 교차하고 있음.",
        "built_space": "어두운 밤의 골목길을 배경으로 하고 있으며, 두 사람 뒤로 돌담이 렌즈에 대해 비스듬하게 올바르게 배치됨.",
        "entities": "박상현은 회색 한복과 갓을 착용했고, 배상도는 검은 페도라와 조끼 등 참조 이미지의 의상과 외형을 잘 반영함.",
        "hard_violations": [],
        "physics": "두 인물이 지면에 서서 서로를 향해 상체를 기울인 자세가 무리 없이 지탱되고 있음."
       },
       {
        "label": "B",
        "direction": "박상현이 배상도의 귓가를 정확히 타겟으로 삼아 입을 바짝 대고 속삭이는 방향성이 뚜렷함.",
        "built_space": "밤의 골목길 배경으로, 지시된 돌담이 인물들 뒤로 비스듬한 각도로 적절히 자리잡고 있음.",
        "entities": "배상도는 페도라와 조끼를 착용해 참조와 일치하나, 박상현은 회색 한복만 입고 갓을 착용하지 않은 상태임.",
        "hard_violations": [],
        "physics": "박상현이 배상도의 뒤쪽 측면에서 다가가 귓속말을 하기 위해 상체를 기울인 자세가 물리적으로 자연스럽게 지지됨."
       }
      ]
     },
     "ok": true
    },
    {
     "model": "gpt-high",
     "order": "reverse",
     "display_to_canonical": {
      "A": "B",
      "B": "A"
     },
     "raw": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 6,
        "verdict_ko": "배상도의 귀 쪽으로 밀착한 박상현의 옆얼굴과 은밀한 야간 클로즈업은 더 가깝지만, 입은 귀보다 아래 뺨을 향하고 참조의 갓도 빠졌다."
       },
       {
        "label": "B",
        "score": 5,
        "verdict_ko": "두 사람의 의상과 외형은 더 충실하지만, 박상현의 입이 배상도의 귀가 아니라 코와 입 앞으로 향해 핵심 귓속말 연출을 놓쳤다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "왼쪽 박상현은 오른쪽 배상도에게 얼굴을 기울이고 입술을 벌린다. 입은 보이는 귀의 앞쪽 아래 뺨과 턱 부근에 가깝고, 귀 자체에 바짝 닿지는 않는다. 박상현의 시선은 배상도의 얼굴 옆으로 향한다. 배상도는 카메라에서 얼굴을 돌려 눈과 정확한 시선이 가려진다.",
        "built_space": "두 사람 뒤에 회색 벽돌과 하부 석재로 된 연속 담장 하나가 비스듬히 이어진다. 오른쪽 위에는 담장 상단의 어두운 덮개가 보인다. 별도의 문이나 등기구, 중복 시설물은 없다. 두 사람은 담장 앞에 밀착해 있으며, 얼굴과 어깨가 화면을 채우고 담장은 주변부에 남는다. 참조 장소의 재료와 어두운 색조가 대체로 유지된다.",
        "entities": "성인 동아시아계 남성 두 명만 보이며 한국인 설정과 충돌하는 외형은 없다. 박상현은 검은 머리, 옅은 수염, 회색 옷과 밝은 깃을 갖췄지만 참조의 갓과 머리띠가 없고 머리 형태도 달라졌다. 배상도는 검은 모자, 검은 조끼, 흰옷과 짧은 검은 머리를 유지한다. 얼굴 대부분이 가려져 배상도의 얼굴 일치는 제한적으로만 확인된다. 읽을 수 있는 글씨나 추가 소품은 없다. 박상현의 표정은 긴장되고 은밀하지만 비굴함은 강하지 않다.",
        "hard_violations": [],
        "physics": "두 사람의 머리는 목과 상체에 자연스럽게 연결되고, 박상현이 상체와 목을 앞으로 기울여 접근하는 자세는 가능하다. 발과 지면 접촉은 클로즈업 밖이므로 확인되지 않지만 공중에 떠 있다는 징후는 없다. 배상도의 모자는 머리에 얹혀 있고 옷은 어깨와 몸을 따라 드리워진다."
       },
       {
        "label": "B",
        "direction": "왼쪽 박상현의 벌어진 입은 오른쪽 배상도의 코와 윗입술 바로 앞으로 향한다. 배상도의 보이는 귀는 화면 오른쪽으로 멀리 떨어져 있어 입과 귀의 관계가 성립하지 않는다. 박상현은 배상도의 눈과 코 쪽을 보고, 배상도도 왼쪽 박상현 쪽으로 눈을 돌린다. 귓속말보다 얼굴을 마주 댄 대화로 읽힌다.",
        "built_space": "배경에는 회색 벽돌과 하부 석재의 담장 하나가 비스듬히 이어지고 오른쪽 위에 어두운 상단 덮개가 보인다. 추가 문, 조명, 중복 시설물이나 반사면은 없다. 두 사람이 담장 앞에 서 있는 상반신 클로즈업이며 배경은 주변부에 남는다. 참조의 담장 재료와 야간 노출을 대체로 유지한다.",
        "entities": "성인 동아시아계 남성 두 명만 보인다. 박상현의 검은 갓, 머리띠, 턱끈, 옅은 콧수염과 턱수염, 회색 옷은 인물 참조에 가깝다. 배상도도 참조와 유사한 얼굴, 짧은 검은 머리, 검은 모자와 조끼, 흰 깃을 유지한다. 두 사람 모두 자연스러운 눈과 피부를 갖췄다. 박상현은 입술을 벌리고 있으나 비굴함보다는 집중한 표정이다. 읽을 수 있는 문자나 추가 인물은 없다.",
        "hard_violations": [],
        "physics": "박상현이 목과 상체를 앞으로 기울이고 배상도가 고개를 돌리는 자세는 신체적으로 가능하다. 두 모자는 각자의 머리에 지지되며 박상현의 턱끈도 턱 아래로 이어진다. 모자 챙은 화면상 겹치지만 명백한 관통은 확인되지 않는다. 하체와 발은 화면 밖이며, 보이는 상체에 부유나 지지 불가능한 자세는 없다."
       }
      ],
      "all_candidates_fail": true
     },
     "normalized": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 6,
        "verdict_ko": "배상도의 귀 쪽으로 밀착한 박상현의 옆얼굴과 은밀한 야간 클로즈업은 더 가깝지만, 입은 귀보다 아래 뺨을 향하고 참조의 갓도 빠졌다."
       },
       {
        "label": "A",
        "score": 5,
        "verdict_ko": "두 사람의 의상과 외형은 더 충실하지만, 박상현의 입이 배상도의 귀가 아니라 코와 입 앞으로 향해 핵심 귓속말 연출을 놓쳤다."
       }
      ],
      "all_candidates_fail": true,
      "readings": [
       {
        "label": "B",
        "direction": "왼쪽 박상현은 오른쪽 배상도에게 얼굴을 기울이고 입술을 벌린다. 입은 보이는 귀의 앞쪽 아래 뺨과 턱 부근에 가깝고, 귀 자체에 바짝 닿지는 않는다. 박상현의 시선은 배상도의 얼굴 옆으로 향한다. 배상도는 카메라에서 얼굴을 돌려 눈과 정확한 시선이 가려진다.",
        "built_space": "두 사람 뒤에 회색 벽돌과 하부 석재로 된 연속 담장 하나가 비스듬히 이어진다. 오른쪽 위에는 담장 상단의 어두운 덮개가 보인다. 별도의 문이나 등기구, 중복 시설물은 없다. 두 사람은 담장 앞에 밀착해 있으며, 얼굴과 어깨가 화면을 채우고 담장은 주변부에 남는다. 참조 장소의 재료와 어두운 색조가 대체로 유지된다.",
        "entities": "성인 동아시아계 남성 두 명만 보이며 한국인 설정과 충돌하는 외형은 없다. 박상현은 검은 머리, 옅은 수염, 회색 옷과 밝은 깃을 갖췄지만 참조의 갓과 머리띠가 없고 머리 형태도 달라졌다. 배상도는 검은 모자, 검은 조끼, 흰옷과 짧은 검은 머리를 유지한다. 얼굴 대부분이 가려져 배상도의 얼굴 일치는 제한적으로만 확인된다. 읽을 수 있는 글씨나 추가 소품은 없다. 박상현의 표정은 긴장되고 은밀하지만 비굴함은 강하지 않다.",
        "hard_violations": [],
        "physics": "두 사람의 머리는 목과 상체에 자연스럽게 연결되고, 박상현이 상체와 목을 앞으로 기울여 접근하는 자세는 가능하다. 발과 지면 접촉은 클로즈업 밖이므로 확인되지 않지만 공중에 떠 있다는 징후는 없다. 배상도의 모자는 머리에 얹혀 있고 옷은 어깨와 몸을 따라 드리워진다."
       },
       {
        "label": "A",
        "direction": "왼쪽 박상현의 벌어진 입은 오른쪽 배상도의 코와 윗입술 바로 앞으로 향한다. 배상도의 보이는 귀는 화면 오른쪽으로 멀리 떨어져 있어 입과 귀의 관계가 성립하지 않는다. 박상현은 배상도의 눈과 코 쪽을 보고, 배상도도 왼쪽 박상현 쪽으로 눈을 돌린다. 귓속말보다 얼굴을 마주 댄 대화로 읽힌다.",
        "built_space": "배경에는 회색 벽돌과 하부 석재의 담장 하나가 비스듬히 이어지고 오른쪽 위에 어두운 상단 덮개가 보인다. 추가 문, 조명, 중복 시설물이나 반사면은 없다. 두 사람이 담장 앞에 서 있는 상반신 클로즈업이며 배경은 주변부에 남는다. 참조의 담장 재료와 야간 노출을 대체로 유지한다.",
        "entities": "성인 동아시아계 남성 두 명만 보인다. 박상현의 검은 갓, 머리띠, 턱끈, 옅은 콧수염과 턱수염, 회색 옷은 인물 참조에 가깝다. 배상도도 참조와 유사한 얼굴, 짧은 검은 머리, 검은 모자와 조끼, 흰 깃을 유지한다. 두 사람 모두 자연스러운 눈과 피부를 갖췄다. 박상현은 입술을 벌리고 있으나 비굴함보다는 집중한 표정이다. 읽을 수 있는 문자나 추가 인물은 없다.",
        "hard_violations": [],
        "physics": "박상현이 목과 상체를 앞으로 기울이고 배상도가 고개를 돌리는 자세는 신체적으로 가능하다. 두 모자는 각자의 머리에 지지되며 박상현의 턱끈도 턱 아래로 이어진다. 모자 챙은 화면상 겹치지만 명백한 관통은 확인되지 않는다. 하체와 발은 화면 밖이며, 보이는 상체에 부유나 지지 불가능한 자세는 없다."
       }
      ]
     },
     "ok": true
    }
   ],
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "slot_winner_match": true,
   "slot_winner": {
    "gemini-pro": "B",
    "gpt-high": "B"
   },
   "route": "cross_slot_agree"
  },
  "dual": {
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "normalized": {
    "A": 1.458,
    "B": 2.0
   },
   "adjusted": {
    "A": 1.458,
    "B": 2.0
   },
   "violations": {},
   "per_model_winner": {
    "gemini-pro": "B",
    "gpt-high": "B"
   },
   "agreed": true
  },
  "totals": {
   "B": 2000,
   "A": 1458
  },
  "selected": "B",
  "ranking": [
   "B",
   "A"
  ],
  "verdicts": [
   {
    "label": "B",
    "score": 2000,
    "verdict_ko": "박상현이 갓을 쓰지 않은 의상 누락(우선순위 4)이 있으나, '배상도의 귓가에 입을 바짝 댄' 핵심 동작과 구도(우선순위 2, 3)를 정확하게 묘사하여 더 우수합니다."
   },
   {
    "label": "A",
    "score": 1458,
    "verdict_ko": "두 인물의 의상은 참조 이미지와 잘 일치하지만, 박상현이 배상도의 귓가가 아닌 얼굴 정면(코와 입 부근)을 향해 입을 대고 있어 프롬프트가 요구한 은밀한 귓속말 동작을 전혀 살리지 못했습니다."
   }
  ],
  "refs": [
   {
    "label": "PREVIOUS SHOT STILL — a visually related earlier shot of this same place: the location's look, materials, fixed features and lighting mood are LOCKED to this photo; never copy its camera framing. The clothing, hair and overall look of 배상도 — who appear both in that photo and in this shot — are LOCKED to that photo. Anyone else visible in that photo is NOT in this shot: never carry their face, body or clothing onto anyone here. If this photo shows a character who cannot move (dead or unconscious), that character's exact pose, position and orientation are ALSO LOCKED — treat their body as a fixed prop of the set that only the camera moves around.",
    "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/db00517b-165b-4005-a287-f28936b59b95/images/272bd7af-9294-4490-98c6-88a40e9deacc/scene/recipe/S42sh4_sel.png",
    "asset_id": "1f067185-9a7b-429e-986d-8463ff3a8e67",
    "role": "prev_still"
   },
   {
    "label": "CHARACTER REFERENCE — 박상현: the exact person appearing in this shot; match face, hair and build exactly.",
    "path": "<bytes:840781>",
    "asset_id": "6187751d-5df3-4eb7-9d30-d6fc98a005e6",
    "role": "character_ref"
   },
   {
    "label": "CHARACTER REFERENCE — 배상도: the exact person appearing in this shot; match face, hair and build exactly.",
    "path": "<bytes:851718>",
    "asset_id": "034f5325-975e-4474-854b-e48359710bf8",
    "role": "character_ref"
   }
  ],
  "critique_skipped": true,
  "shot_run_produced": false,
  "shot_run_uid": "06aa142b-a684-7c27-a463-3f428577d2dc",
  "ref_mode": "prev+엔티티",
  "share_plan": {
   "ref_plan": "prev",
   "prev_anchor_tag": "S42sh4"
  }
 },
 "S42sh9::cine": {
  "applied": true,
  "attempted_at": "2026-09-09T05:43:23.278770+00:00",
  "fingerprint": "15b2f139f32bbdace4daa0078f9aeca9f6fe6b302e71b61fa0dda2d06b813778",
  "fingerprint_version": 2,
  "provider": "mai",
  "endpoint": "openrouter/chat-completions",
  "model": "microsoft/mai-image-2.6",
  "pack": "24.202608252115",
  "source_file": "S42sh9_sel.png",
  "source_sha256": "14309b4b6f7a46bcc4fb7c7a839152e7c02011309ed7ed8d0f374cc401666a33",
  "file": "S42sh9_cine.png",
  "staged_sha256": "3ce3033f03aeeb3ac495e404c9abbaea54ac86589ba189c07792e48558885bf2",
  "latency_ms": 25653
 },
 "S42sh12::signage": {
  "fp": "29f2eb171aea7e6e",
  "inscriptions": [],
  "cues": [],
  "dropped": []
 },
 "S42sh12": {
  "input_fingerprint": "ebc72782c873e84d",
  "prompt": "Create ONE FINAL photorealistic live-action film still of the moment below — early 20th-century Korea, Jeolla region; all people are Korean unless explicitly stated otherwise. TIME OF DAY (lock): night.\n\nSHOT TEXT (authoritative, Korean): 어깨를 잔뜩 움츠린 채 고개를 푹 숙이고 있는 박상현의 꺾인 상체.\n\nLOCATION (lock): The dark alley beneath the government compound's rear wall, outside the buildings from which cries are heard. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: medium shot\n- KEY BACKGROUND ELEMENTS: Government-office boundary wall (Borders the alley beside 박상현) — Seen obliquely from above; used as An upper background boundary contains the lowered figure; Alley ground (Visible beneath 박상현); used as Negative space below his face gives his downward attention a physical destination.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Preserve the dim nighttime tonal range and muted saturation, with sufficient separation to read the bowed head and compressed shoulders.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The clothing, hair and overall look of 박상현 — who appear both in that photo and in this shot — are LOCKED to that photo. Anyone else visible in that photo is NOT in this shot: never carry their face, body or clothing onto anyone here. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The alley beneath the office wall remains dim. 박상현: He stands with his head lowered.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 박상현 (한국인 남성, 성인 나이대, 자연스러운 검은 머리) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.",
  "roll_prompts": {
   "A": "Create ONE FINAL photorealistic live-action film still of the moment below — early 20th-century Korea, Jeolla region; all people are Korean unless explicitly stated otherwise. TIME OF DAY (lock): night.\n\nSHOT TEXT (authoritative, Korean): 어깨를 잔뜩 움츠린 채 고개를 푹 숙이고 있는 박상현의 꺾인 상체.\n\nLOCATION (lock): The dark alley beneath the government compound's rear wall, outside the buildings from which cries are heard. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: medium shot\n- KEY BACKGROUND ELEMENTS: Government-office boundary wall (Borders the alley beside 박상현) — Seen obliquely from above; used as An upper background boundary contains the lowered figure; Alley ground (Visible beneath 박상현); used as Negative space below his face gives his downward attention a physical destination.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Preserve the dim nighttime tonal range and muted saturation, with sufficient separation to read the bowed head and compressed shoulders.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The clothing, hair and overall look of 박상현 — who appear both in that photo and in this shot — are LOCKED to that photo. Anyone else visible in that photo is NOT in this shot: never carry their face, body or clothing onto anyone here. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The alley beneath the office wall remains dim. 박상현: He stands with his head lowered.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 박상현 (한국인 남성, 성인 나이대, 자연스러운 검은 머리) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.",
   "B": "Create ONE FINAL photorealistic live-action film still of the moment below — early 20th-century Korea, Jeolla region; all people are Korean unless explicitly stated otherwise. TIME OF DAY (lock): night.\n\nSHOT TEXT (authoritative, Korean): 어깨를 잔뜩 움츠린 채 고개를 푹 숙이고 있는 박상현의 꺾인 상체.\n\nLOCATION (lock): The dark alley beneath the government compound's rear wall, outside the buildings from which cries are heard. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: medium shot\n- KEY BACKGROUND ELEMENTS: Government-office boundary wall (Borders the alley beside 박상현) — Seen obliquely from above; used as An upper background boundary contains the lowered figure; Alley ground (Visible beneath 박상현); used as Negative space below his face gives his downward attention a physical destination.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Preserve the dim nighttime tonal range and muted saturation, with sufficient separation to read the bowed head and compressed shoulders.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The clothing, hair and overall look of 박상현 — who appear both in that photo and in this shot — are LOCKED to that photo. Anyone else visible in that photo is NOT in this shot: never carry their face, body or clothing onto anyone here. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The alley beneath the office wall remains dim. 박상현: He stands with his head lowered.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 박상현 (한국인 남성, 성인 나이대, 자연스러운 검은 머리) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.\n\nCOMPOSITION VARIATION (this candidate only): keep the place, the people and the moment exactly as specified, and keep from the CAMERA & FRAME contract its subject scale, screen placement and key background elements exactly. For this candidate only, its camera angle and height are deliberately left open: choose a camera position distinctly different from the obvious one, as a film director picking a second setup on the same blocking. Stay true to any explicit framing the SHOT TEXT itself dictates."
  },
  "shot_run_spend_attempt_count": 1,
  "readings": [
   {
    "label": "A",
    "direction": "캐릭터의 고개가 아래를 향해 꺾여 있으며, 시선은 바닥을 향하고 있습니다.",
    "built_space": "왼쪽에 돌담이 길게 늘어서 있고, 바닥은 흙길이며 우측은 어두운 빈 공간으로 이전 샷의 장소와 일치합니다.",
    "entities": "박상현의 얼굴과 머리 스타일이 레퍼런스와 일치하며, 회색 한복과 넓은 진회색 허리띠가 정확히 묘사되었습니다.",
    "hard_violations": [],
    "physics": "캐릭터가 바닥에 두 발로 안정적으로 서 있으며, 팔은 자연스럽게 늘어뜨려져 있습니다."
   },
   {
    "label": "B",
    "direction": "캐릭터의 시선이 바닥을 향하고 있으나, A에 비해 고개가 덜 꺾여 있습니다.",
    "built_space": "왼쪽에 돌담이 있고, 우측 배경에는 전통 건물들의 형태가 어둠 속에 보입니다.",
    "entities": "박상현의 얼굴과 한복의 기본 형태는 일치하지만, 레퍼런스에 있던 넓은 진회색 허리띠가 생략되었습니다.",
    "hard_violations": [],
    "physics": "캐릭터가 두 발로 바닥을 딛고 서 있으며 체중이 자연스럽게 실려 있습니다."
   }
  ],
  "cross_model_order": {
   "policy": "cross_model_order_v2_slot0_fwd_slot1_rev_parallel2",
   "slots": [
    {
     "model": "gemini-pro",
     "order": "forward",
     "display_to_canonical": {
      "A": "A",
      "B": "B"
     },
     "raw": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 9,
        "verdict_ko": "요청된 '잔뜩 움츠린 어깨와 푹 숙인 고개'를 훌륭하게 구현했으며, 레퍼런스 이미지에 있는 넓은 허리띠 등의 의상 디테일까지 정확하게 반영했습니다."
       },
       {
        "label": "B",
        "score": 6,
        "verdict_ko": "고개를 숙이고 있으나 어깨를 잔뜩 움츠린 느낌이 부족하며, 캐릭터 레퍼런스에 있는 넓은 진회색 허리띠가 누락되었습니다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "캐릭터의 고개가 아래를 향해 꺾여 있으며, 시선은 바닥을 향하고 있습니다.",
        "built_space": "왼쪽에 돌담이 길게 늘어서 있고, 바닥은 흙길이며 우측은 어두운 빈 공간으로 이전 샷의 장소와 일치합니다.",
        "entities": "박상현의 얼굴과 머리 스타일이 레퍼런스와 일치하며, 회색 한복과 넓은 진회색 허리띠가 정확히 묘사되었습니다.",
        "hard_violations": [],
        "physics": "캐릭터가 바닥에 두 발로 안정적으로 서 있으며, 팔은 자연스럽게 늘어뜨려져 있습니다."
       },
       {
        "label": "B",
        "direction": "캐릭터의 시선이 바닥을 향하고 있으나, A에 비해 고개가 덜 꺾여 있습니다.",
        "built_space": "왼쪽에 돌담이 있고, 우측 배경에는 전통 건물들의 형태가 어둠 속에 보입니다.",
        "entities": "박상현의 얼굴과 한복의 기본 형태는 일치하지만, 레퍼런스에 있던 넓은 진회색 허리띠가 생략되었습니다.",
        "hard_violations": [],
        "physics": "캐릭터가 두 발로 바닥을 딛고 서 있으며 체중이 자연스럽게 실려 있습니다."
       }
      ],
      "all_candidates_fail": false
     },
     "normalized": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 9,
        "verdict_ko": "요청된 '잔뜩 움츠린 어깨와 푹 숙인 고개'를 훌륭하게 구현했으며, 레퍼런스 이미지에 있는 넓은 허리띠 등의 의상 디테일까지 정확하게 반영했습니다."
       },
       {
        "label": "B",
        "score": 6,
        "verdict_ko": "고개를 숙이고 있으나 어깨를 잔뜩 움츠린 느낌이 부족하며, 캐릭터 레퍼런스에 있는 넓은 진회색 허리띠가 누락되었습니다."
       }
      ],
      "all_candidates_fail": false,
      "readings": [
       {
        "label": "A",
        "direction": "캐릭터의 고개가 아래를 향해 꺾여 있으며, 시선은 바닥을 향하고 있습니다.",
        "built_space": "왼쪽에 돌담이 길게 늘어서 있고, 바닥은 흙길이며 우측은 어두운 빈 공간으로 이전 샷의 장소와 일치합니다.",
        "entities": "박상현의 얼굴과 머리 스타일이 레퍼런스와 일치하며, 회색 한복과 넓은 진회색 허리띠가 정확히 묘사되었습니다.",
        "hard_violations": [],
        "physics": "캐릭터가 바닥에 두 발로 안정적으로 서 있으며, 팔은 자연스럽게 늘어뜨려져 있습니다."
       },
       {
        "label": "B",
        "direction": "캐릭터의 시선이 바닥을 향하고 있으나, A에 비해 고개가 덜 꺾여 있습니다.",
        "built_space": "왼쪽에 돌담이 있고, 우측 배경에는 전통 건물들의 형태가 어둠 속에 보입니다.",
        "entities": "박상현의 얼굴과 한복의 기본 형태는 일치하지만, 레퍼런스에 있던 넓은 진회색 허리띠가 생략되었습니다.",
        "hard_violations": [],
        "physics": "캐릭터가 두 발로 바닥을 딛고 서 있으며 체중이 자연스럽게 실려 있습니다."
       }
      ]
     },
     "ok": true
    },
    {
     "model": "gpt-high",
     "order": "reverse",
     "display_to_canonical": {
      "A": "B",
      "B": "A"
     },
     "raw": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 7,
        "verdict_ko": "밤의 담장 골목과 인물 외형은 잘 이어지지만, 상체가 비교적 곧고 어깨의 수축이 약해 ‘잔뜩 움츠린 어깨와 꺾인 상체’라는 핵심 동작이 부족하다."
       },
       {
        "label": "B",
        "score": 9,
        "verdict_ko": "비스듬히 내려다보는 중간 거리 구도에서 깊이 숙인 머리, 압축된 어깨와 앞으로 꺾인 상체를 더 정확히 보여 주며 담장과 바닥의 공간 관계도 유지한다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "머리와 눈은 아래로 향하며 시선의 목적지는 몸 앞 골목 바닥이다. 카메라를 바라보지 않는다. 다만 얼굴이 상당히 드러나고 목과 상체의 굽힘은 비교적 얕다. 무기나 방향을 확인할 휴대 물건은 없다.",
        "built_space": "왼쪽 전경에서 오른쪽 후경으로 이어지는 담장 하나가 보이며, 작은 사각 석재의 상단과 큰 석재 기단이라는 참고 장소의 구성을 따른다. 오른쪽 아래에는 흙바닥이 넓게 보이고, 골목 끝에는 기와지붕 건물과 작은 밝은 창이 보인다. 인물은 담장 바깥 골목에 혼자 서 있다. 약간 높은 시점이지만 담장이 상단 경계라기보다 왼쪽 측면 경계로 강하게 읽힌다. 반사면이나 중복된 고정 설비는 없다.",
        "entities": "성인 한국인 남성으로 설정된 박상현 한 명만 보인다. 검은 가르마 머리, 옅은 수염, 회색 전통 겉옷과 밝은 깃은 이전 장면의 외형과 대체로 맞는다. 이전 장면에 함께 있던 모자 쓴 남성은 없다. 머리부터 허벅지 부근까지 포함되어 상체 중심 미디엄 숏보다 다소 넓게 느껴진다. 읽을 수 있는 글자나 별도 소품은 없다.",
        "hard_violations": [],
        "physics": "상체는 화면 아래로 이어지는 골반과 하체 위에 놓여 있고, 두 팔과 빈손은 중력 방향으로 내려온다. 발은 프레임 밖이라 바닥 접촉을 직접 확인할 수 없지만, 부유나 불가능한 지지 상태를 나타내는 단서는 없다. 서서 고개를 숙일 수 있는 자연스러운 자세이나 어깨가 강하게 움츠러들거나 허리가 크게 꺾인 상태는 아니다."
       },
       {
        "label": "B",
        "direction": "머리가 두 어깨 사이로 깊이 내려가 있고 얼굴과 시선은 몸 바로 앞의 골목 바닥을 향한다. 정수리가 많이 보이며 카메라를 향한 시선은 없다. 아래를 향하는 주의가 요청된 동작과 일치한다. 손에 든 물건이나 무기는 없다.",
        "built_space": "담장 하나가 왼쪽과 상단을 채우며 오른쪽 먼 곳으로 사선 후퇴한다. 작은 사각 석재 상단, 큰 돌 기단, 위쪽의 어두운 덮개가 보이며 참고 장소의 재료 구성을 유지한다. 오른쪽과 인물 아래에는 골목 흙바닥이 남아 있다. 박상현은 담장 바깥에 혼자 서 있고, 높은 사선 시점에서 담장이 숙인 인물의 윗배경을 감싼다. 먼 오른쪽에는 어두운 건물 일부가 보인다. 중복 설비나 광학적으로 문제 되는 반사는 없다.",
        "entities": "박상현으로 제시된 성인 한국인 남성 한 명만 있으며 검은 가르마 머리와 회색 전통 겉옷, 밝은 깃이 이전 장면과 이어진다. 얼굴은 숙인 각도로 일부 가려져 세부 동일성을 완전히 확인하기 어렵지만 보이는 외형은 부합한다. 허리에는 어두운 띠가 보인다. 다른 남성이나 모자, 휴대 소품은 없고 읽을 수 있는 글자도 없다. 머리부터 허벅지 부근까지 담되 굽힌 상체가 중심이다.",
        "hard_violations": [],
        "physics": "허리에서 앞으로 기울어진 몸통이 화면 아래로 이어지는 하체 위에 놓이고, 어깨가 목 쪽으로 올라와 압축되어 있다. 팔과 빈손은 자연스럽게 아래로 처져 있다. 발의 접촉점은 프레임 밖이지만 몸이 뜬 흔적은 없으며, 서 있는 사람이 몸을 굽히고 어깨를 움츠려 만들 수 있는 자세다. 옷의 주름도 굽힌 허리와 중력 방향을 따른다."
       }
      ],
      "all_candidates_fail": false
     },
     "normalized": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 7,
        "verdict_ko": "밤의 담장 골목과 인물 외형은 잘 이어지지만, 상체가 비교적 곧고 어깨의 수축이 약해 ‘잔뜩 움츠린 어깨와 꺾인 상체’라는 핵심 동작이 부족하다."
       },
       {
        "label": "A",
        "score": 9,
        "verdict_ko": "비스듬히 내려다보는 중간 거리 구도에서 깊이 숙인 머리, 압축된 어깨와 앞으로 꺾인 상체를 더 정확히 보여 주며 담장과 바닥의 공간 관계도 유지한다."
       }
      ],
      "all_candidates_fail": false,
      "readings": [
       {
        "label": "B",
        "direction": "머리와 눈은 아래로 향하며 시선의 목적지는 몸 앞 골목 바닥이다. 카메라를 바라보지 않는다. 다만 얼굴이 상당히 드러나고 목과 상체의 굽힘은 비교적 얕다. 무기나 방향을 확인할 휴대 물건은 없다.",
        "built_space": "왼쪽 전경에서 오른쪽 후경으로 이어지는 담장 하나가 보이며, 작은 사각 석재의 상단과 큰 석재 기단이라는 참고 장소의 구성을 따른다. 오른쪽 아래에는 흙바닥이 넓게 보이고, 골목 끝에는 기와지붕 건물과 작은 밝은 창이 보인다. 인물은 담장 바깥 골목에 혼자 서 있다. 약간 높은 시점이지만 담장이 상단 경계라기보다 왼쪽 측면 경계로 강하게 읽힌다. 반사면이나 중복된 고정 설비는 없다.",
        "entities": "성인 한국인 남성으로 설정된 박상현 한 명만 보인다. 검은 가르마 머리, 옅은 수염, 회색 전통 겉옷과 밝은 깃은 이전 장면의 외형과 대체로 맞는다. 이전 장면에 함께 있던 모자 쓴 남성은 없다. 머리부터 허벅지 부근까지 포함되어 상체 중심 미디엄 숏보다 다소 넓게 느껴진다. 읽을 수 있는 글자나 별도 소품은 없다.",
        "hard_violations": [],
        "physics": "상체는 화면 아래로 이어지는 골반과 하체 위에 놓여 있고, 두 팔과 빈손은 중력 방향으로 내려온다. 발은 프레임 밖이라 바닥 접촉을 직접 확인할 수 없지만, 부유나 불가능한 지지 상태를 나타내는 단서는 없다. 서서 고개를 숙일 수 있는 자연스러운 자세이나 어깨가 강하게 움츠러들거나 허리가 크게 꺾인 상태는 아니다."
       },
       {
        "label": "A",
        "direction": "머리가 두 어깨 사이로 깊이 내려가 있고 얼굴과 시선은 몸 바로 앞의 골목 바닥을 향한다. 정수리가 많이 보이며 카메라를 향한 시선은 없다. 아래를 향하는 주의가 요청된 동작과 일치한다. 손에 든 물건이나 무기는 없다.",
        "built_space": "담장 하나가 왼쪽과 상단을 채우며 오른쪽 먼 곳으로 사선 후퇴한다. 작은 사각 석재 상단, 큰 돌 기단, 위쪽의 어두운 덮개가 보이며 참고 장소의 재료 구성을 유지한다. 오른쪽과 인물 아래에는 골목 흙바닥이 남아 있다. 박상현은 담장 바깥에 혼자 서 있고, 높은 사선 시점에서 담장이 숙인 인물의 윗배경을 감싼다. 먼 오른쪽에는 어두운 건물 일부가 보인다. 중복 설비나 광학적으로 문제 되는 반사는 없다.",
        "entities": "박상현으로 제시된 성인 한국인 남성 한 명만 있으며 검은 가르마 머리와 회색 전통 겉옷, 밝은 깃이 이전 장면과 이어진다. 얼굴은 숙인 각도로 일부 가려져 세부 동일성을 완전히 확인하기 어렵지만 보이는 외형은 부합한다. 허리에는 어두운 띠가 보인다. 다른 남성이나 모자, 휴대 소품은 없고 읽을 수 있는 글자도 없다. 머리부터 허벅지 부근까지 담되 굽힌 상체가 중심이다.",
        "hard_violations": [],
        "physics": "허리에서 앞으로 기울어진 몸통이 화면 아래로 이어지는 하체 위에 놓이고, 어깨가 목 쪽으로 올라와 압축되어 있다. 팔과 빈손은 자연스럽게 아래로 처져 있다. 발의 접촉점은 프레임 밖이지만 몸이 뜬 흔적은 없으며, 서 있는 사람이 몸을 굽히고 어깨를 움츠려 만들 수 있는 자세다. 옷의 주름도 굽힌 허리와 중력 방향을 따른다."
       }
      ]
     },
     "ok": true
    }
   ],
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "slot_winner_match": true,
   "slot_winner": {
    "gemini-pro": "A",
    "gpt-high": "A"
   },
   "route": "cross_slot_agree"
  },
  "dual": {
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "normalized": {
    "A": 2.0,
    "B": 1.444
   },
   "adjusted": {
    "A": 2.0,
    "B": 1.444
   },
   "violations": {},
   "per_model_winner": {
    "gemini-pro": "A",
    "gpt-high": "A"
   },
   "agreed": true
  },
  "totals": {
   "A": 2000,
   "B": 1444
  },
  "selected": "A",
  "ranking": [
   "A",
   "B"
  ],
  "verdicts": [
   {
    "label": "A",
    "score": 2000,
    "verdict_ko": "요청된 '잔뜩 움츠린 어깨와 푹 숙인 고개'를 훌륭하게 구현했으며, 레퍼런스 이미지에 있는 넓은 허리띠 등의 의상 디테일까지 정확하게 반영했습니다."
   },
   {
    "label": "B",
    "score": 1444,
    "verdict_ko": "고개를 숙이고 있으나 어깨를 잔뜩 움츠린 느낌이 부족하며, 캐릭터 레퍼런스에 있는 넓은 진회색 허리띠가 누락되었습니다."
   }
  ],
  "refs": [
   {
    "label": "PREVIOUS SHOT STILL — a visually related earlier shot of this same place: the location's look, materials, fixed features and lighting mood are LOCKED to this photo; never copy its camera framing. The clothing, hair and overall look of 박상현 — who appear both in that photo and in this shot — are LOCKED to that photo. Anyone else visible in that photo is NOT in this shot: never carry their face, body or clothing onto anyone here. If this photo shows a character who cannot move (dead or unconscious), that character's exact pose, position and orientation are ALSO LOCKED — treat their body as a fixed prop of the set that only the camera moves around.",
    "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/db00517b-165b-4005-a287-f28936b59b95/images/272bd7af-9294-4490-98c6-88a40e9deacc/scene/recipe/S42sh9_sel.png",
    "asset_id": "7bc203e9-1d84-4755-8af5-9e71cd9ecf7b",
    "role": "prev_still"
   },
   {
    "label": "CHARACTER REFERENCE — 박상현: the exact person appearing in this shot; match face, hair and build exactly.",
    "path": "<bytes:840781>",
    "asset_id": "6187751d-5df3-4eb7-9d30-d6fc98a005e6",
    "role": "character_ref"
   }
  ],
  "critique_skipped": true,
  "shot_run_produced": false,
  "shot_run_uid": "06aa142b-a829-71c8-9e3a-d74242c700ef",
  "ref_mode": "prev+엔티티",
  "share_plan": {
   "ref_plan": "prev",
   "prev_anchor_tag": "S42sh9"
  }
 },
 "S42sh12::cine": {
  "applied": true,
  "attempted_at": "2026-09-09T05:44:33.185693+00:00",
  "fingerprint": "c9291c7acf38c02bb37be4cb0d6b5e15d6dcee60ec29a214cb090cdd62711f66",
  "fingerprint_version": 2,
  "provider": "mai",
  "endpoint": "openrouter/chat-completions",
  "model": "microsoft/mai-image-2.6",
  "pack": "24.202608252115",
  "source_file": "S42sh12_sel.png",
  "source_sha256": "51ec02fc246c6946dddfb24aad56942071a98758590c03e6dd4aded14c8e78d8",
  "file": "S42sh12_cine.png",
  "staged_sha256": "aaccb4f5cd761e3ee03e51102f99fdd073160c2df9268584da418610800cf8cf",
  "latency_ms": 26024
 },
 "S43sh8::signage": {
  "fp": "6ea5013244c575a4",
  "inscriptions": [],
  "cues": [],
  "dropped": []
 },
 "era_assess::b627bc404143c3b7": {
  "subjects": [
   {
    "subject_native": "20세기 초 한국 전라도 지역 관아 정문 및 솟을삼문",
    "search_terms_native": [
     "구한말 관아 정문",
     "조선 관아 외삼문 솟을대문",
     "20세기 초 관청 정문 사진"
    ],
    "language_lock_native": "반드시 한국어로만 검색어를 사용해야 하며 다른 언어로 번역하거나 추가하지 마십시오.",
    "reason_ko": "조선 말기~20세기 초 관아 정문의 솟을삼문 구조, 판문과 높은 문턱 형태는 중국·일본식 대문이나 현대 재현물과 확연히 달라 고증이 필요함."
   }
  ],
  "subject_text": "관청 앞·정문 문턱\n관청 건물의 큰 정문과 높인 문턱 앞 공간. 문 양옆으로 담장이 이어지고, 아침빛이 출입구와 바닥을 비춘다.",
  "identity": "canonical",
  "scope_id": "L32",
  "scope_role": "location_exterior",
  "scope_sha": "8f9d10e51ca815ef"
 },
 "era_fail::36525aa0383d99f2": {
  "stage": "research",
  "subject": "20세기 초 한국 전라도 지역 관아 정문 및 솟을삼문",
  "terms": [
   "구한말 관아 정문",
   "조선 관아 외삼문 솟을대문",
   "20세기 초 관청 정문 사진"
  ],
  "status": "no_usable",
  "queries": [
   [
    "전라도 관아 정문 옛사진 외삼문",
    "전주 전라감영 정문 옛 사진 선화당 포정루",
    "나주목 관아 외삼문 옛사진"
   ],
   [
    "국립중앙박물관 유리건판 전남 관아 문",
    "전라감영 포정루 1909 사진",
    "전남 동헌 삼문 옛사진"
   ],
   [
    "\"전라감영\" \"포정루\" \"사진\"",
    "\"관아\" \"삼문\" \"유리건판\"",
    "\"나주\" \"정수루\" \"유리건판\"",
    "\"전라남도\" \"정문\" \"유리건판\""
   ],
   [
    "\"전라감영\" \"포정문\" \"옛\" \"사진\"",
    "\"나주 정수루\" \"사진\" \"일제\"",
    "\"관아\" \"솟을삼문\" \"사진\""
   ],
   [
    "\"전주역사박물관\" \"포정루\"",
    "\"전라감영\" \"정문\" \"1909\"",
    "\"나주\" \"정수루\" \"옛사진\""
   ],
   [
    "\"솟을삼문\" \"가운데\"",
    "\"독진대아문\" \"솟을\""
   ]
  ],
  "candidate_urls": [
   "https://www.imaeil.com/photos/2022/03/29/2022032916403316751_l.jpg",
   "https://mblogthumb-phinf.pstatic.net/20160326_10/ohyh45_14589972207646DRWA_JPEG/5.jpg?type=w420",
   "https://blog.kakaocdn.net/dna/bkqLbO/btqGuyaR2J4/AAAAAAAAAAAAAAAAAAAAAE0dS_yqWX8BUe4Jk3o2-UWO1LCE-gDfdqX8odcWc9fu/img.jpg?allow_ip=&allow_referer=&credential=yqXZFxpELC7KVnFOS48ylbz2pIh7yKj8&expires=1777561199&signature=2rRpdUU21%2FCWErswLBQmFllU1ko%3D",
   "https://umq1gm7a8386.edge.naverncp.com/data2/content/image/2020/06/16/.cache/512/2020061600000065.jpg"
  ],
  "coarse": {
   "eligible": [],
   "chosen_index": 0,
   "reason": "종류·보임·기준을 다 만족하는 후보가 없다 — 이 라운드에선 안 고른다",
   "single_judge": true,
   "rejected_judges": {}
  },
  "verdicts": [
   {
    "index": 1,
    "object_type_match": "yes",
    "visible": true,
    "criteria_match": "unsure",
    "similarity": 100
   },
   {
    "index": 2,
    "object_type_match": "no",
    "visible": true,
    "criteria_match": "unsure",
    "similarity": 30
   },
   {
    "index": 3,
    "object_type_match": "no",
    "visible": true,
    "criteria_match": "unsure",
    "similarity": 50
   },
   {
    "index": 4,
    "object_type_match": "yes",
    "visible": true,
    "criteria_match": "unsure",
    "similarity": 90
   }
  ],
  "chosen_reason_ko": "",
  "attempts": 5
 },
 "groupbg::office_front_gate": {
  "input_fingerprint": "1c3d4acf4eaf6747",
  "meta": {
   "model": "gpt-image-2",
   "size": "1536x864",
   "pack": "11.202607220237",
   "contract": "bgfirst_full_v3",
   "group_sig": {
    "key": "office_front_gate",
    "tags": [
     "S43sh12",
     "S43sh16",
     "S43sh8"
    ]
   },
   "context_sig": "223177ccfb8d9f89"
  },
  "prompt": "Create ONE empty live-action location background photograph — NO PEOPLE, no figures, no body parts, no silhouettes, no shadows or reflections of people anywhere.\n\"Empty\" means no people only: KEEP the location's inherent occupants and stock that define the place — animals in an animal shelter, pen or farm, goods and produce in a market, moored boats in a harbour — unless the shot text explicitly removes them.\nThe FIRST attached image is a thin-line storyboard sketch of a shot that happens at this location: use it ONLY as spatial evidence — what this place contains, how its ground, structures and landmarks are arranged and proportioned. Ignore the sketched people and arrows entirely, and do NOT copy its line style: render a fully photographic, physically plausible real place that fits THE LOCATION text below.\nHUMAN-SCALE CALIBRATION: derive every structure's true size from human-scale elements — a door ≈ 2m, a window ≈ 1–1.5m wide, one storey ≈ 2.5–3m; never inflate a small structure or shrink a large one.\n\nTHE LOCATION — early 20th-century Korea, Jeolla region; all people are Korean unless explicitly stated otherwise: The outdoor forecourt just beyond the government compound's main doorway in morning light.\n\nLOCATION DETAIL (descriptive evidence about this place from the production's location records — evidence, not a staging order): use it to understand what kind of place this is and what it permanently contains. Carry over ONLY its enduring physical features — layout, structures, ground and materials, fixed equipment and installations. Do NOT treat any lighting, weather or time-of-day wording, momentary object states, depicted events or subjective impressions in it as fixed facts of the place; the TIME OF DAY lock below is the sole authority for time and lighting.\n관청 앞·정문 문턱: 육중한 관청 정문이 서 있는 출입 공간과 그 앞의 흙바닥 길. (특징: 육중하고 큰 목조 관청 정문; 높은 문턱; 정문 앞 개방된 흙바닥 공간)\n\nSCENE EVIDENCE (verbatim quotes from the screenplay about this place — treat them as evidence of what the location physically contains and looks like; stage the PLACE those moments happen in, but do NOT depict the momentary actions, people or staged props themselves):\n- S#43. 관청 앞 / 아침\n- 문이 열린다. 수척해진 박제현이 나온다. 걸음이 휘청인다.\n- 그때, 순사들에게 잡혀서 끌려오는 달래.\n\nTIME OF DAY (lock): morning.\n\nRender ONE photorealistic empty location photograph, 16:9, neutral enough that every shot of this place can be staged from it later. No readable writing anywhere: surfaces that would carry writing may be present, but stage any wording out of legibility — an oblique angle, distance, shallow focus. No captions, watermarks or overlay text.",
  "effective_prompt": "Create ONE empty live-action location background photograph — NO PEOPLE, no figures, no body parts, no silhouettes, no shadows or reflections of people anywhere.\n\"Empty\" means no people only: KEEP the location's inherent occupants and stock that define the place — animals in an animal shelter, pen or farm, goods and produce in a market, moored boats in a harbour — unless the shot text explicitly removes them.\nThe FIRST attached image is a thin-line storyboard sketch of a shot that happens at this location: use it ONLY as spatial evidence — what this place contains, how its ground, structures and landmarks are arranged and proportioned. Ignore the sketched people and arrows entirely, and do NOT copy its line style: render a fully photographic, physically plausible real place that fits THE LOCATION text below.\nHUMAN-SCALE CALIBRATION: derive every structure's true size from human-scale elements — a door ≈ 2m, a window ≈ 1–1.5m wide, one storey ≈ 2.5–3m; never inflate a small structure or shrink a large one.\n\nTHE LOCATION — early 20th-century Korea, Jeolla region; all people are Korean unless explicitly stated otherwise: The outdoor forecourt just beyond the government compound's main doorway in morning light.\n\nLOCATION DETAIL (descriptive evidence about this place from the production's location records — evidence, not a staging order): use it to understand what kind of place this is and what it permanently contains. Carry over ONLY its enduring physical features — layout, structures, ground and materials, fixed equipment and installations. Do NOT treat any lighting, weather or time-of-day wording, momentary object states, depicted events or subjective impressions in it as fixed facts of the place; the TIME OF DAY lock below is the sole authority for time and lighting.\n관청 앞·정문 문턱: 육중한 관청 정문이 서 있는 출입 공간과 그 앞의 흙바닥 길. (특징: 육중하고 큰 목조 관청 정문; 높은 문턱; 정문 앞 개방된 흙바닥 공간)\n\nSCENE EVIDENCE (verbatim quotes from the screenplay about this place — treat them as evidence of what the location physically contains and looks like; stage the PLACE those moments happen in, but do NOT depict the momentary actions, people or staged props themselves):\n- S#43. 관청 앞 / 아침\n- 문이 열린다. 수척해진 박제현이 나온다. 걸음이 휘청인다.\n- 그때, 순사들에게 잡혀서 끌려오는 달래.\n\nTIME OF DAY (lock): morning.\n\nRender ONE photorealistic empty location photograph, 16:9, neutral enough that every shot of this place can be staged from it later. No readable writing anywhere: surfaces that would carry writing may be present, but stage any wording out of legibility — an oblique angle, distance, shallow focus. No captions, watermarks or overlay text.",
  "bg_path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/db00517b-165b-4005-a287-f28936b59b95/images/272bd7af-9294-4490-98c6-88a40e9deacc/scene/recipe/groupbg_office_front_gate_53d32d.png",
  "asset_id": "5d413ac3-7277-4beb-978a-5b907907f994",
  "input_asset_ids": [
   "58ede9ba-47c8-4b35-8c1a-2368a6f071b4"
  ],
  "origin_tag": "S43sh8",
  "place_text": "The outdoor forecourt just beyond the government compound's main doorway in morning light.",
  "origin_inputs": {
   "place_text": "The outdoor forecourt just beyond the government compound's main doorway in morning light.",
   "time_of_day_en": "morning",
   "conti_asset_id": "58ede9ba-47c8-4b35-8c1a-2368a6f071b4"
  }
 },
 "S43sh8::bgfirst_bg": {
  "input_fingerprint": "7311499a7ddf8618",
  "prompt": "Create the EMPTY BACKGROUND PLATE for one film shot — NO PEOPLE, no figures, no body parts, no sketch lines, no arrows anywhere.\n\"Empty\" means no people only: KEEP the location's inherent occupants and stock that define the place — animals in an animal shelter, pen or farm, goods in a market, moored boats in a harbour — unless the shot text explicitly removes them.\nThe FIRST attached image is a thin-line storyboard sketch: use ONLY its camera angle, horizon, perspective and the placement/size of buildings and set masses — ignore the sketched people and arrows entirely. The SECOND attached image (LOCATION PHOTOGRAPH) is the real place: take its architecture, materials, signage and fixed features, and RE-PROJECT them into the sketch's camera. If the photograph's camera differs from the sketch's, the sketch's camera wins.\nHUMAN-SCALE CALIBRATION: derive every structure's true size from human-scale elements — a door ≈ 2m, a window ≈ 1–1.5m wide, one storey ≈ 2.5–3m; never inflate a small structure or shrink a large one.\n\nSHOT TEXT this background must serve (Korean): 박제현의 손이 박상현의 어깨 위에 툭 얹어진 순간의 손 클로즈업.\n\nLOCATION (lock): The outdoor forecourt just beyond the government compound's main doorway in morning light.\n\nTIME OF DAY (lock): morning.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: close-up\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Use natural ambient morning light with restrained saturation and gentle, realistic contrast that leaves the trembling fingertips legible.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nRender ONE photorealistic empty location photograph, 16:9, that this shot can be staged inside later. No readable writing anywhere: surfaces that would carry writing may be present, but stage any wording out of legibility — an oblique angle, distance, shallow focus. No captions, watermarks or overlay text.",
  "effective_prompt": "Create the EMPTY BACKGROUND PLATE for one film shot — NO PEOPLE, no figures, no body parts, no sketch lines, no arrows anywhere.\n\"Empty\" means no people only: KEEP the location's inherent occupants and stock that define the place — animals in an animal shelter, pen or farm, goods in a market, moored boats in a harbour — unless the shot text explicitly removes them.\nThe FIRST attached image is a thin-line storyboard sketch: use ONLY its camera angle, horizon, perspective and the placement/size of buildings and set masses — ignore the sketched people and arrows entirely. The SECOND attached image (LOCATION PHOTOGRAPH) is the real place: take its architecture, materials, signage and fixed features, and RE-PROJECT them into the sketch's camera. If the photograph's camera differs from the sketch's, the sketch's camera wins.\nHUMAN-SCALE CALIBRATION: derive every structure's true size from human-scale elements — a door ≈ 2m, a window ≈ 1–1.5m wide, one storey ≈ 2.5–3m; never inflate a small structure or shrink a large one.\n\nSHOT TEXT this background must serve (Korean): 박제현의 손이 박상현의 어깨 위에 툭 얹어진 순간의 손 클로즈업.\n\nLOCATION (lock): The outdoor forecourt just beyond the government compound's main doorway in morning light.\n\nTIME OF DAY (lock): morning.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: close-up\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Use natural ambient morning light with restrained saturation and gentle, realistic contrast that leaves the trembling fingertips legible.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nRender ONE photorealistic empty location photograph, 16:9, that this shot can be staged inside later. No readable writing anywhere: surfaces that would carry writing may be present, but stage any wording out of legibility — an oblique angle, distance, shallow focus. No captions, watermarks or overlay text.",
  "bg_path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/db00517b-165b-4005-a287-f28936b59b95/images/272bd7af-9294-4490-98c6-88a40e9deacc/scene/recipe/S43sh8__bgfirst_bg.png",
  "asset_id": "4d64b926-d55a-4fae-b72c-26e4d55d50bd",
  "input_asset_ids": [
   "58ede9ba-47c8-4b35-8c1a-2368a6f071b4",
   "5d413ac3-7277-4beb-978a-5b907907f994"
  ]
 },
 "S43sh8": {
  "input_fingerprint": "d9ffce7eb85e5d28",
  "prompt": "Create ONE FINAL photorealistic live-action film still of the moment below — early 20th-century Korea, Jeolla region; all people are Korean unless explicitly stated otherwise. TIME OF DAY (lock): morning.\n\nSHOT TEXT (authoritative, Korean): 박제현의 손이 박상현의 어깨 위에 툭 얹어진 순간의 손 클로즈업.\n\nLOCATION (lock): The outdoor forecourt just beyond the government compound's main doorway in morning light. The shot takes place here — the attached LOCATION PHOTOGRAPH shows the exact spot.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: close-up\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Use natural ambient morning light with restrained saturation and gentle, realistic contrast that leaves the trembling fingertips legible.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The government-office entrance is open in morning light. 박제현: He is gaunt and unsteady, with a trembling hand extended at shoulder height. 박상현: He stands nearby, avoiding eye contact.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 박제현 (한국인 남성, 중년 이상 나이대, 자연스러운 연령대의 얼굴); 박상현 (한국인 남성, 성인 나이대, 자연스러운 검은 머리) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.",
  "roll_prompts": {
   "A": "Stage the shot. The FIRST attached image (SHOT BACKGROUND) is the finished empty background of this shot — keep it EXACTLY: its camera, perspective, architecture, lighting and every fixed feature stay untouched. The SECOND attached image (LAYOUT SKETCH) tells you ONLY where the people go: each sketched person's position, screen size, pose and the gaze/motion arrows. Ignore the sketch's background lines. The CHARACTER REFERENCE photographs show the real people.\nPlace the real people into the background at exactly the sketched positions, sizes and poses, following the arrow directions. No sketch lines or arrows may remain.\n\nCreate ONE FINAL photorealistic live-action film still of the moment below — early 20th-century Korea, Jeolla region; all people are Korean unless explicitly stated otherwise. TIME OF DAY (lock): morning.\n\nSHOT TEXT (authoritative, Korean): 박제현의 손이 박상현의 어깨 위에 툭 얹어진 순간의 손 클로즈업.\n\nLOCATION (lock): The outdoor forecourt just beyond the government compound's main doorway in morning light. The shot takes place here — the FIRST attached image (SHOT BACKGROUND) is this exact place, already built: its ground, structures, horizon, materials and lighting are the finished truth of this location and must not be redesigned or replaced. No location photograph is attached — read the place from that image alone, and add no scenery, structure, vehicle or fixture that it does not already show. This lock governs the place only; the figures in the shot follow the staging and pose instructions.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: close-up\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Use natural ambient morning light with restrained saturation and gentle, realistic contrast that leaves the trembling fingertips legible.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The government-office entrance is open in morning light. 박제현: He is gaunt and unsteady, with a trembling hand extended at shoulder height. 박상현: He stands nearby, avoiding eye contact.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 박제현 (한국인 남성, 중년 이상 나이대, 자연스러운 연령대의 얼굴); 박상현 (한국인 남성, 성인 나이대, 자연스러운 검은 머리) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.",
   "B": "Create ONE FINAL photorealistic live-action film still of the moment below — early 20th-century Korea, Jeolla region; all people are Korean unless explicitly stated otherwise. TIME OF DAY (lock): morning.\n\nSHOT TEXT (authoritative, Korean): 박제현의 손이 박상현의 어깨 위에 툭 얹어진 순간의 손 클로즈업.\n\nLOCATION (lock): The outdoor forecourt just beyond the government compound's main doorway in morning light. The shot takes place here — the attached LOCATION PHOTOGRAPH shows the exact spot.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: close-up\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Use natural ambient morning light with restrained saturation and gentle, realistic contrast that leaves the trembling fingertips legible.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The government-office entrance is open in morning light. 박제현: He is gaunt and unsteady, with a trembling hand extended at shoulder height. 박상현: He stands nearby, avoiding eye contact.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 박제현 (한국인 남성, 중년 이상 나이대, 자연스러운 연령대의 얼굴); 박상현 (한국인 남성, 성인 나이대, 자연스러운 검은 머리) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.\n\nCOMPOSITION VARIATION (this candidate only): keep the place, the people and the moment exactly as specified, and keep from the CAMERA & FRAME contract its subject scale, screen placement and key background elements exactly. For this candidate only, its camera angle and height are deliberately left open: choose a camera position distinctly different from the obvious one, as a film director picking a second setup on the same blocking. Stay true to any explicit framing the SHOT TEXT itself dictates."
  },
  "roll_refs": {
   "A": [
    {
     "label": "SHOT BACKGROUND",
     "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/db00517b-165b-4005-a287-f28936b59b95/images/272bd7af-9294-4490-98c6-88a40e9deacc/scene/recipe/S43sh8__bgfirst_bg.png",
     "asset_id": "4d64b926-d55a-4fae-b72c-26e4d55d50bd",
     "role": "bgfirst_bg"
    },
    {
     "label": "LAYOUT SKETCH (people placement only)",
     "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/db00517b-165b-4005-a287-f28936b59b95/images/272bd7af-9294-4490-98c6-88a40e9deacc/conti/conti_S43sh8.png",
     "asset_id": "58ede9ba-47c8-4b35-8c1a-2368a6f071b4",
     "role": "conti_light"
    },
    {
     "label": "CHARACTER REFERENCE — 박제현: the exact person appearing in this shot; match face, hair and build exactly.",
     "path": "<bytes:870304>",
     "asset_id": "02a56dab-2641-421d-89df-af22b5de63df",
     "role": "character_ref"
    },
    {
     "label": "CHARACTER REFERENCE — 박상현: the exact person appearing in this shot; match face, hair and build exactly.",
     "path": "<bytes:840781>",
     "asset_id": "6187751d-5df3-4eb7-9d30-d6fc98a005e6",
     "role": "character_ref"
    }
   ],
   "B": [
    {
     "label": "LOCATION PHOTOGRAPH — the exact place of this shot: its architecture, materials, fixed features and lighting mood are spatial truth; stage the moment inside this place. Never copy its camera framing.",
     "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/db00517b-165b-4005-a287-f28936b59b95/images/272bd7af-9294-4490-98c6-88a40e9deacc/scene/recipe/groupbg_office_front_gate_53d32d.png",
     "asset_id": "5d413ac3-7277-4beb-978a-5b907907f994",
     "role": "bgfirst_group_bg"
    },
    {
     "label": "CHARACTER REFERENCE — 박제현: the exact person appearing in this shot; match face, hair and build exactly.",
     "path": "<bytes:870304>",
     "asset_id": "02a56dab-2641-421d-89df-af22b5de63df",
     "role": "character_ref"
    },
    {
     "label": "CHARACTER REFERENCE — 박상현: the exact person appearing in this shot; match face, hair and build exactly.",
     "path": "<bytes:840781>",
     "asset_id": "6187751d-5df3-4eb7-9d30-d6fc98a005e6",
     "role": "character_ref"
    }
   ]
  },
  "shot_run_spend_attempt_count": 1,
  "judge_flip_skipped": "judge_fn_owns_order",
  "readings": [
   {
    "label": "A",
    "direction": "박제현의 시선은 앞을 향하고, 박상현은 고개를 약간 돌린 상태. 박제현의 오른손이 박상현의 왼쪽 어깨를 향해 뻗어 닿아 있음.",
    "built_space": "위치 레퍼런스와 일치하는 관아 건물 입구, 열린 나무 문, 마당, 먼 배경의 담장과 건물이 보임.",
    "entities": "박제현(흰색 한복을 입은 중년 남성, 옆모습), 박상현(회색 한복을 입은 남성, 뒷모습/옆모습, 갓 누락됨).",
    "hard_violations": [],
    "physics": "박제현의 손은 박상현의 어깨에 의해 자연스럽게 지지되고 있으며, 두 사람 모두 바닥에 안정적으로 서 있음."
   },
   {
    "label": "B",
    "direction": "박상현은 시선을 아래로 내리깔고 피하고 있음. 화면 왼쪽에서 뻗어 나온 박제현의 손이 박상현의 어깨에 얹혀 있음.",
    "built_space": "아웃포커싱된 배경에 관아 건물의 벽면, 문틀, 마당 바닥이 보임.",
    "entities": "박제현(흰색 한복 소매와 주름진 손만 프레임에 포함됨), 박상현(회색 한복과 갓을 착용한 성인 남성).",
    "hard_violations": [],
    "physics": "뻗은 팔과 손은 어깨 위에 안정적으로 얹혀 지지받고 있음."
   }
  ],
  "cross_model_order": {
   "policy": "cross_model_order_v2_slot0_fwd_slot1_rev_parallel2",
   "slots": [
    {
     "model": "gemini-pro",
     "order": "forward",
     "display_to_canonical": {
      "A": "A",
      "B": "B"
     },
     "raw": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 9,
        "verdict_ko": "지시된 '손 클로즈업' 프레이밍을 정확히 구현하였으며, 시선을 피하는 박상현의 표정과 의상(갓)이 레퍼런스와 잘 일치합니다."
       },
       {
        "label": "A",
        "score": 5,
        "verdict_ko": "'손 클로즈업'이라는 숏 크기 지시를 어기고 인물의 상반신과 배경이 너무 많이 드러나는 넓은 프레임으로 렌더링되었으며, 박상현의 갓이 누락되었습니다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "박제현의 시선은 앞을 향하고, 박상현은 고개를 약간 돌린 상태. 박제현의 오른손이 박상현의 왼쪽 어깨를 향해 뻗어 닿아 있음.",
        "built_space": "위치 레퍼런스와 일치하는 관아 건물 입구, 열린 나무 문, 마당, 먼 배경의 담장과 건물이 보임.",
        "entities": "박제현(흰색 한복을 입은 중년 남성, 옆모습), 박상현(회색 한복을 입은 남성, 뒷모습/옆모습, 갓 누락됨).",
        "hard_violations": [],
        "physics": "박제현의 손은 박상현의 어깨에 의해 자연스럽게 지지되고 있으며, 두 사람 모두 바닥에 안정적으로 서 있음."
       },
       {
        "label": "B",
        "direction": "박상현은 시선을 아래로 내리깔고 피하고 있음. 화면 왼쪽에서 뻗어 나온 박제현의 손이 박상현의 어깨에 얹혀 있음.",
        "built_space": "아웃포커싱된 배경에 관아 건물의 벽면, 문틀, 마당 바닥이 보임.",
        "entities": "박제현(흰색 한복 소매와 주름진 손만 프레임에 포함됨), 박상현(회색 한복과 갓을 착용한 성인 남성).",
        "hard_violations": [],
        "physics": "뻗은 팔과 손은 어깨 위에 안정적으로 얹혀 지지받고 있음."
       }
      ],
      "all_candidates_fail": false
     },
     "normalized": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 9,
        "verdict_ko": "지시된 '손 클로즈업' 프레이밍을 정확히 구현하였으며, 시선을 피하는 박상현의 표정과 의상(갓)이 레퍼런스와 잘 일치합니다."
       },
       {
        "label": "A",
        "score": 5,
        "verdict_ko": "'손 클로즈업'이라는 숏 크기 지시를 어기고 인물의 상반신과 배경이 너무 많이 드러나는 넓은 프레임으로 렌더링되었으며, 박상현의 갓이 누락되었습니다."
       }
      ],
      "all_candidates_fail": false,
      "readings": [
       {
        "label": "A",
        "direction": "박제현의 시선은 앞을 향하고, 박상현은 고개를 약간 돌린 상태. 박제현의 오른손이 박상현의 왼쪽 어깨를 향해 뻗어 닿아 있음.",
        "built_space": "위치 레퍼런스와 일치하는 관아 건물 입구, 열린 나무 문, 마당, 먼 배경의 담장과 건물이 보임.",
        "entities": "박제현(흰색 한복을 입은 중년 남성, 옆모습), 박상현(회색 한복을 입은 남성, 뒷모습/옆모습, 갓 누락됨).",
        "hard_violations": [],
        "physics": "박제현의 손은 박상현의 어깨에 의해 자연스럽게 지지되고 있으며, 두 사람 모두 바닥에 안정적으로 서 있음."
       },
       {
        "label": "B",
        "direction": "박상현은 시선을 아래로 내리깔고 피하고 있음. 화면 왼쪽에서 뻗어 나온 박제현의 손이 박상현의 어깨에 얹혀 있음.",
        "built_space": "아웃포커싱된 배경에 관아 건물의 벽면, 문틀, 마당 바닥이 보임.",
        "entities": "박제현(흰색 한복 소매와 주름진 손만 프레임에 포함됨), 박상현(회색 한복과 갓을 착용한 성인 남성).",
        "hard_violations": [],
        "physics": "뻗은 팔과 손은 어깨 위에 안정적으로 얹혀 지지받고 있음."
       }
      ]
     },
     "ok": true
    },
    {
     "model": "gpt-high",
     "order": "reverse",
     "display_to_canonical": {
      "A": "B",
      "B": "A"
     },
     "raw": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 9,
        "verdict_ko": "손과 어깨의 접촉을 중심으로 밀착한 클로즈업이며, 박제현의 나이 든 손과 박상현의 시선 회피·회색 도포·갓이 잘 맞는다."
       },
       {
        "label": "B",
        "score": 7,
        "verdict_ko": "어깨 위에 손을 얹는 접촉과 장소는 정확하지만, 두 사람의 몸통과 배경까지 넓게 담아 손 클로즈업의 집중도가 낮고 박상현의 갓이 빠졌다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "왼쪽 아래에서 뻗은 박제현의 손이 오른쪽 박상현의 어깨 바깥 윗부분에 닿는다. 손가락은 오른쪽으로 펼쳐져 어깨를 따라 놓여 있다. 박상현은 고개를 오른쪽 아래로 숙여 박제현과 눈을 마주치지 않는다. 박제현의 눈은 화면 밖이다.",
        "built_space": "왼쪽 배경에 열린 출입구 하나와 목재 문짝 하나, 그 앞의 낮은 석재 계단이 보인다. 계단에는 여러 수평 단차가 드러나며 오른쪽은 흙마당이다. 두 사람은 문턱 밖 마당 쪽에 있고, 밝은 석재 문틀과 어두운 목재 문은 장소 참조에 부합한다. 먼 담장과 건물은 이 밀착 구도에서 제외되어 있으며, 중복 시설이나 반사는 없다.",
        "entities": "박제현은 주름과 힘줄이 뚜렷한 마른 성인 남성의 손 및 아이보리색 소매로 나타나 중년 이상이라는 설정에 맞는다. 얼굴과 머리는 보이지 않는다. 박상현은 검은 머리와 짧은 수염이 있는 동아시아계 성인 남성의 옆얼굴 일부로 보이며, 회색 도포와 검은 갓·갓끈이 참조와 맞는다. 부분 얼굴만으로 정확한 동일인 여부를 확정하기는 어렵다. 추가 인물, 소품, 읽을 수 있는 글자는 없다.",
        "hard_violations": [],
        "physics": "손목은 소매 안의 팔로 자연스럽게 이어지고 손바닥과 손가락은 실제 어깨 표면에 접촉한다. 손이 어깨 정상보다는 바깥쪽에 조금 더 걸쳐 있지만 충분히 가능한 얹기 동작이다. 박상현의 숙인 상체도 자연스럽고, 발은 구도 밖이므로 지면 접촉은 판단하지 않는다. 공중에 떠 있는 신체나 물체는 없다. 손끝은 선명하지만 떨림 자체는 정지 화면에서 뚜렷하지 않다."
       },
       {
        "label": "B",
        "direction": "왼쪽 박제현의 팔이 오른쪽으로 뻗어 박상현의 어깨 정상에 손을 얹는다. 손가락은 어깨 뒤쪽 곡면을 따라 아래로 굽어 있다. 박상현은 박제현에게 등을 비스듬히 보이고 마당 방향으로 머리를 돌려 직접적인 눈맞춤을 피한다. 두 사람의 눈은 보이지 않는다.",
        "built_space": "왼쪽 중앙에 열린 출입구 하나와 목재 문짝 하나, 석재 문틀 및 문턱 계단 일부가 보인다. 뒤에는 검은 철제 울타리 한 구간, 돌난간의 기둥 세 개, 기와 건물 한 동과 나무가 드러나 참조 장소를 잘 재현한다. 두 사람은 출입구 밖에 서 있으며 공간 충돌이나 불가능한 반사는 없다. 다만 인물 몸통과 배경을 많이 포함해 요구된 손 중심 구도보다 넓다.",
        "entities": "박제현은 왼쪽의 나이 든 남성 하관, 아이보리색 도포와 붉은 허리띠 일부로 보인다. 손은 성인 남성의 손이지만 A보다 덜 수척하고 주름도 약하다. 박상현은 검은 머리의 동아시아계 성인 남성이며 회색 도포는 맞지만, 드러난 머리에 참조의 검은 갓이 없다. 두 얼굴이 대부분 제외되어 정확한 얼굴 일치는 확인하기 어렵다. 추가 인물이나 읽을 수 있는 문자는 없다.",
        "hard_violations": [],
        "physics": "박제현의 팔은 왼쪽 몸통에서 소매와 손목으로 연결되고, 굽힌 손가락과 손바닥은 박상현의 어깨에 지지된다. 어깨에 가볍게 손을 얹을 때 가능한 관절 각도와 접촉이다. 두 상체의 위치 관계도 자연스럽고, 발은 화면 밖이다. 지지 없이 떠 있는 물체나 신체는 없지만 손의 떨림은 뚜렷하게 읽히지 않는다."
       }
      ],
      "all_candidates_fail": false
     },
     "normalized": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 9,
        "verdict_ko": "손과 어깨의 접촉을 중심으로 밀착한 클로즈업이며, 박제현의 나이 든 손과 박상현의 시선 회피·회색 도포·갓이 잘 맞는다."
       },
       {
        "label": "A",
        "score": 7,
        "verdict_ko": "어깨 위에 손을 얹는 접촉과 장소는 정확하지만, 두 사람의 몸통과 배경까지 넓게 담아 손 클로즈업의 집중도가 낮고 박상현의 갓이 빠졌다."
       }
      ],
      "all_candidates_fail": false,
      "readings": [
       {
        "label": "B",
        "direction": "왼쪽 아래에서 뻗은 박제현의 손이 오른쪽 박상현의 어깨 바깥 윗부분에 닿는다. 손가락은 오른쪽으로 펼쳐져 어깨를 따라 놓여 있다. 박상현은 고개를 오른쪽 아래로 숙여 박제현과 눈을 마주치지 않는다. 박제현의 눈은 화면 밖이다.",
        "built_space": "왼쪽 배경에 열린 출입구 하나와 목재 문짝 하나, 그 앞의 낮은 석재 계단이 보인다. 계단에는 여러 수평 단차가 드러나며 오른쪽은 흙마당이다. 두 사람은 문턱 밖 마당 쪽에 있고, 밝은 석재 문틀과 어두운 목재 문은 장소 참조에 부합한다. 먼 담장과 건물은 이 밀착 구도에서 제외되어 있으며, 중복 시설이나 반사는 없다.",
        "entities": "박제현은 주름과 힘줄이 뚜렷한 마른 성인 남성의 손 및 아이보리색 소매로 나타나 중년 이상이라는 설정에 맞는다. 얼굴과 머리는 보이지 않는다. 박상현은 검은 머리와 짧은 수염이 있는 동아시아계 성인 남성의 옆얼굴 일부로 보이며, 회색 도포와 검은 갓·갓끈이 참조와 맞는다. 부분 얼굴만으로 정확한 동일인 여부를 확정하기는 어렵다. 추가 인물, 소품, 읽을 수 있는 글자는 없다.",
        "hard_violations": [],
        "physics": "손목은 소매 안의 팔로 자연스럽게 이어지고 손바닥과 손가락은 실제 어깨 표면에 접촉한다. 손이 어깨 정상보다는 바깥쪽에 조금 더 걸쳐 있지만 충분히 가능한 얹기 동작이다. 박상현의 숙인 상체도 자연스럽고, 발은 구도 밖이므로 지면 접촉은 판단하지 않는다. 공중에 떠 있는 신체나 물체는 없다. 손끝은 선명하지만 떨림 자체는 정지 화면에서 뚜렷하지 않다."
       },
       {
        "label": "A",
        "direction": "왼쪽 박제현의 팔이 오른쪽으로 뻗어 박상현의 어깨 정상에 손을 얹는다. 손가락은 어깨 뒤쪽 곡면을 따라 아래로 굽어 있다. 박상현은 박제현에게 등을 비스듬히 보이고 마당 방향으로 머리를 돌려 직접적인 눈맞춤을 피한다. 두 사람의 눈은 보이지 않는다.",
        "built_space": "왼쪽 중앙에 열린 출입구 하나와 목재 문짝 하나, 석재 문틀 및 문턱 계단 일부가 보인다. 뒤에는 검은 철제 울타리 한 구간, 돌난간의 기둥 세 개, 기와 건물 한 동과 나무가 드러나 참조 장소를 잘 재현한다. 두 사람은 출입구 밖에 서 있으며 공간 충돌이나 불가능한 반사는 없다. 다만 인물 몸통과 배경을 많이 포함해 요구된 손 중심 구도보다 넓다.",
        "entities": "박제현은 왼쪽의 나이 든 남성 하관, 아이보리색 도포와 붉은 허리띠 일부로 보인다. 손은 성인 남성의 손이지만 A보다 덜 수척하고 주름도 약하다. 박상현은 검은 머리의 동아시아계 성인 남성이며 회색 도포는 맞지만, 드러난 머리에 참조의 검은 갓이 없다. 두 얼굴이 대부분 제외되어 정확한 얼굴 일치는 확인하기 어렵다. 추가 인물이나 읽을 수 있는 문자는 없다.",
        "hard_violations": [],
        "physics": "박제현의 팔은 왼쪽 몸통에서 소매와 손목으로 연결되고, 굽힌 손가락과 손바닥은 박상현의 어깨에 지지된다. 어깨에 가볍게 손을 얹을 때 가능한 관절 각도와 접촉이다. 두 상체의 위치 관계도 자연스럽고, 발은 화면 밖이다. 지지 없이 떠 있는 물체나 신체는 없지만 손의 떨림은 뚜렷하게 읽히지 않는다."
       }
      ]
     },
     "ok": true
    }
   ],
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "slot_winner_match": true,
   "slot_winner": {
    "gemini-pro": "B",
    "gpt-high": "B"
   },
   "route": "cross_slot_agree"
  },
  "dual": {
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "normalized": {
    "A": 1.333,
    "B": 2.0
   },
   "adjusted": {
    "A": 1.333,
    "B": 2.0
   },
   "violations": {},
   "per_model_winner": {
    "gemini-pro": "B",
    "gpt-high": "B"
   },
   "agreed": true
  },
  "totals": {
   "B": 2000,
   "A": 1333
  },
  "selected": "B",
  "ranking": [
   "B",
   "A"
  ],
  "verdicts": [
   {
    "label": "B",
    "score": 2000,
    "verdict_ko": "지시된 '손 클로즈업' 프레이밍을 정확히 구현하였으며, 시선을 피하는 박상현의 표정과 의상(갓)이 레퍼런스와 잘 일치합니다."
   },
   {
    "label": "A",
    "score": 1333,
    "verdict_ko": "'손 클로즈업'이라는 숏 크기 지시를 어기고 인물의 상반신과 배경이 너무 많이 드러나는 넓은 프레임으로 렌더링되었으며, 박상현의 갓이 누락되었습니다."
   }
  ],
  "refs": [
   {
    "label": "LOCATION PHOTOGRAPH — the exact place of this shot: its architecture, materials, fixed features and lighting mood are spatial truth; stage the moment inside this place. Never copy its camera framing.",
    "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/db00517b-165b-4005-a287-f28936b59b95/images/272bd7af-9294-4490-98c6-88a40e9deacc/scene/recipe/groupbg_office_front_gate_53d32d.png",
    "asset_id": "5d413ac3-7277-4beb-978a-5b907907f994",
    "role": "bgfirst_group_bg"
   },
   {
    "label": "CHARACTER REFERENCE — 박제현: the exact person appearing in this shot; match face, hair and build exactly.",
    "path": "<bytes:870304>",
    "asset_id": "02a56dab-2641-421d-89df-af22b5de63df",
    "role": "character_ref"
   },
   {
    "label": "CHARACTER REFERENCE — 박상현: the exact person appearing in this shot; match face, hair and build exactly.",
    "path": "<bytes:840781>",
    "asset_id": "6187751d-5df3-4eb7-9d30-d6fc98a005e6",
    "role": "character_ref"
   }
  ],
  "critique_skipped": true,
  "shot_run_produced": false,
  "shot_run_uid": "06aa142b-a9be-790f-bd53-d4537c92694c",
  "bgfirst": {
   "bg_path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/db00517b-165b-4005-a287-f28936b59b95/images/272bd7af-9294-4490-98c6-88a40e9deacc/scene/recipe/S43sh8__bgfirst_bg.png",
   "bg_asset_id": "4d64b926-d55a-4fae-b72c-26e4d55d50bd",
   "bg_record_key": "S43sh8::bgfirst_bg",
   "chain_winner": false,
   "authority": "groupbg",
   "group_key": "office_front_gate",
   "groupbg_asset_id": "5d413ac3-7277-4beb-978a-5b907907f994"
  },
  "ref_mode": "그룹 배경+엔티티 (2택1: 무콘티 승)",
  "share_plan": {
   "ref_plan": "background"
  }
 },
 "S43sh8::cine": {
  "applied": true,
  "attempted_at": "2026-09-09T05:50:18.969809+00:00",
  "fingerprint": "b62e07e7fb4199ebc75fcb58a654b0fd048d49c01bb9302cbe562b6d25136d3f",
  "fingerprint_version": 2,
  "provider": "mai",
  "endpoint": "openrouter/chat-completions",
  "model": "microsoft/mai-image-2.6",
  "pack": "24.202608252115",
  "source_file": "S43sh8_sel.png",
  "source_sha256": "11a066cb8bd91fb76ec7035dca4d62bcb5ca047c52475a400bf9ee0df056b49a",
  "file": "S43sh8_cine.png",
  "staged_sha256": "ec28fcf1db24c75849e7aa3b62605def68fdd91a559d013835ed6de57caaeb70",
  "latency_ms": 27175
 },
 "S43sh12::signage": {
  "fp": "89d0b6698100298a",
  "inscriptions": [],
  "cues": [],
  "dropped": []
 },
 "S43sh12": {
  "input_fingerprint": "5ddaf2b74ca1cb60",
  "prompt": "Create ONE FINAL photorealistic live-action film still of the moment below — early 20th-century Korea, Jeolla region; all people are Korean unless explicitly stated otherwise. TIME OF DAY (lock): morning.\n\nSHOT TEXT (authoritative, Korean): 헝클어진 머리에 양손이 포박당한 달래가 두 명의 일본인 순사에게 양팔을 붙들린 채 앞으로 한 걸음 내딛고 있는 미드 액션 전신.\n\nLOCATION (lock): The morning approach to the government compound's main entrance, where a bound prisoner is being brought in from outside. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: wide shot\n- FRAME LAYOUT: Approach route toward the government-office entrance in the lower-left of the frame, foreground.\n- KEY BACKGROUND ELEMENTS: Government-office entrance (Door open following 박제현's release) — The opening is seen obliquely on the approaching group's destination side; used as A peripheral spatial reference establishes where the forced movement leads without crowding the full figures.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Natural ambient morning light reveals the bound hands and disordered hair with muted saturation and unsentimental, realistic contrast.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The people visible in that photo are NOT in this shot: never carry their faces, bodies or clothing onto anyone here. Each person in THIS shot is defined solely by the PEOPLE list, their own CHARACTER REFERENCE images and their wardrobe notes. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The office entrance remains open in morning light. 달래: She is being led in with disheveled hair and bound hands; the half jade ring remains on a finger.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 달래 (한국인 여성, 젊은 성인, 자연스러운 검은 머리); 일본 순사 (일본인 남성, 성인 나이대, 자연스러운 검은 머리) — wearing: 검은색에 가까운 짙은 감색의 순사 제복. 금장 단추와 순사 모자, 순사도를 찼다. — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.",
  "roll_prompts": {
   "A": "Create ONE FINAL photorealistic live-action film still of the moment below — early 20th-century Korea, Jeolla region; all people are Korean unless explicitly stated otherwise. TIME OF DAY (lock): morning.\n\nSHOT TEXT (authoritative, Korean): 헝클어진 머리에 양손이 포박당한 달래가 두 명의 일본인 순사에게 양팔을 붙들린 채 앞으로 한 걸음 내딛고 있는 미드 액션 전신.\n\nLOCATION (lock): The morning approach to the government compound's main entrance, where a bound prisoner is being brought in from outside. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: wide shot\n- FRAME LAYOUT: Approach route toward the government-office entrance in the lower-left of the frame, foreground.\n- KEY BACKGROUND ELEMENTS: Government-office entrance (Door open following 박제현's release) — The opening is seen obliquely on the approaching group's destination side; used as A peripheral spatial reference establishes where the forced movement leads without crowding the full figures.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Natural ambient morning light reveals the bound hands and disordered hair with muted saturation and unsentimental, realistic contrast.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The people visible in that photo are NOT in this shot: never carry their faces, bodies or clothing onto anyone here. Each person in THIS shot is defined solely by the PEOPLE list, their own CHARACTER REFERENCE images and their wardrobe notes. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The office entrance remains open in morning light. 달래: She is being led in with disheveled hair and bound hands; the half jade ring remains on a finger.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 달래 (한국인 여성, 젊은 성인, 자연스러운 검은 머리); 일본 순사 (일본인 남성, 성인 나이대, 자연스러운 검은 머리) — wearing: 검은색에 가까운 짙은 감색의 순사 제복. 금장 단추와 순사 모자, 순사도를 찼다. — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.",
   "B": "Create ONE FINAL photorealistic live-action film still of the moment below — early 20th-century Korea, Jeolla region; all people are Korean unless explicitly stated otherwise. TIME OF DAY (lock): morning.\n\nSHOT TEXT (authoritative, Korean): 헝클어진 머리에 양손이 포박당한 달래가 두 명의 일본인 순사에게 양팔을 붙들린 채 앞으로 한 걸음 내딛고 있는 미드 액션 전신.\n\nLOCATION (lock): The morning approach to the government compound's main entrance, where a bound prisoner is being brought in from outside. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: wide shot\n- FRAME LAYOUT: Approach route toward the government-office entrance in the lower-left of the frame, foreground.\n- KEY BACKGROUND ELEMENTS: Government-office entrance (Door open following 박제현's release) — The opening is seen obliquely on the approaching group's destination side; used as A peripheral spatial reference establishes where the forced movement leads without crowding the full figures.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Natural ambient morning light reveals the bound hands and disordered hair with muted saturation and unsentimental, realistic contrast.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The people visible in that photo are NOT in this shot: never carry their faces, bodies or clothing onto anyone here. Each person in THIS shot is defined solely by the PEOPLE list, their own CHARACTER REFERENCE images and their wardrobe notes. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The office entrance remains open in morning light. 달래: She is being led in with disheveled hair and bound hands; the half jade ring remains on a finger.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 달래 (한국인 여성, 젊은 성인, 자연스러운 검은 머리); 일본 순사 (일본인 남성, 성인 나이대, 자연스러운 검은 머리) — wearing: 검은색에 가까운 짙은 감색의 순사 제복. 금장 단추와 순사 모자, 순사도를 찼다. — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.\n\nCOMPOSITION VARIATION (this candidate only): keep the place, the people and the moment exactly as specified, and keep from the CAMERA & FRAME contract its subject scale, screen placement and key background elements exactly. For this candidate only, its camera angle and height are deliberately left open: choose a camera position distinctly different from the obvious one, as a film director picking a second setup on the same blocking. Stay true to any explicit framing the SHOT TEXT itself dictates."
  },
  "shot_run_spend_attempt_count": 1,
  "readings": [
   {
    "label": "A",
    "direction": "세 인물 모두 정면을 향해 걸어오고 있으며, 달래의 시선은 약간 아래를 향해 자연스럽게 이어집니다.",
    "built_space": "화면 좌측 전경에 레퍼런스와 일치하는 석조 기단 및 열린 목조 문이 정확히 배치되었습니다. 하지만 벽면에 프롬프트 텍스트가 유출된 흰색 현판이 걸려 있습니다.",
    "entities": "달래는 지정된 녹색 저고리와 검은 치마, 옥반지를 정확히 착용했습니다. 두 명의 일본 순사 역시 짙은 감색 제복과 금장 단추, 모자 등 지정된 복장을 갖추고 있습니다.",
    "hard_violations": [
     "[gemini-pro] 프롬프트 지시문('박제현')이 배경 간판의 명확히 읽히는 한글 텍스트로 유출됨 (leaked text 위반 및 읽을 수 있는 글자 생성 금지 위반)",
     "[gpt-high] 출입구 옆 표지의 한글을 읽을 수 있어, 판독 가능한 글자를 일절 금지한 조건을 위반한다."
    ],
    "physics": "세 사람 모두 바닥을 디디고 걷고 있으며 무게 중심이 타당합니다. 달래를 억압하는 순사들의 손과 팔 위치 및 지탱이 물리적으로 자연스럽게 묘사되었습니다."
   },
   {
    "label": "B",
    "direction": "세 인물 모두 카메라를 향해 정면으로 걸어오고 있으며, 달래는 고개를 푹 숙이고 있습니다.",
    "built_space": "레퍼런스의 단일 문 구조 대신, 프롬프트의 지침과 맞지 않는 거대한 대칭형 목조 대문 구조물이 중앙 배경을 꽉 채우고 있습니다.",
    "entities": "두 순사의 제복 형태는 맞으나, 달래가 지정된 녹색 저고리와 검은 치마가 아닌 회색 무명옷을 입고 있어 복장 지침에 심각하게 실패했습니다.",
    "hard_violations": [],
    "physics": "걷는 자세와 발이 지면에 닿아 있는 모습, 그리고 순사들이 양옆에서 달래의 팔을 붙잡는 손의 방향과 형태가 물리적 무리 없이 지탱되고 있습니다."
   }
  ],
  "cross_model_order": {
   "policy": "cross_model_order_v2_slot0_fwd_slot1_rev_parallel2",
   "slots": [
    {
     "model": "gemini-pro",
     "order": "forward",
     "display_to_canonical": {
      "A": "A",
      "B": "B"
     },
     "raw": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 3,
        "verdict_ko": "장소(거대한 목조 대문)와 달래의 복장(회색 무명옷)이 레퍼런스 및 프롬프트와 완전히 어긋나지만, 실격에 해당하는 하드 위반 사항이 없어 A보다 우선합니다."
       },
       {
        "label": "A",
        "score": 0,
        "verdict_ko": "프레임 구성과 복장 등 대부분의 지침을 완벽히 따랐으나, 프롬프트의 지시문('박제현')이 벽면 현판의 읽을 수 있는 한글 텍스트로 그대로 유출되어 실격(Hard Violation) 처리됩니다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "세 인물 모두 정면을 향해 걸어오고 있으며, 달래의 시선은 약간 아래를 향해 자연스럽게 이어집니다.",
        "built_space": "화면 좌측 전경에 레퍼런스와 일치하는 석조 기단 및 열린 목조 문이 정확히 배치되었습니다. 하지만 벽면에 프롬프트 텍스트가 유출된 흰색 현판이 걸려 있습니다.",
        "entities": "달래는 지정된 녹색 저고리와 검은 치마, 옥반지를 정확히 착용했습니다. 두 명의 일본 순사 역시 짙은 감색 제복과 금장 단추, 모자 등 지정된 복장을 갖추고 있습니다.",
        "hard_violations": [
         "프롬프트 지시문('박제현')이 배경 간판의 명확히 읽히는 한글 텍스트로 유출됨 (leaked text 위반 및 읽을 수 있는 글자 생성 금지 위반)"
        ],
        "physics": "세 사람 모두 바닥을 디디고 걷고 있으며 무게 중심이 타당합니다. 달래를 억압하는 순사들의 손과 팔 위치 및 지탱이 물리적으로 자연스럽게 묘사되었습니다."
       },
       {
        "label": "B",
        "direction": "세 인물 모두 카메라를 향해 정면으로 걸어오고 있으며, 달래는 고개를 푹 숙이고 있습니다.",
        "built_space": "레퍼런스의 단일 문 구조 대신, 프롬프트의 지침과 맞지 않는 거대한 대칭형 목조 대문 구조물이 중앙 배경을 꽉 채우고 있습니다.",
        "entities": "두 순사의 제복 형태는 맞으나, 달래가 지정된 녹색 저고리와 검은 치마가 아닌 회색 무명옷을 입고 있어 복장 지침에 심각하게 실패했습니다.",
        "hard_violations": [],
        "physics": "걷는 자세와 발이 지면에 닿아 있는 모습, 그리고 순사들이 양옆에서 달래의 팔을 붙잡는 손의 방향과 형태가 물리적 무리 없이 지탱되고 있습니다."
       }
      ],
      "all_candidates_fail": false
     },
     "normalized": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 3,
        "verdict_ko": "장소(거대한 목조 대문)와 달래의 복장(회색 무명옷)이 레퍼런스 및 프롬프트와 완전히 어긋나지만, 실격에 해당하는 하드 위반 사항이 없어 A보다 우선합니다."
       },
       {
        "label": "A",
        "score": 0,
        "verdict_ko": "프레임 구성과 복장 등 대부분의 지침을 완벽히 따랐으나, 프롬프트의 지시문('박제현')이 벽면 현판의 읽을 수 있는 한글 텍스트로 그대로 유출되어 실격(Hard Violation) 처리됩니다."
       }
      ],
      "all_candidates_fail": false,
      "readings": [
       {
        "label": "A",
        "direction": "세 인물 모두 정면을 향해 걸어오고 있으며, 달래의 시선은 약간 아래를 향해 자연스럽게 이어집니다.",
        "built_space": "화면 좌측 전경에 레퍼런스와 일치하는 석조 기단 및 열린 목조 문이 정확히 배치되었습니다. 하지만 벽면에 프롬프트 텍스트가 유출된 흰색 현판이 걸려 있습니다.",
        "entities": "달래는 지정된 녹색 저고리와 검은 치마, 옥반지를 정확히 착용했습니다. 두 명의 일본 순사 역시 짙은 감색 제복과 금장 단추, 모자 등 지정된 복장을 갖추고 있습니다.",
        "hard_violations": [
         "프롬프트 지시문('박제현')이 배경 간판의 명확히 읽히는 한글 텍스트로 유출됨 (leaked text 위반 및 읽을 수 있는 글자 생성 금지 위반)"
        ],
        "physics": "세 사람 모두 바닥을 디디고 걷고 있으며 무게 중심이 타당합니다. 달래를 억압하는 순사들의 손과 팔 위치 및 지탱이 물리적으로 자연스럽게 묘사되었습니다."
       },
       {
        "label": "B",
        "direction": "세 인물 모두 카메라를 향해 정면으로 걸어오고 있으며, 달래는 고개를 푹 숙이고 있습니다.",
        "built_space": "레퍼런스의 단일 문 구조 대신, 프롬프트의 지침과 맞지 않는 거대한 대칭형 목조 대문 구조물이 중앙 배경을 꽉 채우고 있습니다.",
        "entities": "두 순사의 제복 형태는 맞으나, 달래가 지정된 녹색 저고리와 검은 치마가 아닌 회색 무명옷을 입고 있어 복장 지침에 심각하게 실패했습니다.",
        "hard_violations": [],
        "physics": "걷는 자세와 발이 지면에 닿아 있는 모습, 그리고 순사들이 양옆에서 달래의 팔을 붙잡는 손의 방향과 형태가 물리적 무리 없이 지탱되고 있습니다."
       }
      ]
     },
     "ok": true
    },
    {
     "model": "gpt-high",
     "order": "reverse",
     "display_to_canonical": {
      "A": "B",
      "B": "A"
     },
     "raw": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 4,
        "verdict_ko": "전신 보행과 양팔을 붙잡은 포박 호송은 구현했지만, 목적지 입구의 배치가 불명확하고 달래의 의상이 참조와 다르다."
       },
       {
        "label": "B",
        "score": 2,
        "verdict_ko": "달래의 의상과 왼쪽 출입구는 참조에 더 가깝지만, 읽을 수 있는 한글 표지가 명시적 금지사항을 위반하며 보행도 입구로 향하지 않는다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "세 사람은 화면 앞쪽으로 걷고 순사들은 전방, 달래는 전방 아래를 본다. 크게 열린 문은 일행 뒤에 있어 그 문으로 들어가기보다 멀어지는 동작으로 읽힌다. 왼쪽 전경에는 기단과 기둥이 있지만 목적지인 열린 입구는 확인되지 않는다. 순사도는 칼집에 들어가 허리에서 아래로 향한다.",
        "built_space": "중앙 뒤에 열린 대문 한 곳과 좌우 문짝 두 장, 그 너머 건물의 창호와 낮은 기단이 보인다. 왼쪽 전경에는 목재 기둥과 석재 기단이 있다. 세 사람은 기단 옆 흙바닥에 나란히 서서 걷는다. 참조의 목재·회벽·돌·흙 재질은 이어지지만, 목적지 쪽 주변부에 비스듬히 보여야 할 출입구 대신 뒤쪽 대문이 크게 강조된다. 반사는 없다.",
        "entities": "젊은 동아시아계 여성 한 명과 성인 동아시아계 남성 순사 두 명으로 인원은 맞으며, 외모만으로 한국인과 일본인 국적을 구별할 수는 없다. 달래는 검은 머리가 헝클어지고 양손목이 밧줄로 묶여 있다. 얼굴은 머리카락과 숙인 각도로 일부 가려져 정확한 동일성 확인이 제한된다. 참조의 연녹색 저고리와 검은 치마 대신 회백색 저고리와 회갈색 치마를 입었다. 손가락의 반쪽 옥가락지는 명확하지 않고 소매 근처에 녹색 장식이 보인다. 두 순사는 짙은 감색 제복, 금장 단추, 제모와 허리에 찬 칼을 갖췄다. 읽을 수 있는 글자는 없다.",
        "hard_violations": [],
        "physics": "달래는 화면 왼쪽 신발로 땅을 딛고 반대 발을 들어 한 걸음을 진행한다. 양옆 순사의 손이 각각 달래의 위팔에 접촉해 붙잡고 있으며, 순사들도 지면에 닿은 발로 체중을 지지한다. 포박 밧줄은 손목에 감겨 있고 칼집은 허리 장구에 매달려 있다. 지지 없이 떠 있는 몸이나 물체는 보이지 않는다."
       },
       {
        "label": "B",
        "direction": "세 사람의 몸과 발걸음은 카메라 쪽을 향한다. 달래는 전방 아래, 순사들은 대체로 전방을 본다. 열린 출입구는 일행의 왼쪽 뒤에 있어, 그쪽으로 들어가는 동작보다는 지나쳐 앞으로 나오는 동작이다. 두 칼은 칼집에 꽂힌 채 아래를 향하고 오른쪽 순사는 칼자루 부근에 손을 댄다.",
        "built_space": "왼쪽에 열린 출입구 한 곳과 돌계단, 화면 가장자리의 격자창 한 곳이 보인다. 뒤편에는 기와지붕 건물과 격자창 두 곳, 오른쪽에는 목책이 있다. 세 사람은 계단 옆 흙마당에 있어 발 디딜 공간은 충분하다. 왼쪽 출입구의 목재 문틀·회벽·돌 문턱은 참조와 비교적 가깝고 입구도 비스듬히 보이지만, 일행의 진행 목적지로 연결되지 않는다. 반사는 없다.",
        "entities": "젊은 동아시아계 여성 한 명과 성인 동아시아계 남성 순사 두 명이 보이며 추가 인물은 없다. 국적은 외모만으로 확정할 수 없다. 달래의 연녹색 저고리, 검은 치마와 검은 신발은 참조에 가깝고 검은 머리도 헝클어져 있다. 양손목에는 밧줄이 있으며 손가락에 밝은 고리 같은 부분이 보이지만 반쪽 옥가락지의 형태까지 확정하기는 어렵다. 두 순사의 감색 제복, 금장 단추, 모자와 순사도는 요구에 부합한다. 출입구 옆 흰 세로 표지에는 판독 가능한 한글이 있다.",
        "hard_violations": [
         "출입구 옆 표지의 한글을 읽을 수 있어, 판독 가능한 글자를 일절 금지한 조건을 위반한다."
        ],
        "physics": "달래는 한 발로 흙바닥을 딛고 다른 발을 뒤에서 들어 옮기고 있으며 양쪽 순사가 각각 위팔을 잡는다. 순사들의 보행에도 지면을 딛는 발이 있어 체중 지지가 가능하다. 밧줄은 손목에 감겨 있고 칼과 칼집은 허리띠에 연결되어 있다. 지지 없이 떠 있는 몸이나 물체는 보이지 않는다."
       }
      ],
      "all_candidates_fail": true
     },
     "normalized": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 4,
        "verdict_ko": "전신 보행과 양팔을 붙잡은 포박 호송은 구현했지만, 목적지 입구의 배치가 불명확하고 달래의 의상이 참조와 다르다."
       },
       {
        "label": "A",
        "score": 2,
        "verdict_ko": "달래의 의상과 왼쪽 출입구는 참조에 더 가깝지만, 읽을 수 있는 한글 표지가 명시적 금지사항을 위반하며 보행도 입구로 향하지 않는다."
       }
      ],
      "all_candidates_fail": true,
      "readings": [
       {
        "label": "B",
        "direction": "세 사람은 화면 앞쪽으로 걷고 순사들은 전방, 달래는 전방 아래를 본다. 크게 열린 문은 일행 뒤에 있어 그 문으로 들어가기보다 멀어지는 동작으로 읽힌다. 왼쪽 전경에는 기단과 기둥이 있지만 목적지인 열린 입구는 확인되지 않는다. 순사도는 칼집에 들어가 허리에서 아래로 향한다.",
        "built_space": "중앙 뒤에 열린 대문 한 곳과 좌우 문짝 두 장, 그 너머 건물의 창호와 낮은 기단이 보인다. 왼쪽 전경에는 목재 기둥과 석재 기단이 있다. 세 사람은 기단 옆 흙바닥에 나란히 서서 걷는다. 참조의 목재·회벽·돌·흙 재질은 이어지지만, 목적지 쪽 주변부에 비스듬히 보여야 할 출입구 대신 뒤쪽 대문이 크게 강조된다. 반사는 없다.",
        "entities": "젊은 동아시아계 여성 한 명과 성인 동아시아계 남성 순사 두 명으로 인원은 맞으며, 외모만으로 한국인과 일본인 국적을 구별할 수는 없다. 달래는 검은 머리가 헝클어지고 양손목이 밧줄로 묶여 있다. 얼굴은 머리카락과 숙인 각도로 일부 가려져 정확한 동일성 확인이 제한된다. 참조의 연녹색 저고리와 검은 치마 대신 회백색 저고리와 회갈색 치마를 입었다. 손가락의 반쪽 옥가락지는 명확하지 않고 소매 근처에 녹색 장식이 보인다. 두 순사는 짙은 감색 제복, 금장 단추, 제모와 허리에 찬 칼을 갖췄다. 읽을 수 있는 글자는 없다.",
        "hard_violations": [],
        "physics": "달래는 화면 왼쪽 신발로 땅을 딛고 반대 발을 들어 한 걸음을 진행한다. 양옆 순사의 손이 각각 달래의 위팔에 접촉해 붙잡고 있으며, 순사들도 지면에 닿은 발로 체중을 지지한다. 포박 밧줄은 손목에 감겨 있고 칼집은 허리 장구에 매달려 있다. 지지 없이 떠 있는 몸이나 물체는 보이지 않는다."
       },
       {
        "label": "A",
        "direction": "세 사람의 몸과 발걸음은 카메라 쪽을 향한다. 달래는 전방 아래, 순사들은 대체로 전방을 본다. 열린 출입구는 일행의 왼쪽 뒤에 있어, 그쪽으로 들어가는 동작보다는 지나쳐 앞으로 나오는 동작이다. 두 칼은 칼집에 꽂힌 채 아래를 향하고 오른쪽 순사는 칼자루 부근에 손을 댄다.",
        "built_space": "왼쪽에 열린 출입구 한 곳과 돌계단, 화면 가장자리의 격자창 한 곳이 보인다. 뒤편에는 기와지붕 건물과 격자창 두 곳, 오른쪽에는 목책이 있다. 세 사람은 계단 옆 흙마당에 있어 발 디딜 공간은 충분하다. 왼쪽 출입구의 목재 문틀·회벽·돌 문턱은 참조와 비교적 가깝고 입구도 비스듬히 보이지만, 일행의 진행 목적지로 연결되지 않는다. 반사는 없다.",
        "entities": "젊은 동아시아계 여성 한 명과 성인 동아시아계 남성 순사 두 명이 보이며 추가 인물은 없다. 국적은 외모만으로 확정할 수 없다. 달래의 연녹색 저고리, 검은 치마와 검은 신발은 참조에 가깝고 검은 머리도 헝클어져 있다. 양손목에는 밧줄이 있으며 손가락에 밝은 고리 같은 부분이 보이지만 반쪽 옥가락지의 형태까지 확정하기는 어렵다. 두 순사의 감색 제복, 금장 단추, 모자와 순사도는 요구에 부합한다. 출입구 옆 흰 세로 표지에는 판독 가능한 한글이 있다.",
        "hard_violations": [
         "출입구 옆 표지의 한글을 읽을 수 있어, 판독 가능한 글자를 일절 금지한 조건을 위반한다."
        ],
        "physics": "달래는 한 발로 흙바닥을 딛고 다른 발을 뒤에서 들어 옮기고 있으며 양쪽 순사가 각각 위팔을 잡는다. 순사들의 보행에도 지면을 딛는 발이 있어 체중 지지가 가능하다. 밧줄은 손목에 감겨 있고 칼과 칼집은 허리띠에 연결되어 있다. 지지 없이 떠 있는 몸이나 물체는 보이지 않는다."
       }
      ]
     },
     "ok": true
    }
   ],
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "slot_winner_match": true,
   "slot_winner": {
    "gemini-pro": "B",
    "gpt-high": "B"
   },
   "route": "cross_slot_agree"
  },
  "dual": {
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "normalized": {
    "A": 0.5,
    "B": 2.0
   },
   "adjusted": {
    "A": 0.25,
    "B": 2.0
   },
   "violations": {
    "A": [
     "[gemini-pro] 프롬프트 지시문('박제현')이 배경 간판의 명확히 읽히는 한글 텍스트로 유출됨 (leaked text 위반 및 읽을 수 있는 글자 생성 금지 위반)",
     "[gpt-high] 출입구 옆 표지의 한글을 읽을 수 있어, 판독 가능한 글자를 일절 금지한 조건을 위반한다."
    ]
   },
   "per_model_winner": {
    "gemini-pro": "B",
    "gpt-high": "B"
   },
   "agreed": true
  },
  "totals": {
   "B": 2000,
   "A": 250
  },
  "selected": "B",
  "ranking": [
   "B",
   "A"
  ],
  "verdicts": [
   {
    "label": "B",
    "score": 2000,
    "verdict_ko": "장소(거대한 목조 대문)와 달래의 복장(회색 무명옷)이 레퍼런스 및 프롬프트와 완전히 어긋나지만, 실격에 해당하는 하드 위반 사항이 없어 A보다 우선합니다."
   },
   {
    "label": "A",
    "score": 250,
    "verdict_ko": "프레임 구성과 복장 등 대부분의 지침을 완벽히 따랐으나, 프롬프트의 지시문('박제현')이 벽면 현판의 읽을 수 있는 한글 텍스트로 그대로 유출되어 실격(Hard Violation) 처리됩니다.  ★위반: [gemini-pro] 프롬프트 지시문('박제현')이 배경 간판의 명확히 읽히는 한글 텍스트로 유출됨 (leaked text 위반 및 읽을 수 있는 글자 생성 금지 위반) / [gpt-high] 출입구 옆 표지의 한글을 읽을 수 있어, 판독 가능한 글자를 일절 금지한 조건을 위반한다."
   }
  ],
  "refs": [
   {
    "label": "PREVIOUS SHOT STILL — a visually related earlier shot of this same place: the location's look, materials, fixed features and lighting mood are LOCKED to this photo; never copy its camera framing. The people visible in that photo are NOT in this shot: never carry their faces, bodies or clothing onto anyone here. Each person in THIS shot is defined solely by the PEOPLE list, their own CHARACTER REFERENCE images and their wardrobe notes. If this photo shows a character who cannot move (dead or unconscious), that character's exact pose, position and orientation are ALSO LOCKED — treat their body as a fixed prop of the set that only the camera moves around.",
    "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/db00517b-165b-4005-a287-f28936b59b95/images/272bd7af-9294-4490-98c6-88a40e9deacc/scene/recipe/S43sh8_sel.png",
    "asset_id": "b4854894-b59d-4f5f-bd3b-69707c619eb3",
    "role": "prev_still"
   },
   {
    "label": "CHARACTER REFERENCE — 달래: the exact person appearing in this shot; match face, hair and build exactly.",
    "path": "<bytes:879791>",
    "asset_id": "83024e69-b375-4f3e-98ec-820d0e2ff1f0",
    "role": "character_ref"
   }
  ],
  "critique_skipped": true,
  "shot_run_produced": false,
  "shot_run_uid": "06aa1431-3e09-758d-84e3-621bacb10cf5",
  "ref_mode": "prev+엔티티",
  "share_plan": {
   "ref_plan": "prev",
   "prev_anchor_tag": "S43sh8"
  },
  "lane_policy": "ab_select_bypass:prev"
 },
 "S43sh12::cine": {
  "applied": true,
  "attempted_at": "2026-09-09T05:52:42.678053+00:00",
  "fingerprint": "f0c715329892a7556c1d0032ebab0feefde1a7c0d688ef40e6c0947ef4cd6dbd",
  "fingerprint_version": 2,
  "provider": "mai",
  "endpoint": "openrouter/chat-completions",
  "model": "microsoft/mai-image-2.6",
  "pack": "24.202608252115",
  "source_file": "S43sh12_sel.png",
  "source_sha256": "431a0ef34f3401c21d0e7d8ff0c1b5b002a1cdb8057523076de245b528617999",
  "file": "S43sh12_cine.png",
  "staged_sha256": "cc3718fb641d5b8ced4b239bf27985b53efdc53df7277c446566e29d73dd1104",
  "latency_ms": 30535
 },
 "S43sh16::signage": {
  "fp": "4d215237f0ea73fe",
  "inscriptions": [],
  "cues": [],
  "dropped": []
 },
 "S43sh16": {
  "input_fingerprint": "7bda82565328eb75",
  "prompt": "Create ONE FINAL photorealistic live-action film still of the moment below — early 20th-century Korea, Jeolla region; all people are Korean unless explicitly stated otherwise. TIME OF DAY (lock): morning.\n\nSHOT TEXT (authoritative, Korean): 박상현을 향해 한쪽 입꼬리를 올린 채 씩 비웃고 있는 배상도의 비열한 얼굴 클로즈업.\n\nLOCATION (lock): Immediately outside the government building's entrance, along the arriving prisoner's route across the forecourt. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: close-up\n- KEY BACKGROUND ELEMENTS: Government-office entrance (Open) — A partial oblique view of the opening remains behind his far side; used as A softly resolved edge anticipates his departure into the building.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Maintain natural ambient morning light and muted saturation, allowing ordinary facial contrast rather than stylized shadow to expose the sneer.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The people visible in that photo are NOT in this shot: never carry their faces, bodies or clothing onto anyone here. Each person in THIS shot is defined solely by the PEOPLE list, their own CHARACTER REFERENCE images and their wardrobe notes. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nPREVIOUS STILL USAGE (follow exactly — what to take from the attached still and what to exclude; it governs place and objects only, never who is in this shot or how the camera sees them): Take the official building entrance, threshold materials, and morning light. Exclude the earlier shoulder-touching interaction and any bystanders not remaining at the entrance.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The office entrance remains open. 배상도: He is entering the building with a slight smile.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 배상도 (한국인 남성, 성인 나이대, 자연스러운 검은 머리) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.",
  "roll_prompts": {
   "A": "Create ONE FINAL photorealistic live-action film still of the moment below — early 20th-century Korea, Jeolla region; all people are Korean unless explicitly stated otherwise. TIME OF DAY (lock): morning.\n\nSHOT TEXT (authoritative, Korean): 박상현을 향해 한쪽 입꼬리를 올린 채 씩 비웃고 있는 배상도의 비열한 얼굴 클로즈업.\n\nLOCATION (lock): Immediately outside the government building's entrance, along the arriving prisoner's route across the forecourt. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: close-up\n- KEY BACKGROUND ELEMENTS: Government-office entrance (Open) — A partial oblique view of the opening remains behind his far side; used as A softly resolved edge anticipates his departure into the building.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Maintain natural ambient morning light and muted saturation, allowing ordinary facial contrast rather than stylized shadow to expose the sneer.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The people visible in that photo are NOT in this shot: never carry their faces, bodies or clothing onto anyone here. Each person in THIS shot is defined solely by the PEOPLE list, their own CHARACTER REFERENCE images and their wardrobe notes. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nPREVIOUS STILL USAGE (follow exactly — what to take from the attached still and what to exclude; it governs place and objects only, never who is in this shot or how the camera sees them): Take the official building entrance, threshold materials, and morning light. Exclude the earlier shoulder-touching interaction and any bystanders not remaining at the entrance.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The office entrance remains open. 배상도: He is entering the building with a slight smile.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 배상도 (한국인 남성, 성인 나이대, 자연스러운 검은 머리) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.",
   "B": "Create ONE FINAL photorealistic live-action film still of the moment below — early 20th-century Korea, Jeolla region; all people are Korean unless explicitly stated otherwise. TIME OF DAY (lock): morning.\n\nSHOT TEXT (authoritative, Korean): 박상현을 향해 한쪽 입꼬리를 올린 채 씩 비웃고 있는 배상도의 비열한 얼굴 클로즈업.\n\nLOCATION (lock): Immediately outside the government building's entrance, along the arriving prisoner's route across the forecourt. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: close-up\n- KEY BACKGROUND ELEMENTS: Government-office entrance (Open) — A partial oblique view of the opening remains behind his far side; used as A softly resolved edge anticipates his departure into the building.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Maintain natural ambient morning light and muted saturation, allowing ordinary facial contrast rather than stylized shadow to expose the sneer.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The people visible in that photo are NOT in this shot: never carry their faces, bodies or clothing onto anyone here. Each person in THIS shot is defined solely by the PEOPLE list, their own CHARACTER REFERENCE images and their wardrobe notes. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nPREVIOUS STILL USAGE (follow exactly — what to take from the attached still and what to exclude; it governs place and objects only, never who is in this shot or how the camera sees them): Take the official building entrance, threshold materials, and morning light. Exclude the earlier shoulder-touching interaction and any bystanders not remaining at the entrance.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The office entrance remains open. 배상도: He is entering the building with a slight smile.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 배상도 (한국인 남성, 성인 나이대, 자연스러운 검은 머리) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.\n\nCOMPOSITION VARIATION (this candidate only): keep the place, the people and the moment exactly as specified, and keep from the CAMERA & FRAME contract its subject scale, screen placement and key background elements exactly. For this candidate only, its camera angle and height are deliberately left open: choose a camera position distinctly different from the obvious one, as a film director picking a second setup on the same blocking. Stay true to any explicit framing the SHOT TEXT itself dictates."
  },
  "shot_run_spend_attempt_count": 1,
  "readings": [
   {
    "label": "A",
    "direction": "배상도는 화면 왼쪽 앞부분에 위치한 인물의 머리를 향해 시선을 던지고 있습니다.",
    "built_space": "관아 출입구의 열린 나무 문과 마당이 배경에 나타납니다.",
    "entities": "배상도의 외모와 의상은 레퍼런스와 일치하나, 화면 왼쪽 전경에 프롬프트에 없는 다른 사람의 뒷머리와 귀가 나타나 있습니다.",
    "hard_violations": [
     "[gemini-pro] invented people (프롬프트에 허용되지 않은 다른 인물의 뒷모습이 화면 왼쪽 전경에 등장)",
     "[gpt-high] 배상도만 허용된 장면에 두 번째 인물의 머리와 얼굴 일부를 왼쪽 전경으로 추가했다."
    ],
    "physics": "자연스럽게 서 있는 자세입니다."
   },
   {
    "label": "B",
    "direction": "배상도는 프레임 바깥 왼쪽을 향해 시선을 두고 있으며, 텍스트가 암시하는 박상현을 쳐다보는 것으로 보입니다.",
    "built_space": "이전 샷과 동일한 관아 출입구의 열린 나무 문과 내부 마당이 배경으로 정확히 배치되어 있습니다.",
    "entities": "배상도는 레퍼런스와 일치하는 얼굴과 의상을 입고 있으며, 한쪽 입꼬리를 올린 표정을 잘 짓고 있습니다. 다른 인물은 없습니다.",
    "hard_violations": [],
    "physics": "자연스럽게 서 있는 자세입니다."
   }
  ],
  "cross_model_order": {
   "policy": "cross_model_order_v2_slot0_fwd_slot1_rev_parallel2",
   "slots": [
    {
     "model": "gemini-pro",
     "order": "forward",
     "display_to_canonical": {
      "A": "A",
      "B": "B"
     },
     "raw": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 10,
        "verdict_ko": "배상도의 비열한 미소와 클로즈업 앵글을 완벽하게 구현했으며, 프롬프트의 지시대로 다른 인물을 배제하고 배경을 정확하게 표현했습니다."
       },
       {
        "label": "A",
        "score": 2,
        "verdict_ko": "프롬프트에 명시되지 않은 인물(화면 왼쪽 앞의 뒤통수)이 등장하여 엄격한 등장인물 제한 규칙을 위반했습니다."
       }
      ],
      "readings": [
       {
        "label": "B",
        "direction": "배상도는 프레임 바깥 왼쪽을 향해 시선을 두고 있으며, 텍스트가 암시하는 박상현을 쳐다보는 것으로 보입니다.",
        "built_space": "이전 샷과 동일한 관아 출입구의 열린 나무 문과 내부 마당이 배경으로 정확히 배치되어 있습니다.",
        "entities": "배상도는 레퍼런스와 일치하는 얼굴과 의상을 입고 있으며, 한쪽 입꼬리를 올린 표정을 잘 짓고 있습니다. 다른 인물은 없습니다.",
        "hard_violations": [],
        "physics": "자연스럽게 서 있는 자세입니다."
       },
       {
        "label": "A",
        "direction": "배상도는 화면 왼쪽 앞부분에 위치한 인물의 머리를 향해 시선을 던지고 있습니다.",
        "built_space": "관아 출입구의 열린 나무 문과 마당이 배경에 나타납니다.",
        "entities": "배상도의 외모와 의상은 레퍼런스와 일치하나, 화면 왼쪽 전경에 프롬프트에 없는 다른 사람의 뒷머리와 귀가 나타나 있습니다.",
        "hard_violations": [
         "invented people (프롬프트에 허용되지 않은 다른 인물의 뒷모습이 화면 왼쪽 전경에 등장)"
        ],
        "physics": "자연스럽게 서 있는 자세입니다."
       }
      ],
      "all_candidates_fail": false
     },
     "normalized": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 10,
        "verdict_ko": "배상도의 비열한 미소와 클로즈업 앵글을 완벽하게 구현했으며, 프롬프트의 지시대로 다른 인물을 배제하고 배경을 정확하게 표현했습니다."
       },
       {
        "label": "A",
        "score": 2,
        "verdict_ko": "프롬프트에 명시되지 않은 인물(화면 왼쪽 앞의 뒤통수)이 등장하여 엄격한 등장인물 제한 규칙을 위반했습니다."
       }
      ],
      "all_candidates_fail": false,
      "readings": [
       {
        "label": "B",
        "direction": "배상도는 프레임 바깥 왼쪽을 향해 시선을 두고 있으며, 텍스트가 암시하는 박상현을 쳐다보는 것으로 보입니다.",
        "built_space": "이전 샷과 동일한 관아 출입구의 열린 나무 문과 내부 마당이 배경으로 정확히 배치되어 있습니다.",
        "entities": "배상도는 레퍼런스와 일치하는 얼굴과 의상을 입고 있으며, 한쪽 입꼬리를 올린 표정을 잘 짓고 있습니다. 다른 인물은 없습니다.",
        "hard_violations": [],
        "physics": "자연스럽게 서 있는 자세입니다."
       },
       {
        "label": "A",
        "direction": "배상도는 화면 왼쪽 앞부분에 위치한 인물의 머리를 향해 시선을 던지고 있습니다.",
        "built_space": "관아 출입구의 열린 나무 문과 마당이 배경에 나타납니다.",
        "entities": "배상도의 외모와 의상은 레퍼런스와 일치하나, 화면 왼쪽 전경에 프롬프트에 없는 다른 사람의 뒷머리와 귀가 나타나 있습니다.",
        "hard_violations": [
         "invented people (프롬프트에 허용되지 않은 다른 인물의 뒷모습이 화면 왼쪽 전경에 등장)"
        ],
        "physics": "자연스럽게 서 있는 자세입니다."
       }
      ]
     },
     "ok": true
    },
    {
     "model": "gpt-high",
     "order": "reverse",
     "display_to_canonical": {
      "A": "B",
      "B": "A"
     },
     "raw": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 7,
        "verdict_ko": "배상도만 담은 얼굴 클로즈업과 비대칭 비웃음, 열린 관아 입구를 충실히 구현했으나 참조의 검은 모자가 없고 박상현을 향한 시선은 다소 불명확하다."
       },
       {
        "label": "B",
        "score": 2,
        "verdict_ko": "상대를 향한 비웃음은 명확하지만 허용되지 않은 두 번째 인물의 머리를 전경에 추가해 단독 얼굴 클로즈업을 어긴다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "얼굴은 화면 오른쪽으로 조금 돌아가 있고 눈은 카메라 부근의 화면 밖을 향한다. 박상현은 보이지 않아 그를 정확히 바라보는지는 확정할 수 없지만, 화면 밖 상대에게 보내는 눈길로 읽을 수 있다. 한쪽 입꼬리가 더 올라가 있다. 무기나 방향성 있는 소품은 없다.",
        "built_space": "인물은 목조 출입구 바로 앞에 있으며, 화면 오른쪽에 가로 보강대가 있는 낡은 문짝 하나, 그 안쪽에 열린 통로 하나와 뒤편 목조 건물이 보인다. 참조의 목재·회벽과 열린 입구 구성이 이어지며 중복된 설비는 없다. 입구가 인물 뒤에 비스듬히 보이지만 요청한 부드러운 가장자리보다 배경에서 차지하는 비중은 크다. 반사는 없다.",
        "entities": "성인 한국인 남성으로 설정된 배상도 한 명만 보이며 검은 머리, 흰 한복 깃, 검은 조끼가 확인된다. 참조와 의상 계열은 같지만 참조의 검은 모자가 없고, 드러난 헤어스타일과 얼굴 윤곽도 완전히 같지는 않다. 허리 장식과 하의는 클로즈업 밖이므로 평가하지 않는다. 추가 인물이나 읽을 수 있는 글자는 없다. 빛은 자연스럽고 채도가 낮아 아침 설정과 충돌하지 않는다.",
        "hard_violations": [],
        "physics": "머리는 목과 어깨에 자연스럽게 연결되고 옷은 상체에 걸쳐 있다. 얼굴 근육과 눈의 형태도 정상적인 비웃음 연기로 성립한다. 발과 지면 접촉은 프레임 밖이며 부유를 나타내는 증거는 없다. 건물로 들어가는 발걸음은 이 얼굴 클로즈업만으로 확인할 수 없다."
       },
       {
        "label": "B",
        "direction": "배상도의 얼굴과 눈은 화면 왼쪽 전경 인물의 얼굴 쪽을 향한다. 조롱의 대상은 화면에서 명확하며 입꼬리를 올리고 이를 조금 드러낸다. 다만 그 대상을 보이게 만들기 위해 별도의 인물을 추가했다. 무기나 지향성 소품은 없다.",
        "built_space": "뒤쪽 왼편에 열린 낡은 목제 문짝 하나, 인물 뒤와 오른편에 이어지는 출입구 개구부 하나, 아래쪽에 돌 단과 흙바닥이 보인다. 목조·회벽 재료는 참조와 대체로 맞고 중복 설비나 불가능한 반사는 없다. 배상도는 입구 바로 바깥에 있으나, 왼쪽 전경 인물이 화면을 크게 차지해 단독 클로즈업이 대면 구도로 바뀌었다.",
        "entities": "배상도로 보이는 성인 한국인 남성은 검은 머리, 흰 한복, 검은 조끼를 착용했다. 참조의 검은 모자는 없으며 머리 모양과 얼굴 윤곽에도 차이가 있다. 왼쪽에는 검은 머리와 귀·뺨이 보이는 두 번째 인물이 추가되어 있다. 이 부분 인물의 정확한 신원·나이·성별은 확정할 수 없으며, 등장 허용 목록은 배상도뿐이다. 읽을 수 있는 글자는 없고 자연광과 낮은 채도는 아침 분위기에 부합한다.",
        "hard_violations": [
         "배상도만 허용된 장면에 두 번째 인물의 머리와 얼굴 일부를 왼쪽 전경으로 추가했다."
        ],
        "physics": "배상도의 머리와 목, 어깨는 정상적으로 이어지고 의복도 몸에 자연스럽게 걸려 있다. 전경 인물은 프레임 밖으로 잘려 있어 지지점을 확인할 수 없지만 공중에 떠 있다는 증거는 없다. 비웃는 표정은 실제 얼굴 근육으로 가능한 형태이며, 클로즈업 밖의 발걸음이나 입장 동작은 확인할 수 없다."
       }
      ],
      "all_candidates_fail": false
     },
     "normalized": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 7,
        "verdict_ko": "배상도만 담은 얼굴 클로즈업과 비대칭 비웃음, 열린 관아 입구를 충실히 구현했으나 참조의 검은 모자가 없고 박상현을 향한 시선은 다소 불명확하다."
       },
       {
        "label": "A",
        "score": 2,
        "verdict_ko": "상대를 향한 비웃음은 명확하지만 허용되지 않은 두 번째 인물의 머리를 전경에 추가해 단독 얼굴 클로즈업을 어긴다."
       }
      ],
      "all_candidates_fail": false,
      "readings": [
       {
        "label": "B",
        "direction": "얼굴은 화면 오른쪽으로 조금 돌아가 있고 눈은 카메라 부근의 화면 밖을 향한다. 박상현은 보이지 않아 그를 정확히 바라보는지는 확정할 수 없지만, 화면 밖 상대에게 보내는 눈길로 읽을 수 있다. 한쪽 입꼬리가 더 올라가 있다. 무기나 방향성 있는 소품은 없다.",
        "built_space": "인물은 목조 출입구 바로 앞에 있으며, 화면 오른쪽에 가로 보강대가 있는 낡은 문짝 하나, 그 안쪽에 열린 통로 하나와 뒤편 목조 건물이 보인다. 참조의 목재·회벽과 열린 입구 구성이 이어지며 중복된 설비는 없다. 입구가 인물 뒤에 비스듬히 보이지만 요청한 부드러운 가장자리보다 배경에서 차지하는 비중은 크다. 반사는 없다.",
        "entities": "성인 한국인 남성으로 설정된 배상도 한 명만 보이며 검은 머리, 흰 한복 깃, 검은 조끼가 확인된다. 참조와 의상 계열은 같지만 참조의 검은 모자가 없고, 드러난 헤어스타일과 얼굴 윤곽도 완전히 같지는 않다. 허리 장식과 하의는 클로즈업 밖이므로 평가하지 않는다. 추가 인물이나 읽을 수 있는 글자는 없다. 빛은 자연스럽고 채도가 낮아 아침 설정과 충돌하지 않는다.",
        "hard_violations": [],
        "physics": "머리는 목과 어깨에 자연스럽게 연결되고 옷은 상체에 걸쳐 있다. 얼굴 근육과 눈의 형태도 정상적인 비웃음 연기로 성립한다. 발과 지면 접촉은 프레임 밖이며 부유를 나타내는 증거는 없다. 건물로 들어가는 발걸음은 이 얼굴 클로즈업만으로 확인할 수 없다."
       },
       {
        "label": "A",
        "direction": "배상도의 얼굴과 눈은 화면 왼쪽 전경 인물의 얼굴 쪽을 향한다. 조롱의 대상은 화면에서 명확하며 입꼬리를 올리고 이를 조금 드러낸다. 다만 그 대상을 보이게 만들기 위해 별도의 인물을 추가했다. 무기나 지향성 소품은 없다.",
        "built_space": "뒤쪽 왼편에 열린 낡은 목제 문짝 하나, 인물 뒤와 오른편에 이어지는 출입구 개구부 하나, 아래쪽에 돌 단과 흙바닥이 보인다. 목조·회벽 재료는 참조와 대체로 맞고 중복 설비나 불가능한 반사는 없다. 배상도는 입구 바로 바깥에 있으나, 왼쪽 전경 인물이 화면을 크게 차지해 단독 클로즈업이 대면 구도로 바뀌었다.",
        "entities": "배상도로 보이는 성인 한국인 남성은 검은 머리, 흰 한복, 검은 조끼를 착용했다. 참조의 검은 모자는 없으며 머리 모양과 얼굴 윤곽에도 차이가 있다. 왼쪽에는 검은 머리와 귀·뺨이 보이는 두 번째 인물이 추가되어 있다. 이 부분 인물의 정확한 신원·나이·성별은 확정할 수 없으며, 등장 허용 목록은 배상도뿐이다. 읽을 수 있는 글자는 없고 자연광과 낮은 채도는 아침 분위기에 부합한다.",
        "hard_violations": [
         "배상도만 허용된 장면에 두 번째 인물의 머리와 얼굴 일부를 왼쪽 전경으로 추가했다."
        ],
        "physics": "배상도의 머리와 목, 어깨는 정상적으로 이어지고 의복도 몸에 자연스럽게 걸려 있다. 전경 인물은 프레임 밖으로 잘려 있어 지지점을 확인할 수 없지만 공중에 떠 있다는 증거는 없다. 비웃는 표정은 실제 얼굴 근육으로 가능한 형태이며, 클로즈업 밖의 발걸음이나 입장 동작은 확인할 수 없다."
       }
      ]
     },
     "ok": true
    }
   ],
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "slot_winner_match": true,
   "slot_winner": {
    "gemini-pro": "B",
    "gpt-high": "B"
   },
   "route": "cross_slot_agree"
  },
  "dual": {
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "normalized": {
    "A": 0.486,
    "B": 2.0
   },
   "adjusted": {
    "A": 0.236,
    "B": 2.0
   },
   "violations": {
    "A": [
     "[gemini-pro] invented people (프롬프트에 허용되지 않은 다른 인물의 뒷모습이 화면 왼쪽 전경에 등장)",
     "[gpt-high] 배상도만 허용된 장면에 두 번째 인물의 머리와 얼굴 일부를 왼쪽 전경으로 추가했다."
    ]
   },
   "per_model_winner": {
    "gemini-pro": "B",
    "gpt-high": "B"
   },
   "agreed": true
  },
  "totals": {
   "B": 2000,
   "A": 236
  },
  "selected": "B",
  "ranking": [
   "B",
   "A"
  ],
  "verdicts": [
   {
    "label": "B",
    "score": 2000,
    "verdict_ko": "배상도의 비열한 미소와 클로즈업 앵글을 완벽하게 구현했으며, 프롬프트의 지시대로 다른 인물을 배제하고 배경을 정확하게 표현했습니다."
   },
   {
    "label": "A",
    "score": 236,
    "verdict_ko": "프롬프트에 명시되지 않은 인물(화면 왼쪽 앞의 뒤통수)이 등장하여 엄격한 등장인물 제한 규칙을 위반했습니다.  ★위반: [gemini-pro] invented people (프롬프트에 허용되지 않은 다른 인물의 뒷모습이 화면 왼쪽 전경에 등장) / [gpt-high] 배상도만 허용된 장면에 두 번째 인물의 머리와 얼굴 일부를 왼쪽 전경으로 추가했다."
   }
  ],
  "refs": [
   {
    "label": "PREVIOUS SHOT STILL — a visually related earlier shot of this same place: the location's look, materials, fixed features and lighting mood are LOCKED to this photo; never copy its camera framing. The people visible in that photo are NOT in this shot: never carry their faces, bodies or clothing onto anyone here. Each person in THIS shot is defined solely by the PEOPLE list, their own CHARACTER REFERENCE images and their wardrobe notes. If this photo shows a character who cannot move (dead or unconscious), that character's exact pose, position and orientation are ALSO LOCKED — treat their body as a fixed prop of the set that only the camera moves around.",
    "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/db00517b-165b-4005-a287-f28936b59b95/images/272bd7af-9294-4490-98c6-88a40e9deacc/scene/recipe/S43sh12_sel.png",
    "asset_id": "cb84027e-d99f-4541-9a5c-544fd37c724b",
    "role": "prev_still"
   },
   {
    "label": "CHARACTER REFERENCE — 배상도: the exact person appearing in this shot; match face, hair and build exactly.",
    "path": "<bytes:851718>",
    "asset_id": "034f5325-975e-4474-854b-e48359710bf8",
    "role": "character_ref"
   }
  ],
  "critique_skipped": true,
  "shot_run_produced": false,
  "shot_run_uid": "06aa1431-3fb4-78a1-9883-a6f31041fa30",
  "ref_mode": "prev+엔티티",
  "share_plan": {
   "ref_plan": "prev",
   "prev_anchor_tag": "S43sh12"
  }
 },
 "S43sh16::cine": {
  "applied": true,
  "attempted_at": "2026-09-09T05:53:58.327613+00:00",
  "fingerprint": "74e5ce4cf3b1f9f52115d4453549eedea80dff0fadeb75f2de52cef60a2a499c",
  "fingerprint_version": 2,
  "provider": "mai",
  "endpoint": "openrouter/chat-completions",
  "model": "microsoft/mai-image-2.6",
  "pack": "24.202608252115",
  "source_file": "S43sh16_sel.png",
  "source_sha256": "545d79872e3df2c050b80b335f46d66abca14beab409c2ab98b38affcd0ffe25",
  "file": "S43sh16_cine.png",
  "staged_sha256": "b8c233e8fb791e801f459b40fa7dc1df85e7c468f14894b62cdb139050414902",
  "latency_ms": 24539
 },
 "S44sh6::signage": {
  "fp": "a5a33e8d0299c18e",
  "inscriptions": [],
  "cues": [],
  "dropped": []
 },
 "era_assess::44ce40bdd1afc543": {
  "subjects": [
   {
    "subject_native": "20세기 초 한국 형무소 취조실 내부",
    "search_terms_native": [
     "일제강점기 형무소 취조실",
     "서대문형무소 취조실 내부",
     "근대 감옥 취조실 복원"
    ],
    "language_lock_native": "반드시 한국어 검색어만 사용하고 다른 언어로 번역하거나 추가하지 마십시오.",
    "reason_ko": "현대식 취조실(단방향 거울, 현대식 철제 책상)로 묘사되기 쉬우므로, 20세기 초 형무소 특유의 목재 마감 및 가구 형식을 확인해야 합니다."
   }
  ],
  "subject_text": "감옥 취조실 내부\n탁자와 마주 놓인 의자가 있는 어두운 취조실. 뒤편 벽은 짙은 그림자에 잠기며, 손잡이 달린 문이 한쪽에 있다.",
  "identity": "canonical",
  "scope_id": "L33",
  "scope_role": "location_interior",
  "scope_sha": "98b83b914237bc14"
 },
 "era_fail::8fefe9b1e707a533": {
  "stage": "research",
  "subject": "20세기 초 한국 형무소 취조실 내부",
  "terms": [
   "일제강점기 형무소 취조실",
   "서대문형무소 취조실 내부",
   "근대 감옥 취조실 복원"
  ],
  "status": "no_usable",
  "queries": [
   [
    "서대문형무소역사관 지하 취조실 고문실 복원"
   ]
  ],
  "candidate_urls": [
   "https://mediahub.seoul.go.kr/uploads/mediahub/2024/02/tWsczQtCUUsNqihHfiuxqvIxJVbqUCtQ.jpg",
   "https://img.koreatimes.co.kr/upload/newsV2/images/m111%280%29.jpg?w=728",
   "https://ak-d.tripcdn.com/images/0EQ6j12000dvh1qso33EF.jpg?proc=source%2Ftrip",
   "https://www.korean-culture.org/eng/webzine/202204/img/sub/morephotos07.jpg"
  ],
  "coarse": {
   "eligible": [],
   "chosen_index": 0,
   "reason": "종류·보임·기준을 다 만족하는 후보가 없다 — 이 라운드에선 안 고른다",
   "single_judge": true,
   "rejected_judges": {}
  },
  "verdicts": [
   {
    "index": 1,
    "object_type_match": "yes",
    "visible": true,
    "criteria_match": "unsure",
    "similarity": 95
   },
   {
    "index": 2,
    "object_type_match": "yes",
    "visible": true,
    "criteria_match": "unsure",
    "similarity": 90
   },
   {
    "index": 3,
    "object_type_match": "no",
    "visible": true,
    "criteria_match": "unsure",
    "similarity": 10
   },
   {
    "index": 4,
    "object_type_match": "yes",
    "visible": true,
    "criteria_match": "unsure",
    "similarity": 85
   }
  ],
  "chosen_reason_ko": "",
  "attempts": 5
 },
 "S44sh6::bgfirst_bg": {
  "input_fingerprint": "c4562c41d54844e9",
  "prompt": "Create the EMPTY BACKGROUND PLATE for one film shot — NO PEOPLE, no figures, no body parts, no sketch lines, no arrows anywhere.\n\"Empty\" means no people only: KEEP the location's inherent occupants and stock that define the place — animals in an animal shelter, pen or farm, goods in a market, moored boats in a harbour — unless the shot text explicitly removes them.\nThe FIRST attached image is a thin-line storyboard sketch: use ONLY its camera angle, horizon, perspective and the placement/size of buildings and set masses — ignore the sketched people and arrows entirely. The SECOND attached image (LOCATION PHOTOGRAPH) is the real place: take its architecture, materials, signage and fixed features, and RE-PROJECT them into the sketch's camera. If the photograph's camera differs from the sketch's, the sketch's camera wins.\nHUMAN-SCALE CALIBRATION: derive every structure's true size from human-scale elements — a door ≈ 2m, a window ≈ 1–1.5m wide, one storey ≈ 2.5–3m; never inflate a small structure or shrink a large one.\n\nSHOT TEXT this background must serve (Korean): 토미이시의 손에 든 지휘봉 끝이 그림자 속 배상도 쪽을 비스듬히 가리킨 찰나.\n\nLOCATION (lock): The table area inside a dim prison interrogation room at night, with a deep shadowed standing space behind the detainee.\n\nTIME OF DAY (lock): night.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: medium shot\n- FRAME LAYOUT: Rear position behind the seated pair toward which the baton points in the upper-right of the frame, background.\n- KEY BACKGROUND ELEMENTS: Interrogation table (Between the seated interrogator and prisoner) — Its top and near edge are viewed obliquely at a shallow downward angle; used as The lower edge establishes hand height and connects the near gesture to the room beyond; Command baton (Held by 토미이시 and pointing toward 배상도) — The tip angles from the lower-left foreground toward the upper-right rear; used as A directional bridge rather than an oversized isolated object.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Keep the interrogation room dark with deep rear shadows and restrained saturation, retaining just enough tonal separation to connect the baton tip to 배상도.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nRender ONE photorealistic empty location photograph, 16:9, that this shot can be staged inside later. No readable writing anywhere: surfaces that would carry writing may be present, but stage any wording out of legibility — an oblique angle, distance, shallow focus. No captions, watermarks or overlay text.",
  "effective_prompt": "Create the EMPTY BACKGROUND PLATE for one film shot — NO PEOPLE, no figures, no body parts, no sketch lines, no arrows anywhere.\n\"Empty\" means no people only: KEEP the location's inherent occupants and stock that define the place — animals in an animal shelter, pen or farm, goods in a market, moored boats in a harbour — unless the shot text explicitly removes them.\nThe FIRST attached image is a thin-line storyboard sketch: use ONLY its camera angle, horizon, perspective and the placement/size of buildings and set masses — ignore the sketched people and arrows entirely. The SECOND attached image (LOCATION PHOTOGRAPH) is the real place: take its architecture, materials, signage and fixed features, and RE-PROJECT them into the sketch's camera. If the photograph's camera differs from the sketch's, the sketch's camera wins.\nHUMAN-SCALE CALIBRATION: derive every structure's true size from human-scale elements — a door ≈ 2m, a window ≈ 1–1.5m wide, one storey ≈ 2.5–3m; never inflate a small structure or shrink a large one.\n\nSHOT TEXT this background must serve (Korean): 토미이시의 손에 든 지휘봉 끝이 그림자 속 배상도 쪽을 비스듬히 가리킨 찰나.\n\nLOCATION (lock): The table area inside a dim prison interrogation room at night, with a deep shadowed standing space behind the detainee.\n\nTIME OF DAY (lock): night.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: medium shot\n- FRAME LAYOUT: Rear position behind the seated pair toward which the baton points in the upper-right of the frame, background.\n- KEY BACKGROUND ELEMENTS: Interrogation table (Between the seated interrogator and prisoner) — Its top and near edge are viewed obliquely at a shallow downward angle; used as The lower edge establishes hand height and connects the near gesture to the room beyond; Command baton (Held by 토미이시 and pointing toward 배상도) — The tip angles from the lower-left foreground toward the upper-right rear; used as A directional bridge rather than an oversized isolated object.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Keep the interrogation room dark with deep rear shadows and restrained saturation, retaining just enough tonal separation to connect the baton tip to 배상도.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nRender ONE photorealistic empty location photograph, 16:9, that this shot can be staged inside later. No readable writing anywhere: surfaces that would carry writing may be present, but stage any wording out of legibility — an oblique angle, distance, shallow focus. No captions, watermarks or overlay text.",
  "bg_path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/db00517b-165b-4005-a287-f28936b59b95/images/272bd7af-9294-4490-98c6-88a40e9deacc/scene/recipe/S44sh6__bgfirst_bg.png",
  "asset_id": "eff0cff2-1353-4d8e-91b1-3232859ab071",
  "input_asset_ids": [
   "a89cc4ee-711d-445b-972e-49369d870a7c",
   "00197208-b612-467b-a594-99866513b16a"
  ]
 },
 "S44sh6": {
  "input_fingerprint": "bcc1726d165df4d2",
  "prompt": "Create ONE FINAL photorealistic live-action film still of the moment below — early 20th-century Korea, Jeolla region; all people are Korean unless explicitly stated otherwise. TIME OF DAY (lock): night.\n\nSHOT TEXT (authoritative, Korean): 토미이시의 손에 든 지휘봉 끝이 그림자 속 배상도 쪽을 비스듬히 가리킨 찰나.\n\nLOCATION (lock): The table area inside a dim prison interrogation room at night, with a deep shadowed standing space behind the detainee. The shot takes place here — the attached LOCATION PHOTOGRAPH shows the exact spot.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: medium shot\n- FRAME LAYOUT: Rear position behind the seated pair toward which the baton points in the upper-right of the frame, background.\n- KEY BACKGROUND ELEMENTS: Interrogation table (Between the seated interrogator and prisoner) — Its top and near edge are viewed obliquely at a shallow downward angle; used as The lower edge establishes hand height and connects the near gesture to the room beyond; Command baton (Held by 토미이시 and pointing toward 배상도) — The tip angles from the lower-left foreground toward the upper-right rear; used as A directional bridge rather than an oversized isolated object.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Keep the interrogation room dark with deep rear shadows and restrained saturation, retaining just enough tonal separation to connect the baton tip to 배상도.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): A table divides the dim interrogation room; cigarette smoke hangs in the cold darkness. 토미이시: He sits with a cigarette and a command baton in hand.\n\nTHE HAND THAT IS DOING THIS: the object at the centre of this shot is being held, operated, read, aimed or handed over by 토미이시 right now, so 토미이시's hand — and as much of the wrist and forearm as the framing reaches — is in the frame, gripping or resting on that object exactly the way its real use requires. Match that hand to 토미이시: its size, build, skin, age, grooming, sleeve and anything worn on it belong to that person and to no one else.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 토미이시 (일본인 남성, 성인 나이대, 자연스러운 검은 머리) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.",
  "roll_prompts": {
   "A": "Stage the shot. The FIRST attached image (SHOT BACKGROUND) is the finished empty background of this shot — keep it EXACTLY: its camera, perspective, architecture, lighting and every fixed feature stay untouched. The SECOND attached image (LAYOUT SKETCH) tells you ONLY where the people go: each sketched person's position, screen size, pose and the gaze/motion arrows. Ignore the sketch's background lines. The CHARACTER REFERENCE photographs show the real people.\nPlace the real people into the background at exactly the sketched positions, sizes and poses, following the arrow directions. No sketch lines or arrows may remain.\n\nCreate ONE FINAL photorealistic live-action film still of the moment below — early 20th-century Korea, Jeolla region; all people are Korean unless explicitly stated otherwise. TIME OF DAY (lock): night.\n\nSHOT TEXT (authoritative, Korean): 토미이시의 손에 든 지휘봉 끝이 그림자 속 배상도 쪽을 비스듬히 가리킨 찰나.\n\nLOCATION (lock): The table area inside a dim prison interrogation room at night, with a deep shadowed standing space behind the detainee. The shot takes place here — the FIRST attached image (SHOT BACKGROUND) is this exact place, already built: its ground, structures, horizon, materials and lighting are the finished truth of this location and must not be redesigned or replaced. No location photograph is attached — read the place from that image alone, and add no scenery, structure, vehicle or fixture that it does not already show. This lock governs the place only; the figures in the shot follow the staging and pose instructions.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: medium shot\n- FRAME LAYOUT: Rear position behind the seated pair toward which the baton points in the upper-right of the frame, background.\n- KEY BACKGROUND ELEMENTS: Interrogation table (Between the seated interrogator and prisoner) — Its top and near edge are viewed obliquely at a shallow downward angle; used as The lower edge establishes hand height and connects the near gesture to the room beyond; Command baton (Held by 토미이시 and pointing toward 배상도) — The tip angles from the lower-left foreground toward the upper-right rear; used as A directional bridge rather than an oversized isolated object.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Keep the interrogation room dark with deep rear shadows and restrained saturation, retaining just enough tonal separation to connect the baton tip to 배상도.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): A table divides the dim interrogation room; cigarette smoke hangs in the cold darkness. 토미이시: He sits with a cigarette and a command baton in hand.\n\nTHE HAND THAT IS DOING THIS: the object at the centre of this shot is being held, operated, read, aimed or handed over by 토미이시 right now, so 토미이시's hand — and as much of the wrist and forearm as the framing reaches — is in the frame, gripping or resting on that object exactly the way its real use requires. Match that hand to 토미이시: its size, build, skin, age, grooming, sleeve and anything worn on it belong to that person and to no one else.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 토미이시 (일본인 남성, 성인 나이대, 자연스러운 검은 머리) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.",
   "B": "Create ONE FINAL photorealistic live-action film still of the moment below — early 20th-century Korea, Jeolla region; all people are Korean unless explicitly stated otherwise. TIME OF DAY (lock): night.\n\nSHOT TEXT (authoritative, Korean): 토미이시의 손에 든 지휘봉 끝이 그림자 속 배상도 쪽을 비스듬히 가리킨 찰나.\n\nLOCATION (lock): The table area inside a dim prison interrogation room at night, with a deep shadowed standing space behind the detainee. The shot takes place here — the attached LOCATION PHOTOGRAPH shows the exact spot.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: medium shot\n- FRAME LAYOUT: Rear position behind the seated pair toward which the baton points in the upper-right of the frame, background.\n- KEY BACKGROUND ELEMENTS: Interrogation table (Between the seated interrogator and prisoner) — Its top and near edge are viewed obliquely at a shallow downward angle; used as The lower edge establishes hand height and connects the near gesture to the room beyond; Command baton (Held by 토미이시 and pointing toward 배상도) — The tip angles from the lower-left foreground toward the upper-right rear; used as A directional bridge rather than an oversized isolated object.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Keep the interrogation room dark with deep rear shadows and restrained saturation, retaining just enough tonal separation to connect the baton tip to 배상도.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): A table divides the dim interrogation room; cigarette smoke hangs in the cold darkness. 토미이시: He sits with a cigarette and a command baton in hand.\n\nTHE HAND THAT IS DOING THIS: the object at the centre of this shot is being held, operated, read, aimed or handed over by 토미이시 right now, so 토미이시's hand — and as much of the wrist and forearm as the framing reaches — is in the frame, gripping or resting on that object exactly the way its real use requires. Match that hand to 토미이시: its size, build, skin, age, grooming, sleeve and anything worn on it belong to that person and to no one else.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 토미이시 (일본인 남성, 성인 나이대, 자연스러운 검은 머리) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.\n\nCOMPOSITION VARIATION (this candidate only): keep the place, the people and the moment exactly as specified, and keep from the CAMERA & FRAME contract its subject scale, screen placement and key background elements exactly. For this candidate only, its camera angle and height are deliberately left open: choose a camera position distinctly different from the obvious one, as a film director picking a second setup on the same blocking. Stay true to any explicit framing the SHOT TEXT itself dictates."
  },
  "roll_refs": {
   "A": [
    {
     "label": "SHOT BACKGROUND",
     "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/db00517b-165b-4005-a287-f28936b59b95/images/272bd7af-9294-4490-98c6-88a40e9deacc/scene/recipe/S44sh6__bgfirst_bg.png",
     "asset_id": "eff0cff2-1353-4d8e-91b1-3232859ab071",
     "role": "bgfirst_bg"
    },
    {
     "label": "LAYOUT SKETCH (people placement only)",
     "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/db00517b-165b-4005-a287-f28936b59b95/images/272bd7af-9294-4490-98c6-88a40e9deacc/conti/conti_S44sh6.png",
     "asset_id": "a89cc4ee-711d-445b-972e-49369d870a7c",
     "role": "conti_light"
    },
    {
     "label": "CHARACTER REFERENCE — 토미이시: the exact person appearing in this shot; match face, hair and build exactly.",
     "path": "<bytes:797699>",
     "asset_id": "c861ba0f-8750-4262-8324-25cf5c6c9dcd",
     "role": "character_ref"
    }
   ],
   "B": [
    {
     "label": "LOCATION PHOTOGRAPH — the exact place of this shot: its architecture, materials, fixed features and lighting mood are spatial truth; stage the moment inside this place. Never copy its camera framing.",
     "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/db00517b-165b-4005-a287-f28936b59b95/episodes/272bd7af-9294-4490-98c6-88a40e9deacc/images/background_chain/L33B02.png",
     "asset_id": "00197208-b612-467b-a594-99866513b16a",
     "role": "location_plate"
    },
    {
     "label": "CHARACTER REFERENCE — 토미이시: the exact person appearing in this shot; match face, hair and build exactly.",
     "path": "<bytes:797699>",
     "asset_id": "c861ba0f-8750-4262-8324-25cf5c6c9dcd",
     "role": "character_ref"
    }
   ]
  },
  "shot_run_spend_attempt_count": 1,
  "judge_flip_skipped": "judge_fn_owns_order",
  "initial_roll_all_fail": true,
  "readings": [
   {
    "label": "A",
    "direction": "전경 좌측의 토미이시가 든 지휘봉이 우측 상단을 향하고 있으나, 특정 인물보다는 빈 공간을 가리키는 것으로 보임.",
    "built_space": "어두운 취조실 내부. 테이블이 가로로 넓게 배치되어 있고, 두 명의 인물이 나란히 앞쪽 의자에 앉아 있음. 배경 쪽에 빈 의자가 하나 놓여 있음.",
    "entities": "군복을 입은 토미이시(좌측), 등을 돌리고 앉은 인물, 배경 그림자 속에 서 있는 인물까지 총 3명의 인물이 등장함.",
    "hard_violations": [
     "[gemini-pro] 프롬프트에 명시되지 않은 추가 인물이 렌더링되어 총 3명이 등장함 (Extra bodies).",
     "[gemini-pro] 지휘봉을 쥔 오른쪽 손 아래에 불이 붙은 담배가 아무런 지지 없이 허공에 떠 있음 (Unsupported object).",
     "[gpt-high] 수감자를 토미이시와 탁자의 같은 쪽에 앉혀, 탁자가 심문자와 수감자 사이에 있어야 한다는 배치를 어겼다.",
     "[gpt-high] 참조의 의자 두 개를 전경 두 개와 맞은편 빈 의자 한 개, 총 세 개로 늘렸다.",
     "[gpt-high] 토미이시 입의 담배 외에 탁자 부근에 담배를 하나 더 추가했다."
    ],
    "physics": "오른손 아래에 담배가 허공에 떠 있어 물리적으로 불가능한 묘사가 있음."
   },
   {
    "label": "B",
    "direction": "전경 좌측의 토미이시가 쥔 지휘봉이 테이블 건너편에 앉아 있는 배상도를 정확히 가리키고 있음.",
    "built_space": "어두운 취조실 내부. 테이블을 사이에 두고 전경의 심문자와 배경의 수감자가 마주 보고 있는 올바른 배치임.",
    "entities": "전경에 지휘봉을 든 토미이시, 맞은편에 앉은 수감자(배상도), 그리고 배경 그림자 속에 서 있는 또 다른 일본군 장교까지 총 3명이 등장함.",
    "hard_violations": [
     "[gemini-pro] 프롬프트에 없는 3번째 인물이 추가되었으며, 특히 배경에 서 있는 인물은 군복과 외양이 토미이시와 동일하여 캐릭터가 복제됨 (Duplicated/Extra bodies).",
     "[gpt-high] 배상도로 읽히는 후경 인물을 지정된 오른쪽 위가 아닌 중앙 왼쪽에 배치하여 명시된 인물 위치를 어겼다."
    ],
    "physics": "두 손으로 각각 담배와 지휘봉을 쥐고 있는 모습이나 인물들이 의자에 앉아 있는 자세는 자연스러움."
   }
  ],
  "cross_model_order": {
   "policy": "cross_model_order_v2_slot0_fwd_slot1_rev_parallel2",
   "slots": [
    {
     "model": "gemini-pro",
     "order": "forward",
     "display_to_canonical": {
      "A": "A",
      "B": "B"
     },
     "raw": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 0,
        "verdict_ko": "지휘봉이 테이블 건너편의 대상을 올바르게 가리키고 있으나, 프롬프트에 없는 3번째 인물이 추가되었고 토미이시 캐릭터가 복제되어 심각한 오류가 발생했습니다."
       },
       {
        "label": "A",
        "score": 0,
        "verdict_ko": "지휘봉을 쥔 손 아래에 담배가 허공에 떠 있는 물리적 오류가 있으며, 불필요한 인물이 추가로 배치되어 지침을 위반했습니다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "전경 좌측의 토미이시가 든 지휘봉이 우측 상단을 향하고 있으나, 특정 인물보다는 빈 공간을 가리키는 것으로 보임.",
        "built_space": "어두운 취조실 내부. 테이블이 가로로 넓게 배치되어 있고, 두 명의 인물이 나란히 앞쪽 의자에 앉아 있음. 배경 쪽에 빈 의자가 하나 놓여 있음.",
        "entities": "군복을 입은 토미이시(좌측), 등을 돌리고 앉은 인물, 배경 그림자 속에 서 있는 인물까지 총 3명의 인물이 등장함.",
        "hard_violations": [
         "프롬프트에 명시되지 않은 추가 인물이 렌더링되어 총 3명이 등장함 (Extra bodies).",
         "지휘봉을 쥔 오른쪽 손 아래에 불이 붙은 담배가 아무런 지지 없이 허공에 떠 있음 (Unsupported object)."
        ],
        "physics": "오른손 아래에 담배가 허공에 떠 있어 물리적으로 불가능한 묘사가 있음."
       },
       {
        "label": "B",
        "direction": "전경 좌측의 토미이시가 쥔 지휘봉이 테이블 건너편에 앉아 있는 배상도를 정확히 가리키고 있음.",
        "built_space": "어두운 취조실 내부. 테이블을 사이에 두고 전경의 심문자와 배경의 수감자가 마주 보고 있는 올바른 배치임.",
        "entities": "전경에 지휘봉을 든 토미이시, 맞은편에 앉은 수감자(배상도), 그리고 배경 그림자 속에 서 있는 또 다른 일본군 장교까지 총 3명이 등장함.",
        "hard_violations": [
         "프롬프트에 없는 3번째 인물이 추가되었으며, 특히 배경에 서 있는 인물은 군복과 외양이 토미이시와 동일하여 캐릭터가 복제됨 (Duplicated/Extra bodies)."
        ],
        "physics": "두 손으로 각각 담배와 지휘봉을 쥐고 있는 모습이나 인물들이 의자에 앉아 있는 자세는 자연스러움."
       }
      ],
      "all_candidates_fail": true
     },
     "normalized": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 0,
        "verdict_ko": "지휘봉이 테이블 건너편의 대상을 올바르게 가리키고 있으나, 프롬프트에 없는 3번째 인물이 추가되었고 토미이시 캐릭터가 복제되어 심각한 오류가 발생했습니다."
       },
       {
        "label": "A",
        "score": 0,
        "verdict_ko": "지휘봉을 쥔 손 아래에 담배가 허공에 떠 있는 물리적 오류가 있으며, 불필요한 인물이 추가로 배치되어 지침을 위반했습니다."
       }
      ],
      "all_candidates_fail": true,
      "readings": [
       {
        "label": "A",
        "direction": "전경 좌측의 토미이시가 든 지휘봉이 우측 상단을 향하고 있으나, 특정 인물보다는 빈 공간을 가리키는 것으로 보임.",
        "built_space": "어두운 취조실 내부. 테이블이 가로로 넓게 배치되어 있고, 두 명의 인물이 나란히 앞쪽 의자에 앉아 있음. 배경 쪽에 빈 의자가 하나 놓여 있음.",
        "entities": "군복을 입은 토미이시(좌측), 등을 돌리고 앉은 인물, 배경 그림자 속에 서 있는 인물까지 총 3명의 인물이 등장함.",
        "hard_violations": [
         "프롬프트에 명시되지 않은 추가 인물이 렌더링되어 총 3명이 등장함 (Extra bodies).",
         "지휘봉을 쥔 오른쪽 손 아래에 불이 붙은 담배가 아무런 지지 없이 허공에 떠 있음 (Unsupported object)."
        ],
        "physics": "오른손 아래에 담배가 허공에 떠 있어 물리적으로 불가능한 묘사가 있음."
       },
       {
        "label": "B",
        "direction": "전경 좌측의 토미이시가 쥔 지휘봉이 테이블 건너편에 앉아 있는 배상도를 정확히 가리키고 있음.",
        "built_space": "어두운 취조실 내부. 테이블을 사이에 두고 전경의 심문자와 배경의 수감자가 마주 보고 있는 올바른 배치임.",
        "entities": "전경에 지휘봉을 든 토미이시, 맞은편에 앉은 수감자(배상도), 그리고 배경 그림자 속에 서 있는 또 다른 일본군 장교까지 총 3명이 등장함.",
        "hard_violations": [
         "프롬프트에 없는 3번째 인물이 추가되었으며, 특히 배경에 서 있는 인물은 군복과 외양이 토미이시와 동일하여 캐릭터가 복제됨 (Duplicated/Extra bodies)."
        ],
        "physics": "두 손으로 각각 담배와 지휘봉을 쥐고 있는 모습이나 인물들이 의자에 앉아 있는 자세는 자연스러움."
       }
      ]
     },
     "ok": true
    },
    {
     "model": "gpt-high",
     "order": "reverse",
     "display_to_canonical": {
      "A": "B",
      "B": "A"
     },
     "raw": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 2,
        "verdict_ko": "탁자가 심문자와 수감자 사이에 놓인 점은 맞지만, 지휘봉이 배상도가 아닌 오른쪽 빈 공간을 가리키고 배상도의 후경 위치도 지시와 다릅니다."
       },
       {
        "label": "B",
        "score": 3,
        "verdict_ko": "앉은 두 사람 뒤에서 오른쪽 후경의 배상도로 이어지는 지휘봉 방향은 더 충실하지만, 수감자의 좌석 배치와 추가 의자 때문에 그대로 사용할 수는 없습니다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "토미이시의 오른손에 잡힌 지휘봉은 화면 중앙 아래에서 오른쪽 위로 향하지만, 끝과 그 연장선은 수감자 오른쪽의 빈 어둠을 향한다. 배상도로 읽히는 서 있는 남자는 오히려 중앙 왼쪽 후경에 있어 지휘봉의 지시 대상과 연결되지 않는다. 토미이시는 탁자 건너 수감자를 바라보고, 수감자는 눈을 내리깔며, 뒤의 남자는 전경 쪽을 본다.",
        "built_space": "낡은 직사각형 목재 탁자 한 개, 오른쪽 문틀과 문 한 개, 회벽과 어두운 하단 띠가 보인다. 두 사람은 탁자의 서로 반대편에 앉아 있어 탁자가 심문자와 수감자를 나누는 관계는 맞다. 의자는 몸과 탁자에 가려 정확한 개수를 셀 수 없다. 카메라는 토미이시 뒤에 있지만 수감자는 정면으로 보이므로 앉은 두 사람 모두의 뒤에서 보는 구도가 아니다. 서 있는 남자도 지정된 오른쪽 위가 아닌 중앙 왼쪽에 있다. 반사상은 없다.",
        "entities": "성인 동아시아계 남성 세 명이 보이며, 심문자·수감자·후경의 배상도라는 장면상 역할로 읽힌다. 토미이시는 검은 머리, 짧은 수염, 갈색 군복, 붉은 견장과 가죽 어깨끈을 갖췄지만 머리가 참조보다 훨씬 짧고 군모는 없다. 보이는 얼굴만으로 일본인 여부를 확정할 수는 없다. 수감자는 검은 머리와 짙은 회색 옷, 후경 남자는 군복 차림이다. 손에 든 담배 한 개와 목재 지휘봉 한 개, 담배 연기가 확인된다. 어두운 저채도 실내이며 읽을 수 있는 글자는 없다.",
        "hard_violations": [
         "배상도로 읽히는 후경 인물을 지정된 오른쪽 위가 아닌 중앙 왼쪽에 배치하여 명시된 인물 위치를 어겼다."
        ],
        "physics": "지휘봉은 토미이시의 오른손이 손잡이를 감싸 지지하고, 담배는 왼손 손가락 사이에 잡혀 있다. 손목과 팔의 각도는 앉아서 봉을 들어 가리키는 동작으로 가능하다. 앉은 인물들의 좌면은 가려져 있으나 몸통과 다리 배치에 부유의 증거는 없으며, 뒤의 남자는 바닥에 내려선 다리로 서 있다. 탁자는 바닥으로 내려오는 다리로 지지된다."
       },
       {
        "label": "B",
        "direction": "토미이시가 잡은 지휘봉은 왼쪽 아래에서 오른쪽 위 후경으로 기울어 있다. 봉의 축을 연장하면 오른쪽 뒤에 서 있는 배상도의 상체 부근으로 이어져 목표 인물과 방향이 연결된다. 토미이시의 얼굴도 그쪽으로 돌아가 있다. 수감자는 뒷머리만 보여 정확한 시선은 확인할 수 없고, 배상도는 앉은 사람들 쪽을 보고 있다.",
        "built_space": "목재 탁자 한 개와 오른쪽 출입문 한 개, 회벽과 어두운 하단 띠가 보인다. 의자는 전경에 두 개, 탁자 건너편에 빈 의자 한 개로 총 세 개가 보여 참조의 두 개보다 늘었다. 토미이시와 수감자가 탁자의 같은 쪽에 나란히 앉아 있어 탁자가 둘 사이를 나누지 못한다. 다만 두 사람 뒤에서 탁자 상판과 가까운 가장자리를 비스듬히 내려다보며 오른쪽 위 후경의 배상도를 보는 중간 크기 구도는 A보다 지시에 가깝다. 반사상은 없다.",
        "entities": "성인 동아시아계 남성 세 명이 보인다. 토미이시는 참조와 유사한 검은 머리와 짧은 수염, 갈색 군복, 붉은 견장, 가죽 어깨끈을 갖췄고 군모는 없다. 국적은 외관만으로 확정할 수 없다. 수감자는 검은 머리와 밝은 회색 옷을 입고, 배상도는 어두운 회색 셔츠와 바지 차림으로 얼굴과 몸의 입체감이 남아 있다. 손에는 가느다란 지휘봉이 있으나 끝이 화살촉처럼 뾰족하다. 담배가 입에 한 개 있고 팔 옆 탁자 부근에도 별도의 담배 한 개가 보여 하나를 추가했다. 담배 연기와 어두운 실내는 구현되었고 읽을 수 있는 글자는 없다.",
        "hard_violations": [
         "수감자를 토미이시와 탁자의 같은 쪽에 앉혀, 탁자가 심문자와 수감자 사이에 있어야 한다는 배치를 어겼다.",
         "참조의 의자 두 개를 전경 두 개와 맞은편 빈 의자 한 개, 총 세 개로 늘렸다.",
         "토미이시 입의 담배 외에 탁자 부근에 담배를 하나 더 추가했다."
        ],
        "physics": "지휘봉 손잡이는 토미이시의 오른손에 잡혀 있고 팔과 손목으로 지지된다. 입의 담배는 입술에 물려 있으며, 두 번째 담배는 소매 옆 탁자 면에 놓인 것으로 보인다. 두 전경 인물은 각각 별도 의자에 앉아 등받이를 등 뒤에 두고 있어 착석 자체는 가능하다. 후경 남자의 발은 전경 인물에 가려지지만 다리가 바닥 쪽으로 내려오며 부유하는 자세는 아니다. 탁자와 의자에도 명백한 지지 불가능성은 보이지 않는다."
       }
      ],
      "all_candidates_fail": true
     },
     "normalized": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 2,
        "verdict_ko": "탁자가 심문자와 수감자 사이에 놓인 점은 맞지만, 지휘봉이 배상도가 아닌 오른쪽 빈 공간을 가리키고 배상도의 후경 위치도 지시와 다릅니다."
       },
       {
        "label": "A",
        "score": 3,
        "verdict_ko": "앉은 두 사람 뒤에서 오른쪽 후경의 배상도로 이어지는 지휘봉 방향은 더 충실하지만, 수감자의 좌석 배치와 추가 의자 때문에 그대로 사용할 수는 없습니다."
       }
      ],
      "all_candidates_fail": true,
      "readings": [
       {
        "label": "B",
        "direction": "토미이시의 오른손에 잡힌 지휘봉은 화면 중앙 아래에서 오른쪽 위로 향하지만, 끝과 그 연장선은 수감자 오른쪽의 빈 어둠을 향한다. 배상도로 읽히는 서 있는 남자는 오히려 중앙 왼쪽 후경에 있어 지휘봉의 지시 대상과 연결되지 않는다. 토미이시는 탁자 건너 수감자를 바라보고, 수감자는 눈을 내리깔며, 뒤의 남자는 전경 쪽을 본다.",
        "built_space": "낡은 직사각형 목재 탁자 한 개, 오른쪽 문틀과 문 한 개, 회벽과 어두운 하단 띠가 보인다. 두 사람은 탁자의 서로 반대편에 앉아 있어 탁자가 심문자와 수감자를 나누는 관계는 맞다. 의자는 몸과 탁자에 가려 정확한 개수를 셀 수 없다. 카메라는 토미이시 뒤에 있지만 수감자는 정면으로 보이므로 앉은 두 사람 모두의 뒤에서 보는 구도가 아니다. 서 있는 남자도 지정된 오른쪽 위가 아닌 중앙 왼쪽에 있다. 반사상은 없다.",
        "entities": "성인 동아시아계 남성 세 명이 보이며, 심문자·수감자·후경의 배상도라는 장면상 역할로 읽힌다. 토미이시는 검은 머리, 짧은 수염, 갈색 군복, 붉은 견장과 가죽 어깨끈을 갖췄지만 머리가 참조보다 훨씬 짧고 군모는 없다. 보이는 얼굴만으로 일본인 여부를 확정할 수는 없다. 수감자는 검은 머리와 짙은 회색 옷, 후경 남자는 군복 차림이다. 손에 든 담배 한 개와 목재 지휘봉 한 개, 담배 연기가 확인된다. 어두운 저채도 실내이며 읽을 수 있는 글자는 없다.",
        "hard_violations": [
         "배상도로 읽히는 후경 인물을 지정된 오른쪽 위가 아닌 중앙 왼쪽에 배치하여 명시된 인물 위치를 어겼다."
        ],
        "physics": "지휘봉은 토미이시의 오른손이 손잡이를 감싸 지지하고, 담배는 왼손 손가락 사이에 잡혀 있다. 손목과 팔의 각도는 앉아서 봉을 들어 가리키는 동작으로 가능하다. 앉은 인물들의 좌면은 가려져 있으나 몸통과 다리 배치에 부유의 증거는 없으며, 뒤의 남자는 바닥에 내려선 다리로 서 있다. 탁자는 바닥으로 내려오는 다리로 지지된다."
       },
       {
        "label": "A",
        "direction": "토미이시가 잡은 지휘봉은 왼쪽 아래에서 오른쪽 위 후경으로 기울어 있다. 봉의 축을 연장하면 오른쪽 뒤에 서 있는 배상도의 상체 부근으로 이어져 목표 인물과 방향이 연결된다. 토미이시의 얼굴도 그쪽으로 돌아가 있다. 수감자는 뒷머리만 보여 정확한 시선은 확인할 수 없고, 배상도는 앉은 사람들 쪽을 보고 있다.",
        "built_space": "목재 탁자 한 개와 오른쪽 출입문 한 개, 회벽과 어두운 하단 띠가 보인다. 의자는 전경에 두 개, 탁자 건너편에 빈 의자 한 개로 총 세 개가 보여 참조의 두 개보다 늘었다. 토미이시와 수감자가 탁자의 같은 쪽에 나란히 앉아 있어 탁자가 둘 사이를 나누지 못한다. 다만 두 사람 뒤에서 탁자 상판과 가까운 가장자리를 비스듬히 내려다보며 오른쪽 위 후경의 배상도를 보는 중간 크기 구도는 A보다 지시에 가깝다. 반사상은 없다.",
        "entities": "성인 동아시아계 남성 세 명이 보인다. 토미이시는 참조와 유사한 검은 머리와 짧은 수염, 갈색 군복, 붉은 견장, 가죽 어깨끈을 갖췄고 군모는 없다. 국적은 외관만으로 확정할 수 없다. 수감자는 검은 머리와 밝은 회색 옷을 입고, 배상도는 어두운 회색 셔츠와 바지 차림으로 얼굴과 몸의 입체감이 남아 있다. 손에는 가느다란 지휘봉이 있으나 끝이 화살촉처럼 뾰족하다. 담배가 입에 한 개 있고 팔 옆 탁자 부근에도 별도의 담배 한 개가 보여 하나를 추가했다. 담배 연기와 어두운 실내는 구현되었고 읽을 수 있는 글자는 없다.",
        "hard_violations": [
         "수감자를 토미이시와 탁자의 같은 쪽에 앉혀, 탁자가 심문자와 수감자 사이에 있어야 한다는 배치를 어겼다.",
         "참조의 의자 두 개를 전경 두 개와 맞은편 빈 의자 한 개, 총 세 개로 늘렸다.",
         "토미이시 입의 담배 외에 탁자 부근에 담배를 하나 더 추가했다."
        ],
        "physics": "지휘봉 손잡이는 토미이시의 오른손에 잡혀 있고 팔과 손목으로 지지된다. 입의 담배는 입술에 물려 있으며, 두 번째 담배는 소매 옆 탁자 면에 놓인 것으로 보인다. 두 전경 인물은 각각 별도 의자에 앉아 등받이를 등 뒤에 두고 있어 착석 자체는 가능하다. 후경 남자의 발은 전경 인물에 가려지지만 다리가 바닥 쪽으로 내려오며 부유하는 자세는 아니다. 탁자와 의자에도 명백한 지지 불가능성은 보이지 않는다."
       }
      ]
     },
     "ok": true
    }
   ],
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "slot_winner_match": false,
   "slot_winner": {
    "gemini-pro": "B",
    "gpt-high": "A"
   },
   "route": "cross_slot_combined"
  },
  "dual": {
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "normalized": {
    "A": 1.0,
    "B": 0.667
   },
   "adjusted": {
    "A": 0.75,
    "B": 0.417
   },
   "violations": {
    "A": [
     "[gemini-pro] 프롬프트에 명시되지 않은 추가 인물이 렌더링되어 총 3명이 등장함 (Extra bodies).",
     "[gemini-pro] 지휘봉을 쥔 오른쪽 손 아래에 불이 붙은 담배가 아무런 지지 없이 허공에 떠 있음 (Unsupported object).",
     "[gpt-high] 수감자를 토미이시와 탁자의 같은 쪽에 앉혀, 탁자가 심문자와 수감자 사이에 있어야 한다는 배치를 어겼다.",
     "[gpt-high] 참조의 의자 두 개를 전경 두 개와 맞은편 빈 의자 한 개, 총 세 개로 늘렸다.",
     "[gpt-high] 토미이시 입의 담배 외에 탁자 부근에 담배를 하나 더 추가했다."
    ],
    "B": [
     "[gemini-pro] 프롬프트에 없는 3번째 인물이 추가되었으며, 특히 배경에 서 있는 인물은 군복과 외양이 토미이시와 동일하여 캐릭터가 복제됨 (Duplicated/Extra bodies).",
     "[gpt-high] 배상도로 읽히는 후경 인물을 지정된 오른쪽 위가 아닌 중앙 왼쪽에 배치하여 명시된 인물 위치를 어겼다."
    ]
   },
   "per_model_winner": {
    "gemini-pro": "B",
    "gpt-high": "A"
   },
   "agreed": false
  },
  "totals": {
   "B": 417,
   "A": 750
  },
  "selected": "A",
  "ranking": [
   "A",
   "B"
  ],
  "verdicts": [
   {
    "label": "B",
    "score": 417,
    "verdict_ko": "지휘봉이 테이블 건너편의 대상을 올바르게 가리키고 있으나, 프롬프트에 없는 3번째 인물이 추가되었고 토미이시 캐릭터가 복제되어 심각한 오류가 발생했습니다.  ★위반: [gemini-pro] 프롬프트에 없는 3번째 인물이 추가되었으며, 특히 배경에 서 있는 인물은 군복과 외양이 토미이시와 동일하여 캐릭터가 복제됨 (Duplicated/Extra bodies). / [gpt-high] 배상도로 읽히는 후경 인물을 지정된 오른쪽 위가 아닌 중앙 왼쪽에 배치하여 명시된 인물 위치를 어겼다."
   },
   {
    "label": "A",
    "score": 750,
    "verdict_ko": "지휘봉을 쥔 손 아래에 담배가 허공에 떠 있는 물리적 오류가 있으며, 불필요한 인물이 추가로 배치되어 지침을 위반했습니다.  ★위반: [gemini-pro] 프롬프트에 명시되지 않은 추가 인물이 렌더링되어 총 3명이 등장함 (Extra bodies). / [gemini-pro] 지휘봉을 쥔 오른쪽 손 아래에 불이 붙은 담배가 아무런 지지 없이 허공에 떠 있음 (Unsupported object). / [gpt-high] 수감자를 토미이시와 탁자의 같은 쪽에 앉혀, 탁자가 심문자와 수감자 사이에 있어야 한다는 배치를 어겼다. / [gpt-high] 참조의 의자 두 개를 전경 두 개와 맞은편 빈 의자 한 개, 총 세 개로 늘렸다. / [gpt-high] 토미이시 입의 담배 외에 탁자 부근에 담배를 하나 더 추가했다."
   }
  ],
  "refs": [
   {
    "label": "LOCATION PHOTOGRAPH — the exact place of this shot: its architecture, materials, fixed features and lighting mood are spatial truth; stage the moment inside this place. Never copy its camera framing.",
    "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/db00517b-165b-4005-a287-f28936b59b95/episodes/272bd7af-9294-4490-98c6-88a40e9deacc/images/background_chain/L33B02.png",
    "asset_id": "00197208-b612-467b-a594-99866513b16a",
    "role": "location_plate"
   },
   {
    "label": "CHARACTER REFERENCE — 토미이시: the exact person appearing in this shot; match face, hair and build exactly.",
    "path": "<bytes:797699>",
    "asset_id": "c861ba0f-8750-4262-8324-25cf5c6c9dcd",
    "role": "character_ref"
   }
  ],
  "critique_skipped": true,
  "needs_reshoot": true,
  "shot_run_produced": false,
  "shot_run_uid": "06aa1431-4157-76ff-8606-d5f9aee51fce",
  "bgfirst": {
   "bg_path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/db00517b-165b-4005-a287-f28936b59b95/images/272bd7af-9294-4490-98c6-88a40e9deacc/scene/recipe/S44sh6__bgfirst_bg.png",
   "bg_asset_id": "eff0cff2-1353-4d8e-91b1-3232859ab071",
   "bg_record_key": "S44sh6::bgfirst_bg",
   "chain_winner": true,
   "authority": "plate"
  },
  "ref_mode": "재투영 배경+콘티+엔티티 (2택1: 체인 승)",
  "share_plan": {
   "ref_plan": "background"
  }
 },
 "S44sh6::cine": {
  "applied": true,
  "attempted_at": "2026-09-09T05:57:41.032414+00:00",
  "fingerprint": "762477da6f393dd01c2c94e2c62cc9382f65f15ec3a491cf211da4822cb7415d",
  "fingerprint_version": 2,
  "provider": "mai",
  "endpoint": "openrouter/chat-completions",
  "model": "microsoft/mai-image-2.6",
  "pack": "24.202608252115",
  "source_file": "S44sh6_sel.png",
  "source_sha256": "5cb0588e2ad86be03e27e3e0c6589465431d167e50ab2568af85645c839b8c38",
  "file": "S44sh6_cine.png",
  "staged_sha256": "ec467e0e7826602ea6a08b51530e7686c9299777c8d8d6e6c0badeb1c5dbdeb6",
  "latency_ms": 27855
 },
 "S44sh10::signage": {
  "fp": "1ff938940fa96988",
  "inscriptions": [],
  "cues": [],
  "dropped": []
 },
 "S44sh10": {
  "input_fingerprint": "b00dfe11c3aa21ce",
  "prompt": "Create ONE FINAL photorealistic live-action film still of the moment below — early 20th-century Korea, Jeolla region; all people are Korean unless explicitly stated otherwise. TIME OF DAY (lock): night.\n\nSHOT TEXT (authoritative, Korean): 두 눈을 부릅뜨고 악에 받친 표정으로 토미이시를 매섭게 노려보는 달래의 얼굴.\n\nLOCATION (lock): The detainee's side of the table inside the dark prison interrogation room at night. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: close-up\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Preserve the interrogation room's darkness and muted saturation, using restrained facial separation to reveal the blood, dirt, and defiant eyes without glamorizing the injuries.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The people visible in that photo are NOT in this shot: never carry their faces, bodies or clothing onto anyone here. Each person in THIS shot is defined solely by the PEOPLE list, their own CHARACTER REFERENCE images and their wardrobe notes. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The interrogation table and dim, smoke-filled room remain unchanged. 달래: She remains seated with blood and dirt on her face, bloody lips and disheveled hair; her bound hands clutch the bloodstained skirt, and the half ring remains on a finger.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 달래 (한국인 여성, 젊은 성인, 자연스러운 검은 머리) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.",
  "roll_prompts": {
   "A": "Create ONE FINAL photorealistic live-action film still of the moment below — early 20th-century Korea, Jeolla region; all people are Korean unless explicitly stated otherwise. TIME OF DAY (lock): night.\n\nSHOT TEXT (authoritative, Korean): 두 눈을 부릅뜨고 악에 받친 표정으로 토미이시를 매섭게 노려보는 달래의 얼굴.\n\nLOCATION (lock): The detainee's side of the table inside the dark prison interrogation room at night. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: close-up\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Preserve the interrogation room's darkness and muted saturation, using restrained facial separation to reveal the blood, dirt, and defiant eyes without glamorizing the injuries.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The people visible in that photo are NOT in this shot: never carry their faces, bodies or clothing onto anyone here. Each person in THIS shot is defined solely by the PEOPLE list, their own CHARACTER REFERENCE images and their wardrobe notes. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The interrogation table and dim, smoke-filled room remain unchanged. 달래: She remains seated with blood and dirt on her face, bloody lips and disheveled hair; her bound hands clutch the bloodstained skirt, and the half ring remains on a finger.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 달래 (한국인 여성, 젊은 성인, 자연스러운 검은 머리) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.",
   "B": "Create ONE FINAL photorealistic live-action film still of the moment below — early 20th-century Korea, Jeolla region; all people are Korean unless explicitly stated otherwise. TIME OF DAY (lock): night.\n\nSHOT TEXT (authoritative, Korean): 두 눈을 부릅뜨고 악에 받친 표정으로 토미이시를 매섭게 노려보는 달래의 얼굴.\n\nLOCATION (lock): The detainee's side of the table inside the dark prison interrogation room at night. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: close-up\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Preserve the interrogation room's darkness and muted saturation, using restrained facial separation to reveal the blood, dirt, and defiant eyes without glamorizing the injuries.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The people visible in that photo are NOT in this shot: never carry their faces, bodies or clothing onto anyone here. Each person in THIS shot is defined solely by the PEOPLE list, their own CHARACTER REFERENCE images and their wardrobe notes. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The interrogation table and dim, smoke-filled room remain unchanged. 달래: She remains seated with blood and dirt on her face, bloody lips and disheveled hair; her bound hands clutch the bloodstained skirt, and the half ring remains on a finger.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 달래 (한국인 여성, 젊은 성인, 자연스러운 검은 머리) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.\n\nCOMPOSITION VARIATION (this candidate only): keep the place, the people and the moment exactly as specified, and keep from the CAMERA & FRAME contract its subject scale, screen placement and key background elements exactly. For this candidate only, its camera angle and height are deliberately left open: choose a camera position distinctly different from the obvious one, as a film director picking a second setup on the same blocking. Stay true to any explicit framing the SHOT TEXT itself dictates."
  },
  "shot_run_spend_attempt_count": 1,
  "initial_roll_all_fail": true,
  "readings": [
   {
    "label": "A",
    "direction": "달래는 프레임 왼쪽 전경에 등지고 있는 인물(군복 입은 남자)을 매섭게 노려보고 있음.",
    "built_space": "어두운 취조실 환경. 피사체 앞에 이전 샷과 일치하는 넓은 심문용 테이블 1개가 올바르게 위치함.",
    "entities": "달래는 캐릭터 레퍼런스의 얼굴, 머리 모양, 녹색 한복을 정확히 반영했으며, 피와 흙이 묻은 묘사도 잘 나타남. 그러나 프롬프트에서 등장을 허용하지 않은 추가 인물(군복 입은 남자)의 어깨와 뒷모습이 전경에 등장함.",
    "hard_violations": [
     "[gemini-pro] 허용되지 않은 인물(전경의 군복 입은 남자) 포함",
     "[gpt-high] 달래만 보여야 하는 장면에 군복 남성의 머리와 상체를 추가하여, 명시된 등장인물 제한을 위반했다."
    ],
    "physics": "테이블 앞에 앉아 몸을 약간 기울인 자세가 자연스러우며, 신체를 지지하는 물리적 기반이 안정적임."
   },
   {
    "label": "B",
    "direction": "달래는 정면(카메라 렌즈 방향)을 강하게 노려보고 있음.",
    "built_space": "어두운 취조실. 이전 샷에 설정된 중앙의 커다란 심문용 테이블이 피사체 앞에 있지 않고 왼쪽 배경에 작게 배치되었으며, 피사체 앞에는 좁은 나무 판자만 놓여 있어 스테이징이 완전히 어긋남.",
    "entities": "달래의 얼굴과 상처 묘사는 존재하나, 의상이 캐릭터 레퍼런스의 녹색 한복이 아닌 이전 샷에 앉아 있던 다른 인물의 베이지색 옷으로 잘못 적용됨.",
    "hard_violations": [
     "[gemini-pro] 물리적으로 불가능한 기괴한 해부학 (팔이 없고 옷 뭉치 위에 머리가 얹힌 형태)",
     "[gemini-pro] 스테이징 오류 및 가구 중복 배치 (중앙 테이블 부재, 배경에 엉뚱한 테이블 생성)",
     "[gemini-pro] 이전 샷 인물의 의상을 그대로 차용함 (금지 사항 위반)",
     "[gpt-high] 전경의 심문 탁자 외에 달래 뒤 왼쪽에 별도 탁자를 추가하여, 고정된 장소의 탁자 구조를 중복시켰다."
    ],
    "physics": "앉아 있는 하체나 팔의 형태가 전혀 보이지 않고, 정체불명의 옷 뭉치 덩어리 위에 머리만 놓여 있는 기괴한 형태로 물리적 지지 구조를 상실함."
   }
  ],
  "cross_model_order": {
   "policy": "cross_model_order_v2_slot0_fwd_slot1_rev_parallel2",
   "slots": [
    {
     "model": "gemini-pro",
     "order": "forward",
     "display_to_canonical": {
      "A": "A",
      "B": "B"
     },
     "raw": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 4,
        "verdict_ko": "달래의 표정과 레퍼런스 의상 일치도, 테이블 스테이징은 우수하나, 명시된 허용 인물(달래) 외에는 절대 추가하지 말라는 지시를 어기고 전경에 다른 인물을 배치하여 치명적인 위반을 범했습니다."
       },
       {
        "label": "B",
        "score": 2,
        "verdict_ko": "팔이 없는 듯한 기괴한 신체 구조를 띠고 있으며, 지정된 위치에 있어야 할 중앙 테이블을 무시하고 이전 샷 인물의 의상을 그대로 차용하는 등 다수의 치명적인 위반이 발생했습니다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "달래는 프레임 왼쪽 전경에 등지고 있는 인물(군복 입은 남자)을 매섭게 노려보고 있음.",
        "built_space": "어두운 취조실 환경. 피사체 앞에 이전 샷과 일치하는 넓은 심문용 테이블 1개가 올바르게 위치함.",
        "entities": "달래는 캐릭터 레퍼런스의 얼굴, 머리 모양, 녹색 한복을 정확히 반영했으며, 피와 흙이 묻은 묘사도 잘 나타남. 그러나 프롬프트에서 등장을 허용하지 않은 추가 인물(군복 입은 남자)의 어깨와 뒷모습이 전경에 등장함.",
        "hard_violations": [
         "허용되지 않은 인물(전경의 군복 입은 남자) 포함"
        ],
        "physics": "테이블 앞에 앉아 몸을 약간 기울인 자세가 자연스러우며, 신체를 지지하는 물리적 기반이 안정적임."
       },
       {
        "label": "B",
        "direction": "달래는 정면(카메라 렌즈 방향)을 강하게 노려보고 있음.",
        "built_space": "어두운 취조실. 이전 샷에 설정된 중앙의 커다란 심문용 테이블이 피사체 앞에 있지 않고 왼쪽 배경에 작게 배치되었으며, 피사체 앞에는 좁은 나무 판자만 놓여 있어 스테이징이 완전히 어긋남.",
        "entities": "달래의 얼굴과 상처 묘사는 존재하나, 의상이 캐릭터 레퍼런스의 녹색 한복이 아닌 이전 샷에 앉아 있던 다른 인물의 베이지색 옷으로 잘못 적용됨.",
        "hard_violations": [
         "물리적으로 불가능한 기괴한 해부학 (팔이 없고 옷 뭉치 위에 머리가 얹힌 형태)",
         "스테이징 오류 및 가구 중복 배치 (중앙 테이블 부재, 배경에 엉뚱한 테이블 생성)",
         "이전 샷 인물의 의상을 그대로 차용함 (금지 사항 위반)"
        ],
        "physics": "앉아 있는 하체나 팔의 형태가 전혀 보이지 않고, 정체불명의 옷 뭉치 덩어리 위에 머리만 놓여 있는 기괴한 형태로 물리적 지지 구조를 상실함."
       }
      ],
      "all_candidates_fail": true
     },
     "normalized": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 4,
        "verdict_ko": "달래의 표정과 레퍼런스 의상 일치도, 테이블 스테이징은 우수하나, 명시된 허용 인물(달래) 외에는 절대 추가하지 말라는 지시를 어기고 전경에 다른 인물을 배치하여 치명적인 위반을 범했습니다."
       },
       {
        "label": "B",
        "score": 2,
        "verdict_ko": "팔이 없는 듯한 기괴한 신체 구조를 띠고 있으며, 지정된 위치에 있어야 할 중앙 테이블을 무시하고 이전 샷 인물의 의상을 그대로 차용하는 등 다수의 치명적인 위반이 발생했습니다."
       }
      ],
      "all_candidates_fail": true,
      "readings": [
       {
        "label": "A",
        "direction": "달래는 프레임 왼쪽 전경에 등지고 있는 인물(군복 입은 남자)을 매섭게 노려보고 있음.",
        "built_space": "어두운 취조실 환경. 피사체 앞에 이전 샷과 일치하는 넓은 심문용 테이블 1개가 올바르게 위치함.",
        "entities": "달래는 캐릭터 레퍼런스의 얼굴, 머리 모양, 녹색 한복을 정확히 반영했으며, 피와 흙이 묻은 묘사도 잘 나타남. 그러나 프롬프트에서 등장을 허용하지 않은 추가 인물(군복 입은 남자)의 어깨와 뒷모습이 전경에 등장함.",
        "hard_violations": [
         "허용되지 않은 인물(전경의 군복 입은 남자) 포함"
        ],
        "physics": "테이블 앞에 앉아 몸을 약간 기울인 자세가 자연스러우며, 신체를 지지하는 물리적 기반이 안정적임."
       },
       {
        "label": "B",
        "direction": "달래는 정면(카메라 렌즈 방향)을 강하게 노려보고 있음.",
        "built_space": "어두운 취조실. 이전 샷에 설정된 중앙의 커다란 심문용 테이블이 피사체 앞에 있지 않고 왼쪽 배경에 작게 배치되었으며, 피사체 앞에는 좁은 나무 판자만 놓여 있어 스테이징이 완전히 어긋남.",
        "entities": "달래의 얼굴과 상처 묘사는 존재하나, 의상이 캐릭터 레퍼런스의 녹색 한복이 아닌 이전 샷에 앉아 있던 다른 인물의 베이지색 옷으로 잘못 적용됨.",
        "hard_violations": [
         "물리적으로 불가능한 기괴한 해부학 (팔이 없고 옷 뭉치 위에 머리가 얹힌 형태)",
         "스테이징 오류 및 가구 중복 배치 (중앙 테이블 부재, 배경에 엉뚱한 테이블 생성)",
         "이전 샷 인물의 의상을 그대로 차용함 (금지 사항 위반)"
        ],
        "physics": "앉아 있는 하체나 팔의 형태가 전혀 보이지 않고, 정체불명의 옷 뭉치 덩어리 위에 머리만 놓여 있는 기괴한 형태로 물리적 지지 구조를 상실함."
       }
      ]
     },
     "ok": true
    },
    {
     "model": "gpt-high",
     "order": "reverse",
     "display_to_canonical": {
      "A": "B",
      "B": "A"
     },
     "raw": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 3,
        "verdict_ko": "달래만 등장하고 얼굴과 악에 받친 표정을 더 크게 잡았지만, 뒤쪽에 별도 탁자를 추가했으며 얼굴 클로즈업보다 넓고 의상 색도 참조와 다르다."
       },
       {
        "label": "B",
        "score": 2,
        "verdict_ko": "부릅뜬 눈의 시선은 앞의 남성을 향하지만, 등장하면 안 되는 군복 남성을 추가해 얼굴 클로즈업을 어깨너머 구도로 바꾼 결정적 위반이 있다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "달래는 고개와 상체를 앞으로 숙이고 눈을 치켜떠 카메라 근처, 화면 왼쪽으로 약간 벗어난 지점을 노려본다. 토미이시는 화면 밖이므로 실제 시선 도착점은 확인할 수 없지만 맞은편 상대를 노려보는 연기로 읽힌다. 눈을 크게 벌리기보다는 눈썹을 낮추고 입에 힘을 준 분노가 두드러진다. 무기나 방향성 있는 소품은 없다.",
        "built_space": "하단 전경에 낡은 심문 탁자 상판 하나, 달래 뒤 왼쪽에 별도의 목제 탁자 상판과 다리 구조 하나가 보인다. 오른쪽 어깨 뒤에는 의자 등받이 일부가 보인다. 거친 회색 벽과 어두운 배경은 장소 분위기에 맞지만, 참조의 탁자 하나에 더해 뒤쪽 탁자가 생겼다. 얼굴뿐 아니라 상체와 크게 뭉친 무릎 부근 옷감까지 포함해 지정된 얼굴 클로즈업보다 넓다.",
        "entities": "보이는 사람은 젊은 한국인 여성으로 설정된 달래 한 명뿐이며, 검은 머리와 얼굴 특징은 인물 참조에 대체로 부합한다. 헝클어진 머리, 얼굴의 흙과 피, 피 묻은 입술이 확인된다. 다만 저고리는 참조의 연녹색보다 회갈색이고, 하단 옷감도 참조의 검은 치마와 다르다. 묶인 손과 반지는 보이지 않아 판정하지 않는다. 읽을 수 있는 글자는 없다.",
        "hard_violations": [
         "전경의 심문 탁자 외에 달래 뒤 왼쪽에 별도 탁자를 추가하여, 고정된 장소의 탁자 구조를 중복시켰다."
        ],
        "physics": "머리는 목에, 앞으로 기울인 상체는 화면 아래 몸통에 자연스럽게 이어진다. 어깨 뒤 의자 일부가 보여 앉은 자세와 양립하며, 엉덩이와 좌판의 접촉은 화면 밖이다. 하단의 뭉친 옷감은 무릎 또는 허벅지 위에 놓인 것으로 읽혀 부유한다고 볼 근거가 없다. 손이 가려져 치마를 움켜쥔 접촉은 확인할 수 없다."
       },
       {
        "label": "B",
        "direction": "달래의 두 눈은 크게 열려 있고 화면 왼쪽 전경의 군복 남성 얼굴 쪽을 향한다. 토미이시를 노려보는 관계는 명확하지만, 그 상대를 실제 화면에 등장시킨 것이 인물 제한에 어긋난다. 남성은 달래 쪽으로 머리를 돌리고 있으나 눈은 보이지 않는다. 무기나 움직이는 소품은 없다.",
        "built_space": "하단에 낡은 심문 탁자 하나, 달래의 오른쪽 어깨 뒤에 의자 등받이 일부가 보인다. 어두운 벽과 옅은 연무는 참조의 실내 분위기와 부합하며, 중복된 고정 설비는 보이지 않는다. 달래는 탁자 건너편에 앉아 있지만 왼쪽 전경 남성의 머리와 어깨가 크게 들어와, 요청한 얼굴 클로즈업 대신 상체를 포함한 어깨너머 구도가 되었다.",
        "entities": "달래는 젊은 한국인 여성의 외형, 자연스러운 검은 머리, 오른쪽으로 내려온 땋은 머리, 연녹색 저고리 등 참조의 특징을 대체로 유지한다. 얼굴의 흙과 상처, 피 묻은 입술도 보인다. 묶인 손, 치마, 반지는 화면 밖이므로 평가하지 않는다. 그러나 전경에 짧은 검은 머리와 군복을 입은 성인 남성의 머리·목·어깨가 추가되었으며, 이는 허용된 달래의 신체 일부가 아니다. 읽을 수 있는 글자는 없다.",
        "hard_violations": [
         "달래만 보여야 하는 장면에 군복 남성의 머리와 상체를 추가하여, 명시된 등장인물 제한을 위반했다."
        ],
        "physics": "달래의 머리와 상체는 자연스럽게 연결되고 뒤의 의자와 앉은 자세가 양립한다. 좌판 접촉과 발은 화면 밖이지만 공중에 떠 있는 징후는 없다. 전경 남성도 목과 어깨가 화면 아래 몸통으로 이어져 물리적으로 불가능한 자세는 아니다. 머리카락과 옷감은 중력에 따라 내려오며 떠 있는 물체는 없다."
       }
      ],
      "all_candidates_fail": true
     },
     "normalized": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 3,
        "verdict_ko": "달래만 등장하고 얼굴과 악에 받친 표정을 더 크게 잡았지만, 뒤쪽에 별도 탁자를 추가했으며 얼굴 클로즈업보다 넓고 의상 색도 참조와 다르다."
       },
       {
        "label": "A",
        "score": 2,
        "verdict_ko": "부릅뜬 눈의 시선은 앞의 남성을 향하지만, 등장하면 안 되는 군복 남성을 추가해 얼굴 클로즈업을 어깨너머 구도로 바꾼 결정적 위반이 있다."
       }
      ],
      "all_candidates_fail": true,
      "readings": [
       {
        "label": "B",
        "direction": "달래는 고개와 상체를 앞으로 숙이고 눈을 치켜떠 카메라 근처, 화면 왼쪽으로 약간 벗어난 지점을 노려본다. 토미이시는 화면 밖이므로 실제 시선 도착점은 확인할 수 없지만 맞은편 상대를 노려보는 연기로 읽힌다. 눈을 크게 벌리기보다는 눈썹을 낮추고 입에 힘을 준 분노가 두드러진다. 무기나 방향성 있는 소품은 없다.",
        "built_space": "하단 전경에 낡은 심문 탁자 상판 하나, 달래 뒤 왼쪽에 별도의 목제 탁자 상판과 다리 구조 하나가 보인다. 오른쪽 어깨 뒤에는 의자 등받이 일부가 보인다. 거친 회색 벽과 어두운 배경은 장소 분위기에 맞지만, 참조의 탁자 하나에 더해 뒤쪽 탁자가 생겼다. 얼굴뿐 아니라 상체와 크게 뭉친 무릎 부근 옷감까지 포함해 지정된 얼굴 클로즈업보다 넓다.",
        "entities": "보이는 사람은 젊은 한국인 여성으로 설정된 달래 한 명뿐이며, 검은 머리와 얼굴 특징은 인물 참조에 대체로 부합한다. 헝클어진 머리, 얼굴의 흙과 피, 피 묻은 입술이 확인된다. 다만 저고리는 참조의 연녹색보다 회갈색이고, 하단 옷감도 참조의 검은 치마와 다르다. 묶인 손과 반지는 보이지 않아 판정하지 않는다. 읽을 수 있는 글자는 없다.",
        "hard_violations": [
         "전경의 심문 탁자 외에 달래 뒤 왼쪽에 별도 탁자를 추가하여, 고정된 장소의 탁자 구조를 중복시켰다."
        ],
        "physics": "머리는 목에, 앞으로 기울인 상체는 화면 아래 몸통에 자연스럽게 이어진다. 어깨 뒤 의자 일부가 보여 앉은 자세와 양립하며, 엉덩이와 좌판의 접촉은 화면 밖이다. 하단의 뭉친 옷감은 무릎 또는 허벅지 위에 놓인 것으로 읽혀 부유한다고 볼 근거가 없다. 손이 가려져 치마를 움켜쥔 접촉은 확인할 수 없다."
       },
       {
        "label": "A",
        "direction": "달래의 두 눈은 크게 열려 있고 화면 왼쪽 전경의 군복 남성 얼굴 쪽을 향한다. 토미이시를 노려보는 관계는 명확하지만, 그 상대를 실제 화면에 등장시킨 것이 인물 제한에 어긋난다. 남성은 달래 쪽으로 머리를 돌리고 있으나 눈은 보이지 않는다. 무기나 움직이는 소품은 없다.",
        "built_space": "하단에 낡은 심문 탁자 하나, 달래의 오른쪽 어깨 뒤에 의자 등받이 일부가 보인다. 어두운 벽과 옅은 연무는 참조의 실내 분위기와 부합하며, 중복된 고정 설비는 보이지 않는다. 달래는 탁자 건너편에 앉아 있지만 왼쪽 전경 남성의 머리와 어깨가 크게 들어와, 요청한 얼굴 클로즈업 대신 상체를 포함한 어깨너머 구도가 되었다.",
        "entities": "달래는 젊은 한국인 여성의 외형, 자연스러운 검은 머리, 오른쪽으로 내려온 땋은 머리, 연녹색 저고리 등 참조의 특징을 대체로 유지한다. 얼굴의 흙과 상처, 피 묻은 입술도 보인다. 묶인 손, 치마, 반지는 화면 밖이므로 평가하지 않는다. 그러나 전경에 짧은 검은 머리와 군복을 입은 성인 남성의 머리·목·어깨가 추가되었으며, 이는 허용된 달래의 신체 일부가 아니다. 읽을 수 있는 글자는 없다.",
        "hard_violations": [
         "달래만 보여야 하는 장면에 군복 남성의 머리와 상체를 추가하여, 명시된 등장인물 제한을 위반했다."
        ],
        "physics": "달래의 머리와 상체는 자연스럽게 연결되고 뒤의 의자와 앉은 자세가 양립한다. 좌판 접촉과 발은 화면 밖이지만 공중에 떠 있는 징후는 없다. 전경 남성도 목과 어깨가 화면 아래 몸통으로 이어져 물리적으로 불가능한 자세는 아니다. 머리카락과 옷감은 중력에 따라 내려오며 떠 있는 물체는 없다."
       }
      ]
     },
     "ok": true
    }
   ],
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "slot_winner_match": false,
   "slot_winner": {
    "gemini-pro": "A",
    "gpt-high": "B"
   },
   "route": "cross_slot_combined"
  },
  "dual": {
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "normalized": {
    "A": 1.667,
    "B": 1.5
   },
   "adjusted": {
    "A": 1.417,
    "B": 1.25
   },
   "violations": {
    "A": [
     "[gemini-pro] 허용되지 않은 인물(전경의 군복 입은 남자) 포함",
     "[gpt-high] 달래만 보여야 하는 장면에 군복 남성의 머리와 상체를 추가하여, 명시된 등장인물 제한을 위반했다."
    ],
    "B": [
     "[gemini-pro] 물리적으로 불가능한 기괴한 해부학 (팔이 없고 옷 뭉치 위에 머리가 얹힌 형태)",
     "[gemini-pro] 스테이징 오류 및 가구 중복 배치 (중앙 테이블 부재, 배경에 엉뚱한 테이블 생성)",
     "[gemini-pro] 이전 샷 인물의 의상을 그대로 차용함 (금지 사항 위반)",
     "[gpt-high] 전경의 심문 탁자 외에 달래 뒤 왼쪽에 별도 탁자를 추가하여, 고정된 장소의 탁자 구조를 중복시켰다."
    ]
   },
   "per_model_winner": {
    "gemini-pro": "A",
    "gpt-high": "B"
   },
   "agreed": false
  },
  "totals": {
   "A": 1417,
   "B": 1250
  },
  "selected": "A",
  "ranking": [
   "A",
   "B"
  ],
  "verdicts": [
   {
    "label": "A",
    "score": 1417,
    "verdict_ko": "달래의 표정과 레퍼런스 의상 일치도, 테이블 스테이징은 우수하나, 명시된 허용 인물(달래) 외에는 절대 추가하지 말라는 지시를 어기고 전경에 다른 인물을 배치하여 치명적인 위반을 범했습니다.  ★위반: [gemini-pro] 허용되지 않은 인물(전경의 군복 입은 남자) 포함 / [gpt-high] 달래만 보여야 하는 장면에 군복 남성의 머리와 상체를 추가하여, 명시된 등장인물 제한을 위반했다."
   },
   {
    "label": "B",
    "score": 1250,
    "verdict_ko": "팔이 없는 듯한 기괴한 신체 구조를 띠고 있으며, 지정된 위치에 있어야 할 중앙 테이블을 무시하고 이전 샷 인물의 의상을 그대로 차용하는 등 다수의 치명적인 위반이 발생했습니다.  ★위반: [gemini-pro] 물리적으로 불가능한 기괴한 해부학 (팔이 없고 옷 뭉치 위에 머리가 얹힌 형태) / [gemini-pro] 스테이징 오류 및 가구 중복 배치 (중앙 테이블 부재, 배경에 엉뚱한 테이블 생성) / [gemini-pro] 이전 샷 인물의 의상을 그대로 차용함 (금지 사항 위반) / [gpt-high] 전경의 심문 탁자 외에 달래 뒤 왼쪽에 별도 탁자를 추가하여, 고정된 장소의 탁자 구조를 중복시켰다."
   }
  ],
  "refs": [
   {
    "label": "PREVIOUS SHOT STILL — a visually related earlier shot of this same place: the location's look, materials, fixed features and lighting mood are LOCKED to this photo; never copy its camera framing. The people visible in that photo are NOT in this shot: never carry their faces, bodies or clothing onto anyone here. Each person in THIS shot is defined solely by the PEOPLE list, their own CHARACTER REFERENCE images and their wardrobe notes. If this photo shows a character who cannot move (dead or unconscious), that character's exact pose, position and orientation are ALSO LOCKED — treat their body as a fixed prop of the set that only the camera moves around.",
    "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/db00517b-165b-4005-a287-f28936b59b95/images/272bd7af-9294-4490-98c6-88a40e9deacc/scene/recipe/S44sh6_sel.png",
    "asset_id": "fe5b378b-db22-4707-b5df-97ee3efdfa6b",
    "role": "prev_still"
   },
   {
    "label": "CHARACTER REFERENCE — 달래: the exact person appearing in this shot; match face, hair and build exactly.",
    "path": "<bytes:879791>",
    "asset_id": "83024e69-b375-4f3e-98ec-820d0e2ff1f0",
    "role": "character_ref"
   }
  ],
  "critique_skipped": true,
  "needs_reshoot": true,
  "shot_run_produced": false,
  "shot_run_uid": "06aa1433-1f7f-78fc-b84e-58df1e0fe7f6",
  "ref_mode": "prev+엔티티",
  "share_plan": {
   "ref_plan": "prev",
   "prev_anchor_tag": "S44sh6"
  }
 },
 "S44sh10::cine": {
  "applied": false,
  "attempted_at": "2026-09-09T09:11:54.946803+00:00",
  "fingerprint": "40a1d335f62fc94b4f12193dc7c2bab6bd521ab5dbb697fab289d319b9727891",
  "fingerprint_version": 2,
  "provider": "mai",
  "endpoint": "openrouter/chat-completions",
  "model": "microsoft/mai-image-2.6",
  "pack": "24.202608252115",
  "source_file": "S44sh10_sel.png",
  "source_sha256": "b07313fbf4cb4ee61da376bd6b6eeb1722197a05708258eaadd338cdb0db44f8",
  "error": "RuntimeError: moderation blocked: {\"message\": \"Response content blocked by label 'MultiSeverity_ViolenceScore'.\", \"code\": 400}",
  "moderation_refusals": 2,
  "declined": true,
  "declined_reason": "moderation"
 },
 "S44sh17::signage": {
  "fp": "7a338158eef8e13c",
  "inscriptions": [],
  "cues": [],
  "dropped": []
 },
 "S44sh17::bgfirst_bg": {
  "input_fingerprint": "ca136fba377217a9",
  "prompt": "Create the EMPTY BACKGROUND PLATE for one film shot — NO PEOPLE, no figures, no body parts, no sketch lines, no arrows anywhere.\n\"Empty\" means no people only: KEEP the location's inherent occupants and stock that define the place — animals in an animal shelter, pen or farm, goods in a market, moored boats in a harbour — unless the shot text explicitly removes them.\nThe FIRST attached image is a thin-line storyboard sketch: use ONLY its camera angle, horizon, perspective and the placement/size of buildings and set masses — ignore the sketched people and arrows entirely. The SECOND attached image (LOCATION PHOTOGRAPH) is the real place: take its architecture, materials, signage and fixed features, and RE-PROJECT them into the sketch's camera. If the photograph's camera differs from the sketch's, the sketch's camera wins.\nHUMAN-SCALE CALIBRATION: derive every structure's true size from human-scale elements — a door ≈ 2m, a window ≈ 1–1.5m wide, one storey ≈ 2.5–3m; never inflate a small structure or shrink a large one.\n\nSHOT TEXT this background must serve (Korean): 거대한 바위 위에 우뚝 서서 산길 아래쪽을 뚫어져라 내려다보는 안규홍의 굳은 전신.\n\nLOCATION (lock): The top of a large rock beside the mountain camp's forest approach, overlooking a trail covered in dense dawn mist.\n\nTIME OF DAY (lock): night.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: wide shot\n- FRAME LAYOUT: Lower mountain path continuing beyond the frame in the lower-left of the frame, foreground; Rock supporting the waiting figure in the lower-right of the frame, midground.\n- KEY BACKGROUND ELEMENTS: Large rock beneath 안규홍 (Supports his full figure) — Its upper surface and path-facing side are visible from below and to one side; used as An elevated support confined to the lower-right third, establishing his lookout position; Lower mountain path (Extends through the forest beneath the lookout) — Recedes from the lower-left toward the terrain below the rock; used as Provides look room and a plausible downward viewing direction; Forest bordering the path (Present along the mountain route); used as Surrounding terrain gives scale and depth to the isolated waiting figure.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Soft dawn illumination filtered through the scripted dense mountain fog lowers distant contrast while preserving sober separation around 안규홍's waiting figure.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nRender ONE photorealistic empty location photograph, 16:9, that this shot can be staged inside later. No readable writing anywhere: surfaces that would carry writing may be present, but stage any wording out of legibility — an oblique angle, distance, shallow focus. No captions, watermarks or overlay text.",
  "effective_prompt": "Create the EMPTY BACKGROUND PLATE for one film shot — NO PEOPLE, no figures, no body parts, no sketch lines, no arrows anywhere.\n\"Empty\" means no people only: KEEP the location's inherent occupants and stock that define the place — animals in an animal shelter, pen or farm, goods in a market, moored boats in a harbour — unless the shot text explicitly removes them.\nThe FIRST attached image is a thin-line storyboard sketch: use ONLY its camera angle, horizon, perspective and the placement/size of buildings and set masses — ignore the sketched people and arrows entirely. The SECOND attached image (LOCATION PHOTOGRAPH) is the real place: take its architecture, materials, signage and fixed features, and RE-PROJECT them into the sketch's camera. If the photograph's camera differs from the sketch's, the sketch's camera wins.\nHUMAN-SCALE CALIBRATION: derive every structure's true size from human-scale elements — a door ≈ 2m, a window ≈ 1–1.5m wide, one storey ≈ 2.5–3m; never inflate a small structure or shrink a large one.\n\nSHOT TEXT this background must serve (Korean): 거대한 바위 위에 우뚝 서서 산길 아래쪽을 뚫어져라 내려다보는 안규홍의 굳은 전신.\n\nLOCATION (lock): The top of a large rock beside the mountain camp's forest approach, overlooking a trail covered in dense dawn mist.\n\nTIME OF DAY (lock): night.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: wide shot\n- FRAME LAYOUT: Lower mountain path continuing beyond the frame in the lower-left of the frame, foreground; Rock supporting the waiting figure in the lower-right of the frame, midground.\n- KEY BACKGROUND ELEMENTS: Large rock beneath 안규홍 (Supports his full figure) — Its upper surface and path-facing side are visible from below and to one side; used as An elevated support confined to the lower-right third, establishing his lookout position; Lower mountain path (Extends through the forest beneath the lookout) — Recedes from the lower-left toward the terrain below the rock; used as Provides look room and a plausible downward viewing direction; Forest bordering the path (Present along the mountain route); used as Surrounding terrain gives scale and depth to the isolated waiting figure.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Soft dawn illumination filtered through the scripted dense mountain fog lowers distant contrast while preserving sober separation around 안규홍's waiting figure.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nRender ONE photorealistic empty location photograph, 16:9, that this shot can be staged inside later. No readable writing anywhere: surfaces that would carry writing may be present, but stage any wording out of legibility — an oblique angle, distance, shallow focus. No captions, watermarks or overlay text.",
  "bg_path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/db00517b-165b-4005-a287-f28936b59b95/images/272bd7af-9294-4490-98c6-88a40e9deacc/scene/recipe/S44sh17__bgfirst_bg.png",
  "asset_id": "2999c7ec-7f92-4b23-a5eb-929f4127c479",
  "input_asset_ids": [
   "e1be15bd-1561-470f-8e86-af3cd0283875",
   "e675b1e4-43c3-411f-ba1f-3b6957aa6dab"
  ]
 },
 "S44sh17": {
  "input_fingerprint": "ef68d305f0e68d03",
  "prompt": "Create ONE FINAL photorealistic live-action film still of the moment below — early 20th-century Korea, Jeolla region; all people are Korean unless explicitly stated otherwise. TIME OF DAY (lock): night.\n\nSHOT TEXT (authoritative, Korean): 거대한 바위 위에 우뚝 서서 산길 아래쪽을 뚫어져라 내려다보는 안규홍의 굳은 전신.\n\nLOCATION (lock): The top of a large rock beside the mountain camp's forest approach, overlooking a trail covered in dense dawn mist. The shot takes place here — the attached LOCATION PHOTOGRAPH shows the exact spot.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: wide shot\n- FRAME LAYOUT: Lower mountain path continuing beyond the frame in the lower-left of the frame, foreground; Rock supporting the waiting figure in the lower-right of the frame, midground.\n- KEY BACKGROUND ELEMENTS: Large rock beneath 안규홍 (Supports his full figure) — Its upper surface and path-facing side are visible from below and to one side; used as An elevated support confined to the lower-right third, establishing his lookout position; Lower mountain path (Extends through the forest beneath the lookout) — Recedes from the lower-left toward the terrain below the rock; used as Provides look room and a plausible downward viewing direction; Forest bordering the path (Present along the mountain route); used as Surrounding terrain gives scale and depth to the isolated waiting figure.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Soft dawn illumination filtered through the scripted dense mountain fog lowers distant contrast while preserving sober separation around 안규홍's waiting figure.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): Dense dawn fog covers the mountain path; a cart approaches through the forest. 안규홍: He stands on a rock waiting, with his half jade ring hanging at his neck.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 안규홍 (한국인 남성, 청년 나이대, 검은 머리, 단단히 묶은 머리) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.",
  "roll_prompts": {
   "A": "Stage the shot. The FIRST attached image (SHOT BACKGROUND) is the finished empty background of this shot — keep it EXACTLY: its camera, perspective, architecture, lighting and every fixed feature stay untouched. The SECOND attached image (LAYOUT SKETCH) tells you ONLY where the people go: each sketched person's position, screen size, pose and the gaze/motion arrows. Ignore the sketch's background lines. The CHARACTER REFERENCE photographs show the real people.\nPlace the real people into the background at exactly the sketched positions, sizes and poses, following the arrow directions. No sketch lines or arrows may remain.\n\nCreate ONE FINAL photorealistic live-action film still of the moment below — early 20th-century Korea, Jeolla region; all people are Korean unless explicitly stated otherwise. TIME OF DAY (lock): night.\n\nSHOT TEXT (authoritative, Korean): 거대한 바위 위에 우뚝 서서 산길 아래쪽을 뚫어져라 내려다보는 안규홍의 굳은 전신.\n\nLOCATION (lock): The top of a large rock beside the mountain camp's forest approach, overlooking a trail covered in dense dawn mist. The shot takes place here — the FIRST attached image (SHOT BACKGROUND) is this exact place, already built: its ground, structures, horizon, materials and lighting are the finished truth of this location and must not be redesigned or replaced. No location photograph is attached — read the place from that image alone, and add no scenery, structure, vehicle or fixture that it does not already show. This lock governs the place only; the figures in the shot follow the staging and pose instructions.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: wide shot\n- FRAME LAYOUT: Lower mountain path continuing beyond the frame in the lower-left of the frame, foreground; Rock supporting the waiting figure in the lower-right of the frame, midground.\n- KEY BACKGROUND ELEMENTS: Large rock beneath 안규홍 (Supports his full figure) — Its upper surface and path-facing side are visible from below and to one side; used as An elevated support confined to the lower-right third, establishing his lookout position; Lower mountain path (Extends through the forest beneath the lookout) — Recedes from the lower-left toward the terrain below the rock; used as Provides look room and a plausible downward viewing direction; Forest bordering the path (Present along the mountain route); used as Surrounding terrain gives scale and depth to the isolated waiting figure.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Soft dawn illumination filtered through the scripted dense mountain fog lowers distant contrast while preserving sober separation around 안규홍's waiting figure.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): Dense dawn fog covers the mountain path; a cart approaches through the forest. 안규홍: He stands on a rock waiting, with his half jade ring hanging at his neck.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 안규홍 (한국인 남성, 청년 나이대, 검은 머리, 단단히 묶은 머리) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.",
   "B": "Create ONE FINAL photorealistic live-action film still of the moment below — early 20th-century Korea, Jeolla region; all people are Korean unless explicitly stated otherwise. TIME OF DAY (lock): night.\n\nSHOT TEXT (authoritative, Korean): 거대한 바위 위에 우뚝 서서 산길 아래쪽을 뚫어져라 내려다보는 안규홍의 굳은 전신.\n\nLOCATION (lock): The top of a large rock beside the mountain camp's forest approach, overlooking a trail covered in dense dawn mist. The shot takes place here — the attached LOCATION PHOTOGRAPH shows the exact spot.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: wide shot\n- FRAME LAYOUT: Lower mountain path continuing beyond the frame in the lower-left of the frame, foreground; Rock supporting the waiting figure in the lower-right of the frame, midground.\n- KEY BACKGROUND ELEMENTS: Large rock beneath 안규홍 (Supports his full figure) — Its upper surface and path-facing side are visible from below and to one side; used as An elevated support confined to the lower-right third, establishing his lookout position; Lower mountain path (Extends through the forest beneath the lookout) — Recedes from the lower-left toward the terrain below the rock; used as Provides look room and a plausible downward viewing direction; Forest bordering the path (Present along the mountain route); used as Surrounding terrain gives scale and depth to the isolated waiting figure.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Soft dawn illumination filtered through the scripted dense mountain fog lowers distant contrast while preserving sober separation around 안규홍's waiting figure.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): Dense dawn fog covers the mountain path; a cart approaches through the forest. 안규홍: He stands on a rock waiting, with his half jade ring hanging at his neck.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 안규홍 (한국인 남성, 청년 나이대, 검은 머리, 단단히 묶은 머리) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.\n\nCOMPOSITION VARIATION (this candidate only): keep the place, the people and the moment exactly as specified, and keep from the CAMERA & FRAME contract its subject scale, screen placement and key background elements exactly. For this candidate only, its camera angle and height are deliberately left open: choose a camera position distinctly different from the obvious one, as a film director picking a second setup on the same blocking. Stay true to any explicit framing the SHOT TEXT itself dictates."
  },
  "roll_refs": {
   "A": [
    {
     "label": "SHOT BACKGROUND",
     "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/db00517b-165b-4005-a287-f28936b59b95/images/272bd7af-9294-4490-98c6-88a40e9deacc/scene/recipe/S44sh17__bgfirst_bg.png",
     "asset_id": "2999c7ec-7f92-4b23-a5eb-929f4127c479",
     "role": "bgfirst_bg"
    },
    {
     "label": "LAYOUT SKETCH (people placement only)",
     "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/db00517b-165b-4005-a287-f28936b59b95/images/272bd7af-9294-4490-98c6-88a40e9deacc/conti/conti_S44sh17.png",
     "asset_id": "e1be15bd-1561-470f-8e86-af3cd0283875",
     "role": "conti_light"
    },
    {
     "label": "CHARACTER REFERENCE — 안규홍: the exact person appearing in this shot; match face, hair and build exactly.",
     "path": "<bytes:1028384>",
     "asset_id": "3b5894db-ce55-4f34-b966-5e725c3aa952",
     "role": "character_ref"
    }
   ],
   "B": [
    {
     "label": "LOCATION PHOTOGRAPH — the exact place of this shot: its architecture, materials, fixed features and lighting mood are spatial truth; stage the moment inside this place. Never copy its camera framing.",
     "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/db00517b-165b-4005-a287-f28936b59b95/episodes/272bd7af-9294-4490-98c6-88a40e9deacc/images/background_chain/L33B01.png",
     "asset_id": "e675b1e4-43c3-411f-ba1f-3b6957aa6dab",
     "role": "location_plate"
    },
    {
     "label": "CHARACTER REFERENCE — 안규홍: the exact person appearing in this shot; match face, hair and build exactly.",
     "path": "<bytes:1028384>",
     "asset_id": "3b5894db-ce55-4f34-b966-5e725c3aa952",
     "role": "character_ref"
    }
   ]
  },
  "shot_run_spend_attempt_count": 1,
  "judge_flip_skipped": "judge_fn_owns_order",
  "readings": [
   {
    "label": "A",
    "direction": "안규홍이 우측 하단의 바위 위에서 좌측 하단의 산길과 접근하는 수레를 향해 시선을 아래로 향하고 있습니다.",
    "built_space": "우측 하단에 안규홍이 서 있는 거대한 바위가 배치되고, 좌측 하단으로 안개 낀 산길이 이어져 프레임 밖으로 나가는 구도가 레퍼런스의 지형과 잘 일치합니다.",
    "entities": "안규홍의 외모와 의상은 레퍼런스와 일치하지만 요구된 반옥 가락지가 목에 없습니다. 산길에 수레가 나타나 있으나, 샷 텍스트에 전혀 언급되지 않은 수레를 끄는 인물(마부)이 추가되어 나타납니다.",
    "hard_violations": [
     "[gemini-pro] 샷 텍스트에 명시되지 않은 인물(수레를 모는 사람)이 프레임에 추가됨",
     "[gpt-high] 안규홍만 보여야 하는 장면에 수레 뒤의 별도 인물 한 명을 추가했다.",
     "[gpt-high] 브리프에 없는 소 한 마리를 수레의 견인 동물로 추가했다."
    ],
    "physics": "안규홍의 두 발이 바위 표면에 안정적으로 지지되어 있으며, 소와 수레 역시 길 위에 물리적으로 정상적으로 위치해 있습니다."
   },
   {
    "label": "B",
    "direction": "안규홍이 우측 바위 위에서 좌측 산길 쪽을 뚫어져라 내려다보는 시선과 자세가 명확하게 연출되었습니다.",
    "built_space": "지정된 와이드 샷 프레이밍에 맞게 우측 하단에 바위, 좌측 하단에 안개 낀 산길이 정확히 배치되었으며 레퍼런스 사진의 지형을 충실히 따르고 있습니다.",
    "entities": "안규홍의 인상착의가 캐릭터 레퍼런스와 일치하며, 샷 텍스트 외의 추가 인물이 등장하지 않았습니다. 안개 속에 수레가 접근하는 모습이 묘사되었으나 목에 걸려 있어야 할 반옥 가락지는 보이지 않습니다.",
    "hard_violations": [],
    "physics": "안규홍은 거대한 바위 위에 굳건히 서 있으며 체중이 두 발에 올바르게 지지되어 불안정하거나 떠 있는 요소가 없습니다."
   }
  ],
  "cross_model_order": {
   "policy": "cross_model_order_v2_slot0_fwd_slot1_rev_parallel2",
   "slots": [
    {
     "model": "gemini-pro",
     "order": "forward",
     "display_to_canonical": {
      "A": "A",
      "B": "B"
     },
     "raw": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 2,
        "verdict_ko": "지정된 안규홍 외에 산길의 수레에 명시되지 않은 인물을 임의로 추가하여 인물 제한 규칙을 심각하게 위반했습니다."
       },
       {
        "label": "B",
        "score": 9,
        "verdict_ko": "명시되지 않은 인물을 추가하지 않고 요구된 카메라 구도와 산길을 내려다보는 시선, 안개 낀 분위기를 훌륭하게 구현했으나 반옥 가락지 소품이 누락되었습니다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "안규홍이 우측 하단의 바위 위에서 좌측 하단의 산길과 접근하는 수레를 향해 시선을 아래로 향하고 있습니다.",
        "built_space": "우측 하단에 안규홍이 서 있는 거대한 바위가 배치되고, 좌측 하단으로 안개 낀 산길이 이어져 프레임 밖으로 나가는 구도가 레퍼런스의 지형과 잘 일치합니다.",
        "entities": "안규홍의 외모와 의상은 레퍼런스와 일치하지만 요구된 반옥 가락지가 목에 없습니다. 산길에 수레가 나타나 있으나, 샷 텍스트에 전혀 언급되지 않은 수레를 끄는 인물(마부)이 추가되어 나타납니다.",
        "hard_violations": [
         "샷 텍스트에 명시되지 않은 인물(수레를 모는 사람)이 프레임에 추가됨"
        ],
        "physics": "안규홍의 두 발이 바위 표면에 안정적으로 지지되어 있으며, 소와 수레 역시 길 위에 물리적으로 정상적으로 위치해 있습니다."
       },
       {
        "label": "B",
        "direction": "안규홍이 우측 바위 위에서 좌측 산길 쪽을 뚫어져라 내려다보는 시선과 자세가 명확하게 연출되었습니다.",
        "built_space": "지정된 와이드 샷 프레이밍에 맞게 우측 하단에 바위, 좌측 하단에 안개 낀 산길이 정확히 배치되었으며 레퍼런스 사진의 지형을 충실히 따르고 있습니다.",
        "entities": "안규홍의 인상착의가 캐릭터 레퍼런스와 일치하며, 샷 텍스트 외의 추가 인물이 등장하지 않았습니다. 안개 속에 수레가 접근하는 모습이 묘사되었으나 목에 걸려 있어야 할 반옥 가락지는 보이지 않습니다.",
        "hard_violations": [],
        "physics": "안규홍은 거대한 바위 위에 굳건히 서 있으며 체중이 두 발에 올바르게 지지되어 불안정하거나 떠 있는 요소가 없습니다."
       }
      ],
      "all_candidates_fail": false
     },
     "normalized": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 2,
        "verdict_ko": "지정된 안규홍 외에 산길의 수레에 명시되지 않은 인물을 임의로 추가하여 인물 제한 규칙을 심각하게 위반했습니다."
       },
       {
        "label": "B",
        "score": 9,
        "verdict_ko": "명시되지 않은 인물을 추가하지 않고 요구된 카메라 구도와 산길을 내려다보는 시선, 안개 낀 분위기를 훌륭하게 구현했으나 반옥 가락지 소품이 누락되었습니다."
       }
      ],
      "all_candidates_fail": false,
      "readings": [
       {
        "label": "A",
        "direction": "안규홍이 우측 하단의 바위 위에서 좌측 하단의 산길과 접근하는 수레를 향해 시선을 아래로 향하고 있습니다.",
        "built_space": "우측 하단에 안규홍이 서 있는 거대한 바위가 배치되고, 좌측 하단으로 안개 낀 산길이 이어져 프레임 밖으로 나가는 구도가 레퍼런스의 지형과 잘 일치합니다.",
        "entities": "안규홍의 외모와 의상은 레퍼런스와 일치하지만 요구된 반옥 가락지가 목에 없습니다. 산길에 수레가 나타나 있으나, 샷 텍스트에 전혀 언급되지 않은 수레를 끄는 인물(마부)이 추가되어 나타납니다.",
        "hard_violations": [
         "샷 텍스트에 명시되지 않은 인물(수레를 모는 사람)이 프레임에 추가됨"
        ],
        "physics": "안규홍의 두 발이 바위 표면에 안정적으로 지지되어 있으며, 소와 수레 역시 길 위에 물리적으로 정상적으로 위치해 있습니다."
       },
       {
        "label": "B",
        "direction": "안규홍이 우측 바위 위에서 좌측 산길 쪽을 뚫어져라 내려다보는 시선과 자세가 명확하게 연출되었습니다.",
        "built_space": "지정된 와이드 샷 프레이밍에 맞게 우측 하단에 바위, 좌측 하단에 안개 낀 산길이 정확히 배치되었으며 레퍼런스 사진의 지형을 충실히 따르고 있습니다.",
        "entities": "안규홍의 인상착의가 캐릭터 레퍼런스와 일치하며, 샷 텍스트 외의 추가 인물이 등장하지 않았습니다. 안개 속에 수레가 접근하는 모습이 묘사되었으나 목에 걸려 있어야 할 반옥 가락지는 보이지 않습니다.",
        "hard_violations": [],
        "physics": "안규홍은 거대한 바위 위에 굳건히 서 있으며 체중이 두 발에 올바르게 지지되어 불안정하거나 떠 있는 요소가 없습니다."
       }
      ]
     },
     "ok": true
    },
    {
     "model": "gpt-high",
     "order": "reverse",
     "display_to_canonical": {
      "A": "B",
      "B": "A"
     },
     "raw": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 7,
        "verdict_ko": "바위 위 전신과 왼쪽 아래 산길을 향한 응시를 구현하고 추가 인물이 없지만, 바위의 화면 점유가 크며 밤보다는 새벽으로 보이고 반쪽 옥가락지는 확인되지 않는다."
       },
       {
        "label": "B",
        "score": 2,
        "verdict_ko": "전신과 산길의 배치는 대체로 맞지만, 허용되지 않은 수레꾼과 소를 추가한 것이 결정적 위반이며 밤 시간대와 반쪽 옥가락지도 구현되지 않았다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "안규홍은 몸을 거의 정면으로 두고 얼굴과 눈을 왼쪽 아래로 향한다. 응시 방향은 바위 아래 안개 낀 산길 쪽에 닿는다. 산길 중앙의 수레는 전경 쪽을 향한 것으로 보이나 안개 때문에 진행 여부는 확정하기 어렵다. 무기나 손에 든 지시물은 없다.",
        "built_space": "인공 구조물이나 고정 설비는 없다. 주된 바위 하나가 오른쪽에 있고 그 윗면과 길을 향한 절벽면이 함께 보인다. 안규홍은 그 윗면 가장자리에 서 있으며, 산길 하나가 왼쪽 아래 전경에서 바위 아래 숲으로 이어진다. 왼쪽의 굵은 나무와 낙엽 덮인 암반은 장소 참조와 잘 대응하지만 바위는 지정된 오른쪽 아래 3분의 1보다 크게, 화면 오른쪽 하반부에 걸쳐 있다.",
        "entities": "식별 가능한 사람은 안규홍 한 명이다. 청년 동아시아계 남성의 외양, 뒤로 단단히 정리한 검은 머리, 마른 체격, 때 묻은 밝은 저고리와 바지, 천 허리띠와 짚신이 참조에 대체로 부합한다. 얼굴의 세부 동일성은 이 거리에서 제한적으로만 확인된다. 목에 매단 반쪽 옥가락지는 보이지 않는다. 안개 속에 수레 한 대가 있지만 별도의 사람이나 동물은 명확히 식별되지 않는다. 큰 바위와 숲, 짙은 산길 안개는 있으며, 밝은 하늘과 확산광은 잠금 조건인 밤보다 새벽에 가깝다. 읽을 수 있는 글자는 없다.",
        "hard_violations": [],
        "physics": "두 발의 짚신이 바위 윗면에 닿아 몸을 지지하며, 약간 벌린 발과 내려놓은 팔로 굳게 기다리는 자세가 물리적으로 가능하다. 옷과 허리띠는 몸에 걸려 자연스럽게 처진다. 수레 하부는 안개에 가려지지만 바퀴 윤곽이 길 높이에 있고 공중에 떠 있다는 증거는 없다."
       },
       {
        "label": "B",
        "direction": "안규홍의 얼굴은 왼쪽으로 돌아가 있고 눈은 약간 아래를 향해 바위 아래 산길 방향을 본다. 다만 고개 숙임은 크지 않다. 길 위의 소와 수레는 왼쪽 아래 전경 방향으로 향한다. 수레 뒤 인물의 정확한 시선은 크기가 작아 확인되지 않는다.",
        "built_space": "인공 건물이나 고정 설비는 없다. 오른쪽의 큰 바위 하나 위에 안규홍이 서 있고, 바위 윗면과 산길 쪽 측면이 보인다. 길 하나가 왼쪽 아래에서 숲 안쪽으로 이어져 기본 공간 관계는 맞는다. 장소 참조의 굵은 왼쪽 나무와 낙엽 덮인 바위 지형도 대응한다. 다만 바위가 오른쪽 아래 3분의 1을 넘어 화면 오른쪽 하반부를 크게 차지한다.",
        "entities": "바위 위에는 청년 동아시아계 남성 안규홍이 있으며, 검은 묶음머리와 밝은 낡은 한복, 천 허리띠, 짚신은 참조와 대체로 맞는다. 목의 반쪽 옥가락지는 확인되지 않는다. 산길에는 수레 한 대뿐 아니라 이를 끄는 소 한 마리와 수레 뒤의 별도 사람 한 명이 뚜렷이 보인다. 이 추가 인물의 나이와 세부 외양은 판별하기 어렵지만 안규홍 외에는 아무도 보여서는 안 된다는 조건을 어긴다. 숲과 짙은 안개는 맞지만 조명은 밤보다 푸른 새벽으로 읽힌다. 읽을 수 있는 글자는 없다.",
        "hard_violations": [
         "안규홍만 보여야 하는 장면에 수레 뒤의 별도 인물 한 명을 추가했다.",
         "브리프에 없는 소 한 마리를 수레의 견인 동물로 추가했다."
        ],
        "physics": "안규홍의 두 발은 바위에 접지되어 있고 체중을 지지할 수 있는 자세다. 소의 다리는 길에 닿으며 수레는 바퀴로 지지되고 견인 연결도 보인다. 추가 인물의 하체는 수레에 가려 정확한 접촉점은 확인하기 어렵지만, 명백히 공중에 떠 있는 모습은 아니다. 물리적 불가능성보다 허용되지 않은 존재의 추가가 문제다."
       }
      ],
      "all_candidates_fail": true
     },
     "normalized": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 7,
        "verdict_ko": "바위 위 전신과 왼쪽 아래 산길을 향한 응시를 구현하고 추가 인물이 없지만, 바위의 화면 점유가 크며 밤보다는 새벽으로 보이고 반쪽 옥가락지는 확인되지 않는다."
       },
       {
        "label": "A",
        "score": 2,
        "verdict_ko": "전신과 산길의 배치는 대체로 맞지만, 허용되지 않은 수레꾼과 소를 추가한 것이 결정적 위반이며 밤 시간대와 반쪽 옥가락지도 구현되지 않았다."
       }
      ],
      "all_candidates_fail": true,
      "readings": [
       {
        "label": "B",
        "direction": "안규홍은 몸을 거의 정면으로 두고 얼굴과 눈을 왼쪽 아래로 향한다. 응시 방향은 바위 아래 안개 낀 산길 쪽에 닿는다. 산길 중앙의 수레는 전경 쪽을 향한 것으로 보이나 안개 때문에 진행 여부는 확정하기 어렵다. 무기나 손에 든 지시물은 없다.",
        "built_space": "인공 구조물이나 고정 설비는 없다. 주된 바위 하나가 오른쪽에 있고 그 윗면과 길을 향한 절벽면이 함께 보인다. 안규홍은 그 윗면 가장자리에 서 있으며, 산길 하나가 왼쪽 아래 전경에서 바위 아래 숲으로 이어진다. 왼쪽의 굵은 나무와 낙엽 덮인 암반은 장소 참조와 잘 대응하지만 바위는 지정된 오른쪽 아래 3분의 1보다 크게, 화면 오른쪽 하반부에 걸쳐 있다.",
        "entities": "식별 가능한 사람은 안규홍 한 명이다. 청년 동아시아계 남성의 외양, 뒤로 단단히 정리한 검은 머리, 마른 체격, 때 묻은 밝은 저고리와 바지, 천 허리띠와 짚신이 참조에 대체로 부합한다. 얼굴의 세부 동일성은 이 거리에서 제한적으로만 확인된다. 목에 매단 반쪽 옥가락지는 보이지 않는다. 안개 속에 수레 한 대가 있지만 별도의 사람이나 동물은 명확히 식별되지 않는다. 큰 바위와 숲, 짙은 산길 안개는 있으며, 밝은 하늘과 확산광은 잠금 조건인 밤보다 새벽에 가깝다. 읽을 수 있는 글자는 없다.",
        "hard_violations": [],
        "physics": "두 발의 짚신이 바위 윗면에 닿아 몸을 지지하며, 약간 벌린 발과 내려놓은 팔로 굳게 기다리는 자세가 물리적으로 가능하다. 옷과 허리띠는 몸에 걸려 자연스럽게 처진다. 수레 하부는 안개에 가려지지만 바퀴 윤곽이 길 높이에 있고 공중에 떠 있다는 증거는 없다."
       },
       {
        "label": "A",
        "direction": "안규홍의 얼굴은 왼쪽으로 돌아가 있고 눈은 약간 아래를 향해 바위 아래 산길 방향을 본다. 다만 고개 숙임은 크지 않다. 길 위의 소와 수레는 왼쪽 아래 전경 방향으로 향한다. 수레 뒤 인물의 정확한 시선은 크기가 작아 확인되지 않는다.",
        "built_space": "인공 건물이나 고정 설비는 없다. 오른쪽의 큰 바위 하나 위에 안규홍이 서 있고, 바위 윗면과 산길 쪽 측면이 보인다. 길 하나가 왼쪽 아래에서 숲 안쪽으로 이어져 기본 공간 관계는 맞는다. 장소 참조의 굵은 왼쪽 나무와 낙엽 덮인 바위 지형도 대응한다. 다만 바위가 오른쪽 아래 3분의 1을 넘어 화면 오른쪽 하반부를 크게 차지한다.",
        "entities": "바위 위에는 청년 동아시아계 남성 안규홍이 있으며, 검은 묶음머리와 밝은 낡은 한복, 천 허리띠, 짚신은 참조와 대체로 맞는다. 목의 반쪽 옥가락지는 확인되지 않는다. 산길에는 수레 한 대뿐 아니라 이를 끄는 소 한 마리와 수레 뒤의 별도 사람 한 명이 뚜렷이 보인다. 이 추가 인물의 나이와 세부 외양은 판별하기 어렵지만 안규홍 외에는 아무도 보여서는 안 된다는 조건을 어긴다. 숲과 짙은 안개는 맞지만 조명은 밤보다 푸른 새벽으로 읽힌다. 읽을 수 있는 글자는 없다.",
        "hard_violations": [
         "안규홍만 보여야 하는 장면에 수레 뒤의 별도 인물 한 명을 추가했다.",
         "브리프에 없는 소 한 마리를 수레의 견인 동물로 추가했다."
        ],
        "physics": "안규홍의 두 발은 바위에 접지되어 있고 체중을 지지할 수 있는 자세다. 소의 다리는 길에 닿으며 수레는 바퀴로 지지되고 견인 연결도 보인다. 추가 인물의 하체는 수레에 가려 정확한 접촉점은 확인하기 어렵지만, 명백히 공중에 떠 있는 모습은 아니다. 물리적 불가능성보다 허용되지 않은 존재의 추가가 문제다."
       }
      ]
     },
     "ok": true
    }
   ],
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "slot_winner_match": true,
   "slot_winner": {
    "gemini-pro": "B",
    "gpt-high": "B"
   },
   "route": "cross_slot_agree"
  },
  "dual": {
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "normalized": {
    "A": 0.508,
    "B": 2.0
   },
   "adjusted": {
    "A": 0.258,
    "B": 2.0
   },
   "violations": {
    "A": [
     "[gemini-pro] 샷 텍스트에 명시되지 않은 인물(수레를 모는 사람)이 프레임에 추가됨",
     "[gpt-high] 안규홍만 보여야 하는 장면에 수레 뒤의 별도 인물 한 명을 추가했다.",
     "[gpt-high] 브리프에 없는 소 한 마리를 수레의 견인 동물로 추가했다."
    ]
   },
   "per_model_winner": {
    "gemini-pro": "B",
    "gpt-high": "B"
   },
   "agreed": true
  },
  "totals": {
   "A": 258,
   "B": 2000
  },
  "selected": "B",
  "ranking": [
   "B",
   "A"
  ],
  "verdicts": [
   {
    "label": "A",
    "score": 258,
    "verdict_ko": "지정된 안규홍 외에 산길의 수레에 명시되지 않은 인물을 임의로 추가하여 인물 제한 규칙을 심각하게 위반했습니다.  ★위반: [gemini-pro] 샷 텍스트에 명시되지 않은 인물(수레를 모는 사람)이 프레임에 추가됨 / [gpt-high] 안규홍만 보여야 하는 장면에 수레 뒤의 별도 인물 한 명을 추가했다. / [gpt-high] 브리프에 없는 소 한 마리를 수레의 견인 동물로 추가했다."
   },
   {
    "label": "B",
    "score": 2000,
    "verdict_ko": "명시되지 않은 인물을 추가하지 않고 요구된 카메라 구도와 산길을 내려다보는 시선, 안개 낀 분위기를 훌륭하게 구현했으나 반옥 가락지 소품이 누락되었습니다."
   }
  ],
  "refs": [
   {
    "label": "LOCATION PHOTOGRAPH — the exact place of this shot: its architecture, materials, fixed features and lighting mood are spatial truth; stage the moment inside this place. Never copy its camera framing.",
    "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/db00517b-165b-4005-a287-f28936b59b95/episodes/272bd7af-9294-4490-98c6-88a40e9deacc/images/background_chain/L33B01.png",
    "asset_id": "e675b1e4-43c3-411f-ba1f-3b6957aa6dab",
    "role": "location_plate"
   },
   {
    "label": "CHARACTER REFERENCE — 안규홍: the exact person appearing in this shot; match face, hair and build exactly.",
    "path": "<bytes:1028384>",
    "asset_id": "3b5894db-ce55-4f34-b966-5e725c3aa952",
    "role": "character_ref"
   }
  ],
  "critique_skipped": true,
  "shot_run_produced": false,
  "shot_run_uid": "06aa1433-f5c0-74d9-a03e-6129431a1ff7",
  "bgfirst": {
   "bg_path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/db00517b-165b-4005-a287-f28936b59b95/images/272bd7af-9294-4490-98c6-88a40e9deacc/scene/recipe/S44sh17__bgfirst_bg.png",
   "bg_asset_id": "2999c7ec-7f92-4b23-a5eb-929f4127c479",
   "bg_record_key": "S44sh17::bgfirst_bg",
   "chain_winner": false,
   "authority": "plate"
  },
  "ref_mode": "플레이트+엔티티 (2택1: 무콘티 승)",
  "share_plan": {
   "ref_plan": "background"
  }
 },
 "S44sh17::cine": {
  "applied": true,
  "attempted_at": "2026-09-09T06:02:23.898043+00:00",
  "fingerprint": "61c3c325b78c448afb0c7f9b3fe5ccd8acbbf947c24a8743ceacf44766b8382b",
  "fingerprint_version": 2,
  "provider": "mai",
  "endpoint": "openrouter/chat-completions",
  "model": "microsoft/mai-image-2.6",
  "pack": "24.202608252115",
  "source_file": "S44sh17_sel.png",
  "source_sha256": "18d5f6135c8f0e944a9c7ef9055868e8e7f8ef513c9f6041188154c01c19fe01",
  "file": "S44sh17_cine.png",
  "staged_sha256": "89bdc04ef6d7dce787105926c636904350964e0479bb2d8571925febc291ed6e",
  "latency_ms": 26809
 },
 "S45sh5::signage": {
  "fp": "1b9aa9fbcc1b79e4",
  "inscriptions": [],
  "cues": [],
  "dropped": []
 },
 "S45sh5::bgfirst_bg": {
  "input_fingerprint": "35316c16368d52c6",
  "prompt": "Create the EMPTY BACKGROUND PLATE for one film shot — NO PEOPLE, no figures, no body parts, no sketch lines, no arrows anywhere.\n\"Empty\" means no people only: KEEP the location's inherent occupants and stock that define the place — animals in an animal shelter, pen or farm, goods in a market, moored boats in a harbour — unless the shot text explicitly removes them.\nThe FIRST attached image is a thin-line storyboard sketch: use ONLY its camera angle, horizon, perspective and the placement/size of buildings and set masses — ignore the sketched people and arrows entirely. The SECOND attached image (LOCATION PHOTOGRAPH) is the real place: take its architecture, materials, signage and fixed features, and RE-PROJECT them into the sketch's camera. If the photograph's camera differs from the sketch's, the sketch's camera wins.\nHUMAN-SCALE CALIBRATION: derive every structure's true size from human-scale elements — a door ≈ 2m, a window ≈ 1–1.5m wide, one storey ≈ 2.5–3m; never inflate a small structure or shrink a large one.\n\nSHOT TEXT this background must serve (Korean): 달래의 시신 옆에 무릎을 꿇고 앉아 피 묻은 달래의 손을 들어 올린 안규홍의 상체.\n\nLOCATION (lock): The outdoor mourning place around a straw mat laid in the mountain camp's clearing at night.\n\nTIME OF DAY (lock): night.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: medium shot\n- KEY BACKGROUND ELEMENTS: 짚멍석 (달래의 시신이 위에 뉘어져 있다) — Only a narrow portion of its upper surface appears beneath her body; used as Grounds the lower edge without competing with the lifted hand; 달래의 반쪽 옥가락지 (달래의 손가락에 걸려 있다) — Seen obliquely on her supported finger; used as A small narrative detail within the joined-hand area; 안규홍의 반쪽 가락지 (안규홍의 목에 걸려 있다) — Visible against his upper body from the established side angle; used as Quietly establishes the second half without isolating it.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Subdued nighttime illumination preserves readable hands and facial contours with muted saturation and sober, naturalistic contrast.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nRender ONE photorealistic empty location photograph, 16:9, that this shot can be staged inside later. No readable writing anywhere: surfaces that would carry writing may be present, but stage any wording out of legibility — an oblique angle, distance, shallow focus. No captions, watermarks or overlay text.",
  "effective_prompt": "Create the EMPTY BACKGROUND PLATE for one film shot — NO PEOPLE, no figures, no body parts, no sketch lines, no arrows anywhere.\n\"Empty\" means no people only: KEEP the location's inherent occupants and stock that define the place — animals in an animal shelter, pen or farm, goods in a market, moored boats in a harbour — unless the shot text explicitly removes them.\nThe FIRST attached image is a thin-line storyboard sketch: use ONLY its camera angle, horizon, perspective and the placement/size of buildings and set masses — ignore the sketched people and arrows entirely. The SECOND attached image (LOCATION PHOTOGRAPH) is the real place: take its architecture, materials, signage and fixed features, and RE-PROJECT them into the sketch's camera. If the photograph's camera differs from the sketch's, the sketch's camera wins.\nHUMAN-SCALE CALIBRATION: derive every structure's true size from human-scale elements — a door ≈ 2m, a window ≈ 1–1.5m wide, one storey ≈ 2.5–3m; never inflate a small structure or shrink a large one.\n\nSHOT TEXT this background must serve (Korean): 달래의 시신 옆에 무릎을 꿇고 앉아 피 묻은 달래의 손을 들어 올린 안규홍의 상체.\n\nLOCATION (lock): The outdoor mourning place around a straw mat laid in the mountain camp's clearing at night.\n\nTIME OF DAY (lock): night.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: medium shot\n- KEY BACKGROUND ELEMENTS: 짚멍석 (달래의 시신이 위에 뉘어져 있다) — Only a narrow portion of its upper surface appears beneath her body; used as Grounds the lower edge without competing with the lifted hand; 달래의 반쪽 옥가락지 (달래의 손가락에 걸려 있다) — Seen obliquely on her supported finger; used as A small narrative detail within the joined-hand area; 안규홍의 반쪽 가락지 (안규홍의 목에 걸려 있다) — Visible against his upper body from the established side angle; used as Quietly establishes the second half without isolating it.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Subdued nighttime illumination preserves readable hands and facial contours with muted saturation and sober, naturalistic contrast.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nRender ONE photorealistic empty location photograph, 16:9, that this shot can be staged inside later. No readable writing anywhere: surfaces that would carry writing may be present, but stage any wording out of legibility — an oblique angle, distance, shallow focus. No captions, watermarks or overlay text.",
  "bg_path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/db00517b-165b-4005-a287-f28936b59b95/images/272bd7af-9294-4490-98c6-88a40e9deacc/scene/recipe/S45sh5__bgfirst_bg.png",
  "asset_id": "92eb3e0b-4760-4f89-86cb-984991ffab7c",
  "input_asset_ids": [
   "f9e33123-9dd2-43a4-8218-3df03202a68f",
   "c5df4442-a576-4bdd-9e42-441cb4544a5e"
  ]
 },
 "S45sh5": {
  "input_fingerprint": "2d1dd1e991301a6d",
  "prompt": "Create ONE FINAL photorealistic live-action film still of the moment below — early 20th-century Korea, Jeolla region; all people are Korean unless explicitly stated otherwise. TIME OF DAY (lock): night.\n\nSHOT TEXT (authoritative, Korean): 달래의 시신 옆에 무릎을 꿇고 앉아 피 묻은 달래의 손을 들어 올린 안규홍의 상체.\n\nLOCATION (lock): The outdoor mourning place around a straw mat laid in the mountain camp's clearing at night. The shot takes place here — the attached LOCATION PHOTOGRAPH shows the exact spot.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: medium shot\n- KEY BACKGROUND ELEMENTS: 짚멍석 (달래의 시신이 위에 뉘어져 있다) — Only a narrow portion of its upper surface appears beneath her body; used as Grounds the lower edge without competing with the lifted hand; 달래의 반쪽 옥가락지 (달래의 손가락에 걸려 있다) — Seen obliquely on her supported finger; used as A small narrative detail within the joined-hand area; 안규홍의 반쪽 가락지 (안규홍의 목에 걸려 있다) — Visible against his upper body from the established side angle; used as Quietly establishes the second half without isolating it.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Subdued nighttime illumination preserves readable hands and facial contours with muted saturation and sober, naturalistic contrast.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nIMMOBILE CHARACTER POSE — CANONICAL (identical wherever this character appears in ANY panel; on any conflict THIS POSE WINS): Her dead body is laid out on a straw mat while An Gyuhong lifts and supports one of her blood- and dirt-stained hands, raising that forearm from the mat. The screenplay does not specify which hand he holds, her face direction, or the arrangement of her other arm and legs.\n\nIMMOBILE BODIES OBEY GRAVITY: a person who is dead or unconscious\nexerts NO muscular effort. Every part of their body — head, torso,\narms, hands, fingers, legs — rests fully on whatever supports it\n(floor, wall, furniture, their own lap) and hangs or slumps with\ngravity. NEVER show any part of an immobile person's body lifted,\nraised, held up in the air, or posed as if presenting something:\nan object in their grip stays clenched in a hand that itself lies\nfallen on a support — the hand does not hold the object up. If the\ncanonical pose leaves a body part unspecified, resolve it as the\nmost gravity-compliant, fully-supported position.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): A straw mat lies laid out in the camp at night. 달래: She lies dead on the mat with blood and dirt on her face, skirt and hands, disheveled hair, and a half jade ring on a finger. 안규홍: He kneels with one hand carefully raised; the other half jade ring hangs at his neck.\n\nTHE HAND THAT IS DOING THIS: the object at the centre of this shot is being held, operated, read, aimed or handed over by 안규홍 right now, so 안규홍's hand — and as much of the wrist and forearm as the framing reaches — is in the frame, gripping or resting on that object exactly the way its real use requires. Match that hand to 안규홍: its size, build, skin, age, grooming, sleeve and anything worn on it belong to that person and to no one else.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 안규홍 (한국인 남성, 청년 나이대, 검은 머리, 단단히 묶은 머리); 달래 (한국인 여성, 젊은 성인, 자연스러운 검은 머리) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.",
  "roll_prompts": {
   "A": "Stage the shot. The FIRST attached image (SHOT BACKGROUND) is the finished empty background of this shot — keep it EXACTLY: its camera, perspective, architecture, lighting and every fixed feature stay untouched. The SECOND attached image (LAYOUT SKETCH) tells you ONLY where the people go: each sketched person's position, screen size, pose and the gaze/motion arrows. Ignore the sketch's background lines. The CHARACTER REFERENCE photographs show the real people.\nPlace the real people into the background at exactly the sketched positions, sizes and poses, following the arrow directions. No sketch lines or arrows may remain.\n\nCreate ONE FINAL photorealistic live-action film still of the moment below — early 20th-century Korea, Jeolla region; all people are Korean unless explicitly stated otherwise. TIME OF DAY (lock): night.\n\nSHOT TEXT (authoritative, Korean): 달래의 시신 옆에 무릎을 꿇고 앉아 피 묻은 달래의 손을 들어 올린 안규홍의 상체.\n\nLOCATION (lock): The outdoor mourning place around a straw mat laid in the mountain camp's clearing at night. The shot takes place here — the FIRST attached image (SHOT BACKGROUND) is this exact place, already built: its ground, structures, horizon, materials and lighting are the finished truth of this location and must not be redesigned or replaced. No location photograph is attached — read the place from that image alone, and add no scenery, structure, vehicle or fixture that it does not already show. This lock governs the place only; the figures in the shot follow the staging and pose instructions.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: medium shot\n- KEY BACKGROUND ELEMENTS: 짚멍석 (달래의 시신이 위에 뉘어져 있다) — Only a narrow portion of its upper surface appears beneath her body; used as Grounds the lower edge without competing with the lifted hand; 달래의 반쪽 옥가락지 (달래의 손가락에 걸려 있다) — Seen obliquely on her supported finger; used as A small narrative detail within the joined-hand area; 안규홍의 반쪽 가락지 (안규홍의 목에 걸려 있다) — Visible against his upper body from the established side angle; used as Quietly establishes the second half without isolating it.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Subdued nighttime illumination preserves readable hands and facial contours with muted saturation and sober, naturalistic contrast.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nIMMOBILE CHARACTER POSE — CANONICAL (identical wherever this character appears in ANY panel; on any conflict THIS POSE WINS): Her dead body is laid out on a straw mat while An Gyuhong lifts and supports one of her blood- and dirt-stained hands, raising that forearm from the mat. The screenplay does not specify which hand he holds, her face direction, or the arrangement of her other arm and legs.\n\nIMMOBILE BODIES OBEY GRAVITY: a person who is dead or unconscious\nexerts NO muscular effort. Every part of their body — head, torso,\narms, hands, fingers, legs — rests fully on whatever supports it\n(floor, wall, furniture, their own lap) and hangs or slumps with\ngravity. NEVER show any part of an immobile person's body lifted,\nraised, held up in the air, or posed as if presenting something:\nan object in their grip stays clenched in a hand that itself lies\nfallen on a support — the hand does not hold the object up. If the\ncanonical pose leaves a body part unspecified, resolve it as the\nmost gravity-compliant, fully-supported position.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): A straw mat lies laid out in the camp at night. 달래: She lies dead on the mat with blood and dirt on her face, skirt and hands, disheveled hair, and a half jade ring on a finger. 안규홍: He kneels with one hand carefully raised; the other half jade ring hangs at his neck.\n\nTHE HAND THAT IS DOING THIS: the object at the centre of this shot is being held, operated, read, aimed or handed over by 안규홍 right now, so 안규홍's hand — and as much of the wrist and forearm as the framing reaches — is in the frame, gripping or resting on that object exactly the way its real use requires. Match that hand to 안규홍: its size, build, skin, age, grooming, sleeve and anything worn on it belong to that person and to no one else.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 안규홍 (한국인 남성, 청년 나이대, 검은 머리, 단단히 묶은 머리); 달래 (한국인 여성, 젊은 성인, 자연스러운 검은 머리) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.",
   "B": "Create ONE FINAL photorealistic live-action film still of the moment below — early 20th-century Korea, Jeolla region; all people are Korean unless explicitly stated otherwise. TIME OF DAY (lock): night.\n\nSHOT TEXT (authoritative, Korean): 달래의 시신 옆에 무릎을 꿇고 앉아 피 묻은 달래의 손을 들어 올린 안규홍의 상체.\n\nLOCATION (lock): The outdoor mourning place around a straw mat laid in the mountain camp's clearing at night. The shot takes place here — the attached LOCATION PHOTOGRAPH shows the exact spot.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: medium shot\n- KEY BACKGROUND ELEMENTS: 짚멍석 (달래의 시신이 위에 뉘어져 있다) — Only a narrow portion of its upper surface appears beneath her body; used as Grounds the lower edge without competing with the lifted hand; 달래의 반쪽 옥가락지 (달래의 손가락에 걸려 있다) — Seen obliquely on her supported finger; used as A small narrative detail within the joined-hand area; 안규홍의 반쪽 가락지 (안규홍의 목에 걸려 있다) — Visible against his upper body from the established side angle; used as Quietly establishes the second half without isolating it.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Subdued nighttime illumination preserves readable hands and facial contours with muted saturation and sober, naturalistic contrast.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nIMMOBILE CHARACTER POSE — CANONICAL (identical wherever this character appears in ANY panel; on any conflict THIS POSE WINS): Her dead body is laid out on a straw mat while An Gyuhong lifts and supports one of her blood- and dirt-stained hands, raising that forearm from the mat. The screenplay does not specify which hand he holds, her face direction, or the arrangement of her other arm and legs.\n\nIMMOBILE BODIES OBEY GRAVITY: a person who is dead or unconscious\nexerts NO muscular effort. Every part of their body — head, torso,\narms, hands, fingers, legs — rests fully on whatever supports it\n(floor, wall, furniture, their own lap) and hangs or slumps with\ngravity. NEVER show any part of an immobile person's body lifted,\nraised, held up in the air, or posed as if presenting something:\nan object in their grip stays clenched in a hand that itself lies\nfallen on a support — the hand does not hold the object up. If the\ncanonical pose leaves a body part unspecified, resolve it as the\nmost gravity-compliant, fully-supported position.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): A straw mat lies laid out in the camp at night. 달래: She lies dead on the mat with blood and dirt on her face, skirt and hands, disheveled hair, and a half jade ring on a finger. 안규홍: He kneels with one hand carefully raised; the other half jade ring hangs at his neck.\n\nTHE HAND THAT IS DOING THIS: the object at the centre of this shot is being held, operated, read, aimed or handed over by 안규홍 right now, so 안규홍's hand — and as much of the wrist and forearm as the framing reaches — is in the frame, gripping or resting on that object exactly the way its real use requires. Match that hand to 안규홍: its size, build, skin, age, grooming, sleeve and anything worn on it belong to that person and to no one else.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 안규홍 (한국인 남성, 청년 나이대, 검은 머리, 단단히 묶은 머리); 달래 (한국인 여성, 젊은 성인, 자연스러운 검은 머리) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.\n\nCOMPOSITION VARIATION (this candidate only): keep the place, the people and the moment exactly as specified, and keep from the CAMERA & FRAME contract its subject scale, screen placement and key background elements exactly. For this candidate only, its camera angle and height are deliberately left open: choose a camera position distinctly different from the obvious one, as a film director picking a second setup on the same blocking. Stay true to any explicit framing the SHOT TEXT itself dictates."
  },
  "roll_refs": {
   "A": [
    {
     "label": "SHOT BACKGROUND",
     "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/db00517b-165b-4005-a287-f28936b59b95/images/272bd7af-9294-4490-98c6-88a40e9deacc/scene/recipe/S45sh5__bgfirst_bg.png",
     "asset_id": "92eb3e0b-4760-4f89-86cb-984991ffab7c",
     "role": "bgfirst_bg"
    },
    {
     "label": "LAYOUT SKETCH (people placement only)",
     "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/db00517b-165b-4005-a287-f28936b59b95/images/272bd7af-9294-4490-98c6-88a40e9deacc/conti/conti_S45sh5.png",
     "asset_id": "f9e33123-9dd2-43a4-8218-3df03202a68f",
     "role": "conti_light"
    },
    {
     "label": "CHARACTER REFERENCE — 안규홍: the exact person appearing in this shot; match face, hair and build exactly.",
     "path": "<bytes:1028384>",
     "asset_id": "3b5894db-ce55-4f34-b966-5e725c3aa952",
     "role": "character_ref"
    },
    {
     "label": "CHARACTER REFERENCE — 달래: the exact person appearing in this shot; match face, hair and build exactly.",
     "path": "<bytes:959810>",
     "asset_id": "a51dee5b-8b43-4345-b827-da2ff15b545d",
     "role": "character_ref"
    }
   ],
   "B": [
    {
     "label": "LOCATION PHOTOGRAPH — the exact place of this shot: its architecture, materials, fixed features and lighting mood are spatial truth; stage the moment inside this place. Never copy its camera framing.",
     "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/db00517b-165b-4005-a287-f28936b59b95/images/272bd7af-9294-4490-98c6-88a40e9deacc/scene/recipe/groupbg_mountain_camp_clearing_cb9567.png",
     "asset_id": "c5df4442-a576-4bdd-9e42-441cb4544a5e",
     "role": "bgfirst_group_bg"
    },
    {
     "label": "CHARACTER REFERENCE — 안규홍: the exact person appearing in this shot; match face, hair and build exactly.",
     "path": "<bytes:1028384>",
     "asset_id": "3b5894db-ce55-4f34-b966-5e725c3aa952",
     "role": "character_ref"
    },
    {
     "label": "CHARACTER REFERENCE — 달래: the exact person appearing in this shot; match face, hair and build exactly.",
     "path": "<bytes:959810>",
     "asset_id": "a51dee5b-8b43-4345-b827-da2ff15b545d",
     "role": "character_ref"
    }
   ]
  },
  "shot_run_spend_attempt_count": 1,
  "judge_flip_skipped": "judge_fn_owns_order",
  "initial_roll_all_fail": true,
  "readings": [
   {
    "label": "A",
    "direction": "안규홍의 시선은 자신이 잡고 들어 올린 피 묻은 팔을 향하고 있다.",
    "built_space": "넓은 야영지 배경에 목책과 초가 지붕, 가마솥 등이 풀 샷으로 널찍하게 배치되어 있다.",
    "entities": "안규홍(참조 이미지 복장, 목에 가락지)과 달래(녹색 저고리, 검은 치마). 그러나 달래의 원래 두 팔이 몸 위에 놓여 있음에도, 안규홍이 잡고 있는 피 묻은 세 번째 맨팔이 존재한다.",
    "hard_violations": [
     "[gemini-pro] 달래에게 세 번째 팔이 존재하는 해부학적 불가능성 (Extra limbs)",
     "[gemini-pro] 미디엄 샷 및 상체 위주의 프레이밍 지시를 위반하고 배경을 지나치게 넓게 잡은 프레이밍"
    ],
    "physics": "달래의 두 팔은 몸 위에 지탱되어 늘어져 있으나, 안규홍이 잡고 있는 세 번째 팔은 어깨와의 연결 없이 허공에서 솟아나 들려 있다."
   },
   {
    "label": "B",
    "direction": "안규홍의 시선은 들어 올린 달래의 손과 얼굴 쪽을 향하고 있다.",
    "built_space": "어두운 밤의 야영지로, 뒤편에 목책과 불이 켜진 랜턴이 있어 지시된 공간의 분위기와 구조에 부합한다.",
    "entities": "안규홍(상체 위주, 목에 가락지)과 달래(피 묻은 얼굴, 녹색 저고리). 달래의 손목을 쥐고 있는 손(녹색 소매)에 옥가락지가 보인다.",
    "hard_violations": [
     "[gemini-pro] 시신인 달래의 오른손(녹색 소매)이 허공으로 들려 스스로 자신의 왼팔 손목을 꽉 쥐고 있는 물리적/부동성 위반 (Dead body exerting muscular effort)",
     "[gpt-high] 장소 참조와 장면 지시에 없는 등불 세 개 및 타오르는 불을 추가했다."
    ],
    "physics": "안규홍의 오른손이 달래의 피 묻은 왼손을 정상적으로 잡고 있으나, 그 아래에서 죽은 달래의 오른손이 중력을 거스르고 자신의 왼팔을 쥐어 버티는 불가능한 근육 작용이 나타났다."
   }
  ],
  "cross_model_order": {
   "policy": "cross_model_order_v2_slot0_fwd_slot1_rev_parallel2",
   "slots": [
    {
     "model": "gemini-pro",
     "order": "forward",
     "display_to_canonical": {
      "A": "A",
      "B": "B"
     },
     "raw": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 3,
        "verdict_ko": "미디엄 샷 프레이밍과 조명은 훌륭하게 구현되었으나, 시신인 달래의 다른 손이 스스로 자신의 팔을 꽉 쥐고 있는 치명적인 물리적 오류가 발생했습니다."
       },
       {
        "label": "A",
        "score": 1,
        "verdict_ko": "프레이밍 지시(미디엄 샷)를 완전히 무시하고 풀 샷으로 렌더링했으며, 달래에게 세 번째 팔이 돋아난 심각한 해부학적 오류가 있습니다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "안규홍의 시선은 자신이 잡고 들어 올린 피 묻은 팔을 향하고 있다.",
        "built_space": "넓은 야영지 배경에 목책과 초가 지붕, 가마솥 등이 풀 샷으로 널찍하게 배치되어 있다.",
        "entities": "안규홍(참조 이미지 복장, 목에 가락지)과 달래(녹색 저고리, 검은 치마). 그러나 달래의 원래 두 팔이 몸 위에 놓여 있음에도, 안규홍이 잡고 있는 피 묻은 세 번째 맨팔이 존재한다.",
        "hard_violations": [
         "달래에게 세 번째 팔이 존재하는 해부학적 불가능성 (Extra limbs)",
         "미디엄 샷 및 상체 위주의 프레이밍 지시를 위반하고 배경을 지나치게 넓게 잡은 프레이밍"
        ],
        "physics": "달래의 두 팔은 몸 위에 지탱되어 늘어져 있으나, 안규홍이 잡고 있는 세 번째 팔은 어깨와의 연결 없이 허공에서 솟아나 들려 있다."
       },
       {
        "label": "B",
        "direction": "안규홍의 시선은 들어 올린 달래의 손과 얼굴 쪽을 향하고 있다.",
        "built_space": "어두운 밤의 야영지로, 뒤편에 목책과 불이 켜진 랜턴이 있어 지시된 공간의 분위기와 구조에 부합한다.",
        "entities": "안규홍(상체 위주, 목에 가락지)과 달래(피 묻은 얼굴, 녹색 저고리). 달래의 손목을 쥐고 있는 손(녹색 소매)에 옥가락지가 보인다.",
        "hard_violations": [
         "시신인 달래의 오른손(녹색 소매)이 허공으로 들려 스스로 자신의 왼팔 손목을 꽉 쥐고 있는 물리적/부동성 위반 (Dead body exerting muscular effort)"
        ],
        "physics": "안규홍의 오른손이 달래의 피 묻은 왼손을 정상적으로 잡고 있으나, 그 아래에서 죽은 달래의 오른손이 중력을 거스르고 자신의 왼팔을 쥐어 버티는 불가능한 근육 작용이 나타났다."
       }
      ],
      "all_candidates_fail": true
     },
     "normalized": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 3,
        "verdict_ko": "미디엄 샷 프레이밍과 조명은 훌륭하게 구현되었으나, 시신인 달래의 다른 손이 스스로 자신의 팔을 꽉 쥐고 있는 치명적인 물리적 오류가 발생했습니다."
       },
       {
        "label": "A",
        "score": 1,
        "verdict_ko": "프레이밍 지시(미디엄 샷)를 완전히 무시하고 풀 샷으로 렌더링했으며, 달래에게 세 번째 팔이 돋아난 심각한 해부학적 오류가 있습니다."
       }
      ],
      "all_candidates_fail": true,
      "readings": [
       {
        "label": "A",
        "direction": "안규홍의 시선은 자신이 잡고 들어 올린 피 묻은 팔을 향하고 있다.",
        "built_space": "넓은 야영지 배경에 목책과 초가 지붕, 가마솥 등이 풀 샷으로 널찍하게 배치되어 있다.",
        "entities": "안규홍(참조 이미지 복장, 목에 가락지)과 달래(녹색 저고리, 검은 치마). 그러나 달래의 원래 두 팔이 몸 위에 놓여 있음에도, 안규홍이 잡고 있는 피 묻은 세 번째 맨팔이 존재한다.",
        "hard_violations": [
         "달래에게 세 번째 팔이 존재하는 해부학적 불가능성 (Extra limbs)",
         "미디엄 샷 및 상체 위주의 프레이밍 지시를 위반하고 배경을 지나치게 넓게 잡은 프레이밍"
        ],
        "physics": "달래의 두 팔은 몸 위에 지탱되어 늘어져 있으나, 안규홍이 잡고 있는 세 번째 팔은 어깨와의 연결 없이 허공에서 솟아나 들려 있다."
       },
       {
        "label": "B",
        "direction": "안규홍의 시선은 들어 올린 달래의 손과 얼굴 쪽을 향하고 있다.",
        "built_space": "어두운 밤의 야영지로, 뒤편에 목책과 불이 켜진 랜턴이 있어 지시된 공간의 분위기와 구조에 부합한다.",
        "entities": "안규홍(상체 위주, 목에 가락지)과 달래(피 묻은 얼굴, 녹색 저고리). 달래의 손목을 쥐고 있는 손(녹색 소매)에 옥가락지가 보인다.",
        "hard_violations": [
         "시신인 달래의 오른손(녹색 소매)이 허공으로 들려 스스로 자신의 왼팔 손목을 꽉 쥐고 있는 물리적/부동성 위반 (Dead body exerting muscular effort)"
        ],
        "physics": "안규홍의 오른손이 달래의 피 묻은 왼손을 정상적으로 잡고 있으나, 그 아래에서 죽은 달래의 오른손이 중력을 거스르고 자신의 왼팔을 쥐어 버티는 불가능한 근육 작용이 나타났다."
       }
      ]
     },
     "ok": true
    },
    {
     "model": "gpt-high",
     "order": "reverse",
     "display_to_canonical": {
      "A": "B",
      "B": "A"
     },
     "raw": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 3,
        "verdict_ko": "상체와 받쳐 든 손을 강조한 미디엄 숏은 더 충실하지만, 참조 장소에 없는 등불 세 개와 불을 추가했으며 옥가락지도 달래가 아닌 안규홍의 손가락에 보인다."
       },
       {
        "label": "B",
        "score": 5,
        "verdict_ko": "장소와 신체 지지는 충실하지만, 시신과 멍석 대부분까지 펼친 넓은 구도가 상체 중심 미디엄 숏을 어기며 두 반쪽 옥가락지의 구현도 부족하다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "안규홍은 고개와 눈을 아래로 향해 자신이 받친 달래의 피 묻은 손을 본다. 달래의 얼굴은 위쪽으로 향하고 눈은 감겨 있다. 들어 올린 팔은 두 사람 사이에 놓여 행동의 대상이 명확하다.",
        "built_space": "흙바닥 뒤로 통나무 울타리와 왼쪽 초가 구조물 일부가 보인다. 멍석은 한 장이며 달래가 그 위에 눕고 안규홍이 오른쪽에서 무릎을 꿇는다. 하단에는 요구한 좁은 띠보다 넓은 멍석 표면이 노출된다. 왼쪽 위와 왼쪽 아래, 중앙 뒤에 등불이 각각 하나씩 있고 손 뒤에는 작은 불이 보이는데, 이 조명 물체들은 장소 참조에 없다.",
        "entities": "인물은 젊은 한국인 남녀로 제시된 두 사람뿐이다. 안규홍의 묶은 검은 머리와 낡은 미색 옷, 달래의 검은 머리와 녹색 저고리·검은 치마는 참조와 대체로 맞는다. 달래의 얼굴과 들어 올린 손에는 피와 흙이 보인다. 녹색 가락지는 달래의 손목을 감싼 안규홍의 손가락에 보이며, 달래 자신의 손가락에서는 확인되지 않는다. 목걸이는 반쪽보다 닫힌 원형 옥고리로 보인다. 읽을 수 있는 글자는 없다.",
        "hard_violations": [
         "장소 참조와 장면 지시에 없는 등불 세 개 및 타오르는 불을 추가했다."
        ],
        "physics": "안규홍은 접힌 다리로 바닥에 몸을 지지하고 상체를 숙인다. 한 손이 달래의 손바닥과 손을 받치고 다른 손이 손목을 감싸므로 올라간 전완의 지지점이 명확하다. 달래의 머리와 몸통은 멍석에 놓이고 손가락은 느슨하게 굽어 있어, 지지 없이 떠 있는 신체 부위는 보이지 않는다. 목의 옥고리는 끈에 매달려 있다."
       },
       {
        "label": "B",
        "direction": "안규홍의 시선은 아래쪽의 들어 올린 달래 손과 손목을 향한다. 달래는 눈을 감고 얼굴을 위로 향한 채 누워 있다. 안규홍이 잡은 팔은 두 사람 사이로 올라가며 다른 손은 자기 허벅지 위에 놓인다.",
        "built_space": "좌우 초가 구조물 두 동, 뒤쪽 통나무 울타리, 오른쪽 뒤 망루 한 기, 왼쪽 앞 가마솥과 흙·돌 화덕이 보여 참조 장소의 주요 배치가 유지된다. 멍석 한 장 위에 달래가 가로로 눕고 안규홍이 오른쪽에서 무릎을 꿇는다. 다만 멍석 대부분과 넓은 공터, 안규홍의 접힌 다리까지 보여 요구된 상체 중심 미디엄 숏보다 상당히 넓다.",
        "entities": "추가 인물 없이 젊은 한국인 남성과 여성 두 명이 보이며, 검은 머리와 미색 남성 복식, 녹색 저고리·검은 치마는 참조에 부합한다. 달래의 흐트러진 머리, 얼굴과 손 및 치마의 피·흙 얼룩이 보인다. 달래의 들어 올린 손가락은 비스듬하지만 반쪽 옥가락지는 식별되지 않는다. 안규홍의 목에는 녹색 고리가 있으나 반쪽이 아니라 닫힌 원형으로 읽힌다. 밤의 어두운 색조가 유지되며 읽을 수 있는 글자는 없다.",
        "hard_violations": [],
        "physics": "안규홍의 무릎과 접힌 정강이가 멍석에 닿아 체중을 지지한다. 달래의 머리·몸통·나머지 팔은 멍석에 놓여 있고, 올라간 전완은 안규홍이 손목을 감싸 잡아 지지한다. 손목 위 손가락은 다소 세워져 있지만 약간 굽어 있으며, 이 모습만으로 무지지 부유나 불가능한 자세라고 단정할 수는 없다. 목걸이는 목의 끈이 지지한다."
       }
      ],
      "all_candidates_fail": true
     },
     "normalized": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 3,
        "verdict_ko": "상체와 받쳐 든 손을 강조한 미디엄 숏은 더 충실하지만, 참조 장소에 없는 등불 세 개와 불을 추가했으며 옥가락지도 달래가 아닌 안규홍의 손가락에 보인다."
       },
       {
        "label": "A",
        "score": 5,
        "verdict_ko": "장소와 신체 지지는 충실하지만, 시신과 멍석 대부분까지 펼친 넓은 구도가 상체 중심 미디엄 숏을 어기며 두 반쪽 옥가락지의 구현도 부족하다."
       }
      ],
      "all_candidates_fail": true,
      "readings": [
       {
        "label": "B",
        "direction": "안규홍은 고개와 눈을 아래로 향해 자신이 받친 달래의 피 묻은 손을 본다. 달래의 얼굴은 위쪽으로 향하고 눈은 감겨 있다. 들어 올린 팔은 두 사람 사이에 놓여 행동의 대상이 명확하다.",
        "built_space": "흙바닥 뒤로 통나무 울타리와 왼쪽 초가 구조물 일부가 보인다. 멍석은 한 장이며 달래가 그 위에 눕고 안규홍이 오른쪽에서 무릎을 꿇는다. 하단에는 요구한 좁은 띠보다 넓은 멍석 표면이 노출된다. 왼쪽 위와 왼쪽 아래, 중앙 뒤에 등불이 각각 하나씩 있고 손 뒤에는 작은 불이 보이는데, 이 조명 물체들은 장소 참조에 없다.",
        "entities": "인물은 젊은 한국인 남녀로 제시된 두 사람뿐이다. 안규홍의 묶은 검은 머리와 낡은 미색 옷, 달래의 검은 머리와 녹색 저고리·검은 치마는 참조와 대체로 맞는다. 달래의 얼굴과 들어 올린 손에는 피와 흙이 보인다. 녹색 가락지는 달래의 손목을 감싼 안규홍의 손가락에 보이며, 달래 자신의 손가락에서는 확인되지 않는다. 목걸이는 반쪽보다 닫힌 원형 옥고리로 보인다. 읽을 수 있는 글자는 없다.",
        "hard_violations": [
         "장소 참조와 장면 지시에 없는 등불 세 개 및 타오르는 불을 추가했다."
        ],
        "physics": "안규홍은 접힌 다리로 바닥에 몸을 지지하고 상체를 숙인다. 한 손이 달래의 손바닥과 손을 받치고 다른 손이 손목을 감싸므로 올라간 전완의 지지점이 명확하다. 달래의 머리와 몸통은 멍석에 놓이고 손가락은 느슨하게 굽어 있어, 지지 없이 떠 있는 신체 부위는 보이지 않는다. 목의 옥고리는 끈에 매달려 있다."
       },
       {
        "label": "A",
        "direction": "안규홍의 시선은 아래쪽의 들어 올린 달래 손과 손목을 향한다. 달래는 눈을 감고 얼굴을 위로 향한 채 누워 있다. 안규홍이 잡은 팔은 두 사람 사이로 올라가며 다른 손은 자기 허벅지 위에 놓인다.",
        "built_space": "좌우 초가 구조물 두 동, 뒤쪽 통나무 울타리, 오른쪽 뒤 망루 한 기, 왼쪽 앞 가마솥과 흙·돌 화덕이 보여 참조 장소의 주요 배치가 유지된다. 멍석 한 장 위에 달래가 가로로 눕고 안규홍이 오른쪽에서 무릎을 꿇는다. 다만 멍석 대부분과 넓은 공터, 안규홍의 접힌 다리까지 보여 요구된 상체 중심 미디엄 숏보다 상당히 넓다.",
        "entities": "추가 인물 없이 젊은 한국인 남성과 여성 두 명이 보이며, 검은 머리와 미색 남성 복식, 녹색 저고리·검은 치마는 참조에 부합한다. 달래의 흐트러진 머리, 얼굴과 손 및 치마의 피·흙 얼룩이 보인다. 달래의 들어 올린 손가락은 비스듬하지만 반쪽 옥가락지는 식별되지 않는다. 안규홍의 목에는 녹색 고리가 있으나 반쪽이 아니라 닫힌 원형으로 읽힌다. 밤의 어두운 색조가 유지되며 읽을 수 있는 글자는 없다.",
        "hard_violations": [],
        "physics": "안규홍의 무릎과 접힌 정강이가 멍석에 닿아 체중을 지지한다. 달래의 머리·몸통·나머지 팔은 멍석에 놓여 있고, 올라간 전완은 안규홍이 손목을 감싸 잡아 지지한다. 손목 위 손가락은 다소 세워져 있지만 약간 굽어 있으며, 이 모습만으로 무지지 부유나 불가능한 자세라고 단정할 수는 없다. 목걸이는 목의 끈이 지지한다."
       }
      ]
     },
     "ok": true
    }
   ],
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "slot_winner_match": false,
   "slot_winner": {
    "gemini-pro": "B",
    "gpt-high": "A"
   },
   "route": "cross_slot_combined"
  },
  "dual": {
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "normalized": {
    "A": 1.333,
    "B": 1.6
   },
   "adjusted": {
    "A": 1.083,
    "B": 1.35
   },
   "violations": {
    "A": [
     "[gemini-pro] 달래에게 세 번째 팔이 존재하는 해부학적 불가능성 (Extra limbs)",
     "[gemini-pro] 미디엄 샷 및 상체 위주의 프레이밍 지시를 위반하고 배경을 지나치게 넓게 잡은 프레이밍"
    ],
    "B": [
     "[gemini-pro] 시신인 달래의 오른손(녹색 소매)이 허공으로 들려 스스로 자신의 왼팔 손목을 꽉 쥐고 있는 물리적/부동성 위반 (Dead body exerting muscular effort)",
     "[gpt-high] 장소 참조와 장면 지시에 없는 등불 세 개 및 타오르는 불을 추가했다."
    ]
   },
   "per_model_winner": {
    "gemini-pro": "B",
    "gpt-high": "A"
   },
   "agreed": false
  },
  "totals": {
   "B": 1350,
   "A": 1083
  },
  "selected": "B",
  "ranking": [
   "B",
   "A"
  ],
  "verdicts": [
   {
    "label": "B",
    "score": 1350,
    "verdict_ko": "미디엄 샷 프레이밍과 조명은 훌륭하게 구현되었으나, 시신인 달래의 다른 손이 스스로 자신의 팔을 꽉 쥐고 있는 치명적인 물리적 오류가 발생했습니다.  ★위반: [gemini-pro] 시신인 달래의 오른손(녹색 소매)이 허공으로 들려 스스로 자신의 왼팔 손목을 꽉 쥐고 있는 물리적/부동성 위반 (Dead body exerting muscular effort) / [gpt-high] 장소 참조와 장면 지시에 없는 등불 세 개 및 타오르는 불을 추가했다."
   },
   {
    "label": "A",
    "score": 1083,
    "verdict_ko": "프레이밍 지시(미디엄 샷)를 완전히 무시하고 풀 샷으로 렌더링했으며, 달래에게 세 번째 팔이 돋아난 심각한 해부학적 오류가 있습니다.  ★위반: [gemini-pro] 달래에게 세 번째 팔이 존재하는 해부학적 불가능성 (Extra limbs) / [gemini-pro] 미디엄 샷 및 상체 위주의 프레이밍 지시를 위반하고 배경을 지나치게 넓게 잡은 프레이밍"
   }
  ],
  "refs": [
   {
    "label": "LOCATION PHOTOGRAPH — the exact place of this shot: its architecture, materials, fixed features and lighting mood are spatial truth; stage the moment inside this place. Never copy its camera framing.",
    "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/db00517b-165b-4005-a287-f28936b59b95/images/272bd7af-9294-4490-98c6-88a40e9deacc/scene/recipe/groupbg_mountain_camp_clearing_cb9567.png",
    "asset_id": "c5df4442-a576-4bdd-9e42-441cb4544a5e",
    "role": "bgfirst_group_bg"
   },
   {
    "label": "CHARACTER REFERENCE — 안규홍: the exact person appearing in this shot; match face, hair and build exactly.",
    "path": "<bytes:1028384>",
    "asset_id": "3b5894db-ce55-4f34-b966-5e725c3aa952",
    "role": "character_ref"
   },
   {
    "label": "CHARACTER REFERENCE — 달래: the exact person appearing in this shot; match face, hair and build exactly.",
    "path": "<bytes:959810>",
    "asset_id": "a51dee5b-8b43-4345-b827-da2ff15b545d",
    "role": "character_ref"
   }
  ],
  "critique_skipped": true,
  "needs_reshoot": true,
  "shot_run_produced": false,
  "shot_run_uid": "06aa1433-f8cc-76f7-a563-ffe581b1bc66",
  "bgfirst": {
   "bg_path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/db00517b-165b-4005-a287-f28936b59b95/images/272bd7af-9294-4490-98c6-88a40e9deacc/scene/recipe/S45sh5__bgfirst_bg.png",
   "bg_asset_id": "92eb3e0b-4760-4f89-86cb-984991ffab7c",
   "bg_record_key": "S45sh5::bgfirst_bg",
   "chain_winner": false,
   "authority": "groupbg",
   "group_key": "mountain_camp_clearing",
   "groupbg_asset_id": "c5df4442-a576-4bdd-9e42-441cb4544a5e"
  },
  "ref_mode": "그룹 배경+엔티티 (2택1: 무콘티 승)",
  "share_plan": {
   "ref_plan": "background"
  }
 },
 "S45sh5::cine": {
  "applied": false,
  "attempted_at": "2026-09-09T09:30:46.421787+00:00",
  "fingerprint": "6e3f493737f1b62efb0c0b1284cd811f8baf0cd01d4691c7eaf0f46b3fcb5373",
  "fingerprint_version": 2,
  "provider": "mai",
  "endpoint": "openrouter/chat-completions",
  "model": "microsoft/mai-image-2.6",
  "multiroll_tag": "still_S45sh5_cine",
  "slot": "",
  "pack": "24.202608252115",
  "source_file": "S45sh5_sel.png",
  "source_sha256": "a0d58518c872e0ee560399a834309fc74069cc76215891e900f97c4ec8168ca0",
  "error": "RuntimeError: moderation blocked: {\"message\": \"Output content violated imagegen safety policies.\", \"code\": 400}",
  "moderation_refusals": 2,
  "declined": true,
  "declined_reason": "moderation"
 },
 "S45sh20::signage": {
  "fp": "a0b343d8c0eeda7d",
  "inscriptions": [],
  "cues": [],
  "dropped": []
 },
 "S45sh20": {
  "input_fingerprint": "510ffee5fdb2b536",
  "prompt": "Create ONE FINAL photorealistic live-action film still of the moment below — early 20th-century Korea, Jeolla region; all people are Korean unless explicitly stated otherwise. TIME OF DAY (lock): night.\n\nSHOT TEXT (authoritative, Korean): 다시 앞으로 튀어나가려는 안규홍의 양팔을 염재보와 오주일이 양옆에서 꽉 부여잡은 역동적인 모습.\n\nLOCATION (lock): The open space immediately beside the mourning mat in the mountain camp, along the attempted departure route. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: medium shot\n- KEY BACKGROUND ELEMENTS: 낫 (아직 안규홍의 손에 들려 있다) — Partially visible below the restraining arms, following the angle of his lowered hand; used as Retains the threat at the lower edge without obscuring either grip.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Maintain the preceding nighttime exposure and muted palette, allowing naturalistic contrast to carry the struggle without theatrical highlights.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The clothing, hair and overall look of 안규홍 — who appear both in that photo and in this shot — are LOCKED to that photo. Anyone else visible in that photo is NOT in this shot: never carry their face, body or clothing onto anyone here. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The straw mat remains spread in the camp. 안규홍: He stands struggling with his arms restrained, still holding the sickle and wearing his half jade ring at his neck. 염재보: He stands braced with an extended restraining grip. 오주일: He stands braced with an extended restraining grip.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 안규홍 (한국인 남성, 청년 나이대, 검은 머리, 단단히 묶은 머리); 염재보 (한국인 남성, 성인 나이대, 자연스러운 검은 머리); 오주일 (한국인 남성, 성인 나이대, 자연스러운 검은 머리) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.",
  "roll_prompts": {
   "A": "Create ONE FINAL photorealistic live-action film still of the moment below — early 20th-century Korea, Jeolla region; all people are Korean unless explicitly stated otherwise. TIME OF DAY (lock): night.\n\nSHOT TEXT (authoritative, Korean): 다시 앞으로 튀어나가려는 안규홍의 양팔을 염재보와 오주일이 양옆에서 꽉 부여잡은 역동적인 모습.\n\nLOCATION (lock): The open space immediately beside the mourning mat in the mountain camp, along the attempted departure route. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: medium shot\n- KEY BACKGROUND ELEMENTS: 낫 (아직 안규홍의 손에 들려 있다) — Partially visible below the restraining arms, following the angle of his lowered hand; used as Retains the threat at the lower edge without obscuring either grip.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Maintain the preceding nighttime exposure and muted palette, allowing naturalistic contrast to carry the struggle without theatrical highlights.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The clothing, hair and overall look of 안규홍 — who appear both in that photo and in this shot — are LOCKED to that photo. Anyone else visible in that photo is NOT in this shot: never carry their face, body or clothing onto anyone here. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The straw mat remains spread in the camp. 안규홍: He stands struggling with his arms restrained, still holding the sickle and wearing his half jade ring at his neck. 염재보: He stands braced with an extended restraining grip. 오주일: He stands braced with an extended restraining grip.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 안규홍 (한국인 남성, 청년 나이대, 검은 머리, 단단히 묶은 머리); 염재보 (한국인 남성, 성인 나이대, 자연스러운 검은 머리); 오주일 (한국인 남성, 성인 나이대, 자연스러운 검은 머리) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.",
   "B": "Create ONE FINAL photorealistic live-action film still of the moment below — early 20th-century Korea, Jeolla region; all people are Korean unless explicitly stated otherwise. TIME OF DAY (lock): night.\n\nSHOT TEXT (authoritative, Korean): 다시 앞으로 튀어나가려는 안규홍의 양팔을 염재보와 오주일이 양옆에서 꽉 부여잡은 역동적인 모습.\n\nLOCATION (lock): The open space immediately beside the mourning mat in the mountain camp, along the attempted departure route. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: medium shot\n- KEY BACKGROUND ELEMENTS: 낫 (아직 안규홍의 손에 들려 있다) — Partially visible below the restraining arms, following the angle of his lowered hand; used as Retains the threat at the lower edge without obscuring either grip.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Maintain the preceding nighttime exposure and muted palette, allowing naturalistic contrast to carry the struggle without theatrical highlights.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The clothing, hair and overall look of 안규홍 — who appear both in that photo and in this shot — are LOCKED to that photo. Anyone else visible in that photo is NOT in this shot: never carry their face, body or clothing onto anyone here. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The straw mat remains spread in the camp. 안규홍: He stands struggling with his arms restrained, still holding the sickle and wearing his half jade ring at his neck. 염재보: He stands braced with an extended restraining grip. 오주일: He stands braced with an extended restraining grip.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 안규홍 (한국인 남성, 청년 나이대, 검은 머리, 단단히 묶은 머리); 염재보 (한국인 남성, 성인 나이대, 자연스러운 검은 머리); 오주일 (한국인 남성, 성인 나이대, 자연스러운 검은 머리) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.\n\nCOMPOSITION VARIATION (this candidate only): keep the place, the people and the moment exactly as specified, and keep from the CAMERA & FRAME contract its subject scale, screen placement and key background elements exactly. For this candidate only, its camera angle and height are deliberately left open: choose a camera position distinctly different from the obvious one, as a film director picking a second setup on the same blocking. Stay true to any explicit framing the SHOT TEXT itself dictates."
  },
  "shot_run_spend_attempt_count": 1,
  "readings": [
   {
    "label": "A",
    "direction": "안규홍은 시선을 아래로 향하고 있으며, 염재보와 오주일은 안규홍을 바라보며 양옆에서 그의 팔을 단단히 제압하고 있다.",
    "built_space": "야간의 산채 공터. 배경에 나무 울타리가 있고 바닥에는 멍석이 깔려 있다. 이전 샷과 유사한 위치에 호롱불들이 배치되어 공간의 연속성이 유지된다.",
    "entities": "안규홍(중앙)은 흰색 한복을 입고 오른손에 낫을 쥐고 있으나 목의 옥반지가 보이지 않는다. 염재보(왼쪽)는 삿갓과 누비옷, 오주일(오른쪽)은 제복과 모자 등 레퍼런스의 복장을 정확히 착용하고 있다.",
    "hard_violations": [],
    "physics": "세 사람 모두 바닥에 발을 안정적으로 딛고 있으며, 양옆에서 팔을 꽉 부여잡아 앞으로 튀어나가려는 것을 막는 물리적인 힘과 무게 중심이 자연스럽게 표현되었다."
   },
   {
    "label": "B",
    "direction": "안규홍은 정면을 강렬하게 노려보고 있으며, 양옆의 인물들은 긴장한 표정으로 그를 제압하며 전방 및 안규홍 쪽을 향해 시선을 두고 있다.",
    "built_space": "야간의 산채. 나무 울타리와 바닥의 멍석, 뒤편에 호롱불이 놓여 있어 지정된 장소의 요소를 갖추고 있다.",
    "entities": "안규홍(중앙)은 옥반지를 목에 걸고 낫을 들고 있다. 그러나 왼쪽의 염재보는 삿갓을 쓰지 않은 채 안규홍과 완벽히 동일한 얼굴을 하고 있으며, 오른쪽의 오주일 역시 제복 모자가 누락되었다.",
    "hard_violations": [
     "[gemini-pro] 안규홍의 얼굴이 왼쪽 인물에게 그대로 중복되어 나타남 (Duplicated bodies/faces)"
    ],
    "physics": "양팔을 붙잡는 손의 위치와 당기는 자세는 성립하나, 안규홍이 낫을 쥔 오른손의 손가락 형태가 다소 뭉개져 있고 어색하게 결합되어 있다."
   }
  ],
  "cross_model_order": {
   "policy": "cross_model_order_v2_slot0_fwd_slot1_rev_parallel2",
   "slots": [
    {
     "model": "gemini-pro",
     "order": "forward",
     "display_to_canonical": {
      "A": "A",
      "B": "B"
     },
     "raw": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 7,
        "verdict_ko": "지시된 미디엄 샷 대신 전신이 보이는 풀 샷으로 연출되고 옥반지가 누락되었으나, 세 인물의 복장과 물리적 제압 동작을 레퍼런스에 맞게 안정적으로 구현했습니다."
       },
       {
        "label": "B",
        "score": 2,
        "verdict_ko": "미디엄 샷 프레이밍과 옥반지 묘사는 지시를 따랐으나, 왼쪽 인물의 얼굴이 주인공과 똑같이 중복 생성되고 주요 인물들의 모자가 모두 누락되는 치명적인 오류가 있습니다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "안규홍은 시선을 아래로 향하고 있으며, 염재보와 오주일은 안규홍을 바라보며 양옆에서 그의 팔을 단단히 제압하고 있다.",
        "built_space": "야간의 산채 공터. 배경에 나무 울타리가 있고 바닥에는 멍석이 깔려 있다. 이전 샷과 유사한 위치에 호롱불들이 배치되어 공간의 연속성이 유지된다.",
        "entities": "안규홍(중앙)은 흰색 한복을 입고 오른손에 낫을 쥐고 있으나 목의 옥반지가 보이지 않는다. 염재보(왼쪽)는 삿갓과 누비옷, 오주일(오른쪽)은 제복과 모자 등 레퍼런스의 복장을 정확히 착용하고 있다.",
        "hard_violations": [],
        "physics": "세 사람 모두 바닥에 발을 안정적으로 딛고 있으며, 양옆에서 팔을 꽉 부여잡아 앞으로 튀어나가려는 것을 막는 물리적인 힘과 무게 중심이 자연스럽게 표현되었다."
       },
       {
        "label": "B",
        "direction": "안규홍은 정면을 강렬하게 노려보고 있으며, 양옆의 인물들은 긴장한 표정으로 그를 제압하며 전방 및 안규홍 쪽을 향해 시선을 두고 있다.",
        "built_space": "야간의 산채. 나무 울타리와 바닥의 멍석, 뒤편에 호롱불이 놓여 있어 지정된 장소의 요소를 갖추고 있다.",
        "entities": "안규홍(중앙)은 옥반지를 목에 걸고 낫을 들고 있다. 그러나 왼쪽의 염재보는 삿갓을 쓰지 않은 채 안규홍과 완벽히 동일한 얼굴을 하고 있으며, 오른쪽의 오주일 역시 제복 모자가 누락되었다.",
        "hard_violations": [
         "안규홍의 얼굴이 왼쪽 인물에게 그대로 중복되어 나타남 (Duplicated bodies/faces)"
        ],
        "physics": "양팔을 붙잡는 손의 위치와 당기는 자세는 성립하나, 안규홍이 낫을 쥔 오른손의 손가락 형태가 다소 뭉개져 있고 어색하게 결합되어 있다."
       }
      ],
      "all_candidates_fail": false
     },
     "normalized": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 7,
        "verdict_ko": "지시된 미디엄 샷 대신 전신이 보이는 풀 샷으로 연출되고 옥반지가 누락되었으나, 세 인물의 복장과 물리적 제압 동작을 레퍼런스에 맞게 안정적으로 구현했습니다."
       },
       {
        "label": "B",
        "score": 2,
        "verdict_ko": "미디엄 샷 프레이밍과 옥반지 묘사는 지시를 따랐으나, 왼쪽 인물의 얼굴이 주인공과 똑같이 중복 생성되고 주요 인물들의 모자가 모두 누락되는 치명적인 오류가 있습니다."
       }
      ],
      "all_candidates_fail": false,
      "readings": [
       {
        "label": "A",
        "direction": "안규홍은 시선을 아래로 향하고 있으며, 염재보와 오주일은 안규홍을 바라보며 양옆에서 그의 팔을 단단히 제압하고 있다.",
        "built_space": "야간의 산채 공터. 배경에 나무 울타리가 있고 바닥에는 멍석이 깔려 있다. 이전 샷과 유사한 위치에 호롱불들이 배치되어 공간의 연속성이 유지된다.",
        "entities": "안규홍(중앙)은 흰색 한복을 입고 오른손에 낫을 쥐고 있으나 목의 옥반지가 보이지 않는다. 염재보(왼쪽)는 삿갓과 누비옷, 오주일(오른쪽)은 제복과 모자 등 레퍼런스의 복장을 정확히 착용하고 있다.",
        "hard_violations": [],
        "physics": "세 사람 모두 바닥에 발을 안정적으로 딛고 있으며, 양옆에서 팔을 꽉 부여잡아 앞으로 튀어나가려는 것을 막는 물리적인 힘과 무게 중심이 자연스럽게 표현되었다."
       },
       {
        "label": "B",
        "direction": "안규홍은 정면을 강렬하게 노려보고 있으며, 양옆의 인물들은 긴장한 표정으로 그를 제압하며 전방 및 안규홍 쪽을 향해 시선을 두고 있다.",
        "built_space": "야간의 산채. 나무 울타리와 바닥의 멍석, 뒤편에 호롱불이 놓여 있어 지정된 장소의 요소를 갖추고 있다.",
        "entities": "안규홍(중앙)은 옥반지를 목에 걸고 낫을 들고 있다. 그러나 왼쪽의 염재보는 삿갓을 쓰지 않은 채 안규홍과 완벽히 동일한 얼굴을 하고 있으며, 오른쪽의 오주일 역시 제복 모자가 누락되었다.",
        "hard_violations": [
         "안규홍의 얼굴이 왼쪽 인물에게 그대로 중복되어 나타남 (Duplicated bodies/faces)"
        ],
        "physics": "양팔을 붙잡는 손의 위치와 당기는 자세는 성립하나, 안규홍이 낫을 쥔 오른손의 손가락 형태가 다소 뭉개져 있고 어색하게 결합되어 있다."
       }
      ]
     },
     "ok": true
    },
    {
     "model": "gpt-high",
     "order": "reverse",
     "display_to_canonical": {
      "A": "B",
      "B": "A"
     },
     "raw": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 8,
        "verdict_ko": "모자 등 인물 참조의 일부 차이는 있지만, 더 밀착된 프레이밍에서 앞으로 튀어나가려는 안규홍과 양팔을 붙드는 힘, 하단에 일부만 보이는 낫을 가장 충실하게 구현했다."
       },
       {
        "label": "B",
        "score": 6,
        "verdict_ko": "염재보와 오주일의 모자·복식은 더 충실하지만, 발까지 드러나는 넓은 구도와 아래를 보는 안규홍의 자세가 지정된 미디엄 숏 및 역동적인 재돌진 순간에서 멀어진다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "안규홍은 화면 왼쪽 앞을 노려보며 상체를 앞으로 내민다. 양옆 남성은 안규홍을 바라보면서 각각 그의 위팔을 붙잡는다. 안규홍의 오른손에 잡힌 낫은 허리 아래에서 비스듬히 내려가며 날이 화면 하단 밖으로 이어진다. 낫으로 특정 사람을 겨누지는 않으며, 낮춰 든 위협이라는 지시와 맞는다.",
        "built_space": "뒤쪽에 세로 목책과 가로 지지대, 왼쪽에 초가지붕 구조물 일부가 있다. 등불은 왼쪽 구조물의 것 하나, 왼쪽 지면의 것 하나, 중앙에서 몸에 일부 가려진 것 하나로 세 개 보인다. 중앙 뒤와 오른쪽 가장자리에 불빛이 있고, 돗자리 한 장이 인물들 뒤 오른쪽 지면에 펼쳐져 있다. 세 사람은 돗자리 앞의 흙바닥 공간을 차지한다. 참조의 야간 산중 야영지 재료와 시설을 대체로 유지하며 반사면은 없다.",
        "entities": "세 명 모두 참조 인물들의 연령대에 부합하는 동아시아계 남성으로 보이며 추가 인물은 없다. 중앙 안규홍의 묶은 검은 머리, 얼굴 윤곽, 때 묻은 밝은 한복과 허리띠가 참조에 가깝다. 목에는 참조처럼 녹색 고리 장식이 있지만 반쪽보다는 완전한 고리처럼 보인다. 왼쪽 염재보는 어두운 누비옷을 입었으나 참조의 삿갓이 없고 머리를 뒤로 묶었다. 오른쪽 오주일은 어두운 군복과 허리의 칼집을 갖췄으나 군모가 없으며 붉은 선 장식도 참조와 다르다. 낫은 실제 나무 손잡이와 금속 날로 보인다. 읽을 수 있는 글자는 없다.",
        "hard_violations": [],
        "physics": "안규홍은 몸을 앞으로 기울이고 다리를 벌린 채 버티며, 양옆 남성의 손이 실제로 양쪽 위팔을 감싸 잡고 있다. 팔과 옷의 당김이 서로 반대되는 힘을 전달한다. 발은 프레임 밖이지만 다리가 아래로 이어지는 서 있는 자세이며 공중에 뜬 신체는 없다. 낫 손잡이는 오른손이 쥐고, 목걸이는 목의 끈에, 칼집은 오주일의 허리띠에 지지되어 있다. 돗자리와 등불은 지면에 놓여 있다."
       },
       {
        "label": "B",
        "direction": "안규홍은 고개와 시선을 앞쪽 아래로 떨어뜨린 채 상체를 숙인다. 염재보와 오주일은 중앙의 안규홍을 바라보며 양쪽 팔을 붙든다. 낫은 오른손 아래로 내려오고 굽은 날끝은 화면 오른쪽 위를 향하지만 누구를 직접 겨누지는 않는다. 낮춘 낫의 방향은 맞으나, 전방을 향한 재돌진보다 멈춰 서서 버티는 순간에 가깝다.",
        "built_space": "세로 목책과 가로 지지대, 왼쪽 초가지붕 일부가 보이며 등불은 왼쪽 구조물에 하나, 왼쪽 지면에 하나, 염재보 뒤 지면에 하나로 세 개다. 오른쪽 뒤에는 불을 피운 자리가 있고 중앙 뒤에도 작은 불빛이 보인다. 돗자리 한 장이 세 사람 뒤에 넓게 펼쳐져 있다. 사람들은 그 앞 흙바닥에 서 있다. 야영지의 재료와 야간 분위기는 이어지지만, 발과 넓은 바닥까지 포함하여 요구된 미디엄 숏보다 훨씬 넓다. 반사면은 없다.",
        "entities": "추가 인물 없이 성인 동아시아계 남성 세 명이 등장한다. 안규홍의 묶은 검은 머리와 밝은색 낡은 한복은 참조와 가깝다. 가슴에서 반옥 목걸이는 식별되지 않는다. 염재보의 삿갓·어두운 누비옷, 오주일의 군모·단추 달린 군복·허리 칼집은 각 참조에 더 가깝다. 안규홍의 손에는 굽은 금속 날과 짧은 나무 손잡이를 가진 낫이 있으며 날 전체가 화면 안에 보인다. 읽을 수 있는 글자는 없다.",
        "hard_violations": [],
        "physics": "안규홍의 화면 왼쪽 발은 흙바닥에 닿아 있고 반대쪽 다리는 하단 밖으로 이어진다. 염재보는 벌린 두 발로 지면을 딛고, 오주일도 보이는 발을 지면에 댄 채 무릎을 굽힌다. 양옆의 손은 안규홍의 위팔과 팔꿈치 부근을 실제로 붙잡고 있어 저지가 가능하다. 낫은 안규홍의 오른손에, 칼집은 오주일의 허리띠에 지지된다. 떠 있는 물체나 불가능한 신체 지지는 보이지 않는다."
       }
      ],
      "all_candidates_fail": false
     },
     "normalized": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 8,
        "verdict_ko": "모자 등 인물 참조의 일부 차이는 있지만, 더 밀착된 프레이밍에서 앞으로 튀어나가려는 안규홍과 양팔을 붙드는 힘, 하단에 일부만 보이는 낫을 가장 충실하게 구현했다."
       },
       {
        "label": "A",
        "score": 6,
        "verdict_ko": "염재보와 오주일의 모자·복식은 더 충실하지만, 발까지 드러나는 넓은 구도와 아래를 보는 안규홍의 자세가 지정된 미디엄 숏 및 역동적인 재돌진 순간에서 멀어진다."
       }
      ],
      "all_candidates_fail": false,
      "readings": [
       {
        "label": "B",
        "direction": "안규홍은 화면 왼쪽 앞을 노려보며 상체를 앞으로 내민다. 양옆 남성은 안규홍을 바라보면서 각각 그의 위팔을 붙잡는다. 안규홍의 오른손에 잡힌 낫은 허리 아래에서 비스듬히 내려가며 날이 화면 하단 밖으로 이어진다. 낫으로 특정 사람을 겨누지는 않으며, 낮춰 든 위협이라는 지시와 맞는다.",
        "built_space": "뒤쪽에 세로 목책과 가로 지지대, 왼쪽에 초가지붕 구조물 일부가 있다. 등불은 왼쪽 구조물의 것 하나, 왼쪽 지면의 것 하나, 중앙에서 몸에 일부 가려진 것 하나로 세 개 보인다. 중앙 뒤와 오른쪽 가장자리에 불빛이 있고, 돗자리 한 장이 인물들 뒤 오른쪽 지면에 펼쳐져 있다. 세 사람은 돗자리 앞의 흙바닥 공간을 차지한다. 참조의 야간 산중 야영지 재료와 시설을 대체로 유지하며 반사면은 없다.",
        "entities": "세 명 모두 참조 인물들의 연령대에 부합하는 동아시아계 남성으로 보이며 추가 인물은 없다. 중앙 안규홍의 묶은 검은 머리, 얼굴 윤곽, 때 묻은 밝은 한복과 허리띠가 참조에 가깝다. 목에는 참조처럼 녹색 고리 장식이 있지만 반쪽보다는 완전한 고리처럼 보인다. 왼쪽 염재보는 어두운 누비옷을 입었으나 참조의 삿갓이 없고 머리를 뒤로 묶었다. 오른쪽 오주일은 어두운 군복과 허리의 칼집을 갖췄으나 군모가 없으며 붉은 선 장식도 참조와 다르다. 낫은 실제 나무 손잡이와 금속 날로 보인다. 읽을 수 있는 글자는 없다.",
        "hard_violations": [],
        "physics": "안규홍은 몸을 앞으로 기울이고 다리를 벌린 채 버티며, 양옆 남성의 손이 실제로 양쪽 위팔을 감싸 잡고 있다. 팔과 옷의 당김이 서로 반대되는 힘을 전달한다. 발은 프레임 밖이지만 다리가 아래로 이어지는 서 있는 자세이며 공중에 뜬 신체는 없다. 낫 손잡이는 오른손이 쥐고, 목걸이는 목의 끈에, 칼집은 오주일의 허리띠에 지지되어 있다. 돗자리와 등불은 지면에 놓여 있다."
       },
       {
        "label": "A",
        "direction": "안규홍은 고개와 시선을 앞쪽 아래로 떨어뜨린 채 상체를 숙인다. 염재보와 오주일은 중앙의 안규홍을 바라보며 양쪽 팔을 붙든다. 낫은 오른손 아래로 내려오고 굽은 날끝은 화면 오른쪽 위를 향하지만 누구를 직접 겨누지는 않는다. 낮춘 낫의 방향은 맞으나, 전방을 향한 재돌진보다 멈춰 서서 버티는 순간에 가깝다.",
        "built_space": "세로 목책과 가로 지지대, 왼쪽 초가지붕 일부가 보이며 등불은 왼쪽 구조물에 하나, 왼쪽 지면에 하나, 염재보 뒤 지면에 하나로 세 개다. 오른쪽 뒤에는 불을 피운 자리가 있고 중앙 뒤에도 작은 불빛이 보인다. 돗자리 한 장이 세 사람 뒤에 넓게 펼쳐져 있다. 사람들은 그 앞 흙바닥에 서 있다. 야영지의 재료와 야간 분위기는 이어지지만, 발과 넓은 바닥까지 포함하여 요구된 미디엄 숏보다 훨씬 넓다. 반사면은 없다.",
        "entities": "추가 인물 없이 성인 동아시아계 남성 세 명이 등장한다. 안규홍의 묶은 검은 머리와 밝은색 낡은 한복은 참조와 가깝다. 가슴에서 반옥 목걸이는 식별되지 않는다. 염재보의 삿갓·어두운 누비옷, 오주일의 군모·단추 달린 군복·허리 칼집은 각 참조에 더 가깝다. 안규홍의 손에는 굽은 금속 날과 짧은 나무 손잡이를 가진 낫이 있으며 날 전체가 화면 안에 보인다. 읽을 수 있는 글자는 없다.",
        "hard_violations": [],
        "physics": "안규홍의 화면 왼쪽 발은 흙바닥에 닿아 있고 반대쪽 다리는 하단 밖으로 이어진다. 염재보는 벌린 두 발로 지면을 딛고, 오주일도 보이는 발을 지면에 댄 채 무릎을 굽힌다. 양옆의 손은 안규홍의 위팔과 팔꿈치 부근을 실제로 붙잡고 있어 저지가 가능하다. 낫은 안규홍의 오른손에, 칼집은 오주일의 허리띠에 지지된다. 떠 있는 물체나 불가능한 신체 지지는 보이지 않는다."
       }
      ]
     },
     "ok": true
    }
   ],
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "slot_winner_match": false,
   "slot_winner": {
    "gemini-pro": "A",
    "gpt-high": "B"
   },
   "route": "cross_slot_combined"
  },
  "dual": {
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "normalized": {
    "A": 1.75,
    "B": 1.286
   },
   "adjusted": {
    "A": 1.75,
    "B": 1.036
   },
   "violations": {
    "B": [
     "[gemini-pro] 안규홍의 얼굴이 왼쪽 인물에게 그대로 중복되어 나타남 (Duplicated bodies/faces)"
    ]
   },
   "per_model_winner": {
    "gemini-pro": "A",
    "gpt-high": "B"
   },
   "agreed": false
  },
  "totals": {
   "A": 1750,
   "B": 1036
  },
  "selected": "A",
  "ranking": [
   "A",
   "B"
  ],
  "verdicts": [
   {
    "label": "A",
    "score": 1750,
    "verdict_ko": "지시된 미디엄 샷 대신 전신이 보이는 풀 샷으로 연출되고 옥반지가 누락되었으나, 세 인물의 복장과 물리적 제압 동작을 레퍼런스에 맞게 안정적으로 구현했습니다."
   },
   {
    "label": "B",
    "score": 1036,
    "verdict_ko": "미디엄 샷 프레이밍과 옥반지 묘사는 지시를 따랐으나, 왼쪽 인물의 얼굴이 주인공과 똑같이 중복 생성되고 주요 인물들의 모자가 모두 누락되는 치명적인 오류가 있습니다.  ★위반: [gemini-pro] 안규홍의 얼굴이 왼쪽 인물에게 그대로 중복되어 나타남 (Duplicated bodies/faces)"
   }
  ],
  "refs": [
   {
    "label": "PREVIOUS SHOT STILL — a visually related earlier shot of this same place: the location's look, materials, fixed features and lighting mood are LOCKED to this photo; never copy its camera framing. The clothing, hair and overall look of 안규홍 — who appear both in that photo and in this shot — are LOCKED to that photo. Anyone else visible in that photo is NOT in this shot: never carry their face, body or clothing onto anyone here. If this photo shows a character who cannot move (dead or unconscious), that character's exact pose, position and orientation are ALSO LOCKED — treat their body as a fixed prop of the set that only the camera moves around.",
    "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/db00517b-165b-4005-a287-f28936b59b95/images/272bd7af-9294-4490-98c6-88a40e9deacc/scene/recipe/S45sh5_sel.png",
    "asset_id": "9a25d7f4-fa8e-4b50-87cd-708edf72baa5",
    "role": "prev_still"
   },
   {
    "label": "CHARACTER REFERENCE — 안규홍: the exact person appearing in this shot; match face, hair and build exactly.",
    "path": "<bytes:1028384>",
    "asset_id": "3b5894db-ce55-4f34-b966-5e725c3aa952",
    "role": "character_ref"
   },
   {
    "label": "CHARACTER REFERENCE — 염재보: the exact person appearing in this shot; match face, hair and build exactly.",
    "path": "<bytes:964351>",
    "asset_id": "ab287289-d0b7-436d-8f70-68452f898f73",
    "role": "character_ref"
   },
   {
    "label": "CHARACTER REFERENCE — 오주일: the exact person appearing in this shot; match face, hair and build exactly.",
    "path": "<bytes:809716>",
    "asset_id": "7c31e825-ceef-4208-915f-71c4f0e15dd6",
    "role": "character_ref"
   }
  ],
  "critique_skipped": true,
  "shot_run_produced": false,
  "shot_run_uid": "06aa1435-3979-7ee0-9535-a95941743fef",
  "ref_mode": "prev+엔티티",
  "share_plan": {
   "ref_plan": "prev",
   "prev_anchor_tag": "S45sh5"
  }
 },
 "S45sh20::cine": {
  "applied": true,
  "attempted_at": "2026-09-09T06:07:40.436764+00:00",
  "fingerprint": "ce6c72ffd6bbdddc64233fedfa60a6f24658130c9034bf3a6e568b2ea6212283",
  "fingerprint_version": 2,
  "provider": "mai",
  "endpoint": "openrouter/chat-completions",
  "model": "microsoft/mai-image-2.6",
  "pack": "24.202608252115",
  "source_file": "S45sh20_sel.png",
  "source_sha256": "e76ae44972f76402de49e34accbd5b16ec70b40b4a9feb533903e75fc2dbc22c",
  "file": "S45sh20_cine.png",
  "staged_sha256": "d68d274d33867510e0e0d38e75450c5033b0e63d6df3238ae4c57aa6d2b93f9e",
  "latency_ms": 28548
 },
 "S45sh33::signage": {
  "fp": "e2cd429504d4ee96",
  "inscriptions": [],
  "cues": [],
  "dropped": []
 },
 "S45sh33": {
  "input_fingerprint": "acb1ac342d16cde8",
  "prompt": "Create ONE FINAL photorealistic live-action film still of the moment below — early 20th-century Korea, Jeolla region; all people are Korean unless explicitly stated otherwise. TIME OF DAY (lock): night.\n\nSHOT TEXT (authoritative, Korean): 펴진 달래의 손바닥 위로 안규홍의 손끝이 반쪽 옥가락지를 살포시 맞닿게 한 순간.\n\nLOCATION (lock): At the straw mat on the ground in the mountain camp's nighttime mourning circle. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: close-up\n- KEY BACKGROUND ELEMENTS: 두 반쪽 옥가락지 (펴진 달래의 손바닥 안에서 서로 맞닿는 순간이다) — Both halves are visible from above at an oblique angle within the palm; used as Small central evidence of the reunion, scaled naturally against the fingers; 짚멍석 (달래의 몸 아래에 놓여 있다) — A limited portion of the upper surface surrounds the wrists; used as Preserves physical context at the frame margins.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Keep the established subdued nighttime illumination, with gentle tonal separation across skin and jade rather than an isolated decorative glint.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The clothing, hair and overall look of 달래, 안규홍 — who appear both in that photo and in this shot — are LOCKED to that photo. Anyone else visible in that photo is NOT in this shot: never carry their face, body or clothing onto anyone here. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nIMMOBILE CHARACTER POSE — CANONICAL (identical wherever this character appears in ANY panel; on any conflict THIS POSE WINS): Her dead body remains laid out on the straw mat while An Gyuhong holds one hand open, its palm exposed to receive the jade-ring pieces. Her fingers are open at this stage; the screenplay does not specify her head direction, the height of the held forearm, or the positions of her other arm and legs.\n\nIMMOBILE BODIES OBEY GRAVITY: a person who is dead or unconscious\nexerts NO muscular effort. Every part of their body — head, torso,\narms, hands, fingers, legs — rests fully on whatever supports it\n(floor, wall, furniture, their own lap) and hangs or slumps with\ngravity. NEVER show any part of an immobile person's body lifted,\nraised, held up in the air, or posed as if presenting something:\nan object in their grip stays clenched in a hand that itself lies\nfallen on a support — the hand does not hold the object up. If the\ncanonical pose leaves a body part unspecified, resolve it as the\nmost gravity-compliant, fully-supported position.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The dropped sickle lies on the ground beside the straw mat. 안규홍: He sits beside the mat with his necklace removed, his hands extended in the ring transfer; his neck is now bare. 달래: She remains dead on the mat, bloodied and dirt-stained; her fingers are open at this stage, with both halves of the jade ring resting on her exposed palm.\n\nTHE HAND THAT IS DOING THIS: the object at the centre of this shot is being held, operated, read, aimed or handed over by 안규홍 right now, so 안규홍's hand — and as much of the wrist and forearm as the framing reaches — is in the frame, gripping or resting on that object exactly the way its real use requires. Match that hand to 안규홍: its size, build, skin, age, grooming, sleeve and anything worn on it belong to that person and to no one else.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 안규홍 (한국인 남성, 청년 나이대, 검은 머리, 단단히 묶은 머리); 달래 (한국인 여성, 젊은 성인, 자연스러운 검은 머리) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.",
  "roll_prompts": {
   "A": "Create ONE FINAL photorealistic live-action film still of the moment below — early 20th-century Korea, Jeolla region; all people are Korean unless explicitly stated otherwise. TIME OF DAY (lock): night.\n\nSHOT TEXT (authoritative, Korean): 펴진 달래의 손바닥 위로 안규홍의 손끝이 반쪽 옥가락지를 살포시 맞닿게 한 순간.\n\nLOCATION (lock): At the straw mat on the ground in the mountain camp's nighttime mourning circle. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: close-up\n- KEY BACKGROUND ELEMENTS: 두 반쪽 옥가락지 (펴진 달래의 손바닥 안에서 서로 맞닿는 순간이다) — Both halves are visible from above at an oblique angle within the palm; used as Small central evidence of the reunion, scaled naturally against the fingers; 짚멍석 (달래의 몸 아래에 놓여 있다) — A limited portion of the upper surface surrounds the wrists; used as Preserves physical context at the frame margins.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Keep the established subdued nighttime illumination, with gentle tonal separation across skin and jade rather than an isolated decorative glint.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The clothing, hair and overall look of 달래, 안규홍 — who appear both in that photo and in this shot — are LOCKED to that photo. Anyone else visible in that photo is NOT in this shot: never carry their face, body or clothing onto anyone here. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nIMMOBILE CHARACTER POSE — CANONICAL (identical wherever this character appears in ANY panel; on any conflict THIS POSE WINS): Her dead body remains laid out on the straw mat while An Gyuhong holds one hand open, its palm exposed to receive the jade-ring pieces. Her fingers are open at this stage; the screenplay does not specify her head direction, the height of the held forearm, or the positions of her other arm and legs.\n\nIMMOBILE BODIES OBEY GRAVITY: a person who is dead or unconscious\nexerts NO muscular effort. Every part of their body — head, torso,\narms, hands, fingers, legs — rests fully on whatever supports it\n(floor, wall, furniture, their own lap) and hangs or slumps with\ngravity. NEVER show any part of an immobile person's body lifted,\nraised, held up in the air, or posed as if presenting something:\nan object in their grip stays clenched in a hand that itself lies\nfallen on a support — the hand does not hold the object up. If the\ncanonical pose leaves a body part unspecified, resolve it as the\nmost gravity-compliant, fully-supported position.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The dropped sickle lies on the ground beside the straw mat. 안규홍: He sits beside the mat with his necklace removed, his hands extended in the ring transfer; his neck is now bare. 달래: She remains dead on the mat, bloodied and dirt-stained; her fingers are open at this stage, with both halves of the jade ring resting on her exposed palm.\n\nTHE HAND THAT IS DOING THIS: the object at the centre of this shot is being held, operated, read, aimed or handed over by 안규홍 right now, so 안규홍's hand — and as much of the wrist and forearm as the framing reaches — is in the frame, gripping or resting on that object exactly the way its real use requires. Match that hand to 안규홍: its size, build, skin, age, grooming, sleeve and anything worn on it belong to that person and to no one else.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 안규홍 (한국인 남성, 청년 나이대, 검은 머리, 단단히 묶은 머리); 달래 (한국인 여성, 젊은 성인, 자연스러운 검은 머리) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.",
   "B": "Create ONE FINAL photorealistic live-action film still of the moment below — early 20th-century Korea, Jeolla region; all people are Korean unless explicitly stated otherwise. TIME OF DAY (lock): night.\n\nSHOT TEXT (authoritative, Korean): 펴진 달래의 손바닥 위로 안규홍의 손끝이 반쪽 옥가락지를 살포시 맞닿게 한 순간.\n\nLOCATION (lock): At the straw mat on the ground in the mountain camp's nighttime mourning circle. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: close-up\n- KEY BACKGROUND ELEMENTS: 두 반쪽 옥가락지 (펴진 달래의 손바닥 안에서 서로 맞닿는 순간이다) — Both halves are visible from above at an oblique angle within the palm; used as Small central evidence of the reunion, scaled naturally against the fingers; 짚멍석 (달래의 몸 아래에 놓여 있다) — A limited portion of the upper surface surrounds the wrists; used as Preserves physical context at the frame margins.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Keep the established subdued nighttime illumination, with gentle tonal separation across skin and jade rather than an isolated decorative glint.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The clothing, hair and overall look of 달래, 안규홍 — who appear both in that photo and in this shot — are LOCKED to that photo. Anyone else visible in that photo is NOT in this shot: never carry their face, body or clothing onto anyone here. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nIMMOBILE CHARACTER POSE — CANONICAL (identical wherever this character appears in ANY panel; on any conflict THIS POSE WINS): Her dead body remains laid out on the straw mat while An Gyuhong holds one hand open, its palm exposed to receive the jade-ring pieces. Her fingers are open at this stage; the screenplay does not specify her head direction, the height of the held forearm, or the positions of her other arm and legs.\n\nIMMOBILE BODIES OBEY GRAVITY: a person who is dead or unconscious\nexerts NO muscular effort. Every part of their body — head, torso,\narms, hands, fingers, legs — rests fully on whatever supports it\n(floor, wall, furniture, their own lap) and hangs or slumps with\ngravity. NEVER show any part of an immobile person's body lifted,\nraised, held up in the air, or posed as if presenting something:\nan object in their grip stays clenched in a hand that itself lies\nfallen on a support — the hand does not hold the object up. If the\ncanonical pose leaves a body part unspecified, resolve it as the\nmost gravity-compliant, fully-supported position.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The dropped sickle lies on the ground beside the straw mat. 안규홍: He sits beside the mat with his necklace removed, his hands extended in the ring transfer; his neck is now bare. 달래: She remains dead on the mat, bloodied and dirt-stained; her fingers are open at this stage, with both halves of the jade ring resting on her exposed palm.\n\nTHE HAND THAT IS DOING THIS: the object at the centre of this shot is being held, operated, read, aimed or handed over by 안규홍 right now, so 안규홍's hand — and as much of the wrist and forearm as the framing reaches — is in the frame, gripping or resting on that object exactly the way its real use requires. Match that hand to 안규홍: its size, build, skin, age, grooming, sleeve and anything worn on it belong to that person and to no one else.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 안규홍 (한국인 남성, 청년 나이대, 검은 머리, 단단히 묶은 머리); 달래 (한국인 여성, 젊은 성인, 자연스러운 검은 머리) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.\n\nCOMPOSITION VARIATION (this candidate only): keep the place, the people and the moment exactly as specified, and keep from the CAMERA & FRAME contract its subject scale, screen placement and key background elements exactly. For this candidate only, its camera angle and height are deliberately left open: choose a camera position distinctly different from the obvious one, as a film director picking a second setup on the same blocking. Stay true to any explicit framing the SHOT TEXT itself dictates."
  },
  "shot_run_spend_attempt_count": 1,
  "readings": [
   {
    "label": "A",
    "direction": "위쪽 손이 펴진 손바닥에 이미 놓인 옥가락지 반쪽 옆으로 나머지 반쪽을 내려놓으려 향하고 있다.",
    "built_space": "바닥에 짚멍석이 깔려 있으며 프레임 전체의 배경을 채우고 있다.",
    "entities": "펴진 손바닥과 팔목이 보이나, 소매가 달래의 옥색 저고리가 아닌 거친 삼베옷 재질로 잘못 묘사되었다. 위쪽 손의 팔목에는 소매가 보이지 않는다. 두 개의 옥가락지 반쪽이 있다.",
    "hard_violations": [
     "[gemini-pro] 달래의 손이어야 할 아래쪽 팔에 레퍼런스의 옥색 저고리가 아닌 안규홍의 의상(삼베옷)이 렌더링됨"
    ],
    "physics": "아래쪽 팔은 멍석 위에 뉘여 있고, 위쪽 손은 공중에서 가락지를 쥐고 내려오는 동작을 취하고 있다."
   },
   {
    "label": "B",
    "direction": "위쪽 손의 손끝이 펴진 손바닥 위에 놓인 두 개의 옥가락지 반쪽이 서로 맞닿는 지점을 정확히 향해 짚고 있다.",
    "built_space": "바닥에 짚멍석이 깔려 있으며 프레임 전반을 덮고 있고, 배경에는 멍석 위에 앉아 있는 사람의 하반신이 일부 보인다.",
    "entities": "아래쪽 펴진 손은 피가 묻은 옥색 소매를 입고 있어 달래와 정확히 일치하며, 위쪽 손과 배경의 몸은 흙 묻은 삼베옷을 입고 있어 안규홍과 일치한다. 옥가락지 두 반쪽이 중심에 명확히 묘사되었다.",
    "hard_violations": [],
    "physics": "아래쪽 팔과 배경의 앉은 몸은 멍석 위에 안정적으로 지탱되어 있으며, 위쪽 손은 중력을 거스르지 않고 가락지를 짚는 자연스러운 동작을 취하고 있다."
   }
  ],
  "cross_model_order": {
   "policy": "cross_model_order_v2_slot0_fwd_slot1_rev_parallel2",
   "slots": [
    {
     "model": "gemini-pro",
     "order": "forward",
     "display_to_canonical": {
      "A": "A",
      "B": "B"
     },
     "raw": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 10,
        "verdict_ko": "달래의 옥색 소매와 안규홍의 삼베옷을 정확히 구분하여 렌더링했으며, 두 반쪽 옥가락지가 손바닥 위에서 맞닿는 순간의 클로즈업 샷을 완벽하게 구현했습니다."
       },
       {
        "label": "A",
        "score": 3,
        "verdict_ko": "바닥에 놓인 달래의 손에 옥색 저고리가 아닌 안규홍의 삼베옷 소매가 입혀져 캐릭터 의상 레퍼런스를 심각하게 위반했습니다."
       }
      ],
      "readings": [
       {
        "label": "B",
        "direction": "위쪽 손의 손끝이 펴진 손바닥 위에 놓인 두 개의 옥가락지 반쪽이 서로 맞닿는 지점을 정확히 향해 짚고 있다.",
        "built_space": "바닥에 짚멍석이 깔려 있으며 프레임 전반을 덮고 있고, 배경에는 멍석 위에 앉아 있는 사람의 하반신이 일부 보인다.",
        "entities": "아래쪽 펴진 손은 피가 묻은 옥색 소매를 입고 있어 달래와 정확히 일치하며, 위쪽 손과 배경의 몸은 흙 묻은 삼베옷을 입고 있어 안규홍과 일치한다. 옥가락지 두 반쪽이 중심에 명확히 묘사되었다.",
        "hard_violations": [],
        "physics": "아래쪽 팔과 배경의 앉은 몸은 멍석 위에 안정적으로 지탱되어 있으며, 위쪽 손은 중력을 거스르지 않고 가락지를 짚는 자연스러운 동작을 취하고 있다."
       },
       {
        "label": "A",
        "direction": "위쪽 손이 펴진 손바닥에 이미 놓인 옥가락지 반쪽 옆으로 나머지 반쪽을 내려놓으려 향하고 있다.",
        "built_space": "바닥에 짚멍석이 깔려 있으며 프레임 전체의 배경을 채우고 있다.",
        "entities": "펴진 손바닥과 팔목이 보이나, 소매가 달래의 옥색 저고리가 아닌 거친 삼베옷 재질로 잘못 묘사되었다. 위쪽 손의 팔목에는 소매가 보이지 않는다. 두 개의 옥가락지 반쪽이 있다.",
        "hard_violations": [
         "달래의 손이어야 할 아래쪽 팔에 레퍼런스의 옥색 저고리가 아닌 안규홍의 의상(삼베옷)이 렌더링됨"
        ],
        "physics": "아래쪽 팔은 멍석 위에 뉘여 있고, 위쪽 손은 공중에서 가락지를 쥐고 내려오는 동작을 취하고 있다."
       }
      ],
      "all_candidates_fail": false
     },
     "normalized": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 10,
        "verdict_ko": "달래의 옥색 소매와 안규홍의 삼베옷을 정확히 구분하여 렌더링했으며, 두 반쪽 옥가락지가 손바닥 위에서 맞닿는 순간의 클로즈업 샷을 완벽하게 구현했습니다."
       },
       {
        "label": "A",
        "score": 3,
        "verdict_ko": "바닥에 놓인 달래의 손에 옥색 저고리가 아닌 안규홍의 삼베옷 소매가 입혀져 캐릭터 의상 레퍼런스를 심각하게 위반했습니다."
       }
      ],
      "all_candidates_fail": false,
      "readings": [
       {
        "label": "B",
        "direction": "위쪽 손의 손끝이 펴진 손바닥 위에 놓인 두 개의 옥가락지 반쪽이 서로 맞닿는 지점을 정확히 향해 짚고 있다.",
        "built_space": "바닥에 짚멍석이 깔려 있으며 프레임 전반을 덮고 있고, 배경에는 멍석 위에 앉아 있는 사람의 하반신이 일부 보인다.",
        "entities": "아래쪽 펴진 손은 피가 묻은 옥색 소매를 입고 있어 달래와 정확히 일치하며, 위쪽 손과 배경의 몸은 흙 묻은 삼베옷을 입고 있어 안규홍과 일치한다. 옥가락지 두 반쪽이 중심에 명확히 묘사되었다.",
        "hard_violations": [],
        "physics": "아래쪽 팔과 배경의 앉은 몸은 멍석 위에 안정적으로 지탱되어 있으며, 위쪽 손은 중력을 거스르지 않고 가락지를 짚는 자연스러운 동작을 취하고 있다."
       },
       {
        "label": "A",
        "direction": "위쪽 손이 펴진 손바닥에 이미 놓인 옥가락지 반쪽 옆으로 나머지 반쪽을 내려놓으려 향하고 있다.",
        "built_space": "바닥에 짚멍석이 깔려 있으며 프레임 전체의 배경을 채우고 있다.",
        "entities": "펴진 손바닥과 팔목이 보이나, 소매가 달래의 옥색 저고리가 아닌 거친 삼베옷 재질로 잘못 묘사되었다. 위쪽 손의 팔목에는 소매가 보이지 않는다. 두 개의 옥가락지 반쪽이 있다.",
        "hard_violations": [
         "달래의 손이어야 할 아래쪽 팔에 레퍼런스의 옥색 저고리가 아닌 안규홍의 의상(삼베옷)이 렌더링됨"
        ],
        "physics": "아래쪽 팔은 멍석 위에 뉘여 있고, 위쪽 손은 공중에서 가락지를 쥐고 내려오는 동작을 취하고 있다."
       }
      ]
     },
     "ok": true
    },
    {
     "model": "gpt-high",
     "order": "reverse",
     "display_to_canonical": {
      "A": "B",
      "B": "A"
     },
     "raw": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 7,
        "verdict_ko": "의상과 멍석의 연속성은 좋지만 손 주변과 무릎이 차지하는 면적이 크고, 두 옥 조각이 실제로 맞닿은 순간은 보이지 않는다."
       },
       {
        "label": "B",
        "score": 8,
        "verdict_ko": "손바닥과 손끝에 집중한 근접 구도와 비스듬한 상부 시점이 더 정확하지만, 두 반쪽 사이의 틈 때문에 재회의 접촉 순간은 미완성이다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "안규홍의 손이 오른쪽 위에서 달래의 손바닥 중앙으로 내려오며 엄지와 검지로 옥 조각 끝을 조작한다. 동작의 목표는 손바닥 위 다른 반쪽으로 맞지만, 위쪽 절단면 사이에도 가느다란 피부색 틈이 보이고 아래쪽 끝은 확실히 떨어져 있다. 얼굴과 시선은 화면 밖이다.",
        "built_space": "짚멍석 한 장의 윗면이 배경 대부분을 차지하고 왼쪽 위에 흙바닥이 조금 보인다. 위쪽에는 안규홍의 베이지색 무릎 부분이 들어온다. 달래의 손목은 오른쪽 아래에서 멍석 위로 이어진다. 장소의 재질은 참조와 맞지만, 요구한 손 중심의 제한된 주변부보다 멍석과 무릎의 비중이 크다. 고정 설비나 반사는 보이지 않는다.",
        "entities": "달래의 오염된 청록색 소매와 더 작고 가는 손, 안규홍의 거친 베이지색 소매와 더 큰 손이 보여 참조 의상 및 인물 관계와 대체로 맞는다. 손만으로 정확한 얼굴 정체성이나 한국인 여부는 확인할 수 없다. 연녹색 옥 반쪽 두 개가 손바닥 안에 자연스러운 크기로 놓였으나 절단면은 매우 매끈하다. 낫과 목걸이 상태는 이 구도 밖이며, 추가 인물이나 읽을 수 있는 글자는 없다.",
        "hard_violations": [],
        "physics": "달래의 손목과 손바닥 아랫면은 멍석에 놓여 있고 손가락은 약하게 굽어 있다. 옥 조각은 손바닥이 받치며 안규홍의 손끝도 접촉하고 있어 떠 있는 물체는 없다. 다만 달래의 손을 받쳐 펴 주는 안규홍의 다른 손은 보이지 않아, 명시된 보조 동작은 확인되지 않는다."
       },
       {
        "label": "B",
        "direction": "안규홍의 엄지와 검지가 오른쪽 위에서 손바닥 중앙의 위쪽 옥 반쪽을 다른 반쪽으로 가져간다. 손끝의 목표는 정확하지만 양쪽 파단면 사이에 틈이 남아 두 조각이 맞닿기 직전으로 읽힌다. 달래의 손바닥은 위로 노출되고 손가락은 왼쪽 위로 펼쳐져 있다. 시선은 화면 밖이다.",
        "built_space": "짚멍석 한 장의 윗면이 손 주변을 둘러싸고 왼쪽 위 가장자리 너머로 흙이 조금 보인다. 달래의 손목은 오른쪽 아래, 안규홍의 손목은 오른쪽 위에 배치된다. 무릎이나 몸통 없이 두 손을 크게 잡은 비스듬한 상부 근접 구도여서 지정된 프레이밍에 더 가깝다. 중복 설비나 반사는 없다.",
        "entities": "달래의 손목과 청록색 소매에 피와 때가 있고, 안규홍의 더 큰 손이 옥을 조작한다. 안규홍의 소매는 끝부분만 보여 참조의 베이지색 의상과의 일치 여부가 A보다 덜 분명하다. 얼굴과 머리, 정확한 민족적 외형은 확인할 수 없다. 연녹색 옥 반쪽 두 개의 불규칙한 파단면과 입체적인 표면이 보이며 손가락 대비 크기도 자연스럽다. 추가 인물과 글자는 없고 낫 및 목 부분은 구도 밖이다.",
        "hard_violations": [],
        "physics": "달래의 손목과 손바닥 아래가 멍석에 닿아 받쳐지고, 펼쳐진 손가락도 멍석 가까이에 이어진다. 시신의 팔 전체가 공중에 들린 배치는 아니다. 아래쪽 옥 조각은 손바닥 위에 놓이고 위쪽 조각은 손바닥과 안규홍의 엄지·검지가 지지한다. 다만 안규홍이 달래의 손을 직접 받쳐 펴 주는 다른 손은 보이지 않는다."
       }
      ],
      "all_candidates_fail": true
     },
     "normalized": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 7,
        "verdict_ko": "의상과 멍석의 연속성은 좋지만 손 주변과 무릎이 차지하는 면적이 크고, 두 옥 조각이 실제로 맞닿은 순간은 보이지 않는다."
       },
       {
        "label": "A",
        "score": 8,
        "verdict_ko": "손바닥과 손끝에 집중한 근접 구도와 비스듬한 상부 시점이 더 정확하지만, 두 반쪽 사이의 틈 때문에 재회의 접촉 순간은 미완성이다."
       }
      ],
      "all_candidates_fail": true,
      "readings": [
       {
        "label": "B",
        "direction": "안규홍의 손이 오른쪽 위에서 달래의 손바닥 중앙으로 내려오며 엄지와 검지로 옥 조각 끝을 조작한다. 동작의 목표는 손바닥 위 다른 반쪽으로 맞지만, 위쪽 절단면 사이에도 가느다란 피부색 틈이 보이고 아래쪽 끝은 확실히 떨어져 있다. 얼굴과 시선은 화면 밖이다.",
        "built_space": "짚멍석 한 장의 윗면이 배경 대부분을 차지하고 왼쪽 위에 흙바닥이 조금 보인다. 위쪽에는 안규홍의 베이지색 무릎 부분이 들어온다. 달래의 손목은 오른쪽 아래에서 멍석 위로 이어진다. 장소의 재질은 참조와 맞지만, 요구한 손 중심의 제한된 주변부보다 멍석과 무릎의 비중이 크다. 고정 설비나 반사는 보이지 않는다.",
        "entities": "달래의 오염된 청록색 소매와 더 작고 가는 손, 안규홍의 거친 베이지색 소매와 더 큰 손이 보여 참조 의상 및 인물 관계와 대체로 맞는다. 손만으로 정확한 얼굴 정체성이나 한국인 여부는 확인할 수 없다. 연녹색 옥 반쪽 두 개가 손바닥 안에 자연스러운 크기로 놓였으나 절단면은 매우 매끈하다. 낫과 목걸이 상태는 이 구도 밖이며, 추가 인물이나 읽을 수 있는 글자는 없다.",
        "hard_violations": [],
        "physics": "달래의 손목과 손바닥 아랫면은 멍석에 놓여 있고 손가락은 약하게 굽어 있다. 옥 조각은 손바닥이 받치며 안규홍의 손끝도 접촉하고 있어 떠 있는 물체는 없다. 다만 달래의 손을 받쳐 펴 주는 안규홍의 다른 손은 보이지 않아, 명시된 보조 동작은 확인되지 않는다."
       },
       {
        "label": "A",
        "direction": "안규홍의 엄지와 검지가 오른쪽 위에서 손바닥 중앙의 위쪽 옥 반쪽을 다른 반쪽으로 가져간다. 손끝의 목표는 정확하지만 양쪽 파단면 사이에 틈이 남아 두 조각이 맞닿기 직전으로 읽힌다. 달래의 손바닥은 위로 노출되고 손가락은 왼쪽 위로 펼쳐져 있다. 시선은 화면 밖이다.",
        "built_space": "짚멍석 한 장의 윗면이 손 주변을 둘러싸고 왼쪽 위 가장자리 너머로 흙이 조금 보인다. 달래의 손목은 오른쪽 아래, 안규홍의 손목은 오른쪽 위에 배치된다. 무릎이나 몸통 없이 두 손을 크게 잡은 비스듬한 상부 근접 구도여서 지정된 프레이밍에 더 가깝다. 중복 설비나 반사는 없다.",
        "entities": "달래의 손목과 청록색 소매에 피와 때가 있고, 안규홍의 더 큰 손이 옥을 조작한다. 안규홍의 소매는 끝부분만 보여 참조의 베이지색 의상과의 일치 여부가 A보다 덜 분명하다. 얼굴과 머리, 정확한 민족적 외형은 확인할 수 없다. 연녹색 옥 반쪽 두 개의 불규칙한 파단면과 입체적인 표면이 보이며 손가락 대비 크기도 자연스럽다. 추가 인물과 글자는 없고 낫 및 목 부분은 구도 밖이다.",
        "hard_violations": [],
        "physics": "달래의 손목과 손바닥 아래가 멍석에 닿아 받쳐지고, 펼쳐진 손가락도 멍석 가까이에 이어진다. 시신의 팔 전체가 공중에 들린 배치는 아니다. 아래쪽 옥 조각은 손바닥 위에 놓이고 위쪽 조각은 손바닥과 안규홍의 엄지·검지가 지지한다. 다만 안규홍이 달래의 손을 직접 받쳐 펴 주는 다른 손은 보이지 않는다."
       }
      ]
     },
     "ok": true
    }
   ],
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "slot_winner_match": false,
   "slot_winner": {
    "gemini-pro": "B",
    "gpt-high": "A"
   },
   "route": "cross_slot_combined"
  },
  "dual": {
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "normalized": {
    "A": 1.3,
    "B": 1.875
   },
   "adjusted": {
    "A": 1.05,
    "B": 1.875
   },
   "violations": {
    "A": [
     "[gemini-pro] 달래의 손이어야 할 아래쪽 팔에 레퍼런스의 옥색 저고리가 아닌 안규홍의 의상(삼베옷)이 렌더링됨"
    ]
   },
   "per_model_winner": {
    "gemini-pro": "B",
    "gpt-high": "A"
   },
   "agreed": false
  },
  "totals": {
   "B": 1875,
   "A": 1050
  },
  "selected": "B",
  "ranking": [
   "B",
   "A"
  ],
  "verdicts": [
   {
    "label": "B",
    "score": 1875,
    "verdict_ko": "달래의 옥색 소매와 안규홍의 삼베옷을 정확히 구분하여 렌더링했으며, 두 반쪽 옥가락지가 손바닥 위에서 맞닿는 순간의 클로즈업 샷을 완벽하게 구현했습니다."
   },
   {
    "label": "A",
    "score": 1050,
    "verdict_ko": "바닥에 놓인 달래의 손에 옥색 저고리가 아닌 안규홍의 삼베옷 소매가 입혀져 캐릭터 의상 레퍼런스를 심각하게 위반했습니다.  ★위반: [gemini-pro] 달래의 손이어야 할 아래쪽 팔에 레퍼런스의 옥색 저고리가 아닌 안규홍의 의상(삼베옷)이 렌더링됨"
   }
  ],
  "refs": [
   {
    "label": "PREVIOUS SHOT STILL — a visually related earlier shot of this same place: the location's look, materials, fixed features and lighting mood are LOCKED to this photo; never copy its camera framing. The clothing, hair and overall look of 달래, 안규홍 — who appear both in that photo and in this shot — are LOCKED to that photo. Anyone else visible in that photo is NOT in this shot: never carry their face, body or clothing onto anyone here. If this photo shows a character who cannot move (dead or unconscious), that character's exact pose, position and orientation are ALSO LOCKED — treat their body as a fixed prop of the set that only the camera moves around.",
    "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/db00517b-165b-4005-a287-f28936b59b95/images/272bd7af-9294-4490-98c6-88a40e9deacc/scene/recipe/S45sh5_sel.png",
    "asset_id": "9a25d7f4-fa8e-4b50-87cd-708edf72baa5",
    "role": "prev_still"
   },
   {
    "label": "CHARACTER REFERENCE — 안규홍: the exact person appearing in this shot; match face, hair and build exactly.",
    "path": "<bytes:1028384>",
    "asset_id": "3b5894db-ce55-4f34-b966-5e725c3aa952",
    "role": "character_ref"
   },
   {
    "label": "CHARACTER REFERENCE — 달래: the exact person appearing in this shot; match face, hair and build exactly.",
    "path": "<bytes:959810>",
    "asset_id": "a51dee5b-8b43-4345-b827-da2ff15b545d",
    "role": "character_ref"
   }
  ],
  "critique_skipped": true,
  "shot_run_produced": false,
  "shot_run_uid": "06aa1435-3b2e-7b2d-a8cc-7d129adb1301",
  "ref_mode": "prev+엔티티",
  "share_plan": {
   "ref_plan": "prev",
   "prev_anchor_tag": "S45sh5"
  }
 },
 "S45sh33::cine": {
  "applied": true,
  "attempted_at": "2026-09-09T06:09:01.477746+00:00",
  "fingerprint": "dc37b5ceaf95fbb85411cde5ccc7120d1cabf47d4da7fedf7107cfe6dc9e20c4",
  "fingerprint_version": 2,
  "provider": "mai",
  "endpoint": "openrouter/chat-completions",
  "model": "microsoft/mai-image-2.6",
  "pack": "24.202608252115",
  "source_file": "S45sh33_sel.png",
  "source_sha256": "35901ff53f8ada238992568590c7fdcd46b05c8169655f6d4cc81073ff6464a5",
  "file": "S45sh33_cine.png",
  "staged_sha256": "4b9f7f940740e03e5f1842e4986a6dd4ce8bb8f99ccc9f4e273672222549350a",
  "latency_ms": 28413
 },
 "S46sh3::signage": {
  "fp": "e83fef2c6a4e3a3a",
  "inscriptions": [],
  "cues": [],
  "dropped": []
 },
 "S46sh3::bgfirst_bg": {
  "input_fingerprint": "3c37600905f87e7b",
  "prompt": "Create the EMPTY BACKGROUND PLATE for one film shot — NO PEOPLE, no figures, no body parts, no sketch lines, no arrows anywhere.\n\"Empty\" means no people only: KEEP the location's inherent occupants and stock that define the place — animals in an animal shelter, pen or farm, goods in a market, moored boats in a harbour — unless the shot text explicitly removes them.\nThe FIRST attached image is a thin-line storyboard sketch: use ONLY its camera angle, horizon, perspective and the placement/size of buildings and set masses — ignore the sketched people and arrows entirely. The SECOND attached image (LOCATION PHOTOGRAPH) is the real place: take its architecture, materials, signage and fixed features, and RE-PROJECT them into the sketch's camera. If the photograph's camera differs from the sketch's, the sketch's camera wins.\nHUMAN-SCALE CALIBRATION: derive every structure's true size from human-scale elements — a door ≈ 2m, a window ≈ 1–1.5m wide, one storey ≈ 2.5–3m; never inflate a small structure or shrink a large one.\n\nSHOT TEXT this background must serve (Korean): 산채 한가운데 둥글게 둘러앉은 백정, 목수, 염재보, 임정현, 오주일, 복만, 이관회를 훑어보는 안규홍의 정면.\n\nLOCATION (lock): The central outdoor assembly circle of the mountain camp, with morning sunlight slanting through cypress trees.\n\nTIME OF DAY (lock): morning, slanting sunlight.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: medium shot\n- KEY BACKGROUND ELEMENTS: 편백 (산채 둘레에 서 있다) — Partial trunks occupy the distant margins behind the seated figure; used as Provides restrained vertical context beyond the intimate shoulder framing.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Oblique morning sunlight through the cypresses reveals 안규홍's swollen eyes and dried facial earth in a muted, naturally contrasted palette.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nRender ONE photorealistic empty location photograph, 16:9, that this shot can be staged inside later. No readable writing anywhere: surfaces that would carry writing may be present, but stage any wording out of legibility — an oblique angle, distance, shallow focus. No captions, watermarks or overlay text.",
  "effective_prompt": "Create the EMPTY BACKGROUND PLATE for one film shot — NO PEOPLE, no figures, no body parts, no sketch lines, no arrows anywhere.\n\"Empty\" means no people only: KEEP the location's inherent occupants and stock that define the place — animals in an animal shelter, pen or farm, goods in a market, moored boats in a harbour — unless the shot text explicitly removes them.\nThe FIRST attached image is a thin-line storyboard sketch: use ONLY its camera angle, horizon, perspective and the placement/size of buildings and set masses — ignore the sketched people and arrows entirely. The SECOND attached image (LOCATION PHOTOGRAPH) is the real place: take its architecture, materials, signage and fixed features, and RE-PROJECT them into the sketch's camera. If the photograph's camera differs from the sketch's, the sketch's camera wins.\nHUMAN-SCALE CALIBRATION: derive every structure's true size from human-scale elements — a door ≈ 2m, a window ≈ 1–1.5m wide, one storey ≈ 2.5–3m; never inflate a small structure or shrink a large one.\n\nSHOT TEXT this background must serve (Korean): 산채 한가운데 둥글게 둘러앉은 백정, 목수, 염재보, 임정현, 오주일, 복만, 이관회를 훑어보는 안규홍의 정면.\n\nLOCATION (lock): The central outdoor assembly circle of the mountain camp, with morning sunlight slanting through cypress trees.\n\nTIME OF DAY (lock): morning, slanting sunlight.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: medium shot\n- KEY BACKGROUND ELEMENTS: 편백 (산채 둘레에 서 있다) — Partial trunks occupy the distant margins behind the seated figure; used as Provides restrained vertical context beyond the intimate shoulder framing.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Oblique morning sunlight through the cypresses reveals 안규홍's swollen eyes and dried facial earth in a muted, naturally contrasted palette.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nRender ONE photorealistic empty location photograph, 16:9, that this shot can be staged inside later. No readable writing anywhere: surfaces that would carry writing may be present, but stage any wording out of legibility — an oblique angle, distance, shallow focus. No captions, watermarks or overlay text.",
  "bg_path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/db00517b-165b-4005-a287-f28936b59b95/images/272bd7af-9294-4490-98c6-88a40e9deacc/scene/recipe/S46sh3__bgfirst_bg.png",
  "asset_id": "f0c0037a-64d8-4768-bc16-70d7921aacb0",
  "input_asset_ids": [
   "27ce28a9-6143-468b-9182-2c27d802d80d",
   "c5df4442-a576-4bdd-9e42-441cb4544a5e"
  ]
 },
 "S46sh3": {
  "input_fingerprint": "9f6a1ab4ac891f15",
  "prompt": "Create ONE FINAL photorealistic live-action film still of the moment below — early 20th-century Korea, Jeolla region; all people are Korean unless explicitly stated otherwise. TIME OF DAY (lock): morning, slanting sunlight.\n\nSHOT TEXT (authoritative, Korean): 산채 한가운데 둥글게 둘러앉은 백정, 목수, 염재보, 임정현, 오주일, 복만, 이관회를 훑어보는 안규홍의 정면.\n\nLOCATION (lock): The central outdoor assembly circle of the mountain camp, with morning sunlight slanting through cypress trees. The shot takes place here — the attached LOCATION PHOTOGRAPH shows the exact spot.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: medium shot\n- KEY BACKGROUND ELEMENTS: 편백 (산채 둘레에 서 있다) — Partial trunks occupy the distant margins behind the seated figure; used as Provides restrained vertical context beyond the intimate shoulder framing.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Oblique morning sunlight through the cypresses reveals 안규홍's swollen eyes and dried facial earth in a muted, naturally contrasted palette.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): Morning sunlight slants through the cypress trees into the camp. 안규홍: His eyes are swollen, dried dirt remains on his unwashed face, his hair is firmly retied and his neck is bare.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 안규홍 (한국인 남성, 청년 나이대, 검은 머리, 단단히 묶은 머리) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.",
  "roll_prompts": {
   "A": "Stage the shot. The FIRST attached image (SHOT BACKGROUND) is the finished empty background of this shot — keep it EXACTLY: its camera, perspective, architecture, lighting and every fixed feature stay untouched. The SECOND attached image (LAYOUT SKETCH) tells you ONLY where the people go: each sketched person's position, screen size, pose and the gaze/motion arrows. Ignore the sketch's background lines. The CHARACTER REFERENCE photographs show the real people.\nPlace the real people into the background at exactly the sketched positions, sizes and poses, following the arrow directions. No sketch lines or arrows may remain.\n\nCreate ONE FINAL photorealistic live-action film still of the moment below — early 20th-century Korea, Jeolla region; all people are Korean unless explicitly stated otherwise. TIME OF DAY (lock): morning, slanting sunlight.\n\nSHOT TEXT (authoritative, Korean): 산채 한가운데 둥글게 둘러앉은 백정, 목수, 염재보, 임정현, 오주일, 복만, 이관회를 훑어보는 안규홍의 정면.\n\nLOCATION (lock): The central outdoor assembly circle of the mountain camp, with morning sunlight slanting through cypress trees. The shot takes place here — the FIRST attached image (SHOT BACKGROUND) is this exact place, already built: its ground, structures, horizon, materials and lighting are the finished truth of this location and must not be redesigned or replaced. No location photograph is attached — read the place from that image alone, and add no scenery, structure, vehicle or fixture that it does not already show. This lock governs the place only; the figures in the shot follow the staging and pose instructions.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: medium shot\n- KEY BACKGROUND ELEMENTS: 편백 (산채 둘레에 서 있다) — Partial trunks occupy the distant margins behind the seated figure; used as Provides restrained vertical context beyond the intimate shoulder framing.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Oblique morning sunlight through the cypresses reveals 안규홍's swollen eyes and dried facial earth in a muted, naturally contrasted palette.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): Morning sunlight slants through the cypress trees into the camp. 안규홍: His eyes are swollen, dried dirt remains on his unwashed face, his hair is firmly retied and his neck is bare.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 안규홍 (한국인 남성, 청년 나이대, 검은 머리, 단단히 묶은 머리) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.",
   "B": "Create ONE FINAL photorealistic live-action film still of the moment below — early 20th-century Korea, Jeolla region; all people are Korean unless explicitly stated otherwise. TIME OF DAY (lock): morning, slanting sunlight.\n\nSHOT TEXT (authoritative, Korean): 산채 한가운데 둥글게 둘러앉은 백정, 목수, 염재보, 임정현, 오주일, 복만, 이관회를 훑어보는 안규홍의 정면.\n\nLOCATION (lock): The central outdoor assembly circle of the mountain camp, with morning sunlight slanting through cypress trees. The shot takes place here — the attached LOCATION PHOTOGRAPH shows the exact spot.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: medium shot\n- KEY BACKGROUND ELEMENTS: 편백 (산채 둘레에 서 있다) — Partial trunks occupy the distant margins behind the seated figure; used as Provides restrained vertical context beyond the intimate shoulder framing.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Oblique morning sunlight through the cypresses reveals 안규홍's swollen eyes and dried facial earth in a muted, naturally contrasted palette.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): Morning sunlight slants through the cypress trees into the camp. 안규홍: His eyes are swollen, dried dirt remains on his unwashed face, his hair is firmly retied and his neck is bare.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 안규홍 (한국인 남성, 청년 나이대, 검은 머리, 단단히 묶은 머리) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.\n\nCOMPOSITION VARIATION (this candidate only): keep the place, the people and the moment exactly as specified, and keep from the CAMERA & FRAME contract its subject scale, screen placement and key background elements exactly. For this candidate only, its camera angle and height are deliberately left open: choose a camera position distinctly different from the obvious one, as a film director picking a second setup on the same blocking. Stay true to any explicit framing the SHOT TEXT itself dictates."
  },
  "roll_refs": {
   "A": [
    {
     "label": "SHOT BACKGROUND",
     "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/db00517b-165b-4005-a287-f28936b59b95/images/272bd7af-9294-4490-98c6-88a40e9deacc/scene/recipe/S46sh3__bgfirst_bg.png",
     "asset_id": "f0c0037a-64d8-4768-bc16-70d7921aacb0",
     "role": "bgfirst_bg"
    },
    {
     "label": "LAYOUT SKETCH (people placement only)",
     "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/db00517b-165b-4005-a287-f28936b59b95/images/272bd7af-9294-4490-98c6-88a40e9deacc/conti/conti_S46sh3.png",
     "asset_id": "27ce28a9-6143-468b-9182-2c27d802d80d",
     "role": "conti_light"
    },
    {
     "label": "CHARACTER REFERENCE — 안규홍: the exact person appearing in this shot; match face, hair and build exactly.",
     "path": "<bytes:1028384>",
     "asset_id": "3b5894db-ce55-4f34-b966-5e725c3aa952",
     "role": "character_ref"
    }
   ],
   "B": [
    {
     "label": "LOCATION PHOTOGRAPH — the exact place of this shot: its architecture, materials, fixed features and lighting mood are spatial truth; stage the moment inside this place. Never copy its camera framing.",
     "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/db00517b-165b-4005-a287-f28936b59b95/images/272bd7af-9294-4490-98c6-88a40e9deacc/scene/recipe/groupbg_mountain_camp_clearing_cb9567.png",
     "asset_id": "c5df4442-a576-4bdd-9e42-441cb4544a5e",
     "role": "bgfirst_group_bg"
    },
    {
     "label": "CHARACTER REFERENCE — 안규홍: the exact person appearing in this shot; match face, hair and build exactly.",
     "path": "<bytes:1028384>",
     "asset_id": "3b5894db-ce55-4f34-b966-5e725c3aa952",
     "role": "character_ref"
    }
   ]
  },
  "shot_run_spend_attempt_count": 1,
  "judge_flip_skipped": "judge_fn_owns_order",
  "readings": [
   {
    "label": "A",
    "direction": "안규홍은 카메라 정면을 향해 앉은 채로 고개를 오른쪽으로 돌려 프레임 밖의 인물들을 훑어보고 있으며, 전경의 인물은 안규홍을 향해 위치해 있음.",
    "built_space": "로케이션 사진의 넓은 흙마당, 뒤쪽의 목책, 좌우 초가 지붕 구조를 충실히 재현함. 원경에 나무들이 적절히 배치되었으며, 피사체와 공간의 스케일감이 일치함.",
    "entities": "안규홍은 레퍼런스와 일치하는 이목구비에 흙 묻은 피부, 부은 눈, 묶은 머리를 하고 있음. 텍스트에 언급된 무리 중 1명이 전경에 뒷모습으로 등장함.",
    "hard_violations": [],
    "physics": "안규홍과 전경의 남성은 흙바닥에 자연스럽고 안정적으로 앉아 있으며, 신체를 지탱하는 물리적 구조에 어색함이나 오류가 없음."
   },
   {
    "label": "B",
    "direction": "안규홍은 정면을 응시하며 둥글게 앉은 주변 인물들을 향해 시선을 두고 있으며, 주변의 남성들은 안규홍을 향해 원형으로 앉아 있음.",
    "built_space": "마당 중앙은 빈 공터여야 한다는 로케이션 래퍼런스를 무시하고, 마당 한가운데 인물들 사이로 다수의 거대한 편백나무들이 솟아 있어 공간 구조를 심각하게 왜곡함.",
    "entities": "안규홍을 중심으로 다수의 남성들이 둥글게 둘러앉아 있음. 안규홍의 얼굴과 흙 묻은 분장 등은 레퍼런스와 대체로 일치함.",
    "hard_violations": [
     "[gemini-pro] 로케이션 레퍼런스 상 완전히 빈 공터인 산채 마당 중앙에 커다란 편백나무 다수를 임의로 생성하여 배치함 (Invented objects 및 공간 구조 왜곡)"
    ],
    "physics": "인물들은 바닥에 무리 없이 앉아 있으나, 거대한 나무들이 주변 구조물과 어울리지 않게 흙바닥 중앙에서 갑작스럽게 솟아나 있어 물리적인 공간감이 부자연스러움."
   }
  ],
  "cross_model_order": {
   "policy": "cross_model_order_v2_slot0_fwd_slot1_rev_parallel2",
   "slots": [
    {
     "model": "gemini-pro",
     "order": "forward",
     "display_to_canonical": {
      "A": "A",
      "B": "B"
     },
     "raw": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 9,
        "verdict_ko": "지정된 미디엄 샷과 숄더 프레이밍의 지시를 정확하게 구현하여 최우선 순위인 촬영 구도를 완벽하게 충족했으며, 안규홍의 외양과 로케이션 구조도 훌륭하게 반영했습니다."
       },
       {
        "label": "B",
        "score": 3,
        "verdict_ko": "미디엄 샷 프레이밍을 무시하고 전체 인물이 보이도록 와이드 샷으로 임의 변경하였으며, 빈 공터여야 할 마당 중앙에 나무들을 빽빽하게 배치하여 로케이션 공간 제약을 크게 위반했습니다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "안규홍은 카메라 정면을 향해 앉은 채로 고개를 오른쪽으로 돌려 프레임 밖의 인물들을 훑어보고 있으며, 전경의 인물은 안규홍을 향해 위치해 있음.",
        "built_space": "로케이션 사진의 넓은 흙마당, 뒤쪽의 목책, 좌우 초가 지붕 구조를 충실히 재현함. 원경에 나무들이 적절히 배치되었으며, 피사체와 공간의 스케일감이 일치함.",
        "entities": "안규홍은 레퍼런스와 일치하는 이목구비에 흙 묻은 피부, 부은 눈, 묶은 머리를 하고 있음. 텍스트에 언급된 무리 중 1명이 전경에 뒷모습으로 등장함.",
        "hard_violations": [],
        "physics": "안규홍과 전경의 남성은 흙바닥에 자연스럽고 안정적으로 앉아 있으며, 신체를 지탱하는 물리적 구조에 어색함이나 오류가 없음."
       },
       {
        "label": "B",
        "direction": "안규홍은 정면을 응시하며 둥글게 앉은 주변 인물들을 향해 시선을 두고 있으며, 주변의 남성들은 안규홍을 향해 원형으로 앉아 있음.",
        "built_space": "마당 중앙은 빈 공터여야 한다는 로케이션 래퍼런스를 무시하고, 마당 한가운데 인물들 사이로 다수의 거대한 편백나무들이 솟아 있어 공간 구조를 심각하게 왜곡함.",
        "entities": "안규홍을 중심으로 다수의 남성들이 둥글게 둘러앉아 있음. 안규홍의 얼굴과 흙 묻은 분장 등은 레퍼런스와 대체로 일치함.",
        "hard_violations": [
         "로케이션 레퍼런스 상 완전히 빈 공터인 산채 마당 중앙에 커다란 편백나무 다수를 임의로 생성하여 배치함 (Invented objects 및 공간 구조 왜곡)"
        ],
        "physics": "인물들은 바닥에 무리 없이 앉아 있으나, 거대한 나무들이 주변 구조물과 어울리지 않게 흙바닥 중앙에서 갑작스럽게 솟아나 있어 물리적인 공간감이 부자연스러움."
       }
      ],
      "all_candidates_fail": false
     },
     "normalized": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 9,
        "verdict_ko": "지정된 미디엄 샷과 숄더 프레이밍의 지시를 정확하게 구현하여 최우선 순위인 촬영 구도를 완벽하게 충족했으며, 안규홍의 외양과 로케이션 구조도 훌륭하게 반영했습니다."
       },
       {
        "label": "B",
        "score": 3,
        "verdict_ko": "미디엄 샷 프레이밍을 무시하고 전체 인물이 보이도록 와이드 샷으로 임의 변경하였으며, 빈 공터여야 할 마당 중앙에 나무들을 빽빽하게 배치하여 로케이션 공간 제약을 크게 위반했습니다."
       }
      ],
      "all_candidates_fail": false,
      "readings": [
       {
        "label": "A",
        "direction": "안규홍은 카메라 정면을 향해 앉은 채로 고개를 오른쪽으로 돌려 프레임 밖의 인물들을 훑어보고 있으며, 전경의 인물은 안규홍을 향해 위치해 있음.",
        "built_space": "로케이션 사진의 넓은 흙마당, 뒤쪽의 목책, 좌우 초가 지붕 구조를 충실히 재현함. 원경에 나무들이 적절히 배치되었으며, 피사체와 공간의 스케일감이 일치함.",
        "entities": "안규홍은 레퍼런스와 일치하는 이목구비에 흙 묻은 피부, 부은 눈, 묶은 머리를 하고 있음. 텍스트에 언급된 무리 중 1명이 전경에 뒷모습으로 등장함.",
        "hard_violations": [],
        "physics": "안규홍과 전경의 남성은 흙바닥에 자연스럽고 안정적으로 앉아 있으며, 신체를 지탱하는 물리적 구조에 어색함이나 오류가 없음."
       },
       {
        "label": "B",
        "direction": "안규홍은 정면을 응시하며 둥글게 앉은 주변 인물들을 향해 시선을 두고 있으며, 주변의 남성들은 안규홍을 향해 원형으로 앉아 있음.",
        "built_space": "마당 중앙은 빈 공터여야 한다는 로케이션 래퍼런스를 무시하고, 마당 한가운데 인물들 사이로 다수의 거대한 편백나무들이 솟아 있어 공간 구조를 심각하게 왜곡함.",
        "entities": "안규홍을 중심으로 다수의 남성들이 둥글게 둘러앉아 있음. 안규홍의 얼굴과 흙 묻은 분장 등은 레퍼런스와 대체로 일치함.",
        "hard_violations": [
         "로케이션 레퍼런스 상 완전히 빈 공터인 산채 마당 중앙에 커다란 편백나무 다수를 임의로 생성하여 배치함 (Invented objects 및 공간 구조 왜곡)"
        ],
        "physics": "인물들은 바닥에 무리 없이 앉아 있으나, 거대한 나무들이 주변 구조물과 어울리지 않게 흙바닥 중앙에서 갑작스럽게 솟아나 있어 물리적인 공간감이 부자연스러움."
       }
      ]
     },
     "ok": true
    },
    {
     "model": "gpt-high",
     "order": "reverse",
     "display_to_canonical": {
      "A": "B",
      "B": "A"
     },
     "raw": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 4,
        "verdict_ko": "안규홍의 정면과 원형 모임은 보이지만, 집단을 넓게 담아 지정된 미디엄 숏을 벗어나고 굵은 나무들이 모임 사이를 차지한다."
       },
       {
        "label": "B",
        "score": 6,
        "verdict_ko": "인물 중심 미디엄 숏과 옆 사람들을 살피는 시선, 부은 눈과 얼굴의 흙은 더 충실하지만, 정면 대신 측면에 가까운 얼굴과 가까운 나무 배치는 어긋난다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "중앙 안규홍은 몸과 얼굴을 카메라 쪽으로 두고 거의 정면을 본다. 시선은 렌즈 또는 바로 앞에 등을 보인 남성 쪽이며, 여러 사람을 훑어보는 순간은 뚜렷하지 않다. 주변 남성들은 대체로 원 안쪽을 향한다. 겨냥하거나 움직이는 소품은 없다.",
        "built_space": "좌우 초가 각 한 채, 뒤쪽 목책 한 줄과 오른쪽 망루 한 기가 보여 참조 산채의 주요 구조는 대응한다. 그러나 굵은 나무줄기 약 일곱 개가 전경과 모임 사이까지 들어와, 먼 가장자리의 부분적인 편백 줄기라는 지시와 다르다. 중앙 인물보다 원형 집단과 장소가 크게 강조된 넓은 구도다.",
        "entities": "중앙 청년 남성과 주변 남성 일곱 명으로 총 여덟 명이 보이며, 인원수는 숏 문장의 안규홍과 일곱 동료에 대응한다. 동료들의 개별 신원은 외형만으로 구분할 수 없다. 중앙 인물은 참조와 유사한 얼굴, 뒤로 단단히 정리한 검은 머리, 때 묻은 밝은 전통 의복과 맨목을 갖췄다. 얼굴 흙은 보이지만 눈의 부기는 비교적 약하다. 읽을 수 있는 글자는 없다.",
        "hard_violations": [],
        "physics": "좌우 남성들의 다리는 굽혀져 있고 보이는 발은 지면에 닿는다. 손은 무릎이나 허벅지에 놓여 있다. 중앙과 전경 인물들의 엉덩이 및 하체 지지점은 다른 몸이나 화면 경계에 가려져 있어 앉은 상태를 완전히 확인할 수 없지만, 공중에 떠 있다고 볼 근거도 없다. 항아리와 그릇은 지면이나 받침 위에 놓여 있다."
       },
       {
        "label": "B",
        "direction": "안규홍의 몸은 카메라를 향하지만 얼굴과 눈은 화면 왼쪽으로 돌아가 있다. 시선은 왼쪽 전경 남성의 머리보다 위쪽을 지나 화면 밖의 원 구성원 쪽으로 향하는 것으로 읽힌다. 주변을 살피는 동작에는 맞지만, 요구한 정면 얼굴은 아니다. 전경 남성은 안규홍 쪽으로 고개를 향한다.",
        "built_space": "좌우 초가 각 한 채, 뒤 목책 한 줄, 오른쪽 망루 한 기와 흙마당이 참조 장소에 대응한다. 좌우에 큰 나무줄기 네 개가 두드러지고 뒤로 가는 줄기들이 이어지는데, 특히 오른쪽 나무들은 먼 배경의 절제된 수직 요소보다 가깝고 크다. 안규홍의 상체가 주된 면적을 차지하는 미디엄 숏으로, 왼쪽 전경 동료의 어깨 너머에서 본 구도다.",
        "entities": "안규홍과 왼쪽 전경 남성 한 명이 보인다. 전경 인물은 숏에 명시된 동료 중 누구인지는 판별할 수 없으며, 나머지 동료는 프레임 밖일 수 있다. 안규홍은 청년 남성의 외형, 검은 묶음머리, 때 묻은 밝은 저고리와 허리띠, 드러난 목을 갖췄다. 눈꺼풀의 부기와 볼에 붙은 마른 흙이 명확하다. 측면에 가까워 참조 얼굴과의 정확한 일치는 확인이 제한된다. 읽을 수 있는 글자는 없다.",
        "hard_violations": [],
        "physics": "안규홍은 다리를 교차해 낮게 앉아 있으며 접힌 다리와 하체가 체중을 받는 자연스러운 자세다. 정확한 엉덩이 접촉면은 가려져 있지만 떠 있는 간격이나 불가능한 자세는 보이지 않는다. 양손은 무릎과 다리 쪽에 내려놓았고, 전경 남성의 하체는 화면 밖이다. 그릇과 솥, 뒤쪽 장작은 지면 또는 흙받침에 지지되어 있다."
       }
      ],
      "all_candidates_fail": true
     },
     "normalized": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 4,
        "verdict_ko": "안규홍의 정면과 원형 모임은 보이지만, 집단을 넓게 담아 지정된 미디엄 숏을 벗어나고 굵은 나무들이 모임 사이를 차지한다."
       },
       {
        "label": "A",
        "score": 6,
        "verdict_ko": "인물 중심 미디엄 숏과 옆 사람들을 살피는 시선, 부은 눈과 얼굴의 흙은 더 충실하지만, 정면 대신 측면에 가까운 얼굴과 가까운 나무 배치는 어긋난다."
       }
      ],
      "all_candidates_fail": true,
      "readings": [
       {
        "label": "B",
        "direction": "중앙 안규홍은 몸과 얼굴을 카메라 쪽으로 두고 거의 정면을 본다. 시선은 렌즈 또는 바로 앞에 등을 보인 남성 쪽이며, 여러 사람을 훑어보는 순간은 뚜렷하지 않다. 주변 남성들은 대체로 원 안쪽을 향한다. 겨냥하거나 움직이는 소품은 없다.",
        "built_space": "좌우 초가 각 한 채, 뒤쪽 목책 한 줄과 오른쪽 망루 한 기가 보여 참조 산채의 주요 구조는 대응한다. 그러나 굵은 나무줄기 약 일곱 개가 전경과 모임 사이까지 들어와, 먼 가장자리의 부분적인 편백 줄기라는 지시와 다르다. 중앙 인물보다 원형 집단과 장소가 크게 강조된 넓은 구도다.",
        "entities": "중앙 청년 남성과 주변 남성 일곱 명으로 총 여덟 명이 보이며, 인원수는 숏 문장의 안규홍과 일곱 동료에 대응한다. 동료들의 개별 신원은 외형만으로 구분할 수 없다. 중앙 인물은 참조와 유사한 얼굴, 뒤로 단단히 정리한 검은 머리, 때 묻은 밝은 전통 의복과 맨목을 갖췄다. 얼굴 흙은 보이지만 눈의 부기는 비교적 약하다. 읽을 수 있는 글자는 없다.",
        "hard_violations": [],
        "physics": "좌우 남성들의 다리는 굽혀져 있고 보이는 발은 지면에 닿는다. 손은 무릎이나 허벅지에 놓여 있다. 중앙과 전경 인물들의 엉덩이 및 하체 지지점은 다른 몸이나 화면 경계에 가려져 있어 앉은 상태를 완전히 확인할 수 없지만, 공중에 떠 있다고 볼 근거도 없다. 항아리와 그릇은 지면이나 받침 위에 놓여 있다."
       },
       {
        "label": "A",
        "direction": "안규홍의 몸은 카메라를 향하지만 얼굴과 눈은 화면 왼쪽으로 돌아가 있다. 시선은 왼쪽 전경 남성의 머리보다 위쪽을 지나 화면 밖의 원 구성원 쪽으로 향하는 것으로 읽힌다. 주변을 살피는 동작에는 맞지만, 요구한 정면 얼굴은 아니다. 전경 남성은 안규홍 쪽으로 고개를 향한다.",
        "built_space": "좌우 초가 각 한 채, 뒤 목책 한 줄, 오른쪽 망루 한 기와 흙마당이 참조 장소에 대응한다. 좌우에 큰 나무줄기 네 개가 두드러지고 뒤로 가는 줄기들이 이어지는데, 특히 오른쪽 나무들은 먼 배경의 절제된 수직 요소보다 가깝고 크다. 안규홍의 상체가 주된 면적을 차지하는 미디엄 숏으로, 왼쪽 전경 동료의 어깨 너머에서 본 구도다.",
        "entities": "안규홍과 왼쪽 전경 남성 한 명이 보인다. 전경 인물은 숏에 명시된 동료 중 누구인지는 판별할 수 없으며, 나머지 동료는 프레임 밖일 수 있다. 안규홍은 청년 남성의 외형, 검은 묶음머리, 때 묻은 밝은 저고리와 허리띠, 드러난 목을 갖췄다. 눈꺼풀의 부기와 볼에 붙은 마른 흙이 명확하다. 측면에 가까워 참조 얼굴과의 정확한 일치는 확인이 제한된다. 읽을 수 있는 글자는 없다.",
        "hard_violations": [],
        "physics": "안규홍은 다리를 교차해 낮게 앉아 있으며 접힌 다리와 하체가 체중을 받는 자연스러운 자세다. 정확한 엉덩이 접촉면은 가려져 있지만 떠 있는 간격이나 불가능한 자세는 보이지 않는다. 양손은 무릎과 다리 쪽에 내려놓았고, 전경 남성의 하체는 화면 밖이다. 그릇과 솥, 뒤쪽 장작은 지면 또는 흙받침에 지지되어 있다."
       }
      ]
     },
     "ok": true
    }
   ],
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "slot_winner_match": true,
   "slot_winner": {
    "gemini-pro": "A",
    "gpt-high": "A"
   },
   "route": "cross_slot_agree"
  },
  "dual": {
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "normalized": {
    "A": 2.0,
    "B": 1.0
   },
   "adjusted": {
    "A": 2.0,
    "B": 0.75
   },
   "violations": {
    "B": [
     "[gemini-pro] 로케이션 레퍼런스 상 완전히 빈 공터인 산채 마당 중앙에 커다란 편백나무 다수를 임의로 생성하여 배치함 (Invented objects 및 공간 구조 왜곡)"
    ]
   },
   "per_model_winner": {
    "gemini-pro": "A",
    "gpt-high": "A"
   },
   "agreed": true
  },
  "totals": {
   "A": 2000,
   "B": 750
  },
  "selected": "A",
  "ranking": [
   "A",
   "B"
  ],
  "verdicts": [
   {
    "label": "A",
    "score": 2000,
    "verdict_ko": "지정된 미디엄 샷과 숄더 프레이밍의 지시를 정확하게 구현하여 최우선 순위인 촬영 구도를 완벽하게 충족했으며, 안규홍의 외양과 로케이션 구조도 훌륭하게 반영했습니다."
   },
   {
    "label": "B",
    "score": 750,
    "verdict_ko": "미디엄 샷 프레이밍을 무시하고 전체 인물이 보이도록 와이드 샷으로 임의 변경하였으며, 빈 공터여야 할 마당 중앙에 나무들을 빽빽하게 배치하여 로케이션 공간 제약을 크게 위반했습니다.  ★위반: [gemini-pro] 로케이션 레퍼런스 상 완전히 빈 공터인 산채 마당 중앙에 커다란 편백나무 다수를 임의로 생성하여 배치함 (Invented objects 및 공간 구조 왜곡)"
   }
  ],
  "refs": [
   {
    "label": "LOCATION PHOTOGRAPH — the exact place of this shot: its architecture, materials, fixed features and lighting mood are spatial truth; stage the moment inside this place. Never copy its camera framing.",
    "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/db00517b-165b-4005-a287-f28936b59b95/images/272bd7af-9294-4490-98c6-88a40e9deacc/scene/recipe/groupbg_mountain_camp_clearing_cb9567.png",
    "asset_id": "c5df4442-a576-4bdd-9e42-441cb4544a5e",
    "role": "bgfirst_group_bg"
   },
   {
    "label": "CHARACTER REFERENCE — 안규홍: the exact person appearing in this shot; match face, hair and build exactly.",
    "path": "<bytes:1028384>",
    "asset_id": "3b5894db-ce55-4f34-b966-5e725c3aa952",
    "role": "character_ref"
   }
  ],
  "critique_skipped": true,
  "shot_run_produced": false,
  "shot_run_uid": "06aa1435-3cc7-74b8-9ab5-1541542fc3a1",
  "bgfirst": {
   "bg_path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/db00517b-165b-4005-a287-f28936b59b95/images/272bd7af-9294-4490-98c6-88a40e9deacc/scene/recipe/S46sh3__bgfirst_bg.png",
   "bg_asset_id": "f0c0037a-64d8-4768-bc16-70d7921aacb0",
   "bg_record_key": "S46sh3::bgfirst_bg",
   "chain_winner": true,
   "authority": "groupbg",
   "group_key": "mountain_camp_clearing",
   "groupbg_asset_id": "c5df4442-a576-4bdd-9e42-441cb4544a5e"
  },
  "ref_mode": "재투영 배경+콘티+엔티티 (2택1: 체인 승)",
  "share_plan": {
   "ref_plan": "background"
  }
 },
 "S46sh3::cine": {
  "applied": true,
  "attempted_at": "2026-09-09T06:12:37.024405+00:00",
  "fingerprint": "45ab33318d48df570fc76b6b9960fb05cc3ec871fa8a076e8a5eb1530c75c6b3",
  "fingerprint_version": 2,
  "provider": "mai",
  "endpoint": "openrouter/chat-completions",
  "model": "microsoft/mai-image-2.6",
  "pack": "24.202608252115",
  "source_file": "S46sh3_sel.png",
  "source_sha256": "9d0b512c5d9a41df58a5efba0f706a953eb9bc97d30222891d936ce15f5ca20e",
  "file": "S46sh3_cine.png",
  "staged_sha256": "0431e80a51fa50ae45efd33e0ab76c83b0c4246abb803796d39ee9874e610fda",
  "latency_ms": 28740
 },
 "S46sh7::signage": {
  "fp": "4d379ae5f5982ad9",
  "inscriptions": [],
  "cues": [],
  "dropped": []
 },
 "S46sh7": {
  "input_fingerprint": "e1cd0a142cbc3545",
  "prompt": "Create ONE FINAL photorealistic live-action film still of the moment below — early 20th-century Korea, Jeolla region; all people are Korean unless explicitly stated otherwise. TIME OF DAY (lock): morning, slanting sunlight.\n\nSHOT TEXT (authoritative, Korean): 미소를 지우고 결연한 눈빛으로 의병들을 응시하는 안규홍의 단단한 정면 클로즈업.\n\nLOCATION (lock): The speaking position before the seated assembly in the mountain camp's central sunlit clearing. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: close-up\n- KEY BACKGROUND ELEMENTS: 편백 (안규홍 뒤쪽 산채 둘레에 있다) — Only softly separated trunk fragments remain at the frame margins; used as Maintains location continuity without introducing another focal subject.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Preserve the same oblique morning sunlight and muted natural contrast, keeping the dried earth and swollen eyes visible without beautifying them.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The clothing, hair and overall look of 안규홍 — who appear both in that photo and in this shot — are LOCKED to that photo. Anyone else visible in that photo is NOT in this shot: never carry their face, body or clothing onto anyone here. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nPREVIOUS STILL USAGE (follow exactly — what to take from the attached still and what to exclude; it governs place and objects only, never who is in this shot or how the camera sees them): Take the clearing's dirt ground, surrounding cypress vegetation, and morning sunlight. Exclude the funeral mat and the woman's body from the previous night's scene.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): Slanting morning sunlight continues through the cypress trees. 안규홍: He retains swollen eyes, dried facial dirt, firmly tied hair and a bare neck.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 안규홍 (한국인 남성, 청년 나이대, 검은 머리, 단단히 묶은 머리) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.",
  "roll_prompts": {
   "A": "Create ONE FINAL photorealistic live-action film still of the moment below — early 20th-century Korea, Jeolla region; all people are Korean unless explicitly stated otherwise. TIME OF DAY (lock): morning, slanting sunlight.\n\nSHOT TEXT (authoritative, Korean): 미소를 지우고 결연한 눈빛으로 의병들을 응시하는 안규홍의 단단한 정면 클로즈업.\n\nLOCATION (lock): The speaking position before the seated assembly in the mountain camp's central sunlit clearing. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: close-up\n- KEY BACKGROUND ELEMENTS: 편백 (안규홍 뒤쪽 산채 둘레에 있다) — Only softly separated trunk fragments remain at the frame margins; used as Maintains location continuity without introducing another focal subject.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Preserve the same oblique morning sunlight and muted natural contrast, keeping the dried earth and swollen eyes visible without beautifying them.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The clothing, hair and overall look of 안규홍 — who appear both in that photo and in this shot — are LOCKED to that photo. Anyone else visible in that photo is NOT in this shot: never carry their face, body or clothing onto anyone here. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nPREVIOUS STILL USAGE (follow exactly — what to take from the attached still and what to exclude; it governs place and objects only, never who is in this shot or how the camera sees them): Take the clearing's dirt ground, surrounding cypress vegetation, and morning sunlight. Exclude the funeral mat and the woman's body from the previous night's scene.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): Slanting morning sunlight continues through the cypress trees. 안규홍: He retains swollen eyes, dried facial dirt, firmly tied hair and a bare neck.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 안규홍 (한국인 남성, 청년 나이대, 검은 머리, 단단히 묶은 머리) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.",
   "B": "Create ONE FINAL photorealistic live-action film still of the moment below — early 20th-century Korea, Jeolla region; all people are Korean unless explicitly stated otherwise. TIME OF DAY (lock): morning, slanting sunlight.\n\nSHOT TEXT (authoritative, Korean): 미소를 지우고 결연한 눈빛으로 의병들을 응시하는 안규홍의 단단한 정면 클로즈업.\n\nLOCATION (lock): The speaking position before the seated assembly in the mountain camp's central sunlit clearing. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: close-up\n- KEY BACKGROUND ELEMENTS: 편백 (안규홍 뒤쪽 산채 둘레에 있다) — Only softly separated trunk fragments remain at the frame margins; used as Maintains location continuity without introducing another focal subject.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Preserve the same oblique morning sunlight and muted natural contrast, keeping the dried earth and swollen eyes visible without beautifying them.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The clothing, hair and overall look of 안규홍 — who appear both in that photo and in this shot — are LOCKED to that photo. Anyone else visible in that photo is NOT in this shot: never carry their face, body or clothing onto anyone here. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nPREVIOUS STILL USAGE (follow exactly — what to take from the attached still and what to exclude; it governs place and objects only, never who is in this shot or how the camera sees them): Take the clearing's dirt ground, surrounding cypress vegetation, and morning sunlight. Exclude the funeral mat and the woman's body from the previous night's scene.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): Slanting morning sunlight continues through the cypress trees. 안규홍: He retains swollen eyes, dried facial dirt, firmly tied hair and a bare neck.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 안규홍 (한국인 남성, 청년 나이대, 검은 머리, 단단히 묶은 머리) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.\n\nCOMPOSITION VARIATION (this candidate only): keep the place, the people and the moment exactly as specified, and keep from the CAMERA & FRAME contract its subject scale, screen placement and key background elements exactly. For this candidate only, its camera angle and height are deliberately left open: choose a camera position distinctly different from the obvious one, as a film director picking a second setup on the same blocking. Stay true to any explicit framing the SHOT TEXT itself dictates."
  },
  "shot_run_spend_attempt_count": 1,
  "readings": [
   {
    "label": "A",
    "direction": "안규홍의 시선이 카메라 정면이 아닌 화면 왼쪽 바깥을 향하고 있습니다.",
    "built_space": "아침 햇살이 비치는 흙바닥과 편백나무 숲 배경이 지시에 맞게 구성되어 있습니다.",
    "entities": "안규홍의 복장과 더러워진 얼굴, 머리 모양은 일치하지만, 지시된 '부은 눈'의 상태가 나타나지 않습니다.",
    "hard_violations": [],
    "physics": "인물의 형태와 옷의 주름 등이 자연스럽고 물리적으로 어긋남이 없습니다."
   },
   {
    "label": "B",
    "direction": "안규홍의 시선이 화면 정면을 향해 있으며, 결연한 표정으로 앞을 응시하고 있습니다.",
    "built_space": "햇살이 비치는 산채 공터의 흙바닥과 배경에 자리한 편백나무 줄기들이 올바르게 묘사되어 있습니다.",
    "entities": "안규홍(젊은 한국인 남성)의 굳게 묶은 머리, 흙먼지가 묻은 얼굴, 붉게 부어오른 눈, 지정된 복장이 정확히 일치합니다.",
    "hard_violations": [],
    "physics": "인물의 자세와 옷차림에 자연스러운 중력과 조명이 적용되어 지면에 안정적으로 위치해 있습니다."
   }
  ],
  "cross_model_order": {
   "policy": "cross_model_order_v2_slot0_fwd_slot1_rev_parallel2",
   "slots": [
    {
     "model": "gemini-pro",
     "order": "forward",
     "display_to_canonical": {
      "A": "A",
      "B": "B"
     },
     "raw": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 10,
        "verdict_ko": "요청된 '정면 클로즈업' 구도를 정확히 구현하였으며, 이전 샷에서 이어지는 부은 눈과 얼굴의 흙먼지 등 디테일을 훌륭하게 유지했습니다."
       },
       {
        "label": "A",
        "score": 4,
        "verdict_ko": "'정면 클로즈업'이라는 핵심 지시와 달리 고개를 돌린 측면 앵글로 렌더링되었으며, 유지되어야 할 부은 눈의 묘사가 누락되었습니다."
       }
      ],
      "readings": [
       {
        "label": "B",
        "direction": "안규홍의 시선이 화면 정면을 향해 있으며, 결연한 표정으로 앞을 응시하고 있습니다.",
        "built_space": "햇살이 비치는 산채 공터의 흙바닥과 배경에 자리한 편백나무 줄기들이 올바르게 묘사되어 있습니다.",
        "entities": "안규홍(젊은 한국인 남성)의 굳게 묶은 머리, 흙먼지가 묻은 얼굴, 붉게 부어오른 눈, 지정된 복장이 정확히 일치합니다.",
        "hard_violations": [],
        "physics": "인물의 자세와 옷차림에 자연스러운 중력과 조명이 적용되어 지면에 안정적으로 위치해 있습니다."
       },
       {
        "label": "A",
        "direction": "안규홍의 시선이 카메라 정면이 아닌 화면 왼쪽 바깥을 향하고 있습니다.",
        "built_space": "아침 햇살이 비치는 흙바닥과 편백나무 숲 배경이 지시에 맞게 구성되어 있습니다.",
        "entities": "안규홍의 복장과 더러워진 얼굴, 머리 모양은 일치하지만, 지시된 '부은 눈'의 상태가 나타나지 않습니다.",
        "hard_violations": [],
        "physics": "인물의 형태와 옷의 주름 등이 자연스럽고 물리적으로 어긋남이 없습니다."
       }
      ],
      "all_candidates_fail": false
     },
     "normalized": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 10,
        "verdict_ko": "요청된 '정면 클로즈업' 구도를 정확히 구현하였으며, 이전 샷에서 이어지는 부은 눈과 얼굴의 흙먼지 등 디테일을 훌륭하게 유지했습니다."
       },
       {
        "label": "A",
        "score": 4,
        "verdict_ko": "'정면 클로즈업'이라는 핵심 지시와 달리 고개를 돌린 측면 앵글로 렌더링되었으며, 유지되어야 할 부은 눈의 묘사가 누락되었습니다."
       }
      ],
      "all_candidates_fail": false,
      "readings": [
       {
        "label": "B",
        "direction": "안규홍의 시선이 화면 정면을 향해 있으며, 결연한 표정으로 앞을 응시하고 있습니다.",
        "built_space": "햇살이 비치는 산채 공터의 흙바닥과 배경에 자리한 편백나무 줄기들이 올바르게 묘사되어 있습니다.",
        "entities": "안규홍(젊은 한국인 남성)의 굳게 묶은 머리, 흙먼지가 묻은 얼굴, 붉게 부어오른 눈, 지정된 복장이 정확히 일치합니다.",
        "hard_violations": [],
        "physics": "인물의 자세와 옷차림에 자연스러운 중력과 조명이 적용되어 지면에 안정적으로 위치해 있습니다."
       },
       {
        "label": "A",
        "direction": "안규홍의 시선이 카메라 정면이 아닌 화면 왼쪽 바깥을 향하고 있습니다.",
        "built_space": "아침 햇살이 비치는 흙바닥과 편백나무 숲 배경이 지시에 맞게 구성되어 있습니다.",
        "entities": "안규홍의 복장과 더러워진 얼굴, 머리 모양은 일치하지만, 지시된 '부은 눈'의 상태가 나타나지 않습니다.",
        "hard_violations": [],
        "physics": "인물의 형태와 옷의 주름 등이 자연스럽고 물리적으로 어긋남이 없습니다."
       }
      ]
     },
     "ok": true
    },
    {
     "model": "gpt-high",
     "order": "reverse",
     "display_to_canonical": {
      "A": "B",
      "B": "A"
     },
     "raw": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 7,
        "verdict_ko": "B보다 정면에 가까운 얼굴과 부은 눈, 마른 흙때가 지시에 충실하지만, 가슴까지 담은 넓은 구도와 크게 드러난 나무들은 요구된 정면 클로즈업에 못 미친다."
       },
       {
        "label": "B",
        "score": 6,
        "verdict_ko": "결연한 표정과 인물 외형은 잘 맞지만, 얼굴이 왼쪽으로 더 돌아가 있고 가슴과 숲을 넓게 보여 주어 핵심인 단단한 정면 클로즈업에서 더 멀다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "얼굴은 거의 정면이며 눈은 카메라보다 약간 화면 왼쪽의 화면 밖 대상을 향한다. 의병들은 보이지 않아 실제 시선 도착점은 확인할 수 없지만, 앞에 모인 의병들을 바라보는 방향으로 해석 가능하다. 입은 다물려 있고 미소는 없으나 눈빛은 결연함보다 피로감이 조금 더 강하다. 무기나 방향성 있는 소품은 없다.",
        "built_space": "흙바닥과 왼쪽·오른쪽의 굵은 나무줄기 두 개, 그 사이로 여러 가는 줄기가 보인다. 뒤쪽 상단에는 흐릿한 초가지붕과 목조 구조 일부가 있어 산채의 연속성을 뒷받침한다. 인물은 화면 오른쪽에 치우쳐 머리부터 가슴까지 나온다. 나무가 가장자리의 작은 흐린 조각으로만 남지 않고 뿌리와 넓은 줄기까지 차지하며, 배경 노출량도 요청보다 많다. 복제된 시설이나 불가능한 반사는 보이지 않는다.",
        "entities": "청년 동아시아계 남성 한 명이며, 참조의 안규홍과 얼굴형·검은 머리·체격이 대체로 일치한다. 머리는 뒤로 단단히 넘겨 묶은 형태이고, 목은 드러나며, 때 묻은 미색 저고리와 앞섶 매듭이 참조와 대응한다. 눈두덩의 붓기와 붉은 기, 볼과 턱의 마른 흙때가 뚜렷하다. 사선의 따뜻한 자연광이 들어온다. 추가 인물, 장례용 깔개, 여성의 몸, 읽을 수 있는 글자는 없다.",
        "hard_violations": [],
        "physics": "머리는 목과 어깨에 자연스럽게 연결되고 저고리는 어깨와 몸통에 걸쳐 주름진다. 하체와 지면 접촉점은 프레임 밖이므로 앉았는지 서 있는지는 확인할 수 없다. 보이는 상체에는 부유나 불가능한 자세가 없으며 공중에 뜬 소품도 없다."
       },
       {
        "label": "B",
        "direction": "얼굴과 눈이 모두 화면 왼쪽의 화면 밖 대상을 향하며 A보다 얼굴 회전이 크다. 의병들을 향한 응시로 해석할 수 있지만 대상은 프레임 밖이라 직접 확인되지 않는다. 찌푸린 눈썹과 굳게 다문 입은 미소를 지운 결연함을 잘 보여 준다. 다만 요청된 정면성은 약하다. 무기나 손에 든 소품은 없다.",
        "built_space": "흙바닥 위로 왼쪽 중앙의 굵은 줄기 한 개, 왼쪽 가장자리 안쪽의 가는 줄기 한 개, 오른쪽 가장자리의 굵은 줄기 한 개가 두드러지고 뒤로 여러 흐린 줄기가 이어진다. 건물이나 고정 설비는 식별되지 않는다. 인물은 오른쪽에 배치되어 머리부터 가슴까지 보인다. 특히 왼쪽의 큰 줄기가 화면 내부를 넓게 차지해, 가장자리에만 부드러운 줄기 조각을 남기라는 배경 지시와 다르다. 불가능한 반사나 시설 중복은 보이지 않는다.",
        "entities": "청년 동아시아계 남성 한 명으로 참조 인물의 얼굴과 검은 머리, 미색의 낡고 더러운 저고리를 대체로 유지한다. 뒤로 모아 묶은 머리와 드러난 목이 보이며 얼굴에 마른 흙때가 남아 있다. 눈 주변은 다소 부어 있지만 A보다 붓기와 붉은 기가 덜 두드러진다. 숲과 흙바닥에 사선 햇빛이 들어온다. 추가 인물, 장례용 깔개, 여성의 몸, 읽을 수 있는 글자는 없다.",
        "hard_violations": [],
        "physics": "목과 어깨가 머리를 자연스럽게 지탱하고 옷은 몸통에 걸려 아래로 처진다. 하체와 접촉면은 구도 밖이므로 지지 자세 자체는 확인할 수 없지만, 보이는 범위에 부유나 해부학적 불가능성은 없다. 움직이거나 공중에 떠 있는 물체도 없다."
       }
      ],
      "all_candidates_fail": true
     },
     "normalized": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 7,
        "verdict_ko": "B보다 정면에 가까운 얼굴과 부은 눈, 마른 흙때가 지시에 충실하지만, 가슴까지 담은 넓은 구도와 크게 드러난 나무들은 요구된 정면 클로즈업에 못 미친다."
       },
       {
        "label": "A",
        "score": 6,
        "verdict_ko": "결연한 표정과 인물 외형은 잘 맞지만, 얼굴이 왼쪽으로 더 돌아가 있고 가슴과 숲을 넓게 보여 주어 핵심인 단단한 정면 클로즈업에서 더 멀다."
       }
      ],
      "all_candidates_fail": true,
      "readings": [
       {
        "label": "B",
        "direction": "얼굴은 거의 정면이며 눈은 카메라보다 약간 화면 왼쪽의 화면 밖 대상을 향한다. 의병들은 보이지 않아 실제 시선 도착점은 확인할 수 없지만, 앞에 모인 의병들을 바라보는 방향으로 해석 가능하다. 입은 다물려 있고 미소는 없으나 눈빛은 결연함보다 피로감이 조금 더 강하다. 무기나 방향성 있는 소품은 없다.",
        "built_space": "흙바닥과 왼쪽·오른쪽의 굵은 나무줄기 두 개, 그 사이로 여러 가는 줄기가 보인다. 뒤쪽 상단에는 흐릿한 초가지붕과 목조 구조 일부가 있어 산채의 연속성을 뒷받침한다. 인물은 화면 오른쪽에 치우쳐 머리부터 가슴까지 나온다. 나무가 가장자리의 작은 흐린 조각으로만 남지 않고 뿌리와 넓은 줄기까지 차지하며, 배경 노출량도 요청보다 많다. 복제된 시설이나 불가능한 반사는 보이지 않는다.",
        "entities": "청년 동아시아계 남성 한 명이며, 참조의 안규홍과 얼굴형·검은 머리·체격이 대체로 일치한다. 머리는 뒤로 단단히 넘겨 묶은 형태이고, 목은 드러나며, 때 묻은 미색 저고리와 앞섶 매듭이 참조와 대응한다. 눈두덩의 붓기와 붉은 기, 볼과 턱의 마른 흙때가 뚜렷하다. 사선의 따뜻한 자연광이 들어온다. 추가 인물, 장례용 깔개, 여성의 몸, 읽을 수 있는 글자는 없다.",
        "hard_violations": [],
        "physics": "머리는 목과 어깨에 자연스럽게 연결되고 저고리는 어깨와 몸통에 걸쳐 주름진다. 하체와 지면 접촉점은 프레임 밖이므로 앉았는지 서 있는지는 확인할 수 없다. 보이는 상체에는 부유나 불가능한 자세가 없으며 공중에 뜬 소품도 없다."
       },
       {
        "label": "A",
        "direction": "얼굴과 눈이 모두 화면 왼쪽의 화면 밖 대상을 향하며 A보다 얼굴 회전이 크다. 의병들을 향한 응시로 해석할 수 있지만 대상은 프레임 밖이라 직접 확인되지 않는다. 찌푸린 눈썹과 굳게 다문 입은 미소를 지운 결연함을 잘 보여 준다. 다만 요청된 정면성은 약하다. 무기나 손에 든 소품은 없다.",
        "built_space": "흙바닥 위로 왼쪽 중앙의 굵은 줄기 한 개, 왼쪽 가장자리 안쪽의 가는 줄기 한 개, 오른쪽 가장자리의 굵은 줄기 한 개가 두드러지고 뒤로 여러 흐린 줄기가 이어진다. 건물이나 고정 설비는 식별되지 않는다. 인물은 오른쪽에 배치되어 머리부터 가슴까지 보인다. 특히 왼쪽의 큰 줄기가 화면 내부를 넓게 차지해, 가장자리에만 부드러운 줄기 조각을 남기라는 배경 지시와 다르다. 불가능한 반사나 시설 중복은 보이지 않는다.",
        "entities": "청년 동아시아계 남성 한 명으로 참조 인물의 얼굴과 검은 머리, 미색의 낡고 더러운 저고리를 대체로 유지한다. 뒤로 모아 묶은 머리와 드러난 목이 보이며 얼굴에 마른 흙때가 남아 있다. 눈 주변은 다소 부어 있지만 A보다 붓기와 붉은 기가 덜 두드러진다. 숲과 흙바닥에 사선 햇빛이 들어온다. 추가 인물, 장례용 깔개, 여성의 몸, 읽을 수 있는 글자는 없다.",
        "hard_violations": [],
        "physics": "목과 어깨가 머리를 자연스럽게 지탱하고 옷은 몸통에 걸려 아래로 처진다. 하체와 접촉면은 구도 밖이므로 지지 자세 자체는 확인할 수 없지만, 보이는 범위에 부유나 해부학적 불가능성은 없다. 움직이거나 공중에 떠 있는 물체도 없다."
       }
      ]
     },
     "ok": true
    }
   ],
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "slot_winner_match": true,
   "slot_winner": {
    "gemini-pro": "B",
    "gpt-high": "B"
   },
   "route": "cross_slot_agree"
  },
  "dual": {
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "normalized": {
    "A": 1.257,
    "B": 2.0
   },
   "adjusted": {
    "A": 1.257,
    "B": 2.0
   },
   "violations": {},
   "per_model_winner": {
    "gemini-pro": "B",
    "gpt-high": "B"
   },
   "agreed": true
  },
  "totals": {
   "B": 2000,
   "A": 1257
  },
  "selected": "B",
  "ranking": [
   "B",
   "A"
  ],
  "verdicts": [
   {
    "label": "B",
    "score": 2000,
    "verdict_ko": "요청된 '정면 클로즈업' 구도를 정확히 구현하였으며, 이전 샷에서 이어지는 부은 눈과 얼굴의 흙먼지 등 디테일을 훌륭하게 유지했습니다."
   },
   {
    "label": "A",
    "score": 1257,
    "verdict_ko": "'정면 클로즈업'이라는 핵심 지시와 달리 고개를 돌린 측면 앵글로 렌더링되었으며, 유지되어야 할 부은 눈의 묘사가 누락되었습니다."
   }
  ],
  "refs": [
   {
    "label": "PREVIOUS SHOT STILL — a visually related earlier shot of this same place: the location's look, materials, fixed features and lighting mood are LOCKED to this photo; never copy its camera framing. The clothing, hair and overall look of 안규홍 — who appear both in that photo and in this shot — are LOCKED to that photo. Anyone else visible in that photo is NOT in this shot: never carry their face, body or clothing onto anyone here. If this photo shows a character who cannot move (dead or unconscious), that character's exact pose, position and orientation are ALSO LOCKED — treat their body as a fixed prop of the set that only the camera moves around.",
    "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/db00517b-165b-4005-a287-f28936b59b95/images/272bd7af-9294-4490-98c6-88a40e9deacc/scene/recipe/S46sh3_sel.png",
    "asset_id": "1c18dc96-4b29-4622-81e5-05aeaaa60239",
    "role": "prev_still"
   },
   {
    "label": "CHARACTER REFERENCE — 안규홍: the exact person appearing in this shot; match face, hair and build exactly.",
    "path": "<bytes:1028384>",
    "asset_id": "3b5894db-ce55-4f34-b966-5e725c3aa952",
    "role": "character_ref"
   }
  ],
  "critique_skipped": true,
  "shot_run_produced": false,
  "shot_run_uid": "06aa1435-40ee-785b-b04c-a0c66a659b61",
  "ref_mode": "prev+엔티티",
  "share_plan": {
   "ref_plan": "prev",
   "prev_anchor_tag": "S46sh3"
  }
 },
 "S46sh7::cine": {
  "applied": true,
  "attempted_at": "2026-09-09T06:13:51.697219+00:00",
  "fingerprint": "9bd01da6be6f8027c5b43360c95bef3ce0f5bb78e5c09eefc0b6c6aa8c3c01c3",
  "fingerprint_version": 2,
  "provider": "mai",
  "endpoint": "openrouter/chat-completions",
  "model": "microsoft/mai-image-2.6",
  "pack": "24.202608252115",
  "source_file": "S46sh7_sel.png",
  "source_sha256": "d98db0f1da7d57bdd3d4945ead191b53683c01e071897b391bd6a6bf4c2a0413",
  "file": "S46sh7_cine.png",
  "staged_sha256": "60431faab47938003d8bef8788c871ea487abb5d13f502fc722faed00c494e71",
  "latency_ms": 27911
 },
 "S46sh16::signage": {
  "fp": "0aa8b9ae999c6abe",
  "inscriptions": [],
  "cues": [],
  "dropped": []
 },
 "S46sh16": {
  "input_fingerprint": "db5cb2869263aae3",
  "prompt": "Create ONE FINAL photorealistic live-action film still of the moment below — early 20th-century Korea, Jeolla region; all people are Korean unless explicitly stated otherwise. TIME OF DAY (lock): morning, slanting sunlight.\n\nSHOT TEXT (authoritative, Korean): 자리에서 일어나 각자의 위치로 발을 내딛는 mid-stride 자세인 낡은 옷차림의 조선인 사내들의 역동적인 뒷모습.\n\nLOCATION (lock): The open routes leading away from the mountain camp's central assembly circle toward individual posts among the trees. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: wide shot\n- KEY BACKGROUND ELEMENTS: 산채 한가운데의 자리 (둘러앉았던 사람들이 일어나 각자의 자리로 흩어지고 있다) — The vacated central ground remains along the lower edge behind the departing feet; used as Makes the dispersal readable through increasing gaps between individuals; 편백 (산채 주변에 서 있다) — Trunks remain behind and between the departing figures; used as Provides environmental depth without specifying unestablished destinations.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: The established oblique morning sunlight through the cypresses maintains restrained warmth and natural contrast across the worn clothing.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The people visible in that photo are NOT in this shot: never carry their faces, bodies or clothing onto anyone here. Each person in THIS shot is defined solely by the PEOPLE list, their own CHARACTER REFERENCE images and their wardrobe notes. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nPREVIOUS STILL USAGE (follow exactly — what to take from the attached still and what to exclude; it governs place and objects only, never who is in this shot or how the camera sees them): Take the central clearing's ground, camp structures where visible, and morning light through the trees. Exclude the previous night's funeral objects and do not duplicate equipment as the gathering disperses.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): Morning light persists across the camp; ammunition is to be dispersed among separate positions rather than gathered in one place. 백정: He rises and moves to his position carrying his gun. 목수: He rises and moves to his position carrying his gun. 염재보: He moves to inspect firing positions with his matchlock. 임정현: He moves to prepare dispersed ammunition positions. 오주일: He moves to organize positions, retaining his gun. 복만: He moves toward the rear route with his matchlock. 이관회: He rises from the gathering with his gun.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 이관회 (한국인 남성, 청년 나이대, 자연스러운 검은 머리); 염재보 (한국인 남성, 성인 나이대, 자연스러운 검은 머리); 복만 (한국인 남성, 성인 나이대, 거칠게 자란 수염, 자연스러운 검은 머리); 오주일 (한국인 남성, 성인 나이대, 자연스러운 검은 머리); 임정현 (한국인 남성, 성인 나이대, 자연스러운 검은 머리); 백정 (한국인 남성, 성인 나이대, 자연스러운 검은 머리); 목수 (한국인 남성, 성인 나이대, 자연스러운 검은 머리) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.",
  "roll_prompts": {
   "A": "Create ONE FINAL photorealistic live-action film still of the moment below — early 20th-century Korea, Jeolla region; all people are Korean unless explicitly stated otherwise. TIME OF DAY (lock): morning, slanting sunlight.\n\nSHOT TEXT (authoritative, Korean): 자리에서 일어나 각자의 위치로 발을 내딛는 mid-stride 자세인 낡은 옷차림의 조선인 사내들의 역동적인 뒷모습.\n\nLOCATION (lock): The open routes leading away from the mountain camp's central assembly circle toward individual posts among the trees. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: wide shot\n- KEY BACKGROUND ELEMENTS: 산채 한가운데의 자리 (둘러앉았던 사람들이 일어나 각자의 자리로 흩어지고 있다) — The vacated central ground remains along the lower edge behind the departing feet; used as Makes the dispersal readable through increasing gaps between individuals; 편백 (산채 주변에 서 있다) — Trunks remain behind and between the departing figures; used as Provides environmental depth without specifying unestablished destinations.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: The established oblique morning sunlight through the cypresses maintains restrained warmth and natural contrast across the worn clothing.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The people visible in that photo are NOT in this shot: never carry their faces, bodies or clothing onto anyone here. Each person in THIS shot is defined solely by the PEOPLE list, their own CHARACTER REFERENCE images and their wardrobe notes. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nPREVIOUS STILL USAGE (follow exactly — what to take from the attached still and what to exclude; it governs place and objects only, never who is in this shot or how the camera sees them): Take the central clearing's ground, camp structures where visible, and morning light through the trees. Exclude the previous night's funeral objects and do not duplicate equipment as the gathering disperses.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): Morning light persists across the camp; ammunition is to be dispersed among separate positions rather than gathered in one place. 백정: He rises and moves to his position carrying his gun. 목수: He rises and moves to his position carrying his gun. 염재보: He moves to inspect firing positions with his matchlock. 임정현: He moves to prepare dispersed ammunition positions. 오주일: He moves to organize positions, retaining his gun. 복만: He moves toward the rear route with his matchlock. 이관회: He rises from the gathering with his gun.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 이관회 (한국인 남성, 청년 나이대, 자연스러운 검은 머리); 염재보 (한국인 남성, 성인 나이대, 자연스러운 검은 머리); 복만 (한국인 남성, 성인 나이대, 거칠게 자란 수염, 자연스러운 검은 머리); 오주일 (한국인 남성, 성인 나이대, 자연스러운 검은 머리); 임정현 (한국인 남성, 성인 나이대, 자연스러운 검은 머리); 백정 (한국인 남성, 성인 나이대, 자연스러운 검은 머리); 목수 (한국인 남성, 성인 나이대, 자연스러운 검은 머리) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.",
   "B": "Create ONE FINAL photorealistic live-action film still of the moment below — early 20th-century Korea, Jeolla region; all people are Korean unless explicitly stated otherwise. TIME OF DAY (lock): morning, slanting sunlight.\n\nSHOT TEXT (authoritative, Korean): 자리에서 일어나 각자의 위치로 발을 내딛는 mid-stride 자세인 낡은 옷차림의 조선인 사내들의 역동적인 뒷모습.\n\nLOCATION (lock): The open routes leading away from the mountain camp's central assembly circle toward individual posts among the trees. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: wide shot\n- KEY BACKGROUND ELEMENTS: 산채 한가운데의 자리 (둘러앉았던 사람들이 일어나 각자의 자리로 흩어지고 있다) — The vacated central ground remains along the lower edge behind the departing feet; used as Makes the dispersal readable through increasing gaps between individuals; 편백 (산채 주변에 서 있다) — Trunks remain behind and between the departing figures; used as Provides environmental depth without specifying unestablished destinations.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: The established oblique morning sunlight through the cypresses maintains restrained warmth and natural contrast across the worn clothing.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The people visible in that photo are NOT in this shot: never carry their faces, bodies or clothing onto anyone here. Each person in THIS shot is defined solely by the PEOPLE list, their own CHARACTER REFERENCE images and their wardrobe notes. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nPREVIOUS STILL USAGE (follow exactly — what to take from the attached still and what to exclude; it governs place and objects only, never who is in this shot or how the camera sees them): Take the central clearing's ground, camp structures where visible, and morning light through the trees. Exclude the previous night's funeral objects and do not duplicate equipment as the gathering disperses.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): Morning light persists across the camp; ammunition is to be dispersed among separate positions rather than gathered in one place. 백정: He rises and moves to his position carrying his gun. 목수: He rises and moves to his position carrying his gun. 염재보: He moves to inspect firing positions with his matchlock. 임정현: He moves to prepare dispersed ammunition positions. 오주일: He moves to organize positions, retaining his gun. 복만: He moves toward the rear route with his matchlock. 이관회: He rises from the gathering with his gun.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 이관회 (한국인 남성, 청년 나이대, 자연스러운 검은 머리); 염재보 (한국인 남성, 성인 나이대, 자연스러운 검은 머리); 복만 (한국인 남성, 성인 나이대, 거칠게 자란 수염, 자연스러운 검은 머리); 오주일 (한국인 남성, 성인 나이대, 자연스러운 검은 머리); 임정현 (한국인 남성, 성인 나이대, 자연스러운 검은 머리); 백정 (한국인 남성, 성인 나이대, 자연스러운 검은 머리); 목수 (한국인 남성, 성인 나이대, 자연스러운 검은 머리) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.\n\nCOMPOSITION VARIATION (this candidate only): keep the place, the people and the moment exactly as specified, and keep from the CAMERA & FRAME contract its subject scale, screen placement and key background elements exactly. For this candidate only, its camera angle and height are deliberately left open: choose a camera position distinctly different from the obvious one, as a film director picking a second setup on the same blocking. Stay true to any explicit framing the SHOT TEXT itself dictates."
  },
  "shot_run_spend_attempt_count": 1,
  "readings": [
   {
    "label": "A",
    "direction": "11명의 인물들이 카메라를 등지고 화면 안쪽 숲길을 향해 시선과 방향을 두고 걸어가고 있다.",
    "built_space": "화면 왼쪽에 오두막들이 늘어서 있고 오른쪽에 나무들이 있는 직선 형태의 숲길이다. 이전 샷 레퍼런스에 명시된 넓은 중앙 공터와 나무 울타리(palisade)가 전혀 반영되지 않아 장소가 불일치한다.",
    "entities": "총 11명의 인물이 등장해 지정된 인원수(7명)를 크게 초과함. 목수, 이관회, 임정현, 염재보와 유사한 인물이 있으나, 나무 상자를 든 남자나 우측의 곱슬머리 남자 등 텍스트에 없는 인물들이 다수 섞여 있다.",
    "hard_violations": [
     "[gemini-pro] 초과된 인물 수 (11명)",
     "[gemini-pro] 텍스트에 없는 창조된 인물 및 사물 (나무 상자를 든 사람, 곱슬머리 남자 등)",
     "[gemini-pro] 명시된 위치(LOCATION)와 완전히 다른 공간 구성 (중앙 공터 부재)",
     "[gpt-high] 전경의 일곱 명 외에 배경 통로에 적어도 다섯 명의 추가 인물을 생성하여, 허용된 인물 외에는 아무도 등장시키지 말라는 제한을 위반했다."
    ],
    "physics": "인물들의 발이 땅에 닿아 있고 걷는 자세가 자연스럽게 지지되고 있다. 나무 상자와 총기류 모두 인물의 손에 쥐어져 있다."
   },
   {
    "label": "B",
    "direction": "7명의 인물들이 카메라를 등지고 산채 중앙 공터에서 각자의 위치인 숲 쪽을 향해 발을 내딛고 있다.",
    "built_space": "하단에 인물들이 떠난 빈 중앙 공터가 넓게 배치되어 있고, 우측 원경의 나무 울타리(palisade)와 주변을 둘러싼 큰 나무들이 레퍼런스의 장소 및 사선으로 떨어지는 아침 햇살과 정확히 일치한다.",
    "entities": "정확히 7명의 인물이 등장하며 모두 지시된 대로 총기를 소지함. 목수, 임정현, 이관회, 복만, 염재보의 외형이 일치하나, 백정과 오주일 대신 파란 상의의 인물(제공된 8번째 레퍼런스)과 짙은 남색 옷의 인물이 렌더링되어 아이덴티티 오류가 있다.",
    "hard_violations": [
     "[gemini-pro] 텍스트 목록에 없는 인물 등장 (파란 상의를 입은 인물, 짙은 남색 옷의 인물)"
    ],
    "physics": "모든 인물이 공중에 떠 있지 않고 발을 흙바닥에 굳건히 딛은 채 mid-stride 자세를 취하고 있다. 무기들은 모두 손에 쥐어지거나 어깨 끈으로 지지되어 있으며, 왼쪽 바위 위 총기 역시 표면 위에 안정적으로 놓여 있다."
   }
  ],
  "cross_model_order": {
   "policy": "cross_model_order_v2_slot0_fwd_slot1_rev_parallel2",
   "slots": [
    {
     "model": "gemini-pro",
     "order": "forward",
     "display_to_canonical": {
      "A": "A",
      "B": "B"
     },
     "raw": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 2,
        "verdict_ko": "지정된 7명 외에 불필요한 인물을 다수 추가하고 금지된 사물(나무 상자)을 렌더링했으며, 산채 공터라는 장소(LOCATION) 설정마저 숲길로 바꾸어 프롬프트를 크게 위반했습니다."
       },
       {
        "label": "B",
        "score": 7,
        "verdict_ko": "산채 중앙 공터의 흙바닥과 나무 울타리 배경을 완벽하게 재현하고 인원수(7명)와 역동적인 구도를 잘 구현했으나, 일부 인물의 복장이 텍스트 목록 대신 잘못된 레퍼런스를 따른 점이 유일한 흠입니다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "11명의 인물들이 카메라를 등지고 화면 안쪽 숲길을 향해 시선과 방향을 두고 걸어가고 있다.",
        "built_space": "화면 왼쪽에 오두막들이 늘어서 있고 오른쪽에 나무들이 있는 직선 형태의 숲길이다. 이전 샷 레퍼런스에 명시된 넓은 중앙 공터와 나무 울타리(palisade)가 전혀 반영되지 않아 장소가 불일치한다.",
        "entities": "총 11명의 인물이 등장해 지정된 인원수(7명)를 크게 초과함. 목수, 이관회, 임정현, 염재보와 유사한 인물이 있으나, 나무 상자를 든 남자나 우측의 곱슬머리 남자 등 텍스트에 없는 인물들이 다수 섞여 있다.",
        "hard_violations": [
         "초과된 인물 수 (11명)",
         "텍스트에 없는 창조된 인물 및 사물 (나무 상자를 든 사람, 곱슬머리 남자 등)",
         "명시된 위치(LOCATION)와 완전히 다른 공간 구성 (중앙 공터 부재)"
        ],
        "physics": "인물들의 발이 땅에 닿아 있고 걷는 자세가 자연스럽게 지지되고 있다. 나무 상자와 총기류 모두 인물의 손에 쥐어져 있다."
       },
       {
        "label": "B",
        "direction": "7명의 인물들이 카메라를 등지고 산채 중앙 공터에서 각자의 위치인 숲 쪽을 향해 발을 내딛고 있다.",
        "built_space": "하단에 인물들이 떠난 빈 중앙 공터가 넓게 배치되어 있고, 우측 원경의 나무 울타리(palisade)와 주변을 둘러싼 큰 나무들이 레퍼런스의 장소 및 사선으로 떨어지는 아침 햇살과 정확히 일치한다.",
        "entities": "정확히 7명의 인물이 등장하며 모두 지시된 대로 총기를 소지함. 목수, 임정현, 이관회, 복만, 염재보의 외형이 일치하나, 백정과 오주일 대신 파란 상의의 인물(제공된 8번째 레퍼런스)과 짙은 남색 옷의 인물이 렌더링되어 아이덴티티 오류가 있다.",
        "hard_violations": [
         "텍스트 목록에 없는 인물 등장 (파란 상의를 입은 인물, 짙은 남색 옷의 인물)"
        ],
        "physics": "모든 인물이 공중에 떠 있지 않고 발을 흙바닥에 굳건히 딛은 채 mid-stride 자세를 취하고 있다. 무기들은 모두 손에 쥐어지거나 어깨 끈으로 지지되어 있으며, 왼쪽 바위 위 총기 역시 표면 위에 안정적으로 놓여 있다."
       }
      ],
      "all_candidates_fail": false
     },
     "normalized": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 2,
        "verdict_ko": "지정된 7명 외에 불필요한 인물을 다수 추가하고 금지된 사물(나무 상자)을 렌더링했으며, 산채 공터라는 장소(LOCATION) 설정마저 숲길로 바꾸어 프롬프트를 크게 위반했습니다."
       },
       {
        "label": "B",
        "score": 7,
        "verdict_ko": "산채 중앙 공터의 흙바닥과 나무 울타리 배경을 완벽하게 재현하고 인원수(7명)와 역동적인 구도를 잘 구현했으나, 일부 인물의 복장이 텍스트 목록 대신 잘못된 레퍼런스를 따른 점이 유일한 흠입니다."
       }
      ],
      "all_candidates_fail": false,
      "readings": [
       {
        "label": "A",
        "direction": "11명의 인물들이 카메라를 등지고 화면 안쪽 숲길을 향해 시선과 방향을 두고 걸어가고 있다.",
        "built_space": "화면 왼쪽에 오두막들이 늘어서 있고 오른쪽에 나무들이 있는 직선 형태의 숲길이다. 이전 샷 레퍼런스에 명시된 넓은 중앙 공터와 나무 울타리(palisade)가 전혀 반영되지 않아 장소가 불일치한다.",
        "entities": "총 11명의 인물이 등장해 지정된 인원수(7명)를 크게 초과함. 목수, 이관회, 임정현, 염재보와 유사한 인물이 있으나, 나무 상자를 든 남자나 우측의 곱슬머리 남자 등 텍스트에 없는 인물들이 다수 섞여 있다.",
        "hard_violations": [
         "초과된 인물 수 (11명)",
         "텍스트에 없는 창조된 인물 및 사물 (나무 상자를 든 사람, 곱슬머리 남자 등)",
         "명시된 위치(LOCATION)와 완전히 다른 공간 구성 (중앙 공터 부재)"
        ],
        "physics": "인물들의 발이 땅에 닿아 있고 걷는 자세가 자연스럽게 지지되고 있다. 나무 상자와 총기류 모두 인물의 손에 쥐어져 있다."
       },
       {
        "label": "B",
        "direction": "7명의 인물들이 카메라를 등지고 산채 중앙 공터에서 각자의 위치인 숲 쪽을 향해 발을 내딛고 있다.",
        "built_space": "하단에 인물들이 떠난 빈 중앙 공터가 넓게 배치되어 있고, 우측 원경의 나무 울타리(palisade)와 주변을 둘러싼 큰 나무들이 레퍼런스의 장소 및 사선으로 떨어지는 아침 햇살과 정확히 일치한다.",
        "entities": "정확히 7명의 인물이 등장하며 모두 지시된 대로 총기를 소지함. 목수, 임정현, 이관회, 복만, 염재보의 외형이 일치하나, 백정과 오주일 대신 파란 상의의 인물(제공된 8번째 레퍼런스)과 짙은 남색 옷의 인물이 렌더링되어 아이덴티티 오류가 있다.",
        "hard_violations": [
         "텍스트 목록에 없는 인물 등장 (파란 상의를 입은 인물, 짙은 남색 옷의 인물)"
        ],
        "physics": "모든 인물이 공중에 떠 있지 않고 발을 흙바닥에 굳건히 딛은 채 mid-stride 자세를 취하고 있다. 무기들은 모두 손에 쥐어지거나 어깨 끈으로 지지되어 있으며, 왼쪽 바위 위 총기 역시 표면 위에 안정적으로 놓여 있다."
       }
      ]
     },
     "ok": true
    },
    {
     "model": "gpt-high",
     "order": "reverse",
     "display_to_canonical": {
      "A": "B",
      "B": "A"
     },
     "raw": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 7,
        "verdict_ko": "일곱 사내의 보행 중 뒷모습과 발뒤에 남은 빈 집결지를 와이드 숏으로 구현했으나, 목수의 모자와 임정현·복만의 의상 등 인물별 참조 일치도가 떨어진다."
       },
       {
        "label": "B",
        "score": 2,
        "verdict_ko": "보행과 아침빛은 자연스럽지만 일곱 전경 인물 외에 여러 사람을 추가했고, 각자 흩어지기보다 같은 통로를 따라가는 행렬로 보여 핵심 지시를 위반한다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "일곱 명 모두 카메라에 등을 보이고 중앙 공터에서 숲과 울타리 사이 통로 쪽으로 걷는다. 왼쪽과 오른쪽 인물의 발걸음은 바깥쪽으로 조금 벌어지고, 가운데 군복 인물은 더 깊숙이 전진해 간격과 깊이 차이가 생긴다. 다만 대부분 같은 중앙 출입구로 향하는 성격도 있다. 얼굴과 눈은 거의 보이지 않는다. 왼쪽 두 명의 총구는 어깨 위쪽을 향하고, 오른쪽 청년의 총구는 화면 왼쪽 아래를 향하며 누구를 겨냥하는 동작은 아니다.",
        "built_space": "좌우 가장자리에 초가지붕 구조물이 하나씩 보이고, 중경에는 중앙 통로 양옆으로 목책이 이어진다. 굵은 나무줄기가 인물 뒤와 사이에 서 있으며, 하단에는 떠나는 발 뒤로 빈 흙바닥이 남는다. 왼쪽 아래에는 그루터기 하나와 솥·그릇류, 오른쪽 아래에는 목재 받침과 긴 목재들이 있다. 참고 장소의 흙바닥, 초가, 목책, 수목이라는 주요 재료와 구성이 이어진다. 왼쪽 그루터기 위 총 한 자루는 보이지만 이전 장면의 장비와 같은 물건인지는 확인되지 않는다. 반사면은 없다.",
        "entities": "사람은 정확히 일곱 명이며, 보이는 피부와 체형은 청년 또는 성인 동아시아 남성 설정에 어긋나지 않는다. 뒷모습이라 한국인 여부, 정확한 얼굴 동일성, 복만의 수염은 검증할 수 없다. 왼쪽 백정으로 읽히는 인물은 어두운 두건과 낡은 겉옷이 맞는다. 그 옆 목수는 청색 상의와 누런 바지는 맞지만 짚모자가 없다. 중앙 왼쪽 염재보로 읽히는 인물은 어두운 옷이지만 삿갓이 없다. 먼 중앙 오주일은 군모·군복·밝은 각반이 맞는다. 흰옷 인물은 임정현 참조의 긴 도포 대신 짧은 상의를 입었다. 오른쪽 중경의 복만으로 읽히는 인물은 챙모자가 있으나 긴 겉옷이 짧아졌다. 맨 오른쪽 이관회는 짧은 검은 머리, 갈색 상의, 밝은 짧은 바지가 대체로 맞는다. 여러 인물의 총은 확인되지만 염재보·복만의 화승총 세부 형식과 탄약 분산 준비는 명확하지 않다. 읽을 수 있는 글자나 장례 물건은 없다.",
        "hard_violations": [],
        "physics": "각 인물은 한쪽 발로 흙바닥을 딛고 다른 발의 뒤꿈치나 발 전체를 들어 보행하고 있다. 특히 가까운 세 인물은 디딤발과 들린 발이 구별되어 체중 이동이 가능하다. 어깨에 멘 총은 몸에 걸린 멜빵과 어깨가 지지하고, 오른쪽 청년의 가로로 든 총은 손이 받친다. 그루터기 위 총과 바닥의 그릇·목재는 표면에 놓여 있다. 지지 없이 떠 있는 인체나 물체는 보이지 않는다."
       },
       {
        "label": "B",
        "direction": "가까운 일곱 명은 모두 등을 보이며 화면 중앙 깊숙한 하나의 통로로 전진한다. 각자의 방향으로 벌어지는 분산보다 같은 길을 따라가는 행렬에 가깝다. 흰옷의 가까운 총잡이는 고개를 아래로 숙인다. 왼쪽 인물의 총구는 위로, 목수의 총구는 땅으로, 가운데 흰옷 인물의 총구는 오른쪽 아래로 향한다. 맨 오른쪽 청년의 총구는 왼쪽을 향하며 발사 대상을 겨누지는 않는다. 배경의 추가 인물들도 통로 주변에 모여 있다.",
        "built_space": "하단에 빈 흙바닥이 있고, 나무줄기들이 인물 사이와 뒤에 놓인다. 왼쪽에는 큰 초가 한 동과 그 뒤로 최소 세 개의 작은 지붕 구간이 이어지며, 오른쪽에도 작은 지붕 구조물이 보인다. 왼쪽 전경에는 밧줄을 두른 독립 목재 기둥 하나가 있다. 참고 이미지의 중앙 목책 출입구보다 건물들이 길게 늘어선 산채 내부 통로가 강조되어 동일 장소의 고정 배치를 확인하기 어렵다. 모든 인물은 지상 통로에 있으며 불가능한 반사는 없다.",
        "entities": "가까운 인물 일곱 명 외에 통로 깊숙한 곳에 적어도 다섯 명이 더 보인다. 따라서 허용된 일곱 인물만 등장해야 한다는 조건을 충족하지 않는다. 왼쪽 백정의 두건과 어두운 옷, 목수의 짚모자·청색 상의·누런 바지, 오주일의 군복과 각반, 오른쪽 이관회의 갈색 상의와 밝은 바지는 참조와 대체로 맞는다. 반면 염재보에 대응해야 할 인물은 흰옷과 상투 차림으로 읽혀 어두운 옷과 삿갓 참조에 맞지 않는다. 상자를 든 흰옷 인물 역시 임정현의 긴 도포와 다르다. 오른쪽의 긴 어두운 겉옷 인물은 복만에 가까우나 참조의 챙모자가 없다. 뒷모습이라 얼굴 동일성과 수염은 확인할 수 없다. 작은 나무상자는 보이지만 내용물이 탄약인지는 알 수 없다. 읽을 수 있는 글자는 없다.",
        "hard_violations": [
         "전경의 일곱 명 외에 배경 통로에 적어도 다섯 명의 추가 인물을 생성하여, 허용된 인물 외에는 아무도 등장시키지 말라는 제한을 위반했다."
        ],
        "physics": "가까운 인물들은 디딤발을 지면에 두고 반대쪽 발을 들어 자연스러운 보행 자세를 보인다. 총은 손으로 잡거나 어깨의 멜빵으로 지지하며, 가운데 나무상자는 손잡이를 쥔 손에 매달려 있다. 옷자락과 총 멜빵은 중력에 따라 아래로 처진다. 멀리 있는 사람들의 접지 세부는 작아서 불명확하지만, 지지 없이 공중에 떠 있는 인물이나 물체는 보이지 않는다."
       }
      ],
      "all_candidates_fail": false
     },
     "normalized": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 7,
        "verdict_ko": "일곱 사내의 보행 중 뒷모습과 발뒤에 남은 빈 집결지를 와이드 숏으로 구현했으나, 목수의 모자와 임정현·복만의 의상 등 인물별 참조 일치도가 떨어진다."
       },
       {
        "label": "A",
        "score": 2,
        "verdict_ko": "보행과 아침빛은 자연스럽지만 일곱 전경 인물 외에 여러 사람을 추가했고, 각자 흩어지기보다 같은 통로를 따라가는 행렬로 보여 핵심 지시를 위반한다."
       }
      ],
      "all_candidates_fail": false,
      "readings": [
       {
        "label": "B",
        "direction": "일곱 명 모두 카메라에 등을 보이고 중앙 공터에서 숲과 울타리 사이 통로 쪽으로 걷는다. 왼쪽과 오른쪽 인물의 발걸음은 바깥쪽으로 조금 벌어지고, 가운데 군복 인물은 더 깊숙이 전진해 간격과 깊이 차이가 생긴다. 다만 대부분 같은 중앙 출입구로 향하는 성격도 있다. 얼굴과 눈은 거의 보이지 않는다. 왼쪽 두 명의 총구는 어깨 위쪽을 향하고, 오른쪽 청년의 총구는 화면 왼쪽 아래를 향하며 누구를 겨냥하는 동작은 아니다.",
        "built_space": "좌우 가장자리에 초가지붕 구조물이 하나씩 보이고, 중경에는 중앙 통로 양옆으로 목책이 이어진다. 굵은 나무줄기가 인물 뒤와 사이에 서 있으며, 하단에는 떠나는 발 뒤로 빈 흙바닥이 남는다. 왼쪽 아래에는 그루터기 하나와 솥·그릇류, 오른쪽 아래에는 목재 받침과 긴 목재들이 있다. 참고 장소의 흙바닥, 초가, 목책, 수목이라는 주요 재료와 구성이 이어진다. 왼쪽 그루터기 위 총 한 자루는 보이지만 이전 장면의 장비와 같은 물건인지는 확인되지 않는다. 반사면은 없다.",
        "entities": "사람은 정확히 일곱 명이며, 보이는 피부와 체형은 청년 또는 성인 동아시아 남성 설정에 어긋나지 않는다. 뒷모습이라 한국인 여부, 정확한 얼굴 동일성, 복만의 수염은 검증할 수 없다. 왼쪽 백정으로 읽히는 인물은 어두운 두건과 낡은 겉옷이 맞는다. 그 옆 목수는 청색 상의와 누런 바지는 맞지만 짚모자가 없다. 중앙 왼쪽 염재보로 읽히는 인물은 어두운 옷이지만 삿갓이 없다. 먼 중앙 오주일은 군모·군복·밝은 각반이 맞는다. 흰옷 인물은 임정현 참조의 긴 도포 대신 짧은 상의를 입었다. 오른쪽 중경의 복만으로 읽히는 인물은 챙모자가 있으나 긴 겉옷이 짧아졌다. 맨 오른쪽 이관회는 짧은 검은 머리, 갈색 상의, 밝은 짧은 바지가 대체로 맞는다. 여러 인물의 총은 확인되지만 염재보·복만의 화승총 세부 형식과 탄약 분산 준비는 명확하지 않다. 읽을 수 있는 글자나 장례 물건은 없다.",
        "hard_violations": [],
        "physics": "각 인물은 한쪽 발로 흙바닥을 딛고 다른 발의 뒤꿈치나 발 전체를 들어 보행하고 있다. 특히 가까운 세 인물은 디딤발과 들린 발이 구별되어 체중 이동이 가능하다. 어깨에 멘 총은 몸에 걸린 멜빵과 어깨가 지지하고, 오른쪽 청년의 가로로 든 총은 손이 받친다. 그루터기 위 총과 바닥의 그릇·목재는 표면에 놓여 있다. 지지 없이 떠 있는 인체나 물체는 보이지 않는다."
       },
       {
        "label": "A",
        "direction": "가까운 일곱 명은 모두 등을 보이며 화면 중앙 깊숙한 하나의 통로로 전진한다. 각자의 방향으로 벌어지는 분산보다 같은 길을 따라가는 행렬에 가깝다. 흰옷의 가까운 총잡이는 고개를 아래로 숙인다. 왼쪽 인물의 총구는 위로, 목수의 총구는 땅으로, 가운데 흰옷 인물의 총구는 오른쪽 아래로 향한다. 맨 오른쪽 청년의 총구는 왼쪽을 향하며 발사 대상을 겨누지는 않는다. 배경의 추가 인물들도 통로 주변에 모여 있다.",
        "built_space": "하단에 빈 흙바닥이 있고, 나무줄기들이 인물 사이와 뒤에 놓인다. 왼쪽에는 큰 초가 한 동과 그 뒤로 최소 세 개의 작은 지붕 구간이 이어지며, 오른쪽에도 작은 지붕 구조물이 보인다. 왼쪽 전경에는 밧줄을 두른 독립 목재 기둥 하나가 있다. 참고 이미지의 중앙 목책 출입구보다 건물들이 길게 늘어선 산채 내부 통로가 강조되어 동일 장소의 고정 배치를 확인하기 어렵다. 모든 인물은 지상 통로에 있으며 불가능한 반사는 없다.",
        "entities": "가까운 인물 일곱 명 외에 통로 깊숙한 곳에 적어도 다섯 명이 더 보인다. 따라서 허용된 일곱 인물만 등장해야 한다는 조건을 충족하지 않는다. 왼쪽 백정의 두건과 어두운 옷, 목수의 짚모자·청색 상의·누런 바지, 오주일의 군복과 각반, 오른쪽 이관회의 갈색 상의와 밝은 바지는 참조와 대체로 맞는다. 반면 염재보에 대응해야 할 인물은 흰옷과 상투 차림으로 읽혀 어두운 옷과 삿갓 참조에 맞지 않는다. 상자를 든 흰옷 인물 역시 임정현의 긴 도포와 다르다. 오른쪽의 긴 어두운 겉옷 인물은 복만에 가까우나 참조의 챙모자가 없다. 뒷모습이라 얼굴 동일성과 수염은 확인할 수 없다. 작은 나무상자는 보이지만 내용물이 탄약인지는 알 수 없다. 읽을 수 있는 글자는 없다.",
        "hard_violations": [
         "전경의 일곱 명 외에 배경 통로에 적어도 다섯 명의 추가 인물을 생성하여, 허용된 인물 외에는 아무도 등장시키지 말라는 제한을 위반했다."
        ],
        "physics": "가까운 인물들은 디딤발을 지면에 두고 반대쪽 발을 들어 자연스러운 보행 자세를 보인다. 총은 손으로 잡거나 어깨의 멜빵으로 지지하며, 가운데 나무상자는 손잡이를 쥔 손에 매달려 있다. 옷자락과 총 멜빵은 중력에 따라 아래로 처진다. 멀리 있는 사람들의 접지 세부는 작아서 불명확하지만, 지지 없이 공중에 떠 있는 인물이나 물체는 보이지 않는다."
       }
      ]
     },
     "ok": true
    }
   ],
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "slot_winner_match": true,
   "slot_winner": {
    "gemini-pro": "B",
    "gpt-high": "B"
   },
   "route": "cross_slot_agree"
  },
  "dual": {
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "normalized": {
    "A": 0.571,
    "B": 2.0
   },
   "adjusted": {
    "A": 0.321,
    "B": 1.75
   },
   "violations": {
    "A": [
     "[gemini-pro] 초과된 인물 수 (11명)",
     "[gemini-pro] 텍스트에 없는 창조된 인물 및 사물 (나무 상자를 든 사람, 곱슬머리 남자 등)",
     "[gemini-pro] 명시된 위치(LOCATION)와 완전히 다른 공간 구성 (중앙 공터 부재)",
     "[gpt-high] 전경의 일곱 명 외에 배경 통로에 적어도 다섯 명의 추가 인물을 생성하여, 허용된 인물 외에는 아무도 등장시키지 말라는 제한을 위반했다."
    ],
    "B": [
     "[gemini-pro] 텍스트 목록에 없는 인물 등장 (파란 상의를 입은 인물, 짙은 남색 옷의 인물)"
    ]
   },
   "per_model_winner": {
    "gemini-pro": "B",
    "gpt-high": "B"
   },
   "agreed": true
  },
  "totals": {
   "A": 321,
   "B": 1750
  },
  "selected": "B",
  "ranking": [
   "B",
   "A"
  ],
  "verdicts": [
   {
    "label": "A",
    "score": 321,
    "verdict_ko": "지정된 7명 외에 불필요한 인물을 다수 추가하고 금지된 사물(나무 상자)을 렌더링했으며, 산채 공터라는 장소(LOCATION) 설정마저 숲길로 바꾸어 프롬프트를 크게 위반했습니다.  ★위반: [gemini-pro] 초과된 인물 수 (11명) / [gemini-pro] 텍스트에 없는 창조된 인물 및 사물 (나무 상자를 든 사람, 곱슬머리 남자 등) / [gemini-pro] 명시된 위치(LOCATION)와 완전히 다른 공간 구성 (중앙 공터 부재) / [gpt-high] 전경의 일곱 명 외에 배경 통로에 적어도 다섯 명의 추가 인물을 생성하여, 허용된 인물 외에는 아무도 등장시키지 말라는 제한을 위반했다."
   },
   {
    "label": "B",
    "score": 1750,
    "verdict_ko": "산채 중앙 공터의 흙바닥과 나무 울타리 배경을 완벽하게 재현하고 인원수(7명)와 역동적인 구도를 잘 구현했으나, 일부 인물의 복장이 텍스트 목록 대신 잘못된 레퍼런스를 따른 점이 유일한 흠입니다.  ★위반: [gemini-pro] 텍스트 목록에 없는 인물 등장 (파란 상의를 입은 인물, 짙은 남색 옷의 인물)"
   }
  ],
  "refs": [
   {
    "label": "PREVIOUS SHOT STILL — a visually related earlier shot of this same place: the location's look, materials, fixed features and lighting mood are LOCKED to this photo; never copy its camera framing. The people visible in that photo are NOT in this shot: never carry their faces, bodies or clothing onto anyone here. Each person in THIS shot is defined solely by the PEOPLE list, their own CHARACTER REFERENCE images and their wardrobe notes. If this photo shows a character who cannot move (dead or unconscious), that character's exact pose, position and orientation are ALSO LOCKED — treat their body as a fixed prop of the set that only the camera moves around.",
    "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/db00517b-165b-4005-a287-f28936b59b95/images/272bd7af-9294-4490-98c6-88a40e9deacc/scene/recipe/S46sh3_sel.png",
    "asset_id": "1c18dc96-4b29-4622-81e5-05aeaaa60239",
    "role": "prev_still"
   },
   {
    "label": "CHARACTER REFERENCE — 이관회: the exact person appearing in this shot; match face, hair and build exactly.",
    "path": "<bytes:933540>",
    "asset_id": "851b450f-ac84-47fe-ac5a-03a56f63fcae",
    "role": "character_ref"
   },
   {
    "label": "CHARACTER REFERENCE — 염재보: the exact person appearing in this shot; match face, hair and build exactly.",
    "path": "<bytes:964351>",
    "asset_id": "ab287289-d0b7-436d-8f70-68452f898f73",
    "role": "character_ref"
   },
   {
    "label": "CHARACTER REFERENCE — 복만: the exact person appearing in this shot; match face, hair and build exactly.",
    "path": "<bytes:1027479>",
    "asset_id": "b7e69570-d15d-4241-b5ea-4cb8d4ece1cf",
    "role": "character_ref"
   },
   {
    "label": "CHARACTER REFERENCE — 오주일: the exact person appearing in this shot; match face, hair and build exactly.",
    "path": "<bytes:809716>",
    "asset_id": "7c31e825-ceef-4208-915f-71c4f0e15dd6",
    "role": "character_ref"
   },
   {
    "label": "CHARACTER REFERENCE — 임정현: the exact person appearing in this shot; match face, hair and build exactly.",
    "path": "<bytes:858188>",
    "asset_id": "789c3359-d7f8-4832-9a6f-7e4ad2695c02",
    "role": "character_ref"
   },
   {
    "label": "CHARACTER REFERENCE — 백정: the exact person appearing in this shot; match face, hair and build exactly.",
    "path": "<bytes:980360>",
    "asset_id": "29282584-ae7c-4722-8d77-100c4a5b2a3d",
    "role": "character_ref"
   },
   {
    "label": "CHARACTER REFERENCE — 목수: the exact person appearing in this shot; match face, hair and build exactly.",
    "path": "<bytes:904990>",
    "asset_id": "6b02463a-e818-4e3d-9816-215435a7aa0e",
    "role": "character_ref"
   }
  ],
  "critique_skipped": true,
  "shot_run_produced": false,
  "shot_run_uid": "06aa1435-428b-7bd3-a256-fe0c35eb9076",
  "ref_mode": "prev+엔티티",
  "share_plan": {
   "ref_plan": "prev",
   "prev_anchor_tag": "S46sh3"
  }
 },
 "S46sh16::cine": {
  "applied": true,
  "attempted_at": "2026-09-09T06:16:26.201100+00:00",
  "fingerprint": "b5a0f35c4634419ba6b07750529b140d586987fa1ba4b6ff72450c58c616805f",
  "fingerprint_version": 2,
  "provider": "mai",
  "endpoint": "openrouter/chat-completions",
  "model": "microsoft/mai-image-2.6",
  "pack": "24.202608252115",
  "source_file": "S46sh16_sel.png",
  "source_sha256": "d718dc1673243c9ab9ee9178d200e7ed979085da6f754cf6d78da918db359fe2",
  "file": "S46sh16_cine.png",
  "staged_sha256": "ef917930095fc8a450a393a91d36a29f21ad5952116e20a3f404fc8ed1982134",
  "latency_ms": 28705
 },
 "S47sh3::signage": {
  "fp": "a8f47b27a449074d",
  "inscriptions": [],
  "cues": [],
  "dropped": []
 },
 "S47sh3::bgfirst_bg": {
  "input_fingerprint": "2ff5bf6b335b5cc6",
  "prompt": "Create the EMPTY BACKGROUND PLATE for one film shot — NO PEOPLE, no figures, no body parts, no sketch lines, no arrows anywhere.\n\"Empty\" means no people only: KEEP the location's inherent occupants and stock that define the place — animals in an animal shelter, pen or farm, goods in a market, moored boats in a harbour — unless the shot text explicitly removes them.\nThe FIRST attached image is a thin-line storyboard sketch: use ONLY its camera angle, horizon, perspective and the placement/size of buildings and set masses — ignore the sketched people and arrows entirely. The SECOND attached image (LOCATION PHOTOGRAPH) is the real place: take its architecture, materials, signage and fixed features, and RE-PROJECT them into the sketch's camera. If the photograph's camera differs from the sketch's, the sketch's camera wins.\nHUMAN-SCALE CALIBRATION: derive every structure's true size from human-scale elements — a door ≈ 2m, a window ≈ 1–1.5m wide, one storey ≈ 2.5–3m; never inflate a small structure or shrink a large one.\n\nSHOT TEXT this background must serve (Korean): 쌀자루와 반찬이 실린 지게를 멘 조선인 짐꾼들과 함께 숲길을 오르는 mid-stride 자세의 박상현의 전신.\n\nLOCATION (lock): The uphill forest approach immediately below the mountain camp, along the porters' morning arrival route.\n\nTIME OF DAY (lock): morning.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: wide shot\n- FRAME LAYOUT: uphill continuation toward the camp in the upper-right of the frame, background.\n- KEY BACKGROUND ELEMENTS: 숲속 산길 (박상현과 짐꾼 둘이 산채 쪽으로 올라오고 있다) — The visible route runs diagonally uphill from lower left to upper right; used as Establishes the direction and effort of the arrival; 짐꾼들의 지게 (쌀자루와 반찬 꾸러미, 술 한 병을 나누어 싣고 있다) — Seen from the side and rear, with the loads visible above and behind the porters' shoulders; used as Explains the visitors' apparent purpose while preserving full-body scale.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Natural morning illumination with muted saturation and grounded contrast keeps the arrival legible without specifying an unsupported light source or atmospheric effect.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nRender ONE photorealistic empty location photograph, 16:9, that this shot can be staged inside later. No readable writing anywhere: surfaces that would carry writing may be present, but stage any wording out of legibility — an oblique angle, distance, shallow focus. No captions, watermarks or overlay text.",
  "effective_prompt": "Create the EMPTY BACKGROUND PLATE for one film shot — NO PEOPLE, no figures, no body parts, no sketch lines, no arrows anywhere.\n\"Empty\" means no people only: KEEP the location's inherent occupants and stock that define the place — animals in an animal shelter, pen or farm, goods in a market, moored boats in a harbour — unless the shot text explicitly removes them.\nThe FIRST attached image is a thin-line storyboard sketch: use ONLY its camera angle, horizon, perspective and the placement/size of buildings and set masses — ignore the sketched people and arrows entirely. The SECOND attached image (LOCATION PHOTOGRAPH) is the real place: take its architecture, materials, signage and fixed features, and RE-PROJECT them into the sketch's camera. If the photograph's camera differs from the sketch's, the sketch's camera wins.\nHUMAN-SCALE CALIBRATION: derive every structure's true size from human-scale elements — a door ≈ 2m, a window ≈ 1–1.5m wide, one storey ≈ 2.5–3m; never inflate a small structure or shrink a large one.\n\nSHOT TEXT this background must serve (Korean): 쌀자루와 반찬이 실린 지게를 멘 조선인 짐꾼들과 함께 숲길을 오르는 mid-stride 자세의 박상현의 전신.\n\nLOCATION (lock): The uphill forest approach immediately below the mountain camp, along the porters' morning arrival route.\n\nTIME OF DAY (lock): morning.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: wide shot\n- FRAME LAYOUT: uphill continuation toward the camp in the upper-right of the frame, background.\n- KEY BACKGROUND ELEMENTS: 숲속 산길 (박상현과 짐꾼 둘이 산채 쪽으로 올라오고 있다) — The visible route runs diagonally uphill from lower left to upper right; used as Establishes the direction and effort of the arrival; 짐꾼들의 지게 (쌀자루와 반찬 꾸러미, 술 한 병을 나누어 싣고 있다) — Seen from the side and rear, with the loads visible above and behind the porters' shoulders; used as Explains the visitors' apparent purpose while preserving full-body scale.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Natural morning illumination with muted saturation and grounded contrast keeps the arrival legible without specifying an unsupported light source or atmospheric effect.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nRender ONE photorealistic empty location photograph, 16:9, that this shot can be staged inside later. No readable writing anywhere: surfaces that would carry writing may be present, but stage any wording out of legibility — an oblique angle, distance, shallow focus. No captions, watermarks or overlay text.",
  "bg_path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/db00517b-165b-4005-a287-f28936b59b95/images/272bd7af-9294-4490-98c6-88a40e9deacc/scene/recipe/S47sh3__bgfirst_bg.png",
  "asset_id": "1b70e544-b0b3-4b29-95e4-4170a384c845",
  "input_asset_ids": [
   "4e1e988e-f16d-4e34-a0ee-4d33e37f1724",
   "69bb9661-49a8-4a75-a056-1e48e42c72b5"
  ]
 },
 "S47sh3": {
  "input_fingerprint": "cd6284af19b5ccb7",
  "prompt": "Create ONE FINAL photorealistic live-action film still of the moment below — early 20th-century Korea, Jeolla region; all people are Korean unless explicitly stated otherwise. TIME OF DAY (lock): morning.\n\nSHOT TEXT (authoritative, Korean): 쌀자루와 반찬이 실린 지게를 멘 조선인 짐꾼들과 함께 숲길을 오르는 mid-stride 자세의 박상현의 전신.\n\nLOCATION (lock): The uphill forest approach immediately below the mountain camp, along the porters' morning arrival route. The shot takes place here — the attached LOCATION PHOTOGRAPH shows the exact spot.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: wide shot\n- FRAME LAYOUT: uphill continuation toward the camp in the upper-right of the frame, background.\n- KEY BACKGROUND ELEMENTS: 숲속 산길 (박상현과 짐꾼 둘이 산채 쪽으로 올라오고 있다) — The visible route runs diagonally uphill from lower left to upper right; used as Establishes the direction and effort of the arrival; 짐꾼들의 지게 (쌀자루와 반찬 꾸러미, 술 한 병을 나누어 싣고 있다) — Seen from the side and rear, with the loads visible above and behind the porters' shoulders; used as Explains the visitors' apparent purpose while preserving full-body scale.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Natural morning illumination with muted saturation and grounded contrast keeps the arrival legible without specifying an unsupported light source or atmospheric effect.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): Morning light fills the camp approach. 박상현: He walks into the camp from the mountain path.\n\nTHE HAND THAT IS DOING THIS: the object at the centre of this shot is being held, operated, read, aimed or handed over by 조선인 짐꾼들 right now, so 조선인 짐꾼들's hand — and as much of the wrist and forearm as the framing reaches — is in the frame, gripping or resting on that object exactly the way its real use requires. Match that hand to 조선인 짐꾼들: its size, build, skin, age, grooming, sleeve and anything worn on it belong to that person and to no one else.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 박상현 (한국인 남성, 성인 나이대, 자연스러운 검은 머리) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.",
  "roll_prompts": {
   "A": "Stage the shot. The FIRST attached image (SHOT BACKGROUND) is the finished empty background of this shot — keep it EXACTLY: its camera, perspective, architecture, lighting and every fixed feature stay untouched. The SECOND attached image (LAYOUT SKETCH) tells you ONLY where the people go: each sketched person's position, screen size, pose and the gaze/motion arrows. Ignore the sketch's background lines. The CHARACTER REFERENCE photographs show the real people.\nPlace the real people into the background at exactly the sketched positions, sizes and poses, following the arrow directions. No sketch lines or arrows may remain.\n\nCreate ONE FINAL photorealistic live-action film still of the moment below — early 20th-century Korea, Jeolla region; all people are Korean unless explicitly stated otherwise. TIME OF DAY (lock): morning.\n\nSHOT TEXT (authoritative, Korean): 쌀자루와 반찬이 실린 지게를 멘 조선인 짐꾼들과 함께 숲길을 오르는 mid-stride 자세의 박상현의 전신.\n\nLOCATION (lock): The uphill forest approach immediately below the mountain camp, along the porters' morning arrival route. The shot takes place here — the FIRST attached image (SHOT BACKGROUND) is this exact place, already built: its ground, structures, horizon, materials and lighting are the finished truth of this location and must not be redesigned or replaced. No location photograph is attached — read the place from that image alone, and add no scenery, structure, vehicle or fixture that it does not already show. This lock governs the place only; the figures in the shot follow the staging and pose instructions.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: wide shot\n- FRAME LAYOUT: uphill continuation toward the camp in the upper-right of the frame, background.\n- KEY BACKGROUND ELEMENTS: 숲속 산길 (박상현과 짐꾼 둘이 산채 쪽으로 올라오고 있다) — The visible route runs diagonally uphill from lower left to upper right; used as Establishes the direction and effort of the arrival; 짐꾼들의 지게 (쌀자루와 반찬 꾸러미, 술 한 병을 나누어 싣고 있다) — Seen from the side and rear, with the loads visible above and behind the porters' shoulders; used as Explains the visitors' apparent purpose while preserving full-body scale.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Natural morning illumination with muted saturation and grounded contrast keeps the arrival legible without specifying an unsupported light source or atmospheric effect.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): Morning light fills the camp approach. 박상현: He walks into the camp from the mountain path.\n\nTHE HAND THAT IS DOING THIS: the object at the centre of this shot is being held, operated, read, aimed or handed over by 조선인 짐꾼들 right now, so 조선인 짐꾼들's hand — and as much of the wrist and forearm as the framing reaches — is in the frame, gripping or resting on that object exactly the way its real use requires. Match that hand to 조선인 짐꾼들: its size, build, skin, age, grooming, sleeve and anything worn on it belong to that person and to no one else.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 박상현 (한국인 남성, 성인 나이대, 자연스러운 검은 머리) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.",
   "B": "Create ONE FINAL photorealistic live-action film still of the moment below — early 20th-century Korea, Jeolla region; all people are Korean unless explicitly stated otherwise. TIME OF DAY (lock): morning.\n\nSHOT TEXT (authoritative, Korean): 쌀자루와 반찬이 실린 지게를 멘 조선인 짐꾼들과 함께 숲길을 오르는 mid-stride 자세의 박상현의 전신.\n\nLOCATION (lock): The uphill forest approach immediately below the mountain camp, along the porters' morning arrival route. The shot takes place here — the attached LOCATION PHOTOGRAPH shows the exact spot.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: wide shot\n- FRAME LAYOUT: uphill continuation toward the camp in the upper-right of the frame, background.\n- KEY BACKGROUND ELEMENTS: 숲속 산길 (박상현과 짐꾼 둘이 산채 쪽으로 올라오고 있다) — The visible route runs diagonally uphill from lower left to upper right; used as Establishes the direction and effort of the arrival; 짐꾼들의 지게 (쌀자루와 반찬 꾸러미, 술 한 병을 나누어 싣고 있다) — Seen from the side and rear, with the loads visible above and behind the porters' shoulders; used as Explains the visitors' apparent purpose while preserving full-body scale.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Natural morning illumination with muted saturation and grounded contrast keeps the arrival legible without specifying an unsupported light source or atmospheric effect.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): Morning light fills the camp approach. 박상현: He walks into the camp from the mountain path.\n\nTHE HAND THAT IS DOING THIS: the object at the centre of this shot is being held, operated, read, aimed or handed over by 조선인 짐꾼들 right now, so 조선인 짐꾼들's hand — and as much of the wrist and forearm as the framing reaches — is in the frame, gripping or resting on that object exactly the way its real use requires. Match that hand to 조선인 짐꾼들: its size, build, skin, age, grooming, sleeve and anything worn on it belong to that person and to no one else.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 박상현 (한국인 남성, 성인 나이대, 자연스러운 검은 머리) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.\n\nCOMPOSITION VARIATION (this candidate only): keep the place, the people and the moment exactly as specified, and keep from the CAMERA & FRAME contract its subject scale, screen placement and key background elements exactly. For this candidate only, its camera angle and height are deliberately left open: choose a camera position distinctly different from the obvious one, as a film director picking a second setup on the same blocking. Stay true to any explicit framing the SHOT TEXT itself dictates."
  },
  "roll_refs": {
   "A": [
    {
     "label": "SHOT BACKGROUND",
     "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/db00517b-165b-4005-a287-f28936b59b95/images/272bd7af-9294-4490-98c6-88a40e9deacc/scene/recipe/S47sh3__bgfirst_bg.png",
     "asset_id": "1b70e544-b0b3-4b29-95e4-4170a384c845",
     "role": "bgfirst_bg"
    },
    {
     "label": "LAYOUT SKETCH (people placement only)",
     "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/db00517b-165b-4005-a287-f28936b59b95/images/272bd7af-9294-4490-98c6-88a40e9deacc/conti/conti_S47sh3.png",
     "asset_id": "4e1e988e-f16d-4e34-a0ee-4d33e37f1724",
     "role": "conti_light"
    },
    {
     "label": "CHARACTER REFERENCE — 박상현: the exact person appearing in this shot; match face, hair and build exactly.",
     "path": "<bytes:840781>",
     "asset_id": "6187751d-5df3-4eb7-9d30-d6fc98a005e6",
     "role": "character_ref"
    },
    {
     "label": "PROP REFERENCE — 지게: the exact object appearing in this shot; match its look, material and wear exactly.",
     "path": "<bytes:1157225>",
     "asset_id": "bdfa122b-f62f-4ee2-807f-3e7997ed8125",
     "role": "prop_ref"
    },
    {
     "label": "PROP REFERENCE — 보자기 꾸러미·봇짐: the exact object appearing in this shot; match its look, material and wear exactly.",
     "path": "<bytes:1465734>",
     "asset_id": "3ca92b24-68b4-4883-b851-cabc06c00d1a",
     "role": "prop_ref"
    },
    {
     "label": "PROP REFERENCE — 곡식 자루·가마니: the exact object appearing in this shot; match its look, material and wear exactly.",
     "path": "<bytes:1807398>",
     "asset_id": "e30c319e-fdc7-4c01-a301-290ddd45d27a",
     "role": "prop_ref"
    }
   ],
   "B": [
    {
     "label": "LOCATION PHOTOGRAPH — the exact place of this shot: its architecture, materials, fixed features and lighting mood are spatial truth; stage the moment inside this place. Never copy its camera framing.",
     "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/db00517b-165b-4005-a287-f28936b59b95/images/272bd7af-9294-4490-98c6-88a40e9deacc/scene/recipe/groupbg_camp_supply_approach_8601af.png",
     "asset_id": "69bb9661-49a8-4a75-a056-1e48e42c72b5",
     "role": "bgfirst_group_bg"
    },
    {
     "label": "CHARACTER REFERENCE — 박상현: the exact person appearing in this shot; match face, hair and build exactly.",
     "path": "<bytes:840781>",
     "asset_id": "6187751d-5df3-4eb7-9d30-d6fc98a005e6",
     "role": "character_ref"
    },
    {
     "label": "PROP REFERENCE — 지게: the exact object appearing in this shot; match its look, material and wear exactly.",
     "path": "<bytes:1157225>",
     "asset_id": "bdfa122b-f62f-4ee2-807f-3e7997ed8125",
     "role": "prop_ref"
    },
    {
     "label": "PROP REFERENCE — 보자기 꾸러미·봇짐: the exact object appearing in this shot; match its look, material and wear exactly.",
     "path": "<bytes:1465734>",
     "asset_id": "3ca92b24-68b4-4883-b851-cabc06c00d1a",
     "role": "prop_ref"
    },
    {
     "label": "PROP REFERENCE — 곡식 자루·가마니: the exact object appearing in this shot; match its look, material and wear exactly.",
     "path": "<bytes:1807398>",
     "asset_id": "e30c319e-fdc7-4c01-a301-290ddd45d27a",
     "role": "prop_ref"
    }
   ]
  },
  "shot_run_spend_attempt_count": 1,
  "judge_flip_skipped": "judge_fn_owns_order",
  "readings": [
   {
    "label": "A",
    "direction": "박상현과 두 짐꾼 모두 화면 우측 상단의 산채를 향해 오르막길을 걷고 있으며, 인물들의 시선과 이동 방향이 목적지를 향해 정확히 일치함.",
    "built_space": "배경의 산채 목책, 지붕 구조물은 물론 우측에 짐이 실려 있는 수레까지 장소 참조 사진의 요소와 배치를 완벽하고 정확하게 재현함.",
    "entities": "박상현은 회색 한복을 입고 얼굴이 참조와 일치하나 갓을 쓰지 않음. 짐꾼들은 지게에 쌀자루와 보자기 꾸러미를 실었으나 지시된 술병은 보이지 않음.",
    "hard_violations": [],
    "physics": "세 사람 모두 오르막을 오르는 mid-stride 자세에 맞게 지면을 안정적으로 딛고 체중을 싣고 있으며, 지게와 짐, 손에 든 꾸러미 모두 신체에 의해 정상적으로 지탱됨."
   },
   {
    "label": "B",
    "direction": "두 짐꾼은 오르막을 오르고 있으나, 박상현은 이들을 마주 보며 내리막길 방향으로 서 있어 '함께 오르는' 텍스트 지시와 정반대로 향하고 있음.",
    "built_space": "산채 구조물과 목책은 장소 참조와 대체로 일치하지만, 원본 장소 우측에 명확히 존재하던 수레가 배경에서 생략됨.",
    "entities": "박상현은 얼굴과 회색 한복, 갓까지 참조와 모두 일치함. 짐꾼들의 지게에는 쌀자루, 보자기 꾸러미, 술병이 묘사됨.",
    "hard_violations": [
     "[gemini-pro] 텍스트는 짐꾼들과 '함께 오르막길을 오르는' 박상현을 명시했으나, 해당 후보는 박상현이 내리막 방향으로 마주 선 모습을 렌더링하여 경로 및 방향성을 반대로 왜곡함 (방향 지시 및 이동 동선 위반)."
    ],
    "physics": "인물들은 지면에 발을 딛고 있으나 박상현은 걷는 자세(mid-stride)가 아닌 멈춰 선 자세임. 짐꾼들의 지게가 A자 형태가 아니라 네 개의 다리를 가진 접이식 사다리 형태로 물리적으로 왜곡되어 묘사됨."
   }
  ],
  "cross_model_order": {
   "policy": "cross_model_order_v2_slot0_fwd_slot1_rev_parallel2",
   "slots": [
    {
     "model": "gemini-pro",
     "order": "forward",
     "display_to_canonical": {
      "A": "A",
      "B": "B"
     },
     "raw": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 8,
        "verdict_ko": "박상현의 갓과 지게의 술병이 누락되는 아쉬움이 있으나, 일행이 함께 오르막길을 오르는 핵심 행동 지시와 장소 참조 이미지의 수레까지 완벽하게 구현하여 훌륭한 결과물입니다."
       },
       {
        "label": "B",
        "score": 3,
        "verdict_ko": "인물의 복장과 소품 구성은 충실히 반영했으나, 박상현이 일행과 반대되는 내리막길 방향으로 렌더링되어 프롬프트의 가장 중요한 동선 지시를 심각하게 위반했습니다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "박상현과 두 짐꾼 모두 화면 우측 상단의 산채를 향해 오르막길을 걷고 있으며, 인물들의 시선과 이동 방향이 목적지를 향해 정확히 일치함.",
        "built_space": "배경의 산채 목책, 지붕 구조물은 물론 우측에 짐이 실려 있는 수레까지 장소 참조 사진의 요소와 배치를 완벽하고 정확하게 재현함.",
        "entities": "박상현은 회색 한복을 입고 얼굴이 참조와 일치하나 갓을 쓰지 않음. 짐꾼들은 지게에 쌀자루와 보자기 꾸러미를 실었으나 지시된 술병은 보이지 않음.",
        "hard_violations": [],
        "physics": "세 사람 모두 오르막을 오르는 mid-stride 자세에 맞게 지면을 안정적으로 딛고 체중을 싣고 있으며, 지게와 짐, 손에 든 꾸러미 모두 신체에 의해 정상적으로 지탱됨."
       },
       {
        "label": "B",
        "direction": "두 짐꾼은 오르막을 오르고 있으나, 박상현은 이들을 마주 보며 내리막길 방향으로 서 있어 '함께 오르는' 텍스트 지시와 정반대로 향하고 있음.",
        "built_space": "산채 구조물과 목책은 장소 참조와 대체로 일치하지만, 원본 장소 우측에 명확히 존재하던 수레가 배경에서 생략됨.",
        "entities": "박상현은 얼굴과 회색 한복, 갓까지 참조와 모두 일치함. 짐꾼들의 지게에는 쌀자루, 보자기 꾸러미, 술병이 묘사됨.",
        "hard_violations": [
         "텍스트는 짐꾼들과 '함께 오르막길을 오르는' 박상현을 명시했으나, 해당 후보는 박상현이 내리막 방향으로 마주 선 모습을 렌더링하여 경로 및 방향성을 반대로 왜곡함 (방향 지시 및 이동 동선 위반)."
        ],
        "physics": "인물들은 지면에 발을 딛고 있으나 박상현은 걷는 자세(mid-stride)가 아닌 멈춰 선 자세임. 짐꾼들의 지게가 A자 형태가 아니라 네 개의 다리를 가진 접이식 사다리 형태로 물리적으로 왜곡되어 묘사됨."
       }
      ],
      "all_candidates_fail": false
     },
     "normalized": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 8,
        "verdict_ko": "박상현의 갓과 지게의 술병이 누락되는 아쉬움이 있으나, 일행이 함께 오르막길을 오르는 핵심 행동 지시와 장소 참조 이미지의 수레까지 완벽하게 구현하여 훌륭한 결과물입니다."
       },
       {
        "label": "B",
        "score": 3,
        "verdict_ko": "인물의 복장과 소품 구성은 충실히 반영했으나, 박상현이 일행과 반대되는 내리막길 방향으로 렌더링되어 프롬프트의 가장 중요한 동선 지시를 심각하게 위반했습니다."
       }
      ],
      "all_candidates_fail": false,
      "readings": [
       {
        "label": "A",
        "direction": "박상현과 두 짐꾼 모두 화면 우측 상단의 산채를 향해 오르막길을 걷고 있으며, 인물들의 시선과 이동 방향이 목적지를 향해 정확히 일치함.",
        "built_space": "배경의 산채 목책, 지붕 구조물은 물론 우측에 짐이 실려 있는 수레까지 장소 참조 사진의 요소와 배치를 완벽하고 정확하게 재현함.",
        "entities": "박상현은 회색 한복을 입고 얼굴이 참조와 일치하나 갓을 쓰지 않음. 짐꾼들은 지게에 쌀자루와 보자기 꾸러미를 실었으나 지시된 술병은 보이지 않음.",
        "hard_violations": [],
        "physics": "세 사람 모두 오르막을 오르는 mid-stride 자세에 맞게 지면을 안정적으로 딛고 체중을 싣고 있으며, 지게와 짐, 손에 든 꾸러미 모두 신체에 의해 정상적으로 지탱됨."
       },
       {
        "label": "B",
        "direction": "두 짐꾼은 오르막을 오르고 있으나, 박상현은 이들을 마주 보며 내리막길 방향으로 서 있어 '함께 오르는' 텍스트 지시와 정반대로 향하고 있음.",
        "built_space": "산채 구조물과 목책은 장소 참조와 대체로 일치하지만, 원본 장소 우측에 명확히 존재하던 수레가 배경에서 생략됨.",
        "entities": "박상현은 얼굴과 회색 한복, 갓까지 참조와 모두 일치함. 짐꾼들의 지게에는 쌀자루, 보자기 꾸러미, 술병이 묘사됨.",
        "hard_violations": [
         "텍스트는 짐꾼들과 '함께 오르막길을 오르는' 박상현을 명시했으나, 해당 후보는 박상현이 내리막 방향으로 마주 선 모습을 렌더링하여 경로 및 방향성을 반대로 왜곡함 (방향 지시 및 이동 동선 위반)."
        ],
        "physics": "인물들은 지면에 발을 딛고 있으나 박상현은 걷는 자세(mid-stride)가 아닌 멈춰 선 자세임. 짐꾼들의 지게가 A자 형태가 아니라 네 개의 다리를 가진 접이식 사다리 형태로 물리적으로 왜곡되어 묘사됨."
       }
      ]
     },
     "ok": true
    },
    {
     "model": "gpt-high",
     "order": "reverse",
     "display_to_canonical": {
      "A": "B",
      "B": "A"
     },
     "raw": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 4,
        "verdict_ko": "전신 와이드 구도와 지게의 짐은 충실하지만, 박상현이 짐꾼들과 함께 오르지 않고 아래쪽을 향해 멈춰 서 있어 핵심 동작과 진행 방향이 어긋납니다."
       },
       {
        "label": "B",
        "score": 8,
        "verdict_ko": "세 사람이 좌하단에서 우상단으로 함께 오르는 전신 보행 순간과 산채 공간을 잘 구현했지만, 박상현의 갓이 빠졌고 술 한 병은 확인되지 않습니다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "짐꾼 둘은 등을 비스듬히 보이며 오른쪽 위 산채 방향으로 올라갑니다. 반면 박상현은 몸을 카메라와 내리막 쪽으로 돌리고 다가오는 짐꾼들을 바라봅니다. 길 자체는 좌하단에서 우상단으로 오르지만, 세 사람이 함께 올라가는 방향 관계는 성립하지 않습니다.",
        "built_space": "흙과 돌로 된 경사로 위쪽에 목책으로 둘러싸인 목조 건물 한 채와 오른쪽 개방형 부속 공간이 보입니다. 돌 받침 위 검은 솥 한 개가 뚜렷하고, 오른쪽 가장자리에는 수레 일부가 보입니다. 숲, 목책, 지붕과 바위의 재료는 장소 참조와 대체로 맞습니다. 짐꾼들은 아래쪽 길에 있고 박상현은 그보다 높은 곳에 마주 서 있습니다. 세 사람의 전신이 모두 들어오는 와이드 구도입니다.",
        "entities": "성인 한국인 남성으로 설정에 부합하게 보이는 인물은 박상현과 짐꾼 둘, 총 세 명입니다. 박상현의 검은 갓, 회색 도포, 허리띠와 얼굴의 수염은 참조에 대체로 부합합니다. 짐꾼 둘은 낡은 회색 계열 옷을 입고 나무와 밧줄로 만든 지게를 하나씩 멨습니다. 뒤쪽 짐꾼에게 갈색 자루, 회녹색 꾸러미와 병 한 개가 보이고, 앞쪽 짐꾼에게는 초록 보자기 꾸러미가 보입니다. 자루와 꾸러미의 내용물은 직접 확인할 수 없습니다. 병의 밝은 표면에 읽을 수 있는 글자는 없습니다.",
        "hard_violations": [],
        "physics": "짐꾼들의 손이 지게의 아래쪽 가로 부재를 잡고 있으며, 지게와 짐은 어깨끈 및 나무 받침에 지지됩니다. 앞쪽 짐꾼은 한 발을 땅에 딛고 다른 발을 뒤로 들어 보행 중으로 보입니다. 뒤쪽 짐꾼도 벌어진 발과 기울어진 몸으로 하중을 받습니다. 박상현은 두 발을 가까이 지면에 두고 정지해 있어 요구된 보행 중간 자세가 아닙니다. 지지 없이 떠 있는 인물이나 짐은 보이지 않습니다."
       },
       {
        "label": "B",
        "direction": "박상현이 오른쪽 위로 몸과 얼굴을 향하고 앞장서며, 짐꾼 둘도 같은 방향으로 뒤따릅니다. 세 사람의 시선은 진행할 산길 또는 바로 앞의 발밑을 향합니다. 길은 좌하단에서 우상단 배경으로 이어지고, 인물들의 진행 방향도 그 오르막과 일치합니다.",
        "built_space": "왼쪽 위에 목조 건물 한 채와 열린 부속 공간, 그 주변과 오른쪽 경사에 목책 구간들이 보입니다. 왼쪽 중경에는 검은 솥 두 개가 있는 돌 화덕 구역이 인물과 짐에 일부 가려져 있고, 오른쪽에는 짐을 실은 두 바퀴 수레 한 대가 있습니다. 큰 나무와 바위가 있는 오른쪽 둔덕까지 장소 참조의 공간 관계를 잘 유지합니다. 세 사람은 수레 왼쪽 산길을 따라 오르며, 길의 연속부는 우상단에 남아 있습니다. 전신을 포함한 와이드 구도이고 배경 시설의 크기도 자연스럽습니다.",
        "entities": "박상현과 짐꾼 둘, 총 세 명의 성인 남성이 보이며 외관과 시대 의상은 한국인 설정에 부합합니다. 박상현은 회색 도포와 검은 머리, 머리띠를 갖췄지만 참조의 갓은 쓰지 않았습니다. 옆얼굴만 보여 정확한 얼굴 일치 판단에는 한계가 있습니다. 짐꾼마다 나무 지게 한 개와 갈색 곡식 자루, 초록 보자기 꾸러미가 보입니다. 지게의 나무와 밧줄 재질은 맞지만 참조보다 간소한 형태입니다. 뒤쪽 짐꾼의 꾸러미는 지게 위가 아니라 몸 옆에 매달려 있습니다. 술 한 병은 확인되지 않으며 읽을 수 있는 글자는 없습니다.",
        "hard_violations": [],
        "physics": "박상현은 앞발로 지면을 딛고 뒤쪽 발꿈치를 들어 다음 걸음을 옮기는 자세입니다. 짐꾼들도 각각 지지발과 뒤로 들리거나 앞으로 나아가는 발이 보여 오르막 보행으로 성립합니다. 지게는 어깨끈과 등에 지지되고 자루와 꾸러미는 받침과 묶음줄에 걸려 있습니다. 짐꾼들의 손과 손목은 어깨 앞 운반용 부재나 끈을 붙잡는 위치에 보입니다. 수레는 바퀴로 지면에 놓여 있고 적재물은 짐칸에 받쳐져 있습니다. 지지 없이 떠 있는 물체는 보이지 않습니다."
       }
      ],
      "all_candidates_fail": false
     },
     "normalized": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 4,
        "verdict_ko": "전신 와이드 구도와 지게의 짐은 충실하지만, 박상현이 짐꾼들과 함께 오르지 않고 아래쪽을 향해 멈춰 서 있어 핵심 동작과 진행 방향이 어긋납니다."
       },
       {
        "label": "A",
        "score": 8,
        "verdict_ko": "세 사람이 좌하단에서 우상단으로 함께 오르는 전신 보행 순간과 산채 공간을 잘 구현했지만, 박상현의 갓이 빠졌고 술 한 병은 확인되지 않습니다."
       }
      ],
      "all_candidates_fail": false,
      "readings": [
       {
        "label": "B",
        "direction": "짐꾼 둘은 등을 비스듬히 보이며 오른쪽 위 산채 방향으로 올라갑니다. 반면 박상현은 몸을 카메라와 내리막 쪽으로 돌리고 다가오는 짐꾼들을 바라봅니다. 길 자체는 좌하단에서 우상단으로 오르지만, 세 사람이 함께 올라가는 방향 관계는 성립하지 않습니다.",
        "built_space": "흙과 돌로 된 경사로 위쪽에 목책으로 둘러싸인 목조 건물 한 채와 오른쪽 개방형 부속 공간이 보입니다. 돌 받침 위 검은 솥 한 개가 뚜렷하고, 오른쪽 가장자리에는 수레 일부가 보입니다. 숲, 목책, 지붕과 바위의 재료는 장소 참조와 대체로 맞습니다. 짐꾼들은 아래쪽 길에 있고 박상현은 그보다 높은 곳에 마주 서 있습니다. 세 사람의 전신이 모두 들어오는 와이드 구도입니다.",
        "entities": "성인 한국인 남성으로 설정에 부합하게 보이는 인물은 박상현과 짐꾼 둘, 총 세 명입니다. 박상현의 검은 갓, 회색 도포, 허리띠와 얼굴의 수염은 참조에 대체로 부합합니다. 짐꾼 둘은 낡은 회색 계열 옷을 입고 나무와 밧줄로 만든 지게를 하나씩 멨습니다. 뒤쪽 짐꾼에게 갈색 자루, 회녹색 꾸러미와 병 한 개가 보이고, 앞쪽 짐꾼에게는 초록 보자기 꾸러미가 보입니다. 자루와 꾸러미의 내용물은 직접 확인할 수 없습니다. 병의 밝은 표면에 읽을 수 있는 글자는 없습니다.",
        "hard_violations": [],
        "physics": "짐꾼들의 손이 지게의 아래쪽 가로 부재를 잡고 있으며, 지게와 짐은 어깨끈 및 나무 받침에 지지됩니다. 앞쪽 짐꾼은 한 발을 땅에 딛고 다른 발을 뒤로 들어 보행 중으로 보입니다. 뒤쪽 짐꾼도 벌어진 발과 기울어진 몸으로 하중을 받습니다. 박상현은 두 발을 가까이 지면에 두고 정지해 있어 요구된 보행 중간 자세가 아닙니다. 지지 없이 떠 있는 인물이나 짐은 보이지 않습니다."
       },
       {
        "label": "A",
        "direction": "박상현이 오른쪽 위로 몸과 얼굴을 향하고 앞장서며, 짐꾼 둘도 같은 방향으로 뒤따릅니다. 세 사람의 시선은 진행할 산길 또는 바로 앞의 발밑을 향합니다. 길은 좌하단에서 우상단 배경으로 이어지고, 인물들의 진행 방향도 그 오르막과 일치합니다.",
        "built_space": "왼쪽 위에 목조 건물 한 채와 열린 부속 공간, 그 주변과 오른쪽 경사에 목책 구간들이 보입니다. 왼쪽 중경에는 검은 솥 두 개가 있는 돌 화덕 구역이 인물과 짐에 일부 가려져 있고, 오른쪽에는 짐을 실은 두 바퀴 수레 한 대가 있습니다. 큰 나무와 바위가 있는 오른쪽 둔덕까지 장소 참조의 공간 관계를 잘 유지합니다. 세 사람은 수레 왼쪽 산길을 따라 오르며, 길의 연속부는 우상단에 남아 있습니다. 전신을 포함한 와이드 구도이고 배경 시설의 크기도 자연스럽습니다.",
        "entities": "박상현과 짐꾼 둘, 총 세 명의 성인 남성이 보이며 외관과 시대 의상은 한국인 설정에 부합합니다. 박상현은 회색 도포와 검은 머리, 머리띠를 갖췄지만 참조의 갓은 쓰지 않았습니다. 옆얼굴만 보여 정확한 얼굴 일치 판단에는 한계가 있습니다. 짐꾼마다 나무 지게 한 개와 갈색 곡식 자루, 초록 보자기 꾸러미가 보입니다. 지게의 나무와 밧줄 재질은 맞지만 참조보다 간소한 형태입니다. 뒤쪽 짐꾼의 꾸러미는 지게 위가 아니라 몸 옆에 매달려 있습니다. 술 한 병은 확인되지 않으며 읽을 수 있는 글자는 없습니다.",
        "hard_violations": [],
        "physics": "박상현은 앞발로 지면을 딛고 뒤쪽 발꿈치를 들어 다음 걸음을 옮기는 자세입니다. 짐꾼들도 각각 지지발과 뒤로 들리거나 앞으로 나아가는 발이 보여 오르막 보행으로 성립합니다. 지게는 어깨끈과 등에 지지되고 자루와 꾸러미는 받침과 묶음줄에 걸려 있습니다. 짐꾼들의 손과 손목은 어깨 앞 운반용 부재나 끈을 붙잡는 위치에 보입니다. 수레는 바퀴로 지면에 놓여 있고 적재물은 짐칸에 받쳐져 있습니다. 지지 없이 떠 있는 물체는 보이지 않습니다."
       }
      ]
     },
     "ok": true
    }
   ],
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "slot_winner_match": true,
   "slot_winner": {
    "gemini-pro": "A",
    "gpt-high": "A"
   },
   "route": "cross_slot_agree"
  },
  "dual": {
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "normalized": {
    "A": 2.0,
    "B": 0.875
   },
   "adjusted": {
    "A": 2.0,
    "B": 0.625
   },
   "violations": {
    "B": [
     "[gemini-pro] 텍스트는 짐꾼들과 '함께 오르막길을 오르는' 박상현을 명시했으나, 해당 후보는 박상현이 내리막 방향으로 마주 선 모습을 렌더링하여 경로 및 방향성을 반대로 왜곡함 (방향 지시 및 이동 동선 위반)."
    ]
   },
   "per_model_winner": {
    "gemini-pro": "A",
    "gpt-high": "A"
   },
   "agreed": true
  },
  "totals": {
   "A": 2000,
   "B": 625
  },
  "selected": "A",
  "ranking": [
   "A",
   "B"
  ],
  "verdicts": [
   {
    "label": "A",
    "score": 2000,
    "verdict_ko": "박상현의 갓과 지게의 술병이 누락되는 아쉬움이 있으나, 일행이 함께 오르막길을 오르는 핵심 행동 지시와 장소 참조 이미지의 수레까지 완벽하게 구현하여 훌륭한 결과물입니다."
   },
   {
    "label": "B",
    "score": 625,
    "verdict_ko": "인물의 복장과 소품 구성은 충실히 반영했으나, 박상현이 일행과 반대되는 내리막길 방향으로 렌더링되어 프롬프트의 가장 중요한 동선 지시를 심각하게 위반했습니다.  ★위반: [gemini-pro] 텍스트는 짐꾼들과 '함께 오르막길을 오르는' 박상현을 명시했으나, 해당 후보는 박상현이 내리막 방향으로 마주 선 모습을 렌더링하여 경로 및 방향성을 반대로 왜곡함 (방향 지시 및 이동 동선 위반)."
   }
  ],
  "refs": [
   {
    "label": "LOCATION PHOTOGRAPH — the exact place of this shot: its architecture, materials, fixed features and lighting mood are spatial truth; stage the moment inside this place. Never copy its camera framing.",
    "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/db00517b-165b-4005-a287-f28936b59b95/images/272bd7af-9294-4490-98c6-88a40e9deacc/scene/recipe/groupbg_camp_supply_approach_8601af.png",
    "asset_id": "69bb9661-49a8-4a75-a056-1e48e42c72b5",
    "role": "bgfirst_group_bg"
   },
   {
    "label": "CHARACTER REFERENCE — 박상현: the exact person appearing in this shot; match face, hair and build exactly.",
    "path": "<bytes:840781>",
    "asset_id": "6187751d-5df3-4eb7-9d30-d6fc98a005e6",
    "role": "character_ref"
   },
   {
    "label": "PROP REFERENCE — 지게: the exact object appearing in this shot; match its look, material and wear exactly.",
    "path": "<bytes:1157225>",
    "asset_id": "bdfa122b-f62f-4ee2-807f-3e7997ed8125",
    "role": "prop_ref"
   },
   {
    "label": "PROP REFERENCE — 보자기 꾸러미·봇짐: the exact object appearing in this shot; match its look, material and wear exactly.",
    "path": "<bytes:1465734>",
    "asset_id": "3ca92b24-68b4-4883-b851-cabc06c00d1a",
    "role": "prop_ref"
   },
   {
    "label": "PROP REFERENCE — 곡식 자루·가마니: the exact object appearing in this shot; match its look, material and wear exactly.",
    "path": "<bytes:1807398>",
    "asset_id": "e30c319e-fdc7-4c01-a301-290ddd45d27a",
    "role": "prop_ref"
   }
  ],
  "critique_skipped": true,
  "shot_run_produced": false,
  "shot_run_uid": "06aa1435-446c-7fd8-a088-03b5edfd55cd",
  "bgfirst": {
   "bg_path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/db00517b-165b-4005-a287-f28936b59b95/images/272bd7af-9294-4490-98c6-88a40e9deacc/scene/recipe/S47sh3__bgfirst_bg.png",
   "bg_asset_id": "1b70e544-b0b3-4b29-95e4-4170a384c845",
   "bg_record_key": "S47sh3::bgfirst_bg",
   "chain_winner": true,
   "authority": "groupbg",
   "group_key": "camp_supply_approach",
   "groupbg_asset_id": "69bb9661-49a8-4a75-a056-1e48e42c72b5"
  },
  "ref_mode": "재투영 배경+콘티+엔티티 (2택1: 체인 승)",
  "share_plan": {
   "ref_plan": "background"
  }
 },
 "S47sh3::cine": {
  "applied": true,
  "attempted_at": "2026-09-09T06:20:10.278446+00:00",
  "fingerprint": "3caccce51d15063ef313b6767ec306f06536ddf705756c3f953d24afeffdde12",
  "fingerprint_version": 2,
  "provider": "mai",
  "endpoint": "openrouter/chat-completions",
  "model": "microsoft/mai-image-2.6",
  "pack": "24.202608252115",
  "source_file": "S47sh3_sel.png",
  "source_sha256": "4a2572c5c16f6c45c00da6a64c79162fa7e7d93ab3ba48495e51d61e994ef7ce",
  "file": "S47sh3_cine.png",
  "staged_sha256": "e4d335bbe1d1167688bc698e41107905a96fb5fa09a3c8e600f89559fd19fd75",
  "latency_ms": 29556
 },
 "S47sh13::signage": {
  "fp": "a8371c81394a81ad",
  "inscriptions": [],
  "cues": [],
  "dropped": []
 },
 "S47sh13": {
  "input_fingerprint": "a9eab9640264e5cb",
  "prompt": "Create ONE FINAL photorealistic live-action film still of the moment below — early 20th-century Korea, Jeolla region; all people are Korean unless explicitly stated otherwise. TIME OF DAY (lock): morning.\n\nSHOT TEXT (authoritative, Korean): 쌀자루 위에 손을 멈춘 채 박상현을 향해 고개를 홱 돌린 임정현의 굳은 얼굴.\n\nLOCATION (lock): The outdoor receiving spot beside the newly delivered rice sacks at the mountain camp's entrance. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: medium shot\n- KEY BACKGROUND ELEMENTS: 쌀자루 (임정현이 세다가 그 위에 손을 멈추었다) — The upper portion lies beneath his hand at the lower frame edge; used as Anchors the interrupted task without foreground size exaggeration.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Maintain neutral morning illumination and muted natural contrast, letting the stopped hand and tightened expression carry the unease.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The clothing, hair and overall look of 임정현 — who appear both in that photo and in this shot — are LOCKED to that photo. Anyone else visible in that photo is NOT in this shot: never carry their face, body or clothing onto anyone here. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The delivered rice sacks, side-dish bundles and liquor bottle are now being received in camp. 임정현: He pauses with his hand over the rice sacks during counting.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 임정현 (한국인 남성, 성인 나이대, 자연스러운 검은 머리) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.",
  "roll_prompts": {
   "A": "Create ONE FINAL photorealistic live-action film still of the moment below — early 20th-century Korea, Jeolla region; all people are Korean unless explicitly stated otherwise. TIME OF DAY (lock): morning.\n\nSHOT TEXT (authoritative, Korean): 쌀자루 위에 손을 멈춘 채 박상현을 향해 고개를 홱 돌린 임정현의 굳은 얼굴.\n\nLOCATION (lock): The outdoor receiving spot beside the newly delivered rice sacks at the mountain camp's entrance. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: medium shot\n- KEY BACKGROUND ELEMENTS: 쌀자루 (임정현이 세다가 그 위에 손을 멈추었다) — The upper portion lies beneath his hand at the lower frame edge; used as Anchors the interrupted task without foreground size exaggeration.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Maintain neutral morning illumination and muted natural contrast, letting the stopped hand and tightened expression carry the unease.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The clothing, hair and overall look of 임정현 — who appear both in that photo and in this shot — are LOCKED to that photo. Anyone else visible in that photo is NOT in this shot: never carry their face, body or clothing onto anyone here. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The delivered rice sacks, side-dish bundles and liquor bottle are now being received in camp. 임정현: He pauses with his hand over the rice sacks during counting.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 임정현 (한국인 남성, 성인 나이대, 자연스러운 검은 머리) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.",
   "B": "Create ONE FINAL photorealistic live-action film still of the moment below — early 20th-century Korea, Jeolla region; all people are Korean unless explicitly stated otherwise. TIME OF DAY (lock): morning.\n\nSHOT TEXT (authoritative, Korean): 쌀자루 위에 손을 멈춘 채 박상현을 향해 고개를 홱 돌린 임정현의 굳은 얼굴.\n\nLOCATION (lock): The outdoor receiving spot beside the newly delivered rice sacks at the mountain camp's entrance. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: medium shot\n- KEY BACKGROUND ELEMENTS: 쌀자루 (임정현이 세다가 그 위에 손을 멈추었다) — The upper portion lies beneath his hand at the lower frame edge; used as Anchors the interrupted task without foreground size exaggeration.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Maintain neutral morning illumination and muted natural contrast, letting the stopped hand and tightened expression carry the unease.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The clothing, hair and overall look of 임정현 — who appear both in that photo and in this shot — are LOCKED to that photo. Anyone else visible in that photo is NOT in this shot: never carry their face, body or clothing onto anyone here. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The delivered rice sacks, side-dish bundles and liquor bottle are now being received in camp. 임정현: He pauses with his hand over the rice sacks during counting.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 임정현 (한국인 남성, 성인 나이대, 자연스러운 검은 머리) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.\n\nCOMPOSITION VARIATION (this candidate only): keep the place, the people and the moment exactly as specified, and keep from the CAMERA & FRAME contract its subject scale, screen placement and key background elements exactly. For this candidate only, its camera angle and height are deliberately left open: choose a camera position distinctly different from the obvious one, as a film director picking a second setup on the same blocking. Stay true to any explicit framing the SHOT TEXT itself dictates."
  },
  "shot_run_spend_attempt_count": 1,
  "readings": [
   {
    "label": "A",
    "direction": "임정현이 프레임 왼쪽 밖을 향해 고개를 돌리고 있으며, 시선의 대상은 화면 안에 보이지 않음.",
    "built_space": "야외 캠프 입구의 흙길. 나무 울타리와 소나무, 초가집 등 이전 샷의 환경이 잘 나타남. 화면 하단에 쌀자루들이 쌓여 있음.",
    "entities": "임정현: 얼굴은 비슷하나 깨끗한 흰 옷을 입고 있어 레퍼런스(오염된 옷, 검은 띠)와 불일치함. 배경에 샷 텍스트에 언급되지 않은 인물 4명이 등장함.",
    "hard_violations": [
     "[gemini-pro] 이전 샷 스틸에 있던 인물들(경찰 제복을 입은 남자 등)을 이 샷에 등장시키지 말라는 명시적 금지 지시를 무시하고 배경에 그대로 복사하여 배치함 (초과된/발명된 인물).",
     "[gpt-high] 임정현 외 인물을 제외하라는 지시를 어기고 산채 입구와 자루 옆에 배경 남성 네 명을 추가했다."
    ],
    "physics": "임정현의 오른손이 쌀자루 위에 안정적으로 놓여 있으며, 쌀자루들은 서로 지탱하여 쌓여 있음. 배경 인물들도 지면에 서 있음."
   },
   {
    "label": "B",
    "direction": "임정현이 오른쪽 어깨 너머로 고개를 홱 돌려, 화면 왼쪽 전경에 등지고 있는 남성(박상현)을 똑바로 바라보고 있음.",
    "built_space": "야외 캠프 입구의 흙길. 양옆에 초가집이 있고 오른쪽엔 소나무 기둥이 서 있음. 화면 하단에 쌀자루들이 쌓여 있음.",
    "entities": "임정현: 얼굴과 흙 묻은 의상, 검은 띠가 레퍼런스와 잘 일치함(망건과 상투는 없음). 샷 텍스트에 시선의 대상으로 명시된 박상현이 전경에 등장함. 쌀자루: 화면 하단에 올바르게 배치됨.",
    "hard_violations": [
     "[gpt-high] 임정현만 보여야 하는 화면에 푸른 옷을 입은 다른 남성의 머리와 상체를 왼쪽 전경에 크게 추가했다."
    ],
    "physics": "임정현의 오른손이 쌀자루 위에 자연스럽게 얹혀 있으며, 자루들은 쌓인 채로 안정적임. 두 인물 모두 지면에 올바르게 서 있음."
   }
  ],
  "cross_model_order": {
   "policy": "cross_model_order_v2_slot0_fwd_slot1_rev_parallel2",
   "slots": [
    {
     "model": "gemini-pro",
     "order": "forward",
     "display_to_canonical": {
      "A": "A",
      "B": "B"
     },
     "raw": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 8,
        "verdict_ko": "임정현의 오염된 의상과 검은 띠를 레퍼런스와 동일하게 구현했으며, 전경의 인물(박상현)을 향해 고개를 홱 돌리는 연출을 정확히 포착했습니다."
       },
       {
        "label": "A",
        "score": 2,
        "verdict_ko": "이전 샷의 인물들을 배경에 추가하는 치명적인 금지 사항 위반을 범했고, 임정현의 의상 디테일(오염도, 검은 띠)도 완전히 누락되었습니다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "임정현이 프레임 왼쪽 밖을 향해 고개를 돌리고 있으며, 시선의 대상은 화면 안에 보이지 않음.",
        "built_space": "야외 캠프 입구의 흙길. 나무 울타리와 소나무, 초가집 등 이전 샷의 환경이 잘 나타남. 화면 하단에 쌀자루들이 쌓여 있음.",
        "entities": "임정현: 얼굴은 비슷하나 깨끗한 흰 옷을 입고 있어 레퍼런스(오염된 옷, 검은 띠)와 불일치함. 배경에 샷 텍스트에 언급되지 않은 인물 4명이 등장함.",
        "hard_violations": [
         "이전 샷 스틸에 있던 인물들(경찰 제복을 입은 남자 등)을 이 샷에 등장시키지 말라는 명시적 금지 지시를 무시하고 배경에 그대로 복사하여 배치함 (초과된/발명된 인물)."
        ],
        "physics": "임정현의 오른손이 쌀자루 위에 안정적으로 놓여 있으며, 쌀자루들은 서로 지탱하여 쌓여 있음. 배경 인물들도 지면에 서 있음."
       },
       {
        "label": "B",
        "direction": "임정현이 오른쪽 어깨 너머로 고개를 홱 돌려, 화면 왼쪽 전경에 등지고 있는 남성(박상현)을 똑바로 바라보고 있음.",
        "built_space": "야외 캠프 입구의 흙길. 양옆에 초가집이 있고 오른쪽엔 소나무 기둥이 서 있음. 화면 하단에 쌀자루들이 쌓여 있음.",
        "entities": "임정현: 얼굴과 흙 묻은 의상, 검은 띠가 레퍼런스와 잘 일치함(망건과 상투는 없음). 샷 텍스트에 시선의 대상으로 명시된 박상현이 전경에 등장함. 쌀자루: 화면 하단에 올바르게 배치됨.",
        "hard_violations": [],
        "physics": "임정현의 오른손이 쌀자루 위에 자연스럽게 얹혀 있으며, 자루들은 쌓인 채로 안정적임. 두 인물 모두 지면에 올바르게 서 있음."
       }
      ],
      "all_candidates_fail": false
     },
     "normalized": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 8,
        "verdict_ko": "임정현의 오염된 의상과 검은 띠를 레퍼런스와 동일하게 구현했으며, 전경의 인물(박상현)을 향해 고개를 홱 돌리는 연출을 정확히 포착했습니다."
       },
       {
        "label": "A",
        "score": 2,
        "verdict_ko": "이전 샷의 인물들을 배경에 추가하는 치명적인 금지 사항 위반을 범했고, 임정현의 의상 디테일(오염도, 검은 띠)도 완전히 누락되었습니다."
       }
      ],
      "all_candidates_fail": false,
      "readings": [
       {
        "label": "A",
        "direction": "임정현이 프레임 왼쪽 밖을 향해 고개를 돌리고 있으며, 시선의 대상은 화면 안에 보이지 않음.",
        "built_space": "야외 캠프 입구의 흙길. 나무 울타리와 소나무, 초가집 등 이전 샷의 환경이 잘 나타남. 화면 하단에 쌀자루들이 쌓여 있음.",
        "entities": "임정현: 얼굴은 비슷하나 깨끗한 흰 옷을 입고 있어 레퍼런스(오염된 옷, 검은 띠)와 불일치함. 배경에 샷 텍스트에 언급되지 않은 인물 4명이 등장함.",
        "hard_violations": [
         "이전 샷 스틸에 있던 인물들(경찰 제복을 입은 남자 등)을 이 샷에 등장시키지 말라는 명시적 금지 지시를 무시하고 배경에 그대로 복사하여 배치함 (초과된/발명된 인물)."
        ],
        "physics": "임정현의 오른손이 쌀자루 위에 안정적으로 놓여 있으며, 쌀자루들은 서로 지탱하여 쌓여 있음. 배경 인물들도 지면에 서 있음."
       },
       {
        "label": "B",
        "direction": "임정현이 오른쪽 어깨 너머로 고개를 홱 돌려, 화면 왼쪽 전경에 등지고 있는 남성(박상현)을 똑바로 바라보고 있음.",
        "built_space": "야외 캠프 입구의 흙길. 양옆에 초가집이 있고 오른쪽엔 소나무 기둥이 서 있음. 화면 하단에 쌀자루들이 쌓여 있음.",
        "entities": "임정현: 얼굴과 흙 묻은 의상, 검은 띠가 레퍼런스와 잘 일치함(망건과 상투는 없음). 샷 텍스트에 시선의 대상으로 명시된 박상현이 전경에 등장함. 쌀자루: 화면 하단에 올바르게 배치됨.",
        "hard_violations": [],
        "physics": "임정현의 오른손이 쌀자루 위에 자연스럽게 얹혀 있으며, 자루들은 쌓인 채로 안정적임. 두 인물 모두 지면에 올바르게 서 있음."
       }
      ]
     },
     "ok": true
    },
    {
     "model": "gpt-high",
     "order": "reverse",
     "display_to_canonical": {
      "A": "B",
      "B": "A"
     },
     "raw": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 2,
        "verdict_ko": "쌀자루에 멈춘 손과 돌아본 얼굴은 구현했지만, 금지된 두 번째 인물을 전경에 크게 넣어 어깨너머 구도로 바꾸었고 인물 참조의 머리와 얼굴도 다릅니다."
       },
       {
        "label": "B",
        "score": 3,
        "verdict_ko": "임정현 중심의 미디엄 숏과 산채 입구의 공간 연속성은 더 가깝지만, 배경 인물 네 명을 추가한 실격 오류가 있으며 쌀자루도 하단에서 지나치게 크게 드러납니다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "임정현의 몸은 비스듬히 오른쪽을 향하고 고개와 눈은 왼쪽 전경의 푸른 옷 남자를 향한다. 돌아보는 대상은 명확하지만 그 남자가 박상현인지는 확인할 수 없고, 애초에 이 화면에는 임정현 외 인물을 보여주면 안 된다. 전경 남자도 임정현 쪽으로 얼굴을 향한다.",
        "built_space": "흙마당 뒤에 초가 두 채가 보이며, 왼쪽은 엮은 벽과 기대어 놓은 자루, 오른쪽은 기둥과 낮은 마루가 있다. 임정현은 전경 쌀자루 뒤에 서 있다. 목재·초가·소나무 재료는 참조와 통하지만, 참조에서 두드러진 입구 목책과 통로의 배치는 확인되지 않는다. 반사면이나 명백히 중복된 고정 설비는 없다.",
        "entities": "성인 동아시아계 남성 두 명이 보이며, 임정현 외 남성 한 명은 허용되지 않은 인물이다. 임정현의 검은 머리와 때 묻은 밝은 전통옷, 검은 허리끈은 보이지만, 인물 참조의 상투와 검은 머리쓰개 대신 짧게 가른 머리이고 얼굴도 정확히 일치하지 않는다. 하단에는 마대 쌀자루 윗부분 네 덩이가 보인다. 반찬 보따리와 술병은 보이지 않지만 이 구도에서 필수 노출 대상은 아니다. 읽을 수 있는 글자는 없다.",
        "hard_violations": [
         "임정현만 보여야 하는 화면에 푸른 옷을 입은 다른 남성의 머리와 상체를 왼쪽 전경에 크게 추가했다."
        ],
        "physics": "임정현의 보이는 손은 쌀자루 곡면에 손가락을 대고 있고 팔과 자연스럽게 연결된다. 몸통을 남겨 둔 채 목을 돌린 자세는 가능하다. 두 사람의 발과 자루의 바닥 지지는 화면 밖이지만 공중에 떠 있다는 징후는 없다. 자루는 하단의 적재물로 이어져 보이며, 지지 없는 물체나 불가능한 관절은 보이지 않는다."
       },
       {
        "label": "B",
        "direction": "임정현은 몸통을 거의 정면에 두고 고개와 눈을 화면 왼쪽 밖으로 돌렸다. 박상현은 보이지 않아 실제 시선 도착점은 검증할 수 없지만, 화면 밖 상대를 향해 홱 돌아본 연출로 읽힌다. 배경 네 남자는 각각 통로 안쪽이나 자루 작업 쪽을 향하고 있으며 임정현의 시선 대상은 아니다.",
        "built_space": "양옆의 큰 소나무, 좌우 목책 구간과 그 사이 흙길 입구, 왼쪽 초가 일부가 보인다. 참조의 산채 입구 구성과 재료가 A보다 직접적으로 이어진다. 임정현은 입구 앞 쌀자루 더미 뒤에 서 있다. 다만 자루는 윗부분만 하단에 걸리는 수준을 넘어 화면 아래 약 사분의 일을 차지하고 여러 단으로 노출된다. 불가능한 반사나 확실한 설비 중복은 보이지 않는다.",
        "entities": "전경 임정현 외에 통로 중앙 두 명과 오른쪽 자루 옆 두 명, 총 네 명의 배경 남성이 추가되어 있다. 임정현은 성인 동아시아계 남성으로 검은 머리와 밝은 전통옷을 입었으나, 인물 참조의 상투·검은 머리쓰개가 없고 얼굴도 정확한 일치로 보기 어렵다. 밝은 옷의 겹침과 앞 매듭은 보이지만 참조의 검은 허리끈은 보이지 않는다. 전경에는 윗단 쌀자루 세 개와 아래층 자루 일부가 보인다. 반찬 보따리와 술병은 식별되지 않으며, 읽을 수 있는 글자는 없다.",
        "hard_violations": [
         "임정현 외 인물을 제외하라는 지시를 어기고 산채 입구와 자루 옆에 배경 남성 네 명을 추가했다."
        ],
        "physics": "임정현의 손바닥과 손가락이 가운데 윗단 자루에 자연스럽게 닿아 있어 세다가 멈춘 행동으로 읽힌다. 자루들은 아래층 자루 위에 포개져 지지된다. 임정현의 하체는 가려졌지만 서 있는 몸의 연결은 자연스럽고, 배경 인물들은 흙길에 발을 딛고 있다. 지지 없이 떠 있는 사람이나 물체는 보이지 않는다."
       }
      ],
      "all_candidates_fail": true
     },
     "normalized": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 2,
        "verdict_ko": "쌀자루에 멈춘 손과 돌아본 얼굴은 구현했지만, 금지된 두 번째 인물을 전경에 크게 넣어 어깨너머 구도로 바꾸었고 인물 참조의 머리와 얼굴도 다릅니다."
       },
       {
        "label": "A",
        "score": 3,
        "verdict_ko": "임정현 중심의 미디엄 숏과 산채 입구의 공간 연속성은 더 가깝지만, 배경 인물 네 명을 추가한 실격 오류가 있으며 쌀자루도 하단에서 지나치게 크게 드러납니다."
       }
      ],
      "all_candidates_fail": true,
      "readings": [
       {
        "label": "B",
        "direction": "임정현의 몸은 비스듬히 오른쪽을 향하고 고개와 눈은 왼쪽 전경의 푸른 옷 남자를 향한다. 돌아보는 대상은 명확하지만 그 남자가 박상현인지는 확인할 수 없고, 애초에 이 화면에는 임정현 외 인물을 보여주면 안 된다. 전경 남자도 임정현 쪽으로 얼굴을 향한다.",
        "built_space": "흙마당 뒤에 초가 두 채가 보이며, 왼쪽은 엮은 벽과 기대어 놓은 자루, 오른쪽은 기둥과 낮은 마루가 있다. 임정현은 전경 쌀자루 뒤에 서 있다. 목재·초가·소나무 재료는 참조와 통하지만, 참조에서 두드러진 입구 목책과 통로의 배치는 확인되지 않는다. 반사면이나 명백히 중복된 고정 설비는 없다.",
        "entities": "성인 동아시아계 남성 두 명이 보이며, 임정현 외 남성 한 명은 허용되지 않은 인물이다. 임정현의 검은 머리와 때 묻은 밝은 전통옷, 검은 허리끈은 보이지만, 인물 참조의 상투와 검은 머리쓰개 대신 짧게 가른 머리이고 얼굴도 정확히 일치하지 않는다. 하단에는 마대 쌀자루 윗부분 네 덩이가 보인다. 반찬 보따리와 술병은 보이지 않지만 이 구도에서 필수 노출 대상은 아니다. 읽을 수 있는 글자는 없다.",
        "hard_violations": [
         "임정현만 보여야 하는 화면에 푸른 옷을 입은 다른 남성의 머리와 상체를 왼쪽 전경에 크게 추가했다."
        ],
        "physics": "임정현의 보이는 손은 쌀자루 곡면에 손가락을 대고 있고 팔과 자연스럽게 연결된다. 몸통을 남겨 둔 채 목을 돌린 자세는 가능하다. 두 사람의 발과 자루의 바닥 지지는 화면 밖이지만 공중에 떠 있다는 징후는 없다. 자루는 하단의 적재물로 이어져 보이며, 지지 없는 물체나 불가능한 관절은 보이지 않는다."
       },
       {
        "label": "A",
        "direction": "임정현은 몸통을 거의 정면에 두고 고개와 눈을 화면 왼쪽 밖으로 돌렸다. 박상현은 보이지 않아 실제 시선 도착점은 검증할 수 없지만, 화면 밖 상대를 향해 홱 돌아본 연출로 읽힌다. 배경 네 남자는 각각 통로 안쪽이나 자루 작업 쪽을 향하고 있으며 임정현의 시선 대상은 아니다.",
        "built_space": "양옆의 큰 소나무, 좌우 목책 구간과 그 사이 흙길 입구, 왼쪽 초가 일부가 보인다. 참조의 산채 입구 구성과 재료가 A보다 직접적으로 이어진다. 임정현은 입구 앞 쌀자루 더미 뒤에 서 있다. 다만 자루는 윗부분만 하단에 걸리는 수준을 넘어 화면 아래 약 사분의 일을 차지하고 여러 단으로 노출된다. 불가능한 반사나 확실한 설비 중복은 보이지 않는다.",
        "entities": "전경 임정현 외에 통로 중앙 두 명과 오른쪽 자루 옆 두 명, 총 네 명의 배경 남성이 추가되어 있다. 임정현은 성인 동아시아계 남성으로 검은 머리와 밝은 전통옷을 입었으나, 인물 참조의 상투·검은 머리쓰개가 없고 얼굴도 정확한 일치로 보기 어렵다. 밝은 옷의 겹침과 앞 매듭은 보이지만 참조의 검은 허리끈은 보이지 않는다. 전경에는 윗단 쌀자루 세 개와 아래층 자루 일부가 보인다. 반찬 보따리와 술병은 식별되지 않으며, 읽을 수 있는 글자는 없다.",
        "hard_violations": [
         "임정현 외 인물을 제외하라는 지시를 어기고 산채 입구와 자루 옆에 배경 남성 네 명을 추가했다."
        ],
        "physics": "임정현의 손바닥과 손가락이 가운데 윗단 자루에 자연스럽게 닿아 있어 세다가 멈춘 행동으로 읽힌다. 자루들은 아래층 자루 위에 포개져 지지된다. 임정현의 하체는 가려졌지만 서 있는 몸의 연결은 자연스럽고, 배경 인물들은 흙길에 발을 딛고 있다. 지지 없이 떠 있는 사람이나 물체는 보이지 않는다."
       }
      ]
     },
     "ok": true
    }
   ],
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "slot_winner_match": false,
   "slot_winner": {
    "gemini-pro": "B",
    "gpt-high": "A"
   },
   "route": "cross_slot_combined"
  },
  "dual": {
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "normalized": {
    "A": 1.25,
    "B": 1.667
   },
   "adjusted": {
    "A": 1.0,
    "B": 1.417
   },
   "violations": {
    "A": [
     "[gemini-pro] 이전 샷 스틸에 있던 인물들(경찰 제복을 입은 남자 등)을 이 샷에 등장시키지 말라는 명시적 금지 지시를 무시하고 배경에 그대로 복사하여 배치함 (초과된/발명된 인물).",
     "[gpt-high] 임정현 외 인물을 제외하라는 지시를 어기고 산채 입구와 자루 옆에 배경 남성 네 명을 추가했다."
    ],
    "B": [
     "[gpt-high] 임정현만 보여야 하는 화면에 푸른 옷을 입은 다른 남성의 머리와 상체를 왼쪽 전경에 크게 추가했다."
    ]
   },
   "per_model_winner": {
    "gemini-pro": "B",
    "gpt-high": "A"
   },
   "agreed": false
  },
  "totals": {
   "B": 1417,
   "A": 1000
  },
  "selected": "B",
  "ranking": [
   "B",
   "A"
  ],
  "verdicts": [
   {
    "label": "B",
    "score": 1417,
    "verdict_ko": "임정현의 오염된 의상과 검은 띠를 레퍼런스와 동일하게 구현했으며, 전경의 인물(박상현)을 향해 고개를 홱 돌리는 연출을 정확히 포착했습니다.  ★위반: [gpt-high] 임정현만 보여야 하는 화면에 푸른 옷을 입은 다른 남성의 머리와 상체를 왼쪽 전경에 크게 추가했다."
   },
   {
    "label": "A",
    "score": 1000,
    "verdict_ko": "이전 샷의 인물들을 배경에 추가하는 치명적인 금지 사항 위반을 범했고, 임정현의 의상 디테일(오염도, 검은 띠)도 완전히 누락되었습니다.  ★위반: [gemini-pro] 이전 샷 스틸에 있던 인물들(경찰 제복을 입은 남자 등)을 이 샷에 등장시키지 말라는 명시적 금지 지시를 무시하고 배경에 그대로 복사하여 배치함 (초과된/발명된 인물). / [gpt-high] 임정현 외 인물을 제외하라는 지시를 어기고 산채 입구와 자루 옆에 배경 남성 네 명을 추가했다."
   }
  ],
  "refs": [
   {
    "label": "PREVIOUS SHOT STILL — a visually related earlier shot of this same place: the location's look, materials, fixed features and lighting mood are LOCKED to this photo; never copy its camera framing. The clothing, hair and overall look of 임정현 — who appear both in that photo and in this shot — are LOCKED to that photo. Anyone else visible in that photo is NOT in this shot: never carry their face, body or clothing onto anyone here. If this photo shows a character who cannot move (dead or unconscious), that character's exact pose, position and orientation are ALSO LOCKED — treat their body as a fixed prop of the set that only the camera moves around.",
    "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/db00517b-165b-4005-a287-f28936b59b95/images/272bd7af-9294-4490-98c6-88a40e9deacc/scene/recipe/S46sh16_sel.png",
    "asset_id": "44c9a113-0806-4b07-a450-080e904cdcb1",
    "role": "prev_still"
   },
   {
    "label": "CHARACTER REFERENCE — 임정현: the exact person appearing in this shot; match face, hair and build exactly.",
    "path": "<bytes:858188>",
    "asset_id": "789c3359-d7f8-4832-9a6f-7e4ad2695c02",
    "role": "character_ref"
   }
  ],
  "critique_skipped": true,
  "shot_run_produced": false,
  "shot_run_uid": "06aa1435-48db-7e8f-a766-59cfa184ec76",
  "ref_mode": "prev+엔티티",
  "share_plan": {
   "ref_plan": "prev",
   "prev_anchor_tag": "S46sh16"
  }
 },
 "S47sh13::cine": {
  "applied": true,
  "attempted_at": "2026-09-09T06:22:03.078159+00:00",
  "fingerprint": "824e832982958ac0f4b5e2f97099dab82610a81ba88b8a6a11bb410d9af17d31",
  "fingerprint_version": 2,
  "provider": "mai",
  "endpoint": "openrouter/chat-completions",
  "model": "microsoft/mai-image-2.6",
  "pack": "24.202608252115",
  "source_file": "S47sh13_sel.png",
  "source_sha256": "343378cfd741ac28bd541410f6429931493766f08c06a1e26d2c9643c67bc550",
  "file": "S47sh13_cine.png",
  "staged_sha256": "06c2cb9c608bb18093a0c6a7c944e501a1ec2c9aa21869ea5d2177362e8dbe41",
  "latency_ms": 27575
 },
 "S47sh18::signage": {
  "fp": "b58c10de37e25cf5",
  "inscriptions": [],
  "cues": [],
  "dropped": []
 },
 "S47sh18": {
  "input_fingerprint": "0b6036f7e3028159",
  "prompt": "Create ONE FINAL photorealistic live-action film still of the moment below — early 20th-century Korea, Jeolla region; all people are Korean unless explicitly stated otherwise. TIME OF DAY (lock): morning.\n\nSHOT TEXT (authoritative, Korean): 멀어지는 세 사람의 뒷모습을 뚫어져라 응시하는 임정현의 무거운 정면.\n\nLOCATION (lock): The mountain camp's outdoor supply-receiving area, facing the downhill forest path taken by the departing three men. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: medium shot\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Continue the same muted morning exposure and naturalistic contrast, with no expressive lighting shift or perceptual distortion.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The clothing, hair and overall look of 임정현 — who appear both in that photo and in this shot — are LOCKED to that photo. Anyone else visible in that photo is NOT in this shot: never carry their face, body or clothing onto anyone here. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): Food bundles are being opened beside the delivered rice sacks and liquor bottle. 임정현: He remains beside the food delivery, pausing before joining the meal.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 임정현 (한국인 남성, 성인 나이대, 자연스러운 검은 머리) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.",
  "roll_prompts": {
   "A": "Create ONE FINAL photorealistic live-action film still of the moment below — early 20th-century Korea, Jeolla region; all people are Korean unless explicitly stated otherwise. TIME OF DAY (lock): morning.\n\nSHOT TEXT (authoritative, Korean): 멀어지는 세 사람의 뒷모습을 뚫어져라 응시하는 임정현의 무거운 정면.\n\nLOCATION (lock): The mountain camp's outdoor supply-receiving area, facing the downhill forest path taken by the departing three men. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: medium shot\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Continue the same muted morning exposure and naturalistic contrast, with no expressive lighting shift or perceptual distortion.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The clothing, hair and overall look of 임정현 — who appear both in that photo and in this shot — are LOCKED to that photo. Anyone else visible in that photo is NOT in this shot: never carry their face, body or clothing onto anyone here. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): Food bundles are being opened beside the delivered rice sacks and liquor bottle. 임정현: He remains beside the food delivery, pausing before joining the meal.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 임정현 (한국인 남성, 성인 나이대, 자연스러운 검은 머리) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.",
   "B": "Create ONE FINAL photorealistic live-action film still of the moment below — early 20th-century Korea, Jeolla region; all people are Korean unless explicitly stated otherwise. TIME OF DAY (lock): morning.\n\nSHOT TEXT (authoritative, Korean): 멀어지는 세 사람의 뒷모습을 뚫어져라 응시하는 임정현의 무거운 정면.\n\nLOCATION (lock): The mountain camp's outdoor supply-receiving area, facing the downhill forest path taken by the departing three men. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: medium shot\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Continue the same muted morning exposure and naturalistic contrast, with no expressive lighting shift or perceptual distortion.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The clothing, hair and overall look of 임정현 — who appear both in that photo and in this shot — are LOCKED to that photo. Anyone else visible in that photo is NOT in this shot: never carry their face, body or clothing onto anyone here. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): Food bundles are being opened beside the delivered rice sacks and liquor bottle. 임정현: He remains beside the food delivery, pausing before joining the meal.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 임정현 (한국인 남성, 성인 나이대, 자연스러운 검은 머리) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.\n\nCOMPOSITION VARIATION (this candidate only): keep the place, the people and the moment exactly as specified, and keep from the CAMERA & FRAME contract its subject scale, screen placement and key background elements exactly. For this candidate only, its camera angle and height are deliberately left open: choose a camera position distinctly different from the obvious one, as a film director picking a second setup on the same blocking. Stay true to any explicit framing the SHOT TEXT itself dictates."
  },
  "shot_run_spend_attempt_count": 1,
  "readings": [
   {
    "label": "A",
    "direction": "임정현은 몸을 돌려 정면(카메라 렌즈)을 응시하고 있으며, 배경의 세 남자는 카메라를 등지고 멀어지고 있어 임정현이 그들을 응시한다는 묘사와 맞지 않음.",
    "built_space": "초가집과 나무들이 있는 야외 배경이며, 전경에 쌀가마니가 쌓여 있으나 지문에 명시된 음식 꾸러미와 술병은 존재하지 않음.",
    "entities": "임정현은 지문의 '자연스러운 검은 머리' 지침 및 이전 샷의 상태와 달리 머리에 망건을 쓰고 있음. 배경에 세 명의 남자가 뒷모습으로 등장함.",
    "hard_violations": [
     "[gemini-pro] 이전 샷과 완벽히 일치해야 하는 외모 지침을 어기고 머리에 망건을 착용함 (캐릭터 외모 잠금 위반).",
     "[gpt-high] 떠나는 세 사람을 임정현의 등 뒤에 배치하고 임정현은 반대쪽을 바라보게 하여, 세 사람의 뒷모습을 응시한다는 필수 인물 배치가 성립하지 않습니다."
    ],
    "physics": "인물들이 땅에 잘 서 있고 가마니도 무너지지 않게 지지되어 있으며 특별한 물리적 오류는 없음."
   },
   {
    "label": "B",
    "direction": "임정현은 정면으로 카메라를 응시하고 있고, 배경의 세 사람은 내리막길을 따라 멀어지고 있음. (지문의 시선 방향과는 논리적으로 어긋나나 프레이밍 요구를 따른 결과로 보임).",
    "built_space": "이전 샷과 유사한 헛간 구조물과 내리막 숲길이 보임. 전경에 쌀가마니, 천으로 싼 음식 꾸러미, 술병이 지침대로 정확하게 배치되어 있음.",
    "entities": "임정현은 이전 샷과 일치하는 헝클어진 검은 머리와 의상 상태를 유지함. 배경에 세 명의 남자가 등장함.",
    "hard_violations": [
     "[gpt-high] 떠나는 세 사람을 임정현의 등 뒤에 배치하고 임정현은 반대쪽을 바라보게 하여, 세 사람의 뒷모습을 응시한다는 필수 인물 배치가 성립하지 않습니다."
    ],
    "physics": "임정현의 오른손이 가마니 위에 자연스럽게 놓여 지지받고 있으며, 배경 인물들의 걷는 자세와 사물들의 배치가 안정적임."
   }
  ],
  "cross_model_order": {
   "policy": "cross_model_order_v2_slot0_fwd_slot1_rev_parallel2",
   "slots": [
    {
     "model": "gemini-pro",
     "order": "forward",
     "display_to_canonical": {
      "A": "A",
      "B": "B"
     },
     "raw": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 8,
        "verdict_ko": "카메라와 인물, 배경의 논리적 시선 방향에 물리적 모순이 존재하지만, 지문이 요구한 전경의 소품(술병, 음식 꾸러미)과 인물의 외모(자연스러운 검은 머리)를 이전 샷에 맞춰 매우 훌륭하게 구현했습니다."
       },
       {
        "label": "A",
        "score": 3,
        "verdict_ko": "이전 샷과 레퍼런스에 없는 망건을 머리에 착용하여 캐릭터 외모 잠금 지침을 크게 위반했으며, 요구된 음식 꾸러미와 술병 소품도 누락되었습니다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "임정현은 몸을 돌려 정면(카메라 렌즈)을 응시하고 있으며, 배경의 세 남자는 카메라를 등지고 멀어지고 있어 임정현이 그들을 응시한다는 묘사와 맞지 않음.",
        "built_space": "초가집과 나무들이 있는 야외 배경이며, 전경에 쌀가마니가 쌓여 있으나 지문에 명시된 음식 꾸러미와 술병은 존재하지 않음.",
        "entities": "임정현은 지문의 '자연스러운 검은 머리' 지침 및 이전 샷의 상태와 달리 머리에 망건을 쓰고 있음. 배경에 세 명의 남자가 뒷모습으로 등장함.",
        "hard_violations": [
         "이전 샷과 완벽히 일치해야 하는 외모 지침을 어기고 머리에 망건을 착용함 (캐릭터 외모 잠금 위반)."
        ],
        "physics": "인물들이 땅에 잘 서 있고 가마니도 무너지지 않게 지지되어 있으며 특별한 물리적 오류는 없음."
       },
       {
        "label": "B",
        "direction": "임정현은 정면으로 카메라를 응시하고 있고, 배경의 세 사람은 내리막길을 따라 멀어지고 있음. (지문의 시선 방향과는 논리적으로 어긋나나 프레이밍 요구를 따른 결과로 보임).",
        "built_space": "이전 샷과 유사한 헛간 구조물과 내리막 숲길이 보임. 전경에 쌀가마니, 천으로 싼 음식 꾸러미, 술병이 지침대로 정확하게 배치되어 있음.",
        "entities": "임정현은 이전 샷과 일치하는 헝클어진 검은 머리와 의상 상태를 유지함. 배경에 세 명의 남자가 등장함.",
        "hard_violations": [],
        "physics": "임정현의 오른손이 가마니 위에 자연스럽게 놓여 지지받고 있으며, 배경 인물들의 걷는 자세와 사물들의 배치가 안정적임."
       }
      ],
      "all_candidates_fail": false
     },
     "normalized": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 8,
        "verdict_ko": "카메라와 인물, 배경의 논리적 시선 방향에 물리적 모순이 존재하지만, 지문이 요구한 전경의 소품(술병, 음식 꾸러미)과 인물의 외모(자연스러운 검은 머리)를 이전 샷에 맞춰 매우 훌륭하게 구현했습니다."
       },
       {
        "label": "A",
        "score": 3,
        "verdict_ko": "이전 샷과 레퍼런스에 없는 망건을 머리에 착용하여 캐릭터 외모 잠금 지침을 크게 위반했으며, 요구된 음식 꾸러미와 술병 소품도 누락되었습니다."
       }
      ],
      "all_candidates_fail": false,
      "readings": [
       {
        "label": "A",
        "direction": "임정현은 몸을 돌려 정면(카메라 렌즈)을 응시하고 있으며, 배경의 세 남자는 카메라를 등지고 멀어지고 있어 임정현이 그들을 응시한다는 묘사와 맞지 않음.",
        "built_space": "초가집과 나무들이 있는 야외 배경이며, 전경에 쌀가마니가 쌓여 있으나 지문에 명시된 음식 꾸러미와 술병은 존재하지 않음.",
        "entities": "임정현은 지문의 '자연스러운 검은 머리' 지침 및 이전 샷의 상태와 달리 머리에 망건을 쓰고 있음. 배경에 세 명의 남자가 뒷모습으로 등장함.",
        "hard_violations": [
         "이전 샷과 완벽히 일치해야 하는 외모 지침을 어기고 머리에 망건을 착용함 (캐릭터 외모 잠금 위반)."
        ],
        "physics": "인물들이 땅에 잘 서 있고 가마니도 무너지지 않게 지지되어 있으며 특별한 물리적 오류는 없음."
       },
       {
        "label": "B",
        "direction": "임정현은 정면으로 카메라를 응시하고 있고, 배경의 세 사람은 내리막길을 따라 멀어지고 있음. (지문의 시선 방향과는 논리적으로 어긋나나 프레이밍 요구를 따른 결과로 보임).",
        "built_space": "이전 샷과 유사한 헛간 구조물과 내리막 숲길이 보임. 전경에 쌀가마니, 천으로 싼 음식 꾸러미, 술병이 지침대로 정확하게 배치되어 있음.",
        "entities": "임정현은 이전 샷과 일치하는 헝클어진 검은 머리와 의상 상태를 유지함. 배경에 세 명의 남자가 등장함.",
        "hard_violations": [],
        "physics": "임정현의 오른손이 가마니 위에 자연스럽게 놓여 지지받고 있으며, 배경 인물들의 걷는 자세와 사물들의 배치가 안정적임."
       }
      ]
     },
     "ok": true
    },
    {
     "model": "gpt-high",
     "order": "reverse",
     "display_to_canonical": {
      "A": "B",
      "B": "A"
     },
     "raw": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 2,
        "verdict_ko": "이전 컷의 머리와 의상은 잘 이어지지만, 세 사람을 등진 채 카메라를 바라보며 허벅지까지 넓혀 잡아 핵심 응시 관계와 미디엄 숏을 모두 놓쳤습니다."
       },
       {
        "label": "B",
        "score": 3,
        "verdict_ko": "무거운 정면과 미디엄 숏에는 더 가깝지만, 응시 대상이 등 뒤에 있어 핵심 연출이 성립하지 않고 상투·망건도 이전 컷의 머리와 불일치합니다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "임정현의 얼굴과 눈은 화면 앞쪽, 거의 렌즈를 향합니다. 세 남자는 그의 오른쪽 뒤에서 카메라에 등을 보이고 숲길 안쪽으로 걸어갑니다. 따라서 눈길은 보이는 세 사람의 뒷모습에 닿지 않습니다. 길은 인물들 바로 앞에서 낮아진 뒤 먼 곳에서 다시 올라가는 형태로 보입니다.",
        "built_space": "흙마당 왼쪽에 기둥과 낮은 마루가 있는 초가 한 채, 오른쪽 가장자리에 초가 일부가 보이고 가운데 오른쪽으로 숲길이 열립니다. 전경에는 왼쪽의 세운 포대 하나와 하단·오른쪽의 눕힌 포대 세 개가 부분적으로 보입니다. 임정현은 포대 뒤에 서 있습니다. 초가·목재·흙의 재료는 참고 장소와 유사하지만, 세 사람의 이동로를 임정현 뒤에 배치한 공간 관계는 요구된 응시 장면과 맞지 않습니다.",
        "entities": "성인 한국인 남성으로 보이는 임정현 한 명과 뒷모습의 남성 세 명이 보입니다. 임정현의 풀어 내린 검은 머리, 얼굴 인상, 때 묻은 미색 옷과 검은 허리끈은 이전 컷에 가깝습니다. 세 남자의 얼굴과 세부 신원은 뒷모습과 흐림 때문에 확인할 수 없습니다. 포대, 갈색 술병 한 개, 일부 펼친 포장과 묶인 음식 보따리가 보이지만 보따리를 여는 동작은 없습니다. 읽을 수 있는 글자는 보이지 않습니다.",
        "hard_violations": [
         "떠나는 세 사람을 임정현의 등 뒤에 배치하고 임정현은 반대쪽을 바라보게 하여, 세 사람의 뒷모습을 응시한다는 필수 인물 배치가 성립하지 않습니다."
        ],
        "physics": "임정현의 화면 왼쪽 손은 세운 포대에 얹혀 있고 다른 손은 몸 옆에 내려와 있습니다. 발은 프레임 밖이지만 직립한 몸에 공중 부양 징후는 없습니다. 세 남자는 발을 지면에 딛거나 보행 중 들어 올린 자세로 보입니다. 포대는 바닥이나 하단 적재물에 놓였고 음식 보따리와 병은 전경 적재물 위에 놓인 것으로 보이며, 지지 없이 떠 있는 물체는 없습니다."
       },
       {
        "label": "B",
        "direction": "임정현은 몸을 약간 틀고 얼굴과 눈을 카메라 가까운 앞쪽으로 향합니다. 세 남자는 화면 왼쪽 뒤에서 등을 보인 채 초가 사이 안쪽으로 걸어갑니다. 임정현은 그들과 반대 방향을 보고 있어 응시 대상이 일치하지 않습니다. 보이는 이동 구간은 마당에 가까우며 내리막 숲길이라는 점은 뚜렷하지 않습니다.",
        "built_space": "왼쪽에 벽이 있는 초가 한 채, 중앙부터 오른쪽에 걸쳐 인물과 큰 나무에 가려진 마루형 초가가 보입니다. 전경에는 위쪽 포대 두 개와 하단에서 서로 겹친 포대 네 개가 부분적으로 보입니다. 가까운 포대가 크게 보이는 것은 카메라와의 거리로 설명됩니다. 임정현은 포대 뒤에 서 있고 세 남자는 그의 뒤쪽 건물 사이에 있습니다. 참고의 초가와 흙마당은 이어지지만, 이동로와 응시자의 앞뒤 관계가 잘못되었습니다.",
        "entities": "임정현으로 제시된 성인 한국인 남성 한 명과 멀어지는 남성 세 명이 보입니다. 미색의 낡은 옷과 검은 허리끈은 맞지만, 상투와 망건은 인물 참고에는 있어도 우선적으로 유지해야 할 이전 컷의 풀어 내린 검은 머리와 다릅니다. 얼굴은 참고 인물과 대체로 유사합니다. 전경에 쌀포대로 읽히는 포대들이 있으며 술병과 음식 보따리는 이 구도에서 보이지 않습니다. 세 남자의 얼굴은 확인할 수 없고 읽을 수 있는 글자는 없습니다.",
        "hard_violations": [
         "떠나는 세 사람을 임정현의 등 뒤에 배치하고 임정현은 반대쪽을 바라보게 하여, 세 사람의 뒷모습을 응시한다는 필수 인물 배치가 성립하지 않습니다."
        ],
        "physics": "임정현의 하체와 손은 대부분 포대에 가려져 있지만 상체는 자연스러운 직립 자세입니다. 보행자들은 지면에 닿은 발과 이동 중인 다리로 몸을 지탱합니다. 위쪽 포대들은 아래 포대에 눌려 얹혀 있으며 하단 적재물은 화면 밖으로 이어집니다. 보이는 범위에서 지지 없는 신체나 물체, 불가능한 관절 자세는 없습니다."
       }
      ],
      "all_candidates_fail": true
     },
     "normalized": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 2,
        "verdict_ko": "이전 컷의 머리와 의상은 잘 이어지지만, 세 사람을 등진 채 카메라를 바라보며 허벅지까지 넓혀 잡아 핵심 응시 관계와 미디엄 숏을 모두 놓쳤습니다."
       },
       {
        "label": "A",
        "score": 3,
        "verdict_ko": "무거운 정면과 미디엄 숏에는 더 가깝지만, 응시 대상이 등 뒤에 있어 핵심 연출이 성립하지 않고 상투·망건도 이전 컷의 머리와 불일치합니다."
       }
      ],
      "all_candidates_fail": true,
      "readings": [
       {
        "label": "B",
        "direction": "임정현의 얼굴과 눈은 화면 앞쪽, 거의 렌즈를 향합니다. 세 남자는 그의 오른쪽 뒤에서 카메라에 등을 보이고 숲길 안쪽으로 걸어갑니다. 따라서 눈길은 보이는 세 사람의 뒷모습에 닿지 않습니다. 길은 인물들 바로 앞에서 낮아진 뒤 먼 곳에서 다시 올라가는 형태로 보입니다.",
        "built_space": "흙마당 왼쪽에 기둥과 낮은 마루가 있는 초가 한 채, 오른쪽 가장자리에 초가 일부가 보이고 가운데 오른쪽으로 숲길이 열립니다. 전경에는 왼쪽의 세운 포대 하나와 하단·오른쪽의 눕힌 포대 세 개가 부분적으로 보입니다. 임정현은 포대 뒤에 서 있습니다. 초가·목재·흙의 재료는 참고 장소와 유사하지만, 세 사람의 이동로를 임정현 뒤에 배치한 공간 관계는 요구된 응시 장면과 맞지 않습니다.",
        "entities": "성인 한국인 남성으로 보이는 임정현 한 명과 뒷모습의 남성 세 명이 보입니다. 임정현의 풀어 내린 검은 머리, 얼굴 인상, 때 묻은 미색 옷과 검은 허리끈은 이전 컷에 가깝습니다. 세 남자의 얼굴과 세부 신원은 뒷모습과 흐림 때문에 확인할 수 없습니다. 포대, 갈색 술병 한 개, 일부 펼친 포장과 묶인 음식 보따리가 보이지만 보따리를 여는 동작은 없습니다. 읽을 수 있는 글자는 보이지 않습니다.",
        "hard_violations": [
         "떠나는 세 사람을 임정현의 등 뒤에 배치하고 임정현은 반대쪽을 바라보게 하여, 세 사람의 뒷모습을 응시한다는 필수 인물 배치가 성립하지 않습니다."
        ],
        "physics": "임정현의 화면 왼쪽 손은 세운 포대에 얹혀 있고 다른 손은 몸 옆에 내려와 있습니다. 발은 프레임 밖이지만 직립한 몸에 공중 부양 징후는 없습니다. 세 남자는 발을 지면에 딛거나 보행 중 들어 올린 자세로 보입니다. 포대는 바닥이나 하단 적재물에 놓였고 음식 보따리와 병은 전경 적재물 위에 놓인 것으로 보이며, 지지 없이 떠 있는 물체는 없습니다."
       },
       {
        "label": "A",
        "direction": "임정현은 몸을 약간 틀고 얼굴과 눈을 카메라 가까운 앞쪽으로 향합니다. 세 남자는 화면 왼쪽 뒤에서 등을 보인 채 초가 사이 안쪽으로 걸어갑니다. 임정현은 그들과 반대 방향을 보고 있어 응시 대상이 일치하지 않습니다. 보이는 이동 구간은 마당에 가까우며 내리막 숲길이라는 점은 뚜렷하지 않습니다.",
        "built_space": "왼쪽에 벽이 있는 초가 한 채, 중앙부터 오른쪽에 걸쳐 인물과 큰 나무에 가려진 마루형 초가가 보입니다. 전경에는 위쪽 포대 두 개와 하단에서 서로 겹친 포대 네 개가 부분적으로 보입니다. 가까운 포대가 크게 보이는 것은 카메라와의 거리로 설명됩니다. 임정현은 포대 뒤에 서 있고 세 남자는 그의 뒤쪽 건물 사이에 있습니다. 참고의 초가와 흙마당은 이어지지만, 이동로와 응시자의 앞뒤 관계가 잘못되었습니다.",
        "entities": "임정현으로 제시된 성인 한국인 남성 한 명과 멀어지는 남성 세 명이 보입니다. 미색의 낡은 옷과 검은 허리끈은 맞지만, 상투와 망건은 인물 참고에는 있어도 우선적으로 유지해야 할 이전 컷의 풀어 내린 검은 머리와 다릅니다. 얼굴은 참고 인물과 대체로 유사합니다. 전경에 쌀포대로 읽히는 포대들이 있으며 술병과 음식 보따리는 이 구도에서 보이지 않습니다. 세 남자의 얼굴은 확인할 수 없고 읽을 수 있는 글자는 없습니다.",
        "hard_violations": [
         "떠나는 세 사람을 임정현의 등 뒤에 배치하고 임정현은 반대쪽을 바라보게 하여, 세 사람의 뒷모습을 응시한다는 필수 인물 배치가 성립하지 않습니다."
        ],
        "physics": "임정현의 하체와 손은 대부분 포대에 가려져 있지만 상체는 자연스러운 직립 자세입니다. 보행자들은 지면에 닿은 발과 이동 중인 다리로 몸을 지탱합니다. 위쪽 포대들은 아래 포대에 눌려 얹혀 있으며 하단 적재물은 화면 밖으로 이어집니다. 보이는 범위에서 지지 없는 신체나 물체, 불가능한 관절 자세는 없습니다."
       }
      ]
     },
     "ok": true
    }
   ],
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "slot_winner_match": false,
   "slot_winner": {
    "gemini-pro": "B",
    "gpt-high": "A"
   },
   "route": "cross_slot_combined"
  },
  "dual": {
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "normalized": {
    "A": 1.375,
    "B": 1.667
   },
   "adjusted": {
    "A": 1.125,
    "B": 1.417
   },
   "violations": {
    "A": [
     "[gemini-pro] 이전 샷과 완벽히 일치해야 하는 외모 지침을 어기고 머리에 망건을 착용함 (캐릭터 외모 잠금 위반).",
     "[gpt-high] 떠나는 세 사람을 임정현의 등 뒤에 배치하고 임정현은 반대쪽을 바라보게 하여, 세 사람의 뒷모습을 응시한다는 필수 인물 배치가 성립하지 않습니다."
    ],
    "B": [
     "[gpt-high] 떠나는 세 사람을 임정현의 등 뒤에 배치하고 임정현은 반대쪽을 바라보게 하여, 세 사람의 뒷모습을 응시한다는 필수 인물 배치가 성립하지 않습니다."
    ]
   },
   "per_model_winner": {
    "gemini-pro": "B",
    "gpt-high": "A"
   },
   "agreed": false
  },
  "totals": {
   "B": 1417,
   "A": 1125
  },
  "selected": "B",
  "ranking": [
   "B",
   "A"
  ],
  "verdicts": [
   {
    "label": "B",
    "score": 1417,
    "verdict_ko": "카메라와 인물, 배경의 논리적 시선 방향에 물리적 모순이 존재하지만, 지문이 요구한 전경의 소품(술병, 음식 꾸러미)과 인물의 외모(자연스러운 검은 머리)를 이전 샷에 맞춰 매우 훌륭하게 구현했습니다.  ★위반: [gpt-high] 떠나는 세 사람을 임정현의 등 뒤에 배치하고 임정현은 반대쪽을 바라보게 하여, 세 사람의 뒷모습을 응시한다는 필수 인물 배치가 성립하지 않습니다."
   },
   {
    "label": "A",
    "score": 1125,
    "verdict_ko": "이전 샷과 레퍼런스에 없는 망건을 머리에 착용하여 캐릭터 외모 잠금 지침을 크게 위반했으며, 요구된 음식 꾸러미와 술병 소품도 누락되었습니다.  ★위반: [gemini-pro] 이전 샷과 완벽히 일치해야 하는 외모 지침을 어기고 머리에 망건을 착용함 (캐릭터 외모 잠금 위반). / [gpt-high] 떠나는 세 사람을 임정현의 등 뒤에 배치하고 임정현은 반대쪽을 바라보게 하여, 세 사람의 뒷모습을 응시한다는 필수 인물 배치가 성립하지 않습니다."
   }
  ],
  "refs": [
   {
    "label": "PREVIOUS SHOT STILL — a visually related earlier shot of this same place: the location's look, materials, fixed features and lighting mood are LOCKED to this photo; never copy its camera framing. The clothing, hair and overall look of 임정현 — who appear both in that photo and in this shot — are LOCKED to that photo. Anyone else visible in that photo is NOT in this shot: never carry their face, body or clothing onto anyone here. If this photo shows a character who cannot move (dead or unconscious), that character's exact pose, position and orientation are ALSO LOCKED — treat their body as a fixed prop of the set that only the camera moves around.",
    "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/db00517b-165b-4005-a287-f28936b59b95/images/272bd7af-9294-4490-98c6-88a40e9deacc/scene/recipe/S47sh13_sel.png",
    "asset_id": "bb348fae-53dc-48ba-8268-274034611bf9",
    "role": "prev_still"
   },
   {
    "label": "CHARACTER REFERENCE — 임정현: the exact person appearing in this shot; match face, hair and build exactly.",
    "path": "<bytes:858188>",
    "asset_id": "789c3359-d7f8-4832-9a6f-7e4ad2695c02",
    "role": "character_ref"
   }
  ],
  "critique_skipped": true,
  "shot_run_produced": false,
  "shot_run_uid": "06aa1435-4a74-7373-9b47-bb3edfaa4e57",
  "ref_mode": "prev+엔티티",
  "share_plan": {
   "ref_plan": "prev",
   "prev_anchor_tag": "S47sh13"
  }
 },
 "S47sh18::cine": {
  "applied": true,
  "attempted_at": "2026-09-09T06:23:38.739154+00:00",
  "fingerprint": "eeb6f97a4e6be0b6a2ba8823de1b2817c165c4a8c5a0c8f41cb6b6e393a70eb8",
  "fingerprint_version": 2,
  "provider": "mai",
  "endpoint": "openrouter/chat-completions",
  "model": "microsoft/mai-image-2.6",
  "pack": "24.202608252115",
  "source_file": "S47sh18_sel.png",
  "source_sha256": "59e453a3a3565bd33bbd82a76961413a7016544805dd5e502a9b6018d1c2108e",
  "file": "S47sh18_cine.png",
  "staged_sha256": "cf3c336f404e625b279288cec2d4196fae2ecfdff6a8f1b33bd9d8d738e658db",
  "latency_ms": 27543
 },
 "S48sh10::signage": {
  "fp": "86953ce476735d66",
  "inscriptions": [],
  "cues": [],
  "dropped": []
 },
 "S48sh10::bgfirst_bg": {
  "input_fingerprint": "d5ef5768391cf0a7",
  "prompt": "Create the EMPTY BACKGROUND PLATE for one film shot — NO PEOPLE, no figures, no body parts, no sketch lines, no arrows anywhere.\n\"Empty\" means no people only: KEEP the location's inherent occupants and stock that define the place — animals in an animal shelter, pen or farm, goods in a market, moored boats in a harbour — unless the shot text explicitly removes them.\nThe FIRST attached image is a thin-line storyboard sketch: use ONLY its camera angle, horizon, perspective and the placement/size of buildings and set masses — ignore the sketched people and arrows entirely. The SECOND attached image (LOCATION PHOTOGRAPH) is the real place: take its architecture, materials, signage and fixed features, and RE-PROJECT them into the sketch's camera. If the photograph's camera differs from the sketch's, the sketch's camera wins.\nHUMAN-SCALE CALIBRATION: derive every structure's true size from human-scale elements — a door ≈ 2m, a window ≈ 1–1.5m wide, one storey ≈ 2.5–3m; never inflate a small structure or shrink a large one.\n\nSHOT TEXT this background must serve (Korean): 다급히 뒤를 돌아본 이관회의 시선 너머로, 뒤쪽 골목을 꽉 채우며 나타난 조선인 일진회원들의 전신.\n\nLOCATION (lock): The rear approach alley to the straw-sandal stall in the outdoor market, now blocked across its width.\n\nTIME OF DAY (lock): day.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: wide shot\n- FRAME LAYOUT: rear alley blocked by the arriving Iljinhoe members in the middle-center of the frame, background.\n- KEY BACKGROUND ELEMENTS: 뒤쪽 골목 (나타난 일진회원들에 의해 통로가 막혔다) — The alley extends away from 이관회's near shoulder into the center of the image, with ground visible beneath the blocking group; used as Makes the loss of the retreat route unambiguous; 일진회원들이 찬 칼 (아직 뽑지 않은 채 몸에 차고 있다) — Visible at differing side angles beside individual bodies; used as Establishes the armed threat without advancing to the later sword-drawing beat.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Natural daytime illumination and muted, grounded contrast keep the entire blocked alley readable without adding stylized menace through unsupported lighting.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nRender ONE photorealistic empty location photograph, 16:9, that this shot can be staged inside later. No readable writing anywhere: surfaces that would carry writing may be present, but stage any wording out of legibility — an oblique angle, distance, shallow focus. No captions, watermarks or overlay text.",
  "effective_prompt": "Create the EMPTY BACKGROUND PLATE for one film shot — NO PEOPLE, no figures, no body parts, no sketch lines, no arrows anywhere.\n\"Empty\" means no people only: KEEP the location's inherent occupants and stock that define the place — animals in an animal shelter, pen or farm, goods in a market, moored boats in a harbour — unless the shot text explicitly removes them.\nThe FIRST attached image is a thin-line storyboard sketch: use ONLY its camera angle, horizon, perspective and the placement/size of buildings and set masses — ignore the sketched people and arrows entirely. The SECOND attached image (LOCATION PHOTOGRAPH) is the real place: take its architecture, materials, signage and fixed features, and RE-PROJECT them into the sketch's camera. If the photograph's camera differs from the sketch's, the sketch's camera wins.\nHUMAN-SCALE CALIBRATION: derive every structure's true size from human-scale elements — a door ≈ 2m, a window ≈ 1–1.5m wide, one storey ≈ 2.5–3m; never inflate a small structure or shrink a large one.\n\nSHOT TEXT this background must serve (Korean): 다급히 뒤를 돌아본 이관회의 시선 너머로, 뒤쪽 골목을 꽉 채우며 나타난 조선인 일진회원들의 전신.\n\nLOCATION (lock): The rear approach alley to the straw-sandal stall in the outdoor market, now blocked across its width.\n\nTIME OF DAY (lock): day.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: wide shot\n- FRAME LAYOUT: rear alley blocked by the arriving Iljinhoe members in the middle-center of the frame, background.\n- KEY BACKGROUND ELEMENTS: 뒤쪽 골목 (나타난 일진회원들에 의해 통로가 막혔다) — The alley extends away from 이관회's near shoulder into the center of the image, with ground visible beneath the blocking group; used as Makes the loss of the retreat route unambiguous; 일진회원들이 찬 칼 (아직 뽑지 않은 채 몸에 차고 있다) — Visible at differing side angles beside individual bodies; used as Establishes the armed threat without advancing to the later sword-drawing beat.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Natural daytime illumination and muted, grounded contrast keep the entire blocked alley readable without adding stylized menace through unsupported lighting.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nRender ONE photorealistic empty location photograph, 16:9, that this shot can be staged inside later. No readable writing anywhere: surfaces that would carry writing may be present, but stage any wording out of legibility — an oblique angle, distance, shallow focus. No captions, watermarks or overlay text.",
  "bg_path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/db00517b-165b-4005-a287-f28936b59b95/images/272bd7af-9294-4490-98c6-88a40e9deacc/scene/recipe/S48sh10__bgfirst_bg.png",
  "asset_id": "0adba7e5-5a3a-40f8-899a-5b2cdeef948d",
  "input_asset_ids": [
   "57a75fcf-991e-4505-a6b3-6fa21b051189",
   "32b543ac-1086-4282-b907-3fcd465b580a"
  ]
 },
 "S48sh10": {
  "input_fingerprint": "db9babb85e37789b",
  "prompt": "Create ONE FINAL photorealistic live-action film still of the moment below — early 20th-century Korea, Jeolla region; all people are Korean unless explicitly stated otherwise. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 다급히 뒤를 돌아본 이관회의 시선 너머로, 뒤쪽 골목을 꽉 채우며 나타난 조선인 일진회원들의 전신.\n\nLOCATION (lock): The rear approach alley to the straw-sandal stall in the outdoor market, now blocked across its width. The shot takes place here — the attached LOCATION STRUCTURE PHOTOGRAPH is the single authority for this exact place — its fixed structure and permanent site details are LOCKED to it. No separate location photograph exists for this place. Build everything else strictly from the location text above and the shot text; the layout sketch (when attached) governs framing and placement only, and the shot text governs time of day, lighting and action.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: wide shot\n- FRAME LAYOUT: rear alley blocked by the arriving Iljinhoe members in the middle-center of the frame, background.\n- KEY BACKGROUND ELEMENTS: 뒤쪽 골목 (나타난 일진회원들에 의해 통로가 막혔다) — The alley extends away from 이관회's near shoulder into the center of the image, with ground visible beneath the blocking group; used as Makes the loss of the retreat route unambiguous; 일진회원들이 찬 칼 (아직 뽑지 않은 채 몸에 차고 있다) — Visible at differing side angles beside individual bodies; used as Establishes the armed threat without advancing to the later sword-drawing beat.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Natural daytime illumination and muted, grounded contrast keep the entire blocked alley readable without adding stylized menace through unsupported lighting.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The familiar daytime market lanes are blocked at both ends.\n\nNo people appear unless the shot text itself says so.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.",
  "roll_prompts": {
   "A": "Stage the shot. The FIRST attached image (SHOT BACKGROUND) is the finished empty background of this shot — keep it EXACTLY: its camera, perspective, architecture, lighting and every fixed feature stay untouched. The SECOND attached image (LAYOUT SKETCH) tells you ONLY where the people go: each sketched person's position, screen size, pose and the gaze/motion arrows. Ignore the sketch's background lines. The CHARACTER REFERENCE photographs show the real people.\nPlace the real people into the background at exactly the sketched positions, sizes and poses, following the arrow directions. No sketch lines or arrows may remain.\n\nCreate ONE FINAL photorealistic live-action film still of the moment below — early 20th-century Korea, Jeolla region; all people are Korean unless explicitly stated otherwise. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 다급히 뒤를 돌아본 이관회의 시선 너머로, 뒤쪽 골목을 꽉 채우며 나타난 조선인 일진회원들의 전신.\n\nLOCATION (lock): The rear approach alley to the straw-sandal stall in the outdoor market, now blocked across its width. The shot takes place here — the FIRST attached image (SHOT BACKGROUND) is this exact place, already built: its ground, structures, horizon, materials and lighting are the finished truth of this location and must not be redesigned or replaced. No location photograph is attached — read the place from that image alone, and add no scenery, structure, vehicle or fixture that it does not already show. This lock governs the place only; the figures in the shot follow the staging and pose instructions.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: wide shot\n- FRAME LAYOUT: rear alley blocked by the arriving Iljinhoe members in the middle-center of the frame, background.\n- KEY BACKGROUND ELEMENTS: 뒤쪽 골목 (나타난 일진회원들에 의해 통로가 막혔다) — The alley extends away from 이관회's near shoulder into the center of the image, with ground visible beneath the blocking group; used as Makes the loss of the retreat route unambiguous; 일진회원들이 찬 칼 (아직 뽑지 않은 채 몸에 차고 있다) — Visible at differing side angles beside individual bodies; used as Establishes the armed threat without advancing to the later sword-drawing beat.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Natural daytime illumination and muted, grounded contrast keep the entire blocked alley readable without adding stylized menace through unsupported lighting.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The familiar daytime market lanes are blocked at both ends.\n\nNo people appear unless the shot text itself says so.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.",
   "B": "Create ONE FINAL photorealistic live-action film still of the moment below — early 20th-century Korea, Jeolla region; all people are Korean unless explicitly stated otherwise. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 다급히 뒤를 돌아본 이관회의 시선 너머로, 뒤쪽 골목을 꽉 채우며 나타난 조선인 일진회원들의 전신.\n\nLOCATION (lock): The rear approach alley to the straw-sandal stall in the outdoor market, now blocked across its width. The shot takes place here — the attached LOCATION STRUCTURE PHOTOGRAPH is the single authority for this exact place — its fixed structure and permanent site details are LOCKED to it. No separate location photograph exists for this place. Build everything else strictly from the location text above and the shot text; the layout sketch (when attached) governs framing and placement only, and the shot text governs time of day, lighting and action.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: wide shot\n- FRAME LAYOUT: rear alley blocked by the arriving Iljinhoe members in the middle-center of the frame, background.\n- KEY BACKGROUND ELEMENTS: 뒤쪽 골목 (나타난 일진회원들에 의해 통로가 막혔다) — The alley extends away from 이관회's near shoulder into the center of the image, with ground visible beneath the blocking group; used as Makes the loss of the retreat route unambiguous; 일진회원들이 찬 칼 (아직 뽑지 않은 채 몸에 차고 있다) — Visible at differing side angles beside individual bodies; used as Establishes the armed threat without advancing to the later sword-drawing beat.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Natural daytime illumination and muted, grounded contrast keep the entire blocked alley readable without adding stylized menace through unsupported lighting.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The familiar daytime market lanes are blocked at both ends.\n\nNo people appear unless the shot text itself says so.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.\n\nCOMPOSITION VARIATION (this candidate only): keep the place, the people and the moment exactly as specified, and keep from the CAMERA & FRAME contract its subject scale, screen placement and key background elements exactly. For this candidate only, its camera angle and height are deliberately left open: choose a camera position distinctly different from the obvious one, as a film director picking a second setup on the same blocking. Stay true to any explicit framing the SHOT TEXT itself dictates."
  },
  "roll_refs": {
   "A": [
    {
     "label": "SHOT BACKGROUND",
     "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/db00517b-165b-4005-a287-f28936b59b95/images/272bd7af-9294-4490-98c6-88a40e9deacc/scene/recipe/S48sh10__bgfirst_bg.png",
     "asset_id": "0adba7e5-5a3a-40f8-899a-5b2cdeef948d",
     "role": "bgfirst_bg"
    },
    {
     "label": "LAYOUT SKETCH (people placement only)",
     "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/db00517b-165b-4005-a287-f28936b59b95/images/272bd7af-9294-4490-98c6-88a40e9deacc/conti/conti_S48sh10.png",
     "asset_id": "57a75fcf-991e-4505-a6b3-6fa21b051189",
     "role": "conti_light"
    },
    {
     "label": "PROP REFERENCE — 칼: the exact object appearing in this shot; match its look, material and wear exactly.",
     "path": "<bytes:780122>",
     "asset_id": "82614014-6091-423c-89ca-1e15d209de72",
     "role": "prop_ref"
    }
   ],
   "B": [
    {
     "label": "LOCATION STRUCTURE PHOTOGRAPH — the confirmed photograph of this exact place and its fixed structure: it is the SINGLE authority for the location, the structure's shape, proportions, materials, colors, openings and every permanent site detail. Never copy its camera framing, time of day or lighting — the shot text is the authority for those.",
     "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/db00517b-165b-4005-a287-f28936b59b95/images/background_chain/seed_bg_market_sel.png",
     "asset_id": "32b543ac-1086-4282-b907-3fcd465b580a",
     "role": "location_seed_bg"
    },
    {
     "label": "PROP REFERENCE — 칼: the exact object appearing in this shot; match its look, material and wear exactly.",
     "path": "<bytes:780122>",
     "asset_id": "82614014-6091-423c-89ca-1e15d209de72",
     "role": "prop_ref"
    }
   ]
  },
  "shot_run_spend_attempt_count": 1,
  "judge_flip_skipped": "judge_fn_owns_order",
  "initial_roll_all_fail": true,
  "readings": [
   {
    "label": "A",
    "direction": "카메라는 전경에 있는 이관회의 오른쪽 어깨 너머로 배경의 일진회 무리를 향하고 있으며, 이관회의 시선 역시 무리를 향하는 것으로 보임. 배경의 무리는 이관회 쪽을 마주보고 있음.",
    "built_space": "초가집과 돌담이 늘어선 야외 장터 골목길이며, 참조 사진의 건물 구조 및 배치와 잘 일치함.",
    "entities": "전경의 이관회는 상투를 튼 뒷모습만 보임. 배경의 일진회 무리는 11명으로 모두 검은색 두루마기를 입고 있으나, 여러 인물이 완전히 동일한 얼굴(복제됨)을 하고 있음. 무리가 찬 칼은 대부분 형태가 불완전함.",
    "hard_violations": [
     "[gemini-pro] 배경 인물들의 얼굴이 여러 번 복제됨 (duplicated bodies)",
     "[gemini-pro] 일진회원들의 칼이 허리띠나 끈 등 어떠한 지지대 없이 허공에 떠 있거나 몸을 뚫고 지나감 (physically impossible staging / object with nothing to support it)",
     "[gpt-high] 칼을 차고 있는 인물로 지정된 일진회원들 외에, 전경의 이관회에게도 손으로 받친 별도의 칼을 추가했다. 이는 위협의 주체와 소품 배정을 바꾸는 비지정 소품이다."
    ],
    "physics": "인물들은 모두 땅에 서 있음. 그러나 중앙 및 주변 일진회원들의 칼이 지지 없이 공중에 떠 있거나 두루마기 천을 관통하여 꽂혀 있는 등 물리적으로 불가능한 상태임. 화면 맨 앞의 칼 역시 손으로 온전히 쥐고 있지 않고 떠 있는 듯 보임."
   },
   {
    "label": "B",
    "direction": "전경 좌측의 이관회는 등 뒤의 무리를 보지 않고 화면 우측 바깥을 엉뚱하게 응시하고 있음. 배경의 무리는 이관회가 아닌 카메라 앞쪽을 바라보고 있음.",
    "built_space": "양옆으로 초가집과 흙벽돌담이 늘어선 골목길로 참조 사진의 구조와 부합함.",
    "entities": "이관회는 회색 한복 상의를 입고 좌측에 서 있으며, 배경의 일진회 무리는 양복과 한복 등 20세기 초의 다양한 복장을 입고 있음. 우측 끝 인물 등이 들고 있는 무기는 칼자루가 없는 막대기 형태로 묘사됨.",
    "hard_violations": [
     "[gemini-pro] 지시된 구도와 어긋난 인물 배치: '이관회의 시선 너머로 골목이 이어지는' 구도가 요구되었으나, 이관회를 화면 좌측에 겉돌게 배치하여 카메라 뷰포인트와 샷의 레이아웃을 파괴함 (a person placed where the staging does not put them)"
    ],
    "physics": "인물들은 지면을 딛고 서 있음. 다만 이관회의 왼손 손가락이 심하게 뭉개져 있고 기형적으로 보임."
   }
  ],
  "cross_model_order": {
   "policy": "cross_model_order_v2_slot0_fwd_slot1_rev_parallel2",
   "slots": [
    {
     "model": "gemini-pro",
     "order": "forward",
     "display_to_canonical": {
      "A": "A",
      "B": "B"
     },
     "raw": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 2,
        "verdict_ko": "지시된 어깨 너머(OTS) 구도와 배경 배치는 프롬프트에 부합하나, 배경 인물들의 얼굴이 복제되었고 칼이 지지대 없이 몸을 관통하는 치명적인 물리적 오류가 발생했습니다."
       },
       {
        "label": "B",
        "score": 1,
        "verdict_ko": "인물의 시선이 목표 대상(일진회)을 벗어나 엉뚱한 곳을 향하며, 지시된 어깨 너머 구도를 완전히 무시하고 인물을 잘못된 위치에 배치하여 프레이밍 지침을 심각하게 위반했습니다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "카메라는 전경에 있는 이관회의 오른쪽 어깨 너머로 배경의 일진회 무리를 향하고 있으며, 이관회의 시선 역시 무리를 향하는 것으로 보임. 배경의 무리는 이관회 쪽을 마주보고 있음.",
        "built_space": "초가집과 돌담이 늘어선 야외 장터 골목길이며, 참조 사진의 건물 구조 및 배치와 잘 일치함.",
        "entities": "전경의 이관회는 상투를 튼 뒷모습만 보임. 배경의 일진회 무리는 11명으로 모두 검은색 두루마기를 입고 있으나, 여러 인물이 완전히 동일한 얼굴(복제됨)을 하고 있음. 무리가 찬 칼은 대부분 형태가 불완전함.",
        "hard_violations": [
         "배경 인물들의 얼굴이 여러 번 복제됨 (duplicated bodies)",
         "일진회원들의 칼이 허리띠나 끈 등 어떠한 지지대 없이 허공에 떠 있거나 몸을 뚫고 지나감 (physically impossible staging / object with nothing to support it)"
        ],
        "physics": "인물들은 모두 땅에 서 있음. 그러나 중앙 및 주변 일진회원들의 칼이 지지 없이 공중에 떠 있거나 두루마기 천을 관통하여 꽂혀 있는 등 물리적으로 불가능한 상태임. 화면 맨 앞의 칼 역시 손으로 온전히 쥐고 있지 않고 떠 있는 듯 보임."
       },
       {
        "label": "B",
        "direction": "전경 좌측의 이관회는 등 뒤의 무리를 보지 않고 화면 우측 바깥을 엉뚱하게 응시하고 있음. 배경의 무리는 이관회가 아닌 카메라 앞쪽을 바라보고 있음.",
        "built_space": "양옆으로 초가집과 흙벽돌담이 늘어선 골목길로 참조 사진의 구조와 부합함.",
        "entities": "이관회는 회색 한복 상의를 입고 좌측에 서 있으며, 배경의 일진회 무리는 양복과 한복 등 20세기 초의 다양한 복장을 입고 있음. 우측 끝 인물 등이 들고 있는 무기는 칼자루가 없는 막대기 형태로 묘사됨.",
        "hard_violations": [
         "지시된 구도와 어긋난 인물 배치: '이관회의 시선 너머로 골목이 이어지는' 구도가 요구되었으나, 이관회를 화면 좌측에 겉돌게 배치하여 카메라 뷰포인트와 샷의 레이아웃을 파괴함 (a person placed where the staging does not put them)"
        ],
        "physics": "인물들은 지면을 딛고 서 있음. 다만 이관회의 왼손 손가락이 심하게 뭉개져 있고 기형적으로 보임."
       }
      ],
      "all_candidates_fail": true
     },
     "normalized": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 2,
        "verdict_ko": "지시된 어깨 너머(OTS) 구도와 배경 배치는 프롬프트에 부합하나, 배경 인물들의 얼굴이 복제되었고 칼이 지지대 없이 몸을 관통하는 치명적인 물리적 오류가 발생했습니다."
       },
       {
        "label": "B",
        "score": 1,
        "verdict_ko": "인물의 시선이 목표 대상(일진회)을 벗어나 엉뚱한 곳을 향하며, 지시된 어깨 너머 구도를 완전히 무시하고 인물을 잘못된 위치에 배치하여 프레이밍 지침을 심각하게 위반했습니다."
       }
      ],
      "all_candidates_fail": true,
      "readings": [
       {
        "label": "A",
        "direction": "카메라는 전경에 있는 이관회의 오른쪽 어깨 너머로 배경의 일진회 무리를 향하고 있으며, 이관회의 시선 역시 무리를 향하는 것으로 보임. 배경의 무리는 이관회 쪽을 마주보고 있음.",
        "built_space": "초가집과 돌담이 늘어선 야외 장터 골목길이며, 참조 사진의 건물 구조 및 배치와 잘 일치함.",
        "entities": "전경의 이관회는 상투를 튼 뒷모습만 보임. 배경의 일진회 무리는 11명으로 모두 검은색 두루마기를 입고 있으나, 여러 인물이 완전히 동일한 얼굴(복제됨)을 하고 있음. 무리가 찬 칼은 대부분 형태가 불완전함.",
        "hard_violations": [
         "배경 인물들의 얼굴이 여러 번 복제됨 (duplicated bodies)",
         "일진회원들의 칼이 허리띠나 끈 등 어떠한 지지대 없이 허공에 떠 있거나 몸을 뚫고 지나감 (physically impossible staging / object with nothing to support it)"
        ],
        "physics": "인물들은 모두 땅에 서 있음. 그러나 중앙 및 주변 일진회원들의 칼이 지지 없이 공중에 떠 있거나 두루마기 천을 관통하여 꽂혀 있는 등 물리적으로 불가능한 상태임. 화면 맨 앞의 칼 역시 손으로 온전히 쥐고 있지 않고 떠 있는 듯 보임."
       },
       {
        "label": "B",
        "direction": "전경 좌측의 이관회는 등 뒤의 무리를 보지 않고 화면 우측 바깥을 엉뚱하게 응시하고 있음. 배경의 무리는 이관회가 아닌 카메라 앞쪽을 바라보고 있음.",
        "built_space": "양옆으로 초가집과 흙벽돌담이 늘어선 골목길로 참조 사진의 구조와 부합함.",
        "entities": "이관회는 회색 한복 상의를 입고 좌측에 서 있으며, 배경의 일진회 무리는 양복과 한복 등 20세기 초의 다양한 복장을 입고 있음. 우측 끝 인물 등이 들고 있는 무기는 칼자루가 없는 막대기 형태로 묘사됨.",
        "hard_violations": [
         "지시된 구도와 어긋난 인물 배치: '이관회의 시선 너머로 골목이 이어지는' 구도가 요구되었으나, 이관회를 화면 좌측에 겉돌게 배치하여 카메라 뷰포인트와 샷의 레이아웃을 파괴함 (a person placed where the staging does not put them)"
        ],
        "physics": "인물들은 지면을 딛고 서 있음. 다만 이관회의 왼손 손가락이 심하게 뭉개져 있고 기형적으로 보임."
       }
      ]
     },
     "ok": true
    },
    {
     "model": "gpt-high",
     "order": "reverse",
     "display_to_canonical": {
      "A": "B",
      "B": "A"
     },
     "raw": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 5,
        "verdict_ko": "골목을 가로막은 일진회원들의 전신과 미발도 상태는 구현했지만, 이관회가 그들이 아닌 카메라 쪽을 보아 핵심 시선 연결을 놓쳤다."
       },
       {
        "label": "B",
        "score": 3,
        "verdict_ko": "어깨 너머 시선과 중앙의 전신 배치는 더 정확하지만, 일진회원에게 지정된 칼을 전경의 이관회에게도 추가한 소품 위반으로 탈락한다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "왼쪽 전경의 이관회는 얼굴과 눈을 카메라 쪽으로 돌리고 있다. 화면 중앙 뒤쪽의 일진회원들을 바라보는 시선은 아니다. 일진회원들은 대체로 카메라와 전경 인물 쪽을 정면으로 응시한다. 칼은 허리 옆에 꽂혀 아래나 비스듬한 뒤쪽으로 향하며, 누군가를 향해 뽑아 겨누지는 않는다.",
        "built_space": "흙길 양옆에 초가지붕, 흙벽, 목재 문과 처마 기둥이 이어지고, 후방에도 여러 초가가 보인다. 왼쪽 가까운 처마 아래에는 굵은 기둥 두 개와 뒤쪽의 가는 지지대들이, 오른쪽에는 연속된 처마 지지대들이 보인다. 회원들은 양쪽 건물 사이를 촘촘하게 가로막고 발밑 흙바닥도 드러난다. 다만 참조의 오른쪽 전경을 따라 이어지는 긴 돌기단 흙담 대신 처마와 건물 정면이 강조되어 장소 구조의 일치도는 낮다.",
        "entities": "전경의 성인 남성 한 명과 뒤쪽의 성인 남성 약 16명이 보이며, 외형상 한국인 배역 설정에 어긋나는 뚜렷한 요소는 없다. 이관회는 회색 계열 한복과 머리띠를, 회원들은 한복과 서양식 재킷이 섞인 복장을 입었다. 앞줄 회원들은 머리부터 신발까지 보이고 뒷줄은 다른 회원들에게 가려진다. 여러 회원의 허리에 검은 칼집, 감은 손잡이, 금속 장식과 술이 보인다. 참조 소품의 대체적인 재료와 형태는 따르지만 세부 문양과 마모까지 확인되지는 않는다. 읽을 수 있는 글자는 없다.",
        "hard_violations": [],
        "physics": "앞줄 회원들의 신발이 흙바닥에 닿아 체중을 지탱한다. 뒷줄은 하체가 가려져 있지만 공중에 떠 있다는 징후는 없다. 이관회의 하체 일부는 화면 밖이며, 보이는 몸통의 숙임은 서 있는 자세에서 가능한 범위다. 칼집은 허리띠와 매달림 장치에 연결되어 있어 지지 없는 부유물로 보이지 않는다."
       },
       {
        "label": "B",
        "direction": "왼쪽 전경에서 이관회의 뒤통수와 옆얼굴이 보이고, 얼굴 방향은 골목 중앙의 회원들을 향한다. 회원들도 전경의 이관회 쪽을 보고 있어 시선 관계가 맞는다. 회원들의 칼집은 허리에서 아래로 향한다. 전경 이관회 앞에는 별도의 칼 손잡이가 오른쪽 위로 솟아 있지만, 노출된 칼날이나 상대를 겨누는 동작은 보이지 않는다.",
        "built_space": "가까운 어깨 너머로 흙길이 중앙까지 이어지며, 회원 11명이 길 폭에 걸쳐 두세 겹으로 서 있다. 양옆 초가의 목재 기둥과 문, 오른쪽의 긴 돌기단 흙담 한 줄, 뒤쪽 초가군과 계단식 언덕이 보인다. 오른쪽 담과 왼쪽 처마 아래 통로의 관계는 참조에 더 가깝다. 회원들 사이에 간격은 있지만 양쪽 가장자리까지 점유해 퇴로 차단이 읽힌다. 카메라는 다소 높은 위치이나 프롬프트가 시점을 특정하지 않았으므로 그 자체는 위반이 아니다.",
        "entities": "전경의 성인 남성 한 명과 검은 두루마기·머리쓰개를 착용한 성인 남성 회원 11명이 보인다. 모두 한국인 배역으로 자연스럽게 읽히며 별도의 신원 참조는 없다. 회원들의 전신과 신발이 대부분 드러나고, 검은 칼집과 감은 손잡이의 칼이 허리에 있다. 전경에는 참조와 유사한 마름모꼴 감기와 금속 코등이를 가진 칼이 이관회에게도 추가되어 있다. 읽을 수 있는 글자나 화면 표식은 없다.",
        "hard_violations": [
         "칼을 차고 있는 인물로 지정된 일진회원들 외에, 전경의 이관회에게도 손으로 받친 별도의 칼을 추가했다. 이는 위협의 주체와 소품 배정을 바꾸는 비지정 소품이다."
        ],
        "physics": "회원들은 발을 벌리거나 한 발을 앞으로 내민 채 흙바닥에 서 있고, 이동 직후의 자세로 가능한 체중 배분을 보인다. 허리의 칼은 띠에 매달려 있다. 전경의 추가 칼도 코등이 아래 부분에 손이 닿아 지지되므로 떠 있는 물체는 아니다. 이관회의 화면 밖 하체는 판단 대상이 아니며, 보이는 어깨와 목의 회전에는 명백한 해부학적 불가능성이 없다."
       }
      ],
      "all_candidates_fail": true
     },
     "normalized": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 5,
        "verdict_ko": "골목을 가로막은 일진회원들의 전신과 미발도 상태는 구현했지만, 이관회가 그들이 아닌 카메라 쪽을 보아 핵심 시선 연결을 놓쳤다."
       },
       {
        "label": "A",
        "score": 3,
        "verdict_ko": "어깨 너머 시선과 중앙의 전신 배치는 더 정확하지만, 일진회원에게 지정된 칼을 전경의 이관회에게도 추가한 소품 위반으로 탈락한다."
       }
      ],
      "all_candidates_fail": true,
      "readings": [
       {
        "label": "B",
        "direction": "왼쪽 전경의 이관회는 얼굴과 눈을 카메라 쪽으로 돌리고 있다. 화면 중앙 뒤쪽의 일진회원들을 바라보는 시선은 아니다. 일진회원들은 대체로 카메라와 전경 인물 쪽을 정면으로 응시한다. 칼은 허리 옆에 꽂혀 아래나 비스듬한 뒤쪽으로 향하며, 누군가를 향해 뽑아 겨누지는 않는다.",
        "built_space": "흙길 양옆에 초가지붕, 흙벽, 목재 문과 처마 기둥이 이어지고, 후방에도 여러 초가가 보인다. 왼쪽 가까운 처마 아래에는 굵은 기둥 두 개와 뒤쪽의 가는 지지대들이, 오른쪽에는 연속된 처마 지지대들이 보인다. 회원들은 양쪽 건물 사이를 촘촘하게 가로막고 발밑 흙바닥도 드러난다. 다만 참조의 오른쪽 전경을 따라 이어지는 긴 돌기단 흙담 대신 처마와 건물 정면이 강조되어 장소 구조의 일치도는 낮다.",
        "entities": "전경의 성인 남성 한 명과 뒤쪽의 성인 남성 약 16명이 보이며, 외형상 한국인 배역 설정에 어긋나는 뚜렷한 요소는 없다. 이관회는 회색 계열 한복과 머리띠를, 회원들은 한복과 서양식 재킷이 섞인 복장을 입었다. 앞줄 회원들은 머리부터 신발까지 보이고 뒷줄은 다른 회원들에게 가려진다. 여러 회원의 허리에 검은 칼집, 감은 손잡이, 금속 장식과 술이 보인다. 참조 소품의 대체적인 재료와 형태는 따르지만 세부 문양과 마모까지 확인되지는 않는다. 읽을 수 있는 글자는 없다.",
        "hard_violations": [],
        "physics": "앞줄 회원들의 신발이 흙바닥에 닿아 체중을 지탱한다. 뒷줄은 하체가 가려져 있지만 공중에 떠 있다는 징후는 없다. 이관회의 하체 일부는 화면 밖이며, 보이는 몸통의 숙임은 서 있는 자세에서 가능한 범위다. 칼집은 허리띠와 매달림 장치에 연결되어 있어 지지 없는 부유물로 보이지 않는다."
       },
       {
        "label": "A",
        "direction": "왼쪽 전경에서 이관회의 뒤통수와 옆얼굴이 보이고, 얼굴 방향은 골목 중앙의 회원들을 향한다. 회원들도 전경의 이관회 쪽을 보고 있어 시선 관계가 맞는다. 회원들의 칼집은 허리에서 아래로 향한다. 전경 이관회 앞에는 별도의 칼 손잡이가 오른쪽 위로 솟아 있지만, 노출된 칼날이나 상대를 겨누는 동작은 보이지 않는다.",
        "built_space": "가까운 어깨 너머로 흙길이 중앙까지 이어지며, 회원 11명이 길 폭에 걸쳐 두세 겹으로 서 있다. 양옆 초가의 목재 기둥과 문, 오른쪽의 긴 돌기단 흙담 한 줄, 뒤쪽 초가군과 계단식 언덕이 보인다. 오른쪽 담과 왼쪽 처마 아래 통로의 관계는 참조에 더 가깝다. 회원들 사이에 간격은 있지만 양쪽 가장자리까지 점유해 퇴로 차단이 읽힌다. 카메라는 다소 높은 위치이나 프롬프트가 시점을 특정하지 않았으므로 그 자체는 위반이 아니다.",
        "entities": "전경의 성인 남성 한 명과 검은 두루마기·머리쓰개를 착용한 성인 남성 회원 11명이 보인다. 모두 한국인 배역으로 자연스럽게 읽히며 별도의 신원 참조는 없다. 회원들의 전신과 신발이 대부분 드러나고, 검은 칼집과 감은 손잡이의 칼이 허리에 있다. 전경에는 참조와 유사한 마름모꼴 감기와 금속 코등이를 가진 칼이 이관회에게도 추가되어 있다. 읽을 수 있는 글자나 화면 표식은 없다.",
        "hard_violations": [
         "칼을 차고 있는 인물로 지정된 일진회원들 외에, 전경의 이관회에게도 손으로 받친 별도의 칼을 추가했다. 이는 위협의 주체와 소품 배정을 바꾸는 비지정 소품이다."
        ],
        "physics": "회원들은 발을 벌리거나 한 발을 앞으로 내민 채 흙바닥에 서 있고, 이동 직후의 자세로 가능한 체중 배분을 보인다. 허리의 칼은 띠에 매달려 있다. 전경의 추가 칼도 코등이 아래 부분에 손이 닿아 지지되므로 떠 있는 물체는 아니다. 이관회의 화면 밖 하체는 판단 대상이 아니며, 보이는 어깨와 목의 회전에는 명백한 해부학적 불가능성이 없다."
       }
      ]
     },
     "ok": true
    }
   ],
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "slot_winner_match": false,
   "slot_winner": {
    "gemini-pro": "A",
    "gpt-high": "B"
   },
   "route": "cross_slot_combined"
  },
  "dual": {
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "normalized": {
    "A": 1.6,
    "B": 1.5
   },
   "adjusted": {
    "A": 1.35,
    "B": 1.25
   },
   "violations": {
    "A": [
     "[gemini-pro] 배경 인물들의 얼굴이 여러 번 복제됨 (duplicated bodies)",
     "[gemini-pro] 일진회원들의 칼이 허리띠나 끈 등 어떠한 지지대 없이 허공에 떠 있거나 몸을 뚫고 지나감 (physically impossible staging / object with nothing to support it)",
     "[gpt-high] 칼을 차고 있는 인물로 지정된 일진회원들 외에, 전경의 이관회에게도 손으로 받친 별도의 칼을 추가했다. 이는 위협의 주체와 소품 배정을 바꾸는 비지정 소품이다."
    ],
    "B": [
     "[gemini-pro] 지시된 구도와 어긋난 인물 배치: '이관회의 시선 너머로 골목이 이어지는' 구도가 요구되었으나, 이관회를 화면 좌측에 겉돌게 배치하여 카메라 뷰포인트와 샷의 레이아웃을 파괴함 (a person placed where the staging does not put them)"
    ]
   },
   "per_model_winner": {
    "gemini-pro": "A",
    "gpt-high": "B"
   },
   "agreed": false
  },
  "totals": {
   "A": 1350,
   "B": 1250
  },
  "selected": "A",
  "ranking": [
   "A",
   "B"
  ],
  "verdicts": [
   {
    "label": "A",
    "score": 1350,
    "verdict_ko": "지시된 어깨 너머(OTS) 구도와 배경 배치는 프롬프트에 부합하나, 배경 인물들의 얼굴이 복제되었고 칼이 지지대 없이 몸을 관통하는 치명적인 물리적 오류가 발생했습니다.  ★위반: [gemini-pro] 배경 인물들의 얼굴이 여러 번 복제됨 (duplicated bodies) / [gemini-pro] 일진회원들의 칼이 허리띠나 끈 등 어떠한 지지대 없이 허공에 떠 있거나 몸을 뚫고 지나감 (physically impossible staging / object with nothing to support it) / [gpt-high] 칼을 차고 있는 인물로 지정된 일진회원들 외에, 전경의 이관회에게도 손으로 받친 별도의 칼을 추가했다. 이는 위협의 주체와 소품 배정을 바꾸는 비지정 소품이다."
   },
   {
    "label": "B",
    "score": 1250,
    "verdict_ko": "인물의 시선이 목표 대상(일진회)을 벗어나 엉뚱한 곳을 향하며, 지시된 어깨 너머 구도를 완전히 무시하고 인물을 잘못된 위치에 배치하여 프레이밍 지침을 심각하게 위반했습니다.  ★위반: [gemini-pro] 지시된 구도와 어긋난 인물 배치: '이관회의 시선 너머로 골목이 이어지는' 구도가 요구되었으나, 이관회를 화면 좌측에 겉돌게 배치하여 카메라 뷰포인트와 샷의 레이아웃을 파괴함 (a person placed where the staging does not put them)"
   }
  ],
  "refs": [
   {
    "label": "LOCATION STRUCTURE PHOTOGRAPH — the confirmed photograph of this exact place and its fixed structure: it is the SINGLE authority for the location, the structure's shape, proportions, materials, colors, openings and every permanent site detail. Never copy its camera framing, time of day or lighting — the shot text is the authority for those.",
    "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/db00517b-165b-4005-a287-f28936b59b95/images/background_chain/seed_bg_market_sel.png",
    "asset_id": "32b543ac-1086-4282-b907-3fcd465b580a",
    "role": "location_seed_bg"
   },
   {
    "label": "PROP REFERENCE — 칼: the exact object appearing in this shot; match its look, material and wear exactly.",
    "path": "<bytes:780122>",
    "asset_id": "82614014-6091-423c-89ca-1e15d209de72",
    "role": "prop_ref"
   }
  ],
  "critique_skipped": true,
  "needs_reshoot": true,
  "shot_run_produced": false,
  "shot_run_uid": "06aa1435-4beb-712d-b0f6-8eef29a39242",
  "bgfirst": {
   "bg_path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/db00517b-165b-4005-a287-f28936b59b95/images/272bd7af-9294-4490-98c6-88a40e9deacc/scene/recipe/S48sh10__bgfirst_bg.png",
   "bg_asset_id": "0adba7e5-5a3a-40f8-899a-5b2cdeef948d",
   "bg_record_key": "S48sh10::bgfirst_bg",
   "chain_winner": true,
   "authority": "seed_bg"
  },
  "ref_mode": "재투영 배경+콘티+엔티티 (2택1: 체인 승)",
  "share_plan": {
   "ref_plan": "background"
  },
  "lane_policy": "ab_select_ready"
 },
 "S48sh10::cine": {
  "applied": true,
  "attempted_at": "2026-09-09T06:27:17.065602+00:00",
  "fingerprint": "61dc77b9cac968034d20530f3c6e91725f9c6181e7c370ec6d68070e9cb9f3a2",
  "fingerprint_version": 2,
  "provider": "mai",
  "endpoint": "openrouter/chat-completions",
  "model": "microsoft/mai-image-2.6",
  "pack": "24.202608252115",
  "source_file": "S48sh10_sel.png",
  "source_sha256": "e2d8da613e3ff212676100cfff4f1939ca496db127f13de3d2b7c9ed3aac0000",
  "file": "S48sh10_cine.png",
  "staged_sha256": "c8252552ac540760d0499054bce4fcf8727844ea7a75f3df53b17d63a0d27da6",
  "latency_ms": 29634
 },
 "S48sh12::signage": {
  "fp": "aa9d1c94d9ae110d",
  "inscriptions": [],
  "cues": [],
  "dropped": []
 },
 "S48sh12": {
  "input_fingerprint": "fd9ed24ea3cb6931",
  "prompt": "Create ONE FINAL photorealistic live-action film still of the moment below — early 20th-century Korea, Jeolla region; all people are Korean unless explicitly stated otherwise. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 안규홍의 등짝에 억센 두 손을 거칠게 밀착시켜 미는 힘을 가하고 있는 박상현의 접촉 순간.\n\nLOCATION (lock): The dirt lane directly outside the market's straw-sandal stall, between the blocked front and rear approaches. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: close-up\n- KEY BACKGROUND ELEMENTS: Forward market alley (The route ahead is blocked in the scene; only a narrow portion of the alley is visible beyond the bodies) — The alley recedes beyond 안규홍 in the direction of the shove; used as Provides a restrained depth cue without competing with the hands against his back.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Daylight with restrained saturation and naturalistic contrast keeps the violent contact immediate and unembellished.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The people visible in that photo are NOT in this shot: never carry their faces, bodies or clothing onto anyone here. Each person in THIS shot is defined solely by the PEOPLE list, their own CHARACTER REFERENCE images and their wardrobe notes. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): Both market escape lanes remain blocked. 박상현: He stands with his arms extended in a forceful shove, beginning to flee behind the blockade. 안규홍: He staggers with his sickle retained; his eyes remain swollen, face dirty, hair tied and neck bare.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 안규홍 (한국인 남성, 청년 나이대, 검은 머리, 단단히 묶은 머리); 박상현 (한국인 남성, 성인 나이대, 자연스러운 검은 머리) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.",
  "roll_prompts": {
   "A": "Create ONE FINAL photorealistic live-action film still of the moment below — early 20th-century Korea, Jeolla region; all people are Korean unless explicitly stated otherwise. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 안규홍의 등짝에 억센 두 손을 거칠게 밀착시켜 미는 힘을 가하고 있는 박상현의 접촉 순간.\n\nLOCATION (lock): The dirt lane directly outside the market's straw-sandal stall, between the blocked front and rear approaches. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: close-up\n- KEY BACKGROUND ELEMENTS: Forward market alley (The route ahead is blocked in the scene; only a narrow portion of the alley is visible beyond the bodies) — The alley recedes beyond 안규홍 in the direction of the shove; used as Provides a restrained depth cue without competing with the hands against his back.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Daylight with restrained saturation and naturalistic contrast keeps the violent contact immediate and unembellished.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The people visible in that photo are NOT in this shot: never carry their faces, bodies or clothing onto anyone here. Each person in THIS shot is defined solely by the PEOPLE list, their own CHARACTER REFERENCE images and their wardrobe notes. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): Both market escape lanes remain blocked. 박상현: He stands with his arms extended in a forceful shove, beginning to flee behind the blockade. 안규홍: He staggers with his sickle retained; his eyes remain swollen, face dirty, hair tied and neck bare.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 안규홍 (한국인 남성, 청년 나이대, 검은 머리, 단단히 묶은 머리); 박상현 (한국인 남성, 성인 나이대, 자연스러운 검은 머리) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.",
   "B": "Create ONE FINAL photorealistic live-action film still of the moment below — early 20th-century Korea, Jeolla region; all people are Korean unless explicitly stated otherwise. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 안규홍의 등짝에 억센 두 손을 거칠게 밀착시켜 미는 힘을 가하고 있는 박상현의 접촉 순간.\n\nLOCATION (lock): The dirt lane directly outside the market's straw-sandal stall, between the blocked front and rear approaches. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: close-up\n- KEY BACKGROUND ELEMENTS: Forward market alley (The route ahead is blocked in the scene; only a narrow portion of the alley is visible beyond the bodies) — The alley recedes beyond 안규홍 in the direction of the shove; used as Provides a restrained depth cue without competing with the hands against his back.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Daylight with restrained saturation and naturalistic contrast keeps the violent contact immediate and unembellished.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The people visible in that photo are NOT in this shot: never carry their faces, bodies or clothing onto anyone here. Each person in THIS shot is defined solely by the PEOPLE list, their own CHARACTER REFERENCE images and their wardrobe notes. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): Both market escape lanes remain blocked. 박상현: He stands with his arms extended in a forceful shove, beginning to flee behind the blockade. 안규홍: He staggers with his sickle retained; his eyes remain swollen, face dirty, hair tied and neck bare.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 안규홍 (한국인 남성, 청년 나이대, 검은 머리, 단단히 묶은 머리); 박상현 (한국인 남성, 성인 나이대, 자연스러운 검은 머리) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.\n\nCOMPOSITION VARIATION (this candidate only): keep the place, the people and the moment exactly as specified, and keep from the CAMERA & FRAME contract its subject scale, screen placement and key background elements exactly. For this candidate only, its camera angle and height are deliberately left open: choose a camera position distinctly different from the obvious one, as a film director picking a second setup on the same blocking. Stay true to any explicit framing the SHOT TEXT itself dictates."
  },
  "shot_run_spend_attempt_count": 1,
  "initial_roll_all_fail": true,
  "readings": [
   {
    "label": "A",
    "direction": "박상현의 시선과 두 손이 안규홍의 등(목표물)을 향해 있으며 거칠게 미는 동작을 취하고 있음.",
    "built_space": "흙길과 초가집들이 기준 이미지와 유사하게 배치되어 있으며 카메라 구도상 골목길 원경이 나타남.",
    "entities": "안규홍(묶은 머리, 흰옷)과 박상현(갓, 회색 한복)의 외형은 일치하나, 샷 텍스트에 명시되지 않은 이전 장면의 검은 옷 무리가 배경에 등장함.",
    "hard_violations": [
     "[gemini-pro] 샷 텍스트에 없는 인물 추가 (배경의 검은 옷 무리)",
     "[gemini-pro] 물리적으로 불가능한 해부학 (허공에서 뻗어 나오는 박상현의 왼팔)",
     "[gpt-high] 배경 골목에 검은 옷의 봉쇄 인물 여러 명을 추가했다. 이전 스틸의 사람을 이어 넣지 말고 안규홍과 박상현만 허용한다는 명시적 인물 제한을 위반한다."
    ],
    "physics": "박상현의 두 손이 등에 밀착되어 있으나, 그가 안규홍의 오른편에 서 있음에도 불구하고 왼손과 소매가 안규홍의 왼쪽 뒤편 허공에서 동떨어져 뻗어 나오는 불가능한 해부학을 보임."
   },
   {
    "label": "B",
    "direction": "1인칭 시점의 두 손이 안규홍의 등(목표물)을 밀고 있으며, 정체불명의 얼굴이 카메라 측을 응시하고 있음.",
    "built_space": "흙길과 짚신 가판대 등 요소가 일부 등장하나 기준 이미지의 공간 구조 및 비율과 다소 어긋남.",
    "entities": "안규홍의 뒷모습은 나타나나 박상현의 지정된 기준 이미지(갓, 회색 한복)와 일치하는 인물이 화면에 없으며, 프롬프트에 없는 제3자의 얼굴이 추가됨.",
    "hard_violations": [
     "[gemini-pro] 샷 텍스트에 없는 인물 추가 (정체불명의 낯선 얼굴)",
     "[gemini-pro] 물리적으로 불가능한 해부학 (안규홍의 몸과 제3자의 얼굴이 기괴하게 융합됨)",
     "[gpt-high] 안규홍 앞쪽에 남성의 얼굴과 몸이 있는데 미는 두 팔은 카메라 쪽 아래에서 들어온다. 이를 박상현 한 사람의 몸으로 연결할 수 없으므로, 허용되지 않은 제삼자의 신체 일부가 추가되었거나 불가능한 신체 배치가 된 장면이다."
    ],
    "physics": "1인칭 시점의 손이 등을 밀고 있으나, 피사체의 몸 앞쪽에서 제3자의 얼굴이 튀어나와 있어 두 인물의 신체가 비정상적으로 결합된 불가능한 형태임."
   }
  ],
  "cross_model_order": {
   "policy": "cross_model_order_v2_slot0_fwd_slot1_rev_parallel2",
   "slots": [
    {
     "model": "gemini-pro",
     "order": "forward",
     "display_to_canonical": {
      "A": "A",
      "B": "B"
     },
     "raw": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 2,
        "verdict_ko": "프롬프트에서 엄격히 금지한 이전 샷의 인물들을 배경에 추가하였고, 박상현의 왼팔이 불가능한 위치에서 나오는 해부학적 오류가 발생하여 실격입니다."
       },
       {
        "label": "B",
        "score": 1,
        "verdict_ko": "박상현을 1인칭 시점으로 처리해 외형을 누락시켰으며, 안규홍의 몸 앞으로 명시되지 않은 제3자의 얼굴이 기괴하게 융합되는 치명적인 해부학적 오류가 있습니다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "박상현의 시선과 두 손이 안규홍의 등(목표물)을 향해 있으며 거칠게 미는 동작을 취하고 있음.",
        "built_space": "흙길과 초가집들이 기준 이미지와 유사하게 배치되어 있으며 카메라 구도상 골목길 원경이 나타남.",
        "entities": "안규홍(묶은 머리, 흰옷)과 박상현(갓, 회색 한복)의 외형은 일치하나, 샷 텍스트에 명시되지 않은 이전 장면의 검은 옷 무리가 배경에 등장함.",
        "hard_violations": [
         "샷 텍스트에 없는 인물 추가 (배경의 검은 옷 무리)",
         "물리적으로 불가능한 해부학 (허공에서 뻗어 나오는 박상현의 왼팔)"
        ],
        "physics": "박상현의 두 손이 등에 밀착되어 있으나, 그가 안규홍의 오른편에 서 있음에도 불구하고 왼손과 소매가 안규홍의 왼쪽 뒤편 허공에서 동떨어져 뻗어 나오는 불가능한 해부학을 보임."
       },
       {
        "label": "B",
        "direction": "1인칭 시점의 두 손이 안규홍의 등(목표물)을 밀고 있으며, 정체불명의 얼굴이 카메라 측을 응시하고 있음.",
        "built_space": "흙길과 짚신 가판대 등 요소가 일부 등장하나 기준 이미지의 공간 구조 및 비율과 다소 어긋남.",
        "entities": "안규홍의 뒷모습은 나타나나 박상현의 지정된 기준 이미지(갓, 회색 한복)와 일치하는 인물이 화면에 없으며, 프롬프트에 없는 제3자의 얼굴이 추가됨.",
        "hard_violations": [
         "샷 텍스트에 없는 인물 추가 (정체불명의 낯선 얼굴)",
         "물리적으로 불가능한 해부학 (안규홍의 몸과 제3자의 얼굴이 기괴하게 융합됨)"
        ],
        "physics": "1인칭 시점의 손이 등을 밀고 있으나, 피사체의 몸 앞쪽에서 제3자의 얼굴이 튀어나와 있어 두 인물의 신체가 비정상적으로 결합된 불가능한 형태임."
       }
      ],
      "all_candidates_fail": true
     },
     "normalized": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 2,
        "verdict_ko": "프롬프트에서 엄격히 금지한 이전 샷의 인물들을 배경에 추가하였고, 박상현의 왼팔이 불가능한 위치에서 나오는 해부학적 오류가 발생하여 실격입니다."
       },
       {
        "label": "B",
        "score": 1,
        "verdict_ko": "박상현을 1인칭 시점으로 처리해 외형을 누락시켰으며, 안규홍의 몸 앞으로 명시되지 않은 제3자의 얼굴이 기괴하게 융합되는 치명적인 해부학적 오류가 있습니다."
       }
      ],
      "all_candidates_fail": true,
      "readings": [
       {
        "label": "A",
        "direction": "박상현의 시선과 두 손이 안규홍의 등(목표물)을 향해 있으며 거칠게 미는 동작을 취하고 있음.",
        "built_space": "흙길과 초가집들이 기준 이미지와 유사하게 배치되어 있으며 카메라 구도상 골목길 원경이 나타남.",
        "entities": "안규홍(묶은 머리, 흰옷)과 박상현(갓, 회색 한복)의 외형은 일치하나, 샷 텍스트에 명시되지 않은 이전 장면의 검은 옷 무리가 배경에 등장함.",
        "hard_violations": [
         "샷 텍스트에 없는 인물 추가 (배경의 검은 옷 무리)",
         "물리적으로 불가능한 해부학 (허공에서 뻗어 나오는 박상현의 왼팔)"
        ],
        "physics": "박상현의 두 손이 등에 밀착되어 있으나, 그가 안규홍의 오른편에 서 있음에도 불구하고 왼손과 소매가 안규홍의 왼쪽 뒤편 허공에서 동떨어져 뻗어 나오는 불가능한 해부학을 보임."
       },
       {
        "label": "B",
        "direction": "1인칭 시점의 두 손이 안규홍의 등(목표물)을 밀고 있으며, 정체불명의 얼굴이 카메라 측을 응시하고 있음.",
        "built_space": "흙길과 짚신 가판대 등 요소가 일부 등장하나 기준 이미지의 공간 구조 및 비율과 다소 어긋남.",
        "entities": "안규홍의 뒷모습은 나타나나 박상현의 지정된 기준 이미지(갓, 회색 한복)와 일치하는 인물이 화면에 없으며, 프롬프트에 없는 제3자의 얼굴이 추가됨.",
        "hard_violations": [
         "샷 텍스트에 없는 인물 추가 (정체불명의 낯선 얼굴)",
         "물리적으로 불가능한 해부학 (안규홍의 몸과 제3자의 얼굴이 기괴하게 융합됨)"
        ],
        "physics": "1인칭 시점의 손이 등을 밀고 있으나, 피사체의 몸 앞쪽에서 제3자의 얼굴이 튀어나와 있어 두 인물의 신체가 비정상적으로 결합된 불가능한 형태임."
       }
      ]
     },
     "ok": true
    },
    {
     "model": "gpt-high",
     "order": "reverse",
     "display_to_canonical": {
      "A": "B",
      "B": "A"
     },
     "raw": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 1,
        "verdict_ko": "등과 두 손의 클로즈업은 맞지만, 앞쪽 남성과 뒤쪽 미는 손의 위치가 연결되지 않아 제삼자 또는 불가능한 신체 배치가 발생한다."
       },
       {
        "label": "B",
        "score": 4,
        "verdict_ko": "박상현이 안규홍의 등을 두 손으로 미는 접촉과 의상은 충실하지만, 배경에 금지된 봉쇄 인물들을 추가해 최종 사용 조건을 위반한다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "화면 아래에서 들어온 두 손바닥은 안규홍의 등 중앙에 닿아 화면 안쪽으로 민다. 안규홍은 카메라에 등을 돌리고 양팔을 옆으로 벌렸다. 그런데 다른 남성의 얼굴은 안규홍 앞쪽에서 카메라 쪽을 향하고 있어, 뒤에서 미는 박상현의 위치와 맞지 않는다. 골목의 깊이는 몸 사이보다 왼쪽 가판 뒤로 드러난다.",
        "built_space": "왼쪽에 짚신을 펼친 가판 한 구역, 매달린 짚신 여러 켤레와 바구니들이 있고, 오른쪽에는 목재 문과 기둥이 이어진다. 흙길과 낡은 목재는 시장 재질에 부합하지만, 참고의 초가와 돌기단 배치를 확인할 만큼 배경이 보이지는 않는다. 고정 시설의 명백한 중복이나 반사는 없다.",
        "entities": "등을 보인 인물은 검은 머리를 묶고 때 묻은 밝은 저고리를 입어 안규홍의 보이는 특징과 대체로 맞는다. 얼굴과 눈, 낫은 프레임 밖이므로 확인할 수 없다. 맞은편 남성은 성인 동아시아계 남성으로 보이지만, 흐트러진 머리와 밝은 옷이 박상현 참고의 갓과 회색 도포에 맞지 않는다. 뒤쪽 두 손은 성인 남성의 손처럼 보이나 앞쪽 남성에게 연결되는 팔이 없다. 왼쪽 위 표식은 흐려 판독되지 않는다.",
        "hard_violations": [
         "안규홍 앞쪽에 남성의 얼굴과 몸이 있는데 미는 두 팔은 카메라 쪽 아래에서 들어온다. 이를 박상현 한 사람의 몸으로 연결할 수 없으므로, 허용되지 않은 제삼자의 신체 일부가 추가되었거나 불가능한 신체 배치가 된 장면이다."
        ],
        "physics": "두 손바닥과 손가락은 등판 천에 접촉하며 주름도 보인다. 손에는 각각 손목과 전완이 이어져 있으므로 손 자체가 공중에 떠 있는 것은 아니다. 다만 그 팔을 앞쪽에 보이는 남성의 어깨와 연결할 수 없다. 안규홍의 벌어진 팔은 몸통에 연결되어 있고, 하체는 잘려 있어 발의 지지는 판단할 수 없다."
       },
       {
        "label": "B",
        "direction": "오른쪽 뒤의 박상현이 안규홍의 양쪽 등판에 손바닥을 대고 전방으로 밀고 있다. 그의 시선도 안규홍과 진행 방향을 향한다. 안규홍은 고개와 상체를 앞왼쪽으로 숙여 밀리는 반응을 보인다. 두 사람 너머로 골목과 이를 막은 검은 옷 인물들이 보여 전방 봉쇄 방향은 읽힌다.",
        "built_space": "흙길 양쪽에 초가지붕, 흙벽, 목재 기둥과 낮은 돌기단이 이어지고, 정면에는 골목 끝 건물 한 채가 보인다. 오른쪽에는 낮은 목재 마루가 있다. 참고 장소의 재료와 낮 조명은 잘 이어진다. 짚신 가판은 이 크롭에서 식별되지 않는다. 배경 골목은 인물 사이에 제한되지만 봉쇄 인물 무리가 눈에 띈다. 불가능한 반사나 명백한 시설 중복은 없다.",
        "entities": "안규홍은 묶은 검은 머리, 더러운 밝은 저고리, 드러난 목과 얼굴 옆면으로 제시된다. 눈의 부기는 각도 때문에 확인하기 어렵고 낫은 프레임 밖이다. 박상현은 성인 동아시아계 남성으로, 갓·검은 머리띠·수염·회색 도포가 참고와 대체로 맞는다. 두 사람 외에 배경에 검은 옷을 입은 여러 명의 인물이 명확히 추가되어 있다. 읽을 수 있는 글자는 없다.",
        "hard_violations": [
         "배경 골목에 검은 옷의 봉쇄 인물 여러 명을 추가했다. 이전 스틸의 사람을 이어 넣지 말고 안규홍과 박상현만 허용한다는 명시적 인물 제한을 위반한다."
        ],
        "physics": "두 손은 각각 회색 소매의 손목과 팔을 통해 박상현의 몸에 연결되고, 손바닥은 안규홍의 등판에 밀착되어 있다. 팔을 뻗는 방향과 안규홍의 앞으로 기운 상체가 실제 밀기 동작으로 연결된다. 두 주인공의 발은 클로즈업 밖이므로 지면 접촉을 직접 확인할 수 없지만, 떠 있는 몸으로 묘사되지는 않았다. 배경 인물들도 골목 지면에 서 있다."
       }
      ],
      "all_candidates_fail": true
     },
     "normalized": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 1,
        "verdict_ko": "등과 두 손의 클로즈업은 맞지만, 앞쪽 남성과 뒤쪽 미는 손의 위치가 연결되지 않아 제삼자 또는 불가능한 신체 배치가 발생한다."
       },
       {
        "label": "A",
        "score": 4,
        "verdict_ko": "박상현이 안규홍의 등을 두 손으로 미는 접촉과 의상은 충실하지만, 배경에 금지된 봉쇄 인물들을 추가해 최종 사용 조건을 위반한다."
       }
      ],
      "all_candidates_fail": true,
      "readings": [
       {
        "label": "B",
        "direction": "화면 아래에서 들어온 두 손바닥은 안규홍의 등 중앙에 닿아 화면 안쪽으로 민다. 안규홍은 카메라에 등을 돌리고 양팔을 옆으로 벌렸다. 그런데 다른 남성의 얼굴은 안규홍 앞쪽에서 카메라 쪽을 향하고 있어, 뒤에서 미는 박상현의 위치와 맞지 않는다. 골목의 깊이는 몸 사이보다 왼쪽 가판 뒤로 드러난다.",
        "built_space": "왼쪽에 짚신을 펼친 가판 한 구역, 매달린 짚신 여러 켤레와 바구니들이 있고, 오른쪽에는 목재 문과 기둥이 이어진다. 흙길과 낡은 목재는 시장 재질에 부합하지만, 참고의 초가와 돌기단 배치를 확인할 만큼 배경이 보이지는 않는다. 고정 시설의 명백한 중복이나 반사는 없다.",
        "entities": "등을 보인 인물은 검은 머리를 묶고 때 묻은 밝은 저고리를 입어 안규홍의 보이는 특징과 대체로 맞는다. 얼굴과 눈, 낫은 프레임 밖이므로 확인할 수 없다. 맞은편 남성은 성인 동아시아계 남성으로 보이지만, 흐트러진 머리와 밝은 옷이 박상현 참고의 갓과 회색 도포에 맞지 않는다. 뒤쪽 두 손은 성인 남성의 손처럼 보이나 앞쪽 남성에게 연결되는 팔이 없다. 왼쪽 위 표식은 흐려 판독되지 않는다.",
        "hard_violations": [
         "안규홍 앞쪽에 남성의 얼굴과 몸이 있는데 미는 두 팔은 카메라 쪽 아래에서 들어온다. 이를 박상현 한 사람의 몸으로 연결할 수 없으므로, 허용되지 않은 제삼자의 신체 일부가 추가되었거나 불가능한 신체 배치가 된 장면이다."
        ],
        "physics": "두 손바닥과 손가락은 등판 천에 접촉하며 주름도 보인다. 손에는 각각 손목과 전완이 이어져 있으므로 손 자체가 공중에 떠 있는 것은 아니다. 다만 그 팔을 앞쪽에 보이는 남성의 어깨와 연결할 수 없다. 안규홍의 벌어진 팔은 몸통에 연결되어 있고, 하체는 잘려 있어 발의 지지는 판단할 수 없다."
       },
       {
        "label": "A",
        "direction": "오른쪽 뒤의 박상현이 안규홍의 양쪽 등판에 손바닥을 대고 전방으로 밀고 있다. 그의 시선도 안규홍과 진행 방향을 향한다. 안규홍은 고개와 상체를 앞왼쪽으로 숙여 밀리는 반응을 보인다. 두 사람 너머로 골목과 이를 막은 검은 옷 인물들이 보여 전방 봉쇄 방향은 읽힌다.",
        "built_space": "흙길 양쪽에 초가지붕, 흙벽, 목재 기둥과 낮은 돌기단이 이어지고, 정면에는 골목 끝 건물 한 채가 보인다. 오른쪽에는 낮은 목재 마루가 있다. 참고 장소의 재료와 낮 조명은 잘 이어진다. 짚신 가판은 이 크롭에서 식별되지 않는다. 배경 골목은 인물 사이에 제한되지만 봉쇄 인물 무리가 눈에 띈다. 불가능한 반사나 명백한 시설 중복은 없다.",
        "entities": "안규홍은 묶은 검은 머리, 더러운 밝은 저고리, 드러난 목과 얼굴 옆면으로 제시된다. 눈의 부기는 각도 때문에 확인하기 어렵고 낫은 프레임 밖이다. 박상현은 성인 동아시아계 남성으로, 갓·검은 머리띠·수염·회색 도포가 참고와 대체로 맞는다. 두 사람 외에 배경에 검은 옷을 입은 여러 명의 인물이 명확히 추가되어 있다. 읽을 수 있는 글자는 없다.",
        "hard_violations": [
         "배경 골목에 검은 옷의 봉쇄 인물 여러 명을 추가했다. 이전 스틸의 사람을 이어 넣지 말고 안규홍과 박상현만 허용한다는 명시적 인물 제한을 위반한다."
        ],
        "physics": "두 손은 각각 회색 소매의 손목과 팔을 통해 박상현의 몸에 연결되고, 손바닥은 안규홍의 등판에 밀착되어 있다. 팔을 뻗는 방향과 안규홍의 앞으로 기운 상체가 실제 밀기 동작으로 연결된다. 두 주인공의 발은 클로즈업 밖이므로 지면 접촉을 직접 확인할 수 없지만, 떠 있는 몸으로 묘사되지는 않았다. 배경 인물들도 골목 지면에 서 있다."
       }
      ]
     },
     "ok": true
    }
   ],
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "slot_winner_match": true,
   "slot_winner": {
    "gemini-pro": "A",
    "gpt-high": "A"
   },
   "route": "cross_slot_agree"
  },
  "dual": {
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "normalized": {
    "A": 2.0,
    "B": 0.75
   },
   "adjusted": {
    "A": 1.75,
    "B": 0.5
   },
   "violations": {
    "A": [
     "[gemini-pro] 샷 텍스트에 없는 인물 추가 (배경의 검은 옷 무리)",
     "[gemini-pro] 물리적으로 불가능한 해부학 (허공에서 뻗어 나오는 박상현의 왼팔)",
     "[gpt-high] 배경 골목에 검은 옷의 봉쇄 인물 여러 명을 추가했다. 이전 스틸의 사람을 이어 넣지 말고 안규홍과 박상현만 허용한다는 명시적 인물 제한을 위반한다."
    ],
    "B": [
     "[gemini-pro] 샷 텍스트에 없는 인물 추가 (정체불명의 낯선 얼굴)",
     "[gemini-pro] 물리적으로 불가능한 해부학 (안규홍의 몸과 제3자의 얼굴이 기괴하게 융합됨)",
     "[gpt-high] 안규홍 앞쪽에 남성의 얼굴과 몸이 있는데 미는 두 팔은 카메라 쪽 아래에서 들어온다. 이를 박상현 한 사람의 몸으로 연결할 수 없으므로, 허용되지 않은 제삼자의 신체 일부가 추가되었거나 불가능한 신체 배치가 된 장면이다."
    ]
   },
   "per_model_winner": {
    "gemini-pro": "A",
    "gpt-high": "A"
   },
   "agreed": true
  },
  "totals": {
   "A": 1750,
   "B": 500
  },
  "selected": "A",
  "ranking": [
   "A",
   "B"
  ],
  "verdicts": [
   {
    "label": "A",
    "score": 1750,
    "verdict_ko": "프롬프트에서 엄격히 금지한 이전 샷의 인물들을 배경에 추가하였고, 박상현의 왼팔이 불가능한 위치에서 나오는 해부학적 오류가 발생하여 실격입니다.  ★위반: [gemini-pro] 샷 텍스트에 없는 인물 추가 (배경의 검은 옷 무리) / [gemini-pro] 물리적으로 불가능한 해부학 (허공에서 뻗어 나오는 박상현의 왼팔) / [gpt-high] 배경 골목에 검은 옷의 봉쇄 인물 여러 명을 추가했다. 이전 스틸의 사람을 이어 넣지 말고 안규홍과 박상현만 허용한다는 명시적 인물 제한을 위반한다."
   },
   {
    "label": "B",
    "score": 500,
    "verdict_ko": "박상현을 1인칭 시점으로 처리해 외형을 누락시켰으며, 안규홍의 몸 앞으로 명시되지 않은 제3자의 얼굴이 기괴하게 융합되는 치명적인 해부학적 오류가 있습니다.  ★위반: [gemini-pro] 샷 텍스트에 없는 인물 추가 (정체불명의 낯선 얼굴) / [gemini-pro] 물리적으로 불가능한 해부학 (안규홍의 몸과 제3자의 얼굴이 기괴하게 융합됨) / [gpt-high] 안규홍 앞쪽에 남성의 얼굴과 몸이 있는데 미는 두 팔은 카메라 쪽 아래에서 들어온다. 이를 박상현 한 사람의 몸으로 연결할 수 없으므로, 허용되지 않은 제삼자의 신체 일부가 추가되었거나 불가능한 신체 배치가 된 장면이다."
   }
  ],
  "refs": [
   {
    "label": "PREVIOUS SHOT STILL — a visually related earlier shot of this same place: the location's look, materials, fixed features and lighting mood are LOCKED to this photo; never copy its camera framing. The people visible in that photo are NOT in this shot: never carry their faces, bodies or clothing onto anyone here. Each person in THIS shot is defined solely by the PEOPLE list, their own CHARACTER REFERENCE images and their wardrobe notes. If this photo shows a character who cannot move (dead or unconscious), that character's exact pose, position and orientation are ALSO LOCKED — treat their body as a fixed prop of the set that only the camera moves around.",
    "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/db00517b-165b-4005-a287-f28936b59b95/images/272bd7af-9294-4490-98c6-88a40e9deacc/scene/recipe/S48sh10_sel.png",
    "asset_id": "6faff932-1bef-446b-a34a-f2071142ef9c",
    "role": "prev_still"
   },
   {
    "label": "CHARACTER REFERENCE — 안규홍: the exact person appearing in this shot; match face, hair and build exactly.",
    "path": "<bytes:1028384>",
    "asset_id": "3b5894db-ce55-4f34-b966-5e725c3aa952",
    "role": "character_ref"
   },
   {
    "label": "CHARACTER REFERENCE — 박상현: the exact person appearing in this shot; match face, hair and build exactly.",
    "path": "<bytes:840781>",
    "asset_id": "6187751d-5df3-4eb7-9d30-d6fc98a005e6",
    "role": "character_ref"
   }
  ],
  "critique_skipped": true,
  "needs_reshoot": true,
  "shot_run_produced": false,
  "shot_run_uid": "06aa1435-4ef8-7290-a094-14d7d4d0a878",
  "ref_mode": "prev+엔티티",
  "share_plan": {
   "ref_plan": "prev",
   "prev_anchor_tag": "S48sh10"
  }
 },
 "S48sh12::cine": {
  "applied": true,
  "attempted_at": "2026-09-09T06:29:21.360412+00:00",
  "fingerprint": "ed9e0c6a0a8b50287d935688c5a8eabf17208a7625037dce4846e7201a49b8d0",
  "fingerprint_version": 2,
  "provider": "mai",
  "endpoint": "openrouter/chat-completions",
  "model": "microsoft/mai-image-2.6",
  "pack": "24.202608252115",
  "source_file": "S48sh12_sel.png",
  "source_sha256": "a4d67fd33ba2f652959e643af0258b2f56b6c90826cce7cc9ed0f5c7ce78fa3b",
  "file": "S48sh12_cine.png",
  "staged_sha256": "24c38caab5e90ed63ae1ff491f32284528d35fa02e4ab99ee79583202edcfed7",
  "latency_ms": 26006
 },
 "S48sh23::signage": {
  "fp": "94a945f49c4ce283",
  "inscriptions": [],
  "cues": [],
  "dropped": []
 },
 "S48sh23": {
  "input_fingerprint": "7dc330a4671ec658",
  "prompt": "Create ONE FINAL photorealistic live-action film still of the moment below — early 20th-century Korea, Jeolla region; all people are Korean unless explicitly stated otherwise. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 서로의 등을 꽉 맞댄 채 방어 태세를 갖춘 안규홍과 이관회의 밀착된 상체.\n\nLOCATION (lock): The trapped outdoor standing space before the market's straw-sandal stall, enclosed by attackers rather than walls. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: medium shot\n- KEY BACKGROUND ELEMENTS: Market passage (Both approaches are blocked outside the tight framing) — The opposing approaches continue beyond the left and right edges; used as Small peripheral openings establish that the men are defending opposite directions within one space.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Maintain sober daylight and muted earth-toned grading, preserving facial detail without theatrical changes in contrast.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The clothing, hair and overall look of 안규홍 — who appear both in that photo and in this shot — are LOCKED to that photo. Anyone else visible in that photo is NOT in this shot: never carry their face, body or clothing onto anyone here. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The lanes remain blocked on both sides of the straw-shoe stall. 안규홍: He holds his dirt-gritted sickle handle in a defensive stance, with his back braced; his neck remains bare and his hair tied. 이관회: He stands in a rear-facing defensive stance with his back braced.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 안규홍 (한국인 남성, 청년 나이대, 검은 머리, 단단히 묶은 머리); 이관회 (한국인 남성, 청년 나이대, 자연스러운 검은 머리) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.",
  "roll_prompts": {
   "A": "Create ONE FINAL photorealistic live-action film still of the moment below — early 20th-century Korea, Jeolla region; all people are Korean unless explicitly stated otherwise. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 서로의 등을 꽉 맞댄 채 방어 태세를 갖춘 안규홍과 이관회의 밀착된 상체.\n\nLOCATION (lock): The trapped outdoor standing space before the market's straw-sandal stall, enclosed by attackers rather than walls. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: medium shot\n- KEY BACKGROUND ELEMENTS: Market passage (Both approaches are blocked outside the tight framing) — The opposing approaches continue beyond the left and right edges; used as Small peripheral openings establish that the men are defending opposite directions within one space.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Maintain sober daylight and muted earth-toned grading, preserving facial detail without theatrical changes in contrast.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The clothing, hair and overall look of 안규홍 — who appear both in that photo and in this shot — are LOCKED to that photo. Anyone else visible in that photo is NOT in this shot: never carry their face, body or clothing onto anyone here. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The lanes remain blocked on both sides of the straw-shoe stall. 안규홍: He holds his dirt-gritted sickle handle in a defensive stance, with his back braced; his neck remains bare and his hair tied. 이관회: He stands in a rear-facing defensive stance with his back braced.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 안규홍 (한국인 남성, 청년 나이대, 검은 머리, 단단히 묶은 머리); 이관회 (한국인 남성, 청년 나이대, 자연스러운 검은 머리) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.",
   "B": "Create ONE FINAL photorealistic live-action film still of the moment below — early 20th-century Korea, Jeolla region; all people are Korean unless explicitly stated otherwise. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 서로의 등을 꽉 맞댄 채 방어 태세를 갖춘 안규홍과 이관회의 밀착된 상체.\n\nLOCATION (lock): The trapped outdoor standing space before the market's straw-sandal stall, enclosed by attackers rather than walls. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: medium shot\n- KEY BACKGROUND ELEMENTS: Market passage (Both approaches are blocked outside the tight framing) — The opposing approaches continue beyond the left and right edges; used as Small peripheral openings establish that the men are defending opposite directions within one space.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Maintain sober daylight and muted earth-toned grading, preserving facial detail without theatrical changes in contrast.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The clothing, hair and overall look of 안규홍 — who appear both in that photo and in this shot — are LOCKED to that photo. Anyone else visible in that photo is NOT in this shot: never carry their face, body or clothing onto anyone here. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The lanes remain blocked on both sides of the straw-shoe stall. 안규홍: He holds his dirt-gritted sickle handle in a defensive stance, with his back braced; his neck remains bare and his hair tied. 이관회: He stands in a rear-facing defensive stance with his back braced.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 안규홍 (한국인 남성, 청년 나이대, 검은 머리, 단단히 묶은 머리); 이관회 (한국인 남성, 청년 나이대, 자연스러운 검은 머리) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.\n\nCOMPOSITION VARIATION (this candidate only): keep the place, the people and the moment exactly as specified, and keep from the CAMERA & FRAME contract its subject scale, screen placement and key background elements exactly. For this candidate only, its camera angle and height are deliberately left open: choose a camera position distinctly different from the obvious one, as a film director picking a second setup on the same blocking. Stay true to any explicit framing the SHOT TEXT itself dictates."
  },
  "shot_run_spend_attempt_count": 1,
  "readings": [
   {
    "label": "A",
    "direction": "안규홍은 낫을 쥐고 좌측 전방의 적을 주시하고 있으며, 이관회는 우측을 향해 무기를 겨누고 경계하고 있음.",
    "built_space": "짚신이 널린 시장 가판대 앞 공간으로, 배경 우측 가판대 기둥에 명확한 한글 글씨가 적힌 나무 팻말이 걸려 있음.",
    "entities": "안규홍(흰옷, 묶은 머리)과 이관회(갈색 옷, 짧은 머리)의 외모가 레퍼런스와 일치함. 샷 텍스트에 명시되지 않은 주변 인물(적대자들)이 전경과 배경에 포함됨.",
    "hard_violations": [
     "[gemini-pro] 배경 우측 나무 팻말에 읽을 수 있는 명확한 글씨가 포함되어 텍스트 금지 지시를 위반함",
     "[gpt-high] 지정된 두 사람 외에 좌우 전경과 배경에 다수의 추가 인물이 보인다.",
     "[gpt-high] 이관회에게 지시나 참조에 없는 별도 무기를 추가했다.",
     "[gpt-high] 오른쪽 배경 간판에 읽을 수 있는 글자가 있어 문자 금지 조건을 위반한다."
    ],
    "physics": "안규홍은 오른손으로 낫을 쥐고 있고 이관회는 칼자루 같은 무기를 쥐고 있으며, 두 사람 모두 지면에 서서 자세를 지탱하고 있음."
   },
   {
    "label": "B",
    "direction": "안규홍은 좌측 전방을 향해 낫을 겨누고, 이관회는 우측을 향해 긴 나무 막대를 쥔 채 시선을 두고 있음.",
    "built_space": "짚신이 진열된 시장 가판대 앞 공간으로, 지시된 위치와 환경을 잘 반영하고 있으며 읽을 수 있는 글씨가 없음.",
    "entities": "안규홍과 이관회 모두 지시된 복장과 인상착의를 정확히 묘사함. 전경에 어두운 옷을 입은 적대자들이 배치됨.",
    "hard_violations": [
     "[gpt-high] 안규홍과 이관회 외에는 사람을 보이지 말라는 조건과 달리, 양쪽 전경과 왼쪽 배경에 다수의 추가 인물이 보인다.",
     "[gpt-high] 이관회에게 지시나 참조에 없는 장대를 추가했다."
    ],
    "physics": "안규홍은 양손으로 낫의 자루를 쥐고, 이관회는 양손으로 막대를 쥐고 있으며 두 사람 모두 지면에 발을 딛고 안정적인 방어 자세를 유지함."
   }
  ],
  "cross_model_order": {
   "policy": "cross_model_order_v2_slot0_fwd_slot1_rev_parallel2",
   "slots": [
    {
     "model": "gemini-pro",
     "order": "forward",
     "display_to_canonical": {
      "A": "A",
      "B": "B"
     },
     "raw": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 9,
        "verdict_ko": "지시된 인물들의 외모와 밀착된 방어 태세를 잘 구현했으며, 금지된 텍스트나 물리적 오류 없이 촬영 묘사를 충실히 반영함."
       },
       {
        "label": "A",
        "score": 4,
        "verdict_ko": "배경 우측에 명확히 읽을 수 있는 글씨가 포함된 팻말이 등장하여 화면 내 텍스트 금지 지시를 심각하게 위반함."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "안규홍은 낫을 쥐고 좌측 전방의 적을 주시하고 있으며, 이관회는 우측을 향해 무기를 겨누고 경계하고 있음.",
        "built_space": "짚신이 널린 시장 가판대 앞 공간으로, 배경 우측 가판대 기둥에 명확한 한글 글씨가 적힌 나무 팻말이 걸려 있음.",
        "entities": "안규홍(흰옷, 묶은 머리)과 이관회(갈색 옷, 짧은 머리)의 외모가 레퍼런스와 일치함. 샷 텍스트에 명시되지 않은 주변 인물(적대자들)이 전경과 배경에 포함됨.",
        "hard_violations": [
         "배경 우측 나무 팻말에 읽을 수 있는 명확한 글씨가 포함되어 텍스트 금지 지시를 위반함"
        ],
        "physics": "안규홍은 오른손으로 낫을 쥐고 있고 이관회는 칼자루 같은 무기를 쥐고 있으며, 두 사람 모두 지면에 서서 자세를 지탱하고 있음."
       },
       {
        "label": "B",
        "direction": "안규홍은 좌측 전방을 향해 낫을 겨누고, 이관회는 우측을 향해 긴 나무 막대를 쥔 채 시선을 두고 있음.",
        "built_space": "짚신이 진열된 시장 가판대 앞 공간으로, 지시된 위치와 환경을 잘 반영하고 있으며 읽을 수 있는 글씨가 없음.",
        "entities": "안규홍과 이관회 모두 지시된 복장과 인상착의를 정확히 묘사함. 전경에 어두운 옷을 입은 적대자들이 배치됨.",
        "hard_violations": [],
        "physics": "안규홍은 양손으로 낫의 자루를 쥐고, 이관회는 양손으로 막대를 쥐고 있으며 두 사람 모두 지면에 발을 딛고 안정적인 방어 자세를 유지함."
       }
      ],
      "all_candidates_fail": false
     },
     "normalized": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 9,
        "verdict_ko": "지시된 인물들의 외모와 밀착된 방어 태세를 잘 구현했으며, 금지된 텍스트나 물리적 오류 없이 촬영 묘사를 충실히 반영함."
       },
       {
        "label": "A",
        "score": 4,
        "verdict_ko": "배경 우측에 명확히 읽을 수 있는 글씨가 포함된 팻말이 등장하여 화면 내 텍스트 금지 지시를 심각하게 위반함."
       }
      ],
      "all_candidates_fail": false,
      "readings": [
       {
        "label": "A",
        "direction": "안규홍은 낫을 쥐고 좌측 전방의 적을 주시하고 있으며, 이관회는 우측을 향해 무기를 겨누고 경계하고 있음.",
        "built_space": "짚신이 널린 시장 가판대 앞 공간으로, 배경 우측 가판대 기둥에 명확한 한글 글씨가 적힌 나무 팻말이 걸려 있음.",
        "entities": "안규홍(흰옷, 묶은 머리)과 이관회(갈색 옷, 짧은 머리)의 외모가 레퍼런스와 일치함. 샷 텍스트에 명시되지 않은 주변 인물(적대자들)이 전경과 배경에 포함됨.",
        "hard_violations": [
         "배경 우측 나무 팻말에 읽을 수 있는 명확한 글씨가 포함되어 텍스트 금지 지시를 위반함"
        ],
        "physics": "안규홍은 오른손으로 낫을 쥐고 있고 이관회는 칼자루 같은 무기를 쥐고 있으며, 두 사람 모두 지면에 서서 자세를 지탱하고 있음."
       },
       {
        "label": "B",
        "direction": "안규홍은 좌측 전방을 향해 낫을 겨누고, 이관회는 우측을 향해 긴 나무 막대를 쥔 채 시선을 두고 있음.",
        "built_space": "짚신이 진열된 시장 가판대 앞 공간으로, 지시된 위치와 환경을 잘 반영하고 있으며 읽을 수 있는 글씨가 없음.",
        "entities": "안규홍과 이관회 모두 지시된 복장과 인상착의를 정확히 묘사함. 전경에 어두운 옷을 입은 적대자들이 배치됨.",
        "hard_violations": [],
        "physics": "안규홍은 양손으로 낫의 자루를 쥐고, 이관회는 양손으로 막대를 쥐고 있으며 두 사람 모두 지면에 발을 딛고 안정적인 방어 자세를 유지함."
       }
      ]
     },
     "ok": true
    },
    {
     "model": "gpt-high",
     "order": "reverse",
     "display_to_canonical": {
      "A": "B",
      "B": "A"
     },
     "raw": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 2,
        "verdict_ko": "등을 맞댄 방어 동작은 구현했으나, 금지된 주변 인물들과 이관회의 장대를 추가하고 무릎 부근까지 넓혔으며, 판독 가능한 글자까지 있는 B보다만 낫다."
       },
       {
        "label": "B",
        "score": 1,
        "verdict_ko": "밀착된 상체 중심의 미디엄 숏은 더 가깝지만, 다수의 추가 인물과 이관회의 무기, 읽을 수 있는 간판 글자로 명시적 금지 조건을 위반한다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "안규홍은 화면 왼쪽 전경의 검은 옷 남자를 향해 시선을 낮추고 낫을 왼쪽으로 세워 방어한다. 이관회는 오른쪽 전경 인물을 바라보며 장대를 오른쪽 위로 기울인다. 서로 반대 방향을 경계하지만, 위협 인물은 지시와 달리 화면 안에 들어와 있다.",
        "built_space": "두 사람은 짚신 좌판 앞 흙바닥 공간에서 등을 맞대고 있다. 뒤에는 목재 진열대 한 조와 수평 선반들, 상부에 걸린 짚신 약 세 쌍, 하부에 기대 놓은 한 쌍이 보인다. 왼쪽 가장자리에 좁은 통로가 보이지만 오른쪽은 추가 인물들이 가린다. 목재와 흙, 초가의 재질은 이전 장면과 어울리나, 중앙 인물들을 무릎 부근까지 보여 주어 요구한 밀착 상체 구도보다 넓다.",
        "entities": "중앙에는 청년 한국인 남성으로 보이는 두 사람이 있다. 안규홍의 묶은 검은 머리, 맨목, 때 묻은 미색 옷과 낫은 지시에 부합한다. 이관회의 가르마 탄 검은 머리와 해진 갈색 저고리도 참조와 대체로 맞는다. 그러나 이관회에게 참조와 지시에 없는 장대를 들렸고, 양쪽 전경 및 왼쪽 뒤편에 허용되지 않은 여러 사람을 추가했다. 읽을 수 있는 글자는 보이지 않는다.",
        "hard_violations": [
         "안규홍과 이관회 외에는 사람을 보이지 말라는 조건과 달리, 양쪽 전경과 왼쪽 배경에 다수의 추가 인물이 보인다.",
         "이관회에게 지시나 참조에 없는 장대를 추가했다."
        ],
        "physics": "두 사람은 등을 실제로 접촉시키고 무릎을 굽힌 방어 자세를 취한다. 발은 화면 밖이지만 몸이 떠 있다는 징후는 없다. 안규홍은 두 손으로 낫 손잡이를, 이관회는 두 손으로 장대를 잡아 지지한다. 짚신은 끈에 걸리거나 선반과 받침판 위에 놓여 있어 지지가 설명된다."
       },
       {
        "label": "B",
        "direction": "안규홍은 왼쪽 전경의 머리띠 남자를 바라보고 낫날을 그 방향으로 세운다. 이관회는 오른쪽 전경의 흰 머리띠 남자를 향해 시선을 보내며 손에 든 무기를 오른쪽으로 향한다. 두 사람의 반대 방향 경계는 분명하지만, 원래 화면 밖에 있어야 할 상대들이 화면을 크게 차지한다.",
        "built_space": "중앙 두 사람 뒤에 초가지붕 한 면, 이를 받치는 좌우 목재 기둥, 낮은 목재 좌판 한 조와 짚신 진열 공간이 보인다. 왼쪽 위에는 둥근 짚 물건 하나, 오른쪽 기둥에는 세로 간판 하나가 있다. 상체와 허리 중심의 미디엄 숏은 A보다 요구에 가깝지만, 양쪽 통로의 작은 개방부 대신 추가 인물들이 가장자리를 막는다. 초가와 낡은 목재는 참조의 재질과 대체로 맞는다.",
        "entities": "안규홍과 이관회는 청년 한국인 남성으로 보이며, 각각 묶은 검은 머리와 미색 옷, 자연스러운 가르마 머리와 갈색 저고리를 갖췄다. 안규홍의 맨목과 흙 묻은 손, 손에 든 낫도 보인다. 이관회에게는 목제 손잡이와 검은 금속 부분이 있는 별도 무기를 추가했다. 두 사람 외에 좌우 전경과 배경의 남성들이 여러 명 보이고, 오른쪽 세로 간판에는 판독 가능한 한글이 있다.",
        "hard_violations": [
         "지정된 두 사람 외에 좌우 전경과 배경에 다수의 추가 인물이 보인다.",
         "이관회에게 지시나 참조에 없는 별도 무기를 추가했다.",
         "오른쪽 배경 간판에 읽을 수 있는 글자가 있어 문자 금지 조건을 위반한다."
        ],
        "physics": "두 사람의 등과 어깨 뒤쪽이 맞닿고 상체가 서로 반대쪽을 향한다. 하체와 발의 접지는 화면 밖이므로 확인할 수 없지만 부유나 불가능한 자세는 보이지 않는다. 안규홍의 낫은 오른손이 손잡이를 잡고 있으며, 이관회의 무기도 보이는 손이 목제 자루를 잡아 지지한다. 좌판은 목재 다리와 버팀대로 받쳐지고 진열품은 그 위에 놓여 있다."
       }
      ],
      "all_candidates_fail": true
     },
     "normalized": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 2,
        "verdict_ko": "등을 맞댄 방어 동작은 구현했으나, 금지된 주변 인물들과 이관회의 장대를 추가하고 무릎 부근까지 넓혔으며, 판독 가능한 글자까지 있는 B보다만 낫다."
       },
       {
        "label": "A",
        "score": 1,
        "verdict_ko": "밀착된 상체 중심의 미디엄 숏은 더 가깝지만, 다수의 추가 인물과 이관회의 무기, 읽을 수 있는 간판 글자로 명시적 금지 조건을 위반한다."
       }
      ],
      "all_candidates_fail": true,
      "readings": [
       {
        "label": "B",
        "direction": "안규홍은 화면 왼쪽 전경의 검은 옷 남자를 향해 시선을 낮추고 낫을 왼쪽으로 세워 방어한다. 이관회는 오른쪽 전경 인물을 바라보며 장대를 오른쪽 위로 기울인다. 서로 반대 방향을 경계하지만, 위협 인물은 지시와 달리 화면 안에 들어와 있다.",
        "built_space": "두 사람은 짚신 좌판 앞 흙바닥 공간에서 등을 맞대고 있다. 뒤에는 목재 진열대 한 조와 수평 선반들, 상부에 걸린 짚신 약 세 쌍, 하부에 기대 놓은 한 쌍이 보인다. 왼쪽 가장자리에 좁은 통로가 보이지만 오른쪽은 추가 인물들이 가린다. 목재와 흙, 초가의 재질은 이전 장면과 어울리나, 중앙 인물들을 무릎 부근까지 보여 주어 요구한 밀착 상체 구도보다 넓다.",
        "entities": "중앙에는 청년 한국인 남성으로 보이는 두 사람이 있다. 안규홍의 묶은 검은 머리, 맨목, 때 묻은 미색 옷과 낫은 지시에 부합한다. 이관회의 가르마 탄 검은 머리와 해진 갈색 저고리도 참조와 대체로 맞는다. 그러나 이관회에게 참조와 지시에 없는 장대를 들렸고, 양쪽 전경 및 왼쪽 뒤편에 허용되지 않은 여러 사람을 추가했다. 읽을 수 있는 글자는 보이지 않는다.",
        "hard_violations": [
         "안규홍과 이관회 외에는 사람을 보이지 말라는 조건과 달리, 양쪽 전경과 왼쪽 배경에 다수의 추가 인물이 보인다.",
         "이관회에게 지시나 참조에 없는 장대를 추가했다."
        ],
        "physics": "두 사람은 등을 실제로 접촉시키고 무릎을 굽힌 방어 자세를 취한다. 발은 화면 밖이지만 몸이 떠 있다는 징후는 없다. 안규홍은 두 손으로 낫 손잡이를, 이관회는 두 손으로 장대를 잡아 지지한다. 짚신은 끈에 걸리거나 선반과 받침판 위에 놓여 있어 지지가 설명된다."
       },
       {
        "label": "A",
        "direction": "안규홍은 왼쪽 전경의 머리띠 남자를 바라보고 낫날을 그 방향으로 세운다. 이관회는 오른쪽 전경의 흰 머리띠 남자를 향해 시선을 보내며 손에 든 무기를 오른쪽으로 향한다. 두 사람의 반대 방향 경계는 분명하지만, 원래 화면 밖에 있어야 할 상대들이 화면을 크게 차지한다.",
        "built_space": "중앙 두 사람 뒤에 초가지붕 한 면, 이를 받치는 좌우 목재 기둥, 낮은 목재 좌판 한 조와 짚신 진열 공간이 보인다. 왼쪽 위에는 둥근 짚 물건 하나, 오른쪽 기둥에는 세로 간판 하나가 있다. 상체와 허리 중심의 미디엄 숏은 A보다 요구에 가깝지만, 양쪽 통로의 작은 개방부 대신 추가 인물들이 가장자리를 막는다. 초가와 낡은 목재는 참조의 재질과 대체로 맞는다.",
        "entities": "안규홍과 이관회는 청년 한국인 남성으로 보이며, 각각 묶은 검은 머리와 미색 옷, 자연스러운 가르마 머리와 갈색 저고리를 갖췄다. 안규홍의 맨목과 흙 묻은 손, 손에 든 낫도 보인다. 이관회에게는 목제 손잡이와 검은 금속 부분이 있는 별도 무기를 추가했다. 두 사람 외에 좌우 전경과 배경의 남성들이 여러 명 보이고, 오른쪽 세로 간판에는 판독 가능한 한글이 있다.",
        "hard_violations": [
         "지정된 두 사람 외에 좌우 전경과 배경에 다수의 추가 인물이 보인다.",
         "이관회에게 지시나 참조에 없는 별도 무기를 추가했다.",
         "오른쪽 배경 간판에 읽을 수 있는 글자가 있어 문자 금지 조건을 위반한다."
        ],
        "physics": "두 사람의 등과 어깨 뒤쪽이 맞닿고 상체가 서로 반대쪽을 향한다. 하체와 발의 접지는 화면 밖이므로 확인할 수 없지만 부유나 불가능한 자세는 보이지 않는다. 안규홍의 낫은 오른손이 손잡이를 잡고 있으며, 이관회의 무기도 보이는 손이 목제 자루를 잡아 지지한다. 좌판은 목재 다리와 버팀대로 받쳐지고 진열품은 그 위에 놓여 있다."
       }
      ]
     },
     "ok": true
    }
   ],
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "slot_winner_match": true,
   "slot_winner": {
    "gemini-pro": "B",
    "gpt-high": "B"
   },
   "route": "cross_slot_agree"
  },
  "dual": {
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "normalized": {
    "A": 0.944,
    "B": 2.0
   },
   "adjusted": {
    "A": 0.694,
    "B": 1.75
   },
   "violations": {
    "A": [
     "[gemini-pro] 배경 우측 나무 팻말에 읽을 수 있는 명확한 글씨가 포함되어 텍스트 금지 지시를 위반함",
     "[gpt-high] 지정된 두 사람 외에 좌우 전경과 배경에 다수의 추가 인물이 보인다.",
     "[gpt-high] 이관회에게 지시나 참조에 없는 별도 무기를 추가했다.",
     "[gpt-high] 오른쪽 배경 간판에 읽을 수 있는 글자가 있어 문자 금지 조건을 위반한다."
    ],
    "B": [
     "[gpt-high] 안규홍과 이관회 외에는 사람을 보이지 말라는 조건과 달리, 양쪽 전경과 왼쪽 배경에 다수의 추가 인물이 보인다.",
     "[gpt-high] 이관회에게 지시나 참조에 없는 장대를 추가했다."
    ]
   },
   "per_model_winner": {
    "gemini-pro": "B",
    "gpt-high": "B"
   },
   "agreed": true
  },
  "totals": {
   "B": 1750,
   "A": 694
  },
  "selected": "B",
  "ranking": [
   "B",
   "A"
  ],
  "verdicts": [
   {
    "label": "B",
    "score": 1750,
    "verdict_ko": "지시된 인물들의 외모와 밀착된 방어 태세를 잘 구현했으며, 금지된 텍스트나 물리적 오류 없이 촬영 묘사를 충실히 반영함.  ★위반: [gpt-high] 안규홍과 이관회 외에는 사람을 보이지 말라는 조건과 달리, 양쪽 전경과 왼쪽 배경에 다수의 추가 인물이 보인다. / [gpt-high] 이관회에게 지시나 참조에 없는 장대를 추가했다."
   },
   {
    "label": "A",
    "score": 694,
    "verdict_ko": "배경 우측에 명확히 읽을 수 있는 글씨가 포함된 팻말이 등장하여 화면 내 텍스트 금지 지시를 심각하게 위반함.  ★위반: [gemini-pro] 배경 우측 나무 팻말에 읽을 수 있는 명확한 글씨가 포함되어 텍스트 금지 지시를 위반함 / [gpt-high] 지정된 두 사람 외에 좌우 전경과 배경에 다수의 추가 인물이 보인다. / [gpt-high] 이관회에게 지시나 참조에 없는 별도 무기를 추가했다. / [gpt-high] 오른쪽 배경 간판에 읽을 수 있는 글자가 있어 문자 금지 조건을 위반한다."
   }
  ],
  "refs": [
   {
    "label": "PREVIOUS SHOT STILL — a visually related earlier shot of this same place: the location's look, materials, fixed features and lighting mood are LOCKED to this photo; never copy its camera framing. The clothing, hair and overall look of 안규홍 — who appear both in that photo and in this shot — are LOCKED to that photo. Anyone else visible in that photo is NOT in this shot: never carry their face, body or clothing onto anyone here. If this photo shows a character who cannot move (dead or unconscious), that character's exact pose, position and orientation are ALSO LOCKED — treat their body as a fixed prop of the set that only the camera moves around.",
    "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/db00517b-165b-4005-a287-f28936b59b95/images/272bd7af-9294-4490-98c6-88a40e9deacc/scene/recipe/S48sh12_sel.png",
    "asset_id": "03c5c947-4377-4427-82a0-d2ec9cd843f4",
    "role": "prev_still"
   },
   {
    "label": "CHARACTER REFERENCE — 안규홍: the exact person appearing in this shot; match face, hair and build exactly.",
    "path": "<bytes:1028384>",
    "asset_id": "3b5894db-ce55-4f34-b966-5e725c3aa952",
    "role": "character_ref"
   },
   {
    "label": "CHARACTER REFERENCE — 이관회: the exact person appearing in this shot; match face, hair and build exactly.",
    "path": "<bytes:933540>",
    "asset_id": "851b450f-ac84-47fe-ac5a-03a56f63fcae",
    "role": "character_ref"
   }
  ],
  "critique_skipped": true,
  "shot_run_produced": false,
  "shot_run_uid": "06aa1435-50a5-7918-8a7a-f37ac3d26780",
  "ref_mode": "prev+엔티티",
  "share_plan": {
   "ref_plan": "prev",
   "prev_anchor_tag": "S48sh12"
  }
 },
 "S48sh23::cine": {
  "applied": true,
  "attempted_at": "2026-09-09T06:30:58.514730+00:00",
  "fingerprint": "11bcd8a30e54c9465d7f3d957a37e6a4bbe570c0bed6d35a418f0ee8ade295d8",
  "fingerprint_version": 2,
  "provider": "mai",
  "endpoint": "openrouter/chat-completions",
  "model": "microsoft/mai-image-2.6",
  "pack": "24.202608252115",
  "source_file": "S48sh23_sel.png",
  "source_sha256": "4118f8533a0bc53e1ebb73303fb20a055e6badcc33c8b619aa967d3518a70f37",
  "file": "S48sh23_cine.png",
  "staged_sha256": "e8a5e7c7215cd943527ca3f5dca0ce4b18755bd7f68336beba2e8d025b0baf28",
  "latency_ms": 28634
 },
 "S49sh2::signage": {
  "fp": "2d7fd60e8f647ce0",
  "inscriptions": [],
  "cues": [],
  "dropped": []
 },
 "era_assess::bd158203e521ff11": {
  "subjects": [],
  "subject_text": "파청 산채 인근 편백숲·숲길 훈련 구역·보초 경계 지점\n곧은 편백 줄기가 빽빽한 숲과 좁은 흙길. 낙엽 깔린 바닥에 바위가 드러나고, 나무 사이로 산채 가장자리가 보인다.",
  "identity": "canonical",
  "scope_id": "L24",
  "scope_role": "location_exterior",
  "scope_sha": "1470afee7ce9f0d3"
 },
 "groupbg::outer_cypress_forest": {
  "input_fingerprint": "9fcf039096716f25",
  "meta": {
   "model": "gpt-image-2",
   "size": "1536x864",
   "pack": "11.202607220237",
   "contract": "bgfirst_full_v3",
   "group_sig": {
    "key": "outer_cypress_forest",
    "tags": [
     "S49sh2",
     "S49sh5",
     "S49sh7"
    ]
   },
   "context_sig": "f588cbd59e255e62"
  },
  "prompt": "Create ONE empty live-action location background photograph — NO PEOPLE, no figures, no body parts, no silhouettes, no shadows or reflections of people anywhere.\n\"Empty\" means no people only: KEEP the location's inherent occupants and stock that define the place — animals in an animal shelter, pen or farm, goods and produce in a market, moored boats in a harbour — unless the shot text explicitly removes them.\nThe FIRST attached image is a thin-line storyboard sketch of a shot that happens at this location: use it ONLY as spatial evidence — what this place contains, how its ground, structures and landmarks are arranged and proportioned. Ignore the sketched people and arrows entirely, and do NOT copy its line style: render a fully photographic, physically plausible real place that fits THE LOCATION text below.\nHUMAN-SCALE CALIBRATION: derive every structure's true size from human-scale elements — a door ≈ 2m, a window ≈ 1–1.5m wide, one storey ≈ 2.5–3m; never inflate a small structure or shrink a large one.\n\nTHE LOCATION — early 20th-century Korea, Jeolla region; all people are Korean unless explicitly stated otherwise: The damp, leaf-covered floor of the cypress forest below the mountain camp, between closely spaced trunks.\n\nLOCATION DETAIL (descriptive evidence about this place from the production's location records — evidence, not a staging order): use it to understand what kind of place this is and what it permanently contains. Carry over ONLY its enduring physical features — layout, structures, ground and materials, fixed equipment and installations. Do NOT treat any lighting, weather or time-of-day wording, momentary object states, depicted events or subjective impressions in it as fixed facts of the place; the TIME OF DAY lock below is the sole authority for time and lighting.\n파청 산채 인근 편백숲·숲길 훈련 구역·보초 경계 지점: 곧고 빽빽하게 뻗은 편백나무들이 들어선 울창한 숲 지대. 바닥에는 두꺼운 낙엽이 깔려 있고, 몸을 숨길 수 있는 굵은 나무 밑동, 크고 작은 바위, 나무더미들이 산재해 있다. (특징: 빽빽하고 곧은 편백나무 군락; 이슬에 젖은 두꺼운 낙엽 바닥; 총격과 매복에 쓰이는 큰 바위와 나무더미; 나무 밑 돌로 가려진 탄약 숨기는 장소)\n\nSCENE EVIDENCE (verbatim quotes from the screenplay about this place — treat them as evidence of what the location physically contains and looks like; stage the PLACE those moments happen in, but do NOT depict the momentary actions, people or staged props themselves):\n- S#49. 파청 산채 인근 편백숲 / 낮\n- 나무 사이 저 멀리— 연기 한 줄기가 올라온다. 파청 산채.\n\nTIME OF DAY (lock): day.\n\nRender ONE photorealistic empty location photograph, 16:9, neutral enough that every shot of this place can be staged from it later. No readable writing anywhere: surfaces that would carry writing may be present, but stage any wording out of legibility — an oblique angle, distance, shallow focus. No captions, watermarks or overlay text.",
  "effective_prompt": "Create ONE empty live-action location background photograph — NO PEOPLE, no figures, no body parts, no silhouettes, no shadows or reflections of people anywhere.\n\"Empty\" means no people only: KEEP the location's inherent occupants and stock that define the place — animals in an animal shelter, pen or farm, goods and produce in a market, moored boats in a harbour — unless the shot text explicitly removes them.\nThe FIRST attached image is a thin-line storyboard sketch of a shot that happens at this location: use it ONLY as spatial evidence — what this place contains, how its ground, structures and landmarks are arranged and proportioned. Ignore the sketched people and arrows entirely, and do NOT copy its line style: render a fully photographic, physically plausible real place that fits THE LOCATION text below.\nHUMAN-SCALE CALIBRATION: derive every structure's true size from human-scale elements — a door ≈ 2m, a window ≈ 1–1.5m wide, one storey ≈ 2.5–3m; never inflate a small structure or shrink a large one.\n\nTHE LOCATION — early 20th-century Korea, Jeolla region; all people are Korean unless explicitly stated otherwise: The damp, leaf-covered floor of the cypress forest below the mountain camp, between closely spaced trunks.\n\nLOCATION DETAIL (descriptive evidence about this place from the production's location records — evidence, not a staging order): use it to understand what kind of place this is and what it permanently contains. Carry over ONLY its enduring physical features — layout, structures, ground and materials, fixed equipment and installations. Do NOT treat any lighting, weather or time-of-day wording, momentary object states, depicted events or subjective impressions in it as fixed facts of the place; the TIME OF DAY lock below is the sole authority for time and lighting.\n파청 산채 인근 편백숲·숲길 훈련 구역·보초 경계 지점: 곧고 빽빽하게 뻗은 편백나무들이 들어선 울창한 숲 지대. 바닥에는 두꺼운 낙엽이 깔려 있고, 몸을 숨길 수 있는 굵은 나무 밑동, 크고 작은 바위, 나무더미들이 산재해 있다. (특징: 빽빽하고 곧은 편백나무 군락; 이슬에 젖은 두꺼운 낙엽 바닥; 총격과 매복에 쓰이는 큰 바위와 나무더미; 나무 밑 돌로 가려진 탄약 숨기는 장소)\n\nSCENE EVIDENCE (verbatim quotes from the screenplay about this place — treat them as evidence of what the location physically contains and looks like; stage the PLACE those moments happen in, but do NOT depict the momentary actions, people or staged props themselves):\n- S#49. 파청 산채 인근 편백숲 / 낮\n- 나무 사이 저 멀리— 연기 한 줄기가 올라온다. 파청 산채.\n\nTIME OF DAY (lock): day.\n\nRender ONE photorealistic empty location photograph, 16:9, neutral enough that every shot of this place can be staged from it later. No readable writing anywhere: surfaces that would carry writing may be present, but stage any wording out of legibility — an oblique angle, distance, shallow focus. No captions, watermarks or overlay text.",
  "bg_path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/db00517b-165b-4005-a287-f28936b59b95/images/272bd7af-9294-4490-98c6-88a40e9deacc/scene/recipe/groupbg_outer_cypress_forest_e3184b.png",
  "asset_id": "e9399fca-b21e-44d7-a40a-5ebc8782d296",
  "input_asset_ids": [
   "c27b30b3-1418-418d-b990-faa3668679a4"
  ],
  "origin_tag": "S49sh2",
  "place_text": "The damp, leaf-covered floor of the cypress forest below the mountain camp, between closely spaced trunks.",
  "origin_inputs": {
   "place_text": "The damp, leaf-covered floor of the cypress forest below the mountain camp, between closely spaced trunks.",
   "time_of_day_en": "day",
   "conti_asset_id": "c27b30b3-1418-418d-b990-faa3668679a4"
  }
 },
 "S49sh2::bgfirst_bg": {
  "input_fingerprint": "280fa1357c599a22",
  "prompt": "Create the EMPTY BACKGROUND PLATE for one film shot — NO PEOPLE, no figures, no body parts, no sketch lines, no arrows anywhere.\n\"Empty\" means no people only: KEEP the location's inherent occupants and stock that define the place — animals in an animal shelter, pen or farm, goods in a market, moored boats in a harbour — unless the shot text explicitly removes them.\nThe FIRST attached image is a thin-line storyboard sketch: use ONLY its camera angle, horizon, perspective and the placement/size of buildings and set masses — ignore the sketched people and arrows entirely. The SECOND attached image (LOCATION PHOTOGRAPH) is the real place: take its architecture, materials, signage and fixed features, and RE-PROJECT them into the sketch's camera. If the photograph's camera differs from the sketch's, the sketch's camera wins.\nHUMAN-SCALE CALIBRATION: derive every structure's true size from human-scale elements — a door ≈ 2m, a window ≈ 1–1.5m wide, one storey ≈ 2.5–3m; never inflate a small structure or shrink a large one.\n\nSHOT TEXT this background must serve (Korean): 소총을 가슴 높이로 든 채 빽빽한 편백나무 틈새로 한 발 내딛은 자세의 일본인 군인들의 전신.\n\nLOCATION (lock): The damp, leaf-covered floor of the cypress forest below the mountain camp, between closely spaced trunks.\n\nTIME OF DAY (lock): day.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: wide shot\n- KEY BACKGROUND ELEMENTS: Cypress trees (Dense trees separate the emerging soldiers into two visible branches) — Successive trunks overlap at the frame edges and between figures; used as Create layered partial occlusions while leaving enough gaps to read complete bodies; Fallen leaves (Wet with dew beneath the soldiers' boots); used as Keep the lower frame grounded in the careful placement of each step; Soldiers' rifles (Held at chest height during the advance) — Seen obliquely across individual torsos, with slight differences in carrying angle; used as Repeat the armed threat across the formation without enlarging any single weapon.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Natural daylight, muted saturation, and restrained contrast support the quiet threat without adding atmospheric effects.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nRender ONE photorealistic empty location photograph, 16:9, that this shot can be staged inside later. No readable writing anywhere: surfaces that would carry writing may be present, but stage any wording out of legibility — an oblique angle, distance, shallow focus. No captions, watermarks or overlay text.",
  "effective_prompt": "Create the EMPTY BACKGROUND PLATE for one film shot — NO PEOPLE, no figures, no body parts, no sketch lines, no arrows anywhere.\n\"Empty\" means no people only: KEEP the location's inherent occupants and stock that define the place — animals in an animal shelter, pen or farm, goods in a market, moored boats in a harbour — unless the shot text explicitly removes them.\nThe FIRST attached image is a thin-line storyboard sketch: use ONLY its camera angle, horizon, perspective and the placement/size of buildings and set masses — ignore the sketched people and arrows entirely. The SECOND attached image (LOCATION PHOTOGRAPH) is the real place: take its architecture, materials, signage and fixed features, and RE-PROJECT them into the sketch's camera. If the photograph's camera differs from the sketch's, the sketch's camera wins.\nHUMAN-SCALE CALIBRATION: derive every structure's true size from human-scale elements — a door ≈ 2m, a window ≈ 1–1.5m wide, one storey ≈ 2.5–3m; never inflate a small structure or shrink a large one.\n\nSHOT TEXT this background must serve (Korean): 소총을 가슴 높이로 든 채 빽빽한 편백나무 틈새로 한 발 내딛은 자세의 일본인 군인들의 전신.\n\nLOCATION (lock): The damp, leaf-covered floor of the cypress forest below the mountain camp, between closely spaced trunks.\n\nTIME OF DAY (lock): day.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: wide shot\n- KEY BACKGROUND ELEMENTS: Cypress trees (Dense trees separate the emerging soldiers into two visible branches) — Successive trunks overlap at the frame edges and between figures; used as Create layered partial occlusions while leaving enough gaps to read complete bodies; Fallen leaves (Wet with dew beneath the soldiers' boots); used as Keep the lower frame grounded in the careful placement of each step; Soldiers' rifles (Held at chest height during the advance) — Seen obliquely across individual torsos, with slight differences in carrying angle; used as Repeat the armed threat across the formation without enlarging any single weapon.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Natural daylight, muted saturation, and restrained contrast support the quiet threat without adding atmospheric effects.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nRender ONE photorealistic empty location photograph, 16:9, that this shot can be staged inside later. No readable writing anywhere: surfaces that would carry writing may be present, but stage any wording out of legibility — an oblique angle, distance, shallow focus. No captions, watermarks or overlay text.",
  "bg_path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/db00517b-165b-4005-a287-f28936b59b95/images/272bd7af-9294-4490-98c6-88a40e9deacc/scene/recipe/S49sh2__bgfirst_bg.png",
  "asset_id": "d1dc39c5-ff26-4654-a17c-8e173fecc54c",
  "input_asset_ids": [
   "c27b30b3-1418-418d-b990-faa3668679a4",
   "e9399fca-b21e-44d7-a40a-5ebc8782d296"
  ]
 },
 "S49sh2": {
  "input_fingerprint": "4cc3a20cd209b8b4",
  "prompt": "Create ONE FINAL photorealistic live-action film still of the moment below — early 20th-century Korea, Jeolla region; all people are Korean unless explicitly stated otherwise. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 소총을 가슴 높이로 든 채 빽빽한 편백나무 틈새로 한 발 내딛은 자세의 일본인 군인들의 전신.\n\nLOCATION (lock): The damp, leaf-covered floor of the cypress forest below the mountain camp, between closely spaced trunks. The shot takes place here — the attached LOCATION PHOTOGRAPH shows the exact spot.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: wide shot\n- KEY BACKGROUND ELEMENTS: Cypress trees (Dense trees separate the emerging soldiers into two visible branches) — Successive trunks overlap at the frame edges and between figures; used as Create layered partial occlusions while leaving enough gaps to read complete bodies; Fallen leaves (Wet with dew beneath the soldiers' boots); used as Keep the lower frame grounded in the careful placement of each step; Soldiers' rifles (Held at chest height during the advance) — Seen obliquely across individual torsos, with slight differences in carrying angle; used as Repeat the armed threat across the formation without enlarging any single weapon.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Natural daylight, muted saturation, and restrained contrast support the quiet threat without adding atmospheric effects.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): Dew wets the fallen leaves beneath the cypress trees. 선두 일본인 군인: He advances with his rifle at chest height at the front of the silent formation. A thin plume of camp smoke rises beyond the trees.\n\nTHE HAND THAT IS DOING THIS: the object at the centre of this shot is being held, operated, read, aimed or handed over by 일본인 군인들 right now, so 일본인 군인들's hand — and as much of the wrist and forearm as the framing reaches — is in the frame, gripping or resting on that object exactly the way its real use requires. Match that hand to 일본인 군인들: its size, build, skin, age, grooming, sleeve and anything worn on it belong to that person and to no one else.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 선두 일본인 군인 (일본인 남성, 성인 나이대, 자연스러운 검은 머리) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.",
  "roll_prompts": {
   "A": "Stage the shot. The FIRST attached image (SHOT BACKGROUND) is the finished empty background of this shot — keep it EXACTLY: its camera, perspective, architecture, lighting and every fixed feature stay untouched. The SECOND attached image (LAYOUT SKETCH) tells you ONLY where the people go: each sketched person's position, screen size, pose and the gaze/motion arrows. Ignore the sketch's background lines. The CHARACTER REFERENCE photographs show the real people.\nPlace the real people into the background at exactly the sketched positions, sizes and poses, following the arrow directions. No sketch lines or arrows may remain.\n\nCreate ONE FINAL photorealistic live-action film still of the moment below — early 20th-century Korea, Jeolla region; all people are Korean unless explicitly stated otherwise. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 소총을 가슴 높이로 든 채 빽빽한 편백나무 틈새로 한 발 내딛은 자세의 일본인 군인들의 전신.\n\nLOCATION (lock): The damp, leaf-covered floor of the cypress forest below the mountain camp, between closely spaced trunks. The shot takes place here — the FIRST attached image (SHOT BACKGROUND) is this exact place, already built: its ground, structures, horizon, materials and lighting are the finished truth of this location and must not be redesigned or replaced. No location photograph is attached — read the place from that image alone, and add no scenery, structure, vehicle or fixture that it does not already show. This lock governs the place only; the figures in the shot follow the staging and pose instructions.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: wide shot\n- KEY BACKGROUND ELEMENTS: Cypress trees (Dense trees separate the emerging soldiers into two visible branches) — Successive trunks overlap at the frame edges and between figures; used as Create layered partial occlusions while leaving enough gaps to read complete bodies; Fallen leaves (Wet with dew beneath the soldiers' boots); used as Keep the lower frame grounded in the careful placement of each step; Soldiers' rifles (Held at chest height during the advance) — Seen obliquely across individual torsos, with slight differences in carrying angle; used as Repeat the armed threat across the formation without enlarging any single weapon.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Natural daylight, muted saturation, and restrained contrast support the quiet threat without adding atmospheric effects.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): Dew wets the fallen leaves beneath the cypress trees. 선두 일본인 군인: He advances with his rifle at chest height at the front of the silent formation. A thin plume of camp smoke rises beyond the trees.\n\nTHE HAND THAT IS DOING THIS: the object at the centre of this shot is being held, operated, read, aimed or handed over by 일본인 군인들 right now, so 일본인 군인들's hand — and as much of the wrist and forearm as the framing reaches — is in the frame, gripping or resting on that object exactly the way its real use requires. Match that hand to 일본인 군인들: its size, build, skin, age, grooming, sleeve and anything worn on it belong to that person and to no one else.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 선두 일본인 군인 (일본인 남성, 성인 나이대, 자연스러운 검은 머리) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.",
   "B": "Create ONE FINAL photorealistic live-action film still of the moment below — early 20th-century Korea, Jeolla region; all people are Korean unless explicitly stated otherwise. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 소총을 가슴 높이로 든 채 빽빽한 편백나무 틈새로 한 발 내딛은 자세의 일본인 군인들의 전신.\n\nLOCATION (lock): The damp, leaf-covered floor of the cypress forest below the mountain camp, between closely spaced trunks. The shot takes place here — the attached LOCATION PHOTOGRAPH shows the exact spot.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: wide shot\n- KEY BACKGROUND ELEMENTS: Cypress trees (Dense trees separate the emerging soldiers into two visible branches) — Successive trunks overlap at the frame edges and between figures; used as Create layered partial occlusions while leaving enough gaps to read complete bodies; Fallen leaves (Wet with dew beneath the soldiers' boots); used as Keep the lower frame grounded in the careful placement of each step; Soldiers' rifles (Held at chest height during the advance) — Seen obliquely across individual torsos, with slight differences in carrying angle; used as Repeat the armed threat across the formation without enlarging any single weapon.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Natural daylight, muted saturation, and restrained contrast support the quiet threat without adding atmospheric effects.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): Dew wets the fallen leaves beneath the cypress trees. 선두 일본인 군인: He advances with his rifle at chest height at the front of the silent formation. A thin plume of camp smoke rises beyond the trees.\n\nTHE HAND THAT IS DOING THIS: the object at the centre of this shot is being held, operated, read, aimed or handed over by 일본인 군인들 right now, so 일본인 군인들's hand — and as much of the wrist and forearm as the framing reaches — is in the frame, gripping or resting on that object exactly the way its real use requires. Match that hand to 일본인 군인들: its size, build, skin, age, grooming, sleeve and anything worn on it belong to that person and to no one else.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 선두 일본인 군인 (일본인 남성, 성인 나이대, 자연스러운 검은 머리) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.\n\nCOMPOSITION VARIATION (this candidate only): keep the place, the people and the moment exactly as specified, and keep from the CAMERA & FRAME contract its subject scale, screen placement and key background elements exactly. For this candidate only, its camera angle and height are deliberately left open: choose a camera position distinctly different from the obvious one, as a film director picking a second setup on the same blocking. Stay true to any explicit framing the SHOT TEXT itself dictates."
  },
  "roll_refs": {
   "A": [
    {
     "label": "SHOT BACKGROUND",
     "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/db00517b-165b-4005-a287-f28936b59b95/images/272bd7af-9294-4490-98c6-88a40e9deacc/scene/recipe/S49sh2__bgfirst_bg.png",
     "asset_id": "d1dc39c5-ff26-4654-a17c-8e173fecc54c",
     "role": "bgfirst_bg"
    },
    {
     "label": "LAYOUT SKETCH (people placement only)",
     "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/db00517b-165b-4005-a287-f28936b59b95/images/272bd7af-9294-4490-98c6-88a40e9deacc/conti/conti_S49sh2.png",
     "asset_id": "c27b30b3-1418-418d-b990-faa3668679a4",
     "role": "conti_light"
    },
    {
     "label": "CHARACTER REFERENCE — 선두 일본인 군인: the exact person appearing in this shot; match face, hair and build exactly.",
     "path": "<bytes:1259583>",
     "asset_id": "7a313767-8d10-49c9-93e6-6da394fb77e3",
     "role": "character_ref"
    },
    {
     "label": "PROP REFERENCE — 근대식 소총: the exact object appearing in this shot; match its look, material and wear exactly.",
     "path": "<bytes:828568>",
     "asset_id": "7543806c-8087-4caa-8d7b-56d10864fd3c",
     "role": "prop_ref"
    }
   ],
   "B": [
    {
     "label": "LOCATION PHOTOGRAPH — the exact place of this shot: its architecture, materials, fixed features and lighting mood are spatial truth; stage the moment inside this place. Never copy its camera framing.",
     "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/db00517b-165b-4005-a287-f28936b59b95/images/272bd7af-9294-4490-98c6-88a40e9deacc/scene/recipe/groupbg_outer_cypress_forest_e3184b.png",
     "asset_id": "e9399fca-b21e-44d7-a40a-5ebc8782d296",
     "role": "bgfirst_group_bg"
    },
    {
     "label": "CHARACTER REFERENCE — 선두 일본인 군인: the exact person appearing in this shot; match face, hair and build exactly.",
     "path": "<bytes:1259583>",
     "asset_id": "7a313767-8d10-49c9-93e6-6da394fb77e3",
     "role": "character_ref"
    },
    {
     "label": "PROP REFERENCE — 근대식 소총: the exact object appearing in this shot; match its look, material and wear exactly.",
     "path": "<bytes:828568>",
     "asset_id": "7543806c-8087-4caa-8d7b-56d10864fd3c",
     "role": "prop_ref"
    }
   ]
  },
  "shot_run_spend_attempt_count": 1,
  "judge_flip_skipped": "judge_fn_owns_order",
  "readings": [
   {
    "label": "A",
    "direction": "군인들의 시선이 좌측 전방을 향하고 있으며, 총구 역시 비스듬히 위쪽이나 앞쪽을 향하고 있습니다.",
    "built_space": "편백나무 숲과 낙엽이 덮인 바닥, 그리고 나무 기둥에 의한 전경과 배경의 분리 등 공간적 배경은 잘 구성되어 있습니다.",
    "entities": "군인들이 헬멧을 쓰고 소총을 들고 있으나, 선두 군인의 얼굴이 레퍼런스 인물과 거의 닮지 않았습니다.",
    "hard_violations": [
     "[gemini-pro] 오른쪽 앞부분에 위치한 군인의 몸이 나무 옆 허공에 떠 있음 (발이 땅에 닿지 않음)",
     "[gpt-high] 오른쪽 선두 병사의 두 군화가 나무 밑동 앞 지면에서 떨어져 있으며, 발을 받치는 돌·뿌리·발판이나 나무를 붙잡은 손이 없다. 도약의 추진이나 착지 동작도 없이 보행 자세로 공중에 떠 있어 물리적으로 불가능한 배치다."
    ],
    "physics": "오른쪽 전경에 있는 군인의 양발이 지면에서 떨어져 허공에 떠 있으며, 어떤 물리적 지지대도 없이 나무 옆면에 붙어 있는 불가능한 자세를 취하고 있습니다."
   },
   {
    "label": "B",
    "direction": "군인들의 시선과 가슴 높이로 든 총구들이 전방을 향해 있으며, 앞으로 나아가는 방향성을 뚜렷하게 보여줍니다.",
    "built_space": "빽빽한 편백나무들 사이로 낙엽이 덮인 숲 바닥이 사실적으로 묘사되었고, 프롬프트의 요구대로 나무들이 군인들을 두 무리로 자연스럽게 나누고 있습니다. 배경에는 옅은 연기가 피어오릅니다.",
    "entities": "레퍼런스와 일치하는 얼굴을 가진 선두 군인이 보이며, 나머지 군인들도 복장과 무장을 갖춘 채 묘사되었습니다. 레퍼런스로 주어진 근대식 소총의 외형도 잘 살렸습니다.",
    "hard_violations": [],
    "physics": "모든 군인의 발이 바닥의 낙엽 위에 단단히 지탱되어 있으며, 걷는 자세와 체중 이동, 그리고 가슴 높이에서 총을 쥐고 있는 손의 파지법이 자연스럽습니다."
   }
  ],
  "cross_model_order": {
   "policy": "cross_model_order_v2_slot0_fwd_slot1_rev_parallel2",
   "slots": [
    {
     "model": "gemini-pro",
     "order": "forward",
     "display_to_canonical": {
      "A": "A",
      "B": "B"
     },
     "raw": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 9,
        "verdict_ko": "지정된 숲의 배경 안에서 두 갈래로 나뉘어 진격하는 군인들의 구도를 정확히 담아냈으며, 선두 군인의 외모와 소총 레퍼런스도 충실히 반영되었습니다."
       },
       {
        "label": "A",
        "score": 2,
        "verdict_ko": "화면 우측 전경에 배치된 군인이 땅을 딛지 않고 허공에 떠 있는 물리적 불가능 상태로 묘사되어 치명적인 오류를 범했습니다."
       }
      ],
      "readings": [
       {
        "label": "B",
        "direction": "군인들의 시선과 가슴 높이로 든 총구들이 전방을 향해 있으며, 앞으로 나아가는 방향성을 뚜렷하게 보여줍니다.",
        "built_space": "빽빽한 편백나무들 사이로 낙엽이 덮인 숲 바닥이 사실적으로 묘사되었고, 프롬프트의 요구대로 나무들이 군인들을 두 무리로 자연스럽게 나누고 있습니다. 배경에는 옅은 연기가 피어오릅니다.",
        "entities": "레퍼런스와 일치하는 얼굴을 가진 선두 군인이 보이며, 나머지 군인들도 복장과 무장을 갖춘 채 묘사되었습니다. 레퍼런스로 주어진 근대식 소총의 외형도 잘 살렸습니다.",
        "hard_violations": [],
        "physics": "모든 군인의 발이 바닥의 낙엽 위에 단단히 지탱되어 있으며, 걷는 자세와 체중 이동, 그리고 가슴 높이에서 총을 쥐고 있는 손의 파지법이 자연스럽습니다."
       },
       {
        "label": "A",
        "direction": "군인들의 시선이 좌측 전방을 향하고 있으며, 총구 역시 비스듬히 위쪽이나 앞쪽을 향하고 있습니다.",
        "built_space": "편백나무 숲과 낙엽이 덮인 바닥, 그리고 나무 기둥에 의한 전경과 배경의 분리 등 공간적 배경은 잘 구성되어 있습니다.",
        "entities": "군인들이 헬멧을 쓰고 소총을 들고 있으나, 선두 군인의 얼굴이 레퍼런스 인물과 거의 닮지 않았습니다.",
        "hard_violations": [
         "오른쪽 앞부분에 위치한 군인의 몸이 나무 옆 허공에 떠 있음 (발이 땅에 닿지 않음)"
        ],
        "physics": "오른쪽 전경에 있는 군인의 양발이 지면에서 떨어져 허공에 떠 있으며, 어떤 물리적 지지대도 없이 나무 옆면에 붙어 있는 불가능한 자세를 취하고 있습니다."
       }
      ],
      "all_candidates_fail": false
     },
     "normalized": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 9,
        "verdict_ko": "지정된 숲의 배경 안에서 두 갈래로 나뉘어 진격하는 군인들의 구도를 정확히 담아냈으며, 선두 군인의 외모와 소총 레퍼런스도 충실히 반영되었습니다."
       },
       {
        "label": "A",
        "score": 2,
        "verdict_ko": "화면 우측 전경에 배치된 군인이 땅을 딛지 않고 허공에 떠 있는 물리적 불가능 상태로 묘사되어 치명적인 오류를 범했습니다."
       }
      ],
      "all_candidates_fail": false,
      "readings": [
       {
        "label": "B",
        "direction": "군인들의 시선과 가슴 높이로 든 총구들이 전방을 향해 있으며, 앞으로 나아가는 방향성을 뚜렷하게 보여줍니다.",
        "built_space": "빽빽한 편백나무들 사이로 낙엽이 덮인 숲 바닥이 사실적으로 묘사되었고, 프롬프트의 요구대로 나무들이 군인들을 두 무리로 자연스럽게 나누고 있습니다. 배경에는 옅은 연기가 피어오릅니다.",
        "entities": "레퍼런스와 일치하는 얼굴을 가진 선두 군인이 보이며, 나머지 군인들도 복장과 무장을 갖춘 채 묘사되었습니다. 레퍼런스로 주어진 근대식 소총의 외형도 잘 살렸습니다.",
        "hard_violations": [],
        "physics": "모든 군인의 발이 바닥의 낙엽 위에 단단히 지탱되어 있으며, 걷는 자세와 체중 이동, 그리고 가슴 높이에서 총을 쥐고 있는 손의 파지법이 자연스럽습니다."
       },
       {
        "label": "A",
        "direction": "군인들의 시선이 좌측 전방을 향하고 있으며, 총구 역시 비스듬히 위쪽이나 앞쪽을 향하고 있습니다.",
        "built_space": "편백나무 숲과 낙엽이 덮인 바닥, 그리고 나무 기둥에 의한 전경과 배경의 분리 등 공간적 배경은 잘 구성되어 있습니다.",
        "entities": "군인들이 헬멧을 쓰고 소총을 들고 있으나, 선두 군인의 얼굴이 레퍼런스 인물과 거의 닮지 않았습니다.",
        "hard_violations": [
         "오른쪽 앞부분에 위치한 군인의 몸이 나무 옆 허공에 떠 있음 (발이 땅에 닿지 않음)"
        ],
        "physics": "오른쪽 전경에 있는 군인의 양발이 지면에서 떨어져 허공에 떠 있으며, 어떤 물리적 지지대도 없이 나무 옆면에 붙어 있는 불가능한 자세를 취하고 있습니다."
       }
      ]
     },
     "ok": true
    },
    {
     "model": "gpt-high",
     "order": "reverse",
     "display_to_canonical": {
      "A": "B",
      "B": "A"
     },
     "raw": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 8,
        "verdict_ko": "나무 사이로 전신이 드러나는 두 갈래 대열과 지면을 딛고 전진하는 순간을 충실히 구현했지만, 일부 소총의 높이와 각도 및 비교적 넓게 퍼진 연기는 지시와 차이가 있습니다."
       },
       {
        "label": "B",
        "score": 2,
        "verdict_ko": "전신 와이드 구도와 가슴 높이 소총은 맞지만, 오른쪽 선두 병사의 두 발이 지지 없이 나무 앞 공중에 떠 있어 조심스럽게 한 발 내딛는 장면으로 사용할 수 없습니다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "왼쪽 병사들은 화면 왼쪽 앞의 숲을 보고 총구도 대체로 왼쪽으로 향한다. 중앙 병사는 정면을 경계하면서 소총을 오른쪽 위로 가파르게 들고 있으며, 오른쪽 대열의 총구들은 오른쪽 위의 나무 사이를 향한다. 특정 표적을 겨누라는 지시는 없으므로 표적 누락은 아니지만, 일부 소총은 가슴을 비스듬히 가로지르는 운반 자세보다 세워 든 자세에 가깝다. 병사들의 몸은 대체로 카메라 쪽으로 전진한다.",
        "built_space": "인공 구조물이나 고정 설비는 없다. 화면 왼쪽 중앙과 오른쪽 가장자리의 굵은 줄기, 그 사이의 연속된 가는 줄기들이 대열을 가리고 나누며, 여러 앞줄 병사의 머리부터 군화까지 틈새로 보인다. 이끼 낀 뿌리와 돌, 오른쪽 쓰러진 나무 더미, 낙엽 덮인 경사면은 장소 참조의 주요 특징과 잘 맞는다. 일부 뒤쪽 병사가 줄기에 가려지는 것은 요청된 층층의 부분 가림에 해당한다.",
        "entities": "여러 성인 동아시아계 남성이 카키색 군복과 군모, 각반, 군화를 착용한 일본군 대열로 묘사된다. 국적 자체는 외모만으로 확인할 수 없으며, 선두의 얼굴은 참조와 일부 닮았지만 작은 얼굴 크기와 군모 때문에 정확한 동일 인물 및 검은 머리 일치는 확인하기 어렵다. 목제 개머리판과 금속 총열을 가진 긴 소총들이 보이나, 참조 소총의 독특한 격발장치와 황동 띠까지 동일하게 재현되지는 않았다. 편백림과 낙엽, 나무 뒤로 올라가는 연기는 보이며, 낙엽의 이슬은 뚜렷하지 않고 연기는 가느다란 한 줄기보다 넓게 퍼져 있다. 읽을 수 있는 글자는 없다.",
        "hard_violations": [],
        "physics": "잘 보이는 앞줄 병사들은 한쪽 군화로 낙엽 지면을 딛고 다른 발을 옮기거나, 벌린 두 발 사이로 체중을 이동하고 있다. 들어 올린 발은 지지하는 반대쪽 다리와 연결되어 보행으로 설명된다. 소총은 개머리판·방아쇠 부근과 앞몸통에 놓인 양손으로 지지되며, 뒤쪽 인물의 가려진 발을 근거로 부유를 단정할 수는 없다. 명백하게 지지 없이 떠 있는 신체나 소품은 보이지 않는다."
       },
       {
        "label": "B",
        "direction": "좌우 앞줄 병사와 뒤따르는 병사들은 대부분 화면 오른쪽의 숲을 바라보고, 소총 총구도 오른쪽 또는 오른쪽 위를 향한다. 소총은 대체로 가슴 아래쪽을 비스듬히 가로지르며 양손으로 운반된다. 특정 인물이나 물체를 조준하는 장면은 아니고, 오른쪽 진행 방향을 경계하는 대열로 읽힌다.",
        "built_space": "인공 구조물이나 고정 설비는 없다. 중앙의 굵은 줄기들이 여덟 병사를 좌우 네 명씩의 두 무리로 나누고, 왼쪽 가장자리의 가까운 줄기가 전경을 가린다. 낙엽, 이끼 낀 뿌리, 돌과 오른쪽 고사목은 참조 장소의 재료와 숲 환경에 부합한다. 다만 오른쪽 선두 병사는 굵은 나무의 밑동보다 높은 위치에 겹쳐 있으며, 그 높이에 발을 놓을 지형이나 발판은 보이지 않는다.",
        "entities": "여덟 명 모두 성인 동아시아계 남성 군인으로 보이며 카키색 군복, 철모, 군화 또는 각반을 착용한다. 선두는 측면 얼굴이고 머리는 철모에 가려져 있어 참조 인물의 얼굴과 머리 일치를 확정하기 어렵다. 소총은 목재와 금속의 긴 총기로 읽히지만 참조의 격발장치·황동 띠·총검 형태를 정확히 따르지는 않는다. 편백나무와 낙엽, 오른쪽 뒤의 비교적 가느다란 연기가 보이고 낮의 자연광도 맞지만, 낙엽 위 이슬은 명확하지 않다. 읽을 수 있는 글자는 없다.",
        "hard_violations": [
         "오른쪽 선두 병사의 두 군화가 나무 밑동 앞 지면에서 떨어져 있으며, 발을 받치는 돌·뿌리·발판이나 나무를 붙잡은 손이 없다. 도약의 추진이나 착지 동작도 없이 보행 자세로 공중에 떠 있어 물리적으로 불가능한 배치다."
        ],
        "physics": "왼쪽 선두와 대부분의 후속 병사들은 낙엽 바닥에 발을 두고 있으며, 소총도 양손으로 지지한다. 그러나 오른쪽 선두는 두 발 모두 나무 앞 공중에 놓여 있고 아무것도 몸을 지지하지 않는다. 한쪽 군화가 수직 줄기와 겹쳐 보이지만 올라설 수 있는 면은 없으며, 다리를 벌린 자세에도 도약을 설명할 추진이나 착지 관계가 없다. 따라서 이 인물은 요청된 한 걸음 전진이 아니라 지지 없는 부유로 읽힌다."
       }
      ],
      "all_candidates_fail": false
     },
     "normalized": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 8,
        "verdict_ko": "나무 사이로 전신이 드러나는 두 갈래 대열과 지면을 딛고 전진하는 순간을 충실히 구현했지만, 일부 소총의 높이와 각도 및 비교적 넓게 퍼진 연기는 지시와 차이가 있습니다."
       },
       {
        "label": "A",
        "score": 2,
        "verdict_ko": "전신 와이드 구도와 가슴 높이 소총은 맞지만, 오른쪽 선두 병사의 두 발이 지지 없이 나무 앞 공중에 떠 있어 조심스럽게 한 발 내딛는 장면으로 사용할 수 없습니다."
       }
      ],
      "all_candidates_fail": false,
      "readings": [
       {
        "label": "B",
        "direction": "왼쪽 병사들은 화면 왼쪽 앞의 숲을 보고 총구도 대체로 왼쪽으로 향한다. 중앙 병사는 정면을 경계하면서 소총을 오른쪽 위로 가파르게 들고 있으며, 오른쪽 대열의 총구들은 오른쪽 위의 나무 사이를 향한다. 특정 표적을 겨누라는 지시는 없으므로 표적 누락은 아니지만, 일부 소총은 가슴을 비스듬히 가로지르는 운반 자세보다 세워 든 자세에 가깝다. 병사들의 몸은 대체로 카메라 쪽으로 전진한다.",
        "built_space": "인공 구조물이나 고정 설비는 없다. 화면 왼쪽 중앙과 오른쪽 가장자리의 굵은 줄기, 그 사이의 연속된 가는 줄기들이 대열을 가리고 나누며, 여러 앞줄 병사의 머리부터 군화까지 틈새로 보인다. 이끼 낀 뿌리와 돌, 오른쪽 쓰러진 나무 더미, 낙엽 덮인 경사면은 장소 참조의 주요 특징과 잘 맞는다. 일부 뒤쪽 병사가 줄기에 가려지는 것은 요청된 층층의 부분 가림에 해당한다.",
        "entities": "여러 성인 동아시아계 남성이 카키색 군복과 군모, 각반, 군화를 착용한 일본군 대열로 묘사된다. 국적 자체는 외모만으로 확인할 수 없으며, 선두의 얼굴은 참조와 일부 닮았지만 작은 얼굴 크기와 군모 때문에 정확한 동일 인물 및 검은 머리 일치는 확인하기 어렵다. 목제 개머리판과 금속 총열을 가진 긴 소총들이 보이나, 참조 소총의 독특한 격발장치와 황동 띠까지 동일하게 재현되지는 않았다. 편백림과 낙엽, 나무 뒤로 올라가는 연기는 보이며, 낙엽의 이슬은 뚜렷하지 않고 연기는 가느다란 한 줄기보다 넓게 퍼져 있다. 읽을 수 있는 글자는 없다.",
        "hard_violations": [],
        "physics": "잘 보이는 앞줄 병사들은 한쪽 군화로 낙엽 지면을 딛고 다른 발을 옮기거나, 벌린 두 발 사이로 체중을 이동하고 있다. 들어 올린 발은 지지하는 반대쪽 다리와 연결되어 보행으로 설명된다. 소총은 개머리판·방아쇠 부근과 앞몸통에 놓인 양손으로 지지되며, 뒤쪽 인물의 가려진 발을 근거로 부유를 단정할 수는 없다. 명백하게 지지 없이 떠 있는 신체나 소품은 보이지 않는다."
       },
       {
        "label": "A",
        "direction": "좌우 앞줄 병사와 뒤따르는 병사들은 대부분 화면 오른쪽의 숲을 바라보고, 소총 총구도 오른쪽 또는 오른쪽 위를 향한다. 소총은 대체로 가슴 아래쪽을 비스듬히 가로지르며 양손으로 운반된다. 특정 인물이나 물체를 조준하는 장면은 아니고, 오른쪽 진행 방향을 경계하는 대열로 읽힌다.",
        "built_space": "인공 구조물이나 고정 설비는 없다. 중앙의 굵은 줄기들이 여덟 병사를 좌우 네 명씩의 두 무리로 나누고, 왼쪽 가장자리의 가까운 줄기가 전경을 가린다. 낙엽, 이끼 낀 뿌리, 돌과 오른쪽 고사목은 참조 장소의 재료와 숲 환경에 부합한다. 다만 오른쪽 선두 병사는 굵은 나무의 밑동보다 높은 위치에 겹쳐 있으며, 그 높이에 발을 놓을 지형이나 발판은 보이지 않는다.",
        "entities": "여덟 명 모두 성인 동아시아계 남성 군인으로 보이며 카키색 군복, 철모, 군화 또는 각반을 착용한다. 선두는 측면 얼굴이고 머리는 철모에 가려져 있어 참조 인물의 얼굴과 머리 일치를 확정하기 어렵다. 소총은 목재와 금속의 긴 총기로 읽히지만 참조의 격발장치·황동 띠·총검 형태를 정확히 따르지는 않는다. 편백나무와 낙엽, 오른쪽 뒤의 비교적 가느다란 연기가 보이고 낮의 자연광도 맞지만, 낙엽 위 이슬은 명확하지 않다. 읽을 수 있는 글자는 없다.",
        "hard_violations": [
         "오른쪽 선두 병사의 두 군화가 나무 밑동 앞 지면에서 떨어져 있으며, 발을 받치는 돌·뿌리·발판이나 나무를 붙잡은 손이 없다. 도약의 추진이나 착지 동작도 없이 보행 자세로 공중에 떠 있어 물리적으로 불가능한 배치다."
        ],
        "physics": "왼쪽 선두와 대부분의 후속 병사들은 낙엽 바닥에 발을 두고 있으며, 소총도 양손으로 지지한다. 그러나 오른쪽 선두는 두 발 모두 나무 앞 공중에 놓여 있고 아무것도 몸을 지지하지 않는다. 한쪽 군화가 수직 줄기와 겹쳐 보이지만 올라설 수 있는 면은 없으며, 다리를 벌린 자세에도 도약을 설명할 추진이나 착지 관계가 없다. 따라서 이 인물은 요청된 한 걸음 전진이 아니라 지지 없는 부유로 읽힌다."
       }
      ]
     },
     "ok": true
    }
   ],
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "slot_winner_match": true,
   "slot_winner": {
    "gemini-pro": "B",
    "gpt-high": "B"
   },
   "route": "cross_slot_agree"
  },
  "dual": {
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "normalized": {
    "A": 0.472,
    "B": 2.0
   },
   "adjusted": {
    "A": 0.222,
    "B": 2.0
   },
   "violations": {
    "A": [
     "[gemini-pro] 오른쪽 앞부분에 위치한 군인의 몸이 나무 옆 허공에 떠 있음 (발이 땅에 닿지 않음)",
     "[gpt-high] 오른쪽 선두 병사의 두 군화가 나무 밑동 앞 지면에서 떨어져 있으며, 발을 받치는 돌·뿌리·발판이나 나무를 붙잡은 손이 없다. 도약의 추진이나 착지 동작도 없이 보행 자세로 공중에 떠 있어 물리적으로 불가능한 배치다."
    ]
   },
   "per_model_winner": {
    "gemini-pro": "B",
    "gpt-high": "B"
   },
   "agreed": true
  },
  "totals": {
   "B": 2000,
   "A": 222
  },
  "selected": "B",
  "ranking": [
   "B",
   "A"
  ],
  "verdicts": [
   {
    "label": "B",
    "score": 2000,
    "verdict_ko": "지정된 숲의 배경 안에서 두 갈래로 나뉘어 진격하는 군인들의 구도를 정확히 담아냈으며, 선두 군인의 외모와 소총 레퍼런스도 충실히 반영되었습니다."
   },
   {
    "label": "A",
    "score": 222,
    "verdict_ko": "화면 우측 전경에 배치된 군인이 땅을 딛지 않고 허공에 떠 있는 물리적 불가능 상태로 묘사되어 치명적인 오류를 범했습니다.  ★위반: [gemini-pro] 오른쪽 앞부분에 위치한 군인의 몸이 나무 옆 허공에 떠 있음 (발이 땅에 닿지 않음) / [gpt-high] 오른쪽 선두 병사의 두 군화가 나무 밑동 앞 지면에서 떨어져 있으며, 발을 받치는 돌·뿌리·발판이나 나무를 붙잡은 손이 없다. 도약의 추진이나 착지 동작도 없이 보행 자세로 공중에 떠 있어 물리적으로 불가능한 배치다."
   }
  ],
  "refs": [
   {
    "label": "LOCATION PHOTOGRAPH — the exact place of this shot: its architecture, materials, fixed features and lighting mood are spatial truth; stage the moment inside this place. Never copy its camera framing.",
    "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/db00517b-165b-4005-a287-f28936b59b95/images/272bd7af-9294-4490-98c6-88a40e9deacc/scene/recipe/groupbg_outer_cypress_forest_e3184b.png",
    "asset_id": "e9399fca-b21e-44d7-a40a-5ebc8782d296",
    "role": "bgfirst_group_bg"
   },
   {
    "label": "CHARACTER REFERENCE — 선두 일본인 군인: the exact person appearing in this shot; match face, hair and build exactly.",
    "path": "<bytes:1259583>",
    "asset_id": "7a313767-8d10-49c9-93e6-6da394fb77e3",
    "role": "character_ref"
   },
   {
    "label": "PROP REFERENCE — 근대식 소총: the exact object appearing in this shot; match its look, material and wear exactly.",
    "path": "<bytes:828568>",
    "asset_id": "7543806c-8087-4caa-8d7b-56d10864fd3c",
    "role": "prop_ref"
   }
  ],
  "critique_skipped": true,
  "shot_run_produced": false,
  "shot_run_uid": "06aa1435-5249-7129-81e3-1d145a9b3594",
  "bgfirst": {
   "bg_path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/db00517b-165b-4005-a287-f28936b59b95/images/272bd7af-9294-4490-98c6-88a40e9deacc/scene/recipe/S49sh2__bgfirst_bg.png",
   "bg_asset_id": "d1dc39c5-ff26-4654-a17c-8e173fecc54c",
   "bg_record_key": "S49sh2::bgfirst_bg",
   "chain_winner": false,
   "authority": "groupbg",
   "group_key": "outer_cypress_forest",
   "groupbg_asset_id": "e9399fca-b21e-44d7-a40a-5ebc8782d296"
  },
  "ref_mode": "그룹 배경+엔티티 (2택1: 무콘티 승)",
  "share_plan": {
   "ref_plan": "background"
  }
 },
 "S49sh2::cine": {
  "applied": true,
  "attempted_at": "2026-09-09T06:36:24.879904+00:00",
  "fingerprint": "df4fa1bf8ddb34f406650b88cbdebb5ccbd9f8e7cad1fba9492b3c5cc744bbce",
  "fingerprint_version": 2,
  "provider": "mai",
  "endpoint": "openrouter/chat-completions",
  "model": "microsoft/mai-image-2.6",
  "pack": "24.202608252115",
  "source_file": "S49sh2_sel.png",
  "source_sha256": "badbaa8e90da15c618cac3b2487a42d7b1372a5f0ce900321ac5ed38d5120e37",
  "file": "S49sh2_cine.png",
  "staged_sha256": "45fb659f0f4794022bc95132b3ab10496a6df13c4d17359337df253ab1b8a2b5",
  "latency_ms": 27102
 },
 "S49sh5::signage": {
  "fp": "90ac3e64144b78e6",
  "inscriptions": [],
  "cues": [],
  "dropped": []
 },
 "S49sh5": {
  "input_fingerprint": "c8e8e7ad3414be49",
  "prompt": "Create ONE FINAL photorealistic live-action film still of the moment below — early 20th-century Korea, Jeolla region; all people are Korean unless explicitly stated otherwise. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 나무 사이 저 멀리 하늘로 솟아오르는 가느다란 하얀 연기 한 줄기 전경.\n\nLOCATION (lock): A gap between cypress trunks in the forest below the mountain camp, framing its distant rising smoke. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: wide shot\n- KEY BACKGROUND ELEMENTS: Distant smoke plume (A single thin white plume rises from the direction of the mountain camp); used as Serves as the small, isolated visual destination within the tree gap; Cypress trees (A gap between the trees permits the distant smoke to be seen) — Nearer trunks and branches border the viewing corridor without covering the plume; used as Maintain foreground-to-distance separation and conceal the camp itself.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Restrained daylight contrast separates the distant white smoke from its surroundings without turning the forest into generalized haze.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The people visible in that photo are NOT in this shot: never carry their faces, bodies or clothing onto anyone here. Each person in THIS shot is defined solely by the PEOPLE list, their own CHARACTER REFERENCE images and their wardrobe notes. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): Dew-covered leaves carpet the dense cypress woodland. A single thin plume of camp smoke rises in the distance between the trees.\n\nNO PEOPLE IN THIS SHOT: the shot text shows only the place and its state — no living person appears in frame, and no face, and no figure standing, sitting or passing through. Do not invent someone to occupy the space.\n\nOne thing this rule does NOT forbid: the hand or arm of whoever is doing what the shot text describes. If the moment is an object being held, carried, operated, read, aimed, worn or handed over, the hand doing it belongs in frame, resting on or gripping that object exactly as its real use requires. An object photographed floating unsupported, with the hand that must be holding it erased, is a worse failure than the person you were avoiding — render the hand.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.",
  "roll_prompts": {
   "A": "Create ONE FINAL photorealistic live-action film still of the moment below — early 20th-century Korea, Jeolla region; all people are Korean unless explicitly stated otherwise. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 나무 사이 저 멀리 하늘로 솟아오르는 가느다란 하얀 연기 한 줄기 전경.\n\nLOCATION (lock): A gap between cypress trunks in the forest below the mountain camp, framing its distant rising smoke. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: wide shot\n- KEY BACKGROUND ELEMENTS: Distant smoke plume (A single thin white plume rises from the direction of the mountain camp); used as Serves as the small, isolated visual destination within the tree gap; Cypress trees (A gap between the trees permits the distant smoke to be seen) — Nearer trunks and branches border the viewing corridor without covering the plume; used as Maintain foreground-to-distance separation and conceal the camp itself.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Restrained daylight contrast separates the distant white smoke from its surroundings without turning the forest into generalized haze.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The people visible in that photo are NOT in this shot: never carry their faces, bodies or clothing onto anyone here. Each person in THIS shot is defined solely by the PEOPLE list, their own CHARACTER REFERENCE images and their wardrobe notes. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): Dew-covered leaves carpet the dense cypress woodland. A single thin plume of camp smoke rises in the distance between the trees.\n\nNO PEOPLE IN THIS SHOT: the shot text shows only the place and its state — no living person appears in frame, and no face, and no figure standing, sitting or passing through. Do not invent someone to occupy the space.\n\nOne thing this rule does NOT forbid: the hand or arm of whoever is doing what the shot text describes. If the moment is an object being held, carried, operated, read, aimed, worn or handed over, the hand doing it belongs in frame, resting on or gripping that object exactly as its real use requires. An object photographed floating unsupported, with the hand that must be holding it erased, is a worse failure than the person you were avoiding — render the hand.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.",
   "B": "Create ONE FINAL photorealistic live-action film still of the moment below — early 20th-century Korea, Jeolla region; all people are Korean unless explicitly stated otherwise. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 나무 사이 저 멀리 하늘로 솟아오르는 가느다란 하얀 연기 한 줄기 전경.\n\nLOCATION (lock): A gap between cypress trunks in the forest below the mountain camp, framing its distant rising smoke. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: wide shot\n- KEY BACKGROUND ELEMENTS: Distant smoke plume (A single thin white plume rises from the direction of the mountain camp); used as Serves as the small, isolated visual destination within the tree gap; Cypress trees (A gap between the trees permits the distant smoke to be seen) — Nearer trunks and branches border the viewing corridor without covering the plume; used as Maintain foreground-to-distance separation and conceal the camp itself.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Restrained daylight contrast separates the distant white smoke from its surroundings without turning the forest into generalized haze.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The people visible in that photo are NOT in this shot: never carry their faces, bodies or clothing onto anyone here. Each person in THIS shot is defined solely by the PEOPLE list, their own CHARACTER REFERENCE images and their wardrobe notes. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): Dew-covered leaves carpet the dense cypress woodland. A single thin plume of camp smoke rises in the distance between the trees.\n\nNO PEOPLE IN THIS SHOT: the shot text shows only the place and its state — no living person appears in frame, and no face, and no figure standing, sitting or passing through. Do not invent someone to occupy the space.\n\nOne thing this rule does NOT forbid: the hand or arm of whoever is doing what the shot text describes. If the moment is an object being held, carried, operated, read, aimed, worn or handed over, the hand doing it belongs in frame, resting on or gripping that object exactly as its real use requires. An object photographed floating unsupported, with the hand that must be holding it erased, is a worse failure than the person you were avoiding — render the hand.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.\n\nCOMPOSITION VARIATION (this candidate only): keep the place, the people and the moment exactly as specified, and keep from the CAMERA & FRAME contract its subject scale, screen placement and key background elements exactly. For this candidate only, its camera angle and height are deliberately left open: choose a camera position distinctly different from the obvious one, as a film director picking a second setup on the same blocking. Stay true to any explicit framing the SHOT TEXT itself dictates."
  },
  "shot_run_spend_attempt_count": 1,
  "readings": [
   {
    "label": "A",
    "direction": "카메라는 편백나무들 사이의 틈을 통해 화면 중앙 멀리서 솟아오르는 연기를 향하고 있습니다.",
    "built_space": "키가 큰 편백나무들이 양옆으로 늘어서 시야를 좁혀주는 숲길 형태이며, 바닥에는 이슬 맺힌 녹색 잎과 식물들이 덮여 있습니다. 캠프나 인공 구조물은 보이지 않습니다.",
    "entities": "빽빽한 편백나무 줄기들, 멀리 배경에서 피어오르는 가느다란 연기 한 줄기가 보이며, 지시대로 사람은 등장하지 않습니다.",
    "hard_violations": [],
    "physics": "연기가 대기 중으로 자연스럽게 퍼지며 올라가고 있으며, 물리적으로 어색하거나 지지되지 않고 떠 있는 물체는 없습니다."
   },
   {
    "label": "B",
    "direction": "카메라는 나무 사이의 틈을 통해 숲 안쪽 멀리 솟아오르는 연기와 그 아래의 베이스캠프를 향하고 있습니다.",
    "built_space": "양옆으로 나무들이 서 있고 바닥에는 양치식물과 이끼가 깔려 있습니다. 숲 안쪽 멀리 연기가 시작되는 지점에 여러 채의 텐트와 캠프 시설들이 가려지지 않고 배치되어 있습니다.",
    "entities": "편백나무들, 다소 두껍고 짙은 흰 연기, 캠프 텐트 구조물이 보입니다. 사람은 등장하지 않습니다.",
    "hard_violations": [
     "[gpt-high] 야영지를 감추도록 명시된 중앙 시야 통로에 여러 천막을 가시적 구조물로 추가해, 보여서는 안 되는 야영지 자체를 노출했습니다."
    ],
    "physics": "연기가 정상적으로 위로 피어오르고 있으며, 물리 법칙에 위배되는 요소는 발견되지 않습니다."
   }
  ],
  "cross_model_order": {
   "policy": "cross_model_order_v2_slot0_fwd_slot1_rev_parallel2",
   "slots": [
    {
     "model": "gemini-pro",
     "order": "forward",
     "display_to_canonical": {
      "A": "A",
      "B": "B"
     },
     "raw": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 8,
        "verdict_ko": "캠프 모습을 나무로 가리고 멀리서 피어오르는 가느다란 연기만 보이도록 하라는 프롬프트의 구도 지시를 충실히 따랐습니다."
       },
       {
        "label": "B",
        "score": 4,
        "verdict_ko": "캠프를 숨기라는 명시적인 지시에도 불구하고 연기 아래에 텐트와 구조물들이 선명하게 노출되었으며, 연기 또한 '가느다란'이라는 묘사와 달리 지나치게 두껍습니다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "카메라는 편백나무들 사이의 틈을 통해 화면 중앙 멀리서 솟아오르는 연기를 향하고 있습니다.",
        "built_space": "키가 큰 편백나무들이 양옆으로 늘어서 시야를 좁혀주는 숲길 형태이며, 바닥에는 이슬 맺힌 녹색 잎과 식물들이 덮여 있습니다. 캠프나 인공 구조물은 보이지 않습니다.",
        "entities": "빽빽한 편백나무 줄기들, 멀리 배경에서 피어오르는 가느다란 연기 한 줄기가 보이며, 지시대로 사람은 등장하지 않습니다.",
        "hard_violations": [],
        "physics": "연기가 대기 중으로 자연스럽게 퍼지며 올라가고 있으며, 물리적으로 어색하거나 지지되지 않고 떠 있는 물체는 없습니다."
       },
       {
        "label": "B",
        "direction": "카메라는 나무 사이의 틈을 통해 숲 안쪽 멀리 솟아오르는 연기와 그 아래의 베이스캠프를 향하고 있습니다.",
        "built_space": "양옆으로 나무들이 서 있고 바닥에는 양치식물과 이끼가 깔려 있습니다. 숲 안쪽 멀리 연기가 시작되는 지점에 여러 채의 텐트와 캠프 시설들이 가려지지 않고 배치되어 있습니다.",
        "entities": "편백나무들, 다소 두껍고 짙은 흰 연기, 캠프 텐트 구조물이 보입니다. 사람은 등장하지 않습니다.",
        "hard_violations": [],
        "physics": "연기가 정상적으로 위로 피어오르고 있으며, 물리 법칙에 위배되는 요소는 발견되지 않습니다."
       }
      ],
      "all_candidates_fail": false
     },
     "normalized": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 8,
        "verdict_ko": "캠프 모습을 나무로 가리고 멀리서 피어오르는 가느다란 연기만 보이도록 하라는 프롬프트의 구도 지시를 충실히 따랐습니다."
       },
       {
        "label": "B",
        "score": 4,
        "verdict_ko": "캠프를 숨기라는 명시적인 지시에도 불구하고 연기 아래에 텐트와 구조물들이 선명하게 노출되었으며, 연기 또한 '가느다란'이라는 묘사와 달리 지나치게 두껍습니다."
       }
      ],
      "all_candidates_fail": false,
      "readings": [
       {
        "label": "A",
        "direction": "카메라는 편백나무들 사이의 틈을 통해 화면 중앙 멀리서 솟아오르는 연기를 향하고 있습니다.",
        "built_space": "키가 큰 편백나무들이 양옆으로 늘어서 시야를 좁혀주는 숲길 형태이며, 바닥에는 이슬 맺힌 녹색 잎과 식물들이 덮여 있습니다. 캠프나 인공 구조물은 보이지 않습니다.",
        "entities": "빽빽한 편백나무 줄기들, 멀리 배경에서 피어오르는 가느다란 연기 한 줄기가 보이며, 지시대로 사람은 등장하지 않습니다.",
        "hard_violations": [],
        "physics": "연기가 대기 중으로 자연스럽게 퍼지며 올라가고 있으며, 물리적으로 어색하거나 지지되지 않고 떠 있는 물체는 없습니다."
       },
       {
        "label": "B",
        "direction": "카메라는 나무 사이의 틈을 통해 숲 안쪽 멀리 솟아오르는 연기와 그 아래의 베이스캠프를 향하고 있습니다.",
        "built_space": "양옆으로 나무들이 서 있고 바닥에는 양치식물과 이끼가 깔려 있습니다. 숲 안쪽 멀리 연기가 시작되는 지점에 여러 채의 텐트와 캠프 시설들이 가려지지 않고 배치되어 있습니다.",
        "entities": "편백나무들, 다소 두껍고 짙은 흰 연기, 캠프 텐트 구조물이 보입니다. 사람은 등장하지 않습니다.",
        "hard_violations": [],
        "physics": "연기가 정상적으로 위로 피어오르고 있으며, 물리 법칙에 위배되는 요소는 발견되지 않습니다."
       }
      ]
     },
     "ok": true
    },
    {
     "model": "gpt-high",
     "order": "reverse",
     "display_to_canonical": {
      "A": "B",
      "B": "A"
     },
     "raw": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 4,
        "verdict_ko": "나무 틈으로 상승하는 연기를 담은 와이드 숏이지만, 숨겨야 할 야영지의 천막들을 드러내고 연기 윗부분도 굵게 부풀려 핵심 구도를 어겼습니다."
       },
       {
        "label": "B",
        "score": 8,
        "verdict_ko": "전경 편백 줄기 사이 먼 연기 한 줄기를 고립시키고 야영지를 완전히 가렸으나, 연기가 흰색보다는 회색으로 보이고 바닥 식생은 참조보다 무성합니다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "중앙의 천막 부근에서 연기가 거의 수직으로 하늘을 향해 올라갑니다. 전경 줄기와 가지는 연기 좌우를 둘러싸며 직접 가리지 않습니다. 사람이나 조준 대상은 없습니다.",
        "built_space": "전경과 중경에 편백 줄기가 빽빽하고, 중앙 틈 아래에는 최소 세 동의 삼각 지붕 형태 천막이 노출되어 있습니다. 야영지 자체를 수목으로 숨기라는 공간 조건에 어긋납니다. 참조의 거친 수피와 이끼 낀 뿌리는 이어지지만, 갈색 낙엽 중심의 바닥이 고사리와 이끼 중심으로 바뀌었습니다.",
        "entities": "사람, 얼굴, 글자 없이 편백 숲과 연기가 보입니다. 연기는 아래쪽의 가는 흰 부분에서 위쪽의 굵은 회색 덩어리로 이어져, 멀리 있는 가느다란 흰 연기 한 줄기라는 지정과 차이가 큽니다. 잎은 풍부하지만 이슬방울은 뚜렷하지 않습니다. 화면에 드러난 천막들은 요구된 가려진 야영지 상태에 맞지 않습니다.",
        "hard_violations": [
         "야영지를 감추도록 명시된 중앙 시야 통로에 여러 천막을 가시적 구조물로 추가해, 보여서는 안 되는 야영지 자체를 노출했습니다."
        ],
        "physics": "나무는 땅에 박힌 뿌리로 지지되고, 낙목과 돌은 숲 바닥에 놓여 있습니다. 천막은 지면에 설치된 것으로 보입니다. 연기는 천막 부근의 가려진 지상 발생원에서 부력으로 상승하는 형태이며, 위로 갈수록 퍼지는 현상 자체는 물리적으로 가능합니다. 지지 없이 떠 있는 고체나 몸은 없습니다."
       },
       {
        "label": "B",
        "direction": "중앙의 먼 수관 뒤에서 연기 한 줄기가 위쪽 하늘로 올라가며 상단에서 조금 퍼집니다. 가까운 줄기들이 양옆에서 그 먼 연기를 향한 시야 통로를 만들고, 연기를 가리지 않습니다. 사람이나 무기는 없습니다.",
        "built_space": "좌우의 굵은 전경 줄기와 그 뒤의 여러 가는 줄기가 깊이 있는 숲의 틈을 형성합니다. 중앙 아래를 먼 수관이 막아 야영지의 건물이나 천막은 전혀 보이지 않습니다. 인공 시설과 반사면도 없습니다. 수피와 이끼 낀 밑동은 참조와 유사하나, 바닥의 녹색 잎이 참조보다 훨씬 많습니다.",
        "entities": "편백 숲, 잎으로 덮인 바닥, 멀리 떨어진 단일 연기 기둥이 보이며 사람이나 읽을 수 있는 글자는 없습니다. 연기는 밑부분이 가늘고 멀리 있는 대상으로 읽히지만, 지정된 흰색보다는 청회색에 가깝습니다. 잎 표면에는 밝은 반사가 있으나 개별 이슬방울은 명확하지 않습니다. 낮의 자연광과 선명한 전경은 유지됩니다.",
        "hard_violations": [],
        "physics": "굵은 나무 밑동과 뿌리가 지면에 연결되고, 잎과 작은 식물은 바닥에 놓이거나 땅에서 자랍니다. 연기의 하단은 먼 수관 뒤로 이어져 발생원이 가려진 상승류로 자연스럽게 읽힙니다. 위쪽 확산도 가능한 연기 운동이며, 지지 없는 물체나 인물은 없습니다."
       }
      ],
      "all_candidates_fail": false
     },
     "normalized": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 4,
        "verdict_ko": "나무 틈으로 상승하는 연기를 담은 와이드 숏이지만, 숨겨야 할 야영지의 천막들을 드러내고 연기 윗부분도 굵게 부풀려 핵심 구도를 어겼습니다."
       },
       {
        "label": "A",
        "score": 8,
        "verdict_ko": "전경 편백 줄기 사이 먼 연기 한 줄기를 고립시키고 야영지를 완전히 가렸으나, 연기가 흰색보다는 회색으로 보이고 바닥 식생은 참조보다 무성합니다."
       }
      ],
      "all_candidates_fail": false,
      "readings": [
       {
        "label": "B",
        "direction": "중앙의 천막 부근에서 연기가 거의 수직으로 하늘을 향해 올라갑니다. 전경 줄기와 가지는 연기 좌우를 둘러싸며 직접 가리지 않습니다. 사람이나 조준 대상은 없습니다.",
        "built_space": "전경과 중경에 편백 줄기가 빽빽하고, 중앙 틈 아래에는 최소 세 동의 삼각 지붕 형태 천막이 노출되어 있습니다. 야영지 자체를 수목으로 숨기라는 공간 조건에 어긋납니다. 참조의 거친 수피와 이끼 낀 뿌리는 이어지지만, 갈색 낙엽 중심의 바닥이 고사리와 이끼 중심으로 바뀌었습니다.",
        "entities": "사람, 얼굴, 글자 없이 편백 숲과 연기가 보입니다. 연기는 아래쪽의 가는 흰 부분에서 위쪽의 굵은 회색 덩어리로 이어져, 멀리 있는 가느다란 흰 연기 한 줄기라는 지정과 차이가 큽니다. 잎은 풍부하지만 이슬방울은 뚜렷하지 않습니다. 화면에 드러난 천막들은 요구된 가려진 야영지 상태에 맞지 않습니다.",
        "hard_violations": [
         "야영지를 감추도록 명시된 중앙 시야 통로에 여러 천막을 가시적 구조물로 추가해, 보여서는 안 되는 야영지 자체를 노출했습니다."
        ],
        "physics": "나무는 땅에 박힌 뿌리로 지지되고, 낙목과 돌은 숲 바닥에 놓여 있습니다. 천막은 지면에 설치된 것으로 보입니다. 연기는 천막 부근의 가려진 지상 발생원에서 부력으로 상승하는 형태이며, 위로 갈수록 퍼지는 현상 자체는 물리적으로 가능합니다. 지지 없이 떠 있는 고체나 몸은 없습니다."
       },
       {
        "label": "A",
        "direction": "중앙의 먼 수관 뒤에서 연기 한 줄기가 위쪽 하늘로 올라가며 상단에서 조금 퍼집니다. 가까운 줄기들이 양옆에서 그 먼 연기를 향한 시야 통로를 만들고, 연기를 가리지 않습니다. 사람이나 무기는 없습니다.",
        "built_space": "좌우의 굵은 전경 줄기와 그 뒤의 여러 가는 줄기가 깊이 있는 숲의 틈을 형성합니다. 중앙 아래를 먼 수관이 막아 야영지의 건물이나 천막은 전혀 보이지 않습니다. 인공 시설과 반사면도 없습니다. 수피와 이끼 낀 밑동은 참조와 유사하나, 바닥의 녹색 잎이 참조보다 훨씬 많습니다.",
        "entities": "편백 숲, 잎으로 덮인 바닥, 멀리 떨어진 단일 연기 기둥이 보이며 사람이나 읽을 수 있는 글자는 없습니다. 연기는 밑부분이 가늘고 멀리 있는 대상으로 읽히지만, 지정된 흰색보다는 청회색에 가깝습니다. 잎 표면에는 밝은 반사가 있으나 개별 이슬방울은 명확하지 않습니다. 낮의 자연광과 선명한 전경은 유지됩니다.",
        "hard_violations": [],
        "physics": "굵은 나무 밑동과 뿌리가 지면에 연결되고, 잎과 작은 식물은 바닥에 놓이거나 땅에서 자랍니다. 연기의 하단은 먼 수관 뒤로 이어져 발생원이 가려진 상승류로 자연스럽게 읽힙니다. 위쪽 확산도 가능한 연기 운동이며, 지지 없는 물체나 인물은 없습니다."
       }
      ]
     },
     "ok": true
    }
   ],
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "slot_winner_match": true,
   "slot_winner": {
    "gemini-pro": "A",
    "gpt-high": "A"
   },
   "route": "cross_slot_agree"
  },
  "dual": {
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "normalized": {
    "A": 2.0,
    "B": 1.0
   },
   "adjusted": {
    "A": 2.0,
    "B": 0.75
   },
   "violations": {
    "B": [
     "[gpt-high] 야영지를 감추도록 명시된 중앙 시야 통로에 여러 천막을 가시적 구조물로 추가해, 보여서는 안 되는 야영지 자체를 노출했습니다."
    ]
   },
   "per_model_winner": {
    "gemini-pro": "A",
    "gpt-high": "A"
   },
   "agreed": true
  },
  "totals": {
   "A": 2000,
   "B": 750
  },
  "selected": "A",
  "ranking": [
   "A",
   "B"
  ],
  "verdicts": [
   {
    "label": "A",
    "score": 2000,
    "verdict_ko": "캠프 모습을 나무로 가리고 멀리서 피어오르는 가느다란 연기만 보이도록 하라는 프롬프트의 구도 지시를 충실히 따랐습니다."
   },
   {
    "label": "B",
    "score": 750,
    "verdict_ko": "캠프를 숨기라는 명시적인 지시에도 불구하고 연기 아래에 텐트와 구조물들이 선명하게 노출되었으며, 연기 또한 '가느다란'이라는 묘사와 달리 지나치게 두껍습니다.  ★위반: [gpt-high] 야영지를 감추도록 명시된 중앙 시야 통로에 여러 천막을 가시적 구조물로 추가해, 보여서는 안 되는 야영지 자체를 노출했습니다."
   }
  ],
  "refs": [
   {
    "label": "PREVIOUS SHOT STILL — a visually related earlier shot of this same place: the location's look, materials, fixed features and lighting mood are LOCKED to this photo; never copy its camera framing. The people visible in that photo are NOT in this shot: never carry their faces, bodies or clothing onto anyone here. Each person in THIS shot is defined solely by the PEOPLE list, their own CHARACTER REFERENCE images and their wardrobe notes. If this photo shows a character who cannot move (dead or unconscious), that character's exact pose, position and orientation are ALSO LOCKED — treat their body as a fixed prop of the set that only the camera moves around.",
    "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/db00517b-165b-4005-a287-f28936b59b95/images/272bd7af-9294-4490-98c6-88a40e9deacc/scene/recipe/S49sh2_sel.png",
    "asset_id": "ed3401b3-86cb-49ee-8486-3da6d4f6cb88",
    "role": "prev_still"
   }
  ],
  "critique_skipped": true,
  "shot_run_produced": false,
  "shot_run_uid": "06aa1435-5676-7049-8f5b-b479f7ed83d6",
  "ref_mode": "prev만 (배경 전용·공유 계획)",
  "share_plan": {
   "ref_plan": "prev",
   "prev_anchor_tag": "S49sh2"
  },
  "lane_policy": "share_plan_prev_bgonly"
 },
 "S49sh5::cine": {
  "applied": true,
  "attempted_at": "2026-09-09T06:37:37.851427+00:00",
  "fingerprint": "2c2ceedc6fd67985a35bd0af3a4e4314c5069d43de24a9c0b946f711af44a127",
  "fingerprint_version": 2,
  "provider": "mai",
  "endpoint": "openrouter/chat-completions",
  "model": "microsoft/mai-image-2.6",
  "pack": "24.202608252115",
  "source_file": "S49sh5_sel.png",
  "source_sha256": "d44a69d274ae13ea4cf47a8baaa8aafb5177004bd3d96f9b5338216387cc9258",
  "file": "S49sh5_cine.png",
  "staged_sha256": "2cf27d8ddd6b36450ce3977602484c5114795da9d8cd1380891bfcb6b7718fb2",
  "latency_ms": 28981
 },
 "S49sh7::signage": {
  "fp": "f5eb0883d34900db",
  "inscriptions": [],
  "cues": [],
  "dropped": []
 },
 "S49sh7": {
  "input_fingerprint": "af40f6d7ba9ca34e",
  "prompt": "Create ONE FINAL photorealistic live-action film still of the moment below — early 20th-century Korea, Jeolla region; all people are Korean unless explicitly stated otherwise. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 손짓에 따라 양갈래 방향으로 몸을 꺾어 앞발을 내딛은 자세의 일본인 군인들 역동적인 뒷모습.\n\nLOCATION (lock): The wooded approach below the mountain camp, where the advancing column separates between cypress trunks to either side. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: wide shot\n- FRAME LAYOUT: Left gap receiving one branch of soldiers in the upper-left of the frame, background; Right gap receiving the other branch of soldiers in the upper-right of the frame, background.\n- KEY BACKGROUND ELEMENTS: Left and right gaps between cypresses (The soldiers divide through separate gaps while the central interval opens) — The gaps recede toward the upper-left and upper-right portions of the frame; used as Make the two-part maneuver readable without suggesting a single forward column; Fallen leaves (Dew-wet leaves lie under the carefully placed boots); used as Provide ground continuity beneath the diverging footfalls; Soldiers' rifles (Remain carried at chest height as the soldiers turn) — Partly visible beside torsos from the rear-oblique viewpoint; used as Preserve the armed posture without interrupting the silhouette of the split.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Keep the preceding daylight exposure and subdued palette consistent as the formation separates.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The clothing, hair and overall look of 선두 일본인 군인 — who appear both in that photo and in this shot — are LOCKED to that photo. Anyone else visible in that photo is NOT in this shot: never carry their face, body or clothing onto anyone here. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nPREVIOUS STILL USAGE (follow exactly — what to take from the attached still and what to exclude; it governs place and objects only, never who is in this shot or how the camera sees them): Take the cypress vegetation, damp forest floor, and daytime light. Exclude duplicated soldiers and do not preserve the earlier smoke plume's exact transient shape.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The cypress floor remains covered in dew-wet leaves. 선두 일본인 군인: He directs the split with hand signals while retaining his rifle. The distant camp-smoke plume remains visible through the trees.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 선두 일본인 군인 (일본인 남성, 성인 나이대, 자연스러운 검은 머리) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.",
  "roll_prompts": {
   "A": "Create ONE FINAL photorealistic live-action film still of the moment below — early 20th-century Korea, Jeolla region; all people are Korean unless explicitly stated otherwise. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 손짓에 따라 양갈래 방향으로 몸을 꺾어 앞발을 내딛은 자세의 일본인 군인들 역동적인 뒷모습.\n\nLOCATION (lock): The wooded approach below the mountain camp, where the advancing column separates between cypress trunks to either side. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: wide shot\n- FRAME LAYOUT: Left gap receiving one branch of soldiers in the upper-left of the frame, background; Right gap receiving the other branch of soldiers in the upper-right of the frame, background.\n- KEY BACKGROUND ELEMENTS: Left and right gaps between cypresses (The soldiers divide through separate gaps while the central interval opens) — The gaps recede toward the upper-left and upper-right portions of the frame; used as Make the two-part maneuver readable without suggesting a single forward column; Fallen leaves (Dew-wet leaves lie under the carefully placed boots); used as Provide ground continuity beneath the diverging footfalls; Soldiers' rifles (Remain carried at chest height as the soldiers turn) — Partly visible beside torsos from the rear-oblique viewpoint; used as Preserve the armed posture without interrupting the silhouette of the split.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Keep the preceding daylight exposure and subdued palette consistent as the formation separates.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The clothing, hair and overall look of 선두 일본인 군인 — who appear both in that photo and in this shot — are LOCKED to that photo. Anyone else visible in that photo is NOT in this shot: never carry their face, body or clothing onto anyone here. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nPREVIOUS STILL USAGE (follow exactly — what to take from the attached still and what to exclude; it governs place and objects only, never who is in this shot or how the camera sees them): Take the cypress vegetation, damp forest floor, and daytime light. Exclude duplicated soldiers and do not preserve the earlier smoke plume's exact transient shape.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The cypress floor remains covered in dew-wet leaves. 선두 일본인 군인: He directs the split with hand signals while retaining his rifle. The distant camp-smoke plume remains visible through the trees.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 선두 일본인 군인 (일본인 남성, 성인 나이대, 자연스러운 검은 머리) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.",
   "B": "Create ONE FINAL photorealistic live-action film still of the moment below — early 20th-century Korea, Jeolla region; all people are Korean unless explicitly stated otherwise. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 손짓에 따라 양갈래 방향으로 몸을 꺾어 앞발을 내딛은 자세의 일본인 군인들 역동적인 뒷모습.\n\nLOCATION (lock): The wooded approach below the mountain camp, where the advancing column separates between cypress trunks to either side. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: wide shot\n- FRAME LAYOUT: Left gap receiving one branch of soldiers in the upper-left of the frame, background; Right gap receiving the other branch of soldiers in the upper-right of the frame, background.\n- KEY BACKGROUND ELEMENTS: Left and right gaps between cypresses (The soldiers divide through separate gaps while the central interval opens) — The gaps recede toward the upper-left and upper-right portions of the frame; used as Make the two-part maneuver readable without suggesting a single forward column; Fallen leaves (Dew-wet leaves lie under the carefully placed boots); used as Provide ground continuity beneath the diverging footfalls; Soldiers' rifles (Remain carried at chest height as the soldiers turn) — Partly visible beside torsos from the rear-oblique viewpoint; used as Preserve the armed posture without interrupting the silhouette of the split.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Keep the preceding daylight exposure and subdued palette consistent as the formation separates.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The clothing, hair and overall look of 선두 일본인 군인 — who appear both in that photo and in this shot — are LOCKED to that photo. Anyone else visible in that photo is NOT in this shot: never carry their face, body or clothing onto anyone here. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nPREVIOUS STILL USAGE (follow exactly — what to take from the attached still and what to exclude; it governs place and objects only, never who is in this shot or how the camera sees them): Take the cypress vegetation, damp forest floor, and daytime light. Exclude duplicated soldiers and do not preserve the earlier smoke plume's exact transient shape.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The cypress floor remains covered in dew-wet leaves. 선두 일본인 군인: He directs the split with hand signals while retaining his rifle. The distant camp-smoke plume remains visible through the trees.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 선두 일본인 군인 (일본인 남성, 성인 나이대, 자연스러운 검은 머리) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.\n\nCOMPOSITION VARIATION (this candidate only): keep the place, the people and the moment exactly as specified, and keep from the CAMERA & FRAME contract its subject scale, screen placement and key background elements exactly. For this candidate only, its camera angle and height are deliberately left open: choose a camera position distinctly different from the obvious one, as a film director picking a second setup on the same blocking. Stay true to any explicit framing the SHOT TEXT itself dictates."
  },
  "shot_run_spend_attempt_count": 1,
  "readings": [
   {
    "label": "A",
    "direction": "중앙의 군인이 양팔을 벌려 좌우로 수신호를 보내며, 나머지 군인들의 동선과 무기 방향이 좌우 나무 사이 빈 공간을 향해 나뉩니다.",
    "built_space": "레퍼런스와 일치하는 산림 배경으로, 낙엽이 덮인 지면 위로 편백나무들이 늘어서 있고 병력들이 갈라져 나갈 수 있는 좌우 공간감이 확보되어 있습니다.",
    "entities": "지시를 내리는 선두 군인 및 나머지 병력들이 레퍼런스 이미지와 일치하는 일본군 군복을 입고 등장합니다.",
    "hard_violations": [
     "[gemini-pro] 수신호를 보내는 중앙 군인의 오른손에 들린 소총 하단부가 지지대 없이 허공에서 사라집니다.",
     "[gemini-pro] 중앙 좌측에 위치한 군인의 몸과 손을 두 개의 소총 총열이 물리적으로 불가능하게 교차하며 관통하고 있습니다."
    ],
    "physics": "군인들이 앞발을 내딛는 역동적인 자세를 취하고 있으나, 소총들이 제대로 파지되지 못하고 인체와 융합되거나 공중에 떠 있는 등 무기 지지 상태가 불가능합니다."
   },
   {
    "label": "B",
    "direction": "병력들이 좌우 양방향으로 몸을 틀어 나아가고 있으나, 이를 수신호로 지시하는 명확한 타겟(선두 군인)이 존재하지 않습니다.",
    "built_space": "낙엽이 깔린 바닥과 편백나무 숲 배경이 제시되며, 좌우로 나뉘어 들어갈 수 있는 공간적 틈새가 표현되어 있습니다.",
    "entities": "일본인 군인 복장을 한 인물들이 등장하지만, 프롬프트가 요구한 수신호를 보내는 선두 군인이 없습니다.",
    "hard_violations": [
     "[gemini-pro] 화면 좌측 군인부터 중앙 우측 군인에 이르기까지 소총들이 분리되지 않고 하나의 비정상적으로 긴 막대 형태로 연결되어 다수의 인체를 그대로 관통합니다."
    ],
    "physics": "이동하는 발걸음 자체는 지면에 닿아 있으나, 다수의 인물이 형태가 기형적으로 융합된 하나의 소총 덩어리를 공유하듯 들고 있어 물리적으로 성립하지 않습니다."
   }
  ],
  "cross_model_order": {
   "policy": "cross_model_order_v2_slot0_fwd_slot1_rev_parallel2",
   "slots": [
    {
     "model": "gemini-pro",
     "order": "forward",
     "display_to_canonical": {
      "A": "A",
      "B": "B"
     },
     "raw": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 4,
        "verdict_ko": "수신호로 방향을 지시하는 선두 군인과 양갈래로 나뉘는 병력의 배치는 프롬프트를 잘 따랐으나, 여러 소총들이 인체를 관통하거나 허공에서 잘리는 치명적인 물리적 오류가 발생했습니다."
       },
       {
        "label": "B",
        "score": 2,
        "verdict_ko": "수신호로 대열 분산을 지시하는 선두 군인의 액션이 누락되었으며, 여러 군인의 소총이 하나의 긴 장대처럼 연결되어 몸을 관통하는 심각한 구조적 오류가 있습니다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "중앙의 군인이 양팔을 벌려 좌우로 수신호를 보내며, 나머지 군인들의 동선과 무기 방향이 좌우 나무 사이 빈 공간을 향해 나뉩니다.",
        "built_space": "레퍼런스와 일치하는 산림 배경으로, 낙엽이 덮인 지면 위로 편백나무들이 늘어서 있고 병력들이 갈라져 나갈 수 있는 좌우 공간감이 확보되어 있습니다.",
        "entities": "지시를 내리는 선두 군인 및 나머지 병력들이 레퍼런스 이미지와 일치하는 일본군 군복을 입고 등장합니다.",
        "hard_violations": [
         "수신호를 보내는 중앙 군인의 오른손에 들린 소총 하단부가 지지대 없이 허공에서 사라집니다.",
         "중앙 좌측에 위치한 군인의 몸과 손을 두 개의 소총 총열이 물리적으로 불가능하게 교차하며 관통하고 있습니다."
        ],
        "physics": "군인들이 앞발을 내딛는 역동적인 자세를 취하고 있으나, 소총들이 제대로 파지되지 못하고 인체와 융합되거나 공중에 떠 있는 등 무기 지지 상태가 불가능합니다."
       },
       {
        "label": "B",
        "direction": "병력들이 좌우 양방향으로 몸을 틀어 나아가고 있으나, 이를 수신호로 지시하는 명확한 타겟(선두 군인)이 존재하지 않습니다.",
        "built_space": "낙엽이 깔린 바닥과 편백나무 숲 배경이 제시되며, 좌우로 나뉘어 들어갈 수 있는 공간적 틈새가 표현되어 있습니다.",
        "entities": "일본인 군인 복장을 한 인물들이 등장하지만, 프롬프트가 요구한 수신호를 보내는 선두 군인이 없습니다.",
        "hard_violations": [
         "화면 좌측 군인부터 중앙 우측 군인에 이르기까지 소총들이 분리되지 않고 하나의 비정상적으로 긴 막대 형태로 연결되어 다수의 인체를 그대로 관통합니다."
        ],
        "physics": "이동하는 발걸음 자체는 지면에 닿아 있으나, 다수의 인물이 형태가 기형적으로 융합된 하나의 소총 덩어리를 공유하듯 들고 있어 물리적으로 성립하지 않습니다."
       }
      ],
      "all_candidates_fail": true
     },
     "normalized": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 4,
        "verdict_ko": "수신호로 방향을 지시하는 선두 군인과 양갈래로 나뉘는 병력의 배치는 프롬프트를 잘 따랐으나, 여러 소총들이 인체를 관통하거나 허공에서 잘리는 치명적인 물리적 오류가 발생했습니다."
       },
       {
        "label": "B",
        "score": 2,
        "verdict_ko": "수신호로 대열 분산을 지시하는 선두 군인의 액션이 누락되었으며, 여러 군인의 소총이 하나의 긴 장대처럼 연결되어 몸을 관통하는 심각한 구조적 오류가 있습니다."
       }
      ],
      "all_candidates_fail": true,
      "readings": [
       {
        "label": "A",
        "direction": "중앙의 군인이 양팔을 벌려 좌우로 수신호를 보내며, 나머지 군인들의 동선과 무기 방향이 좌우 나무 사이 빈 공간을 향해 나뉩니다.",
        "built_space": "레퍼런스와 일치하는 산림 배경으로, 낙엽이 덮인 지면 위로 편백나무들이 늘어서 있고 병력들이 갈라져 나갈 수 있는 좌우 공간감이 확보되어 있습니다.",
        "entities": "지시를 내리는 선두 군인 및 나머지 병력들이 레퍼런스 이미지와 일치하는 일본군 군복을 입고 등장합니다.",
        "hard_violations": [
         "수신호를 보내는 중앙 군인의 오른손에 들린 소총 하단부가 지지대 없이 허공에서 사라집니다.",
         "중앙 좌측에 위치한 군인의 몸과 손을 두 개의 소총 총열이 물리적으로 불가능하게 교차하며 관통하고 있습니다."
        ],
        "physics": "군인들이 앞발을 내딛는 역동적인 자세를 취하고 있으나, 소총들이 제대로 파지되지 못하고 인체와 융합되거나 공중에 떠 있는 등 무기 지지 상태가 불가능합니다."
       },
       {
        "label": "B",
        "direction": "병력들이 좌우 양방향으로 몸을 틀어 나아가고 있으나, 이를 수신호로 지시하는 명확한 타겟(선두 군인)이 존재하지 않습니다.",
        "built_space": "낙엽이 깔린 바닥과 편백나무 숲 배경이 제시되며, 좌우로 나뉘어 들어갈 수 있는 공간적 틈새가 표현되어 있습니다.",
        "entities": "일본인 군인 복장을 한 인물들이 등장하지만, 프롬프트가 요구한 수신호를 보내는 선두 군인이 없습니다.",
        "hard_violations": [
         "화면 좌측 군인부터 중앙 우측 군인에 이르기까지 소총들이 분리되지 않고 하나의 비정상적으로 긴 막대 형태로 연결되어 다수의 인체를 그대로 관통합니다."
        ],
        "physics": "이동하는 발걸음 자체는 지면에 닿아 있으나, 다수의 인물이 형태가 기형적으로 융합된 하나의 소총 덩어리를 공유하듯 들고 있어 물리적으로 성립하지 않습니다."
       }
      ]
     },
     "ok": true
    },
    {
     "model": "gpt-high",
     "order": "reverse",
     "display_to_canonical": {
      "A": "B",
      "B": "A"
     },
     "raw": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 8,
        "verdict_ko": "손신호는 보이지 않지만, 몸을 틀고 앞발을 내딛는 뒷모습과 좌우 뒤편으로 갈라지며 중앙을 비우는 와이드 구도가 B보다 충실하다."
       },
       {
        "label": "B",
        "score": 6,
        "verdict_ko": "소총을 유지한 손신호는 명확하지만, 중앙에 정지한 지휘자를 강조하고 병사들이 주로 화면 옆으로 이동하여 요구된 후방 양갈래 분산 구도가 약하다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "왼쪽 병사들은 머리와 몸을 왼쪽 뒤편 나무 사이로 돌려 이동한다. 오른쪽 무리는 등을 보이며 오른쪽 뒤편으로 이어지지만, 가까운 병사들의 몸과 총구 일부는 왼쪽 중앙을 향해 있어 분기 방향의 명료함이 다소 떨어진다. 소총은 대체로 몸통 옆에서 가로 또는 비스듬히 유지되며 특정 인물을 조준하지 않는다. 명확하게 뻗어 방향을 지시하는 손은 보이지 않는다.",
        "built_space": "인공 구조물이나 고정 설비는 없다. 화면 중앙 왼쪽과 오른쪽 가장자리에 굵은 나무줄기가 있고, 그 사이와 바깥쪽으로 병사들이 지나갈 공간이 나뉜다. 왼쪽 무리는 좌상단 깊이로, 오른쪽 무리는 우측 배경으로 이어지며 두 무리 사이에 낙엽 바닥이 드러난다. 편백 줄기, 이끼 낀 뿌리와 돌, 오른쪽 쓰러진 나무, 뒤편 연기가 이전 장면의 숲과 부합한다.",
        "entities": "여러 성인 남성 병사가 군모, 황갈색 군복, 가죽 교차 멜빵과 탄약 주머니, 각반을 착용한다. 대부분 뒷모습이므로 일본인이라는 설정이나 참조 인물의 정확한 얼굴 일치는 확인할 수 없다. 보이는 옆얼굴과 검은 머리는 설정에 어긋나지 않는다. 목재 개머리판과 긴 금속 총열은 보이지만 소품 참조의 격발 장치와 총검까지 일치하는지는 판별하기 어렵다. 낙엽과 먼 연기는 존재하나 낙엽의 이슬 광택은 약하다. 낮의 억제된 색조를 유지하며 읽을 수 있는 글자는 없다.",
        "hard_violations": [],
        "physics": "가까운 병사들은 한쪽 군화를 바닥에 딛고 다른 발을 옮기며, 무릎 굽힘과 몸통 회전이 방향 전환 보행으로 설명된다. 왼쪽 선두의 뒤로 들린 발도 반대쪽 지지발이 있어 공중 부유가 아니다. 소총은 보이는 손과 몸 앞쪽 팔로 지지되고, 주머니와 배낭은 벨트 및 끈에 연결된다. 명백하게 지지 없이 떠 있는 사람이나 물체는 없다."
       },
       {
        "label": "B",
        "direction": "중앙 병사는 등을 보인 채 왼손을 왼쪽 무리 쪽으로 펼치고 오른손 소총의 총구를 우상단으로 세운다. 왼쪽 병사들은 주로 화면 왼쪽 가장자리로, 오른쪽 병사들은 오른쪽 나무 뒤로 움직여 좌우 분산 자체는 읽힌다. 다만 두 무리가 각각 좌상단과 우상단 깊이로 빠지는 것보다는 횡방향 이동에 가깝다. 중앙 오른쪽 병사 한 명은 카메라 쪽으로 얼굴과 가슴을 보여 역동적인 뒷모습이라는 통일성이 깨진다.",
        "built_space": "인공 구조물이나 고정 설비는 없다. 중앙 왼쪽의 큰 편백과 오른쪽 가장자리의 굵은 줄기, 이끼 낀 돌, 오른쪽 쓰러진 나무가 이전 장면과 매우 유사하게 배치되어 있다. 병사들은 줄기 양옆의 실제 통과 공간을 사용한다. 그러나 중앙 통로를 크게 선 지휘자가 차지하여 두 갈래 사이가 열리는 구도가 약하고, 분기 통로도 상단 배경보다 화면 양옆에서 강조된다. 먼 연기는 나무 사이에 남아 있다.",
        "entities": "성인 남성 병사들의 군모, 황갈색 군복, 가죽 장구와 각반은 참조의 복식 계열과 맞는다. 중앙 인물은 뒷모습이어서 참조 얼굴의 동일성은 검증할 수 없으며, 일부 보이는 얼굴은 동아시아계 성인 남성으로 읽히지만 국적은 외모만으로 확인할 수 없다. 중앙 인물은 소총을 놓지 않고 손신호를 보낸다. 다만 그의 소총은 가슴 높이의 휴대 자세가 아니라 머리 위까지 올라간다. 총들의 목재·금속 재질은 보이지만 참조 소총의 세부 기구 일치는 불명확하다. 낙엽, 숲, 낮빛과 연기는 있으며 읽을 수 있는 글자는 없다.",
        "hard_violations": [],
        "physics": "중앙 지휘자는 두 발을 넓게 땅에 딛고 서 있으며, 오른손으로 세운 소총을 잡는다. 이 자세는 신호 동작으로 가능하지만 몸을 꺾어 앞발을 내딛는 순간은 아니다. 주변 병사들의 걸음은 지면에 닿은 군화와 굽힌 무릎으로 지지되고, 나무 뒤의 신체 일부는 자연스럽게 가려진다. 소총과 배낭도 손이나 끈으로 지지되어 명백한 부유나 불가능한 관절은 보이지 않는다."
       }
      ],
      "all_candidates_fail": false
     },
     "normalized": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 8,
        "verdict_ko": "손신호는 보이지 않지만, 몸을 틀고 앞발을 내딛는 뒷모습과 좌우 뒤편으로 갈라지며 중앙을 비우는 와이드 구도가 B보다 충실하다."
       },
       {
        "label": "A",
        "score": 6,
        "verdict_ko": "소총을 유지한 손신호는 명확하지만, 중앙에 정지한 지휘자를 강조하고 병사들이 주로 화면 옆으로 이동하여 요구된 후방 양갈래 분산 구도가 약하다."
       }
      ],
      "all_candidates_fail": false,
      "readings": [
       {
        "label": "B",
        "direction": "왼쪽 병사들은 머리와 몸을 왼쪽 뒤편 나무 사이로 돌려 이동한다. 오른쪽 무리는 등을 보이며 오른쪽 뒤편으로 이어지지만, 가까운 병사들의 몸과 총구 일부는 왼쪽 중앙을 향해 있어 분기 방향의 명료함이 다소 떨어진다. 소총은 대체로 몸통 옆에서 가로 또는 비스듬히 유지되며 특정 인물을 조준하지 않는다. 명확하게 뻗어 방향을 지시하는 손은 보이지 않는다.",
        "built_space": "인공 구조물이나 고정 설비는 없다. 화면 중앙 왼쪽과 오른쪽 가장자리에 굵은 나무줄기가 있고, 그 사이와 바깥쪽으로 병사들이 지나갈 공간이 나뉜다. 왼쪽 무리는 좌상단 깊이로, 오른쪽 무리는 우측 배경으로 이어지며 두 무리 사이에 낙엽 바닥이 드러난다. 편백 줄기, 이끼 낀 뿌리와 돌, 오른쪽 쓰러진 나무, 뒤편 연기가 이전 장면의 숲과 부합한다.",
        "entities": "여러 성인 남성 병사가 군모, 황갈색 군복, 가죽 교차 멜빵과 탄약 주머니, 각반을 착용한다. 대부분 뒷모습이므로 일본인이라는 설정이나 참조 인물의 정확한 얼굴 일치는 확인할 수 없다. 보이는 옆얼굴과 검은 머리는 설정에 어긋나지 않는다. 목재 개머리판과 긴 금속 총열은 보이지만 소품 참조의 격발 장치와 총검까지 일치하는지는 판별하기 어렵다. 낙엽과 먼 연기는 존재하나 낙엽의 이슬 광택은 약하다. 낮의 억제된 색조를 유지하며 읽을 수 있는 글자는 없다.",
        "hard_violations": [],
        "physics": "가까운 병사들은 한쪽 군화를 바닥에 딛고 다른 발을 옮기며, 무릎 굽힘과 몸통 회전이 방향 전환 보행으로 설명된다. 왼쪽 선두의 뒤로 들린 발도 반대쪽 지지발이 있어 공중 부유가 아니다. 소총은 보이는 손과 몸 앞쪽 팔로 지지되고, 주머니와 배낭은 벨트 및 끈에 연결된다. 명백하게 지지 없이 떠 있는 사람이나 물체는 없다."
       },
       {
        "label": "A",
        "direction": "중앙 병사는 등을 보인 채 왼손을 왼쪽 무리 쪽으로 펼치고 오른손 소총의 총구를 우상단으로 세운다. 왼쪽 병사들은 주로 화면 왼쪽 가장자리로, 오른쪽 병사들은 오른쪽 나무 뒤로 움직여 좌우 분산 자체는 읽힌다. 다만 두 무리가 각각 좌상단과 우상단 깊이로 빠지는 것보다는 횡방향 이동에 가깝다. 중앙 오른쪽 병사 한 명은 카메라 쪽으로 얼굴과 가슴을 보여 역동적인 뒷모습이라는 통일성이 깨진다.",
        "built_space": "인공 구조물이나 고정 설비는 없다. 중앙 왼쪽의 큰 편백과 오른쪽 가장자리의 굵은 줄기, 이끼 낀 돌, 오른쪽 쓰러진 나무가 이전 장면과 매우 유사하게 배치되어 있다. 병사들은 줄기 양옆의 실제 통과 공간을 사용한다. 그러나 중앙 통로를 크게 선 지휘자가 차지하여 두 갈래 사이가 열리는 구도가 약하고, 분기 통로도 상단 배경보다 화면 양옆에서 강조된다. 먼 연기는 나무 사이에 남아 있다.",
        "entities": "성인 남성 병사들의 군모, 황갈색 군복, 가죽 장구와 각반은 참조의 복식 계열과 맞는다. 중앙 인물은 뒷모습이어서 참조 얼굴의 동일성은 검증할 수 없으며, 일부 보이는 얼굴은 동아시아계 성인 남성으로 읽히지만 국적은 외모만으로 확인할 수 없다. 중앙 인물은 소총을 놓지 않고 손신호를 보낸다. 다만 그의 소총은 가슴 높이의 휴대 자세가 아니라 머리 위까지 올라간다. 총들의 목재·금속 재질은 보이지만 참조 소총의 세부 기구 일치는 불명확하다. 낙엽, 숲, 낮빛과 연기는 있으며 읽을 수 있는 글자는 없다.",
        "hard_violations": [],
        "physics": "중앙 지휘자는 두 발을 넓게 땅에 딛고 서 있으며, 오른손으로 세운 소총을 잡는다. 이 자세는 신호 동작으로 가능하지만 몸을 꺾어 앞발을 내딛는 순간은 아니다. 주변 병사들의 걸음은 지면에 닿은 군화와 굽힌 무릎으로 지지되고, 나무 뒤의 신체 일부는 자연스럽게 가려진다. 소총과 배낭도 손이나 끈으로 지지되어 명백한 부유나 불가능한 관절은 보이지 않는다."
       }
      ]
     },
     "ok": true
    }
   ],
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "slot_winner_match": false,
   "slot_winner": {
    "gemini-pro": "A",
    "gpt-high": "B"
   },
   "route": "cross_slot_combined"
  },
  "dual": {
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "normalized": {
    "A": 1.75,
    "B": 1.5
   },
   "adjusted": {
    "A": 1.5,
    "B": 1.25
   },
   "violations": {
    "A": [
     "[gemini-pro] 수신호를 보내는 중앙 군인의 오른손에 들린 소총 하단부가 지지대 없이 허공에서 사라집니다.",
     "[gemini-pro] 중앙 좌측에 위치한 군인의 몸과 손을 두 개의 소총 총열이 물리적으로 불가능하게 교차하며 관통하고 있습니다."
    ],
    "B": [
     "[gemini-pro] 화면 좌측 군인부터 중앙 우측 군인에 이르기까지 소총들이 분리되지 않고 하나의 비정상적으로 긴 막대 형태로 연결되어 다수의 인체를 그대로 관통합니다."
    ]
   },
   "per_model_winner": {
    "gemini-pro": "A",
    "gpt-high": "B"
   },
   "agreed": false
  },
  "totals": {
   "A": 1500,
   "B": 1250
  },
  "selected": "A",
  "ranking": [
   "A",
   "B"
  ],
  "verdicts": [
   {
    "label": "A",
    "score": 1500,
    "verdict_ko": "수신호로 방향을 지시하는 선두 군인과 양갈래로 나뉘는 병력의 배치는 프롬프트를 잘 따랐으나, 여러 소총들이 인체를 관통하거나 허공에서 잘리는 치명적인 물리적 오류가 발생했습니다.  ★위반: [gemini-pro] 수신호를 보내는 중앙 군인의 오른손에 들린 소총 하단부가 지지대 없이 허공에서 사라집니다. / [gemini-pro] 중앙 좌측에 위치한 군인의 몸과 손을 두 개의 소총 총열이 물리적으로 불가능하게 교차하며 관통하고 있습니다."
   },
   {
    "label": "B",
    "score": 1250,
    "verdict_ko": "수신호로 대열 분산을 지시하는 선두 군인의 액션이 누락되었으며, 여러 군인의 소총이 하나의 긴 장대처럼 연결되어 몸을 관통하는 심각한 구조적 오류가 있습니다.  ★위반: [gemini-pro] 화면 좌측 군인부터 중앙 우측 군인에 이르기까지 소총들이 분리되지 않고 하나의 비정상적으로 긴 막대 형태로 연결되어 다수의 인체를 그대로 관통합니다."
   }
  ],
  "refs": [
   {
    "label": "PREVIOUS SHOT STILL — a visually related earlier shot of this same place: the location's look, materials, fixed features and lighting mood are LOCKED to this photo; never copy its camera framing. The clothing, hair and overall look of 선두 일본인 군인 — who appear both in that photo and in this shot — are LOCKED to that photo. Anyone else visible in that photo is NOT in this shot: never carry their face, body or clothing onto anyone here. If this photo shows a character who cannot move (dead or unconscious), that character's exact pose, position and orientation are ALSO LOCKED — treat their body as a fixed prop of the set that only the camera moves around.",
    "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/db00517b-165b-4005-a287-f28936b59b95/images/272bd7af-9294-4490-98c6-88a40e9deacc/scene/recipe/S49sh2_sel.png",
    "asset_id": "ed3401b3-86cb-49ee-8486-3da6d4f6cb88",
    "role": "prev_still"
   },
   {
    "label": "CHARACTER REFERENCE — 선두 일본인 군인: the exact person appearing in this shot; match face, hair and build exactly.",
    "path": "<bytes:905427>",
    "asset_id": "d283786c-b709-4716-919a-04ff7bb729ce",
    "role": "character_ref"
   },
   {
    "label": "PROP REFERENCE — 근대식 소총: the exact object appearing in this shot; match its look, material and wear exactly.",
    "path": "<bytes:828568>",
    "asset_id": "7543806c-8087-4caa-8d7b-56d10864fd3c",
    "role": "prop_ref"
   }
  ],
  "critique_skipped": true,
  "shot_run_produced": false,
  "shot_run_uid": "06aa1435-5823-7f24-b44c-ae9d3dc81285",
  "ref_mode": "prev+엔티티",
  "share_plan": {
   "ref_plan": "prev",
   "prev_anchor_tag": "S49sh2"
  }
 },
 "S49sh7::cine": {
  "applied": true,
  "attempted_at": "2026-09-09T06:39:31.541853+00:00",
  "fingerprint": "df62da0897b40169d5c5a0882983c442dfff2fe21cc66eddb4a525eee49ed044",
  "fingerprint_version": 2,
  "provider": "mai",
  "endpoint": "openrouter/chat-completions",
  "model": "microsoft/mai-image-2.6",
  "pack": "24.202608252115",
  "source_file": "S49sh7_sel.png",
  "source_sha256": "cac59f34cc47b6373afc616831c18a1141fa1adb5dcd9c6de6d141f9f8026265",
  "file": "S49sh7_cine.png",
  "staged_sha256": "a12de31d0a340a746bd46496b5364552982ce6f3db035a1addfb4fa17ccb15c7",
  "latency_ms": 29828
 },
 "S50sh1::signage": {
  "fp": "c5e537482ed9cf74",
  "inscriptions": [],
  "cues": [],
  "dropped": []
 },
 "S50sh1::bgfirst_bg": {
  "input_fingerprint": "57224eb447281043",
  "prompt": "Create the EMPTY BACKGROUND PLATE for one film shot — NO PEOPLE, no figures, no body parts, no sketch lines, no arrows anywhere.\n\"Empty\" means no people only: KEEP the location's inherent occupants and stock that define the place — animals in an animal shelter, pen or farm, goods in a market, moored boats in a harbour — unless the shot text explicitly removes them.\nThe FIRST attached image is a thin-line storyboard sketch: use ONLY its camera angle, horizon, perspective and the placement/size of buildings and set masses — ignore the sketched people and arrows entirely. The SECOND attached image (LOCATION PHOTOGRAPH) is the real place: take its architecture, materials, signage and fixed features, and RE-PROJECT them into the sketch's camera. If the photograph's camera differs from the sketch's, the sketch's camera wins.\nHUMAN-SCALE CALIBRATION: derive every structure's true size from human-scale elements — a door ≈ 2m, a window ≈ 1–1.5m wide, one storey ≈ 2.5–3m; never inflate a small structure or shrink a large one.\n\nSHOT TEXT this background must serve (Korean): 산채 흙바닥에 삼삼오오 둥글게 둘러앉아 사기그릇의 밥을 숟가락으로 뜬 채 입가로 가져가는 낡은 옷차림의 조선인 의병들 전경.\n\nLOCATION (lock): The outdoor communal eating ground on the mountain camp's dirt clearing, near the cooking area.\n\nTIME OF DAY (lock): day.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: wide shot\n- KEY BACKGROUND ELEMENTS: Camp earth floor (The fighters sit directly on the ground in small circles); used as Visible intervals of ground separate the groups and clarify the gathering's layout; Rice bowls (Ceramic bowls contain rice being eaten) — Their open tops are visible from the oblique elevated angle; used as Provide small repeated centers of attention within the seated circles; Spoons (Held at different stages of bringing rice toward mouths) — Angles vary naturally with each fighter's eating motion; used as Supply asynchronous gestures that make the meal feel lived rather than choreographed.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Gentle daylight contrast and muted earth tones preserve an unpolished human warmth without a conspicuous lighting effect.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nRender ONE photorealistic empty location photograph, 16:9, that this shot can be staged inside later. No readable writing anywhere: surfaces that would carry writing may be present, but stage any wording out of legibility — an oblique angle, distance, shallow focus. No captions, watermarks or overlay text.",
  "effective_prompt": "Create the EMPTY BACKGROUND PLATE for one film shot — NO PEOPLE, no figures, no body parts, no sketch lines, no arrows anywhere.\n\"Empty\" means no people only: KEEP the location's inherent occupants and stock that define the place — animals in an animal shelter, pen or farm, goods in a market, moored boats in a harbour — unless the shot text explicitly removes them.\nThe FIRST attached image is a thin-line storyboard sketch: use ONLY its camera angle, horizon, perspective and the placement/size of buildings and set masses — ignore the sketched people and arrows entirely. The SECOND attached image (LOCATION PHOTOGRAPH) is the real place: take its architecture, materials, signage and fixed features, and RE-PROJECT them into the sketch's camera. If the photograph's camera differs from the sketch's, the sketch's camera wins.\nHUMAN-SCALE CALIBRATION: derive every structure's true size from human-scale elements — a door ≈ 2m, a window ≈ 1–1.5m wide, one storey ≈ 2.5–3m; never inflate a small structure or shrink a large one.\n\nSHOT TEXT this background must serve (Korean): 산채 흙바닥에 삼삼오오 둥글게 둘러앉아 사기그릇의 밥을 숟가락으로 뜬 채 입가로 가져가는 낡은 옷차림의 조선인 의병들 전경.\n\nLOCATION (lock): The outdoor communal eating ground on the mountain camp's dirt clearing, near the cooking area.\n\nTIME OF DAY (lock): day.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: wide shot\n- KEY BACKGROUND ELEMENTS: Camp earth floor (The fighters sit directly on the ground in small circles); used as Visible intervals of ground separate the groups and clarify the gathering's layout; Rice bowls (Ceramic bowls contain rice being eaten) — Their open tops are visible from the oblique elevated angle; used as Provide small repeated centers of attention within the seated circles; Spoons (Held at different stages of bringing rice toward mouths) — Angles vary naturally with each fighter's eating motion; used as Supply asynchronous gestures that make the meal feel lived rather than choreographed.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Gentle daylight contrast and muted earth tones preserve an unpolished human warmth without a conspicuous lighting effect.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nRender ONE photorealistic empty location photograph, 16:9, that this shot can be staged inside later. No readable writing anywhere: surfaces that would carry writing may be present, but stage any wording out of legibility — an oblique angle, distance, shallow focus. No captions, watermarks or overlay text.",
  "bg_path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/db00517b-165b-4005-a287-f28936b59b95/images/272bd7af-9294-4490-98c6-88a40e9deacc/scene/recipe/S50sh1__bgfirst_bg.png",
  "asset_id": "b4779050-532d-4553-bac4-0d5064852dbf",
  "input_asset_ids": [
   "2d101188-50be-410c-b655-e8647d121794",
   "87bcdd2b-607c-40d6-8d7f-f4dfbf3ed054"
  ]
 },
 "S50sh1": {
  "input_fingerprint": "24b68ceec9da84c3",
  "prompt": "Create ONE FINAL photorealistic live-action film still of the moment below — early 20th-century Korea, Jeolla region; all people are Korean unless explicitly stated otherwise. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 산채 흙바닥에 삼삼오오 둥글게 둘러앉아 사기그릇의 밥을 숟가락으로 뜬 채 입가로 가져가는 낡은 옷차림의 조선인 의병들 전경.\n\nLOCATION (lock): The outdoor communal eating ground on the mountain camp's dirt clearing, near the cooking area. The shot takes place here — the attached LOCATION PHOTOGRAPH shows the exact spot.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: wide shot\n- KEY BACKGROUND ELEMENTS: Camp earth floor (The fighters sit directly on the ground in small circles); used as Visible intervals of ground separate the groups and clarify the gathering's layout; Rice bowls (Ceramic bowls contain rice being eaten) — Their open tops are visible from the oblique elevated angle; used as Provide small repeated centers of attention within the seated circles; Spoons (Held at different stages of bringing rice toward mouths) — Angles vary naturally with each fighter's eating motion; used as Supply asynchronous gestures that make the meal feel lived rather than choreographed.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Gentle daylight contrast and muted earth tones preserve an unpolished human warmth without a conspicuous lighting effect.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): Bowls contain white rice from the delivered provisions; the cooking pot remains upright, and ammunition bundles are hidden beneath stones near trees.\n\nTHE HAND THAT IS DOING THIS: the object at the centre of this shot is being held, operated, read, aimed or handed over by 낡은 옷차림의 조선인 의병들 right now, so 낡은 옷차림의 조선인 의병들's hand — and as much of the wrist and forearm as the framing reaches — is in the frame, gripping or resting on that object exactly the way its real use requires. Match that hand to 낡은 옷차림의 조선인 의병들: its size, build, skin, age, grooming, sleeve and anything worn on it belong to that person and to no one else.\n\nNo people appear unless the shot text itself says so.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.",
  "roll_prompts": {
   "A": "Stage the shot. The FIRST attached image (SHOT BACKGROUND) is the finished empty background of this shot — keep it EXACTLY: its camera, perspective, architecture, lighting and every fixed feature stay untouched. The SECOND attached image (LAYOUT SKETCH) tells you ONLY where the people go: each sketched person's position, screen size, pose and the gaze/motion arrows. Ignore the sketch's background lines. The CHARACTER REFERENCE photographs show the real people.\nPlace the real people into the background at exactly the sketched positions, sizes and poses, following the arrow directions. No sketch lines or arrows may remain.\n\nCreate ONE FINAL photorealistic live-action film still of the moment below — early 20th-century Korea, Jeolla region; all people are Korean unless explicitly stated otherwise. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 산채 흙바닥에 삼삼오오 둥글게 둘러앉아 사기그릇의 밥을 숟가락으로 뜬 채 입가로 가져가는 낡은 옷차림의 조선인 의병들 전경.\n\nLOCATION (lock): The outdoor communal eating ground on the mountain camp's dirt clearing, near the cooking area. The shot takes place here — the FIRST attached image (SHOT BACKGROUND) is this exact place, already built: its ground, structures, horizon, materials and lighting are the finished truth of this location and must not be redesigned or replaced. No location photograph is attached — read the place from that image alone, and add no scenery, structure, vehicle or fixture that it does not already show. This lock governs the place only; the figures in the shot follow the staging and pose instructions.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: wide shot\n- KEY BACKGROUND ELEMENTS: Camp earth floor (The fighters sit directly on the ground in small circles); used as Visible intervals of ground separate the groups and clarify the gathering's layout; Rice bowls (Ceramic bowls contain rice being eaten) — Their open tops are visible from the oblique elevated angle; used as Provide small repeated centers of attention within the seated circles; Spoons (Held at different stages of bringing rice toward mouths) — Angles vary naturally with each fighter's eating motion; used as Supply asynchronous gestures that make the meal feel lived rather than choreographed.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Gentle daylight contrast and muted earth tones preserve an unpolished human warmth without a conspicuous lighting effect.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): Bowls contain white rice from the delivered provisions; the cooking pot remains upright, and ammunition bundles are hidden beneath stones near trees.\n\nTHE HAND THAT IS DOING THIS: the object at the centre of this shot is being held, operated, read, aimed or handed over by 낡은 옷차림의 조선인 의병들 right now, so 낡은 옷차림의 조선인 의병들's hand — and as much of the wrist and forearm as the framing reaches — is in the frame, gripping or resting on that object exactly the way its real use requires. Match that hand to 낡은 옷차림의 조선인 의병들: its size, build, skin, age, grooming, sleeve and anything worn on it belong to that person and to no one else.\n\nNo people appear unless the shot text itself says so.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.",
   "B": "Create ONE FINAL photorealistic live-action film still of the moment below — early 20th-century Korea, Jeolla region; all people are Korean unless explicitly stated otherwise. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 산채 흙바닥에 삼삼오오 둥글게 둘러앉아 사기그릇의 밥을 숟가락으로 뜬 채 입가로 가져가는 낡은 옷차림의 조선인 의병들 전경.\n\nLOCATION (lock): The outdoor communal eating ground on the mountain camp's dirt clearing, near the cooking area. The shot takes place here — the attached LOCATION PHOTOGRAPH shows the exact spot.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: wide shot\n- KEY BACKGROUND ELEMENTS: Camp earth floor (The fighters sit directly on the ground in small circles); used as Visible intervals of ground separate the groups and clarify the gathering's layout; Rice bowls (Ceramic bowls contain rice being eaten) — Their open tops are visible from the oblique elevated angle; used as Provide small repeated centers of attention within the seated circles; Spoons (Held at different stages of bringing rice toward mouths) — Angles vary naturally with each fighter's eating motion; used as Supply asynchronous gestures that make the meal feel lived rather than choreographed.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Gentle daylight contrast and muted earth tones preserve an unpolished human warmth without a conspicuous lighting effect.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): Bowls contain white rice from the delivered provisions; the cooking pot remains upright, and ammunition bundles are hidden beneath stones near trees.\n\nTHE HAND THAT IS DOING THIS: the object at the centre of this shot is being held, operated, read, aimed or handed over by 낡은 옷차림의 조선인 의병들 right now, so 낡은 옷차림의 조선인 의병들's hand — and as much of the wrist and forearm as the framing reaches — is in the frame, gripping or resting on that object exactly the way its real use requires. Match that hand to 낡은 옷차림의 조선인 의병들: its size, build, skin, age, grooming, sleeve and anything worn on it belong to that person and to no one else.\n\nNo people appear unless the shot text itself says so.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.\n\nCOMPOSITION VARIATION (this candidate only): keep the place, the people and the moment exactly as specified, and keep from the CAMERA & FRAME contract its subject scale, screen placement and key background elements exactly. For this candidate only, its camera angle and height are deliberately left open: choose a camera position distinctly different from the obvious one, as a film director picking a second setup on the same blocking. Stay true to any explicit framing the SHOT TEXT itself dictates."
  },
  "roll_refs": {
   "A": [
    {
     "label": "SHOT BACKGROUND",
     "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/db00517b-165b-4005-a287-f28936b59b95/images/272bd7af-9294-4490-98c6-88a40e9deacc/scene/recipe/S50sh1__bgfirst_bg.png",
     "asset_id": "b4779050-532d-4553-bac4-0d5064852dbf",
     "role": "bgfirst_bg"
    },
    {
     "label": "LAYOUT SKETCH (people placement only)",
     "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/db00517b-165b-4005-a287-f28936b59b95/images/272bd7af-9294-4490-98c6-88a40e9deacc/conti/conti_S50sh1.png",
     "asset_id": "2d101188-50be-410c-b655-e8647d121794",
     "role": "conti_light"
    }
   ],
   "B": [
    {
     "label": "LOCATION PHOTOGRAPH — the exact place of this shot: its architecture, materials, fixed features and lighting mood are spatial truth; stage the moment inside this place. Never copy its camera framing.",
     "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/db00517b-165b-4005-a287-f28936b59b95/images/272bd7af-9294-4490-98c6-88a40e9deacc/scene/recipe/groupbg_mountain_camp_hearth_81c95f.png",
     "asset_id": "87bcdd2b-607c-40d6-8d7f-f4dfbf3ed054",
     "role": "bgfirst_group_bg"
    }
   ]
  },
  "shot_run_spend_attempt_count": 1,
  "judge_flip_skipped": "judge_fn_owns_order",
  "readings": [
   {
    "label": "A",
    "direction": "의병들의 시선과 숟가락의 방향이 각자의 밥그릇이나 입을 향하고 있습니다. 전경 좌측 및 중앙의 인물들이 숟가락으로 밥을 떠서 입으로 가져가는 동작이 명확히 보입니다.",
    "built_space": "배경의 건축물이 레퍼런스 사진과 정확히 일치합니다(좌측의 한쪽으로 기운 지붕, 중앙의 평평한 차양 구조물, 우측의 초가집). 진흙 화덕과 가마솥의 위치도 적절하며, 의병들이 흙바닥에 삼삼오오 무리 지어 앉아 있는 배치가 공간과 자연스럽게 어울립니다.",
    "entities": "낡은 옷차림의 조선인 의병들, 흰 쌀밥이 담긴 사기그릇, 숟가락이 모두 묘사되었습니다. 프롬프트의 지시대로 인원들이 소규모(삼삼오오)로 여러 그룹을 지어 앉아 있습니다.",
    "hard_violations": [
     "[gemini-pro] 물리적으로 불가능한 해부학: 중경 좌측 그룹에서 오른쪽을 보고 앉은 남자는 무릎에 얹은 왼팔 외에 가슴 부위에서 나와 밥그릇을 든 또 다른 팔이 있어 총 세 개의 팔을 가지고 있습니다.",
     "[gemini-pro] 명시적 상태 지시 위반(Carried State): '가마솥이 똑바로 놓여 있어야 한다(the cooking pot remains upright)'는 텍스트 지시가 있었으나, 레퍼런스 사진을 그대로 모방하여 가마솥이 옆으로 비스듬히 눕혀져 있습니다."
    ],
    "physics": "대부분의 인물이 바닥에 체중을 싣고 자연스럽게 앉아 그릇과 숟가락을 들고 있으나, 언급된 바와 같이 일부 손과 팔의 해부학적 구조가 붕괴되어 물리적으로 불가능한 자세가 포함되어 있습니다."
   },
   {
    "label": "B",
    "direction": "의병들의 시선이 각자의 그릇이나 중앙에 놓인 반찬들, 혹은 서로를 향하고 있습니다. 숟가락은 입이나 밥그릇을 향해 있습니다.",
    "built_space": "배경의 건축물이 레퍼런스와 완전히 다릅니다. 좌측의 단층 창고 구조물이 A자형 지붕의 집으로 바뀌었고, 중앙의 구조물 역시 A자형 텐트로 변형되어 로케이션의 공간적 진실성을 훼손했습니다.",
    "entities": "낡은 옷차림의 조선인 의병들, 사기그릇, 숟가락이 묘사되었습니다. 그러나 '삼삼오오(small circles)'로 앉으라는 지시와 달리 전경에 하나의 거대한 원형으로 둘러앉아 있으며, 프롬프트에 없는 중앙 반찬들이 임의로 추가되었습니다.",
    "hard_violations": [
     "[gemini-pro] 발명된 오브제 및 장소 위반: 좌측 및 중앙의 건축물이 레퍼런스에 없는 전혀 다른 형태(A자형 지붕)로 창조되었습니다.",
     "[gemini-pro] 명시적 상태 지시 위반(Carried State): A 후보와 마찬가지로 가마솥을 똑바로 세우라는 지시를 어기고 옆으로 눕혀 놓았습니다.",
     "[gemini-pro] 물리적으로 불가능한 해부학: 전경 좌측 남성의 손이 그릇과 완전히 융합되어 있으며, 우측 남성의 숟가락을 쥔 손가락 구조가 심각하게 뭉개져 있습니다."
    ],
    "physics": "인물들이 흙바닥에 앉아 있으나, 다수의 손과 그릇, 숟가락이 점토처럼 하나로 융합되어 허공에 떠 있거나 제대로 지지되지 않는 물리적 오류가 발생했습니다."
   }
  ],
  "cross_model_order": {
   "policy": "cross_model_order_v2_slot0_fwd_slot1_rev_parallel2",
   "slots": [
    {
     "model": "gemini-pro",
     "order": "forward",
     "display_to_canonical": {
      "A": "A",
      "B": "B"
     },
     "raw": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 4,
        "verdict_ko": "레퍼런스의 장소(건축물)와 '삼삼오오' 앉은 배치를 훌륭하게 구현했으나, 가마솥을 똑바로 세우라는 지시를 누락하고 인물에 여분의 팔이 생성되는 치명적인 해부학적 오류가 있습니다."
       },
       {
        "label": "B",
        "score": 2,
        "verdict_ko": "장소의 건축물을 완전히 다른 형태로 변형하여 로케이션 조건을 위반했으며, '삼삼오오' 배치를 무시하고 대형 원형으로 인물들을 배치한 데다 가마솥 지시 누락 및 손가락 융합 등 다수의 오류가 있습니다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "의병들의 시선과 숟가락의 방향이 각자의 밥그릇이나 입을 향하고 있습니다. 전경 좌측 및 중앙의 인물들이 숟가락으로 밥을 떠서 입으로 가져가는 동작이 명확히 보입니다.",
        "built_space": "배경의 건축물이 레퍼런스 사진과 정확히 일치합니다(좌측의 한쪽으로 기운 지붕, 중앙의 평평한 차양 구조물, 우측의 초가집). 진흙 화덕과 가마솥의 위치도 적절하며, 의병들이 흙바닥에 삼삼오오 무리 지어 앉아 있는 배치가 공간과 자연스럽게 어울립니다.",
        "entities": "낡은 옷차림의 조선인 의병들, 흰 쌀밥이 담긴 사기그릇, 숟가락이 모두 묘사되었습니다. 프롬프트의 지시대로 인원들이 소규모(삼삼오오)로 여러 그룹을 지어 앉아 있습니다.",
        "hard_violations": [
         "물리적으로 불가능한 해부학: 중경 좌측 그룹에서 오른쪽을 보고 앉은 남자는 무릎에 얹은 왼팔 외에 가슴 부위에서 나와 밥그릇을 든 또 다른 팔이 있어 총 세 개의 팔을 가지고 있습니다.",
         "명시적 상태 지시 위반(Carried State): '가마솥이 똑바로 놓여 있어야 한다(the cooking pot remains upright)'는 텍스트 지시가 있었으나, 레퍼런스 사진을 그대로 모방하여 가마솥이 옆으로 비스듬히 눕혀져 있습니다."
        ],
        "physics": "대부분의 인물이 바닥에 체중을 싣고 자연스럽게 앉아 그릇과 숟가락을 들고 있으나, 언급된 바와 같이 일부 손과 팔의 해부학적 구조가 붕괴되어 물리적으로 불가능한 자세가 포함되어 있습니다."
       },
       {
        "label": "B",
        "direction": "의병들의 시선이 각자의 그릇이나 중앙에 놓인 반찬들, 혹은 서로를 향하고 있습니다. 숟가락은 입이나 밥그릇을 향해 있습니다.",
        "built_space": "배경의 건축물이 레퍼런스와 완전히 다릅니다. 좌측의 단층 창고 구조물이 A자형 지붕의 집으로 바뀌었고, 중앙의 구조물 역시 A자형 텐트로 변형되어 로케이션의 공간적 진실성을 훼손했습니다.",
        "entities": "낡은 옷차림의 조선인 의병들, 사기그릇, 숟가락이 묘사되었습니다. 그러나 '삼삼오오(small circles)'로 앉으라는 지시와 달리 전경에 하나의 거대한 원형으로 둘러앉아 있으며, 프롬프트에 없는 중앙 반찬들이 임의로 추가되었습니다.",
        "hard_violations": [
         "발명된 오브제 및 장소 위반: 좌측 및 중앙의 건축물이 레퍼런스에 없는 전혀 다른 형태(A자형 지붕)로 창조되었습니다.",
         "명시적 상태 지시 위반(Carried State): A 후보와 마찬가지로 가마솥을 똑바로 세우라는 지시를 어기고 옆으로 눕혀 놓았습니다.",
         "물리적으로 불가능한 해부학: 전경 좌측 남성의 손이 그릇과 완전히 융합되어 있으며, 우측 남성의 숟가락을 쥔 손가락 구조가 심각하게 뭉개져 있습니다."
        ],
        "physics": "인물들이 흙바닥에 앉아 있으나, 다수의 손과 그릇, 숟가락이 점토처럼 하나로 융합되어 허공에 떠 있거나 제대로 지지되지 않는 물리적 오류가 발생했습니다."
       }
      ],
      "all_candidates_fail": true
     },
     "normalized": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 4,
        "verdict_ko": "레퍼런스의 장소(건축물)와 '삼삼오오' 앉은 배치를 훌륭하게 구현했으나, 가마솥을 똑바로 세우라는 지시를 누락하고 인물에 여분의 팔이 생성되는 치명적인 해부학적 오류가 있습니다."
       },
       {
        "label": "B",
        "score": 2,
        "verdict_ko": "장소의 건축물을 완전히 다른 형태로 변형하여 로케이션 조건을 위반했으며, '삼삼오오' 배치를 무시하고 대형 원형으로 인물들을 배치한 데다 가마솥 지시 누락 및 손가락 융합 등 다수의 오류가 있습니다."
       }
      ],
      "all_candidates_fail": true,
      "readings": [
       {
        "label": "A",
        "direction": "의병들의 시선과 숟가락의 방향이 각자의 밥그릇이나 입을 향하고 있습니다. 전경 좌측 및 중앙의 인물들이 숟가락으로 밥을 떠서 입으로 가져가는 동작이 명확히 보입니다.",
        "built_space": "배경의 건축물이 레퍼런스 사진과 정확히 일치합니다(좌측의 한쪽으로 기운 지붕, 중앙의 평평한 차양 구조물, 우측의 초가집). 진흙 화덕과 가마솥의 위치도 적절하며, 의병들이 흙바닥에 삼삼오오 무리 지어 앉아 있는 배치가 공간과 자연스럽게 어울립니다.",
        "entities": "낡은 옷차림의 조선인 의병들, 흰 쌀밥이 담긴 사기그릇, 숟가락이 모두 묘사되었습니다. 프롬프트의 지시대로 인원들이 소규모(삼삼오오)로 여러 그룹을 지어 앉아 있습니다.",
        "hard_violations": [
         "물리적으로 불가능한 해부학: 중경 좌측 그룹에서 오른쪽을 보고 앉은 남자는 무릎에 얹은 왼팔 외에 가슴 부위에서 나와 밥그릇을 든 또 다른 팔이 있어 총 세 개의 팔을 가지고 있습니다.",
         "명시적 상태 지시 위반(Carried State): '가마솥이 똑바로 놓여 있어야 한다(the cooking pot remains upright)'는 텍스트 지시가 있었으나, 레퍼런스 사진을 그대로 모방하여 가마솥이 옆으로 비스듬히 눕혀져 있습니다."
        ],
        "physics": "대부분의 인물이 바닥에 체중을 싣고 자연스럽게 앉아 그릇과 숟가락을 들고 있으나, 언급된 바와 같이 일부 손과 팔의 해부학적 구조가 붕괴되어 물리적으로 불가능한 자세가 포함되어 있습니다."
       },
       {
        "label": "B",
        "direction": "의병들의 시선이 각자의 그릇이나 중앙에 놓인 반찬들, 혹은 서로를 향하고 있습니다. 숟가락은 입이나 밥그릇을 향해 있습니다.",
        "built_space": "배경의 건축물이 레퍼런스와 완전히 다릅니다. 좌측의 단층 창고 구조물이 A자형 지붕의 집으로 바뀌었고, 중앙의 구조물 역시 A자형 텐트로 변형되어 로케이션의 공간적 진실성을 훼손했습니다.",
        "entities": "낡은 옷차림의 조선인 의병들, 사기그릇, 숟가락이 묘사되었습니다. 그러나 '삼삼오오(small circles)'로 앉으라는 지시와 달리 전경에 하나의 거대한 원형으로 둘러앉아 있으며, 프롬프트에 없는 중앙 반찬들이 임의로 추가되었습니다.",
        "hard_violations": [
         "발명된 오브제 및 장소 위반: 좌측 및 중앙의 건축물이 레퍼런스에 없는 전혀 다른 형태(A자형 지붕)로 창조되었습니다.",
         "명시적 상태 지시 위반(Carried State): A 후보와 마찬가지로 가마솥을 똑바로 세우라는 지시를 어기고 옆으로 눕혀 놓았습니다.",
         "물리적으로 불가능한 해부학: 전경 좌측 남성의 손이 그릇과 완전히 융합되어 있으며, 우측 남성의 숟가락을 쥔 손가락 구조가 심각하게 뭉개져 있습니다."
        ],
        "physics": "인물들이 흙바닥에 앉아 있으나, 다수의 손과 그릇, 숟가락이 점토처럼 하나로 융합되어 허공에 떠 있거나 제대로 지지되지 않는 물리적 오류가 발생했습니다."
       }
      ]
     },
     "ok": true
    },
    {
     "model": "gpt-high",
     "order": "reverse",
     "display_to_canonical": {
      "A": "B",
      "B": "A"
     },
     "raw": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 7,
        "verdict_ko": "밥을 입으로 가져가는 동작은 선명하지만, 큰 전경 무리가 화면을 차지하고 하단에서 잘려 작은 원형 무리들과 흙바닥 간격을 보여주는 와이드 구도가 약하다."
       },
       {
        "label": "B",
        "score": 9,
        "verdict_ko": "흙바닥에 떨어져 앉은 여러 원형 무리와 서로 다른 식사 동작을 와이드 숏으로 담아, 산채 공동 식사터의 공간 배치와 지정 순간을 더 충실하게 구현한다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "전경 중앙의 남자는 흰밥을 뜬 숟가락을 자기 입 바로 아래로 가져가며 내려다본다. 오른쪽 남자들은 자기 그릇에서 밥을 뜨고, 왼쪽 남자들도 그릇과 입 사이로 숟가락을 움직인다. 무리의 몸 방향은 대체로 원 안쪽을 향한다. 카메라를 응시하거나 숟가락이 엉뚱한 상대를 향하는 뚜렷한 사례는 없다.",
        "built_space": "왼쪽 초가 한 채, 중앙의 작은 취사 움막 한 채, 오른쪽 긴 초가 한 채와 뒤편 목책이 보인다. 중앙 취사 구역에는 큰 개방형 화덕 하나와 덮개가 놓인 작은 화덕 하나, 뒤쪽 낮은 원형 시설들이 있으며, 오른쪽에는 큰 철제 용기가 돌에 기대 있다. 참고 장소의 주요 재료와 시설 구성이 유지된다. 전경·왼쪽 중경·중앙 후경·오른쪽 가장자리에 식사 무리가 있지만, 전경 인물들이 크고 하단에서 잘려 무리 사이의 빈 땅보다 한 무리의 식사가 강조된다.",
        "entities": "화면의 인물들은 한국인 의병 설정에 부합하는 외양의 성인·중장년 남성들로, 상투나 두건과 낡은 회청색·갈색 전통 옷을 착용한다. 손에 든 사기그릇의 흰밥과 숟가락이 뚜렷하다. 전경 중앙에는 여러 반찬그릇도 놓여 있다. 무기나 탄약 꾸러미는 노출되지 않으며, 숨겨진 탄약의 위치는 확인할 수 없다. 읽을 수 있는 글자는 없다.",
        "hard_violations": [],
        "physics": "인물들은 엉덩이와 접은 다리를 흙바닥에 대고 앉아 있다. 밥그릇은 손바닥으로 받치거나 땅에 놓였고, 숟가락은 손으로 잡고 있다. 전경에서 일부 신체가 잘리지만 보이는 자세의 지지는 자연스럽다. 큰 철제 용기는 아래쪽 지면과 돌에 기대어 있으며 공중에 떠 있지 않다."
       },
       {
        "label": "B",
        "direction": "앞쪽 세 무리에서 남자들이 자기 밥그릇을 내려다보거나 밥을 뜬 숟가락을 자기 입으로 가져간다. 중앙 앞무리의 왼쪽 남자는 숟가락을 입 가까이 올리고, 중앙 남자는 그릇 위에서 숟가락을 들어 올린 상태다. 다른 사람들은 밥을 뜨거나 먹는 서로 다른 단계에 있다. 각 무리는 몸을 안쪽으로 돌려 둥글게 마주 앉는다.",
        "built_space": "왼쪽에 일부 잘린 초가 한 채, 중앙 취사 움막 한 채, 오른쪽 긴 초가 한 채가 있고 뒤로 목책과 산림이 이어진다. 중앙에는 큰 개방형 화덕 하나, 그 오른쪽 덮개가 놓인 작은 화덕 하나와 낮은 흙 구조물 하나가 보이며, 뒤쪽에도 낮은 원형 시설 두 개가 있다. 큰 철제 용기 하나는 취사 구역 오른쪽 돌 곁에 기대 있다. 앞쪽 세 무리와 뒤쪽 다섯 무리 사이에 흙바닥 간격이 보인다. 비스듬히 내려다보는 와이드 구도에서 전경 무리의 둘레와 그릇 윗면이 함께 드러나며, 참고 장소의 배치도 잘 유지된다.",
        "entities": "인물들은 한국인 의병 설정에 부합하는 외양의 성인·중장년 남성들이며, 낡고 먼지 묻은 베이지색·회갈색 전통 옷과 두건 또는 상투를 하고 있다. 사기그릇에 담긴 흰밥과 손에 든 숟가락이 전경에서 확인된다. 뒤쪽 인물들의 작은 크기는 거리와 와이드 숏에 맞는다. 탄약 꾸러미는 노출되지 않으며, 은닉 상태 자체는 확인할 수 없다. 읽을 수 있는 글자나 화면 표식은 없다.",
        "hard_violations": [],
        "physics": "남자들은 모두 흙바닥에 엉덩이와 접은 다리를 대고 앉아 체중을 지탱한다. 그릇은 손으로 받치거나 지면에 놓여 있고, 숟가락을 든 손과 팔은 식사 동작으로 이어진다. 화덕 위의 덮개는 화덕에 받쳐져 있으며, 큰 철제 용기는 지면과 주변 돌에 기대 있다. 지지 없이 떠 있는 사람이나 물건은 보이지 않는다."
       }
      ],
      "all_candidates_fail": false
     },
     "normalized": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 7,
        "verdict_ko": "밥을 입으로 가져가는 동작은 선명하지만, 큰 전경 무리가 화면을 차지하고 하단에서 잘려 작은 원형 무리들과 흙바닥 간격을 보여주는 와이드 구도가 약하다."
       },
       {
        "label": "A",
        "score": 9,
        "verdict_ko": "흙바닥에 떨어져 앉은 여러 원형 무리와 서로 다른 식사 동작을 와이드 숏으로 담아, 산채 공동 식사터의 공간 배치와 지정 순간을 더 충실하게 구현한다."
       }
      ],
      "all_candidates_fail": false,
      "readings": [
       {
        "label": "B",
        "direction": "전경 중앙의 남자는 흰밥을 뜬 숟가락을 자기 입 바로 아래로 가져가며 내려다본다. 오른쪽 남자들은 자기 그릇에서 밥을 뜨고, 왼쪽 남자들도 그릇과 입 사이로 숟가락을 움직인다. 무리의 몸 방향은 대체로 원 안쪽을 향한다. 카메라를 응시하거나 숟가락이 엉뚱한 상대를 향하는 뚜렷한 사례는 없다.",
        "built_space": "왼쪽 초가 한 채, 중앙의 작은 취사 움막 한 채, 오른쪽 긴 초가 한 채와 뒤편 목책이 보인다. 중앙 취사 구역에는 큰 개방형 화덕 하나와 덮개가 놓인 작은 화덕 하나, 뒤쪽 낮은 원형 시설들이 있으며, 오른쪽에는 큰 철제 용기가 돌에 기대 있다. 참고 장소의 주요 재료와 시설 구성이 유지된다. 전경·왼쪽 중경·중앙 후경·오른쪽 가장자리에 식사 무리가 있지만, 전경 인물들이 크고 하단에서 잘려 무리 사이의 빈 땅보다 한 무리의 식사가 강조된다.",
        "entities": "화면의 인물들은 한국인 의병 설정에 부합하는 외양의 성인·중장년 남성들로, 상투나 두건과 낡은 회청색·갈색 전통 옷을 착용한다. 손에 든 사기그릇의 흰밥과 숟가락이 뚜렷하다. 전경 중앙에는 여러 반찬그릇도 놓여 있다. 무기나 탄약 꾸러미는 노출되지 않으며, 숨겨진 탄약의 위치는 확인할 수 없다. 읽을 수 있는 글자는 없다.",
        "hard_violations": [],
        "physics": "인물들은 엉덩이와 접은 다리를 흙바닥에 대고 앉아 있다. 밥그릇은 손바닥으로 받치거나 땅에 놓였고, 숟가락은 손으로 잡고 있다. 전경에서 일부 신체가 잘리지만 보이는 자세의 지지는 자연스럽다. 큰 철제 용기는 아래쪽 지면과 돌에 기대어 있으며 공중에 떠 있지 않다."
       },
       {
        "label": "A",
        "direction": "앞쪽 세 무리에서 남자들이 자기 밥그릇을 내려다보거나 밥을 뜬 숟가락을 자기 입으로 가져간다. 중앙 앞무리의 왼쪽 남자는 숟가락을 입 가까이 올리고, 중앙 남자는 그릇 위에서 숟가락을 들어 올린 상태다. 다른 사람들은 밥을 뜨거나 먹는 서로 다른 단계에 있다. 각 무리는 몸을 안쪽으로 돌려 둥글게 마주 앉는다.",
        "built_space": "왼쪽에 일부 잘린 초가 한 채, 중앙 취사 움막 한 채, 오른쪽 긴 초가 한 채가 있고 뒤로 목책과 산림이 이어진다. 중앙에는 큰 개방형 화덕 하나, 그 오른쪽 덮개가 놓인 작은 화덕 하나와 낮은 흙 구조물 하나가 보이며, 뒤쪽에도 낮은 원형 시설 두 개가 있다. 큰 철제 용기 하나는 취사 구역 오른쪽 돌 곁에 기대 있다. 앞쪽 세 무리와 뒤쪽 다섯 무리 사이에 흙바닥 간격이 보인다. 비스듬히 내려다보는 와이드 구도에서 전경 무리의 둘레와 그릇 윗면이 함께 드러나며, 참고 장소의 배치도 잘 유지된다.",
        "entities": "인물들은 한국인 의병 설정에 부합하는 외양의 성인·중장년 남성들이며, 낡고 먼지 묻은 베이지색·회갈색 전통 옷과 두건 또는 상투를 하고 있다. 사기그릇에 담긴 흰밥과 손에 든 숟가락이 전경에서 확인된다. 뒤쪽 인물들의 작은 크기는 거리와 와이드 숏에 맞는다. 탄약 꾸러미는 노출되지 않으며, 은닉 상태 자체는 확인할 수 없다. 읽을 수 있는 글자나 화면 표식은 없다.",
        "hard_violations": [],
        "physics": "남자들은 모두 흙바닥에 엉덩이와 접은 다리를 대고 앉아 체중을 지탱한다. 그릇은 손으로 받치거나 지면에 놓여 있고, 숟가락을 든 손과 팔은 식사 동작으로 이어진다. 화덕 위의 덮개는 화덕에 받쳐져 있으며, 큰 철제 용기는 지면과 주변 돌에 기대 있다. 지지 없이 떠 있는 사람이나 물건은 보이지 않는다."
       }
      ]
     },
     "ok": true
    }
   ],
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "slot_winner_match": true,
   "slot_winner": {
    "gemini-pro": "A",
    "gpt-high": "A"
   },
   "route": "cross_slot_agree"
  },
  "dual": {
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "normalized": {
    "A": 2.0,
    "B": 1.278
   },
   "adjusted": {
    "A": 1.75,
    "B": 1.028
   },
   "violations": {
    "A": [
     "[gemini-pro] 물리적으로 불가능한 해부학: 중경 좌측 그룹에서 오른쪽을 보고 앉은 남자는 무릎에 얹은 왼팔 외에 가슴 부위에서 나와 밥그릇을 든 또 다른 팔이 있어 총 세 개의 팔을 가지고 있습니다.",
     "[gemini-pro] 명시적 상태 지시 위반(Carried State): '가마솥이 똑바로 놓여 있어야 한다(the cooking pot remains upright)'는 텍스트 지시가 있었으나, 레퍼런스 사진을 그대로 모방하여 가마솥이 옆으로 비스듬히 눕혀져 있습니다."
    ],
    "B": [
     "[gemini-pro] 발명된 오브제 및 장소 위반: 좌측 및 중앙의 건축물이 레퍼런스에 없는 전혀 다른 형태(A자형 지붕)로 창조되었습니다.",
     "[gemini-pro] 명시적 상태 지시 위반(Carried State): A 후보와 마찬가지로 가마솥을 똑바로 세우라는 지시를 어기고 옆으로 눕혀 놓았습니다.",
     "[gemini-pro] 물리적으로 불가능한 해부학: 전경 좌측 남성의 손이 그릇과 완전히 융합되어 있으며, 우측 남성의 숟가락을 쥔 손가락 구조가 심각하게 뭉개져 있습니다."
    ]
   },
   "per_model_winner": {
    "gemini-pro": "A",
    "gpt-high": "A"
   },
   "agreed": true
  },
  "totals": {
   "A": 1750,
   "B": 1028
  },
  "selected": "A",
  "ranking": [
   "A",
   "B"
  ],
  "verdicts": [
   {
    "label": "A",
    "score": 1750,
    "verdict_ko": "레퍼런스의 장소(건축물)와 '삼삼오오' 앉은 배치를 훌륭하게 구현했으나, 가마솥을 똑바로 세우라는 지시를 누락하고 인물에 여분의 팔이 생성되는 치명적인 해부학적 오류가 있습니다.  ★위반: [gemini-pro] 물리적으로 불가능한 해부학: 중경 좌측 그룹에서 오른쪽을 보고 앉은 남자는 무릎에 얹은 왼팔 외에 가슴 부위에서 나와 밥그릇을 든 또 다른 팔이 있어 총 세 개의 팔을 가지고 있습니다. / [gemini-pro] 명시적 상태 지시 위반(Carried State): '가마솥이 똑바로 놓여 있어야 한다(the cooking pot remains upright)'는 텍스트 지시가 있었으나, 레퍼런스 사진을 그대로 모방하여 가마솥이 옆으로 비스듬히 눕혀져 있습니다."
   },
   {
    "label": "B",
    "score": 1028,
    "verdict_ko": "장소의 건축물을 완전히 다른 형태로 변형하여 로케이션 조건을 위반했으며, '삼삼오오' 배치를 무시하고 대형 원형으로 인물들을 배치한 데다 가마솥 지시 누락 및 손가락 융합 등 다수의 오류가 있습니다.  ★위반: [gemini-pro] 발명된 오브제 및 장소 위반: 좌측 및 중앙의 건축물이 레퍼런스에 없는 전혀 다른 형태(A자형 지붕)로 창조되었습니다. / [gemini-pro] 명시적 상태 지시 위반(Carried State): A 후보와 마찬가지로 가마솥을 똑바로 세우라는 지시를 어기고 옆으로 눕혀 놓았습니다. / [gemini-pro] 물리적으로 불가능한 해부학: 전경 좌측 남성의 손이 그릇과 완전히 융합되어 있으며, 우측 남성의 숟가락을 쥔 손가락 구조가 심각하게 뭉개져 있습니다."
   }
  ],
  "refs": [
   {
    "label": "LOCATION PHOTOGRAPH — the exact place of this shot: its architecture, materials, fixed features and lighting mood are spatial truth; stage the moment inside this place. Never copy its camera framing.",
    "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/db00517b-165b-4005-a287-f28936b59b95/images/272bd7af-9294-4490-98c6-88a40e9deacc/scene/recipe/groupbg_mountain_camp_hearth_81c95f.png",
    "asset_id": "87bcdd2b-607c-40d6-8d7f-f4dfbf3ed054",
    "role": "bgfirst_group_bg"
   }
  ],
  "critique_skipped": true,
  "shot_run_produced": false,
  "shot_run_uid": "06aa1435-59db-77ae-b7f3-a65d6272240a",
  "bgfirst": {
   "bg_path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/db00517b-165b-4005-a287-f28936b59b95/images/272bd7af-9294-4490-98c6-88a40e9deacc/scene/recipe/S50sh1__bgfirst_bg.png",
   "bg_asset_id": "b4779050-532d-4553-bac4-0d5064852dbf",
   "bg_record_key": "S50sh1::bgfirst_bg",
   "chain_winner": true,
   "authority": "groupbg",
   "group_key": "mountain_camp_hearth",
   "groupbg_asset_id": "87bcdd2b-607c-40d6-8d7f-f4dfbf3ed054"
  },
  "ref_mode": "재투영 배경+콘티+엔티티 (2택1: 체인 승)",
  "share_plan": {
   "ref_plan": "background"
  }
 },
 "S50sh1::cine": {
  "applied": true,
  "attempted_at": "2026-09-09T06:43:34.953207+00:00",
  "fingerprint": "f8a00b5f7c33f5549f09d4f4d1e85254c970c6ddda947305b3c29c7fb4827cd0",
  "fingerprint_version": 2,
  "provider": "mai",
  "endpoint": "openrouter/chat-completions",
  "model": "microsoft/mai-image-2.6",
  "pack": "24.202608252115",
  "source_file": "S50sh1_sel.png",
  "source_sha256": "4f71ea2064aa1dca117b9e5dc0c7bdfdef1d17d4cd018d495ecc4bda65e6f1f5",
  "file": "S50sh1_cine.png",
  "staged_sha256": "7efae5ed99dd41bc8630bca0cdce5145a8ba183c91299d9e029c5fa022a141e4",
  "latency_ms": 27745
 },
 "S50sh4::signage": {
  "fp": "ed5f08ba92b2c1c1",
  "inscriptions": [],
  "cues": [],
  "dropped": []
 },
 "S50sh4": {
  "input_fingerprint": "0c363002e117d46f",
  "prompt": "Create ONE FINAL photorealistic live-action film still of the moment below — early 20th-century Korea, Jeolla region; all people are Korean unless explicitly stated otherwise. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 밥을 입안 가득 머금어 빵빵해진 볼로 만족스러운 미소를 짓고 있는 염재보의 클로즈업.\n\nLOCATION (lock): A ground-level place in the mountain camp's outdoor meal circle during the daytime rice meal. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: close-up\n- KEY BACKGROUND ELEMENTS: Camp ground (Visible beyond the seated figure at the edge of the meal gathering); used as A soft peripheral ground plane preserves location without introducing another face into the close framing.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Maintain the meal's gentle daylight and restrained saturation, allowing the smile to supply warmth rather than changing the light.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The people visible in that photo are NOT in this shot: never carry their faces, bodies or clothing onto anyone here. Each person in THIS shot is defined solely by the PEOPLE list, their own CHARACTER REFERENCE images and their wardrobe notes. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nPREVIOUS STILL USAGE (follow exactly — what to take from the attached still and what to exclude; it governs place and objects only, never who is in this shot or how the camera sees them): Take the dirt ground, ceramic bowls, rice, and daylight of the eating area. Exclude supply bundles occupying the meal space and do not duplicate individual servings.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): Rice bowls and the upright cooking pot remain in use; cloth-wrapped ammunition is already concealed beneath stones near trees. 염재보: He sits with rice in his mouth and retains his matchlock in close hand control.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 염재보 (한국인 남성, 성인 나이대, 자연스러운 검은 머리) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.",
  "roll_prompts": {
   "A": "Create ONE FINAL photorealistic live-action film still of the moment below — early 20th-century Korea, Jeolla region; all people are Korean unless explicitly stated otherwise. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 밥을 입안 가득 머금어 빵빵해진 볼로 만족스러운 미소를 짓고 있는 염재보의 클로즈업.\n\nLOCATION (lock): A ground-level place in the mountain camp's outdoor meal circle during the daytime rice meal. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: close-up\n- KEY BACKGROUND ELEMENTS: Camp ground (Visible beyond the seated figure at the edge of the meal gathering); used as A soft peripheral ground plane preserves location without introducing another face into the close framing.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Maintain the meal's gentle daylight and restrained saturation, allowing the smile to supply warmth rather than changing the light.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The people visible in that photo are NOT in this shot: never carry their faces, bodies or clothing onto anyone here. Each person in THIS shot is defined solely by the PEOPLE list, their own CHARACTER REFERENCE images and their wardrobe notes. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nPREVIOUS STILL USAGE (follow exactly — what to take from the attached still and what to exclude; it governs place and objects only, never who is in this shot or how the camera sees them): Take the dirt ground, ceramic bowls, rice, and daylight of the eating area. Exclude supply bundles occupying the meal space and do not duplicate individual servings.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): Rice bowls and the upright cooking pot remain in use; cloth-wrapped ammunition is already concealed beneath stones near trees. 염재보: He sits with rice in his mouth and retains his matchlock in close hand control.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 염재보 (한국인 남성, 성인 나이대, 자연스러운 검은 머리) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.",
   "B": "Create ONE FINAL photorealistic live-action film still of the moment below — early 20th-century Korea, Jeolla region; all people are Korean unless explicitly stated otherwise. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 밥을 입안 가득 머금어 빵빵해진 볼로 만족스러운 미소를 짓고 있는 염재보의 클로즈업.\n\nLOCATION (lock): A ground-level place in the mountain camp's outdoor meal circle during the daytime rice meal. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: close-up\n- KEY BACKGROUND ELEMENTS: Camp ground (Visible beyond the seated figure at the edge of the meal gathering); used as A soft peripheral ground plane preserves location without introducing another face into the close framing.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Maintain the meal's gentle daylight and restrained saturation, allowing the smile to supply warmth rather than changing the light.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The people visible in that photo are NOT in this shot: never carry their faces, bodies or clothing onto anyone here. Each person in THIS shot is defined solely by the PEOPLE list, their own CHARACTER REFERENCE images and their wardrobe notes. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nPREVIOUS STILL USAGE (follow exactly — what to take from the attached still and what to exclude; it governs place and objects only, never who is in this shot or how the camera sees them): Take the dirt ground, ceramic bowls, rice, and daylight of the eating area. Exclude supply bundles occupying the meal space and do not duplicate individual servings.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): Rice bowls and the upright cooking pot remain in use; cloth-wrapped ammunition is already concealed beneath stones near trees. 염재보: He sits with rice in his mouth and retains his matchlock in close hand control.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 염재보 (한국인 남성, 성인 나이대, 자연스러운 검은 머리) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.\n\nCOMPOSITION VARIATION (this candidate only): keep the place, the people and the moment exactly as specified, and keep from the CAMERA & FRAME contract its subject scale, screen placement and key background elements exactly. For this candidate only, its camera angle and height are deliberately left open: choose a camera position distinctly different from the obvious one, as a film director picking a second setup on the same blocking. Stay true to any explicit framing the SHOT TEXT itself dictates."
  },
  "shot_run_spend_attempt_count": 1,
  "initial_roll_all_fail": true,
  "readings": [
   {
    "label": "A",
    "direction": "주인공의 시선은 카메라 왼쪽 바깥을 부드럽게 향하고 있습니다.",
    "built_space": "배경은 흐릿한 흙바닥으로 처리되어 레퍼런스의 장소적 특성만 부드럽게 남겼습니다.",
    "entities": "염재보의 얼굴과 의상이 레퍼런스와 정확히 일치하며, 밥을 가득 머금어 빵빵해진 볼로 미소 짓는 표정이 완벽합니다. 하지만 프롬프트의 지시와 달리 화면 좌우 가장자리와 배경에 다른 인물들의 신체 일부와 형체가 포함되어 있습니다.",
    "hard_violations": [
     "[gemini-pro] 명시되지 않은 인물 및 신체 일부 추가 (화면 양끝 및 배경)",
     "[gpt-high] 염재보만 허용된 화면에 왼쪽·오른쪽 가장자리와 뒤편의 다른 사람 신체를 추가했다."
    ],
    "physics": "앉아 있는 자세가 자연스럽습니다. 화승총은 타이트한 클로즈업 프레이밍으로 인해 보이지 않게 배제되었으며, 물리적으로 어색하거나 지지되지 않은 사물은 없습니다."
   },
   {
    "label": "B",
    "direction": "주인공의 시선은 카메라 왼쪽 바깥을 향하고 있습니다.",
    "built_space": "레퍼런스에 있는 산채의 오두막, 울타리, 산 배경이 명확하게 나타납니다.",
    "entities": "염재보의 외모와 표정은 지시사항과 잘 일치합니다. 그러나 배경에 레퍼런스 이미지에 있던 다수의 식사하는 사람들이 그대로 등장하여 다른 사람을 배제하라는 강력한 지시를 완전히 위반했습니다.",
    "hard_violations": [
     "[gemini-pro] 명시되지 않은 다수의 인물 추가 (배경 전체)",
     "[gpt-high] 허용된 염재보 외에 배경의 여러 식사 인물과 화면 왼쪽의 다른 사람 신체를 추가했다."
    ],
    "physics": "주인공이 양손으로 그릇과 젓가락을 들고 있습니다. 화승총은 손으로 쥐고 있지 않고 왼쪽 팔에 어색하게 기대어져 있어 '가까이 손에 쥐고 있다'는 지시를 위반했습니다."
   }
  ],
  "cross_model_order": {
   "policy": "cross_model_order_v2_slot0_fwd_slot1_rev_parallel2",
   "slots": [
    {
     "model": "gemini-pro",
     "order": "forward",
     "display_to_canonical": {
      "A": "A",
      "B": "B"
     },
     "raw": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 4,
        "verdict_ko": "클로즈업 프레이밍과 주인공이 밥을 머금고 짓는 표정 연기는 매우 훌륭하게 구현되었으나, 화면 좌우 가장자리에 프롬프트에서 금지한 다른 인물의 신체 일부가 등장하여 심각한 위반이 발생했습니다."
       },
       {
        "label": "B",
        "score": 2,
        "verdict_ko": "주인공의 표정은 요구사항을 충족하지만, 절대로 추가하지 말라고 명시된 배경의 다른 인물들이 레퍼런스에서 그대로 복사되어 다수 등장했으며 화승총을 손에 쥐고 있어야 한다는 지시도 무시되었습니다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "주인공의 시선은 카메라 왼쪽 바깥을 부드럽게 향하고 있습니다.",
        "built_space": "배경은 흐릿한 흙바닥으로 처리되어 레퍼런스의 장소적 특성만 부드럽게 남겼습니다.",
        "entities": "염재보의 얼굴과 의상이 레퍼런스와 정확히 일치하며, 밥을 가득 머금어 빵빵해진 볼로 미소 짓는 표정이 완벽합니다. 하지만 프롬프트의 지시와 달리 화면 좌우 가장자리와 배경에 다른 인물들의 신체 일부와 형체가 포함되어 있습니다.",
        "hard_violations": [
         "명시되지 않은 인물 및 신체 일부 추가 (화면 양끝 및 배경)"
        ],
        "physics": "앉아 있는 자세가 자연스럽습니다. 화승총은 타이트한 클로즈업 프레이밍으로 인해 보이지 않게 배제되었으며, 물리적으로 어색하거나 지지되지 않은 사물은 없습니다."
       },
       {
        "label": "B",
        "direction": "주인공의 시선은 카메라 왼쪽 바깥을 향하고 있습니다.",
        "built_space": "레퍼런스에 있는 산채의 오두막, 울타리, 산 배경이 명확하게 나타납니다.",
        "entities": "염재보의 외모와 표정은 지시사항과 잘 일치합니다. 그러나 배경에 레퍼런스 이미지에 있던 다수의 식사하는 사람들이 그대로 등장하여 다른 사람을 배제하라는 강력한 지시를 완전히 위반했습니다.",
        "hard_violations": [
         "명시되지 않은 다수의 인물 추가 (배경 전체)"
        ],
        "physics": "주인공이 양손으로 그릇과 젓가락을 들고 있습니다. 화승총은 손으로 쥐고 있지 않고 왼쪽 팔에 어색하게 기대어져 있어 '가까이 손에 쥐고 있다'는 지시를 위반했습니다."
       }
      ],
      "all_candidates_fail": true
     },
     "normalized": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 4,
        "verdict_ko": "클로즈업 프레이밍과 주인공이 밥을 머금고 짓는 표정 연기는 매우 훌륭하게 구현되었으나, 화면 좌우 가장자리에 프롬프트에서 금지한 다른 인물의 신체 일부가 등장하여 심각한 위반이 발생했습니다."
       },
       {
        "label": "B",
        "score": 2,
        "verdict_ko": "주인공의 표정은 요구사항을 충족하지만, 절대로 추가하지 말라고 명시된 배경의 다른 인물들이 레퍼런스에서 그대로 복사되어 다수 등장했으며 화승총을 손에 쥐고 있어야 한다는 지시도 무시되었습니다."
       }
      ],
      "all_candidates_fail": true,
      "readings": [
       {
        "label": "A",
        "direction": "주인공의 시선은 카메라 왼쪽 바깥을 부드럽게 향하고 있습니다.",
        "built_space": "배경은 흐릿한 흙바닥으로 처리되어 레퍼런스의 장소적 특성만 부드럽게 남겼습니다.",
        "entities": "염재보의 얼굴과 의상이 레퍼런스와 정확히 일치하며, 밥을 가득 머금어 빵빵해진 볼로 미소 짓는 표정이 완벽합니다. 하지만 프롬프트의 지시와 달리 화면 좌우 가장자리와 배경에 다른 인물들의 신체 일부와 형체가 포함되어 있습니다.",
        "hard_violations": [
         "명시되지 않은 인물 및 신체 일부 추가 (화면 양끝 및 배경)"
        ],
        "physics": "앉아 있는 자세가 자연스럽습니다. 화승총은 타이트한 클로즈업 프레이밍으로 인해 보이지 않게 배제되었으며, 물리적으로 어색하거나 지지되지 않은 사물은 없습니다."
       },
       {
        "label": "B",
        "direction": "주인공의 시선은 카메라 왼쪽 바깥을 향하고 있습니다.",
        "built_space": "레퍼런스에 있는 산채의 오두막, 울타리, 산 배경이 명확하게 나타납니다.",
        "entities": "염재보의 외모와 표정은 지시사항과 잘 일치합니다. 그러나 배경에 레퍼런스 이미지에 있던 다수의 식사하는 사람들이 그대로 등장하여 다른 사람을 배제하라는 강력한 지시를 완전히 위반했습니다.",
        "hard_violations": [
         "명시되지 않은 다수의 인물 추가 (배경 전체)"
        ],
        "physics": "주인공이 양손으로 그릇과 젓가락을 들고 있습니다. 화승총은 손으로 쥐고 있지 않고 왼쪽 팔에 어색하게 기대어져 있어 '가까이 손에 쥐고 있다'는 지시를 위반했습니다."
       }
      ]
     },
     "ok": true
    },
    {
     "model": "gpt-high",
     "order": "reverse",
     "display_to_canonical": {
      "A": "B",
      "B": "A"
     },
     "raw": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 2,
        "verdict_ko": "빵빵한 볼은 분명하지만, 주변 인물을 대거 추가하고 흙바닥 대신 산채 전경을 펼쳐 단독 클로즈업의 배경 지시를 어겼다."
       },
       {
        "label": "B",
        "score": 3,
        "verdict_ko": "만족스러운 미소와 흐린 흙바닥 배경은 더 충실하지만, 화면 가장자리와 뒤편의 다른 사람들 때문에 이 후보 역시 사용 불가다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "염재보는 화면 왼쪽의 프레임 밖을 바라보며, 특정 상대는 확인되지 않는다. 젓가락 끝은 밥그릇 안의 밥을 향한다. 오른쪽 장총의 총구는 화면 오른쪽 위로 향하며 사람을 겨누지 않는다.",
        "built_space": "머리와 상반신을 크게 담았지만 뒤로 산, 목책, 좌·중·우의 움막 세 동과 여러 식사 무리가 넓게 보인다. 오른쪽에는 흙으로 만든 취사 시설과 큰 그릇도 보인다. 이전 장소의 재료는 유사하지만, 요구한 부드러운 주변 흙바닥만의 배경이 아니다. 염재보 외에 왼쪽 가장자리의 잘린 몸과 뒤편 여러 인물이 들어왔다.",
        "entities": "성인 동아시아계 남성 한 명이 중심이며 검은 머리와 짙은 누비옷을 입었다. 다만 참고 인물보다 얼굴이 넓고 나이 들어 보이며, 참고의 삿갓은 없고 머리 모양과 어깨 덧댐도 다르다. 부푼 양 볼, 닫힌 입의 미소, 입가의 쌀알이 보인다. 밥이 든 도자기 그릇 하나와 젓가락, 장총이 있으며 총의 화승 기구는 확인하기 어렵다. 읽을 수 있는 글자는 없다.",
        "hard_violations": [
         "허용된 염재보 외에 배경의 여러 식사 인물과 화면 왼쪽의 다른 사람 신체를 추가했다."
        ],
        "physics": "한 손이 밥그릇을 받치고 다른 손이 젓가락을 쥐고 있다. 장총은 몸과 팔에 기대어 있으며 하단은 화면 밖이라 손으로 직접 통제하는지는 확인되지 않는다. 앉은 자세의 하체 접점은 프레임 밖이다. 볼의 팽창은 밥을 머금은 실제 연기로 가능한 수준이며, 명백히 공중에 떠 있는 물체는 없다."
       },
       {
        "label": "B",
        "direction": "염재보의 눈은 화면 왼쪽 위의 프레임 밖을 향한다. 눈을 살짝 좁히고 입꼬리를 올려 만족감을 표현한다. 화면에 총구나 손에 든 도구는 없어 그 방향은 판정할 수 없다.",
        "built_space": "약간 내려다보는 상반신 클로즈업이며 뒤편 대부분은 흐린 흙바닥이다. 건축물이나 고정 취사 시설은 이 구도에 보이지 않아 제외 자체는 문제가 아니다. 그러나 왼쪽 앞의 잘린 상체, 오른쪽의 큰 등과 하체, 뒤편에 앉은 사람들의 몸이 보인다. 배경에는 작은 그릇 하나도 식별된다. 흙바닥 중심의 장소 표현은 지시에 더 가깝다.",
        "entities": "검은 머리를 묶은 성인 동아시아계 남성이 짙은 누비옷을 입고 있다. 참고보다 성숙하고 넓은 얼굴로 보이며 삿갓이 없고 어깨 덧댐과 옷 여밈도 다르다. 양 볼이 부풀고 입가에 쌀알이 붙어 있으며 만족스러운 미소가 뚜렷하다. 본인의 밥그릇, 화승총, 취사 솥은 프레임 밖이라 보유 상태를 확인할 수 없다. 읽을 수 있는 글자는 없다.",
        "hard_violations": [
         "염재보만 허용된 화면에 왼쪽·오른쪽 가장자리와 뒤편의 다른 사람 신체를 추가했다."
        ],
        "physics": "상체는 앉아서 약간 앞으로 기울인 자세로 자연스럽고, 팔은 아래쪽 프레임 밖으로 이어진다. 엉덩이와 지면의 접점은 보이지 않지만 부유를 나타내는 모습은 없다. 배경 인물들은 흙바닥에 앉아 있고 작은 그릇은 바닥에 놓여 있다. 입가의 쌀알과 부푼 볼도 물리적으로 가능한 식사 순간이다."
       }
      ],
      "all_candidates_fail": true
     },
     "normalized": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 2,
        "verdict_ko": "빵빵한 볼은 분명하지만, 주변 인물을 대거 추가하고 흙바닥 대신 산채 전경을 펼쳐 단독 클로즈업의 배경 지시를 어겼다."
       },
       {
        "label": "A",
        "score": 3,
        "verdict_ko": "만족스러운 미소와 흐린 흙바닥 배경은 더 충실하지만, 화면 가장자리와 뒤편의 다른 사람들 때문에 이 후보 역시 사용 불가다."
       }
      ],
      "all_candidates_fail": true,
      "readings": [
       {
        "label": "B",
        "direction": "염재보는 화면 왼쪽의 프레임 밖을 바라보며, 특정 상대는 확인되지 않는다. 젓가락 끝은 밥그릇 안의 밥을 향한다. 오른쪽 장총의 총구는 화면 오른쪽 위로 향하며 사람을 겨누지 않는다.",
        "built_space": "머리와 상반신을 크게 담았지만 뒤로 산, 목책, 좌·중·우의 움막 세 동과 여러 식사 무리가 넓게 보인다. 오른쪽에는 흙으로 만든 취사 시설과 큰 그릇도 보인다. 이전 장소의 재료는 유사하지만, 요구한 부드러운 주변 흙바닥만의 배경이 아니다. 염재보 외에 왼쪽 가장자리의 잘린 몸과 뒤편 여러 인물이 들어왔다.",
        "entities": "성인 동아시아계 남성 한 명이 중심이며 검은 머리와 짙은 누비옷을 입었다. 다만 참고 인물보다 얼굴이 넓고 나이 들어 보이며, 참고의 삿갓은 없고 머리 모양과 어깨 덧댐도 다르다. 부푼 양 볼, 닫힌 입의 미소, 입가의 쌀알이 보인다. 밥이 든 도자기 그릇 하나와 젓가락, 장총이 있으며 총의 화승 기구는 확인하기 어렵다. 읽을 수 있는 글자는 없다.",
        "hard_violations": [
         "허용된 염재보 외에 배경의 여러 식사 인물과 화면 왼쪽의 다른 사람 신체를 추가했다."
        ],
        "physics": "한 손이 밥그릇을 받치고 다른 손이 젓가락을 쥐고 있다. 장총은 몸과 팔에 기대어 있으며 하단은 화면 밖이라 손으로 직접 통제하는지는 확인되지 않는다. 앉은 자세의 하체 접점은 프레임 밖이다. 볼의 팽창은 밥을 머금은 실제 연기로 가능한 수준이며, 명백히 공중에 떠 있는 물체는 없다."
       },
       {
        "label": "A",
        "direction": "염재보의 눈은 화면 왼쪽 위의 프레임 밖을 향한다. 눈을 살짝 좁히고 입꼬리를 올려 만족감을 표현한다. 화면에 총구나 손에 든 도구는 없어 그 방향은 판정할 수 없다.",
        "built_space": "약간 내려다보는 상반신 클로즈업이며 뒤편 대부분은 흐린 흙바닥이다. 건축물이나 고정 취사 시설은 이 구도에 보이지 않아 제외 자체는 문제가 아니다. 그러나 왼쪽 앞의 잘린 상체, 오른쪽의 큰 등과 하체, 뒤편에 앉은 사람들의 몸이 보인다. 배경에는 작은 그릇 하나도 식별된다. 흙바닥 중심의 장소 표현은 지시에 더 가깝다.",
        "entities": "검은 머리를 묶은 성인 동아시아계 남성이 짙은 누비옷을 입고 있다. 참고보다 성숙하고 넓은 얼굴로 보이며 삿갓이 없고 어깨 덧댐과 옷 여밈도 다르다. 양 볼이 부풀고 입가에 쌀알이 붙어 있으며 만족스러운 미소가 뚜렷하다. 본인의 밥그릇, 화승총, 취사 솥은 프레임 밖이라 보유 상태를 확인할 수 없다. 읽을 수 있는 글자는 없다.",
        "hard_violations": [
         "염재보만 허용된 화면에 왼쪽·오른쪽 가장자리와 뒤편의 다른 사람 신체를 추가했다."
        ],
        "physics": "상체는 앉아서 약간 앞으로 기울인 자세로 자연스럽고, 팔은 아래쪽 프레임 밖으로 이어진다. 엉덩이와 지면의 접점은 보이지 않지만 부유를 나타내는 모습은 없다. 배경 인물들은 흙바닥에 앉아 있고 작은 그릇은 바닥에 놓여 있다. 입가의 쌀알과 부푼 볼도 물리적으로 가능한 식사 순간이다."
       }
      ]
     },
     "ok": true
    }
   ],
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "slot_winner_match": true,
   "slot_winner": {
    "gemini-pro": "A",
    "gpt-high": "A"
   },
   "route": "cross_slot_agree"
  },
  "dual": {
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "normalized": {
    "A": 2.0,
    "B": 1.167
   },
   "adjusted": {
    "A": 1.75,
    "B": 0.917
   },
   "violations": {
    "A": [
     "[gemini-pro] 명시되지 않은 인물 및 신체 일부 추가 (화면 양끝 및 배경)",
     "[gpt-high] 염재보만 허용된 화면에 왼쪽·오른쪽 가장자리와 뒤편의 다른 사람 신체를 추가했다."
    ],
    "B": [
     "[gemini-pro] 명시되지 않은 다수의 인물 추가 (배경 전체)",
     "[gpt-high] 허용된 염재보 외에 배경의 여러 식사 인물과 화면 왼쪽의 다른 사람 신체를 추가했다."
    ]
   },
   "per_model_winner": {
    "gemini-pro": "A",
    "gpt-high": "A"
   },
   "agreed": true
  },
  "totals": {
   "A": 1750,
   "B": 917
  },
  "selected": "A",
  "ranking": [
   "A",
   "B"
  ],
  "verdicts": [
   {
    "label": "A",
    "score": 1750,
    "verdict_ko": "클로즈업 프레이밍과 주인공이 밥을 머금고 짓는 표정 연기는 매우 훌륭하게 구현되었으나, 화면 좌우 가장자리에 프롬프트에서 금지한 다른 인물의 신체 일부가 등장하여 심각한 위반이 발생했습니다.  ★위반: [gemini-pro] 명시되지 않은 인물 및 신체 일부 추가 (화면 양끝 및 배경) / [gpt-high] 염재보만 허용된 화면에 왼쪽·오른쪽 가장자리와 뒤편의 다른 사람 신체를 추가했다."
   },
   {
    "label": "B",
    "score": 917,
    "verdict_ko": "주인공의 표정은 요구사항을 충족하지만, 절대로 추가하지 말라고 명시된 배경의 다른 인물들이 레퍼런스에서 그대로 복사되어 다수 등장했으며 화승총을 손에 쥐고 있어야 한다는 지시도 무시되었습니다.  ★위반: [gemini-pro] 명시되지 않은 다수의 인물 추가 (배경 전체) / [gpt-high] 허용된 염재보 외에 배경의 여러 식사 인물과 화면 왼쪽의 다른 사람 신체를 추가했다."
   }
  ],
  "refs": [
   {
    "label": "PREVIOUS SHOT STILL — a visually related earlier shot of this same place: the location's look, materials, fixed features and lighting mood are LOCKED to this photo; never copy its camera framing. The people visible in that photo are NOT in this shot: never carry their faces, bodies or clothing onto anyone here. Each person in THIS shot is defined solely by the PEOPLE list, their own CHARACTER REFERENCE images and their wardrobe notes. If this photo shows a character who cannot move (dead or unconscious), that character's exact pose, position and orientation are ALSO LOCKED — treat their body as a fixed prop of the set that only the camera moves around.",
    "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/db00517b-165b-4005-a287-f28936b59b95/images/272bd7af-9294-4490-98c6-88a40e9deacc/scene/recipe/S50sh1_sel.png",
    "asset_id": "1624ad2e-c774-4af4-b6d5-eefeff78f9d3",
    "role": "prev_still"
   },
   {
    "label": "CHARACTER REFERENCE — 염재보: the exact person appearing in this shot; match face, hair and build exactly.",
    "path": "<bytes:964351>",
    "asset_id": "ab287289-d0b7-436d-8f70-68452f898f73",
    "role": "character_ref"
   }
  ],
  "critique_skipped": true,
  "needs_reshoot": true,
  "shot_run_produced": false,
  "shot_run_uid": "06aa1435-5e2e-7221-8d6d-8d7e244359b2",
  "ref_mode": "prev+엔티티",
  "share_plan": {
   "ref_plan": "prev",
   "prev_anchor_tag": "S50sh1"
  }
 },
 "S50sh4::cine": {
  "applied": true,
  "attempted_at": "2026-09-09T06:45:29.709755+00:00",
  "fingerprint": "5e0f35261a759e8f24718a8d0225d4bc410d4b2c81ff79c49e79f5ce91485e93",
  "fingerprint_version": 2,
  "provider": "mai",
  "endpoint": "openrouter/chat-completions",
  "model": "microsoft/mai-image-2.6",
  "pack": "24.202608252115",
  "source_file": "S50sh4_sel.png",
  "source_sha256": "18708f696c760e1bf8ff490e334597e10383dac20855e38b92f0729f88a3e33b",
  "file": "S50sh4_cine.png",
  "staged_sha256": "d5486c81b1ba2aee40cb945b75de21e791828a8fe725cc2b53c4d0621ef26266",
  "latency_ms": 31990
 },
 "S50sh7::signage": {
  "fp": "0906f58ab137a6e3",
  "inscriptions": [],
  "cues": [],
  "dropped": []
 },
 "groupbg::camp_sentry_edge": {
  "input_fingerprint": "621bcfb6efb2658d",
  "meta": {
   "model": "gpt-image-2",
   "size": "1536x864",
   "pack": "11.202607220237",
   "contract": "bgfirst_full_v3",
   "group_sig": {
    "key": "camp_sentry_edge",
    "tags": [
     "S50sh7",
     "S54sh7"
    ]
   },
   "context_sig": "0fbadb1a4a419abf"
  },
  "prompt": "Create ONE empty live-action location background photograph — NO PEOPLE, no figures, no body parts, no silhouettes, no shadows or reflections of people anywhere.\n\"Empty\" means no people only: KEEP the location's inherent occupants and stock that define the place — animals in an animal shelter, pen or farm, goods and produce in a market, moored boats in a harbour — unless the shot text explicitly removes them.\nThe FIRST attached image is a thin-line storyboard sketch of a shot that happens at this location: use it ONLY as spatial evidence — what this place contains, how its ground, structures and landmarks are arranged and proportioned. Ignore the sketched people and arrows entirely, and do NOT copy its line style: render a fully photographic, physically plausible real place that fits THE LOCATION text below.\nHUMAN-SCALE CALIBRATION: derive every structure's true size from human-scale elements — a door ≈ 2m, a window ≈ 1–1.5m wide, one storey ≈ 2.5–3m; never inflate a small structure or shrink a large one.\n\nTHE LOCATION — early 20th-century Korea, Jeolla region; all people are Korean unless explicitly stated otherwise: The sentry position just outside the mountain camp's eating area, facing downhill into the surrounding forest.\n\nLOCATION DETAIL (descriptive evidence about this place from the production's location records — evidence, not a staging order): use it to understand what kind of place this is and what it permanently contains. Carry over ONLY its enduring physical features — layout, structures, ground and materials, fixed equipment and installations. Do NOT treat any lighting, weather or time-of-day wording, momentary object states, depicted events or subjective impressions in it as fixed facts of the place; the TIME OF DAY lock below is the sole authority for time and lighting.\n파청 산채 공터·훈련장·취사장·막사 주변: 산 중턱에 통나무를 세워 만든 조잡한 목책 안의 넓은 흙바닥 공터. 나무와 거적으로 만든 움막과 막사가 흩어져 있고, 돌과 진흙으로 만든 취사용 아궁이와 무쇠 솥, 모닥불 터가 있다. 전투 후에는 솥이 뒤집히고 막사가 불타거나 무너진다. (특징: 통나무를 박아 세운 목책; 넓게 다져진 흙바닥 (공터/훈련장); 돌과 진흙으로 만든 다수의 아궁이와 무쇠솥; 나무와 거적으로 엮어 만든 움막과 막사; 부서진 토기 밥그릇과 뒤집힌 무쇠솥 (전투 후); 불에 탄 나무 잔해와 흩어진 쌀밥 (전투 후))\n\nSCENE EVIDENCE (verbatim quotes from the screenplay about this place — treat them as evidence of what the location physically contains and looks like; stage the PLACE those moments happen in, but do NOT depict the momentary actions, people or staged props themselves):\n- 조금 떨어진 곳. 보초 의병 하나가 숲을 지켜보고 있다.\n- 산채 바깥. 보초를 서던 의병 하나.\n고요한 숲을 바라본다.\n- 다시 편백 사이. 총구 하나. 반대편에도 하나.\n\nTIME OF DAY (lock): day.\n\nRender ONE photorealistic empty location photograph, 16:9, neutral enough that every shot of this place can be staged from it later. No readable writing anywhere: surfaces that would carry writing may be present, but stage any wording out of legibility — an oblique angle, distance, shallow focus. No captions, watermarks or overlay text.",
  "effective_prompt": "Create ONE empty live-action location background photograph — NO PEOPLE, no figures, no body parts, no silhouettes, no shadows or reflections of people anywhere.\n\"Empty\" means no people only: KEEP the location's inherent occupants and stock that define the place — animals in an animal shelter, pen or farm, goods and produce in a market, moored boats in a harbour — unless the shot text explicitly removes them.\nThe FIRST attached image is a thin-line storyboard sketch of a shot that happens at this location: use it ONLY as spatial evidence — what this place contains, how its ground, structures and landmarks are arranged and proportioned. Ignore the sketched people and arrows entirely, and do NOT copy its line style: render a fully photographic, physically plausible real place that fits THE LOCATION text below.\nHUMAN-SCALE CALIBRATION: derive every structure's true size from human-scale elements — a door ≈ 2m, a window ≈ 1–1.5m wide, one storey ≈ 2.5–3m; never inflate a small structure or shrink a large one.\n\nTHE LOCATION — early 20th-century Korea, Jeolla region; all people are Korean unless explicitly stated otherwise: The sentry position just outside the mountain camp's eating area, facing downhill into the surrounding forest.\n\nLOCATION DETAIL (descriptive evidence about this place from the production's location records — evidence, not a staging order): use it to understand what kind of place this is and what it permanently contains. Carry over ONLY its enduring physical features — layout, structures, ground and materials, fixed equipment and installations. Do NOT treat any lighting, weather or time-of-day wording, momentary object states, depicted events or subjective impressions in it as fixed facts of the place; the TIME OF DAY lock below is the sole authority for time and lighting.\n파청 산채 공터·훈련장·취사장·막사 주변: 산 중턱에 통나무를 세워 만든 조잡한 목책 안의 넓은 흙바닥 공터. 나무와 거적으로 만든 움막과 막사가 흩어져 있고, 돌과 진흙으로 만든 취사용 아궁이와 무쇠 솥, 모닥불 터가 있다. 전투 후에는 솥이 뒤집히고 막사가 불타거나 무너진다. (특징: 통나무를 박아 세운 목책; 넓게 다져진 흙바닥 (공터/훈련장); 돌과 진흙으로 만든 다수의 아궁이와 무쇠솥; 나무와 거적으로 엮어 만든 움막과 막사; 부서진 토기 밥그릇과 뒤집힌 무쇠솥 (전투 후); 불에 탄 나무 잔해와 흩어진 쌀밥 (전투 후))\n\nSCENE EVIDENCE (verbatim quotes from the screenplay about this place — treat them as evidence of what the location physically contains and looks like; stage the PLACE those moments happen in, but do NOT depict the momentary actions, people or staged props themselves):\n- 조금 떨어진 곳. 보초 의병 하나가 숲을 지켜보고 있다.\n- 산채 바깥. 보초를 서던 의병 하나.\n고요한 숲을 바라본다.\n- 다시 편백 사이. 총구 하나. 반대편에도 하나.\n\nTIME OF DAY (lock): day.\n\nRender ONE photorealistic empty location photograph, 16:9, neutral enough that every shot of this place can be staged from it later. No readable writing anywhere: surfaces that would carry writing may be present, but stage any wording out of legibility — an oblique angle, distance, shallow focus. No captions, watermarks or overlay text.",
  "bg_path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/db00517b-165b-4005-a287-f28936b59b95/images/272bd7af-9294-4490-98c6-88a40e9deacc/scene/recipe/groupbg_camp_sentry_edge_f110f8.png",
  "asset_id": "238c405a-c572-403c-957e-0713389524ff",
  "input_asset_ids": [
   "c1e82e04-4ee4-4c86-bc7e-b1b98e10c920"
  ],
  "origin_tag": "S50sh7",
  "place_text": "The sentry position just outside the mountain camp's eating area, facing downhill into the surrounding forest.",
  "origin_inputs": {
   "place_text": "The sentry position just outside the mountain camp's eating area, facing downhill into the surrounding forest.",
   "time_of_day_en": "day",
   "conti_asset_id": "c1e82e04-4ee4-4c86-bc7e-b1b98e10c920"
  }
 },
 "S50sh7::bgfirst_bg": {
  "input_fingerprint": "be9888b3d3c7926a",
  "prompt": "Create the EMPTY BACKGROUND PLATE for one film shot — NO PEOPLE, no figures, no body parts, no sketch lines, no arrows anywhere.\n\"Empty\" means no people only: KEEP the location's inherent occupants and stock that define the place — animals in an animal shelter, pen or farm, goods in a market, moored boats in a harbour — unless the shot text explicitly removes them.\nThe FIRST attached image is a thin-line storyboard sketch: use ONLY its camera angle, horizon, perspective and the placement/size of buildings and set masses — ignore the sketched people and arrows entirely. The SECOND attached image (LOCATION PHOTOGRAPH) is the real place: take its architecture, materials, signage and fixed features, and RE-PROJECT them into the sketch's camera. If the photograph's camera differs from the sketch's, the sketch's camera wins.\nHUMAN-SCALE CALIBRATION: derive every structure's true size from human-scale elements — a door ≈ 2m, a window ≈ 1–1.5m wide, one storey ≈ 2.5–3m; never inflate a small structure or shrink a large one.\n\nSHOT TEXT this background must serve (Korean): 조금 떨어진 곳, 총을 든 채 숲 아래를 뚫어져라 지켜보는 보초 의병의 꼿꼿한 전신.\n\nLOCATION (lock): The sentry position just outside the mountain camp's eating area, facing downhill into the surrounding forest.\n\nTIME OF DAY (lock): day.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: wide shot\n- KEY BACKGROUND ELEMENTS: Forest (Being watched by the sentry from a position apart from the meal gathering) — Visible at the right and continuing beyond the right edge along his viewing direction; used as Provides meaningful look room and the unseen approach that holds his attention; Sentry's gun (Held while he keeps watch) — Seen obliquely alongside his torso without obscuring his face; used as Connects the still full-body pose to active guard duty.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Use consistent natural daylight with subdued saturation and sober contrast, keeping the watchful pause grounded in the same camp environment.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nRender ONE photorealistic empty location photograph, 16:9, that this shot can be staged inside later. No readable writing anywhere: surfaces that would carry writing may be present, but stage any wording out of legibility — an oblique angle, distance, shallow focus. No captions, watermarks or overlay text.",
  "effective_prompt": "Create the EMPTY BACKGROUND PLATE for one film shot — NO PEOPLE, no figures, no body parts, no sketch lines, no arrows anywhere.\n\"Empty\" means no people only: KEEP the location's inherent occupants and stock that define the place — animals in an animal shelter, pen or farm, goods in a market, moored boats in a harbour — unless the shot text explicitly removes them.\nThe FIRST attached image is a thin-line storyboard sketch: use ONLY its camera angle, horizon, perspective and the placement/size of buildings and set masses — ignore the sketched people and arrows entirely. The SECOND attached image (LOCATION PHOTOGRAPH) is the real place: take its architecture, materials, signage and fixed features, and RE-PROJECT them into the sketch's camera. If the photograph's camera differs from the sketch's, the sketch's camera wins.\nHUMAN-SCALE CALIBRATION: derive every structure's true size from human-scale elements — a door ≈ 2m, a window ≈ 1–1.5m wide, one storey ≈ 2.5–3m; never inflate a small structure or shrink a large one.\n\nSHOT TEXT this background must serve (Korean): 조금 떨어진 곳, 총을 든 채 숲 아래를 뚫어져라 지켜보는 보초 의병의 꼿꼿한 전신.\n\nLOCATION (lock): The sentry position just outside the mountain camp's eating area, facing downhill into the surrounding forest.\n\nTIME OF DAY (lock): day.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: wide shot\n- KEY BACKGROUND ELEMENTS: Forest (Being watched by the sentry from a position apart from the meal gathering) — Visible at the right and continuing beyond the right edge along his viewing direction; used as Provides meaningful look room and the unseen approach that holds his attention; Sentry's gun (Held while he keeps watch) — Seen obliquely alongside his torso without obscuring his face; used as Connects the still full-body pose to active guard duty.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Use consistent natural daylight with subdued saturation and sober contrast, keeping the watchful pause grounded in the same camp environment.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nRender ONE photorealistic empty location photograph, 16:9, that this shot can be staged inside later. No readable writing anywhere: surfaces that would carry writing may be present, but stage any wording out of legibility — an oblique angle, distance, shallow focus. No captions, watermarks or overlay text.",
  "bg_path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/db00517b-165b-4005-a287-f28936b59b95/images/272bd7af-9294-4490-98c6-88a40e9deacc/scene/recipe/S50sh7__bgfirst_bg.png",
  "asset_id": "6d9a1640-8042-4007-812b-f4b66a22aba8",
  "input_asset_ids": [
   "c1e82e04-4ee4-4c86-bc7e-b1b98e10c920",
   "238c405a-c572-403c-957e-0713389524ff"
  ]
 },
 "S50sh7": {
  "input_fingerprint": "2d7b9ae64063c2ad",
  "prompt": "Create ONE FINAL photorealistic live-action film still of the moment below — early 20th-century Korea, Jeolla region; all people are Korean unless explicitly stated otherwise. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 조금 떨어진 곳, 총을 든 채 숲 아래를 뚫어져라 지켜보는 보초 의병의 꼿꼿한 전신.\n\nLOCATION (lock): The sentry position just outside the mountain camp's eating area, facing downhill into the surrounding forest. The shot takes place here — the attached LOCATION PHOTOGRAPH shows the exact spot.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: wide shot\n- KEY BACKGROUND ELEMENTS: Forest (Being watched by the sentry from a position apart from the meal gathering) — Visible at the right and continuing beyond the right edge along his viewing direction; used as Provides meaningful look room and the unseen approach that holds his attention; Sentry's gun (Held while he keeps watch) — Seen obliquely alongside his torso without obscuring his face; used as Connects the still full-body pose to active guard duty.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Use consistent natural daylight with subdued saturation and sober contrast, keeping the watchful pause grounded in the same camp environment.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The white-rice meal remains laid out, with the pot upright and ammunition bundles concealed at the prepared tree-and-stone positions. 보초 의병: He stands at his post with a gun, watching the forest.\n\nTHE HAND THAT IS DOING THIS: the object at the centre of this shot is being held, operated, read, aimed or handed over by 보초 의병 right now, so 보초 의병's hand — and as much of the wrist and forearm as the framing reaches — is in the frame, gripping or resting on that object exactly the way its real use requires. Match that hand to 보초 의병: its size, build, skin, age, grooming, sleeve and anything worn on it belong to that person and to no one else.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 보초 의병 (한국인 남성, 성인 나이대, 자연스러운 검은 머리) — wearing: 바람을 막기 위해 겹겹이 껴입은 낡은 베옷과 무명옷. 주변 나뭇잎이나 흙색과 비슷한 어두운 색상. — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.",
  "roll_prompts": {
   "A": "Stage the shot. The FIRST attached image (SHOT BACKGROUND) is the finished empty background of this shot — keep it EXACTLY: its camera, perspective, architecture, lighting and every fixed feature stay untouched. The SECOND attached image (LAYOUT SKETCH) tells you ONLY where the people go: each sketched person's position, screen size, pose and the gaze/motion arrows. Ignore the sketch's background lines. The CHARACTER REFERENCE photographs show the real people.\nPlace the real people into the background at exactly the sketched positions, sizes and poses, following the arrow directions. No sketch lines or arrows may remain.\n\nCreate ONE FINAL photorealistic live-action film still of the moment below — early 20th-century Korea, Jeolla region; all people are Korean unless explicitly stated otherwise. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 조금 떨어진 곳, 총을 든 채 숲 아래를 뚫어져라 지켜보는 보초 의병의 꼿꼿한 전신.\n\nLOCATION (lock): The sentry position just outside the mountain camp's eating area, facing downhill into the surrounding forest. The shot takes place here — the FIRST attached image (SHOT BACKGROUND) is this exact place, already built: its ground, structures, horizon, materials and lighting are the finished truth of this location and must not be redesigned or replaced. No location photograph is attached — read the place from that image alone, and add no scenery, structure, vehicle or fixture that it does not already show. This lock governs the place only; the figures in the shot follow the staging and pose instructions.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: wide shot\n- KEY BACKGROUND ELEMENTS: Forest (Being watched by the sentry from a position apart from the meal gathering) — Visible at the right and continuing beyond the right edge along his viewing direction; used as Provides meaningful look room and the unseen approach that holds his attention; Sentry's gun (Held while he keeps watch) — Seen obliquely alongside his torso without obscuring his face; used as Connects the still full-body pose to active guard duty.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Use consistent natural daylight with subdued saturation and sober contrast, keeping the watchful pause grounded in the same camp environment.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The white-rice meal remains laid out, with the pot upright and ammunition bundles concealed at the prepared tree-and-stone positions. 보초 의병: He stands at his post with a gun, watching the forest.\n\nTHE HAND THAT IS DOING THIS: the object at the centre of this shot is being held, operated, read, aimed or handed over by 보초 의병 right now, so 보초 의병's hand — and as much of the wrist and forearm as the framing reaches — is in the frame, gripping or resting on that object exactly the way its real use requires. Match that hand to 보초 의병: its size, build, skin, age, grooming, sleeve and anything worn on it belong to that person and to no one else.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 보초 의병 (한국인 남성, 성인 나이대, 자연스러운 검은 머리) — wearing: 바람을 막기 위해 겹겹이 껴입은 낡은 베옷과 무명옷. 주변 나뭇잎이나 흙색과 비슷한 어두운 색상. — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.",
   "B": "Create ONE FINAL photorealistic live-action film still of the moment below — early 20th-century Korea, Jeolla region; all people are Korean unless explicitly stated otherwise. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 조금 떨어진 곳, 총을 든 채 숲 아래를 뚫어져라 지켜보는 보초 의병의 꼿꼿한 전신.\n\nLOCATION (lock): The sentry position just outside the mountain camp's eating area, facing downhill into the surrounding forest. The shot takes place here — the attached LOCATION PHOTOGRAPH shows the exact spot.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: wide shot\n- KEY BACKGROUND ELEMENTS: Forest (Being watched by the sentry from a position apart from the meal gathering) — Visible at the right and continuing beyond the right edge along his viewing direction; used as Provides meaningful look room and the unseen approach that holds his attention; Sentry's gun (Held while he keeps watch) — Seen obliquely alongside his torso without obscuring his face; used as Connects the still full-body pose to active guard duty.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Use consistent natural daylight with subdued saturation and sober contrast, keeping the watchful pause grounded in the same camp environment.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The white-rice meal remains laid out, with the pot upright and ammunition bundles concealed at the prepared tree-and-stone positions. 보초 의병: He stands at his post with a gun, watching the forest.\n\nTHE HAND THAT IS DOING THIS: the object at the centre of this shot is being held, operated, read, aimed or handed over by 보초 의병 right now, so 보초 의병's hand — and as much of the wrist and forearm as the framing reaches — is in the frame, gripping or resting on that object exactly the way its real use requires. Match that hand to 보초 의병: its size, build, skin, age, grooming, sleeve and anything worn on it belong to that person and to no one else.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 보초 의병 (한국인 남성, 성인 나이대, 자연스러운 검은 머리) — wearing: 바람을 막기 위해 겹겹이 껴입은 낡은 베옷과 무명옷. 주변 나뭇잎이나 흙색과 비슷한 어두운 색상. — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.\n\nCOMPOSITION VARIATION (this candidate only): keep the place, the people and the moment exactly as specified, and keep from the CAMERA & FRAME contract its subject scale, screen placement and key background elements exactly. For this candidate only, its camera angle and height are deliberately left open: choose a camera position distinctly different from the obvious one, as a film director picking a second setup on the same blocking. Stay true to any explicit framing the SHOT TEXT itself dictates."
  },
  "roll_refs": {
   "A": [
    {
     "label": "SHOT BACKGROUND",
     "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/db00517b-165b-4005-a287-f28936b59b95/images/272bd7af-9294-4490-98c6-88a40e9deacc/scene/recipe/S50sh7__bgfirst_bg.png",
     "asset_id": "6d9a1640-8042-4007-812b-f4b66a22aba8",
     "role": "bgfirst_bg"
    },
    {
     "label": "LAYOUT SKETCH (people placement only)",
     "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/db00517b-165b-4005-a287-f28936b59b95/images/272bd7af-9294-4490-98c6-88a40e9deacc/conti/conti_S50sh7.png",
     "asset_id": "c1e82e04-4ee4-4c86-bc7e-b1b98e10c920",
     "role": "conti_light"
    }
   ],
   "B": [
    {
     "label": "LOCATION PHOTOGRAPH — the exact place of this shot: its architecture, materials, fixed features and lighting mood are spatial truth; stage the moment inside this place. Never copy its camera framing.",
     "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/db00517b-165b-4005-a287-f28936b59b95/images/272bd7af-9294-4490-98c6-88a40e9deacc/scene/recipe/groupbg_camp_sentry_edge_f110f8.png",
     "asset_id": "238c405a-c572-403c-957e-0713389524ff",
     "role": "bgfirst_group_bg"
    }
   ]
  },
  "shot_run_spend_attempt_count": 1,
  "judge_flip_skipped": "judge_fn_owns_order",
  "readings": [
   {
    "label": "A",
    "direction": "의병의 시선이 화면 오른쪽 아래 숲을 향함. 들고 있는 총의 총구는 사선 위를 향하고 있음.",
    "built_space": "레퍼런스와 동일한 흙바닥 경사지와 돌로 둘러싸인 나무 밑동이 위치함. 인공 구조물 없음.",
    "entities": "낡고 어두운 색상의 옷을 입은 한국인 남성, 소총 등 프롬프트의 인물 및 소품 묘사와 일치함.",
    "hard_violations": [],
    "physics": "두 발로 흙바닥 경사지를 안정적으로 딛고 서 있으며, 두 손으로 총을 쥐고 몸을 지탱함."
   },
   {
    "label": "B",
    "direction": "의병의 시선이 화면 오른쪽 숲을 향함. 들고 있는 총의 총구는 아래를 향함.",
    "built_space": "화면 왼쪽에 레퍼런스에 없는 목조 기둥과 지붕으로 이루어진 구조물이 자리 잡고 있음.",
    "entities": "낡은 옷을 입은 한국인 남성과 소총은 일치하나, 배경 구조물 아래에 프롬프트에 없는 여분의 소총과 짐 꾸러미들이 있음.",
    "hard_violations": [
     "[gemini-pro] 레퍼런스 장소에 존재하지 않는 목조 구조물(오두막/정자) 임의 생성",
     "[gemini-pro] 지시되지 않은 여분의 소총을 화면 좌측 바닥에 추가",
     "[gpt-high] 장소 참조에 없는 지붕과 기둥을 갖춘 목조 시설을 화면 왼쪽에 새로 만들었다.",
     "[gpt-high] 보초가 든 총 외에 지시되지 않은 여분의 소총 한 정을 왼쪽 돌무더기에 추가했다."
    ],
    "physics": "두 발로 땅을 딛고 서 있으며, 두 손으로 총을 쥐고 있음."
   }
  ],
  "cross_model_order": {
   "policy": "cross_model_order_v2_slot0_fwd_slot1_rev_parallel2",
   "slots": [
    {
     "model": "gemini-pro",
     "order": "forward",
     "display_to_canonical": {
      "A": "A",
      "B": "B"
     },
     "raw": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 9,
        "verdict_ko": "레퍼런스의 장소 형태를 정확히 유지하면서 지정된 와이드 샷 구도와 인물의 경계하는 자세를 충실히 구현함."
       },
       {
        "label": "B",
        "score": 3,
        "verdict_ko": "지정된 장소에 존재하지 않는 지붕 구조물과 여분의 소총을 임의로 추가하여 장소 설정(Location lock)을 심각하게 위반함."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "의병의 시선이 화면 오른쪽 아래 숲을 향함. 들고 있는 총의 총구는 사선 위를 향하고 있음.",
        "built_space": "레퍼런스와 동일한 흙바닥 경사지와 돌로 둘러싸인 나무 밑동이 위치함. 인공 구조물 없음.",
        "entities": "낡고 어두운 색상의 옷을 입은 한국인 남성, 소총 등 프롬프트의 인물 및 소품 묘사와 일치함.",
        "hard_violations": [],
        "physics": "두 발로 흙바닥 경사지를 안정적으로 딛고 서 있으며, 두 손으로 총을 쥐고 몸을 지탱함."
       },
       {
        "label": "B",
        "direction": "의병의 시선이 화면 오른쪽 숲을 향함. 들고 있는 총의 총구는 아래를 향함.",
        "built_space": "화면 왼쪽에 레퍼런스에 없는 목조 기둥과 지붕으로 이루어진 구조물이 자리 잡고 있음.",
        "entities": "낡은 옷을 입은 한국인 남성과 소총은 일치하나, 배경 구조물 아래에 프롬프트에 없는 여분의 소총과 짐 꾸러미들이 있음.",
        "hard_violations": [
         "레퍼런스 장소에 존재하지 않는 목조 구조물(오두막/정자) 임의 생성",
         "지시되지 않은 여분의 소총을 화면 좌측 바닥에 추가"
        ],
        "physics": "두 발로 땅을 딛고 서 있으며, 두 손으로 총을 쥐고 있음."
       }
      ],
      "all_candidates_fail": false
     },
     "normalized": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 9,
        "verdict_ko": "레퍼런스의 장소 형태를 정확히 유지하면서 지정된 와이드 샷 구도와 인물의 경계하는 자세를 충실히 구현함."
       },
       {
        "label": "B",
        "score": 3,
        "verdict_ko": "지정된 장소에 존재하지 않는 지붕 구조물과 여분의 소총을 임의로 추가하여 장소 설정(Location lock)을 심각하게 위반함."
       }
      ],
      "all_candidates_fail": false,
      "readings": [
       {
        "label": "A",
        "direction": "의병의 시선이 화면 오른쪽 아래 숲을 향함. 들고 있는 총의 총구는 사선 위를 향하고 있음.",
        "built_space": "레퍼런스와 동일한 흙바닥 경사지와 돌로 둘러싸인 나무 밑동이 위치함. 인공 구조물 없음.",
        "entities": "낡고 어두운 색상의 옷을 입은 한국인 남성, 소총 등 프롬프트의 인물 및 소품 묘사와 일치함.",
        "hard_violations": [],
        "physics": "두 발로 흙바닥 경사지를 안정적으로 딛고 서 있으며, 두 손으로 총을 쥐고 몸을 지탱함."
       },
       {
        "label": "B",
        "direction": "의병의 시선이 화면 오른쪽 숲을 향함. 들고 있는 총의 총구는 아래를 향함.",
        "built_space": "화면 왼쪽에 레퍼런스에 없는 목조 기둥과 지붕으로 이루어진 구조물이 자리 잡고 있음.",
        "entities": "낡은 옷을 입은 한국인 남성과 소총은 일치하나, 배경 구조물 아래에 프롬프트에 없는 여분의 소총과 짐 꾸러미들이 있음.",
        "hard_violations": [
         "레퍼런스 장소에 존재하지 않는 목조 구조물(오두막/정자) 임의 생성",
         "지시되지 않은 여분의 소총을 화면 좌측 바닥에 추가"
        ],
        "physics": "두 발로 땅을 딛고 서 있으며, 두 손으로 총을 쥐고 있음."
       }
      ]
     },
     "ok": true
    },
    {
     "model": "gpt-high",
     "order": "reverse",
     "display_to_canonical": {
      "A": "B",
      "B": "A"
     },
     "raw": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 3,
        "verdict_ko": "전신 경계 자세와 총의 지지는 맞지만, 참조에 없는 목조 시설과 여분의 소총을 추가했으며 인물이 너무 커서 지정된 와이드 구도에도 덜 충실하다."
       },
       {
        "label": "B",
        "score": 9,
        "verdict_ko": "참조의 나무·돌무더기·산비탈을 유지하면서 보초의 꼿꼿한 전신을 왼쪽에 두고 오른쪽 감시 숲에 충분한 공간을 준 와이드 숏으로 가장 충실하다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "보초의 얼굴과 시선은 화면 오른쪽 숲을 향하며 카메라를 보지 않는다. 시선의 하향 각도는 크지 않다. 들고 있는 소총의 총구는 오른쪽 아래 비탈과 숲 쪽을 향한다. 사격 조준이 아니라 총을 내려 들고 경계하는 상태로 읽힌다. 왼쪽 돌무더기의 여분 소총은 오른쪽 위로 비스듬히 놓여 있다.",
        "built_space": "왼쪽 큰 나무 한 그루와 그 밑 돌무더기, 흙과 바위로 된 전망 지점, 오른쪽 아래 숲과 먼 산이 보인다. 보초는 비탈 가장자리의 흙바닥에 서 있다. 그러나 왼쪽 가장자리에 지붕·기둥·높인 바닥을 갖춘 목조 시설 한 채를 추가해, 시설이 없는 참조 장소를 바꾸었다. 인물은 화면 높이의 약 8할을 차지하고 중앙 가까이에 있어 떨어져 바라보는 와이드 숏의 여유가 적다.",
        "entities": "사람은 검은 머리의 성인 남성 한 명이며 한국인 보초 설정과 외형상 부합한다. 낡고 거친 흙갈색 겹옷, 바지, 다리 감개와 신발을 착용했다. 두 손으로 목제 개머리판과 금속 총열이 있는 소총 한 정을 들고 있다. 왼쪽에는 별도의 소총 한 정과 천·짐처럼 보이는 물건들이 추가되어 있다. 흰쌀밥과 세워 둔 솥은 보이지 않으며, 천 더미를 은닉 탄약이라고 확인할 수는 없다. 읽을 수 있는 글씨나 추가 인물은 없다.",
        "hard_violations": [
         "장소 참조에 없는 지붕과 기둥을 갖춘 목조 시설을 화면 왼쪽에 새로 만들었다.",
         "보초가 든 총 외에 지시되지 않은 여분의 소총 한 정을 왼쪽 돌무더기에 추가했다."
        ],
        "physics": "두 발이 흙바닥에 닿아 체중을 지탱하고 있으며 곧게 선 자세가 가능하다. 한 손은 소총의 손잡이 부근을 잡고 다른 손은 앞쪽 목재 부분을 받친다. 멜빵은 총에 연결되어 아래로 자연스럽게 처진다. 여분 소총과 짐은 돌과 바닥에 기대어 있어 무지지 부유는 보이지 않는다."
       },
       {
        "label": "B",
        "direction": "보초의 얼굴과 시선은 화면 오른쪽의 숲이 우거진 사면을 향한다. 고개를 크게 숙이지는 않았지만 아래쪽 숲을 감시하는 위치 관계가 읽힌다. 숲은 시선 앞에 넓게 남고 오른쪽 화면 밖으로 이어진다. 소총 총구는 오른쪽 위를 향하지만, 지시된 행동은 숲을 향한 사격이 아니라 총을 든 경계이므로 조준 방향 위반은 아니다.",
        "built_space": "왼쪽 큰 나무 한 그루, 그 밑 돌무더기, 흙과 노출된 바위로 된 비탈, 오른쪽 아래 숲과 겹친 산 능선이 보인다. 참조와 같은 주요 지형 요소를 유지하며 별도 건축물을 추가하지 않았다. 보초는 나무 오른쪽의 지면에 서 있고, 화면 왼쪽 약 3분의 1 지점에 전신이 놓인다. 인물 높이는 화면의 약 6할이며 오른쪽 숲에 넓은 관찰 여백이 확보된다.",
        "entities": "검은 머리의 성인 남성 한 명이 있으며 한국인 보초 설정과 외형상 부합한다. 어두운 회갈색의 거친 겹옷과 바지, 다리 감개, 신발을 착용했다. 목제 개머리판·금속 총열·멜빵이 있는 소총 한 정을 몸통 앞에 비스듬히 들었으며 얼굴을 가리지 않는다. 손과 소매도 해당 인물에 자연스럽게 연결된다. 식사와 솥, 은닉 탄약은 화면에서 확인되지 않지만 이를 보여주기 위해 식사 장소까지 구도를 넓힐 필요는 없다. 추가 인물과 읽을 수 있는 글씨는 없다.",
        "hard_violations": [],
        "physics": "양발이 경사진 흙바닥에 닿고 다리가 몸을 지탱한다. 꼿꼿하게 멈춰 선 자세에 불가능한 균형은 없다. 한 손이 소총의 손잡이 부근을 잡고 다른 손이 앞부분을 받쳐 무게를 지탱한다. 연결된 멜빵은 중력 방향으로 늘어져 있으며 지지 없이 떠 있는 물체는 없다."
       }
      ],
      "all_candidates_fail": false
     },
     "normalized": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 3,
        "verdict_ko": "전신 경계 자세와 총의 지지는 맞지만, 참조에 없는 목조 시설과 여분의 소총을 추가했으며 인물이 너무 커서 지정된 와이드 구도에도 덜 충실하다."
       },
       {
        "label": "A",
        "score": 9,
        "verdict_ko": "참조의 나무·돌무더기·산비탈을 유지하면서 보초의 꼿꼿한 전신을 왼쪽에 두고 오른쪽 감시 숲에 충분한 공간을 준 와이드 숏으로 가장 충실하다."
       }
      ],
      "all_candidates_fail": false,
      "readings": [
       {
        "label": "B",
        "direction": "보초의 얼굴과 시선은 화면 오른쪽 숲을 향하며 카메라를 보지 않는다. 시선의 하향 각도는 크지 않다. 들고 있는 소총의 총구는 오른쪽 아래 비탈과 숲 쪽을 향한다. 사격 조준이 아니라 총을 내려 들고 경계하는 상태로 읽힌다. 왼쪽 돌무더기의 여분 소총은 오른쪽 위로 비스듬히 놓여 있다.",
        "built_space": "왼쪽 큰 나무 한 그루와 그 밑 돌무더기, 흙과 바위로 된 전망 지점, 오른쪽 아래 숲과 먼 산이 보인다. 보초는 비탈 가장자리의 흙바닥에 서 있다. 그러나 왼쪽 가장자리에 지붕·기둥·높인 바닥을 갖춘 목조 시설 한 채를 추가해, 시설이 없는 참조 장소를 바꾸었다. 인물은 화면 높이의 약 8할을 차지하고 중앙 가까이에 있어 떨어져 바라보는 와이드 숏의 여유가 적다.",
        "entities": "사람은 검은 머리의 성인 남성 한 명이며 한국인 보초 설정과 외형상 부합한다. 낡고 거친 흙갈색 겹옷, 바지, 다리 감개와 신발을 착용했다. 두 손으로 목제 개머리판과 금속 총열이 있는 소총 한 정을 들고 있다. 왼쪽에는 별도의 소총 한 정과 천·짐처럼 보이는 물건들이 추가되어 있다. 흰쌀밥과 세워 둔 솥은 보이지 않으며, 천 더미를 은닉 탄약이라고 확인할 수는 없다. 읽을 수 있는 글씨나 추가 인물은 없다.",
        "hard_violations": [
         "장소 참조에 없는 지붕과 기둥을 갖춘 목조 시설을 화면 왼쪽에 새로 만들었다.",
         "보초가 든 총 외에 지시되지 않은 여분의 소총 한 정을 왼쪽 돌무더기에 추가했다."
        ],
        "physics": "두 발이 흙바닥에 닿아 체중을 지탱하고 있으며 곧게 선 자세가 가능하다. 한 손은 소총의 손잡이 부근을 잡고 다른 손은 앞쪽 목재 부분을 받친다. 멜빵은 총에 연결되어 아래로 자연스럽게 처진다. 여분 소총과 짐은 돌과 바닥에 기대어 있어 무지지 부유는 보이지 않는다."
       },
       {
        "label": "A",
        "direction": "보초의 얼굴과 시선은 화면 오른쪽의 숲이 우거진 사면을 향한다. 고개를 크게 숙이지는 않았지만 아래쪽 숲을 감시하는 위치 관계가 읽힌다. 숲은 시선 앞에 넓게 남고 오른쪽 화면 밖으로 이어진다. 소총 총구는 오른쪽 위를 향하지만, 지시된 행동은 숲을 향한 사격이 아니라 총을 든 경계이므로 조준 방향 위반은 아니다.",
        "built_space": "왼쪽 큰 나무 한 그루, 그 밑 돌무더기, 흙과 노출된 바위로 된 비탈, 오른쪽 아래 숲과 겹친 산 능선이 보인다. 참조와 같은 주요 지형 요소를 유지하며 별도 건축물을 추가하지 않았다. 보초는 나무 오른쪽의 지면에 서 있고, 화면 왼쪽 약 3분의 1 지점에 전신이 놓인다. 인물 높이는 화면의 약 6할이며 오른쪽 숲에 넓은 관찰 여백이 확보된다.",
        "entities": "검은 머리의 성인 남성 한 명이 있으며 한국인 보초 설정과 외형상 부합한다. 어두운 회갈색의 거친 겹옷과 바지, 다리 감개, 신발을 착용했다. 목제 개머리판·금속 총열·멜빵이 있는 소총 한 정을 몸통 앞에 비스듬히 들었으며 얼굴을 가리지 않는다. 손과 소매도 해당 인물에 자연스럽게 연결된다. 식사와 솥, 은닉 탄약은 화면에서 확인되지 않지만 이를 보여주기 위해 식사 장소까지 구도를 넓힐 필요는 없다. 추가 인물과 읽을 수 있는 글씨는 없다.",
        "hard_violations": [],
        "physics": "양발이 경사진 흙바닥에 닿고 다리가 몸을 지탱한다. 꼿꼿하게 멈춰 선 자세에 불가능한 균형은 없다. 한 손이 소총의 손잡이 부근을 잡고 다른 손이 앞부분을 받쳐 무게를 지탱한다. 연결된 멜빵은 중력 방향으로 늘어져 있으며 지지 없이 떠 있는 물체는 없다."
       }
      ]
     },
     "ok": true
    }
   ],
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "slot_winner_match": true,
   "slot_winner": {
    "gemini-pro": "A",
    "gpt-high": "A"
   },
   "route": "cross_slot_agree"
  },
  "dual": {
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "normalized": {
    "A": 2.0,
    "B": 0.667
   },
   "adjusted": {
    "A": 2.0,
    "B": 0.417
   },
   "violations": {
    "B": [
     "[gemini-pro] 레퍼런스 장소에 존재하지 않는 목조 구조물(오두막/정자) 임의 생성",
     "[gemini-pro] 지시되지 않은 여분의 소총을 화면 좌측 바닥에 추가",
     "[gpt-high] 장소 참조에 없는 지붕과 기둥을 갖춘 목조 시설을 화면 왼쪽에 새로 만들었다.",
     "[gpt-high] 보초가 든 총 외에 지시되지 않은 여분의 소총 한 정을 왼쪽 돌무더기에 추가했다."
    ]
   },
   "per_model_winner": {
    "gemini-pro": "A",
    "gpt-high": "A"
   },
   "agreed": true
  },
  "totals": {
   "A": 2000,
   "B": 417
  },
  "selected": "A",
  "ranking": [
   "A",
   "B"
  ],
  "verdicts": [
   {
    "label": "A",
    "score": 2000,
    "verdict_ko": "레퍼런스의 장소 형태를 정확히 유지하면서 지정된 와이드 샷 구도와 인물의 경계하는 자세를 충실히 구현함."
   },
   {
    "label": "B",
    "score": 417,
    "verdict_ko": "지정된 장소에 존재하지 않는 지붕 구조물과 여분의 소총을 임의로 추가하여 장소 설정(Location lock)을 심각하게 위반함.  ★위반: [gemini-pro] 레퍼런스 장소에 존재하지 않는 목조 구조물(오두막/정자) 임의 생성 / [gemini-pro] 지시되지 않은 여분의 소총을 화면 좌측 바닥에 추가 / [gpt-high] 장소 참조에 없는 지붕과 기둥을 갖춘 목조 시설을 화면 왼쪽에 새로 만들었다. / [gpt-high] 보초가 든 총 외에 지시되지 않은 여분의 소총 한 정을 왼쪽 돌무더기에 추가했다."
   }
  ],
  "refs": [
   {
    "label": "LOCATION PHOTOGRAPH — the exact place of this shot: its architecture, materials, fixed features and lighting mood are spatial truth; stage the moment inside this place. Never copy its camera framing.",
    "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/db00517b-165b-4005-a287-f28936b59b95/images/272bd7af-9294-4490-98c6-88a40e9deacc/scene/recipe/groupbg_camp_sentry_edge_f110f8.png",
    "asset_id": "238c405a-c572-403c-957e-0713389524ff",
    "role": "bgfirst_group_bg"
   }
  ],
  "critique_skipped": true,
  "shot_run_produced": false,
  "shot_run_uid": "06aa1435-5fc6-7df8-babc-a6804d271f61",
  "bgfirst": {
   "bg_path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/db00517b-165b-4005-a287-f28936b59b95/images/272bd7af-9294-4490-98c6-88a40e9deacc/scene/recipe/S50sh7__bgfirst_bg.png",
   "bg_asset_id": "6d9a1640-8042-4007-812b-f4b66a22aba8",
   "bg_record_key": "S50sh7::bgfirst_bg",
   "chain_winner": true,
   "authority": "groupbg",
   "group_key": "camp_sentry_edge",
   "groupbg_asset_id": "238c405a-c572-403c-957e-0713389524ff"
  },
  "ref_mode": "재투영 배경+콘티+엔티티 (2택1: 체인 승)",
  "share_plan": {
   "ref_plan": "background"
  }
 },
 "S50sh7::cine": {
  "applied": true,
  "attempted_at": "2026-09-09T06:50:37.758811+00:00",
  "fingerprint": "50c36c10afb754a08688eb4c8e001118a772cfed4500385e555feb190c0022ad",
  "fingerprint_version": 2,
  "provider": "mai",
  "endpoint": "openrouter/chat-completions",
  "model": "microsoft/mai-image-2.6",
  "pack": "24.202608252115",
  "source_file": "S50sh7_sel.png",
  "source_sha256": "ee795f4b0fda1685c9719f93703526a22beaefbf2f62468d1cf92b218276d5cf",
  "file": "S50sh7_cine.png",
  "staged_sha256": "2534086566aa90e263c7dedf2672ad4b5afcf9cd23bcb9bd70e77c7aa67689f7",
  "latency_ms": 28136
 },
 "S51sh2::signage": {
  "fp": "2f8c88ecd974755b",
  "inscriptions": [],
  "cues": [],
  "dropped": []
 },
 "S51sh2": {
  "input_fingerprint": "cdd98994ba85c9e1",
  "prompt": "Create ONE FINAL photorealistic live-action film still of the moment below — early 20th-century Korea, Jeolla region; all people are Korean unless explicitly stated otherwise. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 안규홍이 높이 든 낫자루로 떨어지는 칼날을 탁 맞받아 막은 순간.\n\nLOCATION (lock): The outdoor market lane before the straw-sandal stall, within the encircled fighting space beside the vendors' stands. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: medium shot\n- KEY BACKGROUND ELEMENTS: Sickle handle (Raised to block the descending sword) — Runs obliquely across the upper portion of 안규홍's face, with the collision point clear of his eyes; used as Connects both hands to the defensive impact while retaining his expression beneath it; Attacking sword (Arrested against the sickle handle during its downward stroke) — Enters from the upper edge on a downward diagonal; its wielder remains outside frame; used as Makes the immediate threat readable through a bounded intrusion into the human-centered composition; Market passage (The fight is underway within the market) — Only peripheral depth remains visible around 안규홍; used as Provides spatial context without distracting from the block.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Natural daylight with muted saturation and firm but realistic contrast keeps the weapon contact legible without stylized flashes.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The clothing, hair and overall look of 안규홍 — who appear both in that photo and in this shot — are LOCKED to that photo. Anyone else visible in that photo is NOT in this shot: never carry their face, body or clothing onto anyone here. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The market lanes remain blocked at both ends. 안규홍: He holds his dirt-gritted sickle shaft up defensively; his face remains unwashed, hair tied and neck bare.\n\nTHE HAND THAT IS DOING THIS: the object at the centre of this shot is being held, operated, read, aimed or handed over by 안규홍 right now, so 안규홍's hand — and as much of the wrist and forearm as the framing reaches — is in the frame, gripping or resting on that object exactly the way its real use requires. Match that hand to 안규홍: its size, build, skin, age, grooming, sleeve and anything worn on it belong to that person and to no one else.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 안규홍 (한국인 남성, 청년 나이대, 검은 머리, 단단히 묶은 머리) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.",
  "roll_prompts": {
   "A": "Create ONE FINAL photorealistic live-action film still of the moment below — early 20th-century Korea, Jeolla region; all people are Korean unless explicitly stated otherwise. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 안규홍이 높이 든 낫자루로 떨어지는 칼날을 탁 맞받아 막은 순간.\n\nLOCATION (lock): The outdoor market lane before the straw-sandal stall, within the encircled fighting space beside the vendors' stands. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: medium shot\n- KEY BACKGROUND ELEMENTS: Sickle handle (Raised to block the descending sword) — Runs obliquely across the upper portion of 안규홍's face, with the collision point clear of his eyes; used as Connects both hands to the defensive impact while retaining his expression beneath it; Attacking sword (Arrested against the sickle handle during its downward stroke) — Enters from the upper edge on a downward diagonal; its wielder remains outside frame; used as Makes the immediate threat readable through a bounded intrusion into the human-centered composition; Market passage (The fight is underway within the market) — Only peripheral depth remains visible around 안규홍; used as Provides spatial context without distracting from the block.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Natural daylight with muted saturation and firm but realistic contrast keeps the weapon contact legible without stylized flashes.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The clothing, hair and overall look of 안규홍 — who appear both in that photo and in this shot — are LOCKED to that photo. Anyone else visible in that photo is NOT in this shot: never carry their face, body or clothing onto anyone here. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The market lanes remain blocked at both ends. 안규홍: He holds his dirt-gritted sickle shaft up defensively; his face remains unwashed, hair tied and neck bare.\n\nTHE HAND THAT IS DOING THIS: the object at the centre of this shot is being held, operated, read, aimed or handed over by 안규홍 right now, so 안규홍's hand — and as much of the wrist and forearm as the framing reaches — is in the frame, gripping or resting on that object exactly the way its real use requires. Match that hand to 안규홍: its size, build, skin, age, grooming, sleeve and anything worn on it belong to that person and to no one else.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 안규홍 (한국인 남성, 청년 나이대, 검은 머리, 단단히 묶은 머리) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.",
   "B": "Create ONE FINAL photorealistic live-action film still of the moment below — early 20th-century Korea, Jeolla region; all people are Korean unless explicitly stated otherwise. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 안규홍이 높이 든 낫자루로 떨어지는 칼날을 탁 맞받아 막은 순간.\n\nLOCATION (lock): The outdoor market lane before the straw-sandal stall, within the encircled fighting space beside the vendors' stands. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: medium shot\n- KEY BACKGROUND ELEMENTS: Sickle handle (Raised to block the descending sword) — Runs obliquely across the upper portion of 안규홍's face, with the collision point clear of his eyes; used as Connects both hands to the defensive impact while retaining his expression beneath it; Attacking sword (Arrested against the sickle handle during its downward stroke) — Enters from the upper edge on a downward diagonal; its wielder remains outside frame; used as Makes the immediate threat readable through a bounded intrusion into the human-centered composition; Market passage (The fight is underway within the market) — Only peripheral depth remains visible around 안규홍; used as Provides spatial context without distracting from the block.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Natural daylight with muted saturation and firm but realistic contrast keeps the weapon contact legible without stylized flashes.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The clothing, hair and overall look of 안규홍 — who appear both in that photo and in this shot — are LOCKED to that photo. Anyone else visible in that photo is NOT in this shot: never carry their face, body or clothing onto anyone here. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The market lanes remain blocked at both ends. 안규홍: He holds his dirt-gritted sickle shaft up defensively; his face remains unwashed, hair tied and neck bare.\n\nTHE HAND THAT IS DOING THIS: the object at the centre of this shot is being held, operated, read, aimed or handed over by 안규홍 right now, so 안규홍's hand — and as much of the wrist and forearm as the framing reaches — is in the frame, gripping or resting on that object exactly the way its real use requires. Match that hand to 안규홍: its size, build, skin, age, grooming, sleeve and anything worn on it belong to that person and to no one else.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 안규홍 (한국인 남성, 청년 나이대, 검은 머리, 단단히 묶은 머리) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.\n\nCOMPOSITION VARIATION (this candidate only): keep the place, the people and the moment exactly as specified, and keep from the CAMERA & FRAME contract its subject scale, screen placement and key background elements exactly. For this candidate only, its camera angle and height are deliberately left open: choose a camera position distinctly different from the obvious one, as a film director picking a second setup on the same blocking. Stay true to any explicit framing the SHOT TEXT itself dictates."
  },
  "shot_run_spend_attempt_count": 1,
  "readings": [
   {
    "label": "A",
    "direction": "안규홍의 시선은 왼쪽을 향하고 있어 떨어지는 칼이나 공격자를 직접 보지 않고 있습니다. 하지만 공격자의 칼은 우측 상단에서 대각선 아래로 향하며 안규홍이 든 자루에 정확히 맞닿아 물리적 충돌 지점을 형성합니다.",
    "built_space": "배경에 짚신이 걸려 있는 시장 가판대가 위치하여 레퍼런스의 야외 시장 공간과 정확히 일치합니다. 인물과 배경 구조물 간의 거리감 및 공간적 맥락이 지시대로 잘 표현되었습니다.",
    "entities": "안규홍의 얼굴, 상투, 더러워진 복장은 레퍼런스와 일치합니다. 그러나 방어에 사용된 무기는 낫의 날이 긴 장대 중간에 붙어있는 기형적인 창 형태로 변형되었습니다. 또한 프레임 밖에 있어야 할 공격자의 두 손과 팔이 명확히 등장합니다.",
    "hard_violations": [
     "[gemini-pro] 낫이 긴 장대 중간에 날이 달린 형태의 전혀 다른 무기로 창조됨 (Invented objects)",
     "[gemini-pro] 공격자의 손과 팔이 화면에 등장하여 'wielder remains outside frame' 카메라 지시 위반",
     "[gemini-pro] 지시문에 명시되지 않은 배경 인물들이 화면 양쪽에 등장함",
     "[gpt-high] 공격자의 손·팔이 화면 안에 들어와 칼을 휘두르는 사람은 화면 밖에 있어야 한다는 명시적 구도를 위반한다.",
     "[gpt-high] 좌우 배경에 안규홍 외의 인물 두 명이 보여 허용된 등장인물 제한을 위반한다."
    ],
    "physics": "양손으로 하나의 긴 자루를 단단히 쥐고 떨어지는 칼을 막아내는 자세와 충돌의 물리적 저항감은 안정적으로 구현되었으며 신체 지지도 자연스럽습니다."
   },
   {
    "label": "B",
    "direction": "안규홍의 시선은 우측 상단의 공격자를 향해 올바르게 고정되어 있습니다. 공격자의 칼은 우측 상단에서 들어와 수평으로 든 막대에 부딪힙니다.",
    "built_space": "짚신이 걸린 가판대와 야외 시장 골목의 배경 구조물들이 레퍼런스에 맞게 적절히 배치되어 있습니다.",
    "entities": "안규홍의 외형과 복장은 잘 묘사되었습니다. 그러나 손에 쥔 무기는 낫과 단순한 나무 막대가 겹쳐지거나 분리된 알 수 없는 형태입니다. 우측 가장자리에 공격자의 얼굴 일부가 등장합니다.",
    "hard_violations": [
     "[gemini-pro] 하나의 낫자루가 아닌 분리된 막대와 낫을 들고 있으며 낫의 날이 막대와 비정상적으로 겹치는 등 물리적으로 불가능한 무기 구조 발생",
     "[gemini-pro] 화면 우측 가장자리에 공격자의 얼굴이 등장하여 'wielder remains outside frame' 지시 및 인물 등장 제한 위반",
     "[gemini-pro] 하나의 낫자루가 안규홍의 얼굴 윗부분을 대각선으로 가로지른다는 핵심 구도 지시(Runs obliquely across)를 완전히 무시하고 수평으로 막아냄",
     "[gemini-pro] 지시문에 명시되지 않은 다수의 배경 인물 등장",
     "[gpt-high] 안규홍 외의 여러 주변 인물과 공격자의 신체 일부가 노출되어, 안규홍만 보이고 칼을 휘두르는 사람은 화면 밖에 있어야 한다는 제한을 위반한다.",
     "[gpt-high] 낫과 별도로 긴 나무 막대를 추가하여, 하나의 낫자루를 양손으로 들어 막는 필수 소품·동작 구성을 바꿨다."
    ],
    "physics": "왼손에 쥔 막대와 오른손에 쥔 낫의 자루가 별개로 분리되어 허공에서 비정상적으로 겹쳐 있으며, 양손이 결합되어 방어의 충격을 지탱한다는 물리적 구조가 붕괴되었습니다."
   }
  ],
  "cross_model_order": {
   "policy": "cross_model_order_v2_slot0_fwd_slot1_rev_parallel2",
   "slots": [
    {
     "model": "gemini-pro",
     "order": "forward",
     "display_to_canonical": {
      "A": "A",
      "B": "B"
     },
     "raw": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 4,
        "verdict_ko": "지시된 낫자루의 대각선 방어 구도와 안규홍의 외형을 훌륭하게 구현했으나, 낫이 긴 장대처럼 왜곡되고 공격자의 손이 화면에 포함되어 감점되었습니다."
       },
       {
        "label": "B",
        "score": 2,
        "verdict_ko": "무기가 두 개의 막대로 분리되는 심각한 물리적 구조 오류가 발생했으며, 지시된 대각선 방어 프레이밍을 전혀 따르지 않았습니다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "안규홍의 시선은 왼쪽을 향하고 있어 떨어지는 칼이나 공격자를 직접 보지 않고 있습니다. 하지만 공격자의 칼은 우측 상단에서 대각선 아래로 향하며 안규홍이 든 자루에 정확히 맞닿아 물리적 충돌 지점을 형성합니다.",
        "built_space": "배경에 짚신이 걸려 있는 시장 가판대가 위치하여 레퍼런스의 야외 시장 공간과 정확히 일치합니다. 인물과 배경 구조물 간의 거리감 및 공간적 맥락이 지시대로 잘 표현되었습니다.",
        "entities": "안규홍의 얼굴, 상투, 더러워진 복장은 레퍼런스와 일치합니다. 그러나 방어에 사용된 무기는 낫의 날이 긴 장대 중간에 붙어있는 기형적인 창 형태로 변형되었습니다. 또한 프레임 밖에 있어야 할 공격자의 두 손과 팔이 명확히 등장합니다.",
        "hard_violations": [
         "낫이 긴 장대 중간에 날이 달린 형태의 전혀 다른 무기로 창조됨 (Invented objects)",
         "공격자의 손과 팔이 화면에 등장하여 'wielder remains outside frame' 카메라 지시 위반",
         "지시문에 명시되지 않은 배경 인물들이 화면 양쪽에 등장함"
        ],
        "physics": "양손으로 하나의 긴 자루를 단단히 쥐고 떨어지는 칼을 막아내는 자세와 충돌의 물리적 저항감은 안정적으로 구현되었으며 신체 지지도 자연스럽습니다."
       },
       {
        "label": "B",
        "direction": "안규홍의 시선은 우측 상단의 공격자를 향해 올바르게 고정되어 있습니다. 공격자의 칼은 우측 상단에서 들어와 수평으로 든 막대에 부딪힙니다.",
        "built_space": "짚신이 걸린 가판대와 야외 시장 골목의 배경 구조물들이 레퍼런스에 맞게 적절히 배치되어 있습니다.",
        "entities": "안규홍의 외형과 복장은 잘 묘사되었습니다. 그러나 손에 쥔 무기는 낫과 단순한 나무 막대가 겹쳐지거나 분리된 알 수 없는 형태입니다. 우측 가장자리에 공격자의 얼굴 일부가 등장합니다.",
        "hard_violations": [
         "하나의 낫자루가 아닌 분리된 막대와 낫을 들고 있으며 낫의 날이 막대와 비정상적으로 겹치는 등 물리적으로 불가능한 무기 구조 발생",
         "화면 우측 가장자리에 공격자의 얼굴이 등장하여 'wielder remains outside frame' 지시 및 인물 등장 제한 위반",
         "하나의 낫자루가 안규홍의 얼굴 윗부분을 대각선으로 가로지른다는 핵심 구도 지시(Runs obliquely across)를 완전히 무시하고 수평으로 막아냄",
         "지시문에 명시되지 않은 다수의 배경 인물 등장"
        ],
        "physics": "왼손에 쥔 막대와 오른손에 쥔 낫의 자루가 별개로 분리되어 허공에서 비정상적으로 겹쳐 있으며, 양손이 결합되어 방어의 충격을 지탱한다는 물리적 구조가 붕괴되었습니다."
       }
      ],
      "all_candidates_fail": false
     },
     "normalized": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 4,
        "verdict_ko": "지시된 낫자루의 대각선 방어 구도와 안규홍의 외형을 훌륭하게 구현했으나, 낫이 긴 장대처럼 왜곡되고 공격자의 손이 화면에 포함되어 감점되었습니다."
       },
       {
        "label": "B",
        "score": 2,
        "verdict_ko": "무기가 두 개의 막대로 분리되는 심각한 물리적 구조 오류가 발생했으며, 지시된 대각선 방어 프레이밍을 전혀 따르지 않았습니다."
       }
      ],
      "all_candidates_fail": false,
      "readings": [
       {
        "label": "A",
        "direction": "안규홍의 시선은 왼쪽을 향하고 있어 떨어지는 칼이나 공격자를 직접 보지 않고 있습니다. 하지만 공격자의 칼은 우측 상단에서 대각선 아래로 향하며 안규홍이 든 자루에 정확히 맞닿아 물리적 충돌 지점을 형성합니다.",
        "built_space": "배경에 짚신이 걸려 있는 시장 가판대가 위치하여 레퍼런스의 야외 시장 공간과 정확히 일치합니다. 인물과 배경 구조물 간의 거리감 및 공간적 맥락이 지시대로 잘 표현되었습니다.",
        "entities": "안규홍의 얼굴, 상투, 더러워진 복장은 레퍼런스와 일치합니다. 그러나 방어에 사용된 무기는 낫의 날이 긴 장대 중간에 붙어있는 기형적인 창 형태로 변형되었습니다. 또한 프레임 밖에 있어야 할 공격자의 두 손과 팔이 명확히 등장합니다.",
        "hard_violations": [
         "낫이 긴 장대 중간에 날이 달린 형태의 전혀 다른 무기로 창조됨 (Invented objects)",
         "공격자의 손과 팔이 화면에 등장하여 'wielder remains outside frame' 카메라 지시 위반",
         "지시문에 명시되지 않은 배경 인물들이 화면 양쪽에 등장함"
        ],
        "physics": "양손으로 하나의 긴 자루를 단단히 쥐고 떨어지는 칼을 막아내는 자세와 충돌의 물리적 저항감은 안정적으로 구현되었으며 신체 지지도 자연스럽습니다."
       },
       {
        "label": "B",
        "direction": "안규홍의 시선은 우측 상단의 공격자를 향해 올바르게 고정되어 있습니다. 공격자의 칼은 우측 상단에서 들어와 수평으로 든 막대에 부딪힙니다.",
        "built_space": "짚신이 걸린 가판대와 야외 시장 골목의 배경 구조물들이 레퍼런스에 맞게 적절히 배치되어 있습니다.",
        "entities": "안규홍의 외형과 복장은 잘 묘사되었습니다. 그러나 손에 쥔 무기는 낫과 단순한 나무 막대가 겹쳐지거나 분리된 알 수 없는 형태입니다. 우측 가장자리에 공격자의 얼굴 일부가 등장합니다.",
        "hard_violations": [
         "하나의 낫자루가 아닌 분리된 막대와 낫을 들고 있으며 낫의 날이 막대와 비정상적으로 겹치는 등 물리적으로 불가능한 무기 구조 발생",
         "화면 우측 가장자리에 공격자의 얼굴이 등장하여 'wielder remains outside frame' 지시 및 인물 등장 제한 위반",
         "하나의 낫자루가 안규홍의 얼굴 윗부분을 대각선으로 가로지른다는 핵심 구도 지시(Runs obliquely across)를 완전히 무시하고 수평으로 막아냄",
         "지시문에 명시되지 않은 다수의 배경 인물 등장"
        ],
        "physics": "왼손에 쥔 막대와 오른손에 쥔 낫의 자루가 별개로 분리되어 허공에서 비정상적으로 겹쳐 있으며, 양손이 결합되어 방어의 충격을 지탱한다는 물리적 구조가 붕괴되었습니다."
       }
      ]
     },
     "ok": true
    },
    {
     "model": "gpt-high",
     "order": "reverse",
     "display_to_canonical": {
      "A": "B",
      "B": "A"
     },
     "raw": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 2,
        "verdict_ko": "미디엄 숏과 이마 위 충돌은 보이지만, 양손이 같은 낫자루를 잡지 않고 별도 막대와 낫을 나눠 들며 금지된 주변 인물까지 등장한다."
       },
       {
        "label": "B",
        "score": 4,
        "verdict_ko": "양손으로 비스듬히 든 낫자루에 내려오는 칼을 막는 동작은 더 충실하지만, 공격자의 손·팔과 주변 인물 노출로 필수 인원 제한을 위반한다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "안규홍의 시선은 화면 오른쪽 위의 공격자 쪽을 향한다. 칼은 위쪽 오른편에서 왼쪽 아래로 들어와 이마 위 나무 막대에 끝부분이 닿는다. 충돌점은 눈을 가리지 않지만, 막대는 거의 수평이며 오른손의 낫자루와 같은 연속된 자루로 보이지 않는다. 낫날은 오른쪽 손잡이에서 위로 굽어 왼쪽 막대 끝 쪽을 향한다.",
        "built_space": "중앙에 목조 짚신 가판대 한 곳이 있고, 양옆 주기둥 두 개와 뒤쪽 선반·하단 진열대가 보인다. 위에는 여러 켤레의 짚신이 걸려 있고 아래에도 짚신이 놓여 있다. 왼쪽에는 시장 통로, 오른쪽에는 천을 늘어뜨린 인접 가판대가 넓게 보인다. 안규홍은 가판대 앞에 허리까지 잡히며, 기준 사진의 목재·짚신 재질은 이어지지만 시장 주변이 상당히 많이 노출된다. 통로 양끝의 봉쇄는 확인되지 않는다.",
        "entities": "중앙 인물은 검은 머리를 묶은 한국인 청년 남성으로 표현되며, 얼굴과 체격은 참조에 가깝다. 때 묻은 밝은 저고리, 허리끈, 드러난 목도 대체로 맞는다. 칼에는 참조와 유사한 금속 날과 날무늬가 보인다. 그러나 한 손에는 짧은 나무 손잡이의 낫, 다른 손에는 별도 긴 막대가 있어 요구된 하나의 낫자루를 양손으로 드는 구성과 다르다. 왼쪽 통로의 여러 사람, 오른쪽 뒤 인물, 오른쪽 가장자리 인물 일부와 공격자의 손·소매가 추가로 보인다. 읽을 수 있는 글자는 없다.",
        "hard_violations": [
         "안규홍 외의 여러 주변 인물과 공격자의 신체 일부가 노출되어, 안규홍만 보이고 칼을 휘두르는 사람은 화면 밖에 있어야 한다는 제한을 위반한다.",
         "낫과 별도로 긴 나무 막대를 추가하여, 하나의 낫자루를 양손으로 들어 막는 필수 소품·동작 구성을 바꿨다."
        ],
        "physics": "화면 왼쪽 손은 긴 막대를, 오른쪽 손은 낫의 짧은 손잡이를 실제로 잡고 있다. 공격 칼도 위쪽 가장자리의 손에 잡혀 있고 칼끝 부근은 막대에 닿아 있어 무기들이 무지지 상태로 떠 있지는 않다. 다만 두 손의 힘이 하나의 연속된 낫자루를 통해 충돌점으로 전달되는 방어 자세는 아니다. 하체는 프레임 밖이므로 발의 접지는 판정할 수 없으며, 상체 자체에 부유 징후는 없다."
       },
       {
        "label": "B",
        "direction": "안규홍은 오른쪽 위 칼자루와 공격자 방향을 올려다본다. 공격 칼은 위쪽 오른편에서 왼쪽 아래로 향하며, 머리 위에서 낫의 나무 자루와 교차해 접촉한다. 낫자루는 왼쪽 위에서 오른쪽 아래로 기울어 얼굴 상부 앞을 지나고, 충돌점은 두 눈보다 위에 있어 표정이 드러난다. 낫날은 왼쪽 위 프레임 밖으로 이어진다.",
        "built_space": "안규홍 뒤에 목조 짚신 가판대 한 곳이 있고, 전면 주기둥 두 개와 상부 가로대, 뒤쪽 선반, 하단 진열 면이 보인다. 왼쪽에는 짚신 여러 켤레, 오른쪽 기둥에는 걸린 주머니가 있으며 양옆에 인접 판매대가 일부 들어온다. 안규홍은 가판대 바로 앞에 상체 중심의 미디엄 숏으로 배치된다. 기준 사진의 어두운 목조 가판대와 짚신 판매 장소는 유지되지만, 진열과 주변 판매대의 세부 배치는 달라졌다. 통로 끝은 보이지 않아 봉쇄 여부는 확인할 수 없다.",
        "entities": "중앙 인물은 참조와 가까운 얼굴·체격의 한국인 청년 남성이며 검은 머리를 단단히 묶었다. 때 묻은 밝은 저고리, 맨목, 얼굴의 오염이 유지된다. 두 손이 동일한 나무 낫자루를 잡는다. 녹슨 곡선 낫날과 닳은 목재는 참조의 재질과 맞지만 자루는 참조보다 길다. 칼의 완만한 곡선, 금속 날, 장식된 코등이와 감은 손잡이는 참조에 가깝다. 다만 위쪽 오른편에 공격자의 손과 검은 소매가 있고 좌우 배경에도 각각 추가 인물이 보인다. 읽을 수 있는 글자는 없다.",
        "hard_violations": [
         "공격자의 손·팔이 화면 안에 들어와 칼을 휘두르는 사람은 화면 밖에 있어야 한다는 명시적 구도를 위반한다.",
         "좌우 배경에 안규홍 외의 인물 두 명이 보여 허용된 등장인물 제한을 위반한다."
        ],
        "physics": "안규홍의 두 손이 하나의 연속된 나무 자루를 단단히 잡고, 굽힌 팔이 머리 위 접촉점의 힘을 받는 자세다. 칼은 오른쪽 위 손에 잡혀 있으며 낫자루와 맞닿아 있어 내려치기를 막는 순간으로 해석 가능하다. 무기나 손에 지지 없는 부유는 보이지 않는다. 발은 프레임 밖이지만 상체의 기울기와 팔 자세는 지상에서 버티는 동작으로 가능하다."
       }
      ],
      "all_candidates_fail": true
     },
     "normalized": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 2,
        "verdict_ko": "미디엄 숏과 이마 위 충돌은 보이지만, 양손이 같은 낫자루를 잡지 않고 별도 막대와 낫을 나눠 들며 금지된 주변 인물까지 등장한다."
       },
       {
        "label": "A",
        "score": 4,
        "verdict_ko": "양손으로 비스듬히 든 낫자루에 내려오는 칼을 막는 동작은 더 충실하지만, 공격자의 손·팔과 주변 인물 노출로 필수 인원 제한을 위반한다."
       }
      ],
      "all_candidates_fail": true,
      "readings": [
       {
        "label": "B",
        "direction": "안규홍의 시선은 화면 오른쪽 위의 공격자 쪽을 향한다. 칼은 위쪽 오른편에서 왼쪽 아래로 들어와 이마 위 나무 막대에 끝부분이 닿는다. 충돌점은 눈을 가리지 않지만, 막대는 거의 수평이며 오른손의 낫자루와 같은 연속된 자루로 보이지 않는다. 낫날은 오른쪽 손잡이에서 위로 굽어 왼쪽 막대 끝 쪽을 향한다.",
        "built_space": "중앙에 목조 짚신 가판대 한 곳이 있고, 양옆 주기둥 두 개와 뒤쪽 선반·하단 진열대가 보인다. 위에는 여러 켤레의 짚신이 걸려 있고 아래에도 짚신이 놓여 있다. 왼쪽에는 시장 통로, 오른쪽에는 천을 늘어뜨린 인접 가판대가 넓게 보인다. 안규홍은 가판대 앞에 허리까지 잡히며, 기준 사진의 목재·짚신 재질은 이어지지만 시장 주변이 상당히 많이 노출된다. 통로 양끝의 봉쇄는 확인되지 않는다.",
        "entities": "중앙 인물은 검은 머리를 묶은 한국인 청년 남성으로 표현되며, 얼굴과 체격은 참조에 가깝다. 때 묻은 밝은 저고리, 허리끈, 드러난 목도 대체로 맞는다. 칼에는 참조와 유사한 금속 날과 날무늬가 보인다. 그러나 한 손에는 짧은 나무 손잡이의 낫, 다른 손에는 별도 긴 막대가 있어 요구된 하나의 낫자루를 양손으로 드는 구성과 다르다. 왼쪽 통로의 여러 사람, 오른쪽 뒤 인물, 오른쪽 가장자리 인물 일부와 공격자의 손·소매가 추가로 보인다. 읽을 수 있는 글자는 없다.",
        "hard_violations": [
         "안규홍 외의 여러 주변 인물과 공격자의 신체 일부가 노출되어, 안규홍만 보이고 칼을 휘두르는 사람은 화면 밖에 있어야 한다는 제한을 위반한다.",
         "낫과 별도로 긴 나무 막대를 추가하여, 하나의 낫자루를 양손으로 들어 막는 필수 소품·동작 구성을 바꿨다."
        ],
        "physics": "화면 왼쪽 손은 긴 막대를, 오른쪽 손은 낫의 짧은 손잡이를 실제로 잡고 있다. 공격 칼도 위쪽 가장자리의 손에 잡혀 있고 칼끝 부근은 막대에 닿아 있어 무기들이 무지지 상태로 떠 있지는 않다. 다만 두 손의 힘이 하나의 연속된 낫자루를 통해 충돌점으로 전달되는 방어 자세는 아니다. 하체는 프레임 밖이므로 발의 접지는 판정할 수 없으며, 상체 자체에 부유 징후는 없다."
       },
       {
        "label": "A",
        "direction": "안규홍은 오른쪽 위 칼자루와 공격자 방향을 올려다본다. 공격 칼은 위쪽 오른편에서 왼쪽 아래로 향하며, 머리 위에서 낫의 나무 자루와 교차해 접촉한다. 낫자루는 왼쪽 위에서 오른쪽 아래로 기울어 얼굴 상부 앞을 지나고, 충돌점은 두 눈보다 위에 있어 표정이 드러난다. 낫날은 왼쪽 위 프레임 밖으로 이어진다.",
        "built_space": "안규홍 뒤에 목조 짚신 가판대 한 곳이 있고, 전면 주기둥 두 개와 상부 가로대, 뒤쪽 선반, 하단 진열 면이 보인다. 왼쪽에는 짚신 여러 켤레, 오른쪽 기둥에는 걸린 주머니가 있으며 양옆에 인접 판매대가 일부 들어온다. 안규홍은 가판대 바로 앞에 상체 중심의 미디엄 숏으로 배치된다. 기준 사진의 어두운 목조 가판대와 짚신 판매 장소는 유지되지만, 진열과 주변 판매대의 세부 배치는 달라졌다. 통로 끝은 보이지 않아 봉쇄 여부는 확인할 수 없다.",
        "entities": "중앙 인물은 참조와 가까운 얼굴·체격의 한국인 청년 남성이며 검은 머리를 단단히 묶었다. 때 묻은 밝은 저고리, 맨목, 얼굴의 오염이 유지된다. 두 손이 동일한 나무 낫자루를 잡는다. 녹슨 곡선 낫날과 닳은 목재는 참조의 재질과 맞지만 자루는 참조보다 길다. 칼의 완만한 곡선, 금속 날, 장식된 코등이와 감은 손잡이는 참조에 가깝다. 다만 위쪽 오른편에 공격자의 손과 검은 소매가 있고 좌우 배경에도 각각 추가 인물이 보인다. 읽을 수 있는 글자는 없다.",
        "hard_violations": [
         "공격자의 손·팔이 화면 안에 들어와 칼을 휘두르는 사람은 화면 밖에 있어야 한다는 명시적 구도를 위반한다.",
         "좌우 배경에 안규홍 외의 인물 두 명이 보여 허용된 등장인물 제한을 위반한다."
        ],
        "physics": "안규홍의 두 손이 하나의 연속된 나무 자루를 단단히 잡고, 굽힌 팔이 머리 위 접촉점의 힘을 받는 자세다. 칼은 오른쪽 위 손에 잡혀 있으며 낫자루와 맞닿아 있어 내려치기를 막는 순간으로 해석 가능하다. 무기나 손에 지지 없는 부유는 보이지 않는다. 발은 프레임 밖이지만 상체의 기울기와 팔 자세는 지상에서 버티는 동작으로 가능하다."
       }
      ]
     },
     "ok": true
    }
   ],
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "slot_winner_match": true,
   "slot_winner": {
    "gemini-pro": "A",
    "gpt-high": "A"
   },
   "route": "cross_slot_agree"
  },
  "dual": {
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "normalized": {
    "A": 2.0,
    "B": 1.0
   },
   "adjusted": {
    "A": 1.75,
    "B": 0.75
   },
   "violations": {
    "A": [
     "[gemini-pro] 낫이 긴 장대 중간에 날이 달린 형태의 전혀 다른 무기로 창조됨 (Invented objects)",
     "[gemini-pro] 공격자의 손과 팔이 화면에 등장하여 'wielder remains outside frame' 카메라 지시 위반",
     "[gemini-pro] 지시문에 명시되지 않은 배경 인물들이 화면 양쪽에 등장함",
     "[gpt-high] 공격자의 손·팔이 화면 안에 들어와 칼을 휘두르는 사람은 화면 밖에 있어야 한다는 명시적 구도를 위반한다.",
     "[gpt-high] 좌우 배경에 안규홍 외의 인물 두 명이 보여 허용된 등장인물 제한을 위반한다."
    ],
    "B": [
     "[gemini-pro] 하나의 낫자루가 아닌 분리된 막대와 낫을 들고 있으며 낫의 날이 막대와 비정상적으로 겹치는 등 물리적으로 불가능한 무기 구조 발생",
     "[gemini-pro] 화면 우측 가장자리에 공격자의 얼굴이 등장하여 'wielder remains outside frame' 지시 및 인물 등장 제한 위반",
     "[gemini-pro] 하나의 낫자루가 안규홍의 얼굴 윗부분을 대각선으로 가로지른다는 핵심 구도 지시(Runs obliquely across)를 완전히 무시하고 수평으로 막아냄",
     "[gemini-pro] 지시문에 명시되지 않은 다수의 배경 인물 등장",
     "[gpt-high] 안규홍 외의 여러 주변 인물과 공격자의 신체 일부가 노출되어, 안규홍만 보이고 칼을 휘두르는 사람은 화면 밖에 있어야 한다는 제한을 위반한다.",
     "[gpt-high] 낫과 별도로 긴 나무 막대를 추가하여, 하나의 낫자루를 양손으로 들어 막는 필수 소품·동작 구성을 바꿨다."
    ]
   },
   "per_model_winner": {
    "gemini-pro": "A",
    "gpt-high": "A"
   },
   "agreed": true
  },
  "totals": {
   "A": 1750,
   "B": 750
  },
  "selected": "A",
  "ranking": [
   "A",
   "B"
  ],
  "verdicts": [
   {
    "label": "A",
    "score": 1750,
    "verdict_ko": "지시된 낫자루의 대각선 방어 구도와 안규홍의 외형을 훌륭하게 구현했으나, 낫이 긴 장대처럼 왜곡되고 공격자의 손이 화면에 포함되어 감점되었습니다.  ★위반: [gemini-pro] 낫이 긴 장대 중간에 날이 달린 형태의 전혀 다른 무기로 창조됨 (Invented objects) / [gemini-pro] 공격자의 손과 팔이 화면에 등장하여 'wielder remains outside frame' 카메라 지시 위반 / [gemini-pro] 지시문에 명시되지 않은 배경 인물들이 화면 양쪽에 등장함 / [gpt-high] 공격자의 손·팔이 화면 안에 들어와 칼을 휘두르는 사람은 화면 밖에 있어야 한다는 명시적 구도를 위반한다. / [gpt-high] 좌우 배경에 안규홍 외의 인물 두 명이 보여 허용된 등장인물 제한을 위반한다."
   },
   {
    "label": "B",
    "score": 750,
    "verdict_ko": "무기가 두 개의 막대로 분리되는 심각한 물리적 구조 오류가 발생했으며, 지시된 대각선 방어 프레이밍을 전혀 따르지 않았습니다.  ★위반: [gemini-pro] 하나의 낫자루가 아닌 분리된 막대와 낫을 들고 있으며 낫의 날이 막대와 비정상적으로 겹치는 등 물리적으로 불가능한 무기 구조 발생 / [gemini-pro] 화면 우측 가장자리에 공격자의 얼굴이 등장하여 'wielder remains outside frame' 지시 및 인물 등장 제한 위반 / [gemini-pro] 하나의 낫자루가 안규홍의 얼굴 윗부분을 대각선으로 가로지른다는 핵심 구도 지시(Runs obliquely across)를 완전히 무시하고 수평으로 막아냄 / [gemini-pro] 지시문에 명시되지 않은 다수의 배경 인물 등장 / [gpt-high] 안규홍 외의 여러 주변 인물과 공격자의 신체 일부가 노출되어, 안규홍만 보이고 칼을 휘두르는 사람은 화면 밖에 있어야 한다는 제한을 위반한다. / [gpt-high] 낫과 별도로 긴 나무 막대를 추가하여, 하나의 낫자루를 양손으로 들어 막는 필수 소품·동작 구성을 바꿨다."
   }
  ],
  "refs": [
   {
    "label": "PREVIOUS SHOT STILL — a visually related earlier shot of this same place: the location's look, materials, fixed features and lighting mood are LOCKED to this photo; never copy its camera framing. The clothing, hair and overall look of 안규홍 — who appear both in that photo and in this shot — are LOCKED to that photo. Anyone else visible in that photo is NOT in this shot: never carry their face, body or clothing onto anyone here. If this photo shows a character who cannot move (dead or unconscious), that character's exact pose, position and orientation are ALSO LOCKED — treat their body as a fixed prop of the set that only the camera moves around.",
    "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/db00517b-165b-4005-a287-f28936b59b95/images/272bd7af-9294-4490-98c6-88a40e9deacc/scene/recipe/S48sh23_sel.png",
    "asset_id": "42779b0a-279e-4c40-bfe8-c2a14ac603cf",
    "role": "prev_still"
   },
   {
    "label": "CHARACTER REFERENCE — 안규홍: the exact person appearing in this shot; match face, hair and build exactly.",
    "path": "<bytes:1028384>",
    "asset_id": "3b5894db-ce55-4f34-b966-5e725c3aa952",
    "role": "character_ref"
   },
   {
    "label": "PROP REFERENCE — 칼: the exact object appearing in this shot; match its look, material and wear exactly.",
    "path": "<bytes:780122>",
    "asset_id": "82614014-6091-423c-89ca-1e15d209de72",
    "role": "prop_ref"
   },
   {
    "label": "PROP REFERENCE — 낫: the exact object appearing in this shot; match its look, material and wear exactly.",
    "path": "<bytes:977465>",
    "asset_id": "e9884693-b2aa-4769-9ee0-4c7b50db40c2",
    "role": "prop_ref"
   }
  ],
  "critique_skipped": true,
  "shot_run_produced": false,
  "shot_run_uid": "06aa1435-63f5-7caf-89c6-29ddcdbcf339",
  "ref_mode": "prev+엔티티",
  "share_plan": {
   "ref_plan": "prev",
   "prev_anchor_tag": "S48sh23"
  }
 },
 "S51sh2::cine": {
  "applied": false,
  "attempted_at": "2026-09-09T09:31:31.163338+00:00",
  "fingerprint": "84f394b647ea48819afca361d8963f6ee16d2dc648a059ea6aa782f1bea8fea8",
  "fingerprint_version": 2,
  "provider": "mai",
  "endpoint": "openrouter/chat-completions",
  "model": "microsoft/mai-image-2.6",
  "multiroll_tag": "still_S51sh2_cine",
  "slot": "",
  "pack": "24.202608252115",
  "source_file": "S51sh2_sel.png",
  "source_sha256": "56c4af61362044bd408de3a9acef5769e52629ae40dc03cb672348e05dd385e0",
  "error": "RuntimeError: moderation blocked: {\"message\": \"Response content blocked by label 'MultiSeverity_ViolenceScore'.\", \"code\": 400}",
  "moderation_refusals": 2,
  "declined": true,
  "declined_reason": "moderation"
 },
 "S51sh9::signage": {
  "fp": "d7912b957af22ec3",
  "inscriptions": [],
  "cues": [],
  "dropped": []
 },
 "S51sh9": {
  "input_fingerprint": "8a36c4440b974a05",
  "prompt": "Create ONE FINAL photorealistic live-action film still of the moment below — early 20th-century Korea, Jeolla region; all people are Korean unless explicitly stated otherwise. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 안규홍의 등 뒤로 날카로운 칼날이 스윽 긋고 지나가는 찰나.\n\nLOCATION (lock): The dirt fighting strip beside the market stalls, with attackers closing in from both directions. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: close-up\n- KEY BACKGROUND ELEMENTS: Rear attacker's sword (Passing across 안규홍's back and cutting him) — Seen obliquely across his upper back, with the wielder outside the frame; used as Provides the decisive interruption across the exposed torso without showing an additional attacker; Sickle handle (Still engaged in blocking the attack in front) — Partially visible beyond 안규홍's forward shoulder and raised arm; used as Preserves the simultaneous frontal defense that explains his rear vulnerability; Market passage (Combat continues around the tightly framed figure) — A narrow strip of depth remains beside the torso; used as Keeps the close contact shot situated within the ongoing fight.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Retain the fight's daylight exposure and sober natural contrast, conveying the wound through contact and recoil rather than added gore or lighting effects.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The clothing, hair and overall look of 안규홍 — who appear both in that photo and in this shot — are LOCKED to that photo. Anyone else visible in that photo is NOT in this shot: never carry their face, body or clothing onto anyone here. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): A stall has been struck and disrupted during the fight; a dropped sword lies on the ground. 안규홍: A fresh slash opens across his back beneath his jeogori; he retains his sickle and bare neck.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 안규홍 (한국인 남성, 청년 나이대, 검은 머리, 단단히 묶은 머리) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.",
  "roll_prompts": {
   "A": "Create ONE FINAL photorealistic live-action film still of the moment below — early 20th-century Korea, Jeolla region; all people are Korean unless explicitly stated otherwise. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 안규홍의 등 뒤로 날카로운 칼날이 스윽 긋고 지나가는 찰나.\n\nLOCATION (lock): The dirt fighting strip beside the market stalls, with attackers closing in from both directions. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: close-up\n- KEY BACKGROUND ELEMENTS: Rear attacker's sword (Passing across 안규홍's back and cutting him) — Seen obliquely across his upper back, with the wielder outside the frame; used as Provides the decisive interruption across the exposed torso without showing an additional attacker; Sickle handle (Still engaged in blocking the attack in front) — Partially visible beyond 안규홍's forward shoulder and raised arm; used as Preserves the simultaneous frontal defense that explains his rear vulnerability; Market passage (Combat continues around the tightly framed figure) — A narrow strip of depth remains beside the torso; used as Keeps the close contact shot situated within the ongoing fight.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Retain the fight's daylight exposure and sober natural contrast, conveying the wound through contact and recoil rather than added gore or lighting effects.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The clothing, hair and overall look of 안규홍 — who appear both in that photo and in this shot — are LOCKED to that photo. Anyone else visible in that photo is NOT in this shot: never carry their face, body or clothing onto anyone here. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): A stall has been struck and disrupted during the fight; a dropped sword lies on the ground. 안규홍: A fresh slash opens across his back beneath his jeogori; he retains his sickle and bare neck.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 안규홍 (한국인 남성, 청년 나이대, 검은 머리, 단단히 묶은 머리) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.",
   "B": "Create ONE FINAL photorealistic live-action film still of the moment below — early 20th-century Korea, Jeolla region; all people are Korean unless explicitly stated otherwise. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 안규홍의 등 뒤로 날카로운 칼날이 스윽 긋고 지나가는 찰나.\n\nLOCATION (lock): The dirt fighting strip beside the market stalls, with attackers closing in from both directions. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: close-up\n- KEY BACKGROUND ELEMENTS: Rear attacker's sword (Passing across 안규홍's back and cutting him) — Seen obliquely across his upper back, with the wielder outside the frame; used as Provides the decisive interruption across the exposed torso without showing an additional attacker; Sickle handle (Still engaged in blocking the attack in front) — Partially visible beyond 안규홍's forward shoulder and raised arm; used as Preserves the simultaneous frontal defense that explains his rear vulnerability; Market passage (Combat continues around the tightly framed figure) — A narrow strip of depth remains beside the torso; used as Keeps the close contact shot situated within the ongoing fight.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Retain the fight's daylight exposure and sober natural contrast, conveying the wound through contact and recoil rather than added gore or lighting effects.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The clothing, hair and overall look of 안규홍 — who appear both in that photo and in this shot — are LOCKED to that photo. Anyone else visible in that photo is NOT in this shot: never carry their face, body or clothing onto anyone here. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): A stall has been struck and disrupted during the fight; a dropped sword lies on the ground. 안규홍: A fresh slash opens across his back beneath his jeogori; he retains his sickle and bare neck.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 안규홍 (한국인 남성, 청년 나이대, 검은 머리, 단단히 묶은 머리) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.\n\nCOMPOSITION VARIATION (this candidate only): keep the place, the people and the moment exactly as specified, and keep from the CAMERA & FRAME contract its subject scale, screen placement and key background elements exactly. For this candidate only, its camera angle and height are deliberately left open: choose a camera position distinctly different from the obvious one, as a film director picking a second setup on the same blocking. Stay true to any explicit framing the SHOT TEXT itself dictates."
  },
  "shot_run_spend_attempt_count": 1,
  "readings": [
   {
    "label": "A",
    "direction": "안규홍은 고개를 돌려 왼쪽 어깨 너머를 바라보고 있으며, 화면 왼쪽에서 등장한 손이 쥔 칼이 안규홍의 등을 향해 꽂혀 있습니다.",
    "built_space": "짚신이 매달린 나무 가판대와 흙바닥 등 레퍼런스의 시장 배경이 올바르게 구현되었으며, 좁은 깊이감을 보여줍니다.",
    "entities": "안규홍의 외모, 체형, 복장(더러워진 홑옷)이 레퍼런스와 일치합니다. 프롬프트가 요구한 공격자의 칼(소품 레퍼런스 일치)과 안규홍이 쥔 낫 자루가 모두 존재합니다.",
    "hard_violations": [
     "[gpt-high] 왼쪽 화면 안에 공격자의 손과 짙은 소매를 추가했다. 이는 칼을 휘두르는 사람을 화면 밖에 두고 안규홍 외의 인물이나 신체 일부를 노출하지 말라는 명시적 구도·인물 제한을 위반한다."
    ],
    "physics": "공격자의 손이 칼자루를 단단히 쥐고 있으며, 안규홍의 오른손 역시 낫 자루를 지탱하고 있습니다. 다만 칼날이 표면을 스치는 것이 아니라 몸을 관통하듯 꽂혀 있습니다."
   },
   {
    "label": "B",
    "direction": "안규홍은 화면 왼쪽을 향해 시선을 고정하고 방어적인 자세를 취하고 있습니다.",
    "built_space": "나무 가판대, 바구니, 흙바닥 등 시장 배경의 구조물들이 묘사되어 있습니다.",
    "entities": "안규홍의 인상착의는 레퍼런스와 일치하며 낫 자루를 쥐고 있으나, 명시된 핵심 요소인 '등 뒤 공격자의 칼'이 존재하지 않습니다(바닥에 떨어진 칼만 존재).",
    "hard_violations": [],
    "physics": "안규홍의 두 손이 낫 자루를 쥐는 자세를 취하고 있으며, 바닥에 놓인 칼은 지면에 지탱되어 있습니다."
   }
  ],
  "cross_model_order": {
   "policy": "cross_model_order_v2_slot0_fwd_slot1_rev_parallel2",
   "slots": [
    {
     "model": "gemini-pro",
     "order": "forward",
     "display_to_canonical": {
      "A": "A",
      "B": "B"
     },
     "raw": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 6,
        "verdict_ko": "프롬프트가 요구한 등 뒤의 칼 공격과 낫을 이용한 전면 방어 자세를 모두 화면에 담아냈으나, 칼날이 등을 베고 지나가는(스윽 긋는) 형태가 아니라 등 안으로 찔러 들어간 것처럼 묘사된 점이 아쉽습니다."
       },
       {
        "label": "B",
        "score": 2,
        "verdict_ko": "프롬프트의 핵심 액션인 '등 뒤로 칼날이 긋고 지나가는 찰나'와 공격자의 칼 자체를 화면에서 완전히 누락하여 연출 지시를 실패했습니다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "안규홍은 고개를 돌려 왼쪽 어깨 너머를 바라보고 있으며, 화면 왼쪽에서 등장한 손이 쥔 칼이 안규홍의 등을 향해 꽂혀 있습니다.",
        "built_space": "짚신이 매달린 나무 가판대와 흙바닥 등 레퍼런스의 시장 배경이 올바르게 구현되었으며, 좁은 깊이감을 보여줍니다.",
        "entities": "안규홍의 외모, 체형, 복장(더러워진 홑옷)이 레퍼런스와 일치합니다. 프롬프트가 요구한 공격자의 칼(소품 레퍼런스 일치)과 안규홍이 쥔 낫 자루가 모두 존재합니다.",
        "hard_violations": [],
        "physics": "공격자의 손이 칼자루를 단단히 쥐고 있으며, 안규홍의 오른손 역시 낫 자루를 지탱하고 있습니다. 다만 칼날이 표면을 스치는 것이 아니라 몸을 관통하듯 꽂혀 있습니다."
       },
       {
        "label": "B",
        "direction": "안규홍은 화면 왼쪽을 향해 시선을 고정하고 방어적인 자세를 취하고 있습니다.",
        "built_space": "나무 가판대, 바구니, 흙바닥 등 시장 배경의 구조물들이 묘사되어 있습니다.",
        "entities": "안규홍의 인상착의는 레퍼런스와 일치하며 낫 자루를 쥐고 있으나, 명시된 핵심 요소인 '등 뒤 공격자의 칼'이 존재하지 않습니다(바닥에 떨어진 칼만 존재).",
        "hard_violations": [],
        "physics": "안규홍의 두 손이 낫 자루를 쥐는 자세를 취하고 있으며, 바닥에 놓인 칼은 지면에 지탱되어 있습니다."
       }
      ],
      "all_candidates_fail": false
     },
     "normalized": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 6,
        "verdict_ko": "프롬프트가 요구한 등 뒤의 칼 공격과 낫을 이용한 전면 방어 자세를 모두 화면에 담아냈으나, 칼날이 등을 베고 지나가는(스윽 긋는) 형태가 아니라 등 안으로 찔러 들어간 것처럼 묘사된 점이 아쉽습니다."
       },
       {
        "label": "B",
        "score": 2,
        "verdict_ko": "프롬프트의 핵심 액션인 '등 뒤로 칼날이 긋고 지나가는 찰나'와 공격자의 칼 자체를 화면에서 완전히 누락하여 연출 지시를 실패했습니다."
       }
      ],
      "all_candidates_fail": false,
      "readings": [
       {
        "label": "A",
        "direction": "안규홍은 고개를 돌려 왼쪽 어깨 너머를 바라보고 있으며, 화면 왼쪽에서 등장한 손이 쥔 칼이 안규홍의 등을 향해 꽂혀 있습니다.",
        "built_space": "짚신이 매달린 나무 가판대와 흙바닥 등 레퍼런스의 시장 배경이 올바르게 구현되었으며, 좁은 깊이감을 보여줍니다.",
        "entities": "안규홍의 외모, 체형, 복장(더러워진 홑옷)이 레퍼런스와 일치합니다. 프롬프트가 요구한 공격자의 칼(소품 레퍼런스 일치)과 안규홍이 쥔 낫 자루가 모두 존재합니다.",
        "hard_violations": [],
        "physics": "공격자의 손이 칼자루를 단단히 쥐고 있으며, 안규홍의 오른손 역시 낫 자루를 지탱하고 있습니다. 다만 칼날이 표면을 스치는 것이 아니라 몸을 관통하듯 꽂혀 있습니다."
       },
       {
        "label": "B",
        "direction": "안규홍은 화면 왼쪽을 향해 시선을 고정하고 방어적인 자세를 취하고 있습니다.",
        "built_space": "나무 가판대, 바구니, 흙바닥 등 시장 배경의 구조물들이 묘사되어 있습니다.",
        "entities": "안규홍의 인상착의는 레퍼런스와 일치하며 낫 자루를 쥐고 있으나, 명시된 핵심 요소인 '등 뒤 공격자의 칼'이 존재하지 않습니다(바닥에 떨어진 칼만 존재).",
        "hard_violations": [],
        "physics": "안규홍의 두 손이 낫 자루를 쥐는 자세를 취하고 있으며, 바닥에 놓인 칼은 지면에 지탱되어 있습니다."
       }
      ]
     },
     "ok": true
    },
    {
     "model": "gpt-high",
     "order": "reverse",
     "display_to_canonical": {
      "A": "B",
      "B": "A"
     },
     "raw": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 4,
        "verdict_ko": "안규홍만 보이는 등 중심 근접 구도와 장소·복장은 맞지만, 등을 베고 지나가는 칼날과 새 상처가 없어 결정적인 순간을 구현하지 못했다."
       },
       {
        "label": "B",
        "score": 2,
        "verdict_ko": "칼날의 등 접촉과 상처는 보이지만, 반드시 화면 밖에 있어야 할 공격자의 손과 소매를 노출한 중대 위반으로 탈락한다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "안규홍은 고개를 왼쪽으로 돌려 화면 밖을 바라본다. 오른손의 나무 손잡이는 오른쪽 위로 들려 있지만 전방 공격을 막는 접촉점은 보이지 않는다. 등을 향하거나 가로지르는 공격 칼날은 없으며, 오른쪽 아래 칼은 땅에 놓여 있을 뿐이다.",
        "built_space": "등과 어깨가 화면 대부분을 차지하고 오른쪽에 흙 통로가 남는다. 왼쪽에는 목재 좌판 하나와 선반, 기둥에 매달린 바구니 하나가 보이고, 아래에는 손잡이 바구니 하나와 겹쳐진 얕은 바구니 두 개가 보인다. 뒤쪽에도 목재 가판과 물품이 이어진다. 참고의 낡은 목재·직물·바구니 시장과 대체로 이어지며, 고정 시설의 명백한 중복이나 불가능한 반사는 없다.",
        "entities": "보이는 사람은 안규홍 한 명이다. 청년 한국인 남성 설정에 부합하는 외형, 검은 묶음머리, 참고와 유사한 옆얼굴과 체격, 때 묻은 밝은 저고리, 장신구 없는 목이 확인된다. 오른손에는 낫으로 추정되는 나무 손잡이가 있으나 날은 화면 밖이다. 떨어진 칼은 흙바닥에 보인다. 뒤에서 베는 칼과 저고리 아래 새로 열린 등 상처는 보이지 않는다. 좌판 주변은 어수선하지만 충격으로 파손된 상태인지는 명확하지 않다. 읽을 수 있는 글자는 없다.",
        "hard_violations": [],
        "physics": "오른손이 나무 손잡이를 실제로 감싸 쥐며, 바닥의 칼과 바구니는 흙바닥에 지지된다. 매달린 바구니는 끈으로 지지된다. 하체는 근접 구도 밖이므로 발의 접지는 확인할 수 없지만 몸이 공중에 떠 있는 모습은 아니다. 팔을 벌리고 몸통을 튼 자세는 가능하나, 베임의 접촉이나 그 충격에 따른 반동은 확인되지 않는다."
       },
       {
        "label": "B",
        "direction": "왼쪽 가장자리에서 들어온 칼은 오른쪽 아래로 비스듬히 향하고 끝부분이 안규홍의 등 왼쪽에 닿는다. 그 접촉점 아래에 짧은 붉은 상처가 있어 공격 대상은 정확히 등으로 읽힌다. 다만 등 위를 길게 스쳐 가는 동작보다는 칼끝으로 찍거나 짧게 긋는 모습에 가깝다. 안규홍의 눈은 뒤로 돌아 카메라 쪽을 향하며, 오른쪽 위로 든 낫에는 전방 공격과의 접촉이 보이지 않는다.",
        "built_space": "안규홍의 상체 뒤로 좁은 흙 통로와 양쪽 목재 가판이 보인다. 왼쪽에는 처마를 받치는 기둥과 사선 버팀목, 매달린 바구니 하나, 아래쪽의 긴 직조 물품이 보인다. 오른쪽에는 그릇이 놓인 좌판 하나와 매달린 짚신 두 짝이 보인다. 참고와 시장 재료 및 낮의 조명은 유사하지만 세부 배치를 정확히 대조하기는 어렵다. 구조 중복이나 반사 문제는 없다.",
        "entities": "안규홍의 청년 남성 외형, 검은 묶음머리, 얼굴과 체격, 낡고 밝은 저고리, 맨목은 참고와 대체로 맞는다. 오른손에는 나무 자루와 굽은 금속 날을 갖춘 낫이 있다. 공격 칼에는 금속 날과 황동색 장식이 보이지만 참고보다 칼끝이 각지고 날이 곧아 보인다. 등에 작은 절개와 피가 있다. 그러나 왼쪽 가장자리에 별도 공격자의 손과 짙은 소매가 나타난다. 바닥에 떨어진 칼과 좌판 파손은 이 구도에서 확인되지 않으며, 글자는 없다.",
        "hard_violations": [
         "왼쪽 화면 안에 공격자의 손과 짙은 소매를 추가했다. 이는 칼을 휘두르는 사람을 화면 밖에 두고 안규홍 외의 인물이나 신체 일부를 노출하지 말라는 명시적 구도·인물 제한을 위반한다."
        ],
        "physics": "낫은 안규홍의 오른손이, 공격 칼은 왼쪽 가장자리의 다른 사람 손이 지지하므로 무기가 떠 있지는 않다. 칼끝과 저고리의 상처가 맞닿아 있어 물리적 접촉은 읽힌다. 몸통을 비틀고 뒤돌아보는 자세도 가능하지만, 전방 방어가 유지되는 접촉이나 베임에 따른 뚜렷한 반동은 보이지 않는다. 발은 화면 밖이며 공중 부양의 증거는 없다."
       }
      ],
      "all_candidates_fail": true
     },
     "normalized": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 4,
        "verdict_ko": "안규홍만 보이는 등 중심 근접 구도와 장소·복장은 맞지만, 등을 베고 지나가는 칼날과 새 상처가 없어 결정적인 순간을 구현하지 못했다."
       },
       {
        "label": "A",
        "score": 2,
        "verdict_ko": "칼날의 등 접촉과 상처는 보이지만, 반드시 화면 밖에 있어야 할 공격자의 손과 소매를 노출한 중대 위반으로 탈락한다."
       }
      ],
      "all_candidates_fail": true,
      "readings": [
       {
        "label": "B",
        "direction": "안규홍은 고개를 왼쪽으로 돌려 화면 밖을 바라본다. 오른손의 나무 손잡이는 오른쪽 위로 들려 있지만 전방 공격을 막는 접촉점은 보이지 않는다. 등을 향하거나 가로지르는 공격 칼날은 없으며, 오른쪽 아래 칼은 땅에 놓여 있을 뿐이다.",
        "built_space": "등과 어깨가 화면 대부분을 차지하고 오른쪽에 흙 통로가 남는다. 왼쪽에는 목재 좌판 하나와 선반, 기둥에 매달린 바구니 하나가 보이고, 아래에는 손잡이 바구니 하나와 겹쳐진 얕은 바구니 두 개가 보인다. 뒤쪽에도 목재 가판과 물품이 이어진다. 참고의 낡은 목재·직물·바구니 시장과 대체로 이어지며, 고정 시설의 명백한 중복이나 불가능한 반사는 없다.",
        "entities": "보이는 사람은 안규홍 한 명이다. 청년 한국인 남성 설정에 부합하는 외형, 검은 묶음머리, 참고와 유사한 옆얼굴과 체격, 때 묻은 밝은 저고리, 장신구 없는 목이 확인된다. 오른손에는 낫으로 추정되는 나무 손잡이가 있으나 날은 화면 밖이다. 떨어진 칼은 흙바닥에 보인다. 뒤에서 베는 칼과 저고리 아래 새로 열린 등 상처는 보이지 않는다. 좌판 주변은 어수선하지만 충격으로 파손된 상태인지는 명확하지 않다. 읽을 수 있는 글자는 없다.",
        "hard_violations": [],
        "physics": "오른손이 나무 손잡이를 실제로 감싸 쥐며, 바닥의 칼과 바구니는 흙바닥에 지지된다. 매달린 바구니는 끈으로 지지된다. 하체는 근접 구도 밖이므로 발의 접지는 확인할 수 없지만 몸이 공중에 떠 있는 모습은 아니다. 팔을 벌리고 몸통을 튼 자세는 가능하나, 베임의 접촉이나 그 충격에 따른 반동은 확인되지 않는다."
       },
       {
        "label": "A",
        "direction": "왼쪽 가장자리에서 들어온 칼은 오른쪽 아래로 비스듬히 향하고 끝부분이 안규홍의 등 왼쪽에 닿는다. 그 접촉점 아래에 짧은 붉은 상처가 있어 공격 대상은 정확히 등으로 읽힌다. 다만 등 위를 길게 스쳐 가는 동작보다는 칼끝으로 찍거나 짧게 긋는 모습에 가깝다. 안규홍의 눈은 뒤로 돌아 카메라 쪽을 향하며, 오른쪽 위로 든 낫에는 전방 공격과의 접촉이 보이지 않는다.",
        "built_space": "안규홍의 상체 뒤로 좁은 흙 통로와 양쪽 목재 가판이 보인다. 왼쪽에는 처마를 받치는 기둥과 사선 버팀목, 매달린 바구니 하나, 아래쪽의 긴 직조 물품이 보인다. 오른쪽에는 그릇이 놓인 좌판 하나와 매달린 짚신 두 짝이 보인다. 참고와 시장 재료 및 낮의 조명은 유사하지만 세부 배치를 정확히 대조하기는 어렵다. 구조 중복이나 반사 문제는 없다.",
        "entities": "안규홍의 청년 남성 외형, 검은 묶음머리, 얼굴과 체격, 낡고 밝은 저고리, 맨목은 참고와 대체로 맞는다. 오른손에는 나무 자루와 굽은 금속 날을 갖춘 낫이 있다. 공격 칼에는 금속 날과 황동색 장식이 보이지만 참고보다 칼끝이 각지고 날이 곧아 보인다. 등에 작은 절개와 피가 있다. 그러나 왼쪽 가장자리에 별도 공격자의 손과 짙은 소매가 나타난다. 바닥에 떨어진 칼과 좌판 파손은 이 구도에서 확인되지 않으며, 글자는 없다.",
        "hard_violations": [
         "왼쪽 화면 안에 공격자의 손과 짙은 소매를 추가했다. 이는 칼을 휘두르는 사람을 화면 밖에 두고 안규홍 외의 인물이나 신체 일부를 노출하지 말라는 명시적 구도·인물 제한을 위반한다."
        ],
        "physics": "낫은 안규홍의 오른손이, 공격 칼은 왼쪽 가장자리의 다른 사람 손이 지지하므로 무기가 떠 있지는 않다. 칼끝과 저고리의 상처가 맞닿아 있어 물리적 접촉은 읽힌다. 몸통을 비틀고 뒤돌아보는 자세도 가능하지만, 전방 방어가 유지되는 접촉이나 베임에 따른 뚜렷한 반동은 보이지 않는다. 발은 화면 밖이며 공중 부양의 증거는 없다."
       }
      ]
     },
     "ok": true
    }
   ],
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "slot_winner_match": false,
   "slot_winner": {
    "gemini-pro": "A",
    "gpt-high": "B"
   },
   "route": "cross_slot_combined"
  },
  "dual": {
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "normalized": {
    "A": 1.5,
    "B": 1.333
   },
   "adjusted": {
    "A": 1.25,
    "B": 1.333
   },
   "violations": {
    "A": [
     "[gpt-high] 왼쪽 화면 안에 공격자의 손과 짙은 소매를 추가했다. 이는 칼을 휘두르는 사람을 화면 밖에 두고 안규홍 외의 인물이나 신체 일부를 노출하지 말라는 명시적 구도·인물 제한을 위반한다."
    ]
   },
   "per_model_winner": {
    "gemini-pro": "A",
    "gpt-high": "B"
   },
   "agreed": false
  },
  "totals": {
   "A": 1250,
   "B": 1333
  },
  "selected": "B",
  "ranking": [
   "B",
   "A"
  ],
  "verdicts": [
   {
    "label": "A",
    "score": 1250,
    "verdict_ko": "프롬프트가 요구한 등 뒤의 칼 공격과 낫을 이용한 전면 방어 자세를 모두 화면에 담아냈으나, 칼날이 등을 베고 지나가는(스윽 긋는) 형태가 아니라 등 안으로 찔러 들어간 것처럼 묘사된 점이 아쉽습니다.  ★위반: [gpt-high] 왼쪽 화면 안에 공격자의 손과 짙은 소매를 추가했다. 이는 칼을 휘두르는 사람을 화면 밖에 두고 안규홍 외의 인물이나 신체 일부를 노출하지 말라는 명시적 구도·인물 제한을 위반한다."
   },
   {
    "label": "B",
    "score": 1333,
    "verdict_ko": "프롬프트의 핵심 액션인 '등 뒤로 칼날이 긋고 지나가는 찰나'와 공격자의 칼 자체를 화면에서 완전히 누락하여 연출 지시를 실패했습니다."
   }
  ],
  "refs": [
   {
    "label": "PREVIOUS SHOT STILL — a visually related earlier shot of this same place: the location's look, materials, fixed features and lighting mood are LOCKED to this photo; never copy its camera framing. The clothing, hair and overall look of 안규홍 — who appear both in that photo and in this shot — are LOCKED to that photo. Anyone else visible in that photo is NOT in this shot: never carry their face, body or clothing onto anyone here. If this photo shows a character who cannot move (dead or unconscious), that character's exact pose, position and orientation are ALSO LOCKED — treat their body as a fixed prop of the set that only the camera moves around.",
    "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/db00517b-165b-4005-a287-f28936b59b95/images/272bd7af-9294-4490-98c6-88a40e9deacc/scene/recipe/S51sh2_sel.png",
    "asset_id": "5c7f697c-c51b-4c91-b6c9-7bd4ad72472b",
    "role": "prev_still"
   },
   {
    "label": "CHARACTER REFERENCE — 안규홍: the exact person appearing in this shot; match face, hair and build exactly.",
    "path": "<bytes:1028384>",
    "asset_id": "3b5894db-ce55-4f34-b966-5e725c3aa952",
    "role": "character_ref"
   },
   {
    "label": "PROP REFERENCE — 칼: the exact object appearing in this shot; match its look, material and wear exactly.",
    "path": "<bytes:780122>",
    "asset_id": "82614014-6091-423c-89ca-1e15d209de72",
    "role": "prop_ref"
   }
  ],
  "critique_skipped": true,
  "shot_run_produced": false,
  "shot_run_uid": "06aa1436-5657-7add-b773-188f5905ee22",
  "ref_mode": "prev+엔티티",
  "share_plan": {
   "ref_plan": "prev",
   "prev_anchor_tag": "S51sh2"
  }
 },
 "S51sh9::cine": {
  "applied": true,
  "attempted_at": "2026-09-09T06:55:30.222424+00:00",
  "fingerprint": "87f8fbe00508fb9897c6e0d0befbb8f56ec63c8c70847e869a6803490bb5cb81",
  "fingerprint_version": 2,
  "provider": "mai",
  "endpoint": "openrouter/chat-completions",
  "model": "microsoft/mai-image-2.6",
  "pack": "24.202608252115",
  "source_file": "S51sh9_sel.png",
  "source_sha256": "965e228f7bcbc3a0192e15335a23a96d4e390aff17f663fff6bf83c28ee2f696",
  "file": "S51sh9_cine.png",
  "staged_sha256": "ad14494407bde35817d0c16f65483ee12e54623b67a0a39311be765e33e13cf3",
  "latency_ms": 31691
 },
 "S51sh15::signage": {
  "fp": "cc38f419fd4fe9f8",
  "inscriptions": [],
  "cues": [],
  "dropped": []
 },
 "S51sh15": {
  "input_fingerprint": "6cea2f44c9ff3348",
  "prompt": "Create ONE FINAL photorealistic live-action film still of the moment below — early 20th-century Korea, Jeolla region; all people are Korean unless explicitly stated otherwise. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 다시 서로의 넓은 등을 꽉 맞대고 방어 태세를 갖춘 안규홍과 이관회의 밀착된 뒷모습.\n\nLOCATION (lock): The narrow outdoor fighting space between the market stalls in front of the straw-sandal shop. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: medium shot\n- KEY BACKGROUND ELEMENTS: Market stall (The stall involved in the preceding collision remains partly visible) — An oblique portion appears beyond the pair's shoulders; used as Retains a peripheral reference to the fight's established geography without competing with the back-to-back contact.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Maintain restrained daylight, muted saturation, and naturalistic contrast that keeps their defensive contact legible without glamorizing the violence.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The clothing, hair and overall look of 안규홍 — who appear both in that photo and in this shot — are LOCKED to that photo. Anyone else visible in that photo is NOT in this shot: never carry their face, body or clothing onto anyone here. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The disrupted stall and dropped sword remain in the blocked market lane. 안규홍: He braces backward in defense with his sickle; blood begins soaking his jeogori from the back slash, and his neck remains bare. 이관회: He has regained a standing rear-facing defensive posture with his back braced.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 안규홍 (한국인 남성, 청년 나이대, 검은 머리, 단단히 묶은 머리); 이관회 (한국인 남성, 청년 나이대, 자연스러운 검은 머리) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.",
  "roll_prompts": {
   "A": "Create ONE FINAL photorealistic live-action film still of the moment below — early 20th-century Korea, Jeolla region; all people are Korean unless explicitly stated otherwise. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 다시 서로의 넓은 등을 꽉 맞대고 방어 태세를 갖춘 안규홍과 이관회의 밀착된 뒷모습.\n\nLOCATION (lock): The narrow outdoor fighting space between the market stalls in front of the straw-sandal shop. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: medium shot\n- KEY BACKGROUND ELEMENTS: Market stall (The stall involved in the preceding collision remains partly visible) — An oblique portion appears beyond the pair's shoulders; used as Retains a peripheral reference to the fight's established geography without competing with the back-to-back contact.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Maintain restrained daylight, muted saturation, and naturalistic contrast that keeps their defensive contact legible without glamorizing the violence.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The clothing, hair and overall look of 안규홍 — who appear both in that photo and in this shot — are LOCKED to that photo. Anyone else visible in that photo is NOT in this shot: never carry their face, body or clothing onto anyone here. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The disrupted stall and dropped sword remain in the blocked market lane. 안규홍: He braces backward in defense with his sickle; blood begins soaking his jeogori from the back slash, and his neck remains bare. 이관회: He has regained a standing rear-facing defensive posture with his back braced.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 안규홍 (한국인 남성, 청년 나이대, 검은 머리, 단단히 묶은 머리); 이관회 (한국인 남성, 청년 나이대, 자연스러운 검은 머리) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.",
   "B": "Create ONE FINAL photorealistic live-action film still of the moment below — early 20th-century Korea, Jeolla region; all people are Korean unless explicitly stated otherwise. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 다시 서로의 넓은 등을 꽉 맞대고 방어 태세를 갖춘 안규홍과 이관회의 밀착된 뒷모습.\n\nLOCATION (lock): The narrow outdoor fighting space between the market stalls in front of the straw-sandal shop. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: medium shot\n- KEY BACKGROUND ELEMENTS: Market stall (The stall involved in the preceding collision remains partly visible) — An oblique portion appears beyond the pair's shoulders; used as Retains a peripheral reference to the fight's established geography without competing with the back-to-back contact.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Maintain restrained daylight, muted saturation, and naturalistic contrast that keeps their defensive contact legible without glamorizing the violence.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The clothing, hair and overall look of 안규홍 — who appear both in that photo and in this shot — are LOCKED to that photo. Anyone else visible in that photo is NOT in this shot: never carry their face, body or clothing onto anyone here. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The disrupted stall and dropped sword remain in the blocked market lane. 안규홍: He braces backward in defense with his sickle; blood begins soaking his jeogori from the back slash, and his neck remains bare. 이관회: He has regained a standing rear-facing defensive posture with his back braced.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 안규홍 (한국인 남성, 청년 나이대, 검은 머리, 단단히 묶은 머리); 이관회 (한국인 남성, 청년 나이대, 자연스러운 검은 머리) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.\n\nCOMPOSITION VARIATION (this candidate only): keep the place, the people and the moment exactly as specified, and keep from the CAMERA & FRAME contract its subject scale, screen placement and key background elements exactly. For this candidate only, its camera angle and height are deliberately left open: choose a camera position distinctly different from the obvious one, as a film director picking a second setup on the same blocking. Stay true to any explicit framing the SHOT TEXT itself dictates."
  },
  "shot_run_spend_attempt_count": 1,
  "readings": [
   {
    "label": "A",
    "direction": "안규홍은 화면 좌측 밖을, 이관회는 화면 우측 밖을 향해 시선을 두고 방어 태세를 취하고 있음.",
    "built_space": "이전 샷과 동일한 짚신 매대와 바구니가 좌측에 배치된 시장 골목이며, 우측 바닥에는 떨어진 칼이 하나 놓여 있음.",
    "entities": "안규홍은 상투를 틀고 목이 드러난 채 피가 묻은 밝은색 저고리를 입었으며 왼손으로 낫을 들고 있음. 이관회는 레퍼런스와 일치하는 갈색 저고리를 입고 뒷모습과 옆모습을 보임.",
    "hard_violations": [
     "[gpt-high] 오른쪽 가장자리에 제삼자의 검은 신발과 하퇴 일부가 등장한다. 안규홍과 이관회 외에는 신체 일부도 추가하지 말라는 명시적 조건을 위반한다."
    ],
    "physics": "두 사람이 서로의 넓은 등에 체중을 싣고 안정적으로 서 있음. 안규홍의 왼손은 낫의 자루를 물리적으로 자연스럽게 쥐고 있으며, 바닥에 놓인 칼도 지면의 지지를 받고 있음."
   },
   {
    "label": "B",
    "direction": "안규홍은 좌측을, 이관회는 우측을 향해 시선을 두고 서로 등을 맞대고 있음.",
    "built_space": "시장 매대를 배경으로 하고 있으며, 안규홍 뒤편으로 부서진 나무 판자들이 보이고 우측 지면에 떨어진 칼이 놓여 있음.",
    "entities": "안규홍은 등에 피가 묻은 밝은색 저고리를 입고 양손으로 낫을 쥐고 있음. 이관회는 갈색 저고리를 입고 있으나 정면(가슴 부위)이 노출되어 있음.",
    "hard_violations": [
     "[gemini-pro] 카메라 구도 및 연출 위반: 프롬프트에서 '밀착된 뒷모습'을 명시했으나, 이관회의 앞모습(가슴 및 복부)이 정면으로 렌더링됨."
    ],
    "physics": "두 사람 모두 바닥에 발을 딛고 서 있으며, 안규홍의 양손은 낫의 손잡이를 단단히 파지하고 있음."
   }
  ],
  "cross_model_order": {
   "policy": "cross_model_order_v2_slot0_fwd_slot1_rev_parallel2",
   "slots": [
    {
     "model": "gemini-pro",
     "order": "forward",
     "display_to_canonical": {
      "A": "A",
      "B": "B"
     },
     "raw": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 9,
        "verdict_ko": "프롬프트가 요구한 '밀착된 뒷모습'의 카메라 구도를 정확하게 구현하였으며, 이전 샷의 배경 구조물과 혈흔 등 캐릭터의 상태를 완벽하게 유지한 훌륭한 결과물입니다."
       },
       {
        "label": "B",
        "score": 5,
        "verdict_ko": "이관회의 가슴과 앞모습이 화면에 그대로 노출되어, 프롬프트의 가장 중요한 지시인 '뒷모습' 카메라 구도를 위반하였습니다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "안규홍은 화면 좌측 밖을, 이관회는 화면 우측 밖을 향해 시선을 두고 방어 태세를 취하고 있음.",
        "built_space": "이전 샷과 동일한 짚신 매대와 바구니가 좌측에 배치된 시장 골목이며, 우측 바닥에는 떨어진 칼이 하나 놓여 있음.",
        "entities": "안규홍은 상투를 틀고 목이 드러난 채 피가 묻은 밝은색 저고리를 입었으며 왼손으로 낫을 들고 있음. 이관회는 레퍼런스와 일치하는 갈색 저고리를 입고 뒷모습과 옆모습을 보임.",
        "hard_violations": [],
        "physics": "두 사람이 서로의 넓은 등에 체중을 싣고 안정적으로 서 있음. 안규홍의 왼손은 낫의 자루를 물리적으로 자연스럽게 쥐고 있으며, 바닥에 놓인 칼도 지면의 지지를 받고 있음."
       },
       {
        "label": "B",
        "direction": "안규홍은 좌측을, 이관회는 우측을 향해 시선을 두고 서로 등을 맞대고 있음.",
        "built_space": "시장 매대를 배경으로 하고 있으며, 안규홍 뒤편으로 부서진 나무 판자들이 보이고 우측 지면에 떨어진 칼이 놓여 있음.",
        "entities": "안규홍은 등에 피가 묻은 밝은색 저고리를 입고 양손으로 낫을 쥐고 있음. 이관회는 갈색 저고리를 입고 있으나 정면(가슴 부위)이 노출되어 있음.",
        "hard_violations": [
         "카메라 구도 및 연출 위반: 프롬프트에서 '밀착된 뒷모습'을 명시했으나, 이관회의 앞모습(가슴 및 복부)이 정면으로 렌더링됨."
        ],
        "physics": "두 사람 모두 바닥에 발을 딛고 서 있으며, 안규홍의 양손은 낫의 손잡이를 단단히 파지하고 있음."
       }
      ],
      "all_candidates_fail": false
     },
     "normalized": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 9,
        "verdict_ko": "프롬프트가 요구한 '밀착된 뒷모습'의 카메라 구도를 정확하게 구현하였으며, 이전 샷의 배경 구조물과 혈흔 등 캐릭터의 상태를 완벽하게 유지한 훌륭한 결과물입니다."
       },
       {
        "label": "B",
        "score": 5,
        "verdict_ko": "이관회의 가슴과 앞모습이 화면에 그대로 노출되어, 프롬프트의 가장 중요한 지시인 '뒷모습' 카메라 구도를 위반하였습니다."
       }
      ],
      "all_candidates_fail": false,
      "readings": [
       {
        "label": "A",
        "direction": "안규홍은 화면 좌측 밖을, 이관회는 화면 우측 밖을 향해 시선을 두고 방어 태세를 취하고 있음.",
        "built_space": "이전 샷과 동일한 짚신 매대와 바구니가 좌측에 배치된 시장 골목이며, 우측 바닥에는 떨어진 칼이 하나 놓여 있음.",
        "entities": "안규홍은 상투를 틀고 목이 드러난 채 피가 묻은 밝은색 저고리를 입었으며 왼손으로 낫을 들고 있음. 이관회는 레퍼런스와 일치하는 갈색 저고리를 입고 뒷모습과 옆모습을 보임.",
        "hard_violations": [],
        "physics": "두 사람이 서로의 넓은 등에 체중을 싣고 안정적으로 서 있음. 안규홍의 왼손은 낫의 자루를 물리적으로 자연스럽게 쥐고 있으며, 바닥에 놓인 칼도 지면의 지지를 받고 있음."
       },
       {
        "label": "B",
        "direction": "안규홍은 좌측을, 이관회는 우측을 향해 시선을 두고 서로 등을 맞대고 있음.",
        "built_space": "시장 매대를 배경으로 하고 있으며, 안규홍 뒤편으로 부서진 나무 판자들이 보이고 우측 지면에 떨어진 칼이 놓여 있음.",
        "entities": "안규홍은 등에 피가 묻은 밝은색 저고리를 입고 양손으로 낫을 쥐고 있음. 이관회는 갈색 저고리를 입고 있으나 정면(가슴 부위)이 노출되어 있음.",
        "hard_violations": [
         "카메라 구도 및 연출 위반: 프롬프트에서 '밀착된 뒷모습'을 명시했으나, 이관회의 앞모습(가슴 및 복부)이 정면으로 렌더링됨."
        ],
        "physics": "두 사람 모두 바닥에 발을 딛고 서 있으며, 안규홍의 양손은 낫의 손잡이를 단단히 파지하고 있음."
       }
      ]
     },
     "ok": true
    },
    {
     "model": "gpt-high",
     "order": "reverse",
     "display_to_canonical": {
      "A": "B",
      "B": "A"
     },
     "raw": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 7,
        "verdict_ko": "두 사람의 등 접촉과 낫을 든 방어 동작은 명확하지만, 무릎 부근까지 넓힌 구도와 이관회의 정면 노출은 지정된 미디엄 뒷모습에서 벗어난다."
       },
       {
        "label": "B",
        "score": 3,
        "verdict_ko": "밀착된 등과 미디엄 구도는 더 충실하지만, 오른쪽 가장자리에 제삼자의 검은 신발과 하퇴가 보여 두 사람만 등장해야 한다는 조건을 위반한다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "안규홍은 화면 왼쪽 밖을 경계하며 낫을 두 손으로 들어 올렸다. 낫날 끝도 왼쪽을 향하고 특정한 화면 내 인물을 겨누지는 않는다. 이관회는 반대쪽인 화면 오른쪽 밖을 보고 있어 양방향 방어 관계가 읽힌다. 다만 이관회의 얼굴과 가슴이 카메라에 상당히 드러나 두 사람의 뒷모습이라는 지시는 약해진다.",
        "built_space": "왼쪽에 긴 목제 진열대 하나와 어두운 수납부, 그 옆 목제 기둥이 있고, 오른쪽 전경에도 큰 기둥 하나가 보인다. 부서진 판재 한 덩어리가 왼쪽 진열대와 땅 사이에 비스듬히 걸쳐 있으며, 바닥에는 바구니 세 개와 짚신이 보인다. 두 사람은 그 사이 흙길에 서 있다. 목재·천 차양·바구니의 재질은 이전 장소와 잘 이어지지만, 배경과 하체를 넓게 보여 주어 접촉 부위 중심의 미디엄 숏보다 넓다.",
        "entities": "등장인물은 젊은 한국인 남성으로 표현된 두 명뿐이다. 안규홍의 검은 묶음머리, 때 묻은 밝은 저고리, 드러난 목, 등에 번진 피와 낫이 확인된다. 얼굴 윤곽도 이전 스틸과 대체로 이어진다. 이관회는 자연스럽게 가른 검은 머리, 해진 갈색 저고리와 밝은 바지를 입어 인물 참조와 부합한다. 떨어진 칼 한 자루가 오른쪽 흙바닥에 놓여 있다. 읽을 수 있는 글자나 그래픽은 없다.",
        "hard_violations": [],
        "physics": "두 사람은 상체를 반대 방향으로 틀고 등과 허리 부근을 맞대고 있다. 보이는 다리는 아래 프레임 밖으로 이어지며 서 있는 자세에 모순이 없고, 발이 잘렸다는 이유로 부유한다고 볼 수 없다. 낫 손잡이는 안규홍의 두 손에 확실히 잡혀 있다. 부서진 판재는 진열대와 지면에 지지되고, 칼과 바구니는 땅에 놓여 있다."
       },
       {
        "label": "B",
        "direction": "안규홍은 화면 왼쪽 아래 바깥을 보고, 이관회는 오른쪽 뒤편 통로를 향한다. 서로 반대 방향을 경계하며 이관회의 뒤통수와 등이 보인다. 안규홍의 낫은 왼손 쪽 전방에 낮게 들려 있고 날 끝은 왼쪽 아래를 향한다. 화면 안에 직접 겨누는 상대는 없다.",
        "built_space": "왼쪽 목제 진열대 하나, 인접한 굵은 기둥, 걸린 바구니와 바닥의 바구니 세 개가 이전 시장 공간을 이어 준다. 두 사람의 어깨 너머 중앙에 비스듬한 파손 판재가 일부 보이며 오른쪽으로 흙길이 열린다. 인물은 허리와 엉덩이 부근에서 잘려 미디엄 숏에 더 가깝고, 등 접촉이 화면 중심을 차지한다. 오른쪽 통로 가장자리에는 두 주인공에게 속할 수 없는 신발과 하퇴 일부가 보인다.",
        "entities": "중앙의 두 남성은 안규홍과 이관회의 머리 형태와 밝은색·갈색 저고리를 각각 유지한다. 안규홍의 목은 장신구 없이 드러나고 등에 혈흔이 있으며 한 손에 낫을 들었다. 이관회는 얼굴이 거의 가려져 얼굴 동일성은 판단하기 어렵지만 머리와 옷은 참조에 부합한다. 오른쪽 바닥에 떨어진 칼이 보이고, 그보다 뒤에는 별도의 긴 막대형 물체도 놓여 있다. 오른쪽 끝의 검은 신발과 하퇴는 허용된 두 인물 외의 신체 일부다. 읽을 수 있는 글자는 없다.",
        "hard_violations": [
         "오른쪽 가장자리에 제삼자의 검은 신발과 하퇴 일부가 등장한다. 안규홍과 이관회 외에는 신체 일부도 추가하지 말라는 명시적 조건을 위반한다."
        ],
        "physics": "두 사람의 등과 허리 접촉은 자연스럽고, 몸통은 프레임 아래로 이어지는 서 있는 자세다. 발은 구도 밖이므로 직접 확인할 수 없지만 공중에 떠 있다는 징후는 없다. 낫은 보이는 손이 손잡이를 감싸 지지한다. 바구니와 칼은 지면에 놓여 있고 파손 판재는 주변 구조물에 기대어 있다. 오른쪽 추가 인물의 신발 역시 통로 바닥에 닿아 보인다."
       }
      ],
      "all_candidates_fail": false
     },
     "normalized": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 7,
        "verdict_ko": "두 사람의 등 접촉과 낫을 든 방어 동작은 명확하지만, 무릎 부근까지 넓힌 구도와 이관회의 정면 노출은 지정된 미디엄 뒷모습에서 벗어난다."
       },
       {
        "label": "A",
        "score": 3,
        "verdict_ko": "밀착된 등과 미디엄 구도는 더 충실하지만, 오른쪽 가장자리에 제삼자의 검은 신발과 하퇴가 보여 두 사람만 등장해야 한다는 조건을 위반한다."
       }
      ],
      "all_candidates_fail": false,
      "readings": [
       {
        "label": "B",
        "direction": "안규홍은 화면 왼쪽 밖을 경계하며 낫을 두 손으로 들어 올렸다. 낫날 끝도 왼쪽을 향하고 특정한 화면 내 인물을 겨누지는 않는다. 이관회는 반대쪽인 화면 오른쪽 밖을 보고 있어 양방향 방어 관계가 읽힌다. 다만 이관회의 얼굴과 가슴이 카메라에 상당히 드러나 두 사람의 뒷모습이라는 지시는 약해진다.",
        "built_space": "왼쪽에 긴 목제 진열대 하나와 어두운 수납부, 그 옆 목제 기둥이 있고, 오른쪽 전경에도 큰 기둥 하나가 보인다. 부서진 판재 한 덩어리가 왼쪽 진열대와 땅 사이에 비스듬히 걸쳐 있으며, 바닥에는 바구니 세 개와 짚신이 보인다. 두 사람은 그 사이 흙길에 서 있다. 목재·천 차양·바구니의 재질은 이전 장소와 잘 이어지지만, 배경과 하체를 넓게 보여 주어 접촉 부위 중심의 미디엄 숏보다 넓다.",
        "entities": "등장인물은 젊은 한국인 남성으로 표현된 두 명뿐이다. 안규홍의 검은 묶음머리, 때 묻은 밝은 저고리, 드러난 목, 등에 번진 피와 낫이 확인된다. 얼굴 윤곽도 이전 스틸과 대체로 이어진다. 이관회는 자연스럽게 가른 검은 머리, 해진 갈색 저고리와 밝은 바지를 입어 인물 참조와 부합한다. 떨어진 칼 한 자루가 오른쪽 흙바닥에 놓여 있다. 읽을 수 있는 글자나 그래픽은 없다.",
        "hard_violations": [],
        "physics": "두 사람은 상체를 반대 방향으로 틀고 등과 허리 부근을 맞대고 있다. 보이는 다리는 아래 프레임 밖으로 이어지며 서 있는 자세에 모순이 없고, 발이 잘렸다는 이유로 부유한다고 볼 수 없다. 낫 손잡이는 안규홍의 두 손에 확실히 잡혀 있다. 부서진 판재는 진열대와 지면에 지지되고, 칼과 바구니는 땅에 놓여 있다."
       },
       {
        "label": "A",
        "direction": "안규홍은 화면 왼쪽 아래 바깥을 보고, 이관회는 오른쪽 뒤편 통로를 향한다. 서로 반대 방향을 경계하며 이관회의 뒤통수와 등이 보인다. 안규홍의 낫은 왼손 쪽 전방에 낮게 들려 있고 날 끝은 왼쪽 아래를 향한다. 화면 안에 직접 겨누는 상대는 없다.",
        "built_space": "왼쪽 목제 진열대 하나, 인접한 굵은 기둥, 걸린 바구니와 바닥의 바구니 세 개가 이전 시장 공간을 이어 준다. 두 사람의 어깨 너머 중앙에 비스듬한 파손 판재가 일부 보이며 오른쪽으로 흙길이 열린다. 인물은 허리와 엉덩이 부근에서 잘려 미디엄 숏에 더 가깝고, 등 접촉이 화면 중심을 차지한다. 오른쪽 통로 가장자리에는 두 주인공에게 속할 수 없는 신발과 하퇴 일부가 보인다.",
        "entities": "중앙의 두 남성은 안규홍과 이관회의 머리 형태와 밝은색·갈색 저고리를 각각 유지한다. 안규홍의 목은 장신구 없이 드러나고 등에 혈흔이 있으며 한 손에 낫을 들었다. 이관회는 얼굴이 거의 가려져 얼굴 동일성은 판단하기 어렵지만 머리와 옷은 참조에 부합한다. 오른쪽 바닥에 떨어진 칼이 보이고, 그보다 뒤에는 별도의 긴 막대형 물체도 놓여 있다. 오른쪽 끝의 검은 신발과 하퇴는 허용된 두 인물 외의 신체 일부다. 읽을 수 있는 글자는 없다.",
        "hard_violations": [
         "오른쪽 가장자리에 제삼자의 검은 신발과 하퇴 일부가 등장한다. 안규홍과 이관회 외에는 신체 일부도 추가하지 말라는 명시적 조건을 위반한다."
        ],
        "physics": "두 사람의 등과 허리 접촉은 자연스럽고, 몸통은 프레임 아래로 이어지는 서 있는 자세다. 발은 구도 밖이므로 직접 확인할 수 없지만 공중에 떠 있다는 징후는 없다. 낫은 보이는 손이 손잡이를 감싸 지지한다. 바구니와 칼은 지면에 놓여 있고 파손 판재는 주변 구조물에 기대어 있다. 오른쪽 추가 인물의 신발 역시 통로 바닥에 닿아 보인다."
       }
      ]
     },
     "ok": true
    }
   ],
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "slot_winner_match": false,
   "slot_winner": {
    "gemini-pro": "A",
    "gpt-high": "B"
   },
   "route": "cross_slot_combined"
  },
  "dual": {
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "normalized": {
    "A": 1.429,
    "B": 1.556
   },
   "adjusted": {
    "A": 1.179,
    "B": 1.306
   },
   "violations": {
    "B": [
     "[gemini-pro] 카메라 구도 및 연출 위반: 프롬프트에서 '밀착된 뒷모습'을 명시했으나, 이관회의 앞모습(가슴 및 복부)이 정면으로 렌더링됨."
    ],
    "A": [
     "[gpt-high] 오른쪽 가장자리에 제삼자의 검은 신발과 하퇴 일부가 등장한다. 안규홍과 이관회 외에는 신체 일부도 추가하지 말라는 명시적 조건을 위반한다."
    ]
   },
   "per_model_winner": {
    "gemini-pro": "A",
    "gpt-high": "B"
   },
   "agreed": false
  },
  "totals": {
   "A": 1179,
   "B": 1306
  },
  "selected": "B",
  "ranking": [
   "B",
   "A"
  ],
  "verdicts": [
   {
    "label": "A",
    "score": 1179,
    "verdict_ko": "프롬프트가 요구한 '밀착된 뒷모습'의 카메라 구도를 정확하게 구현하였으며, 이전 샷의 배경 구조물과 혈흔 등 캐릭터의 상태를 완벽하게 유지한 훌륭한 결과물입니다.  ★위반: [gpt-high] 오른쪽 가장자리에 제삼자의 검은 신발과 하퇴 일부가 등장한다. 안규홍과 이관회 외에는 신체 일부도 추가하지 말라는 명시적 조건을 위반한다."
   },
   {
    "label": "B",
    "score": 1306,
    "verdict_ko": "이관회의 가슴과 앞모습이 화면에 그대로 노출되어, 프롬프트의 가장 중요한 지시인 '뒷모습' 카메라 구도를 위반하였습니다.  ★위반: [gemini-pro] 카메라 구도 및 연출 위반: 프롬프트에서 '밀착된 뒷모습'을 명시했으나, 이관회의 앞모습(가슴 및 복부)이 정면으로 렌더링됨."
   }
  ],
  "refs": [
   {
    "label": "PREVIOUS SHOT STILL — a visually related earlier shot of this same place: the location's look, materials, fixed features and lighting mood are LOCKED to this photo; never copy its camera framing. The clothing, hair and overall look of 안규홍 — who appear both in that photo and in this shot — are LOCKED to that photo. Anyone else visible in that photo is NOT in this shot: never carry their face, body or clothing onto anyone here. If this photo shows a character who cannot move (dead or unconscious), that character's exact pose, position and orientation are ALSO LOCKED — treat their body as a fixed prop of the set that only the camera moves around.",
    "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/db00517b-165b-4005-a287-f28936b59b95/images/272bd7af-9294-4490-98c6-88a40e9deacc/scene/recipe/S51sh9_sel.png",
    "asset_id": "74cf1ba8-ffc5-4250-b0e6-e0801ea7f913",
    "role": "prev_still"
   },
   {
    "label": "CHARACTER REFERENCE — 안규홍: the exact person appearing in this shot; match face, hair and build exactly.",
    "path": "<bytes:1028384>",
    "asset_id": "3b5894db-ce55-4f34-b966-5e725c3aa952",
    "role": "character_ref"
   },
   {
    "label": "CHARACTER REFERENCE — 이관회: the exact person appearing in this shot; match face, hair and build exactly.",
    "path": "<bytes:933540>",
    "asset_id": "851b450f-ac84-47fe-ac5a-03a56f63fcae",
    "role": "character_ref"
   }
  ],
  "critique_skipped": true,
  "shot_run_produced": false,
  "shot_run_uid": "06aa1436-5805-7a92-8a71-6bfd7b2735b9",
  "ref_mode": "prev+엔티티",
  "share_plan": {
   "ref_plan": "prev",
   "prev_anchor_tag": "S51sh9"
  }
 },
 "S51sh15::cine": {
  "applied": true,
  "attempted_at": "2026-09-09T06:57:23.465049+00:00",
  "fingerprint": "6abc65b31596f292895fabd110a80ce98607c0cc954157a4142dcf68cf5a12d0",
  "fingerprint_version": 2,
  "provider": "mai",
  "endpoint": "openrouter/chat-completions",
  "model": "microsoft/mai-image-2.6",
  "pack": "24.202608252115",
  "source_file": "S51sh15_sel.png",
  "source_sha256": "c83af6ea6edb4f80511df999293ad17b31191d17046039209be414047ce7faab",
  "file": "S51sh15_cine.png",
  "staged_sha256": "6bbeb39c74414beacac94994690ac3117921094e108b276c509a29dda66acbb0",
  "latency_ms": 32930
 },
 "S52sh1::signage": {
  "fp": "71b3dadeb8a1a50c",
  "inscriptions": [],
  "cues": [],
  "dropped": []
 },
 "era_assess::3fe8c75dc5a7013f": {
  "subjects": [],
  "subject_text": "파청 산채를 내려다보는 능선\n숲속 산채 공터와 움막을 내려다보는 높은 능선. 나무 사이로 아래 공간이 열리고, 양옆으로 숲 비탈이 이어진다.",
  "identity": "canonical",
  "scope_id": "L26",
  "scope_role": "location_exterior",
  "scope_sha": "75837cbdd367aa06"
 },
 "S52sh1::bgfirst_bg": {
  "input_fingerprint": "b53177c0bef52bc4",
  "prompt": "Create the EMPTY BACKGROUND PLATE for one film shot — NO PEOPLE, no figures, no body parts, no sketch lines, no arrows anywhere.\n\"Empty\" means no people only: KEEP the location's inherent occupants and stock that define the place — animals in an animal shelter, pen or farm, goods in a market, moored boats in a harbour — unless the shot text explicitly removes them.\nThe FIRST attached image is a thin-line storyboard sketch: use ONLY its camera angle, horizon, perspective and the placement/size of buildings and set masses — ignore the sketched people and arrows entirely. The SECOND attached image (LOCATION PHOTOGRAPH) is the real place: take its architecture, materials, signage and fixed features, and RE-PROJECT them into the sketch's camera. If the photograph's camera differs from the sketch's, the sketch's camera wins.\nHUMAN-SCALE CALIBRATION: derive every structure's true size from human-scale elements — a door ≈ 2m, a window ≈ 1–1.5m wide, one storey ≈ 2.5–3m; never inflate a small structure or shrink a large one.\n\nSHOT TEXT this background must serve (Korean): 산채가 내려다보이는 높은 능선 위, 토미이시와 배상도가 나란히 서서 아래를 굽어보는 전신.\n\nLOCATION (lock): An exposed wooded ridge above the mountain camp, offering a clear downhill view into its clearing.\n\nTIME OF DAY (lock): day.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: wide shot\n- KEY BACKGROUND ELEMENTS: Ridge (Overlooks the camp) — Its downhill slope recedes away from the figures; used as Establishes the commanders' elevated position and the direction of the impending advance; Camp below (Visible beneath the ridge) — Seen from above beyond the two figures; used as Provides the shared destination of their downward attention without introducing unsupported architectural details.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Use natural daylight with subdued saturation and measured contrast, preserving clear spatial depth and an unsentimental sense of threat.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nRender ONE photorealistic empty location photograph, 16:9, that this shot can be staged inside later. No readable writing anywhere: surfaces that would carry writing may be present, but stage any wording out of legibility — an oblique angle, distance, shallow focus. No captions, watermarks or overlay text.",
  "effective_prompt": "Create the EMPTY BACKGROUND PLATE for one film shot — NO PEOPLE, no figures, no body parts, no sketch lines, no arrows anywhere.\n\"Empty\" means no people only: KEEP the location's inherent occupants and stock that define the place — animals in an animal shelter, pen or farm, goods in a market, moored boats in a harbour — unless the shot text explicitly removes them.\nThe FIRST attached image is a thin-line storyboard sketch: use ONLY its camera angle, horizon, perspective and the placement/size of buildings and set masses — ignore the sketched people and arrows entirely. The SECOND attached image (LOCATION PHOTOGRAPH) is the real place: take its architecture, materials, signage and fixed features, and RE-PROJECT them into the sketch's camera. If the photograph's camera differs from the sketch's, the sketch's camera wins.\nHUMAN-SCALE CALIBRATION: derive every structure's true size from human-scale elements — a door ≈ 2m, a window ≈ 1–1.5m wide, one storey ≈ 2.5–3m; never inflate a small structure or shrink a large one.\n\nSHOT TEXT this background must serve (Korean): 산채가 내려다보이는 높은 능선 위, 토미이시와 배상도가 나란히 서서 아래를 굽어보는 전신.\n\nLOCATION (lock): An exposed wooded ridge above the mountain camp, offering a clear downhill view into its clearing.\n\nTIME OF DAY (lock): day.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: wide shot\n- KEY BACKGROUND ELEMENTS: Ridge (Overlooks the camp) — Its downhill slope recedes away from the figures; used as Establishes the commanders' elevated position and the direction of the impending advance; Camp below (Visible beneath the ridge) — Seen from above beyond the two figures; used as Provides the shared destination of their downward attention without introducing unsupported architectural details.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Use natural daylight with subdued saturation and measured contrast, preserving clear spatial depth and an unsentimental sense of threat.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nRender ONE photorealistic empty location photograph, 16:9, that this shot can be staged inside later. No readable writing anywhere: surfaces that would carry writing may be present, but stage any wording out of legibility — an oblique angle, distance, shallow focus. No captions, watermarks or overlay text.",
  "bg_path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/db00517b-165b-4005-a287-f28936b59b95/images/272bd7af-9294-4490-98c6-88a40e9deacc/scene/recipe/S52sh1__bgfirst_bg.png",
  "asset_id": "2bb90254-cd01-44d5-8d2b-7c078f8f33a1",
  "input_asset_ids": [
   "0091b78d-fbc2-42ab-a5cc-2957683bf011",
   "8c17c7ae-efff-4bb5-b88c-7c5528f00716"
  ]
 },
 "S52sh1": {
  "input_fingerprint": "7b07ca0bcbb977af",
  "prompt": "Create ONE FINAL photorealistic live-action film still of the moment below — early 20th-century Korea, Jeolla region; all people are Korean unless explicitly stated otherwise. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 산채가 내려다보이는 높은 능선 위, 토미이시와 배상도가 나란히 서서 아래를 굽어보는 전신.\n\nLOCATION (lock): An exposed wooded ridge above the mountain camp, offering a clear downhill view into its clearing. The shot takes place here — the attached LOCATION STRUCTURE PHOTOGRAPH is the single authority for this exact place — its fixed structure and permanent site details are LOCKED to it. No separate location photograph exists for this place. Build everything else strictly from the location text above and the shot text; the layout sketch (when attached) governs framing and placement only, and the shot text governs time of day, lighting and action.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: wide shot\n- KEY BACKGROUND ELEMENTS: Ridge (Overlooks the camp) — Its downhill slope recedes away from the figures; used as Establishes the commanders' elevated position and the direction of the impending advance; Camp below (Visible beneath the ridge) — Seen from above beyond the two figures; used as Provides the shared destination of their downward attention without introducing unsupported architectural details.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Use natural daylight with subdued saturation and measured contrast, preserving clear spatial depth and an unsentimental sense of threat.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): A wooded ridge overlooks the camp in daylight. 토미이시: He stands on the ridge looking down, carrying the sword used in the later fight. 배상도: He stands on the ridge.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 배상도 (한국인 남성, 성인 나이대, 자연스러운 검은 머리); 토미이시 (일본인 남성, 성인 나이대, 자연스러운 검은 머리) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.",
  "roll_prompts": {
   "A": "Stage the shot. The FIRST attached image (SHOT BACKGROUND) is the finished empty background of this shot — keep it EXACTLY: its camera, perspective, architecture, lighting and every fixed feature stay untouched. The SECOND attached image (LAYOUT SKETCH) tells you ONLY where the people go: each sketched person's position, screen size, pose and the gaze/motion arrows. Ignore the sketch's background lines. The CHARACTER REFERENCE photographs show the real people.\nPlace the real people into the background at exactly the sketched positions, sizes and poses, following the arrow directions. No sketch lines or arrows may remain.\n\nCreate ONE FINAL photorealistic live-action film still of the moment below — early 20th-century Korea, Jeolla region; all people are Korean unless explicitly stated otherwise. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 산채가 내려다보이는 높은 능선 위, 토미이시와 배상도가 나란히 서서 아래를 굽어보는 전신.\n\nLOCATION (lock): An exposed wooded ridge above the mountain camp, offering a clear downhill view into its clearing. The shot takes place here — the FIRST attached image (SHOT BACKGROUND) is this exact place, already built: its ground, structures, horizon, materials and lighting are the finished truth of this location and must not be redesigned or replaced. No location photograph is attached — read the place from that image alone, and add no scenery, structure, vehicle or fixture that it does not already show. This lock governs the place only; the figures in the shot follow the staging and pose instructions.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: wide shot\n- KEY BACKGROUND ELEMENTS: Ridge (Overlooks the camp) — Its downhill slope recedes away from the figures; used as Establishes the commanders' elevated position and the direction of the impending advance; Camp below (Visible beneath the ridge) — Seen from above beyond the two figures; used as Provides the shared destination of their downward attention without introducing unsupported architectural details.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Use natural daylight with subdued saturation and measured contrast, preserving clear spatial depth and an unsentimental sense of threat.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): A wooded ridge overlooks the camp in daylight. 토미이시: He stands on the ridge looking down, carrying the sword used in the later fight. 배상도: He stands on the ridge.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 배상도 (한국인 남성, 성인 나이대, 자연스러운 검은 머리); 토미이시 (일본인 남성, 성인 나이대, 자연스러운 검은 머리) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.",
   "B": "Create ONE FINAL photorealistic live-action film still of the moment below — early 20th-century Korea, Jeolla region; all people are Korean unless explicitly stated otherwise. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 산채가 내려다보이는 높은 능선 위, 토미이시와 배상도가 나란히 서서 아래를 굽어보는 전신.\n\nLOCATION (lock): An exposed wooded ridge above the mountain camp, offering a clear downhill view into its clearing. The shot takes place here — the attached LOCATION STRUCTURE PHOTOGRAPH is the single authority for this exact place — its fixed structure and permanent site details are LOCKED to it. No separate location photograph exists for this place. Build everything else strictly from the location text above and the shot text; the layout sketch (when attached) governs framing and placement only, and the shot text governs time of day, lighting and action.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: wide shot\n- KEY BACKGROUND ELEMENTS: Ridge (Overlooks the camp) — Its downhill slope recedes away from the figures; used as Establishes the commanders' elevated position and the direction of the impending advance; Camp below (Visible beneath the ridge) — Seen from above beyond the two figures; used as Provides the shared destination of their downward attention without introducing unsupported architectural details.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Use natural daylight with subdued saturation and measured contrast, preserving clear spatial depth and an unsentimental sense of threat.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): A wooded ridge overlooks the camp in daylight. 토미이시: He stands on the ridge looking down, carrying the sword used in the later fight. 배상도: He stands on the ridge.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 배상도 (한국인 남성, 성인 나이대, 자연스러운 검은 머리); 토미이시 (일본인 남성, 성인 나이대, 자연스러운 검은 머리) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.\n\nCOMPOSITION VARIATION (this candidate only): keep the place, the people and the moment exactly as specified, and keep from the CAMERA & FRAME contract its subject scale, screen placement and key background elements exactly. For this candidate only, its camera angle and height are deliberately left open: choose a camera position distinctly different from the obvious one, as a film director picking a second setup on the same blocking. Stay true to any explicit framing the SHOT TEXT itself dictates."
  },
  "roll_refs": {
   "A": [
    {
     "label": "SHOT BACKGROUND",
     "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/db00517b-165b-4005-a287-f28936b59b95/images/272bd7af-9294-4490-98c6-88a40e9deacc/scene/recipe/S52sh1__bgfirst_bg.png",
     "asset_id": "2bb90254-cd01-44d5-8d2b-7c078f8f33a1",
     "role": "bgfirst_bg"
    },
    {
     "label": "LAYOUT SKETCH (people placement only)",
     "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/db00517b-165b-4005-a287-f28936b59b95/images/272bd7af-9294-4490-98c6-88a40e9deacc/conti/conti_S52sh1.png",
     "asset_id": "0091b78d-fbc2-42ab-a5cc-2957683bf011",
     "role": "conti_light"
    },
    {
     "label": "CHARACTER REFERENCE — 배상도: the exact person appearing in this shot; match face, hair and build exactly.",
     "path": "<bytes:851718>",
     "asset_id": "034f5325-975e-4474-854b-e48359710bf8",
     "role": "character_ref"
    },
    {
     "label": "CHARACTER REFERENCE — 토미이시: the exact person appearing in this shot; match face, hair and build exactly.",
     "path": "<bytes:797699>",
     "asset_id": "c861ba0f-8750-4262-8324-25cf5c6c9dcd",
     "role": "character_ref"
    }
   ],
   "B": [
    {
     "label": "LOCATION STRUCTURE PHOTOGRAPH — the confirmed photograph of this exact place and its fixed structure: it is the SINGLE authority for the location, the structure's shape, proportions, materials, colors, openings and every permanent site detail. Never copy its camera framing, time of day or lighting — the shot text is the authority for those.",
     "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/db00517b-165b-4005-a287-f28936b59b95/images/background_chain/seed_bg_mountain_camp_sel.png",
     "asset_id": "8c17c7ae-efff-4bb5-b88c-7c5528f00716",
     "role": "location_seed_bg"
    },
    {
     "label": "CHARACTER REFERENCE — 배상도: the exact person appearing in this shot; match face, hair and build exactly.",
     "path": "<bytes:851718>",
     "asset_id": "034f5325-975e-4474-854b-e48359710bf8",
     "role": "character_ref"
    },
    {
     "label": "CHARACTER REFERENCE — 토미이시: the exact person appearing in this shot; match face, hair and build exactly.",
     "path": "<bytes:797699>",
     "asset_id": "c861ba0f-8750-4262-8324-25cf5c6c9dcd",
     "role": "character_ref"
    }
   ]
  },
  "shot_run_spend_attempt_count": 1,
  "judge_flip_skipped": "judge_fn_owns_order",
  "readings": [
   {
    "label": "A",
    "direction": "두 인물의 시선이 표적인 아래쪽 계곡의 산채를 향해 정확히 굽어보고 있음.",
    "built_space": "아래쪽 산채는 원형 목책과 초가집이 레퍼런스와 일치하게 배치됨. 두 인물은 산채가 내려다보이는 바위 능선 위 자연스러운 위치에 서 있으며, 지형의 경사가 인물들 너머로 멀어지는 카메라 구도가 정확히 구현됨.",
    "entities": "배상도(한복, 검은 조끼)와 토미이시(군복) 모두 지정된 레퍼런스의 복장과 외형에 일치함.",
    "hard_violations": [
     "[gemini-pro] 물리적으로 불가능한 형태/사물: 토미이시의 왼손이 허리에 찬 칼집을 쥐고 있음에도 불구하고 동일한 손에서 손잡이가 없는 여분의 금속 칼날이 허공을 향해 비정상적으로 뻗어 나옴."
    ],
    "physics": "두 인물의 발은 바위 지형을 딛고 지지됨. 그러나 토미이시의 왼손에서 자루 없는 칼날만이 기형적으로 뻗어 나와 물리적 형태와 지지 구조에 오류가 있음."
   },
   {
    "label": "B",
    "direction": "두 인물 모두 시선이 아래를 향하지 않고 수평으로 왼쪽 화면 밖을 향하고 있어, 명시된 표적인 산채를 내려다보지 않고 완전히 빗나감.",
    "built_space": "계곡 아래 산채의 초가집과 목책이 레퍼런스와 일치함. 두 인물은 능선 위 흙길에 나란히 서 있으나, 산채가 인물들 너머(beyond)가 아닌 화면 왼쪽에 위치하여 제시된 배경 구도와 다름.",
    "entities": "배상도와 토미이시 모두 레퍼런스의 복장과 특징을 잘 반영함. 배상도는 텍스트에 지시되지 않은 검은 모자를 오른손에 들고 있음.",
    "hard_violations": [],
    "physics": "두 인물 모두 지면에 안정적으로 서 있으며, 배상도의 손에 쥐어진 모자도 자연스럽게 지지됨. 공중에 떠 있거나 구조적으로 불가능한 사물은 없음."
   }
  ],
  "cross_model_order": {
   "policy": "cross_model_order_v2_slot0_fwd_slot1_rev_parallel2",
   "slots": [
    {
     "model": "gemini-pro",
     "order": "forward",
     "display_to_canonical": {
      "A": "A",
      "B": "B"
     },
     "raw": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 5,
        "verdict_ko": "물리적 오류나 치명적 위반은 없으나, 두 인물이 수평으로 딴 곳을 응시하여 '아래의 산채를 굽어보는' 핵심 시선 방향과 행동 지시를 완전히 실패함."
       },
       {
        "label": "A",
        "score": 4,
        "verdict_ko": "프롬프트의 시선 방향과 공간 구도를 완벽하게 구현했으나, 토미이시의 손에서 자루 없는 여분의 칼날이 생성되는 치명적인 물리적 오류로 인해 사용 불가함."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "두 인물의 시선이 표적인 아래쪽 계곡의 산채를 향해 정확히 굽어보고 있음.",
        "built_space": "아래쪽 산채는 원형 목책과 초가집이 레퍼런스와 일치하게 배치됨. 두 인물은 산채가 내려다보이는 바위 능선 위 자연스러운 위치에 서 있으며, 지형의 경사가 인물들 너머로 멀어지는 카메라 구도가 정확히 구현됨.",
        "entities": "배상도(한복, 검은 조끼)와 토미이시(군복) 모두 지정된 레퍼런스의 복장과 외형에 일치함.",
        "hard_violations": [
         "물리적으로 불가능한 형태/사물: 토미이시의 왼손이 허리에 찬 칼집을 쥐고 있음에도 불구하고 동일한 손에서 손잡이가 없는 여분의 금속 칼날이 허공을 향해 비정상적으로 뻗어 나옴."
        ],
        "physics": "두 인물의 발은 바위 지형을 딛고 지지됨. 그러나 토미이시의 왼손에서 자루 없는 칼날만이 기형적으로 뻗어 나와 물리적 형태와 지지 구조에 오류가 있음."
       },
       {
        "label": "B",
        "direction": "두 인물 모두 시선이 아래를 향하지 않고 수평으로 왼쪽 화면 밖을 향하고 있어, 명시된 표적인 산채를 내려다보지 않고 완전히 빗나감.",
        "built_space": "계곡 아래 산채의 초가집과 목책이 레퍼런스와 일치함. 두 인물은 능선 위 흙길에 나란히 서 있으나, 산채가 인물들 너머(beyond)가 아닌 화면 왼쪽에 위치하여 제시된 배경 구도와 다름.",
        "entities": "배상도와 토미이시 모두 레퍼런스의 복장과 특징을 잘 반영함. 배상도는 텍스트에 지시되지 않은 검은 모자를 오른손에 들고 있음.",
        "hard_violations": [],
        "physics": "두 인물 모두 지면에 안정적으로 서 있으며, 배상도의 손에 쥐어진 모자도 자연스럽게 지지됨. 공중에 떠 있거나 구조적으로 불가능한 사물은 없음."
       }
      ],
      "all_candidates_fail": true
     },
     "normalized": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 5,
        "verdict_ko": "물리적 오류나 치명적 위반은 없으나, 두 인물이 수평으로 딴 곳을 응시하여 '아래의 산채를 굽어보는' 핵심 시선 방향과 행동 지시를 완전히 실패함."
       },
       {
        "label": "A",
        "score": 4,
        "verdict_ko": "프롬프트의 시선 방향과 공간 구도를 완벽하게 구현했으나, 토미이시의 손에서 자루 없는 여분의 칼날이 생성되는 치명적인 물리적 오류로 인해 사용 불가함."
       }
      ],
      "all_candidates_fail": true,
      "readings": [
       {
        "label": "A",
        "direction": "두 인물의 시선이 표적인 아래쪽 계곡의 산채를 향해 정확히 굽어보고 있음.",
        "built_space": "아래쪽 산채는 원형 목책과 초가집이 레퍼런스와 일치하게 배치됨. 두 인물은 산채가 내려다보이는 바위 능선 위 자연스러운 위치에 서 있으며, 지형의 경사가 인물들 너머로 멀어지는 카메라 구도가 정확히 구현됨.",
        "entities": "배상도(한복, 검은 조끼)와 토미이시(군복) 모두 지정된 레퍼런스의 복장과 외형에 일치함.",
        "hard_violations": [
         "물리적으로 불가능한 형태/사물: 토미이시의 왼손이 허리에 찬 칼집을 쥐고 있음에도 불구하고 동일한 손에서 손잡이가 없는 여분의 금속 칼날이 허공을 향해 비정상적으로 뻗어 나옴."
        ],
        "physics": "두 인물의 발은 바위 지형을 딛고 지지됨. 그러나 토미이시의 왼손에서 자루 없는 칼날만이 기형적으로 뻗어 나와 물리적 형태와 지지 구조에 오류가 있음."
       },
       {
        "label": "B",
        "direction": "두 인물 모두 시선이 아래를 향하지 않고 수평으로 왼쪽 화면 밖을 향하고 있어, 명시된 표적인 산채를 내려다보지 않고 완전히 빗나감.",
        "built_space": "계곡 아래 산채의 초가집과 목책이 레퍼런스와 일치함. 두 인물은 능선 위 흙길에 나란히 서 있으나, 산채가 인물들 너머(beyond)가 아닌 화면 왼쪽에 위치하여 제시된 배경 구도와 다름.",
        "entities": "배상도와 토미이시 모두 레퍼런스의 복장과 특징을 잘 반영함. 배상도는 텍스트에 지시되지 않은 검은 모자를 오른손에 들고 있음.",
        "hard_violations": [],
        "physics": "두 인물 모두 지면에 안정적으로 서 있으며, 배상도의 손에 쥐어진 모자도 자연스럽게 지지됨. 공중에 떠 있거나 구조적으로 불가능한 사물은 없음."
       }
      ]
     },
     "ok": true
    },
    {
     "model": "gpt-high",
     "order": "reverse",
     "display_to_canonical": {
      "A": "B",
      "B": "A"
     },
     "raw": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 7,
        "verdict_ko": "능선 위 두 사람의 전신과 아래 산채는 잘 담았지만, 두 사람의 시선이 산채보다 왼쪽 먼 곳을 향해 ‘아래를 굽어보는’ 핵심 행동이 약하다."
       },
       {
        "label": "B",
        "score": 9,
        "verdict_ko": "두 사람이 나란히 서서 실제 아래 산채로 고개와 시선을 내리는 전신 와이드 구도가 가장 충실하며, 배상도의 모자 누락은 남는다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "오른쪽 능선에 선 두 사람은 얼굴을 화면 왼쪽으로 돌리고 있다. 산채도 왼쪽 아래에 있지만, 턱과 눈의 방향은 비교적 수평이어서 아래 마당보다 왼쪽 먼 산림을 보는 인상이 강하다. 토미이시의 칼은 칼집에 든 채 허리에서 아래로 내려가며 특정 대상을 겨누지 않는다.",
        "built_space": "전경 오른쪽 흙 능선에 두 사람이 나란히 서 있고, 왼쪽 아래로 경사가 이어져 산채가 내려다보인다. 산채에는 긴 초가 한 동과 작은 초가 한 동, 마당을 두른 목책 한 겹, 전면 출입구와 오른쪽 측면 틈이 보인다. 초가 앞 바위 무리와 오른쪽 둥근 바위, 목책 바깥 산길도 있어 장소 참고의 주요 배치를 잘 따른다. 두 사람 모두 발끝까지 포함되며 산채는 원경 크기로 표현된다.",
        "entities": "성인 동아시아계 남성 두 명만 보이며, 외모만으로 국적을 확정할 수는 없다. 배상도는 검은 머리, 흰 한복, 검은 조끼, 허리 장식과 검은 신발을 갖췄고 참고의 검은 모자는 쓰지 않고 손에 들었다. 토미이시는 수염, 군모, 갈색 군복, 훈장, 가죽 띠와 장화가 참고와 대체로 일치하며 허리에 칼을 찼다. 두 얼굴은 참고 인물과 대체로 부합한다. 읽을 수 있는 글씨나 추가 인물은 없다.",
        "hard_violations": [],
        "physics": "두 사람의 신발과 장화는 흙바닥에 닿아 있고, 벌린 발이 체중을 받친다. 배상도의 모자는 손에 잡혀 있으며 토미이시의 칼집은 허리 장구에 매달려 있다. 떠 있는 신체나 지지 없는 물체는 보이지 않는다. 서 있는 자세는 가능하지만 고개를 숙여 내려다보는 동작은 약하다."
       },
       {
        "label": "B",
        "direction": "왼쪽 전경의 두 사람이 몸을 오른쪽 아래 산채로 향하고 있다. 배상도는 고개를 분명히 숙여 마당을 바라보고, 토미이시 역시 얼굴을 같은 아래쪽 산채로 돌려 두 사람의 공통 관찰 대상이 명확하다. 토미이시가 든 칼날은 뒤쪽 왼쪽 아래로 향하며 산채를 겨누는 자세는 아니다.",
        "built_space": "카메라는 두 사람 뒤의 높은 바위 능선에서 아래 산채를 본다. 두 사람은 나란히 서 있고 머리부터 신발까지 모두 보인다. 전경 암반에서 산채 방향으로 지형이 내려간다. 산채에는 긴 초가 한 동과 작은 초가 한 동, 둘레 목책 한 겹과 전면 출입구 하나가 보인다. 초가 앞 바위 무리, 오른쪽 둥근 바위, 흙마당과 외곽 산길이 참고의 구성을 유지한다. 참고의 오른쪽 측면 개구부는 뚜렷하지 않지만 주요 구조와 재료는 잘 맞는다.",
        "entities": "성인 남성 두 명만 등장한다. 뒤쪽 모습이므로 얼굴과 국적의 정확한 일치는 확인하기 어렵다. 토미이시는 군모, 갈색 군복, 가죽 띠, 장화와 허리의 칼집을 갖추고 손에 뽑은 칼을 들었다. 배상도는 자연스러운 검은 머리, 검은 조끼, 흰 한복과 검은 신발이 참고와 부합하지만 검은 모자는 보이지 않는다. 뒤로 가려진 훈장과 앞쪽 장식은 확인할 수 없다. 추가 인물이나 읽을 수 있는 글씨는 없다.",
        "hard_violations": [],
        "physics": "두 사람의 양발이 흙과 암반에 놓여 체중을 지탱한다. 배상도의 숙인 고개와 약간 앞으로 향한 상체는 능선 아래를 살피는 정지 동작으로 자연스럽다. 토미이시의 오른손이 칼 손잡이를 쥐고 있으며, 칼집은 허리에 매달려 있다. 칼날이 공중으로 뻗어 있어도 손의 지지가 명확하고, 지지 없이 떠 있는 신체나 물체는 없다."
       }
      ],
      "all_candidates_fail": false
     },
     "normalized": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 7,
        "verdict_ko": "능선 위 두 사람의 전신과 아래 산채는 잘 담았지만, 두 사람의 시선이 산채보다 왼쪽 먼 곳을 향해 ‘아래를 굽어보는’ 핵심 행동이 약하다."
       },
       {
        "label": "A",
        "score": 9,
        "verdict_ko": "두 사람이 나란히 서서 실제 아래 산채로 고개와 시선을 내리는 전신 와이드 구도가 가장 충실하며, 배상도의 모자 누락은 남는다."
       }
      ],
      "all_candidates_fail": false,
      "readings": [
       {
        "label": "B",
        "direction": "오른쪽 능선에 선 두 사람은 얼굴을 화면 왼쪽으로 돌리고 있다. 산채도 왼쪽 아래에 있지만, 턱과 눈의 방향은 비교적 수평이어서 아래 마당보다 왼쪽 먼 산림을 보는 인상이 강하다. 토미이시의 칼은 칼집에 든 채 허리에서 아래로 내려가며 특정 대상을 겨누지 않는다.",
        "built_space": "전경 오른쪽 흙 능선에 두 사람이 나란히 서 있고, 왼쪽 아래로 경사가 이어져 산채가 내려다보인다. 산채에는 긴 초가 한 동과 작은 초가 한 동, 마당을 두른 목책 한 겹, 전면 출입구와 오른쪽 측면 틈이 보인다. 초가 앞 바위 무리와 오른쪽 둥근 바위, 목책 바깥 산길도 있어 장소 참고의 주요 배치를 잘 따른다. 두 사람 모두 발끝까지 포함되며 산채는 원경 크기로 표현된다.",
        "entities": "성인 동아시아계 남성 두 명만 보이며, 외모만으로 국적을 확정할 수는 없다. 배상도는 검은 머리, 흰 한복, 검은 조끼, 허리 장식과 검은 신발을 갖췄고 참고의 검은 모자는 쓰지 않고 손에 들었다. 토미이시는 수염, 군모, 갈색 군복, 훈장, 가죽 띠와 장화가 참고와 대체로 일치하며 허리에 칼을 찼다. 두 얼굴은 참고 인물과 대체로 부합한다. 읽을 수 있는 글씨나 추가 인물은 없다.",
        "hard_violations": [],
        "physics": "두 사람의 신발과 장화는 흙바닥에 닿아 있고, 벌린 발이 체중을 받친다. 배상도의 모자는 손에 잡혀 있으며 토미이시의 칼집은 허리 장구에 매달려 있다. 떠 있는 신체나 지지 없는 물체는 보이지 않는다. 서 있는 자세는 가능하지만 고개를 숙여 내려다보는 동작은 약하다."
       },
       {
        "label": "A",
        "direction": "왼쪽 전경의 두 사람이 몸을 오른쪽 아래 산채로 향하고 있다. 배상도는 고개를 분명히 숙여 마당을 바라보고, 토미이시 역시 얼굴을 같은 아래쪽 산채로 돌려 두 사람의 공통 관찰 대상이 명확하다. 토미이시가 든 칼날은 뒤쪽 왼쪽 아래로 향하며 산채를 겨누는 자세는 아니다.",
        "built_space": "카메라는 두 사람 뒤의 높은 바위 능선에서 아래 산채를 본다. 두 사람은 나란히 서 있고 머리부터 신발까지 모두 보인다. 전경 암반에서 산채 방향으로 지형이 내려간다. 산채에는 긴 초가 한 동과 작은 초가 한 동, 둘레 목책 한 겹과 전면 출입구 하나가 보인다. 초가 앞 바위 무리, 오른쪽 둥근 바위, 흙마당과 외곽 산길이 참고의 구성을 유지한다. 참고의 오른쪽 측면 개구부는 뚜렷하지 않지만 주요 구조와 재료는 잘 맞는다.",
        "entities": "성인 남성 두 명만 등장한다. 뒤쪽 모습이므로 얼굴과 국적의 정확한 일치는 확인하기 어렵다. 토미이시는 군모, 갈색 군복, 가죽 띠, 장화와 허리의 칼집을 갖추고 손에 뽑은 칼을 들었다. 배상도는 자연스러운 검은 머리, 검은 조끼, 흰 한복과 검은 신발이 참고와 부합하지만 검은 모자는 보이지 않는다. 뒤로 가려진 훈장과 앞쪽 장식은 확인할 수 없다. 추가 인물이나 읽을 수 있는 글씨는 없다.",
        "hard_violations": [],
        "physics": "두 사람의 양발이 흙과 암반에 놓여 체중을 지탱한다. 배상도의 숙인 고개와 약간 앞으로 향한 상체는 능선 아래를 살피는 정지 동작으로 자연스럽다. 토미이시의 오른손이 칼 손잡이를 쥐고 있으며, 칼집은 허리에 매달려 있다. 칼날이 공중으로 뻗어 있어도 손의 지지가 명확하고, 지지 없이 떠 있는 신체나 물체는 없다."
       }
      ]
     },
     "ok": true
    }
   ],
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "slot_winner_match": false,
   "slot_winner": {
    "gemini-pro": "B",
    "gpt-high": "A"
   },
   "route": "cross_slot_combined"
  },
  "dual": {
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "normalized": {
    "A": 1.8,
    "B": 1.778
   },
   "adjusted": {
    "A": 1.55,
    "B": 1.778
   },
   "violations": {
    "A": [
     "[gemini-pro] 물리적으로 불가능한 형태/사물: 토미이시의 왼손이 허리에 찬 칼집을 쥐고 있음에도 불구하고 동일한 손에서 손잡이가 없는 여분의 금속 칼날이 허공을 향해 비정상적으로 뻗어 나옴."
    ]
   },
   "per_model_winner": {
    "gemini-pro": "B",
    "gpt-high": "A"
   },
   "agreed": false
  },
  "totals": {
   "B": 1778,
   "A": 1550
  },
  "selected": "B",
  "ranking": [
   "B",
   "A"
  ],
  "verdicts": [
   {
    "label": "B",
    "score": 1778,
    "verdict_ko": "물리적 오류나 치명적 위반은 없으나, 두 인물이 수평으로 딴 곳을 응시하여 '아래의 산채를 굽어보는' 핵심 시선 방향과 행동 지시를 완전히 실패함."
   },
   {
    "label": "A",
    "score": 1550,
    "verdict_ko": "프롬프트의 시선 방향과 공간 구도를 완벽하게 구현했으나, 토미이시의 손에서 자루 없는 여분의 칼날이 생성되는 치명적인 물리적 오류로 인해 사용 불가함.  ★위반: [gemini-pro] 물리적으로 불가능한 형태/사물: 토미이시의 왼손이 허리에 찬 칼집을 쥐고 있음에도 불구하고 동일한 손에서 손잡이가 없는 여분의 금속 칼날이 허공을 향해 비정상적으로 뻗어 나옴."
   }
  ],
  "refs": [
   {
    "label": "LOCATION STRUCTURE PHOTOGRAPH — the confirmed photograph of this exact place and its fixed structure: it is the SINGLE authority for the location, the structure's shape, proportions, materials, colors, openings and every permanent site detail. Never copy its camera framing, time of day or lighting — the shot text is the authority for those.",
    "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/db00517b-165b-4005-a287-f28936b59b95/images/background_chain/seed_bg_mountain_camp_sel.png",
    "asset_id": "8c17c7ae-efff-4bb5-b88c-7c5528f00716",
    "role": "location_seed_bg"
   },
   {
    "label": "CHARACTER REFERENCE — 배상도: the exact person appearing in this shot; match face, hair and build exactly.",
    "path": "<bytes:851718>",
    "asset_id": "034f5325-975e-4474-854b-e48359710bf8",
    "role": "character_ref"
   },
   {
    "label": "CHARACTER REFERENCE — 토미이시: the exact person appearing in this shot; match face, hair and build exactly.",
    "path": "<bytes:797699>",
    "asset_id": "c861ba0f-8750-4262-8324-25cf5c6c9dcd",
    "role": "character_ref"
   }
  ],
  "critique_skipped": true,
  "shot_run_produced": false,
  "shot_run_uid": "06aa1436-59a8-7b19-89d2-d978735d47ae",
  "bgfirst": {
   "bg_path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/db00517b-165b-4005-a287-f28936b59b95/images/272bd7af-9294-4490-98c6-88a40e9deacc/scene/recipe/S52sh1__bgfirst_bg.png",
   "bg_asset_id": "2bb90254-cd01-44d5-8d2b-7c078f8f33a1",
   "bg_record_key": "S52sh1::bgfirst_bg",
   "chain_winner": false,
   "authority": "seed_bg"
  },
  "ref_mode": "seed-bg+엔티티 (2택1: 무콘티 승)",
  "share_plan": {
   "ref_plan": "background"
  },
  "lane_policy": "ab_select_ready"
 },
 "S52sh1::cine": {
  "applied": true,
  "attempted_at": "2026-09-09T07:01:47.219143+00:00",
  "fingerprint": "b96dbd1c9a7eb22b026fc3d1afafd66f7c759e6223fc66daface19ce832f56db",
  "fingerprint_version": 2,
  "provider": "mai",
  "endpoint": "openrouter/chat-completions",
  "model": "microsoft/mai-image-2.6",
  "pack": "24.202608252115",
  "source_file": "S52sh1_sel.png",
  "source_sha256": "3faebf788766491844d117bf2a9b626adb935379ad83e4808443790e03612117",
  "file": "S52sh1_cine.png",
  "staged_sha256": "bc29bfc7f7d225e8fe82f056d7f9f435cb318f8ed8f13489006802cb78104b02",
  "latency_ms": 28401
 },
 "S52sh6::signage": {
  "fp": "0c32fdb4c6ae0534",
  "inscriptions": [],
  "cues": [],
  "dropped": []
 },
 "S52sh6": {
  "input_fingerprint": "7d797fe3798744cc",
  "prompt": "Create ONE FINAL photorealistic live-action film still of the moment below — early 20th-century Korea, Jeolla region; all people are Korean unless explicitly stated otherwise. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 토미이시의 수신호에 양쪽 숲길로 흩어지며 몸을 날린 mid-stride 자세의 일본인 군인들의 역동적인 뒷모습들.\n\nLOCATION (lock): The two forest approaches descending from the overlooking ridge around the mountain camp's flanks. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: wide shot\n- FRAME LAYOUT: Left forest route receiving the left soldier group in the middle-left of the frame, background; Right forest route receiving the right soldier group in the middle-right of the frame, background.\n- KEY BACKGROUND ELEMENTS: Left forest route (Taken by one part of the separating force) — Recedes downhill toward screen left; used as Gives the leftward movement a clear spatial destination; Right forest route (Taken by the other part of the separating force) — Recedes downhill toward screen right; used as Completes the encircling geometry while keeping the central gap readable.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Retain the ridge's natural daylight and muted saturation, with consistent contrast as the soldiers move toward the forest routes.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The people visible in that photo are NOT in this shot: never carry their faces, bodies or clothing onto anyone here. Each person in THIS shot is defined solely by the PEOPLE list, their own CHARACTER REFERENCE images and their wardrobe notes. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The wooded approaches form a ring around the camp.\n\nNo people appear unless the shot text itself says so.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.",
  "roll_prompts": {
   "A": "Create ONE FINAL photorealistic live-action film still of the moment below — early 20th-century Korea, Jeolla region; all people are Korean unless explicitly stated otherwise. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 토미이시의 수신호에 양쪽 숲길로 흩어지며 몸을 날린 mid-stride 자세의 일본인 군인들의 역동적인 뒷모습들.\n\nLOCATION (lock): The two forest approaches descending from the overlooking ridge around the mountain camp's flanks. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: wide shot\n- FRAME LAYOUT: Left forest route receiving the left soldier group in the middle-left of the frame, background; Right forest route receiving the right soldier group in the middle-right of the frame, background.\n- KEY BACKGROUND ELEMENTS: Left forest route (Taken by one part of the separating force) — Recedes downhill toward screen left; used as Gives the leftward movement a clear spatial destination; Right forest route (Taken by the other part of the separating force) — Recedes downhill toward screen right; used as Completes the encircling geometry while keeping the central gap readable.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Retain the ridge's natural daylight and muted saturation, with consistent contrast as the soldiers move toward the forest routes.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The people visible in that photo are NOT in this shot: never carry their faces, bodies or clothing onto anyone here. Each person in THIS shot is defined solely by the PEOPLE list, their own CHARACTER REFERENCE images and their wardrobe notes. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The wooded approaches form a ring around the camp.\n\nNo people appear unless the shot text itself says so.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.",
   "B": "Create ONE FINAL photorealistic live-action film still of the moment below — early 20th-century Korea, Jeolla region; all people are Korean unless explicitly stated otherwise. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 토미이시의 수신호에 양쪽 숲길로 흩어지며 몸을 날린 mid-stride 자세의 일본인 군인들의 역동적인 뒷모습들.\n\nLOCATION (lock): The two forest approaches descending from the overlooking ridge around the mountain camp's flanks. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: wide shot\n- FRAME LAYOUT: Left forest route receiving the left soldier group in the middle-left of the frame, background; Right forest route receiving the right soldier group in the middle-right of the frame, background.\n- KEY BACKGROUND ELEMENTS: Left forest route (Taken by one part of the separating force) — Recedes downhill toward screen left; used as Gives the leftward movement a clear spatial destination; Right forest route (Taken by the other part of the separating force) — Recedes downhill toward screen right; used as Completes the encircling geometry while keeping the central gap readable.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Retain the ridge's natural daylight and muted saturation, with consistent contrast as the soldiers move toward the forest routes.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The people visible in that photo are NOT in this shot: never carry their faces, bodies or clothing onto anyone here. Each person in THIS shot is defined solely by the PEOPLE list, their own CHARACTER REFERENCE images and their wardrobe notes. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The wooded approaches form a ring around the camp.\n\nNo people appear unless the shot text itself says so.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.\n\nCOMPOSITION VARIATION (this candidate only): keep the place, the people and the moment exactly as specified, and keep from the CAMERA & FRAME contract its subject scale, screen placement and key background elements exactly. For this candidate only, its camera angle and height are deliberately left open: choose a camera position distinctly different from the obvious one, as a film director picking a second setup on the same blocking. Stay true to any explicit framing the SHOT TEXT itself dictates."
  },
  "shot_run_spend_attempt_count": 1,
  "readings": [
   {
    "label": "A",
    "direction": "중앙의 장교가 우측을 가리키고 있으며, 군인들은 좌우 숲길을 향해 이동하는 방향성을 보여줍니다.",
    "built_space": "초가집과 나무 울타리, 좌우로 나뉘는 숲길 등 레퍼런스의 배경 요소들이 올바른 위치에 구현되어 있습니다.",
    "entities": "일본인 군인들이 등장하나, 이전 샷의 인물들을 이 샷에 포함하지 말라는 지침에도 불구하고 레퍼런스 이미지의 장교가 중앙에 그대로 복제되어 나타났습니다.",
    "hard_violations": [
     "[gemini-pro] 레퍼런스 이미지의 인물을 이 샷에 등장시키지 말라는 지침을 어기고 장교를 포함함",
     "[gemini-pro] 좌측 하단 군인의 오른쪽 다리가 생략되어 해부학적으로 불가능한 신체 구조를 보임",
     "[gpt-high] 이전 사진의 훈장 단 장교와 그 의상을 중앙 인물로 재사용하여, 이전 사진 인물의 얼굴·몸·의상을 이번 장면에 옮기지 말라는 명시적 제외 조건을 위반한다.",
     "[gpt-high] 오른쪽 병력을 아래로 내려가는 숲길이 아니라 능선 위쪽으로 진행하도록 배치하여, 명시된 오른쪽 하강 동선을 반대로 구현한다."
    ],
    "physics": "좌측 하단의 군인은 몸을 지탱해야 할 오른쪽 다리가 사라져 공중에 떠 있는 듯한 오류가 있으며, 우측 끝의 군인 역시 다리와 지면의 접촉이 불분명하여 물리적으로 어색합니다."
   },
   {
    "label": "B",
    "direction": "모든 군인의 몸체, 무기, 시선이 좌측 및 우측 숲길이라는 명확한 목적지를 향해 아래로 향하고 있습니다.",
    "built_space": "산 아래 분지의 막사, 초가집 두 채, 곡선형의 나무 울타리 등 레퍼런스의 고정된 지형지물이 정확한 위치와 스케일로 구현되었으며, 좌우로 이어지는 숲길의 배치도 프롬프트와 일치합니다.",
    "entities": "군장과 소총을 휴대한 일본인 군인들의 뒷모습이 명확하게 묘사되었으며, 프롬프트의 지시에 따라 이전 샷(레퍼런스)에 있던 인물들은 화면에서 배제되었습니다.",
    "hard_violations": [
     "[gpt-high] 왼쪽 배경의 숲길이 화면 왼쪽으로 멀어질수록 올라가며, 명시된 왼쪽 내리막 경로의 수직 방향을 뒤집는다."
    ],
    "physics": "경사진 지형을 뛰어 내려가는 군인들의 mid-stride 자세가 자연스러우며, 모든 인물의 디딤발이 땅에 안정적으로 접지되어 몸을 지탱하고 있습니다."
   }
  ],
  "cross_model_order": {
   "policy": "cross_model_order_v2_slot0_fwd_slot1_rev_parallel2",
   "slots": [
    {
     "model": "gemini-pro",
     "order": "forward",
     "display_to_canonical": {
      "A": "A",
      "B": "B"
     },
     "raw": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 9,
        "verdict_ko": "레퍼런스 이미지의 인물을 배제하라는 부정 지침을 완벽히 준수하였으며, 지시된 프레이밍 내에서 양쪽 숲길로 흩어지는 군인들의 역동적인 뒷모습을 물리적 오류 없이 훌륭하게 구현했습니다."
       },
       {
        "label": "A",
        "score": 2,
        "verdict_ko": "레퍼런스 이미지에 등장한 인물(장교)을 절대 포함하지 말라는 명시적인 지침을 위반하였고, 일부 군인의 신체 구조에 심각한 오류가 발생했습니다."
       }
      ],
      "readings": [
       {
        "label": "B",
        "direction": "모든 군인의 몸체, 무기, 시선이 좌측 및 우측 숲길이라는 명확한 목적지를 향해 아래로 향하고 있습니다.",
        "built_space": "산 아래 분지의 막사, 초가집 두 채, 곡선형의 나무 울타리 등 레퍼런스의 고정된 지형지물이 정확한 위치와 스케일로 구현되었으며, 좌우로 이어지는 숲길의 배치도 프롬프트와 일치합니다.",
        "entities": "군장과 소총을 휴대한 일본인 군인들의 뒷모습이 명확하게 묘사되었으며, 프롬프트의 지시에 따라 이전 샷(레퍼런스)에 있던 인물들은 화면에서 배제되었습니다.",
        "hard_violations": [],
        "physics": "경사진 지형을 뛰어 내려가는 군인들의 mid-stride 자세가 자연스러우며, 모든 인물의 디딤발이 땅에 안정적으로 접지되어 몸을 지탱하고 있습니다."
       },
       {
        "label": "A",
        "direction": "중앙의 장교가 우측을 가리키고 있으며, 군인들은 좌우 숲길을 향해 이동하는 방향성을 보여줍니다.",
        "built_space": "초가집과 나무 울타리, 좌우로 나뉘는 숲길 등 레퍼런스의 배경 요소들이 올바른 위치에 구현되어 있습니다.",
        "entities": "일본인 군인들이 등장하나, 이전 샷의 인물들을 이 샷에 포함하지 말라는 지침에도 불구하고 레퍼런스 이미지의 장교가 중앙에 그대로 복제되어 나타났습니다.",
        "hard_violations": [
         "레퍼런스 이미지의 인물을 이 샷에 등장시키지 말라는 지침을 어기고 장교를 포함함",
         "좌측 하단 군인의 오른쪽 다리가 생략되어 해부학적으로 불가능한 신체 구조를 보임"
        ],
        "physics": "좌측 하단의 군인은 몸을 지탱해야 할 오른쪽 다리가 사라져 공중에 떠 있는 듯한 오류가 있으며, 우측 끝의 군인 역시 다리와 지면의 접촉이 불분명하여 물리적으로 어색합니다."
       }
      ],
      "all_candidates_fail": false
     },
     "normalized": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 9,
        "verdict_ko": "레퍼런스 이미지의 인물을 배제하라는 부정 지침을 완벽히 준수하였으며, 지시된 프레이밍 내에서 양쪽 숲길로 흩어지는 군인들의 역동적인 뒷모습을 물리적 오류 없이 훌륭하게 구현했습니다."
       },
       {
        "label": "A",
        "score": 2,
        "verdict_ko": "레퍼런스 이미지에 등장한 인물(장교)을 절대 포함하지 말라는 명시적인 지침을 위반하였고, 일부 군인의 신체 구조에 심각한 오류가 발생했습니다."
       }
      ],
      "all_candidates_fail": false,
      "readings": [
       {
        "label": "B",
        "direction": "모든 군인의 몸체, 무기, 시선이 좌측 및 우측 숲길이라는 명확한 목적지를 향해 아래로 향하고 있습니다.",
        "built_space": "산 아래 분지의 막사, 초가집 두 채, 곡선형의 나무 울타리 등 레퍼런스의 고정된 지형지물이 정확한 위치와 스케일로 구현되었으며, 좌우로 이어지는 숲길의 배치도 프롬프트와 일치합니다.",
        "entities": "군장과 소총을 휴대한 일본인 군인들의 뒷모습이 명확하게 묘사되었으며, 프롬프트의 지시에 따라 이전 샷(레퍼런스)에 있던 인물들은 화면에서 배제되었습니다.",
        "hard_violations": [],
        "physics": "경사진 지형을 뛰어 내려가는 군인들의 mid-stride 자세가 자연스러우며, 모든 인물의 디딤발이 땅에 안정적으로 접지되어 몸을 지탱하고 있습니다."
       },
       {
        "label": "A",
        "direction": "중앙의 장교가 우측을 가리키고 있으며, 군인들은 좌우 숲길을 향해 이동하는 방향성을 보여줍니다.",
        "built_space": "초가집과 나무 울타리, 좌우로 나뉘는 숲길 등 레퍼런스의 배경 요소들이 올바른 위치에 구현되어 있습니다.",
        "entities": "일본인 군인들이 등장하나, 이전 샷의 인물들을 이 샷에 포함하지 말라는 지침에도 불구하고 레퍼런스 이미지의 장교가 중앙에 그대로 복제되어 나타났습니다.",
        "hard_violations": [
         "레퍼런스 이미지의 인물을 이 샷에 등장시키지 말라는 지침을 어기고 장교를 포함함",
         "좌측 하단 군인의 오른쪽 다리가 생략되어 해부학적으로 불가능한 신체 구조를 보임"
        ],
        "physics": "좌측 하단의 군인은 몸을 지탱해야 할 오른쪽 다리가 사라져 공중에 떠 있는 듯한 오류가 있으며, 우측 끝의 군인 역시 다리와 지면의 접촉이 불분명하여 물리적으로 어색합니다."
       }
      ]
     },
     "ok": true
    },
    {
     "model": "gpt-high",
     "order": "reverse",
     "display_to_canonical": {
      "A": "B",
      "B": "A"
     },
     "raw": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 3,
        "verdict_ko": "병사들의 역동적인 뒷모습과 중앙 틈은 더 충실하지만, 병력이 배경이 아닌 전경을 차지하고 왼쪽 숲길이 내리막과 반대로 이어진다."
       },
       {
        "label": "B",
        "score": 1,
        "verdict_ko": "이전 장면의 장교와 의상을 중앙에 재사용하고 오른쪽 병력을 내리막 숲길 대신 능선 위로 보내어 인물 제외 조건과 양쪽 하강 배치를 어긴다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "왼쪽 세 명은 카메라에서 멀어지며 중앙 왼쪽의 비탈로, 오른쪽 네 명은 중앙 오른쪽으로 움직인다. 다만 두 무리 모두 아직 가까운 능선에 있어 좌우 배경 숲길로 흩어진 상태는 아니다. 왼쪽에 길게 보이는 길은 화면 왼쪽 위로 올라가며, 요구된 왼쪽 하강 목적지와 반대다. 얼굴은 대부분 가려져 정확한 시선은 확인하기 어렵다. 소총은 위쪽 대각선이나 진행 방향 옆으로 운반되며 특정 사람을 조준하지 않는다.",
        "built_space": "아래쪽 야영지에 초가 건물 두 동, 마당 하나, 이를 두른 목책과 앞쪽 출입구 하나가 보인다. 바위와 침엽수림도 참고 장소와 대체로 이어진다. 야영지 양옆 길은 보이지만 왼쪽 길의 원경 구간은 상승한다. 병사들은 중간 왼쪽·오른쪽 배경이 아니라 화면 하단 전경에 크게 배치되었고, 오른쪽 끝 인물은 잘린다. 중앙에는 인물 없는 틈이 남아 있다.",
        "entities": "군복과 군모를 쓴 성인 남성 병사 일곱 명이 보이며, 소총·배낭·침구 묶음·가죽 장구류를 갖추었다. 대부분 뒷모습이라 일본인이라는 국적이나 구체적인 얼굴 정체성은 외형으로 검증할 수 없다. 이전 사진의 흰 한복 인물과 훈장 단 장교는 보이지 않는다. 토미이시의 수신호 자체는 화면에 없지만 신호 이후의 출발 순간으로 읽을 수 있다. 읽을 수 있는 글자는 없다.",
        "hard_violations": [
         "왼쪽 배경의 숲길이 화면 왼쪽으로 멀어질수록 올라가며, 명시된 왼쪽 내리막 경로의 수직 방향을 뒤집는다."
        ],
        "physics": "병사들은 무릎을 굽히고 상체를 앞으로 기울인 달리기 자세다. 여러 인물은 한쪽 발로 흙길을 딛고 있으며, 양발이 잠시 떨어진 인물도 뒤로 찬 다리와 앞으로 내민 다리, 바로 아래의 착지 지면이 있어 달리기의 도약 단계로 설명된다. 소총은 손이나 멜빵으로, 배낭과 침구는 어깨끈과 결속끈으로 지지된다. 근거 없이 공중에 떠 있는 몸이나 장비는 보이지 않는다."
       },
       {
        "label": "B",
        "direction": "중앙 장교는 왼팔과 고개를 화면 왼쪽 야영지 접근 방향으로 향한다. 왼쪽 네 병사는 왼쪽 아래 비탈로 움직이지만, 오른쪽 네 병사는 화면 오른쪽 위의 능선과 나무 사이로 이동하여 아래쪽 야영지 측면으로 내려가는 방향이 아니다. 오른쪽 하강 숲길을 목적지로 삼는 동선이 보이지 않는다. 병사들의 시선은 대부분 이동 방향을 따르며, 소총은 조준하지 않고 휴대한다.",
        "built_space": "초가 두 동, 목책으로 둘러싸인 마당 하나와 출입구 하나, 바위와 능선의 나무 배치가 참고 사진과 매우 가깝다. 그러나 참고 사진의 카메라 구도를 거의 유지한 채 병력을 가까운 능선에 배치했다. 왼쪽 무리는 전경 비탈에, 오른쪽 무리는 전경·중경의 능선 위에 있으며, 요구된 양쪽 배경 숲길 배치가 아니다. 중앙 장교가 비어 있어야 할 분리 공간을 차지한다.",
        "entities": "중앙 장교 한 명과 양쪽 병사 여덟 명으로, 성인 남성 아홉 명이 보인다. 중앙 인물은 참고 사진 장교의 얼굴 인상, 챙 있는 군모, 훈장 배치와 군복을 재현한다. 이는 이전 사진 인물의 얼굴·몸·의상을 이어 쓰지 말라는 조건과 충돌한다. 병사들도 다수가 챙 있는 장교형 군모와 장화를 착용한다. 일본인이라는 국적은 화면만으로 확인할 수 없다. 흰 한복 인물과 읽을 수 있는 글자는 없다.",
        "hard_violations": [
         "이전 사진의 훈장 단 장교와 그 의상을 중앙 인물로 재사용하여, 이전 사진 인물의 얼굴·몸·의상을 이번 장면에 옮기지 말라는 명시적 제외 조건을 위반한다.",
         "오른쪽 병력을 아래로 내려가는 숲길이 아니라 능선 위쪽으로 진행하도록 배치하여, 명시된 오른쪽 하강 동선을 반대로 구현한다."
        ],
        "physics": "중앙 장교는 두 장화로 땅을 딛고 수신호를 보낸다. 왼쪽 병사들은 비탈이나 바위에 발을 디디며 내려가고, 크게 보폭을 벌린 병사도 지면 접촉과 다음 착지 공간이 있다. 오른쪽 병사들의 굽힌 무릎과 뒤로 뻗은 다리 역시 달리기로 설명된다. 소총과 칼집은 손·멜빵·허리띠에 지지된다. 지지나 도약 근거 없이 떠 있는 인물은 보이지 않는다."
       }
      ],
      "all_candidates_fail": true
     },
     "normalized": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 3,
        "verdict_ko": "병사들의 역동적인 뒷모습과 중앙 틈은 더 충실하지만, 병력이 배경이 아닌 전경을 차지하고 왼쪽 숲길이 내리막과 반대로 이어진다."
       },
       {
        "label": "A",
        "score": 1,
        "verdict_ko": "이전 장면의 장교와 의상을 중앙에 재사용하고 오른쪽 병력을 내리막 숲길 대신 능선 위로 보내어 인물 제외 조건과 양쪽 하강 배치를 어긴다."
       }
      ],
      "all_candidates_fail": true,
      "readings": [
       {
        "label": "B",
        "direction": "왼쪽 세 명은 카메라에서 멀어지며 중앙 왼쪽의 비탈로, 오른쪽 네 명은 중앙 오른쪽으로 움직인다. 다만 두 무리 모두 아직 가까운 능선에 있어 좌우 배경 숲길로 흩어진 상태는 아니다. 왼쪽에 길게 보이는 길은 화면 왼쪽 위로 올라가며, 요구된 왼쪽 하강 목적지와 반대다. 얼굴은 대부분 가려져 정확한 시선은 확인하기 어렵다. 소총은 위쪽 대각선이나 진행 방향 옆으로 운반되며 특정 사람을 조준하지 않는다.",
        "built_space": "아래쪽 야영지에 초가 건물 두 동, 마당 하나, 이를 두른 목책과 앞쪽 출입구 하나가 보인다. 바위와 침엽수림도 참고 장소와 대체로 이어진다. 야영지 양옆 길은 보이지만 왼쪽 길의 원경 구간은 상승한다. 병사들은 중간 왼쪽·오른쪽 배경이 아니라 화면 하단 전경에 크게 배치되었고, 오른쪽 끝 인물은 잘린다. 중앙에는 인물 없는 틈이 남아 있다.",
        "entities": "군복과 군모를 쓴 성인 남성 병사 일곱 명이 보이며, 소총·배낭·침구 묶음·가죽 장구류를 갖추었다. 대부분 뒷모습이라 일본인이라는 국적이나 구체적인 얼굴 정체성은 외형으로 검증할 수 없다. 이전 사진의 흰 한복 인물과 훈장 단 장교는 보이지 않는다. 토미이시의 수신호 자체는 화면에 없지만 신호 이후의 출발 순간으로 읽을 수 있다. 읽을 수 있는 글자는 없다.",
        "hard_violations": [
         "왼쪽 배경의 숲길이 화면 왼쪽으로 멀어질수록 올라가며, 명시된 왼쪽 내리막 경로의 수직 방향을 뒤집는다."
        ],
        "physics": "병사들은 무릎을 굽히고 상체를 앞으로 기울인 달리기 자세다. 여러 인물은 한쪽 발로 흙길을 딛고 있으며, 양발이 잠시 떨어진 인물도 뒤로 찬 다리와 앞으로 내민 다리, 바로 아래의 착지 지면이 있어 달리기의 도약 단계로 설명된다. 소총은 손이나 멜빵으로, 배낭과 침구는 어깨끈과 결속끈으로 지지된다. 근거 없이 공중에 떠 있는 몸이나 장비는 보이지 않는다."
       },
       {
        "label": "A",
        "direction": "중앙 장교는 왼팔과 고개를 화면 왼쪽 야영지 접근 방향으로 향한다. 왼쪽 네 병사는 왼쪽 아래 비탈로 움직이지만, 오른쪽 네 병사는 화면 오른쪽 위의 능선과 나무 사이로 이동하여 아래쪽 야영지 측면으로 내려가는 방향이 아니다. 오른쪽 하강 숲길을 목적지로 삼는 동선이 보이지 않는다. 병사들의 시선은 대부분 이동 방향을 따르며, 소총은 조준하지 않고 휴대한다.",
        "built_space": "초가 두 동, 목책으로 둘러싸인 마당 하나와 출입구 하나, 바위와 능선의 나무 배치가 참고 사진과 매우 가깝다. 그러나 참고 사진의 카메라 구도를 거의 유지한 채 병력을 가까운 능선에 배치했다. 왼쪽 무리는 전경 비탈에, 오른쪽 무리는 전경·중경의 능선 위에 있으며, 요구된 양쪽 배경 숲길 배치가 아니다. 중앙 장교가 비어 있어야 할 분리 공간을 차지한다.",
        "entities": "중앙 장교 한 명과 양쪽 병사 여덟 명으로, 성인 남성 아홉 명이 보인다. 중앙 인물은 참고 사진 장교의 얼굴 인상, 챙 있는 군모, 훈장 배치와 군복을 재현한다. 이는 이전 사진 인물의 얼굴·몸·의상을 이어 쓰지 말라는 조건과 충돌한다. 병사들도 다수가 챙 있는 장교형 군모와 장화를 착용한다. 일본인이라는 국적은 화면만으로 확인할 수 없다. 흰 한복 인물과 읽을 수 있는 글자는 없다.",
        "hard_violations": [
         "이전 사진의 훈장 단 장교와 그 의상을 중앙 인물로 재사용하여, 이전 사진 인물의 얼굴·몸·의상을 이번 장면에 옮기지 말라는 명시적 제외 조건을 위반한다.",
         "오른쪽 병력을 아래로 내려가는 숲길이 아니라 능선 위쪽으로 진행하도록 배치하여, 명시된 오른쪽 하강 동선을 반대로 구현한다."
        ],
        "physics": "중앙 장교는 두 장화로 땅을 딛고 수신호를 보낸다. 왼쪽 병사들은 비탈이나 바위에 발을 디디며 내려가고, 크게 보폭을 벌린 병사도 지면 접촉과 다음 착지 공간이 있다. 오른쪽 병사들의 굽힌 무릎과 뒤로 뻗은 다리 역시 달리기로 설명된다. 소총과 칼집은 손·멜빵·허리띠에 지지된다. 지지나 도약 근거 없이 떠 있는 인물은 보이지 않는다."
       }
      ]
     },
     "ok": true
    }
   ],
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "slot_winner_match": true,
   "slot_winner": {
    "gemini-pro": "B",
    "gpt-high": "B"
   },
   "route": "cross_slot_agree"
  },
  "dual": {
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "normalized": {
    "A": 0.556,
    "B": 2.0
   },
   "adjusted": {
    "A": 0.306,
    "B": 1.75
   },
   "violations": {
    "A": [
     "[gemini-pro] 레퍼런스 이미지의 인물을 이 샷에 등장시키지 말라는 지침을 어기고 장교를 포함함",
     "[gemini-pro] 좌측 하단 군인의 오른쪽 다리가 생략되어 해부학적으로 불가능한 신체 구조를 보임",
     "[gpt-high] 이전 사진의 훈장 단 장교와 그 의상을 중앙 인물로 재사용하여, 이전 사진 인물의 얼굴·몸·의상을 이번 장면에 옮기지 말라는 명시적 제외 조건을 위반한다.",
     "[gpt-high] 오른쪽 병력을 아래로 내려가는 숲길이 아니라 능선 위쪽으로 진행하도록 배치하여, 명시된 오른쪽 하강 동선을 반대로 구현한다."
    ],
    "B": [
     "[gpt-high] 왼쪽 배경의 숲길이 화면 왼쪽으로 멀어질수록 올라가며, 명시된 왼쪽 내리막 경로의 수직 방향을 뒤집는다."
    ]
   },
   "per_model_winner": {
    "gemini-pro": "B",
    "gpt-high": "B"
   },
   "agreed": true
  },
  "totals": {
   "B": 1750,
   "A": 306
  },
  "selected": "B",
  "ranking": [
   "B",
   "A"
  ],
  "verdicts": [
   {
    "label": "B",
    "score": 1750,
    "verdict_ko": "레퍼런스 이미지의 인물을 배제하라는 부정 지침을 완벽히 준수하였으며, 지시된 프레이밍 내에서 양쪽 숲길로 흩어지는 군인들의 역동적인 뒷모습을 물리적 오류 없이 훌륭하게 구현했습니다.  ★위반: [gpt-high] 왼쪽 배경의 숲길이 화면 왼쪽으로 멀어질수록 올라가며, 명시된 왼쪽 내리막 경로의 수직 방향을 뒤집는다."
   },
   {
    "label": "A",
    "score": 306,
    "verdict_ko": "레퍼런스 이미지에 등장한 인물(장교)을 절대 포함하지 말라는 명시적인 지침을 위반하였고, 일부 군인의 신체 구조에 심각한 오류가 발생했습니다.  ★위반: [gemini-pro] 레퍼런스 이미지의 인물을 이 샷에 등장시키지 말라는 지침을 어기고 장교를 포함함 / [gemini-pro] 좌측 하단 군인의 오른쪽 다리가 생략되어 해부학적으로 불가능한 신체 구조를 보임 / [gpt-high] 이전 사진의 훈장 단 장교와 그 의상을 중앙 인물로 재사용하여, 이전 사진 인물의 얼굴·몸·의상을 이번 장면에 옮기지 말라는 명시적 제외 조건을 위반한다. / [gpt-high] 오른쪽 병력을 아래로 내려가는 숲길이 아니라 능선 위쪽으로 진행하도록 배치하여, 명시된 오른쪽 하강 동선을 반대로 구현한다."
   }
  ],
  "refs": [
   {
    "label": "PREVIOUS SHOT STILL — a visually related earlier shot of this same place: the location's look, materials, fixed features and lighting mood are LOCKED to this photo; never copy its camera framing. The people visible in that photo are NOT in this shot: never carry their faces, bodies or clothing onto anyone here. Each person in THIS shot is defined solely by the PEOPLE list, their own CHARACTER REFERENCE images and their wardrobe notes. If this photo shows a character who cannot move (dead or unconscious), that character's exact pose, position and orientation are ALSO LOCKED — treat their body as a fixed prop of the set that only the camera moves around.",
    "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/db00517b-165b-4005-a287-f28936b59b95/images/272bd7af-9294-4490-98c6-88a40e9deacc/scene/recipe/S52sh1_sel.png",
    "asset_id": "959a963f-9d07-435d-9add-87cd97f56e93",
    "role": "prev_still"
   }
  ],
  "critique_skipped": true,
  "shot_run_produced": false,
  "shot_run_uid": "06aa1436-5cc7-7086-9f09-6cd139d02707",
  "ref_mode": "prev+엔티티",
  "share_plan": {
   "ref_plan": "prev",
   "prev_anchor_tag": "S52sh1"
  }
 },
 "S52sh6::cine": {
  "applied": true,
  "attempted_at": "2026-09-09T07:03:25.667007+00:00",
  "fingerprint": "7f0a2a9968138c76ed66b7da520c1f63c2049c7e252247e6b353d8b112d39797",
  "fingerprint_version": 2,
  "provider": "mai",
  "endpoint": "openrouter/chat-completions",
  "model": "microsoft/mai-image-2.6",
  "pack": "24.202608252115",
  "source_file": "S52sh6_sel.png",
  "source_sha256": "0f46f690ef17463edabc2cb187056069479e6e8e6a63c4eec599ddfc962d2676",
  "file": "S52sh6_cine.png",
  "staged_sha256": "4e068a7f827bed8d815f88bcc8a208ebfc5e0f41206e194fcc9525e1b1bf824b",
  "latency_ms": 27028
 },
 "S53sh2::signage": {
  "fp": "5376c63b723aa87d",
  "inscriptions": [],
  "cues": [],
  "dropped": []
 },
 "S53sh2": {
  "input_fingerprint": "4912dc150a65619b",
  "prompt": "Create ONE FINAL photorealistic live-action film still of the moment below — early 20th-century Korea, Jeolla region; all people are Korean unless explicitly stated otherwise. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 피할 틈 없이 날아든 칼날이 이관회의 옆구리에 깊게 박힌 순간.\n\nLOCATION (lock): The outdoor market fighting lane beside the straw-sandal stall, amid the continuing encirclement. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: close-up\n- KEY BACKGROUND ELEMENTS: Attacking blade (Cuts across 이관회's side at the moment of contact) — The edge travels obliquely across the visible flank rather than pointing inward as an embedded weapon; used as Provides a small, precise action cue at the lower frame boundary while the body remains the primary subject.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Use restrained daylight and naturalistic contrast, keeping the cutting action readable without heightened color or stylized impact lighting.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The clothing, hair and overall look of 이관회 — who appear both in that photo and in this shot — are LOCKED to that photo. Anyone else visible in that photo is NOT in this shot: never carry their face, body or clothing onto anyone here. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The damaged stall and dropped sword remain in the blocked market lane. 이관회: He receives a fresh slash across his flank.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 이관회 (한국인 남성, 청년 나이대, 자연스러운 검은 머리) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.",
  "roll_prompts": {
   "A": "Create ONE FINAL photorealistic live-action film still of the moment below — early 20th-century Korea, Jeolla region; all people are Korean unless explicitly stated otherwise. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 피할 틈 없이 날아든 칼날이 이관회의 옆구리에 깊게 박힌 순간.\n\nLOCATION (lock): The outdoor market fighting lane beside the straw-sandal stall, amid the continuing encirclement. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: close-up\n- KEY BACKGROUND ELEMENTS: Attacking blade (Cuts across 이관회's side at the moment of contact) — The edge travels obliquely across the visible flank rather than pointing inward as an embedded weapon; used as Provides a small, precise action cue at the lower frame boundary while the body remains the primary subject.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Use restrained daylight and naturalistic contrast, keeping the cutting action readable without heightened color or stylized impact lighting.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The clothing, hair and overall look of 이관회 — who appear both in that photo and in this shot — are LOCKED to that photo. Anyone else visible in that photo is NOT in this shot: never carry their face, body or clothing onto anyone here. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The damaged stall and dropped sword remain in the blocked market lane. 이관회: He receives a fresh slash across his flank.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 이관회 (한국인 남성, 청년 나이대, 자연스러운 검은 머리) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.",
   "B": "Create ONE FINAL photorealistic live-action film still of the moment below — early 20th-century Korea, Jeolla region; all people are Korean unless explicitly stated otherwise. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 피할 틈 없이 날아든 칼날이 이관회의 옆구리에 깊게 박힌 순간.\n\nLOCATION (lock): The outdoor market fighting lane beside the straw-sandal stall, amid the continuing encirclement. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: close-up\n- KEY BACKGROUND ELEMENTS: Attacking blade (Cuts across 이관회's side at the moment of contact) — The edge travels obliquely across the visible flank rather than pointing inward as an embedded weapon; used as Provides a small, precise action cue at the lower frame boundary while the body remains the primary subject.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Use restrained daylight and naturalistic contrast, keeping the cutting action readable without heightened color or stylized impact lighting.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The clothing, hair and overall look of 이관회 — who appear both in that photo and in this shot — are LOCKED to that photo. Anyone else visible in that photo is NOT in this shot: never carry their face, body or clothing onto anyone here. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The damaged stall and dropped sword remain in the blocked market lane. 이관회: He receives a fresh slash across his flank.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 이관회 (한국인 남성, 청년 나이대, 자연스러운 검은 머리) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.\n\nCOMPOSITION VARIATION (this candidate only): keep the place, the people and the moment exactly as specified, and keep from the CAMERA & FRAME contract its subject scale, screen placement and key background elements exactly. For this candidate only, its camera angle and height are deliberately left open: choose a camera position distinctly different from the obvious one, as a film director picking a second setup on the same blocking. Stay true to any explicit framing the SHOT TEXT itself dictates."
  },
  "shot_run_spend_attempt_count": 1,
  "readings": [
   {
    "label": "A",
    "direction": "이관회는 왼쪽을 바라보고 있으며, 그의 몸을 향해 겨냥된 칼날이나 공격 액션은 전혀 보이지 않음.",
    "built_space": "짚신과 바구니가 있는 시장 가판대 공간이나, 이전 샷에서 명시된 부서진 나무 조각이나 바닥에 떨어진 칼 없이 멀쩡한 상태로 묘사됨.",
    "entities": "이관회의 외모와 복장은 기준 이미지와 일치하나, 프롬프트가 요구한 공격하는 칼날은 존재하지 않음.",
    "hard_violations": [],
    "physics": "쪼그려 앉은 자세로 두 다리로 체중을 정상적으로 지탱하고 있으며, 떠 있거나 물리적으로 불가능한 요소는 없음."
   },
   {
    "label": "B",
    "direction": "이관회는 오른쪽 어깨 너머로 뒤를 돌아보고 있으며, 화면 하단 가장자리에 그의 옆구리를 향해 들어오는 칼날이 보임.",
    "built_space": "이전 샷과 일치하는 시장 골목 배경. 화면 왼쪽에 부서진 가판대 나무 조각들이 있고, 오른쪽 바닥에 떨어진 칼이 위치해 연속성이 맞음.",
    "entities": "이관회의 인물 기준이 정확히 반영되었으며, 화면 하단에 프롬프트가 요구한 칼날의 가장자리가 나타남.",
    "hard_violations": [
     "[gpt-high] 기존의 떨어진 칼과 별개로 오른쪽 먼 흙길에 두 번째 칼 또는 칼집 모양 소품을 추가했다."
    ],
    "physics": "서서 몸을 돌리는 자연스러운 자세를 취하고 있으며, 하단의 칼날은 화면 밖에서 들어와 찌르는 형태로 물리적 오류가 없음."
   }
  ],
  "cross_model_order": {
   "policy": "cross_model_order_v2_slot0_fwd_slot1_rev_parallel2",
   "slots": [
    {
     "model": "gemini-pro",
     "order": "forward",
     "display_to_canonical": {
      "A": "A",
      "B": "B"
     },
     "raw": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 3,
        "verdict_ko": "프레이밍은 클로즈업에 가깝지만, 핵심 액션인 옆구리를 찌르는 칼날이 없고 이전 샷의 연속성(부서진 가판대, 떨어진 칼)을 완전히 누락했습니다."
       },
       {
        "label": "B",
        "score": 8,
        "verdict_ko": "화면 하단에 명시된 칼날 액션을 정확히 배치했으며, 부서진 가판대와 바닥에 떨어진 칼 등 이전 샷의 환경 연속성을 훌륭하게 유지했습니다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "이관회는 왼쪽을 바라보고 있으며, 그의 몸을 향해 겨냥된 칼날이나 공격 액션은 전혀 보이지 않음.",
        "built_space": "짚신과 바구니가 있는 시장 가판대 공간이나, 이전 샷에서 명시된 부서진 나무 조각이나 바닥에 떨어진 칼 없이 멀쩡한 상태로 묘사됨.",
        "entities": "이관회의 외모와 복장은 기준 이미지와 일치하나, 프롬프트가 요구한 공격하는 칼날은 존재하지 않음.",
        "hard_violations": [],
        "physics": "쪼그려 앉은 자세로 두 다리로 체중을 정상적으로 지탱하고 있으며, 떠 있거나 물리적으로 불가능한 요소는 없음."
       },
       {
        "label": "B",
        "direction": "이관회는 오른쪽 어깨 너머로 뒤를 돌아보고 있으며, 화면 하단 가장자리에 그의 옆구리를 향해 들어오는 칼날이 보임.",
        "built_space": "이전 샷과 일치하는 시장 골목 배경. 화면 왼쪽에 부서진 가판대 나무 조각들이 있고, 오른쪽 바닥에 떨어진 칼이 위치해 연속성이 맞음.",
        "entities": "이관회의 인물 기준이 정확히 반영되었으며, 화면 하단에 프롬프트가 요구한 칼날의 가장자리가 나타남.",
        "hard_violations": [],
        "physics": "서서 몸을 돌리는 자연스러운 자세를 취하고 있으며, 하단의 칼날은 화면 밖에서 들어와 찌르는 형태로 물리적 오류가 없음."
       }
      ],
      "all_candidates_fail": false
     },
     "normalized": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 3,
        "verdict_ko": "프레이밍은 클로즈업에 가깝지만, 핵심 액션인 옆구리를 찌르는 칼날이 없고 이전 샷의 연속성(부서진 가판대, 떨어진 칼)을 완전히 누락했습니다."
       },
       {
        "label": "B",
        "score": 8,
        "verdict_ko": "화면 하단에 명시된 칼날 액션을 정확히 배치했으며, 부서진 가판대와 바닥에 떨어진 칼 등 이전 샷의 환경 연속성을 훌륭하게 유지했습니다."
       }
      ],
      "all_candidates_fail": false,
      "readings": [
       {
        "label": "A",
        "direction": "이관회는 왼쪽을 바라보고 있으며, 그의 몸을 향해 겨냥된 칼날이나 공격 액션은 전혀 보이지 않음.",
        "built_space": "짚신과 바구니가 있는 시장 가판대 공간이나, 이전 샷에서 명시된 부서진 나무 조각이나 바닥에 떨어진 칼 없이 멀쩡한 상태로 묘사됨.",
        "entities": "이관회의 외모와 복장은 기준 이미지와 일치하나, 프롬프트가 요구한 공격하는 칼날은 존재하지 않음.",
        "hard_violations": [],
        "physics": "쪼그려 앉은 자세로 두 다리로 체중을 정상적으로 지탱하고 있으며, 떠 있거나 물리적으로 불가능한 요소는 없음."
       },
       {
        "label": "B",
        "direction": "이관회는 오른쪽 어깨 너머로 뒤를 돌아보고 있으며, 화면 하단 가장자리에 그의 옆구리를 향해 들어오는 칼날이 보임.",
        "built_space": "이전 샷과 일치하는 시장 골목 배경. 화면 왼쪽에 부서진 가판대 나무 조각들이 있고, 오른쪽 바닥에 떨어진 칼이 위치해 연속성이 맞음.",
        "entities": "이관회의 인물 기준이 정확히 반영되었으며, 화면 하단에 프롬프트가 요구한 칼날의 가장자리가 나타남.",
        "hard_violations": [],
        "physics": "서서 몸을 돌리는 자연스러운 자세를 취하고 있으며, 하단의 칼날은 화면 밖에서 들어와 찌르는 형태로 물리적 오류가 없음."
       }
      ]
     },
     "ok": true
    },
    {
     "model": "gpt-high",
     "order": "reverse",
     "display_to_canonical": {
      "A": "B",
      "B": "A"
     },
     "raw": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 2,
        "verdict_ko": "외모와 시장 재질은 이어지지만 클로즈업보다 넓고, 옆구리 베임 없이 바닥의 칼만 보이며 불필요한 두 번째 칼 모양 소품까지 추가됐다."
       },
       {
        "label": "B",
        "score": 4,
        "verdict_ko": "인물 중심 클로즈업과 외모·의상은 더 충실하지만, 하단을 비스듬히 가르는 칼날과 옆구리의 새 상처가 없어 핵심 피격 순간에는 실패했다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "이관회는 고개와 눈을 화면 왼쪽으로 돌렸으며 시선 대상은 화면 밖이다. 오른쪽 아래 칼은 손잡이가 오른쪽, 날이 왼쪽 아래를 향한 채 바닥에 놓여 있다. 옆구리를 향하거나 접촉하는 공격 칼날은 보이지 않는다.",
        "built_space": "왼쪽에 낮은 목제 진열대 하나와 어두운 선반 공간, 중앙 뒤에 기대어진 파손 목판 하나가 보인다. 왼쪽 아래에는 큰 손잡이 바구니 하나와 겹친 얕은 바구니 두 개가 있고, 뒤에는 목재 점포와 천 차양이 이어진다. 인물은 진열대 옆 흙길에 서 있다. 재료와 파손 목판은 이전 장소와 잘 이어지지만 허리까지 포함한 넓은 구도로 배경 비중이 크다. 반사는 없다.",
        "entities": "젊은 한국인 남성으로 설정된 이관회 한 명만 보이며, 검은 가르마 머리와 얼굴, 해진 갈색 저고리 및 보이는 흰 바지는 참조와 대체로 맞는다. 짚신 좌판과 파손 목판, 떨어진 칼이 보인다. 칼의 세부 장식과 마모는 작아서 확인하기 어렵다. 오른쪽 먼 흙길에는 별도의 검은 칼 또는 칼집 모양 물체가 추가되어 있다. 옆구리에 새 상처나 피, 접촉 중인 칼날은 없다. 읽을 수 있는 글자는 없다.",
        "hard_violations": [
         "기존의 떨어진 칼과 별개로 오른쪽 먼 흙길에 두 번째 칼 또는 칼집 모양 소품을 추가했다."
        ],
        "physics": "몸통은 선 자세로 자연스럽게 연결되며 발은 화면 밖이어서 지면 접촉 자체는 확인할 수 없다. 공중에 떠 있는 자세는 아니다. 두 칼 모양 물체는 흙바닥에 놓여 있고, 파손 목판은 진열대와 지면에 기대어 있다. 공격 동작이나 칼날 충격에 따른 신체 반응은 보이지 않는다."
       },
       {
        "label": "B",
        "direction": "이관회의 얼굴과 시선은 화면 왼쪽의 보이지 않는 대상을 향한다. 칼날이나 다른 방향성 소품이 전혀 보이지 않아, 옆구리를 비스듬히 가르는 공격 방향과 접촉 지점을 확인할 수 없다.",
        "built_space": "왼쪽 뒤에 낮은 목제 진열대 하나와 어두운 선반 공간 하나가 있고, 진열대 위에는 짚신과 천이 놓여 있다. 왼쪽 위에는 매달린 짚신들이, 오른쪽 위에는 흰 천 차양이 보인다. 인물은 좌판 앞에서 한쪽 무릎을 높게 굽힌 모습이다. 얼굴과 상체가 화면을 크게 차지해 클로즈업에 부합한다. 파손 목판과 떨어진 칼은 이 구도에서 확인되지 않으며, 이를 보이기 위해 구도를 넓힐 필요는 없다. 반사는 없다.",
        "entities": "이관회 한 명만 보인다. 젊은 남성의 얼굴, 자연스러운 검은 머리, 해진 갈색 저고리와 무릎 부분의 흰 바지는 참조와 대체로 일치한다. 짚신 좌판의 목재와 직물도 장소의 시대감과 맞는다. 공격 칼날, 옆구리의 새 베임과 출혈은 보이지 않는다. 읽을 수 있는 글자는 없다.",
        "hard_violations": [],
        "physics": "흰 바지를 입은 한쪽 다리가 무릎에서 굽혀져 상체 가까이 올라와 있다. 발과 골반의 지지점은 하단 밖이라 앉았는지 쪼그렸는지는 확정할 수 없지만, 보이는 관절 연결에 명백한 불가능성이나 공중 부유는 없다. 진열 물건은 선반에 놓이고 짚신은 끈에 매달려 있다. 칼날 접촉과 그 충격은 묘사되지 않았다."
       }
      ],
      "all_candidates_fail": true
     },
     "normalized": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 2,
        "verdict_ko": "외모와 시장 재질은 이어지지만 클로즈업보다 넓고, 옆구리 베임 없이 바닥의 칼만 보이며 불필요한 두 번째 칼 모양 소품까지 추가됐다."
       },
       {
        "label": "A",
        "score": 4,
        "verdict_ko": "인물 중심 클로즈업과 외모·의상은 더 충실하지만, 하단을 비스듬히 가르는 칼날과 옆구리의 새 상처가 없어 핵심 피격 순간에는 실패했다."
       }
      ],
      "all_candidates_fail": true,
      "readings": [
       {
        "label": "B",
        "direction": "이관회는 고개와 눈을 화면 왼쪽으로 돌렸으며 시선 대상은 화면 밖이다. 오른쪽 아래 칼은 손잡이가 오른쪽, 날이 왼쪽 아래를 향한 채 바닥에 놓여 있다. 옆구리를 향하거나 접촉하는 공격 칼날은 보이지 않는다.",
        "built_space": "왼쪽에 낮은 목제 진열대 하나와 어두운 선반 공간, 중앙 뒤에 기대어진 파손 목판 하나가 보인다. 왼쪽 아래에는 큰 손잡이 바구니 하나와 겹친 얕은 바구니 두 개가 있고, 뒤에는 목재 점포와 천 차양이 이어진다. 인물은 진열대 옆 흙길에 서 있다. 재료와 파손 목판은 이전 장소와 잘 이어지지만 허리까지 포함한 넓은 구도로 배경 비중이 크다. 반사는 없다.",
        "entities": "젊은 한국인 남성으로 설정된 이관회 한 명만 보이며, 검은 가르마 머리와 얼굴, 해진 갈색 저고리 및 보이는 흰 바지는 참조와 대체로 맞는다. 짚신 좌판과 파손 목판, 떨어진 칼이 보인다. 칼의 세부 장식과 마모는 작아서 확인하기 어렵다. 오른쪽 먼 흙길에는 별도의 검은 칼 또는 칼집 모양 물체가 추가되어 있다. 옆구리에 새 상처나 피, 접촉 중인 칼날은 없다. 읽을 수 있는 글자는 없다.",
        "hard_violations": [
         "기존의 떨어진 칼과 별개로 오른쪽 먼 흙길에 두 번째 칼 또는 칼집 모양 소품을 추가했다."
        ],
        "physics": "몸통은 선 자세로 자연스럽게 연결되며 발은 화면 밖이어서 지면 접촉 자체는 확인할 수 없다. 공중에 떠 있는 자세는 아니다. 두 칼 모양 물체는 흙바닥에 놓여 있고, 파손 목판은 진열대와 지면에 기대어 있다. 공격 동작이나 칼날 충격에 따른 신체 반응은 보이지 않는다."
       },
       {
        "label": "A",
        "direction": "이관회의 얼굴과 시선은 화면 왼쪽의 보이지 않는 대상을 향한다. 칼날이나 다른 방향성 소품이 전혀 보이지 않아, 옆구리를 비스듬히 가르는 공격 방향과 접촉 지점을 확인할 수 없다.",
        "built_space": "왼쪽 뒤에 낮은 목제 진열대 하나와 어두운 선반 공간 하나가 있고, 진열대 위에는 짚신과 천이 놓여 있다. 왼쪽 위에는 매달린 짚신들이, 오른쪽 위에는 흰 천 차양이 보인다. 인물은 좌판 앞에서 한쪽 무릎을 높게 굽힌 모습이다. 얼굴과 상체가 화면을 크게 차지해 클로즈업에 부합한다. 파손 목판과 떨어진 칼은 이 구도에서 확인되지 않으며, 이를 보이기 위해 구도를 넓힐 필요는 없다. 반사는 없다.",
        "entities": "이관회 한 명만 보인다. 젊은 남성의 얼굴, 자연스러운 검은 머리, 해진 갈색 저고리와 무릎 부분의 흰 바지는 참조와 대체로 일치한다. 짚신 좌판의 목재와 직물도 장소의 시대감과 맞는다. 공격 칼날, 옆구리의 새 베임과 출혈은 보이지 않는다. 읽을 수 있는 글자는 없다.",
        "hard_violations": [],
        "physics": "흰 바지를 입은 한쪽 다리가 무릎에서 굽혀져 상체 가까이 올라와 있다. 발과 골반의 지지점은 하단 밖이라 앉았는지 쪼그렸는지는 확정할 수 없지만, 보이는 관절 연결에 명백한 불가능성이나 공중 부유는 없다. 진열 물건은 선반에 놓이고 짚신은 끈에 매달려 있다. 칼날 접촉과 그 충격은 묘사되지 않았다."
       }
      ]
     },
     "ok": true
    }
   ],
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "slot_winner_match": false,
   "slot_winner": {
    "gemini-pro": "B",
    "gpt-high": "A"
   },
   "route": "cross_slot_combined"
  },
  "dual": {
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "normalized": {
    "A": 1.375,
    "B": 1.5
   },
   "adjusted": {
    "A": 1.375,
    "B": 1.25
   },
   "violations": {
    "B": [
     "[gpt-high] 기존의 떨어진 칼과 별개로 오른쪽 먼 흙길에 두 번째 칼 또는 칼집 모양 소품을 추가했다."
    ]
   },
   "per_model_winner": {
    "gemini-pro": "B",
    "gpt-high": "A"
   },
   "agreed": false
  },
  "totals": {
   "A": 1375,
   "B": 1250
  },
  "selected": "A",
  "ranking": [
   "A",
   "B"
  ],
  "verdicts": [
   {
    "label": "A",
    "score": 1375,
    "verdict_ko": "프레이밍은 클로즈업에 가깝지만, 핵심 액션인 옆구리를 찌르는 칼날이 없고 이전 샷의 연속성(부서진 가판대, 떨어진 칼)을 완전히 누락했습니다."
   },
   {
    "label": "B",
    "score": 1250,
    "verdict_ko": "화면 하단에 명시된 칼날 액션을 정확히 배치했으며, 부서진 가판대와 바닥에 떨어진 칼 등 이전 샷의 환경 연속성을 훌륭하게 유지했습니다.  ★위반: [gpt-high] 기존의 떨어진 칼과 별개로 오른쪽 먼 흙길에 두 번째 칼 또는 칼집 모양 소품을 추가했다."
   }
  ],
  "refs": [
   {
    "label": "PREVIOUS SHOT STILL — a visually related earlier shot of this same place: the location's look, materials, fixed features and lighting mood are LOCKED to this photo; never copy its camera framing. The clothing, hair and overall look of 이관회 — who appear both in that photo and in this shot — are LOCKED to that photo. Anyone else visible in that photo is NOT in this shot: never carry their face, body or clothing onto anyone here. If this photo shows a character who cannot move (dead or unconscious), that character's exact pose, position and orientation are ALSO LOCKED — treat their body as a fixed prop of the set that only the camera moves around.",
    "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/db00517b-165b-4005-a287-f28936b59b95/images/272bd7af-9294-4490-98c6-88a40e9deacc/scene/recipe/S51sh15_sel.png",
    "asset_id": "53e5828b-837a-47f2-ad68-e0b0258adca3",
    "role": "prev_still"
   },
   {
    "label": "CHARACTER REFERENCE — 이관회: the exact person appearing in this shot; match face, hair and build exactly.",
    "path": "<bytes:933540>",
    "asset_id": "851b450f-ac84-47fe-ac5a-03a56f63fcae",
    "role": "character_ref"
   },
   {
    "label": "PROP REFERENCE — 칼: the exact object appearing in this shot; match its look, material and wear exactly.",
    "path": "<bytes:780122>",
    "asset_id": "82614014-6091-423c-89ca-1e15d209de72",
    "role": "prop_ref"
   }
  ],
  "critique_skipped": true,
  "shot_run_produced": false,
  "shot_run_uid": "06aa1436-5e72-79b2-9ee0-da78cf679be3",
  "ref_mode": "prev+엔티티",
  "share_plan": {
   "ref_plan": "prev",
   "prev_anchor_tag": "S51sh15"
  }
 },
 "S53sh2::cine": {
  "applied": false,
  "attempted_at": "2026-09-09T09:32:08.028594+00:00",
  "fingerprint": "202082dd996d3c434b5fea91368db2e738adc823fced7600117b8e3acb96b72e",
  "fingerprint_version": 2,
  "provider": "mai",
  "endpoint": "openrouter/chat-completions",
  "model": "microsoft/mai-image-2.6",
  "multiroll_tag": "still_S53sh2_cine",
  "slot": "",
  "pack": "24.202608252115",
  "source_file": "S53sh2_sel.png",
  "source_sha256": "5a7b02e31818e0acda3c9f98561c1eb6b8706540b45b3804d7ba450ac9dab374",
  "error": "RuntimeError: moderation blocked: {\"message\": \"Response content blocked by label 'MultiSeverity_ViolenceScore'.\", \"code\": 400}",
  "moderation_refusals": 2,
  "declined": true,
  "declined_reason": "moderation"
 },
 "S53sh6::signage": {
  "fp": "f5e1dcad827b2f76",
  "inscriptions": [],
  "cues": [],
  "dropped": []
 },
 "S53sh6": {
  "input_fingerprint": "22ea1023a8459ebe",
  "prompt": "Create ONE FINAL photorealistic live-action film still of the moment below — early 20th-century Korea, Jeolla region; all people are Korean unless explicitly stated otherwise. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 안규홍의 앞을 가로막은 채 서늘한 칼날을 들이민 일진회 대장의 굳은 전신.\n\nLOCATION (lock): The strip of dirt market lane between the two separated defenders, blocked by the attackers' leader. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: wide shot\n- KEY BACKGROUND ELEMENTS: Leader's blade (Extended into 안규홍's approach) — Seen obliquely across the gap between the men, not directed toward the lens; used as Draws a restrained diagonal through the route that has just been blocked.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Preserve the market's daylight and subdued saturation, maintaining consistent contrast across both men and the obstructed space.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The people visible in that photo are NOT in this shot: never carry their faces, bodies or clothing onto anyone here. Each person in THIS shot is defined solely by the PEOPLE list, their own CHARACTER REFERENCE images and their wardrobe notes. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The damaged stall and fallen sword remain where the fight has displaced them. 일진회 대장: He blocks the lane holding a drawn sword.\n\nTHE HAND THAT IS DOING THIS: the object at the centre of this shot is being held, operated, read, aimed or handed over by 일진회 대장 right now, so 일진회 대장's hand — and as much of the wrist and forearm as the framing reaches — is in the frame, gripping or resting on that object exactly the way its real use requires. Match that hand to 일진회 대장: its size, build, skin, age, grooming, sleeve and anything worn on it belong to that person and to no one else.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 일진회 대장 (한국인 남성, 성인 나이대, 자연스러운 검은 머리) — wearing: 일본식 하오리나 서양식 조끼를 조선식 바지저고리 위에 덧입은 기형적인 개화기 불량배 복장. 허리에 칼을 찼다. — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.",
  "roll_prompts": {
   "A": "Create ONE FINAL photorealistic live-action film still of the moment below — early 20th-century Korea, Jeolla region; all people are Korean unless explicitly stated otherwise. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 안규홍의 앞을 가로막은 채 서늘한 칼날을 들이민 일진회 대장의 굳은 전신.\n\nLOCATION (lock): The strip of dirt market lane between the two separated defenders, blocked by the attackers' leader. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: wide shot\n- KEY BACKGROUND ELEMENTS: Leader's blade (Extended into 안규홍's approach) — Seen obliquely across the gap between the men, not directed toward the lens; used as Draws a restrained diagonal through the route that has just been blocked.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Preserve the market's daylight and subdued saturation, maintaining consistent contrast across both men and the obstructed space.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The people visible in that photo are NOT in this shot: never carry their faces, bodies or clothing onto anyone here. Each person in THIS shot is defined solely by the PEOPLE list, their own CHARACTER REFERENCE images and their wardrobe notes. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The damaged stall and fallen sword remain where the fight has displaced them. 일진회 대장: He blocks the lane holding a drawn sword.\n\nTHE HAND THAT IS DOING THIS: the object at the centre of this shot is being held, operated, read, aimed or handed over by 일진회 대장 right now, so 일진회 대장's hand — and as much of the wrist and forearm as the framing reaches — is in the frame, gripping or resting on that object exactly the way its real use requires. Match that hand to 일진회 대장: its size, build, skin, age, grooming, sleeve and anything worn on it belong to that person and to no one else.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 일진회 대장 (한국인 남성, 성인 나이대, 자연스러운 검은 머리) — wearing: 일본식 하오리나 서양식 조끼를 조선식 바지저고리 위에 덧입은 기형적인 개화기 불량배 복장. 허리에 칼을 찼다. — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.",
   "B": "Create ONE FINAL photorealistic live-action film still of the moment below — early 20th-century Korea, Jeolla region; all people are Korean unless explicitly stated otherwise. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 안규홍의 앞을 가로막은 채 서늘한 칼날을 들이민 일진회 대장의 굳은 전신.\n\nLOCATION (lock): The strip of dirt market lane between the two separated defenders, blocked by the attackers' leader. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: wide shot\n- KEY BACKGROUND ELEMENTS: Leader's blade (Extended into 안규홍's approach) — Seen obliquely across the gap between the men, not directed toward the lens; used as Draws a restrained diagonal through the route that has just been blocked.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Preserve the market's daylight and subdued saturation, maintaining consistent contrast across both men and the obstructed space.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The people visible in that photo are NOT in this shot: never carry their faces, bodies or clothing onto anyone here. Each person in THIS shot is defined solely by the PEOPLE list, their own CHARACTER REFERENCE images and their wardrobe notes. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The damaged stall and fallen sword remain where the fight has displaced them. 일진회 대장: He blocks the lane holding a drawn sword.\n\nTHE HAND THAT IS DOING THIS: the object at the centre of this shot is being held, operated, read, aimed or handed over by 일진회 대장 right now, so 일진회 대장's hand — and as much of the wrist and forearm as the framing reaches — is in the frame, gripping or resting on that object exactly the way its real use requires. Match that hand to 일진회 대장: its size, build, skin, age, grooming, sleeve and anything worn on it belong to that person and to no one else.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 일진회 대장 (한국인 남성, 성인 나이대, 자연스러운 검은 머리) — wearing: 일본식 하오리나 서양식 조끼를 조선식 바지저고리 위에 덧입은 기형적인 개화기 불량배 복장. 허리에 칼을 찼다. — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.\n\nCOMPOSITION VARIATION (this candidate only): keep the place, the people and the moment exactly as specified, and keep from the CAMERA & FRAME contract its subject scale, screen placement and key background elements exactly. For this candidate only, its camera angle and height are deliberately left open: choose a camera position distinctly different from the obvious one, as a film director picking a second setup on the same blocking. Stay true to any explicit framing the SHOT TEXT itself dictates."
  },
  "shot_run_spend_attempt_count": 1,
  "initial_roll_all_fail": true,
  "readings": [
   {
    "label": "A",
    "direction": "남자의 시선은 화면 밖 왼쪽을 향하고 있으며, 오른손에 쥔 칼은 대각선으로 길을 가로지르며 앞으로 뻗어 있어 차단된 경로를 묘사함.",
    "built_space": "흙바닥으로 된 시장 골목과 양옆의 목조 난전들이 레퍼런스 이미지의 배경과 유사하게 배치되어 있음.",
    "entities": "일진회 대장이 검은색 하오리와 서양식 조끼, 조선식 바지를 입고 있음. 하지만 프롬프트에서 절대 사용하지 말라고 지시한 레퍼런스 이미지(이전 샷) 속 인물의 얼굴을 그대로 복제하여 그려냄.",
    "hard_violations": [
     "[gemini-pro] 레퍼런스 이미지의 인물 얼굴을 그대로 복사하여 사용함 (금지된 인물 등장)",
     "[gemini-pro] 오른손이 칼자루가 아닌 칼날 부분을 쥐고 있으며 코등이(츠바)가 손목 뒤쪽에 위치함 (물리적으로 불가능한 해부학 및 무기 형태)",
     "[gpt-high] 오른손에 이미 칼을 뽑아 들었는데도 허리에 별도의 칼자루와 칼막이가 남아 있어, 지정되지 않은 두 번째 착용 칼이 추가되었다."
    ],
    "physics": "두 발로 땅을 안정적으로 딛고 서 있음. 바닥에 떨어진 칼은 흙 위에 놓여 있으나, 오른손은 손잡이가 아닌 날을 맨손으로 쥐고 있는 기형적인 구조임."
   },
   {
    "label": "B",
    "direction": "남자가 정면(카메라 렌즈)을 응시하고 있으며, 오른손의 칼은 뻗지 않고 바닥을 향해 비스듬히 내려가 있음.",
    "built_space": "시장 골목의 양옆으로 난전들이 늘어서 있으나, 화면 왼쪽 상단 가판대 구조물에 명확히 읽을 수 있는 한글 텍스트(간판)가 그려져 있음.",
    "entities": "일진회 대장이 검은색 하오리와 조선식 저고리 및 바지를 입고 등장함. 이 후보 역시 금지된 레퍼런스 이미지 속 인물의 얼굴을 동일하게 복사하여 적용함.",
    "hard_violations": [
     "[gemini-pro] 레퍼런스 이미지의 인물 얼굴을 그대로 복사하여 사용함 (금지된 인물 등장)",
     "[gemini-pro] 왼쪽 상단 가판대에 읽을 수 있는 한글 텍스트가 뚜렷하게 노출됨 (텍스트 유출)",
     "[gemini-pro] 주먹 쥔 왼손 뒤쪽에 검은색 칼집이 손에 쥐어지거나 고정되지 않은 채 허공에 떠 있음 (물리적으로 지지되지 않은 사물)",
     "[gpt-high] 손에 든 칼과 별개로 허리에 다른 칼자루가 있으며, 지정된 떨어진 칼 이외에도 왼쪽 바닥과 오른쪽 진열대 옆에 추가 무기가 배치되어 있다.",
     "[gpt-high] 왼쪽 상단 간판에 판독 가능한 한글이 노출되어 글자 금지 조건을 위반한다."
    ],
    "physics": "두 발로 바닥을 딛고 서 있고 오른손으로 칼자루를 올바르게 쥐고 있음. 그러나 왼손 주먹 뒤쪽으로 검은색 칼집이 아무런 지지 없이 허공에 떠 있거나 몸통을 통과하는 듯한 불가능한 물리 법칙을 보여줌."
   }
  ],
  "cross_model_order": {
   "policy": "cross_model_order_v2_slot0_fwd_slot1_rev_parallel2",
   "slots": [
    {
     "model": "gemini-pro",
     "order": "forward",
     "display_to_canonical": {
      "A": "A",
      "B": "B"
     },
     "raw": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 2,
        "verdict_ko": "프롬프트가 명시적으로 금지한 이전 장면(레퍼런스) 인물의 얼굴을 그대로 복사해 왔으며, 오른손이 칼자루가 아닌 칼날을 쥐는 물리적으로 불가능한 파지법을 그려내어 탈락 수준의 오류를 보입니다."
       },
       {
        "label": "B",
        "score": 1,
        "verdict_ko": "레퍼런스 인물의 얼굴을 차용한 심각한 위반에 더해, 좌측 상단에 읽을 수 있는 간판 텍스트가 노출되고 왼손 부근에 지지되지 않은 채 떠 있는 칼집이 있어 프롬프트를 전방위로 위반했습니다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "남자의 시선은 화면 밖 왼쪽을 향하고 있으며, 오른손에 쥔 칼은 대각선으로 길을 가로지르며 앞으로 뻗어 있어 차단된 경로를 묘사함.",
        "built_space": "흙바닥으로 된 시장 골목과 양옆의 목조 난전들이 레퍼런스 이미지의 배경과 유사하게 배치되어 있음.",
        "entities": "일진회 대장이 검은색 하오리와 서양식 조끼, 조선식 바지를 입고 있음. 하지만 프롬프트에서 절대 사용하지 말라고 지시한 레퍼런스 이미지(이전 샷) 속 인물의 얼굴을 그대로 복제하여 그려냄.",
        "hard_violations": [
         "레퍼런스 이미지의 인물 얼굴을 그대로 복사하여 사용함 (금지된 인물 등장)",
         "오른손이 칼자루가 아닌 칼날 부분을 쥐고 있으며 코등이(츠바)가 손목 뒤쪽에 위치함 (물리적으로 불가능한 해부학 및 무기 형태)"
        ],
        "physics": "두 발로 땅을 안정적으로 딛고 서 있음. 바닥에 떨어진 칼은 흙 위에 놓여 있으나, 오른손은 손잡이가 아닌 날을 맨손으로 쥐고 있는 기형적인 구조임."
       },
       {
        "label": "B",
        "direction": "남자가 정면(카메라 렌즈)을 응시하고 있으며, 오른손의 칼은 뻗지 않고 바닥을 향해 비스듬히 내려가 있음.",
        "built_space": "시장 골목의 양옆으로 난전들이 늘어서 있으나, 화면 왼쪽 상단 가판대 구조물에 명확히 읽을 수 있는 한글 텍스트(간판)가 그려져 있음.",
        "entities": "일진회 대장이 검은색 하오리와 조선식 저고리 및 바지를 입고 등장함. 이 후보 역시 금지된 레퍼런스 이미지 속 인물의 얼굴을 동일하게 복사하여 적용함.",
        "hard_violations": [
         "레퍼런스 이미지의 인물 얼굴을 그대로 복사하여 사용함 (금지된 인물 등장)",
         "왼쪽 상단 가판대에 읽을 수 있는 한글 텍스트가 뚜렷하게 노출됨 (텍스트 유출)",
         "주먹 쥔 왼손 뒤쪽에 검은색 칼집이 손에 쥐어지거나 고정되지 않은 채 허공에 떠 있음 (물리적으로 지지되지 않은 사물)"
        ],
        "physics": "두 발로 바닥을 딛고 서 있고 오른손으로 칼자루를 올바르게 쥐고 있음. 그러나 왼손 주먹 뒤쪽으로 검은색 칼집이 아무런 지지 없이 허공에 떠 있거나 몸통을 통과하는 듯한 불가능한 물리 법칙을 보여줌."
       }
      ],
      "all_candidates_fail": true
     },
     "normalized": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 2,
        "verdict_ko": "프롬프트가 명시적으로 금지한 이전 장면(레퍼런스) 인물의 얼굴을 그대로 복사해 왔으며, 오른손이 칼자루가 아닌 칼날을 쥐는 물리적으로 불가능한 파지법을 그려내어 탈락 수준의 오류를 보입니다."
       },
       {
        "label": "B",
        "score": 1,
        "verdict_ko": "레퍼런스 인물의 얼굴을 차용한 심각한 위반에 더해, 좌측 상단에 읽을 수 있는 간판 텍스트가 노출되고 왼손 부근에 지지되지 않은 채 떠 있는 칼집이 있어 프롬프트를 전방위로 위반했습니다."
       }
      ],
      "all_candidates_fail": true,
      "readings": [
       {
        "label": "A",
        "direction": "남자의 시선은 화면 밖 왼쪽을 향하고 있으며, 오른손에 쥔 칼은 대각선으로 길을 가로지르며 앞으로 뻗어 있어 차단된 경로를 묘사함.",
        "built_space": "흙바닥으로 된 시장 골목과 양옆의 목조 난전들이 레퍼런스 이미지의 배경과 유사하게 배치되어 있음.",
        "entities": "일진회 대장이 검은색 하오리와 서양식 조끼, 조선식 바지를 입고 있음. 하지만 프롬프트에서 절대 사용하지 말라고 지시한 레퍼런스 이미지(이전 샷) 속 인물의 얼굴을 그대로 복제하여 그려냄.",
        "hard_violations": [
         "레퍼런스 이미지의 인물 얼굴을 그대로 복사하여 사용함 (금지된 인물 등장)",
         "오른손이 칼자루가 아닌 칼날 부분을 쥐고 있으며 코등이(츠바)가 손목 뒤쪽에 위치함 (물리적으로 불가능한 해부학 및 무기 형태)"
        ],
        "physics": "두 발로 땅을 안정적으로 딛고 서 있음. 바닥에 떨어진 칼은 흙 위에 놓여 있으나, 오른손은 손잡이가 아닌 날을 맨손으로 쥐고 있는 기형적인 구조임."
       },
       {
        "label": "B",
        "direction": "남자가 정면(카메라 렌즈)을 응시하고 있으며, 오른손의 칼은 뻗지 않고 바닥을 향해 비스듬히 내려가 있음.",
        "built_space": "시장 골목의 양옆으로 난전들이 늘어서 있으나, 화면 왼쪽 상단 가판대 구조물에 명확히 읽을 수 있는 한글 텍스트(간판)가 그려져 있음.",
        "entities": "일진회 대장이 검은색 하오리와 조선식 저고리 및 바지를 입고 등장함. 이 후보 역시 금지된 레퍼런스 이미지 속 인물의 얼굴을 동일하게 복사하여 적용함.",
        "hard_violations": [
         "레퍼런스 이미지의 인물 얼굴을 그대로 복사하여 사용함 (금지된 인물 등장)",
         "왼쪽 상단 가판대에 읽을 수 있는 한글 텍스트가 뚜렷하게 노출됨 (텍스트 유출)",
         "주먹 쥔 왼손 뒤쪽에 검은색 칼집이 손에 쥐어지거나 고정되지 않은 채 허공에 떠 있음 (물리적으로 지지되지 않은 사물)"
        ],
        "physics": "두 발로 바닥을 딛고 서 있고 오른손으로 칼자루를 올바르게 쥐고 있음. 그러나 왼손 주먹 뒤쪽으로 검은색 칼집이 아무런 지지 없이 허공에 떠 있거나 몸통을 통과하는 듯한 불가능한 물리 법칙을 보여줌."
       }
      ]
     },
     "ok": true
    },
    {
     "model": "gpt-high",
     "order": "reverse",
     "display_to_canonical": {
      "A": "B",
      "B": "A"
     },
     "raw": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 2,
        "verdict_ko": "전신 와이드 구도는 맞지만 칼을 상대의 접근로가 아닌 땅으로 내리고 정면을 응시하며, 추가 무기와 간판 글자도 금지 조건을 어긴다."
       },
       {
        "label": "B",
        "score": 4,
        "verdict_ko": "왼쪽 접근로로 칼을 내밀고 같은 쪽을 경계하는 동작은 더 충실하지만, 발이 잘린 구도와 허리의 별도 칼 때문에 완전한 적합작은 아니다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "남자는 카메라 정면을 바라본다. 오른손의 칼은 화면 왼쪽 아래로 내려가며 칼끝이 가까운 흙바닥을 향한다. 렌즈를 겨누지는 않지만 안규홍의 접근을 향해 칼을 들이미는 동작보다는 칼을 내려 든 자세다. 화면에 상대는 없다.",
        "built_space": "양쪽 목조 점포 사이 흙길 중앙에 한 사람이 서 있다. 왼쪽 가까운 기둥에는 짚신 한 쌍이 걸려 있고 그 아래로 목조 진열대와 바구니들이 있다. 오른쪽에는 천을 덮은 진열대와 기울어 무너진 목조 판넬 하나가 보인다. 양쪽 천막과 낡은 목재는 참고 장소의 재질에 부합하지만, 참고에서 확인되지 않는 긴 시장 골목까지 확장되어 정확한 고정 배치는 확인하기 어렵다.",
        "entities": "검은 머리의 젊은 성인 동아시아계 남성 한 명으로, 한국인 남성 설정과 외형상 충돌하지 않는다. 조선식 저고리와 바지 위에 어두운 하오리를 입었다. 다만 얼굴과 가르마 머리는 이전 장면 인물과 매우 유사하다. 손에 든 곡도는 참고의 금속 칼과 대체로 같은 종류이나 장식과 마모의 정확한 일치는 부족하다. 허리에는 별도 칼자루가 있고 왼손에는 칼집처럼 보이는 물체가 있다. 바닥의 떨어진 칼 외에도 왼쪽 바닥과 오른쪽 진열대 옆에 추가 장검류가 보인다. 상단 왼쪽 간판에는 판독 가능한 한글 글자 일부가 드러난다.",
        "hard_violations": [
         "손에 든 칼과 별개로 허리에 다른 칼자루가 있으며, 지정된 떨어진 칼 이외에도 왼쪽 바닥과 오른쪽 진열대 옆에 추가 무기가 배치되어 있다.",
         "왼쪽 상단 간판에 판독 가능한 한글이 노출되어 글자 금지 조건을 위반한다."
        ],
        "physics": "양발이 흙바닥에 닿아 체중을 지탱한다. 오른손이 칼자루를 쥐고 왼손이 길쭉한 칼집을 잡고 있어 부유하는 소품은 없다. 바닥 무기들은 땅에 놓였고 오른쪽 장검은 바닥과 진열대 쪽에 기대어 있다. 무너진 판넬도 바닥에 닿아 있다. 자세 자체는 가능하지만 팔을 내린 상태라 요구된 칼을 들이미는 행동을 보여주지 못한다."
       },
       {
        "label": "B",
        "direction": "남자는 화면 왼쪽 바깥을 경계하며 오른팔을 그쪽으로 뻗는다. 칼날은 손에서 왼쪽 아래로 비스듬히 이어져 길의 앞쪽 공간을 가로막고, 렌즈를 향하지 않는다. 상대가 보이지 않아 실제 안규홍의 몸을 겨누는지는 확인할 수 없지만, 시선과 칼을 내민 방향이 일치해 화면 밖 접근자를 저지하는 동작으로 읽힌다.",
        "built_space": "한 사람이 목조 점포 사이 흙길을 막고 서 있다. 왼쪽에는 짚신 한 쌍과 낮은 목조 진열대, 바닥 바구니 두 개가 보인다. 오른쪽 벽에도 짚신 한 쌍이 있고 바구니를 얹은 진열대가 있다. 흰 천막, 낡은 목재와 흙길은 참고 장소의 성격을 유지한다. 다만 명백히 파손된 좌판은 보이지 않아 그 상태의 연속성을 확인하기 어렵다. 머리부터 종아리 일부까지 보이고 발은 하단 밖으로 잘려, 요구한 굳은 전신을 온전히 담지 못한다.",
        "entities": "검은 머리의 젊은 성인 동아시아계 남성 한 명이다. 한국인 남성 설정과 외형상 충돌하지 않으며 저고리와 바지 위에 서양식 조끼와 어두운 하오리를 겹쳐 입어 지정 복장에 부합한다. 오른손의 곡도는 금속 날과 장식된 칼막이를 갖추지만 참고의 표면 마모까지 정확히 같지는 않다. 허리에는 손에 든 칼과 별개인 칼자루와 칼막이가 보인다. 뒤쪽 흙바닥에는 떨어진 장검류 하나가 있다. 상단 간판은 일부만 잘려 보여 읽을 수 있는 문구는 확인되지 않는다.",
        "hard_violations": [
         "오른손에 이미 칼을 뽑아 들었는데도 허리에 별도의 칼자루와 칼막이가 남아 있어, 지정되지 않은 두 번째 착용 칼이 추가되었다."
        ],
        "physics": "뻗은 오른손이 칼자루를 확실히 잡고 있으며 손목과 팔의 연결도 자연스럽다. 허리의 칼집은 허리끈 쪽에 매달려 있고 떨어진 무기는 흙바닥에 놓여 있다. 발은 프레임 밖이지만 상체와 다리는 정상적인 입식 자세로 이어져 있어 공중에 떠 있다고 볼 근거는 없다. 팔을 뻗은 채 몸을 굳히고 진로를 막는 동작은 물리적으로 가능하다."
       }
      ],
      "all_candidates_fail": true
     },
     "normalized": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 2,
        "verdict_ko": "전신 와이드 구도는 맞지만 칼을 상대의 접근로가 아닌 땅으로 내리고 정면을 응시하며, 추가 무기와 간판 글자도 금지 조건을 어긴다."
       },
       {
        "label": "A",
        "score": 4,
        "verdict_ko": "왼쪽 접근로로 칼을 내밀고 같은 쪽을 경계하는 동작은 더 충실하지만, 발이 잘린 구도와 허리의 별도 칼 때문에 완전한 적합작은 아니다."
       }
      ],
      "all_candidates_fail": true,
      "readings": [
       {
        "label": "B",
        "direction": "남자는 카메라 정면을 바라본다. 오른손의 칼은 화면 왼쪽 아래로 내려가며 칼끝이 가까운 흙바닥을 향한다. 렌즈를 겨누지는 않지만 안규홍의 접근을 향해 칼을 들이미는 동작보다는 칼을 내려 든 자세다. 화면에 상대는 없다.",
        "built_space": "양쪽 목조 점포 사이 흙길 중앙에 한 사람이 서 있다. 왼쪽 가까운 기둥에는 짚신 한 쌍이 걸려 있고 그 아래로 목조 진열대와 바구니들이 있다. 오른쪽에는 천을 덮은 진열대와 기울어 무너진 목조 판넬 하나가 보인다. 양쪽 천막과 낡은 목재는 참고 장소의 재질에 부합하지만, 참고에서 확인되지 않는 긴 시장 골목까지 확장되어 정확한 고정 배치는 확인하기 어렵다.",
        "entities": "검은 머리의 젊은 성인 동아시아계 남성 한 명으로, 한국인 남성 설정과 외형상 충돌하지 않는다. 조선식 저고리와 바지 위에 어두운 하오리를 입었다. 다만 얼굴과 가르마 머리는 이전 장면 인물과 매우 유사하다. 손에 든 곡도는 참고의 금속 칼과 대체로 같은 종류이나 장식과 마모의 정확한 일치는 부족하다. 허리에는 별도 칼자루가 있고 왼손에는 칼집처럼 보이는 물체가 있다. 바닥의 떨어진 칼 외에도 왼쪽 바닥과 오른쪽 진열대 옆에 추가 장검류가 보인다. 상단 왼쪽 간판에는 판독 가능한 한글 글자 일부가 드러난다.",
        "hard_violations": [
         "손에 든 칼과 별개로 허리에 다른 칼자루가 있으며, 지정된 떨어진 칼 이외에도 왼쪽 바닥과 오른쪽 진열대 옆에 추가 무기가 배치되어 있다.",
         "왼쪽 상단 간판에 판독 가능한 한글이 노출되어 글자 금지 조건을 위반한다."
        ],
        "physics": "양발이 흙바닥에 닿아 체중을 지탱한다. 오른손이 칼자루를 쥐고 왼손이 길쭉한 칼집을 잡고 있어 부유하는 소품은 없다. 바닥 무기들은 땅에 놓였고 오른쪽 장검은 바닥과 진열대 쪽에 기대어 있다. 무너진 판넬도 바닥에 닿아 있다. 자세 자체는 가능하지만 팔을 내린 상태라 요구된 칼을 들이미는 행동을 보여주지 못한다."
       },
       {
        "label": "A",
        "direction": "남자는 화면 왼쪽 바깥을 경계하며 오른팔을 그쪽으로 뻗는다. 칼날은 손에서 왼쪽 아래로 비스듬히 이어져 길의 앞쪽 공간을 가로막고, 렌즈를 향하지 않는다. 상대가 보이지 않아 실제 안규홍의 몸을 겨누는지는 확인할 수 없지만, 시선과 칼을 내민 방향이 일치해 화면 밖 접근자를 저지하는 동작으로 읽힌다.",
        "built_space": "한 사람이 목조 점포 사이 흙길을 막고 서 있다. 왼쪽에는 짚신 한 쌍과 낮은 목조 진열대, 바닥 바구니 두 개가 보인다. 오른쪽 벽에도 짚신 한 쌍이 있고 바구니를 얹은 진열대가 있다. 흰 천막, 낡은 목재와 흙길은 참고 장소의 성격을 유지한다. 다만 명백히 파손된 좌판은 보이지 않아 그 상태의 연속성을 확인하기 어렵다. 머리부터 종아리 일부까지 보이고 발은 하단 밖으로 잘려, 요구한 굳은 전신을 온전히 담지 못한다.",
        "entities": "검은 머리의 젊은 성인 동아시아계 남성 한 명이다. 한국인 남성 설정과 외형상 충돌하지 않으며 저고리와 바지 위에 서양식 조끼와 어두운 하오리를 겹쳐 입어 지정 복장에 부합한다. 오른손의 곡도는 금속 날과 장식된 칼막이를 갖추지만 참고의 표면 마모까지 정확히 같지는 않다. 허리에는 손에 든 칼과 별개인 칼자루와 칼막이가 보인다. 뒤쪽 흙바닥에는 떨어진 장검류 하나가 있다. 상단 간판은 일부만 잘려 보여 읽을 수 있는 문구는 확인되지 않는다.",
        "hard_violations": [
         "오른손에 이미 칼을 뽑아 들었는데도 허리에 별도의 칼자루와 칼막이가 남아 있어, 지정되지 않은 두 번째 착용 칼이 추가되었다."
        ],
        "physics": "뻗은 오른손이 칼자루를 확실히 잡고 있으며 손목과 팔의 연결도 자연스럽다. 허리의 칼집은 허리끈 쪽에 매달려 있고 떨어진 무기는 흙바닥에 놓여 있다. 발은 프레임 밖이지만 상체와 다리는 정상적인 입식 자세로 이어져 있어 공중에 떠 있다고 볼 근거는 없다. 팔을 뻗은 채 몸을 굳히고 진로를 막는 동작은 물리적으로 가능하다."
       }
      ]
     },
     "ok": true
    }
   ],
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "slot_winner_match": true,
   "slot_winner": {
    "gemini-pro": "A",
    "gpt-high": "A"
   },
   "route": "cross_slot_agree"
  },
  "dual": {
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "normalized": {
    "A": 2.0,
    "B": 1.0
   },
   "adjusted": {
    "A": 1.75,
    "B": 0.75
   },
   "violations": {
    "A": [
     "[gemini-pro] 레퍼런스 이미지의 인물 얼굴을 그대로 복사하여 사용함 (금지된 인물 등장)",
     "[gemini-pro] 오른손이 칼자루가 아닌 칼날 부분을 쥐고 있으며 코등이(츠바)가 손목 뒤쪽에 위치함 (물리적으로 불가능한 해부학 및 무기 형태)",
     "[gpt-high] 오른손에 이미 칼을 뽑아 들었는데도 허리에 별도의 칼자루와 칼막이가 남아 있어, 지정되지 않은 두 번째 착용 칼이 추가되었다."
    ],
    "B": [
     "[gemini-pro] 레퍼런스 이미지의 인물 얼굴을 그대로 복사하여 사용함 (금지된 인물 등장)",
     "[gemini-pro] 왼쪽 상단 가판대에 읽을 수 있는 한글 텍스트가 뚜렷하게 노출됨 (텍스트 유출)",
     "[gemini-pro] 주먹 쥔 왼손 뒤쪽에 검은색 칼집이 손에 쥐어지거나 고정되지 않은 채 허공에 떠 있음 (물리적으로 지지되지 않은 사물)",
     "[gpt-high] 손에 든 칼과 별개로 허리에 다른 칼자루가 있으며, 지정된 떨어진 칼 이외에도 왼쪽 바닥과 오른쪽 진열대 옆에 추가 무기가 배치되어 있다.",
     "[gpt-high] 왼쪽 상단 간판에 판독 가능한 한글이 노출되어 글자 금지 조건을 위반한다."
    ]
   },
   "per_model_winner": {
    "gemini-pro": "A",
    "gpt-high": "A"
   },
   "agreed": true
  },
  "totals": {
   "A": 1750,
   "B": 750
  },
  "selected": "A",
  "ranking": [
   "A",
   "B"
  ],
  "verdicts": [
   {
    "label": "A",
    "score": 1750,
    "verdict_ko": "프롬프트가 명시적으로 금지한 이전 장면(레퍼런스) 인물의 얼굴을 그대로 복사해 왔으며, 오른손이 칼자루가 아닌 칼날을 쥐는 물리적으로 불가능한 파지법을 그려내어 탈락 수준의 오류를 보입니다.  ★위반: [gemini-pro] 레퍼런스 이미지의 인물 얼굴을 그대로 복사하여 사용함 (금지된 인물 등장) / [gemini-pro] 오른손이 칼자루가 아닌 칼날 부분을 쥐고 있으며 코등이(츠바)가 손목 뒤쪽에 위치함 (물리적으로 불가능한 해부학 및 무기 형태) / [gpt-high] 오른손에 이미 칼을 뽑아 들었는데도 허리에 별도의 칼자루와 칼막이가 남아 있어, 지정되지 않은 두 번째 착용 칼이 추가되었다."
   },
   {
    "label": "B",
    "score": 750,
    "verdict_ko": "레퍼런스 인물의 얼굴을 차용한 심각한 위반에 더해, 좌측 상단에 읽을 수 있는 간판 텍스트가 노출되고 왼손 부근에 지지되지 않은 채 떠 있는 칼집이 있어 프롬프트를 전방위로 위반했습니다.  ★위반: [gemini-pro] 레퍼런스 이미지의 인물 얼굴을 그대로 복사하여 사용함 (금지된 인물 등장) / [gemini-pro] 왼쪽 상단 가판대에 읽을 수 있는 한글 텍스트가 뚜렷하게 노출됨 (텍스트 유출) / [gemini-pro] 주먹 쥔 왼손 뒤쪽에 검은색 칼집이 손에 쥐어지거나 고정되지 않은 채 허공에 떠 있음 (물리적으로 지지되지 않은 사물) / [gpt-high] 손에 든 칼과 별개로 허리에 다른 칼자루가 있으며, 지정된 떨어진 칼 이외에도 왼쪽 바닥과 오른쪽 진열대 옆에 추가 무기가 배치되어 있다. / [gpt-high] 왼쪽 상단 간판에 판독 가능한 한글이 노출되어 글자 금지 조건을 위반한다."
   }
  ],
  "refs": [
   {
    "label": "PREVIOUS SHOT STILL — a visually related earlier shot of this same place: the location's look, materials, fixed features and lighting mood are LOCKED to this photo; never copy its camera framing. The people visible in that photo are NOT in this shot: never carry their faces, bodies or clothing onto anyone here. Each person in THIS shot is defined solely by the PEOPLE list, their own CHARACTER REFERENCE images and their wardrobe notes. If this photo shows a character who cannot move (dead or unconscious), that character's exact pose, position and orientation are ALSO LOCKED — treat their body as a fixed prop of the set that only the camera moves around.",
    "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/db00517b-165b-4005-a287-f28936b59b95/images/272bd7af-9294-4490-98c6-88a40e9deacc/scene/recipe/S53sh2_sel.png",
    "asset_id": "589ffa1a-6307-4d54-b4ad-2c769f29f91b",
    "role": "prev_still"
   },
   {
    "label": "PROP REFERENCE — 칼: the exact object appearing in this shot; match its look, material and wear exactly.",
    "path": "<bytes:780122>",
    "asset_id": "82614014-6091-423c-89ca-1e15d209de72",
    "role": "prop_ref"
   }
  ],
  "critique_skipped": true,
  "needs_reshoot": true,
  "shot_run_produced": false,
  "shot_run_uid": "06aa1437-8f74-7ca0-a48b-716912193e30",
  "ref_mode": "prev+엔티티",
  "share_plan": {
   "ref_plan": "prev",
   "prev_anchor_tag": "S53sh2"
  }
 },
 "S53sh6::cine": {
  "applied": true,
  "attempted_at": "2026-09-09T07:07:49.417841+00:00",
  "fingerprint": "4c6ae872ecab4ef986d172d767a4fcef67e7ff8f462bf71c45d17b0082f87930",
  "fingerprint_version": 2,
  "provider": "mai",
  "endpoint": "openrouter/chat-completions",
  "model": "microsoft/mai-image-2.6",
  "pack": "24.202608252115",
  "source_file": "S53sh6_sel.png",
  "source_sha256": "110b5cfe1d1c181be909d89041e781dda336537d7e1b0294898aaa641751373e",
  "file": "S53sh6_cine.png",
  "staged_sha256": "ad06b73cfaeac20f3dbffa105a694cc35aa5c0f415ba01240b1923adda7dc5b7",
  "latency_ms": 27588
 },
 "S53sh10::signage": {
  "fp": "15ac69144c8325a7",
  "inscriptions": [],
  "cues": [],
  "dropped": []
 },
 "S53sh10": {
  "input_fingerprint": "c235bbd3983c9bce",
  "prompt": "Create ONE FINAL photorealistic live-action film still of the moment below — early 20th-century Korea, Jeolla region; all people are Korean unless explicitly stated otherwise. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 이를 꽉 악문 채 다시 일진회 대장 쪽으로 매섭게 몸을 돌린 안규홍의 단호한 정면.\n\nLOCATION (lock): The outdoor combat space beside the market stalls, directly facing the leader blocking the lane. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: medium shot\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Hold the established daylight and naturalistic contrast unchanged, allowing the tightened expression rather than a lighting shift to sharpen the confrontation.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The people visible in that photo are NOT in this shot: never carry their faces, bodies or clothing onto anyone here. Each person in THIS shot is defined solely by the PEOPLE list, their own CHARACTER REFERENCE images and their wardrobe notes. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nPREVIOUS STILL USAGE (follow exactly — what to take from the attached still and what to exclude; it governs place and objects only, never who is in this shot or how the camera sees them): Take the sunlit dirt ground and nearby stall structures visible in the reference. Exclude duplicated loose weapons and preserve any stall damage already present.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The market lane remains obstructed and littered with the disrupted stall and dropped sword. 안규홍: He turns back into the fight with his sickle, his back wound continuing to bleed through his jeogori.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 안규홍 (한국인 남성, 청년 나이대, 검은 머리, 단단히 묶은 머리) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.",
  "roll_prompts": {
   "A": "Create ONE FINAL photorealistic live-action film still of the moment below — early 20th-century Korea, Jeolla region; all people are Korean unless explicitly stated otherwise. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 이를 꽉 악문 채 다시 일진회 대장 쪽으로 매섭게 몸을 돌린 안규홍의 단호한 정면.\n\nLOCATION (lock): The outdoor combat space beside the market stalls, directly facing the leader blocking the lane. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: medium shot\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Hold the established daylight and naturalistic contrast unchanged, allowing the tightened expression rather than a lighting shift to sharpen the confrontation.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The people visible in that photo are NOT in this shot: never carry their faces, bodies or clothing onto anyone here. Each person in THIS shot is defined solely by the PEOPLE list, their own CHARACTER REFERENCE images and their wardrobe notes. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nPREVIOUS STILL USAGE (follow exactly — what to take from the attached still and what to exclude; it governs place and objects only, never who is in this shot or how the camera sees them): Take the sunlit dirt ground and nearby stall structures visible in the reference. Exclude duplicated loose weapons and preserve any stall damage already present.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The market lane remains obstructed and littered with the disrupted stall and dropped sword. 안규홍: He turns back into the fight with his sickle, his back wound continuing to bleed through his jeogori.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 안규홍 (한국인 남성, 청년 나이대, 검은 머리, 단단히 묶은 머리) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.",
   "B": "Create ONE FINAL photorealistic live-action film still of the moment below — early 20th-century Korea, Jeolla region; all people are Korean unless explicitly stated otherwise. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 이를 꽉 악문 채 다시 일진회 대장 쪽으로 매섭게 몸을 돌린 안규홍의 단호한 정면.\n\nLOCATION (lock): The outdoor combat space beside the market stalls, directly facing the leader blocking the lane. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: medium shot\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Hold the established daylight and naturalistic contrast unchanged, allowing the tightened expression rather than a lighting shift to sharpen the confrontation.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The people visible in that photo are NOT in this shot: never carry their faces, bodies or clothing onto anyone here. Each person in THIS shot is defined solely by the PEOPLE list, their own CHARACTER REFERENCE images and their wardrobe notes. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nPREVIOUS STILL USAGE (follow exactly — what to take from the attached still and what to exclude; it governs place and objects only, never who is in this shot or how the camera sees them): Take the sunlit dirt ground and nearby stall structures visible in the reference. Exclude duplicated loose weapons and preserve any stall damage already present.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The market lane remains obstructed and littered with the disrupted stall and dropped sword. 안규홍: He turns back into the fight with his sickle, his back wound continuing to bleed through his jeogori.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 안규홍 (한국인 남성, 청년 나이대, 검은 머리, 단단히 묶은 머리) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.\n\nCOMPOSITION VARIATION (this candidate only): keep the place, the people and the moment exactly as specified, and keep from the CAMERA & FRAME contract its subject scale, screen placement and key background elements exactly. For this candidate only, its camera angle and height are deliberately left open: choose a camera position distinctly different from the obvious one, as a film director picking a second setup on the same blocking. Stay true to any explicit framing the SHOT TEXT itself dictates."
  },
  "shot_run_spend_attempt_count": 1,
  "readings": [
   {
    "label": "A",
    "direction": "안규홍의 시선은 화면 우측 전경에 위치한 일진회 대장을 향해 날카롭게 고정되어 있다.",
    "built_space": "시장 가판대 사이의 흙바닥으로 이전 샷의 야외 전투 공간과 동일하며, 바닥에 칼과 칼집이 떨어져 있고 인물과 배경의 공간적 호응이 자연스럽다.",
    "entities": "안규홍은 흰색 저고리에 피가 배어나는 상처, 단단히 묶은 검은 머리를 한 청년으로 레퍼런스와 정확히 일치하며, 지문에 따라 이를 꽉 악문 표정을 보여준다.",
    "hard_violations": [
     "[gpt-high] 화면 오른쪽에 안규홍이 아닌 남자의 머리와 상체를 추가해 인물 제한과 이전 스틸 인물 반입 금지를 위반한다.",
     "[gpt-high] 바닥 앞쪽의 참조와 유사한 칼집형 무기 외에 통로 뒤쪽에 별도의 떨어진 칼을 추가해 중복된 바닥 무기 금지를 위반한다."
    ],
    "physics": "안규홍은 흙바닥에 안정적으로 체중을 싣고 몸을 비튼 자세를 취하고 있으며, 물리적으로 어긋난 부분이나 공중에 떠 있는 요소는 없다."
   },
   {
    "label": "B",
    "direction": "안규홍의 시선은 우측 전경의 일진회 대장을 향하고 있으며, 들고 있는 낫은 앞쪽을 향하고 있다.",
    "built_space": "이전 샷과 동일한 시장 가판대 옆 흙길이며, 바닥에 칼집이 떨어져 있고 우측 전경에 대장의 뒷모습이 배치되어 있다.",
    "entities": "안규홍의 외형(흰 옷, 핏자국, 머리 모양)과 낫을 든 모습은 제시된 조건과 일치하나, 입을 굳게 다물고 있어 '이를 꽉 악문 채'라는 지문과는 다르다.",
    "hard_violations": [
     "[gpt-high] 화면 오른쪽에 안규홍이 아닌 남자의 머리와 상체를 추가했다. 이번 프레임의 인물 제한 및 이전 스틸 인물의 외모·복장 반입 금지를 위반한다."
    ],
    "physics": "인물들은 바닥에 서서 몸을 지탱하고 있으며, 안규홍의 오른손이 낫의 자루를 단단히 쥐고 있다."
   }
  ],
  "cross_model_order": {
   "policy": "cross_model_order_v2_slot0_fwd_slot1_rev_parallel2",
   "slots": [
    {
     "model": "gemini-pro",
     "order": "forward",
     "display_to_canonical": {
      "A": "A",
      "B": "B"
     },
     "raw": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 10,
        "verdict_ko": "지문에서 명시한 '이를 꽉 악문 채'라는 핵심적인 표정 묘사와 매섭게 몸을 돌린 역동적인 자세를 레퍼런스에 맞춰 완벽하게 구현했습니다."
       },
       {
        "label": "B",
        "score": 6,
        "verdict_ko": "낫을 든 모습과 배경 연출은 좋으나, 지문에서 가장 중요한 감정 표현인 '이를 꽉 악문 채'를 전혀 구현하지 못해 매서운 느낌이 부족합니다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "안규홍의 시선은 화면 우측 전경에 위치한 일진회 대장을 향해 날카롭게 고정되어 있다.",
        "built_space": "시장 가판대 사이의 흙바닥으로 이전 샷의 야외 전투 공간과 동일하며, 바닥에 칼과 칼집이 떨어져 있고 인물과 배경의 공간적 호응이 자연스럽다.",
        "entities": "안규홍은 흰색 저고리에 피가 배어나는 상처, 단단히 묶은 검은 머리를 한 청년으로 레퍼런스와 정확히 일치하며, 지문에 따라 이를 꽉 악문 표정을 보여준다.",
        "hard_violations": [],
        "physics": "안규홍은 흙바닥에 안정적으로 체중을 싣고 몸을 비튼 자세를 취하고 있으며, 물리적으로 어긋난 부분이나 공중에 떠 있는 요소는 없다."
       },
       {
        "label": "B",
        "direction": "안규홍의 시선은 우측 전경의 일진회 대장을 향하고 있으며, 들고 있는 낫은 앞쪽을 향하고 있다.",
        "built_space": "이전 샷과 동일한 시장 가판대 옆 흙길이며, 바닥에 칼집이 떨어져 있고 우측 전경에 대장의 뒷모습이 배치되어 있다.",
        "entities": "안규홍의 외형(흰 옷, 핏자국, 머리 모양)과 낫을 든 모습은 제시된 조건과 일치하나, 입을 굳게 다물고 있어 '이를 꽉 악문 채'라는 지문과는 다르다.",
        "hard_violations": [],
        "physics": "인물들은 바닥에 서서 몸을 지탱하고 있으며, 안규홍의 오른손이 낫의 자루를 단단히 쥐고 있다."
       }
      ],
      "all_candidates_fail": false
     },
     "normalized": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 10,
        "verdict_ko": "지문에서 명시한 '이를 꽉 악문 채'라는 핵심적인 표정 묘사와 매섭게 몸을 돌린 역동적인 자세를 레퍼런스에 맞춰 완벽하게 구현했습니다."
       },
       {
        "label": "B",
        "score": 6,
        "verdict_ko": "낫을 든 모습과 배경 연출은 좋으나, 지문에서 가장 중요한 감정 표현인 '이를 꽉 악문 채'를 전혀 구현하지 못해 매서운 느낌이 부족합니다."
       }
      ],
      "all_candidates_fail": false,
      "readings": [
       {
        "label": "A",
        "direction": "안규홍의 시선은 화면 우측 전경에 위치한 일진회 대장을 향해 날카롭게 고정되어 있다.",
        "built_space": "시장 가판대 사이의 흙바닥으로 이전 샷의 야외 전투 공간과 동일하며, 바닥에 칼과 칼집이 떨어져 있고 인물과 배경의 공간적 호응이 자연스럽다.",
        "entities": "안규홍은 흰색 저고리에 피가 배어나는 상처, 단단히 묶은 검은 머리를 한 청년으로 레퍼런스와 정확히 일치하며, 지문에 따라 이를 꽉 악문 표정을 보여준다.",
        "hard_violations": [],
        "physics": "안규홍은 흙바닥에 안정적으로 체중을 싣고 몸을 비튼 자세를 취하고 있으며, 물리적으로 어긋난 부분이나 공중에 떠 있는 요소는 없다."
       },
       {
        "label": "B",
        "direction": "안규홍의 시선은 우측 전경의 일진회 대장을 향하고 있으며, 들고 있는 낫은 앞쪽을 향하고 있다.",
        "built_space": "이전 샷과 동일한 시장 가판대 옆 흙길이며, 바닥에 칼집이 떨어져 있고 우측 전경에 대장의 뒷모습이 배치되어 있다.",
        "entities": "안규홍의 외형(흰 옷, 핏자국, 머리 모양)과 낫을 든 모습은 제시된 조건과 일치하나, 입을 굳게 다물고 있어 '이를 꽉 악문 채'라는 지문과는 다르다.",
        "hard_violations": [],
        "physics": "인물들은 바닥에 서서 몸을 지탱하고 있으며, 안규홍의 오른손이 낫의 자루를 단단히 쥐고 있다."
       }
      ]
     },
     "ok": true
    },
    {
     "model": "gpt-high",
     "order": "reverse",
     "display_to_canonical": {
      "A": "B",
      "B": "A"
     },
     "raw": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 3,
        "verdict_ko": "금지된 상대 인물을 추가해 실격이지만, B와 달리 시선이 상대에게 정확히 닿고 낫을 든 허리 위 회전 동작이 명확하다. 다만 이를 악문 단호한 정면은 부족하다."
       },
       {
        "label": "B",
        "score": 1,
        "verdict_ko": "금지된 상대 인물과 중복된 바닥 무기를 추가했으며, 안규홍의 시선이 오른쪽 상대가 아닌 왼쪽으로 향하고 낫도 식별되지 않아 대치 동작이 어긋난다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "안규홍은 몸통을 비스듬히 둔 채 얼굴을 화면 오른쪽으로 돌려, 오른쪽 전경 남자의 얼굴을 노려본다. 상대 남자도 안규홍 쪽으로 머리를 향한다. 오른손의 낫은 세워져 있고 굽은 날끝은 상대의 몸 쪽인 오른쪽을 향한다. 상대를 향한 방향 관계는 명확하지만 얼굴은 정면보다 사선이며, 입은 거의 닫혀 있어 이를 꽉 악문 표현은 약하다.",
        "built_space": "흙길 양쪽에 낡은 목조 점포가 있고, 왼쪽에는 천을 덮은 판매대 하나, 오른쪽에는 바구니를 올린 판매대 하나가 뚜렷하다. 왼쪽 중경과 오른쪽 인물 뒤에 각각 기울어진 목재 구조물이 있어 흐트러진 좌판 상태가 드러난다. 포대와 바구니, 바닥의 긴 무기 하나가 보이며 참조의 시장 재료와 낮빛을 대체로 유지한다. 안규홍은 통로 중앙, 추가 남자는 오른쪽 전경을 차지한다. 반사면이나 명백히 중복된 고정 설비는 보이지 않는다.",
        "entities": "주인공은 검은 머리를 뒤로 단단히 넘겨 묶은 청년 한국인 남성으로 표현되며, 참조와 유사한 얼굴 윤곽과 체격, 때 묻은 미색 저고리와 허리띠를 갖췄다. 등에 붉은 피가 배어 있고 오른손에는 금속 곡선 날과 나무 손잡이를 갖춘 낫이 있다. 바닥에는 참조와 비슷한 긴 칼집 형태의 무기 하나가 있다. 그러나 오른쪽에 이전 스틸 인물과 유사한 짧은 머리와 짙은 옷의 남자를 추가했다. 읽을 수 있는 글자는 없다.",
        "hard_violations": [
         "화면 오른쪽에 안규홍이 아닌 남자의 머리와 상체를 추가했다. 이번 프레임의 인물 제한 및 이전 스틸 인물의 외모·복장 반입 금지를 위반한다."
        ],
        "physics": "오른손이 낫 손잡이를 실제로 감싸 쥐어 낫의 지지가 명확하다. 어깨와 허리의 비틀림은 서 있는 상태에서 뒤돌아서는 동작으로 가능하다. 발은 프레임 밖이므로 접지는 직접 확인할 수 없지만 공중에 뜬 자세는 아니다. 바닥 무기는 흙 위에 놓여 있고, 기울어진 목재는 지면과 주변 구조물에 기대어 있다. 판매대와 바구니에도 지지면이 보인다."
       },
       {
        "label": "B",
        "direction": "안규홍의 얼굴은 카메라 쪽으로 돌아와 있으나 눈동자는 화면 왼쪽을 노려본다. 실제로 보이는 상대는 화면 오른쪽 전경에 있어 시선이 그 남자에게 닿지 않는다. 상대 남자는 안규홍 쪽으로 머리를 향하고, 그 옆의 직선 칼날은 왼쪽 위를 향하지만 안규홍의 몸에 직접 겨눠진 선은 아니다. 드러낸 치아와 긴장한 턱은 이를 악문 표정에 가깝지만, 대장 쪽으로 돌아선다는 핵심 방향 관계가 어긋난다.",
        "built_space": "왼쪽 목조 점포의 긴 마루와 걸린 짚 물건, 천을 덮은 판매대 하나, 오른쪽의 가까운 판매대 하나와 뒤쪽 판매대 하나, 양쪽 차양이 보인다. 왼쪽 바닥의 바구니 두 개와 오른쪽 포대 등은 참조 배치에 가깝다. 안규홍은 통로 중앙에 서 있고 추가 남자가 오른쪽 전경에 있다. 통로의 좌판 파손이나 큰 장애물은 뚜렷하지 않으며, 바닥에는 앞쪽의 긴 칼집형 무기 외에 뒤쪽의 별도 칼이 보인다.",
        "entities": "주인공은 참조와 대체로 유사한 청년 한국인 남성의 얼굴과 검게 묶은 머리, 미색 저고리와 허리띠를 갖췄고 등에 피가 배어 있다. 얼굴의 치아 노출은 A보다 강하다. 손은 몸 앞쪽 아래에 일부 보이지만 낫의 손잡이나 곡선 날은 식별되지 않는다. 오른쪽에는 허용되지 않은 짙은 옷의 남자와 직선 칼날, 허리 부근 칼자루가 있다. 바닥 앞쪽 무기와 뒤쪽의 별도 칼이 함께 보여 무기 중복 금지에도 어긋난다. 읽을 수 있는 글자는 없다.",
        "hard_violations": [
         "화면 오른쪽에 안규홍이 아닌 남자의 머리와 상체를 추가해 인물 제한과 이전 스틸 인물 반입 금지를 위반한다.",
         "바닥 앞쪽의 참조와 유사한 칼집형 무기 외에 통로 뒤쪽에 별도의 떨어진 칼을 추가해 중복된 바닥 무기 금지를 위반한다."
        ],
        "physics": "몸통과 목을 비튼 자세는 서서 돌아보는 동작으로 가능하며, 공중에 뜬 신체는 없다. 발은 프레임 밖이라 접지는 확인되지 않는다. 바닥 무기 두 개는 흙에 놓여 있고 바구니와 포대도 바닥이나 판매대에 지지된다. 오른쪽 칼날의 손잡이와 쥔 손은 남자의 몸에 가려져 지지 방식을 직접 확인할 수 없으나, 그것만으로 떠 있다고 단정할 수는 없다."
       }
      ],
      "all_candidates_fail": true
     },
     "normalized": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 3,
        "verdict_ko": "금지된 상대 인물을 추가해 실격이지만, B와 달리 시선이 상대에게 정확히 닿고 낫을 든 허리 위 회전 동작이 명확하다. 다만 이를 악문 단호한 정면은 부족하다."
       },
       {
        "label": "A",
        "score": 1,
        "verdict_ko": "금지된 상대 인물과 중복된 바닥 무기를 추가했으며, 안규홍의 시선이 오른쪽 상대가 아닌 왼쪽으로 향하고 낫도 식별되지 않아 대치 동작이 어긋난다."
       }
      ],
      "all_candidates_fail": true,
      "readings": [
       {
        "label": "B",
        "direction": "안규홍은 몸통을 비스듬히 둔 채 얼굴을 화면 오른쪽으로 돌려, 오른쪽 전경 남자의 얼굴을 노려본다. 상대 남자도 안규홍 쪽으로 머리를 향한다. 오른손의 낫은 세워져 있고 굽은 날끝은 상대의 몸 쪽인 오른쪽을 향한다. 상대를 향한 방향 관계는 명확하지만 얼굴은 정면보다 사선이며, 입은 거의 닫혀 있어 이를 꽉 악문 표현은 약하다.",
        "built_space": "흙길 양쪽에 낡은 목조 점포가 있고, 왼쪽에는 천을 덮은 판매대 하나, 오른쪽에는 바구니를 올린 판매대 하나가 뚜렷하다. 왼쪽 중경과 오른쪽 인물 뒤에 각각 기울어진 목재 구조물이 있어 흐트러진 좌판 상태가 드러난다. 포대와 바구니, 바닥의 긴 무기 하나가 보이며 참조의 시장 재료와 낮빛을 대체로 유지한다. 안규홍은 통로 중앙, 추가 남자는 오른쪽 전경을 차지한다. 반사면이나 명백히 중복된 고정 설비는 보이지 않는다.",
        "entities": "주인공은 검은 머리를 뒤로 단단히 넘겨 묶은 청년 한국인 남성으로 표현되며, 참조와 유사한 얼굴 윤곽과 체격, 때 묻은 미색 저고리와 허리띠를 갖췄다. 등에 붉은 피가 배어 있고 오른손에는 금속 곡선 날과 나무 손잡이를 갖춘 낫이 있다. 바닥에는 참조와 비슷한 긴 칼집 형태의 무기 하나가 있다. 그러나 오른쪽에 이전 스틸 인물과 유사한 짧은 머리와 짙은 옷의 남자를 추가했다. 읽을 수 있는 글자는 없다.",
        "hard_violations": [
         "화면 오른쪽에 안규홍이 아닌 남자의 머리와 상체를 추가했다. 이번 프레임의 인물 제한 및 이전 스틸 인물의 외모·복장 반입 금지를 위반한다."
        ],
        "physics": "오른손이 낫 손잡이를 실제로 감싸 쥐어 낫의 지지가 명확하다. 어깨와 허리의 비틀림은 서 있는 상태에서 뒤돌아서는 동작으로 가능하다. 발은 프레임 밖이므로 접지는 직접 확인할 수 없지만 공중에 뜬 자세는 아니다. 바닥 무기는 흙 위에 놓여 있고, 기울어진 목재는 지면과 주변 구조물에 기대어 있다. 판매대와 바구니에도 지지면이 보인다."
       },
       {
        "label": "A",
        "direction": "안규홍의 얼굴은 카메라 쪽으로 돌아와 있으나 눈동자는 화면 왼쪽을 노려본다. 실제로 보이는 상대는 화면 오른쪽 전경에 있어 시선이 그 남자에게 닿지 않는다. 상대 남자는 안규홍 쪽으로 머리를 향하고, 그 옆의 직선 칼날은 왼쪽 위를 향하지만 안규홍의 몸에 직접 겨눠진 선은 아니다. 드러낸 치아와 긴장한 턱은 이를 악문 표정에 가깝지만, 대장 쪽으로 돌아선다는 핵심 방향 관계가 어긋난다.",
        "built_space": "왼쪽 목조 점포의 긴 마루와 걸린 짚 물건, 천을 덮은 판매대 하나, 오른쪽의 가까운 판매대 하나와 뒤쪽 판매대 하나, 양쪽 차양이 보인다. 왼쪽 바닥의 바구니 두 개와 오른쪽 포대 등은 참조 배치에 가깝다. 안규홍은 통로 중앙에 서 있고 추가 남자가 오른쪽 전경에 있다. 통로의 좌판 파손이나 큰 장애물은 뚜렷하지 않으며, 바닥에는 앞쪽의 긴 칼집형 무기 외에 뒤쪽의 별도 칼이 보인다.",
        "entities": "주인공은 참조와 대체로 유사한 청년 한국인 남성의 얼굴과 검게 묶은 머리, 미색 저고리와 허리띠를 갖췄고 등에 피가 배어 있다. 얼굴의 치아 노출은 A보다 강하다. 손은 몸 앞쪽 아래에 일부 보이지만 낫의 손잡이나 곡선 날은 식별되지 않는다. 오른쪽에는 허용되지 않은 짙은 옷의 남자와 직선 칼날, 허리 부근 칼자루가 있다. 바닥 앞쪽 무기와 뒤쪽의 별도 칼이 함께 보여 무기 중복 금지에도 어긋난다. 읽을 수 있는 글자는 없다.",
        "hard_violations": [
         "화면 오른쪽에 안규홍이 아닌 남자의 머리와 상체를 추가해 인물 제한과 이전 스틸 인물 반입 금지를 위반한다.",
         "바닥 앞쪽의 참조와 유사한 칼집형 무기 외에 통로 뒤쪽에 별도의 떨어진 칼을 추가해 중복된 바닥 무기 금지를 위반한다."
        ],
        "physics": "몸통과 목을 비튼 자세는 서서 돌아보는 동작으로 가능하며, 공중에 뜬 신체는 없다. 발은 프레임 밖이라 접지는 확인되지 않는다. 바닥 무기 두 개는 흙에 놓여 있고 바구니와 포대도 바닥이나 판매대에 지지된다. 오른쪽 칼날의 손잡이와 쥔 손은 남자의 몸에 가려져 지지 방식을 직접 확인할 수 없으나, 그것만으로 떠 있다고 단정할 수는 없다."
       }
      ]
     },
     "ok": true
    }
   ],
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "slot_winner_match": false,
   "slot_winner": {
    "gemini-pro": "A",
    "gpt-high": "B"
   },
   "route": "cross_slot_combined"
  },
  "dual": {
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "normalized": {
    "A": 1.333,
    "B": 1.6
   },
   "adjusted": {
    "A": 1.083,
    "B": 1.35
   },
   "violations": {
    "B": [
     "[gpt-high] 화면 오른쪽에 안규홍이 아닌 남자의 머리와 상체를 추가했다. 이번 프레임의 인물 제한 및 이전 스틸 인물의 외모·복장 반입 금지를 위반한다."
    ],
    "A": [
     "[gpt-high] 화면 오른쪽에 안규홍이 아닌 남자의 머리와 상체를 추가해 인물 제한과 이전 스틸 인물 반입 금지를 위반한다.",
     "[gpt-high] 바닥 앞쪽의 참조와 유사한 칼집형 무기 외에 통로 뒤쪽에 별도의 떨어진 칼을 추가해 중복된 바닥 무기 금지를 위반한다."
    ]
   },
   "per_model_winner": {
    "gemini-pro": "A",
    "gpt-high": "B"
   },
   "agreed": false
  },
  "totals": {
   "A": 1083,
   "B": 1350
  },
  "selected": "B",
  "ranking": [
   "B",
   "A"
  ],
  "verdicts": [
   {
    "label": "A",
    "score": 1083,
    "verdict_ko": "지문에서 명시한 '이를 꽉 악문 채'라는 핵심적인 표정 묘사와 매섭게 몸을 돌린 역동적인 자세를 레퍼런스에 맞춰 완벽하게 구현했습니다.  ★위반: [gpt-high] 화면 오른쪽에 안규홍이 아닌 남자의 머리와 상체를 추가해 인물 제한과 이전 스틸 인물 반입 금지를 위반한다. / [gpt-high] 바닥 앞쪽의 참조와 유사한 칼집형 무기 외에 통로 뒤쪽에 별도의 떨어진 칼을 추가해 중복된 바닥 무기 금지를 위반한다."
   },
   {
    "label": "B",
    "score": 1350,
    "verdict_ko": "낫을 든 모습과 배경 연출은 좋으나, 지문에서 가장 중요한 감정 표현인 '이를 꽉 악문 채'를 전혀 구현하지 못해 매서운 느낌이 부족합니다.  ★위반: [gpt-high] 화면 오른쪽에 안규홍이 아닌 남자의 머리와 상체를 추가했다. 이번 프레임의 인물 제한 및 이전 스틸 인물의 외모·복장 반입 금지를 위반한다."
   }
  ],
  "refs": [
   {
    "label": "PREVIOUS SHOT STILL — a visually related earlier shot of this same place: the location's look, materials, fixed features and lighting mood are LOCKED to this photo; never copy its camera framing. The people visible in that photo are NOT in this shot: never carry their faces, bodies or clothing onto anyone here. Each person in THIS shot is defined solely by the PEOPLE list, their own CHARACTER REFERENCE images and their wardrobe notes. If this photo shows a character who cannot move (dead or unconscious), that character's exact pose, position and orientation are ALSO LOCKED — treat their body as a fixed prop of the set that only the camera moves around.",
    "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/db00517b-165b-4005-a287-f28936b59b95/images/272bd7af-9294-4490-98c6-88a40e9deacc/scene/recipe/S53sh6_sel.png",
    "asset_id": "1d718c39-0246-4e90-ae27-67b68e81d242",
    "role": "prev_still"
   },
   {
    "label": "CHARACTER REFERENCE — 안규홍: the exact person appearing in this shot; match face, hair and build exactly.",
    "path": "<bytes:1028384>",
    "asset_id": "3b5894db-ce55-4f34-b966-5e725c3aa952",
    "role": "character_ref"
   }
  ],
  "critique_skipped": true,
  "shot_run_produced": false,
  "shot_run_uid": "06aa1437-910a-77b4-9e47-d7b38e2e5c61",
  "ref_mode": "prev+엔티티",
  "share_plan": {
   "ref_plan": "prev",
   "prev_anchor_tag": "S53sh6"
  }
 },
 "S53sh10::cine": {
  "applied": true,
  "attempted_at": "2026-09-09T07:10:01.731702+00:00",
  "fingerprint": "08a0e8982708d34ca232ad956c44ba9f3378db1c4409af105df2cdb76ac15f9e",
  "fingerprint_version": 2,
  "provider": "mai",
  "endpoint": "openrouter/chat-completions",
  "model": "microsoft/mai-image-2.6",
  "pack": "24.202608252115",
  "source_file": "S53sh10_sel.png",
  "source_sha256": "5b864ac70553418e5e8d803bc70080cee53844189ddd8c93493c5c1565aeec62",
  "file": "S53sh10_cine.png",
  "staged_sha256": "55719411a7b330c0123bb05ca8278080d1d2265ad02ca162320156a6dea60fdf",
  "latency_ms": 27123
 },
 "S54sh4::signage": {
  "fp": "17e302e55bc4f258",
  "inscriptions": [],
  "cues": [],
  "dropped": []
 },
 "S54sh4": {
  "input_fingerprint": "50906e9f85d2f0c4",
  "prompt": "Create ONE FINAL photorealistic live-action film still of the moment below — early 20th-century Korea, Jeolla region; all people are Korean unless explicitly stated otherwise. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 밥을 먹다 말고 숟가락을 허공에 멈춘 채 숲 쪽으로 고개를 확 든 임정현의 정면.\n\nLOCATION (lock): The mountain camp's outdoor meal circle on the dirt clearing, facing the surrounding cypress forest. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: medium shot\n- KEY BACKGROUND ELEMENTS: Spoon (Held motionless during the interrupted meal) — Seen obliquely below 임정현's face; used as Makes the interruption visible through the still hand without enlarging the utensil beyond natural scale; Forest beyond the camp (Quiet as 임정현 lifts his head); used as Remains behind his far shoulder as the source of his attention, without prematurely exposing the gun muzzles.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Use subdued natural daylight with gentle, realistic contrast, preserving the ordinary warmth of the interrupted meal beneath the emerging tension.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The people visible in that photo are NOT in this shot: never carry their faces, bodies or clothing onto anyone here. Each person in THIS shot is defined solely by the PEOPLE list, their own CHARACTER REFERENCE images and their wardrobe notes. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nPREVIOUS STILL USAGE (follow exactly — what to take from the attached still and what to exclude; it governs place and objects only, never who is in this shot or how the camera sees them): Take the dirt ground, ceramic meal bowls, rice, and daytime light. Exclude the sentry's separate forest post and do not add broken dishes before the shooting begins.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The white-rice meal remains intact with the pot upright; hidden ammunition bundles remain beneath stones near trees. 임정현: He sits with his meal interrupted, his head raised toward the forest.\n\nTHE HAND THAT IS DOING THIS: the object at the centre of this shot is being held, operated, read, aimed or handed over by 임정현 right now, so 임정현's hand — and as much of the wrist and forearm as the framing reaches — is in the frame, gripping or resting on that object exactly the way its real use requires. Match that hand to 임정현: its size, build, skin, age, grooming, sleeve and anything worn on it belong to that person and to no one else.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 임정현 (한국인 남성, 성인 나이대, 자연스러운 검은 머리) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.",
  "roll_prompts": {
   "A": "Create ONE FINAL photorealistic live-action film still of the moment below — early 20th-century Korea, Jeolla region; all people are Korean unless explicitly stated otherwise. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 밥을 먹다 말고 숟가락을 허공에 멈춘 채 숲 쪽으로 고개를 확 든 임정현의 정면.\n\nLOCATION (lock): The mountain camp's outdoor meal circle on the dirt clearing, facing the surrounding cypress forest. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: medium shot\n- KEY BACKGROUND ELEMENTS: Spoon (Held motionless during the interrupted meal) — Seen obliquely below 임정현's face; used as Makes the interruption visible through the still hand without enlarging the utensil beyond natural scale; Forest beyond the camp (Quiet as 임정현 lifts his head); used as Remains behind his far shoulder as the source of his attention, without prematurely exposing the gun muzzles.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Use subdued natural daylight with gentle, realistic contrast, preserving the ordinary warmth of the interrupted meal beneath the emerging tension.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The people visible in that photo are NOT in this shot: never carry their faces, bodies or clothing onto anyone here. Each person in THIS shot is defined solely by the PEOPLE list, their own CHARACTER REFERENCE images and their wardrobe notes. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nPREVIOUS STILL USAGE (follow exactly — what to take from the attached still and what to exclude; it governs place and objects only, never who is in this shot or how the camera sees them): Take the dirt ground, ceramic meal bowls, rice, and daytime light. Exclude the sentry's separate forest post and do not add broken dishes before the shooting begins.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The white-rice meal remains intact with the pot upright; hidden ammunition bundles remain beneath stones near trees. 임정현: He sits with his meal interrupted, his head raised toward the forest.\n\nTHE HAND THAT IS DOING THIS: the object at the centre of this shot is being held, operated, read, aimed or handed over by 임정현 right now, so 임정현's hand — and as much of the wrist and forearm as the framing reaches — is in the frame, gripping or resting on that object exactly the way its real use requires. Match that hand to 임정현: its size, build, skin, age, grooming, sleeve and anything worn on it belong to that person and to no one else.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 임정현 (한국인 남성, 성인 나이대, 자연스러운 검은 머리) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.",
   "B": "Create ONE FINAL photorealistic live-action film still of the moment below — early 20th-century Korea, Jeolla region; all people are Korean unless explicitly stated otherwise. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 밥을 먹다 말고 숟가락을 허공에 멈춘 채 숲 쪽으로 고개를 확 든 임정현의 정면.\n\nLOCATION (lock): The mountain camp's outdoor meal circle on the dirt clearing, facing the surrounding cypress forest. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: medium shot\n- KEY BACKGROUND ELEMENTS: Spoon (Held motionless during the interrupted meal) — Seen obliquely below 임정현's face; used as Makes the interruption visible through the still hand without enlarging the utensil beyond natural scale; Forest beyond the camp (Quiet as 임정현 lifts his head); used as Remains behind his far shoulder as the source of his attention, without prematurely exposing the gun muzzles.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Use subdued natural daylight with gentle, realistic contrast, preserving the ordinary warmth of the interrupted meal beneath the emerging tension.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The people visible in that photo are NOT in this shot: never carry their faces, bodies or clothing onto anyone here. Each person in THIS shot is defined solely by the PEOPLE list, their own CHARACTER REFERENCE images and their wardrobe notes. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nPREVIOUS STILL USAGE (follow exactly — what to take from the attached still and what to exclude; it governs place and objects only, never who is in this shot or how the camera sees them): Take the dirt ground, ceramic meal bowls, rice, and daytime light. Exclude the sentry's separate forest post and do not add broken dishes before the shooting begins.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The white-rice meal remains intact with the pot upright; hidden ammunition bundles remain beneath stones near trees. 임정현: He sits with his meal interrupted, his head raised toward the forest.\n\nTHE HAND THAT IS DOING THIS: the object at the centre of this shot is being held, operated, read, aimed or handed over by 임정현 right now, so 임정현's hand — and as much of the wrist and forearm as the framing reaches — is in the frame, gripping or resting on that object exactly the way its real use requires. Match that hand to 임정현: its size, build, skin, age, grooming, sleeve and anything worn on it belong to that person and to no one else.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 임정현 (한국인 남성, 성인 나이대, 자연스러운 검은 머리) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.\n\nCOMPOSITION VARIATION (this candidate only): keep the place, the people and the moment exactly as specified, and keep from the CAMERA & FRAME contract its subject scale, screen placement and key background elements exactly. For this candidate only, its camera angle and height are deliberately left open: choose a camera position distinctly different from the obvious one, as a film director picking a second setup on the same blocking. Stay true to any explicit framing the SHOT TEXT itself dictates."
  },
  "shot_run_spend_attempt_count": 1,
  "initial_roll_all_fail": true,
  "readings": [
   {
    "label": "A",
    "direction": "고개를 들어 숲이 있는 화면 밖 오른쪽 위를 향해 시선을 고정하고 있습니다.",
    "built_space": "참조 이미지에 나타난 야영지의 흙바닥, 나무 울타리, 움막, 가마솥 등이 배경에 올바르게 배치되어 있으며, 인물은 흙바닥에 직접 앉아있습니다.",
    "entities": "임정현의 얼굴과 복장이 참조 이미지와 일치하며, 밥이 담긴 그릇과 숟가락이 묘사되었습니다.",
    "hard_violations": [
     "[gemini-pro] 추가 인물 등장: 프롬프트에 명시되지 않은 다른 사람들의 신체 일부(무릎과 다리)가 화면 좌우 가장자리에 등장함.",
     "[gpt-high] 화면 왼쪽과 오른쪽 가장자리에 임정현이 아닌 두 사람의 하체를 추가했다."
    ],
    "physics": "흙바닥에 안정적으로 앉아 오른손으로 숟가락을 들고 있으나, 화면 가장자리의 다른 인물들 다리 배치가 공간적으로 약간 어색하게 맞닿아 있습니다."
   },
   {
    "label": "B",
    "direction": "고개를 들어 숲이 있는 화면 밖 오른쪽 위를 향해 시선을 두고 있습니다.",
    "built_space": "배경에 울창한 숲이 보이나 참조 이미지의 주요 야영지 구조물(가마솥, 움막 등)이 배제되었고, 인물이 맨바닥이 아닌 흙으로 만들어진 탁자 형태의 구조물 앞에 앉아있습니다.",
    "entities": "임정현의 이목구비와 복장이 일치하며, 밥그릇과 숟가락 등이 놓여 있습니다.",
    "hard_violations": [
     "[gemini-pro] 추가 인물 등장: 프롬프트에 명시되지 않은 여러 인물들이 배경과 화면 우측 전경에 등장함.",
     "[gemini-pro] 잘못된 구조물 배치: 참조 이미지의 야영지 바닥 묘사와 달리 흙으로 된 탁자가 등장함.",
     "[gemini-pro] 물리적 오류: 숟가락의 손잡이가 두 갈래로 나뉘어 허공에 왜곡된 형태로 떠 있음.",
     "[gpt-high] 임정현만 등장해야 하는 장면에 다른 사람 여섯 명을 추가했다.",
     "[gpt-high] 흙바닥에서 이어져야 할 식사 자리에 참조에 없는 식탁형 상판을 추가하고 그 위에 식기를 배치했다."
    ],
    "physics": "흙 탁자 앞에 앉아있으나 숟가락의 손잡이가 물리적으로 불가능한 형태로 갈라져 손에 제대로 쥐어져 있지 않고 허공에 떠 있습니다."
   }
  ],
  "cross_model_order": {
   "policy": "cross_model_order_v2_slot0_fwd_slot1_rev_parallel2",
   "slots": [
    {
     "model": "gemini-pro",
     "order": "forward",
     "display_to_canonical": {
      "A": "A",
      "B": "B"
     },
     "raw": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 4,
        "verdict_ko": "명시되지 않은 다른 인물들의 신체 일부가 화면 가장자리에 나타나 하드 위반에 해당하나, 참조 이미지의 야영지 배경과 바닥에 앉은 설정은 프롬프트를 비교적 잘 구현했습니다."
       },
       {
        "label": "B",
        "score": 2,
        "verdict_ko": "명시되지 않은 다수의 인물들이 등장하고 장소에 맞지 않는 흙 탁자 구조물이 나타났으며, 숟가락의 형태가 크게 왜곡되는 등 다수의 하드 위반이 발생했습니다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "고개를 들어 숲이 있는 화면 밖 오른쪽 위를 향해 시선을 고정하고 있습니다.",
        "built_space": "참조 이미지에 나타난 야영지의 흙바닥, 나무 울타리, 움막, 가마솥 등이 배경에 올바르게 배치되어 있으며, 인물은 흙바닥에 직접 앉아있습니다.",
        "entities": "임정현의 얼굴과 복장이 참조 이미지와 일치하며, 밥이 담긴 그릇과 숟가락이 묘사되었습니다.",
        "hard_violations": [
         "추가 인물 등장: 프롬프트에 명시되지 않은 다른 사람들의 신체 일부(무릎과 다리)가 화면 좌우 가장자리에 등장함."
        ],
        "physics": "흙바닥에 안정적으로 앉아 오른손으로 숟가락을 들고 있으나, 화면 가장자리의 다른 인물들 다리 배치가 공간적으로 약간 어색하게 맞닿아 있습니다."
       },
       {
        "label": "B",
        "direction": "고개를 들어 숲이 있는 화면 밖 오른쪽 위를 향해 시선을 두고 있습니다.",
        "built_space": "배경에 울창한 숲이 보이나 참조 이미지의 주요 야영지 구조물(가마솥, 움막 등)이 배제되었고, 인물이 맨바닥이 아닌 흙으로 만들어진 탁자 형태의 구조물 앞에 앉아있습니다.",
        "entities": "임정현의 이목구비와 복장이 일치하며, 밥그릇과 숟가락 등이 놓여 있습니다.",
        "hard_violations": [
         "추가 인물 등장: 프롬프트에 명시되지 않은 여러 인물들이 배경과 화면 우측 전경에 등장함.",
         "잘못된 구조물 배치: 참조 이미지의 야영지 바닥 묘사와 달리 흙으로 된 탁자가 등장함.",
         "물리적 오류: 숟가락의 손잡이가 두 갈래로 나뉘어 허공에 왜곡된 형태로 떠 있음."
        ],
        "physics": "흙 탁자 앞에 앉아있으나 숟가락의 손잡이가 물리적으로 불가능한 형태로 갈라져 손에 제대로 쥐어져 있지 않고 허공에 떠 있습니다."
       }
      ],
      "all_candidates_fail": true
     },
     "normalized": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 4,
        "verdict_ko": "명시되지 않은 다른 인물들의 신체 일부가 화면 가장자리에 나타나 하드 위반에 해당하나, 참조 이미지의 야영지 배경과 바닥에 앉은 설정은 프롬프트를 비교적 잘 구현했습니다."
       },
       {
        "label": "B",
        "score": 2,
        "verdict_ko": "명시되지 않은 다수의 인물들이 등장하고 장소에 맞지 않는 흙 탁자 구조물이 나타났으며, 숟가락의 형태가 크게 왜곡되는 등 다수의 하드 위반이 발생했습니다."
       }
      ],
      "all_candidates_fail": true,
      "readings": [
       {
        "label": "A",
        "direction": "고개를 들어 숲이 있는 화면 밖 오른쪽 위를 향해 시선을 고정하고 있습니다.",
        "built_space": "참조 이미지에 나타난 야영지의 흙바닥, 나무 울타리, 움막, 가마솥 등이 배경에 올바르게 배치되어 있으며, 인물은 흙바닥에 직접 앉아있습니다.",
        "entities": "임정현의 얼굴과 복장이 참조 이미지와 일치하며, 밥이 담긴 그릇과 숟가락이 묘사되었습니다.",
        "hard_violations": [
         "추가 인물 등장: 프롬프트에 명시되지 않은 다른 사람들의 신체 일부(무릎과 다리)가 화면 좌우 가장자리에 등장함."
        ],
        "physics": "흙바닥에 안정적으로 앉아 오른손으로 숟가락을 들고 있으나, 화면 가장자리의 다른 인물들 다리 배치가 공간적으로 약간 어색하게 맞닿아 있습니다."
       },
       {
        "label": "B",
        "direction": "고개를 들어 숲이 있는 화면 밖 오른쪽 위를 향해 시선을 두고 있습니다.",
        "built_space": "배경에 울창한 숲이 보이나 참조 이미지의 주요 야영지 구조물(가마솥, 움막 등)이 배제되었고, 인물이 맨바닥이 아닌 흙으로 만들어진 탁자 형태의 구조물 앞에 앉아있습니다.",
        "entities": "임정현의 이목구비와 복장이 일치하며, 밥그릇과 숟가락 등이 놓여 있습니다.",
        "hard_violations": [
         "추가 인물 등장: 프롬프트에 명시되지 않은 여러 인물들이 배경과 화면 우측 전경에 등장함.",
         "잘못된 구조물 배치: 참조 이미지의 야영지 바닥 묘사와 달리 흙으로 된 탁자가 등장함.",
         "물리적 오류: 숟가락의 손잡이가 두 갈래로 나뉘어 허공에 왜곡된 형태로 떠 있음."
        ],
        "physics": "흙 탁자 앞에 앉아있으나 숟가락의 손잡이가 물리적으로 불가능한 형태로 갈라져 손에 제대로 쥐어져 있지 않고 허공에 떠 있습니다."
       }
      ]
     },
     "ok": true
    },
    {
     "model": "gpt-high",
     "order": "reverse",
     "display_to_canonical": {
      "A": "B",
      "B": "A"
     },
     "raw": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 2,
        "verdict_ko": "미디엄 숏과 멈춘 숟가락은 맞지만, 금지된 동료들을 다수 추가하고 흙바닥 식사를 식탁 장면으로 바꾸었으며 고개를 확 든 동작도 약하다."
       },
       {
        "label": "B",
        "score": 3,
        "verdict_ko": "기존 야영지와 숟가락을 멈추고 고개를 든 순간은 더 충실하지만, 양옆에 다른 사람의 몸을 추가했고 다리까지 보여 지정된 미디엄 숏보다 넓다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "임정현의 얼굴과 눈은 화면 오른쪽 바깥을 향한다. 숲은 뒤에 있지만 시선이 향하는 실제 지점은 화면 밖이므로 숲을 본다는 연결은 간접적이다. 턱을 급히 들어 올린 모습보다는 식사 중 옆을 보는 모습에 가깝다. 숟가락은 얼굴 아래에서 오른쪽으로 비스듬히 뻗으며 총구는 없다. 주변 사람들은 주로 자기 식기를 내려다본다.",
        "built_space": "흙 공터와 울창한 숲, 양쪽 가장자리의 목재 구조 일부가 보인다. 이전 장면의 움막·목책·화덕 배치는 이 구도로는 확인되지 않는다. 임정현 앞에는 밥그릇과 국그릇을 올린 넓은 식탁형 상판이 있으며, 이는 이전의 흙바닥 식사 배치와 다르다. 주인공 외에 왼쪽 세 명, 뒤쪽 중앙 한 명, 오른쪽 두 명의 인물이 보인다.",
        "entities": "임정현은 검은 머리를 상투와 검은 망건으로 정리한 한국인 성인 남성으로 표현되며, 얼굴과 마른 체격, 때 묻은 밝은색 옷과 검은 허리끈은 인물 참조에 대체로 부합한다. 자연스러운 크기의 금속 숟가락, 흰쌀밥과 도자기 식기가 보인다. 솥과 숨긴 탄약은 화면에서 확인되지 않는다. 허용되지 않은 다른 남성들이 등장한다. 읽을 수 있는 글자나 총기는 없다.",
        "hard_violations": [
         "임정현만 등장해야 하는 장면에 다른 사람 여섯 명을 추가했다.",
         "흙바닥에서 이어져야 할 식사 자리에 참조에 없는 식탁형 상판을 추가하고 그 위에 식기를 배치했다."
        ],
        "physics": "숟가락은 임정현의 오른손 손가락에 실제로 잡혀 있고 손목과 팔뚝이 소매에 자연스럽게 연결된다. 숟가락을 공중에서 멈추는 자세는 가능하다. 식기는 상판이나 지면에 놓여 있으며, 배경 인물들은 흙바닥에 앉아 있다. 임정현의 하체 접촉점은 프레임 아래로 잘리지만 몸이 떠 있다는 증거는 없다."
       },
       {
        "label": "B",
        "direction": "임정현은 몸통을 거의 정면으로 둔 채 고개와 눈을 화면 왼쪽 위로 들어 올렸다. 시선의 끝은 화면 밖이지만 뒤쪽 산림과 이어지는 방향으로 읽혀 숲의 기척에 반응하는 동작이 A보다 분명하다. 숟가락은 얼굴 아래 허리 높이에서 오른쪽으로 비스듬히 뻗어 멈춰 있다. 총구는 노출되지 않았다.",
        "built_space": "흙 공터 뒤로 왼쪽 움막 하나, 중앙 천막형 구조 하나, 오른쪽 움막 하나와 목책이 보인다. 중앙 오른쪽의 흙 화덕, 오른쪽의 큰 원형 금속판 하나와 주변 솥류도 이전 장소의 특징을 이어간다. 임정현은 공터 바닥에 양반다리로 앉고 밥그릇은 바로 앞 지면에 놓여 있다. 다만 화면 양끝에 별도 인물의 다리가 들어왔고, 주인공의 신발까지 포함하여 미디엄 숏보다 넓게 잡았다.",
        "entities": "임정현은 참조와 유사한 얼굴, 검은 상투와 망건, 마른 체격을 가진 한국인 성인 남성으로 표현된다. 밝은색의 해진 겉옷, 가슴의 매듭과 검은 허리끈도 대체로 일치한다. 오른손에 든 금속 숟가락에는 흰쌀밥이 있고 앞에는 온전한 밥그릇이 있다. 오른쪽 전경의 큰 용기는 가장자리가 불규칙하지만 파손 여부는 단정하기 어렵다. 양끝의 다른 사람들은 하체 일부만 보여 신원을 확인할 수 없으나 임정현의 몸은 아니다. 탄약과 읽을 수 있는 글자는 보이지 않는다.",
        "hard_violations": [
         "화면 왼쪽과 오른쪽 가장자리에 임정현이 아닌 두 사람의 하체를 추가했다."
        ],
        "physics": "임정현의 엉덩이와 접힌 다리가 흙바닥에 지지되고 신발도 지면에 닿는다. 오른손이 숟가락 손잡이를 잡고 있으며 밥은 위를 향한 숟가락 오목면에 얹혀 있다. 왼팔은 무릎 위에 걸쳐 손을 자연스럽게 늘어뜨린다. 식기와 솥류는 지면에 놓여 있고 큰 금속판은 기울여 세워져 있다. 지지 없이 떠 있는 신체나 물체는 보이지 않는다."
       }
      ],
      "all_candidates_fail": true
     },
     "normalized": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 2,
        "verdict_ko": "미디엄 숏과 멈춘 숟가락은 맞지만, 금지된 동료들을 다수 추가하고 흙바닥 식사를 식탁 장면으로 바꾸었으며 고개를 확 든 동작도 약하다."
       },
       {
        "label": "A",
        "score": 3,
        "verdict_ko": "기존 야영지와 숟가락을 멈추고 고개를 든 순간은 더 충실하지만, 양옆에 다른 사람의 몸을 추가했고 다리까지 보여 지정된 미디엄 숏보다 넓다."
       }
      ],
      "all_candidates_fail": true,
      "readings": [
       {
        "label": "B",
        "direction": "임정현의 얼굴과 눈은 화면 오른쪽 바깥을 향한다. 숲은 뒤에 있지만 시선이 향하는 실제 지점은 화면 밖이므로 숲을 본다는 연결은 간접적이다. 턱을 급히 들어 올린 모습보다는 식사 중 옆을 보는 모습에 가깝다. 숟가락은 얼굴 아래에서 오른쪽으로 비스듬히 뻗으며 총구는 없다. 주변 사람들은 주로 자기 식기를 내려다본다.",
        "built_space": "흙 공터와 울창한 숲, 양쪽 가장자리의 목재 구조 일부가 보인다. 이전 장면의 움막·목책·화덕 배치는 이 구도로는 확인되지 않는다. 임정현 앞에는 밥그릇과 국그릇을 올린 넓은 식탁형 상판이 있으며, 이는 이전의 흙바닥 식사 배치와 다르다. 주인공 외에 왼쪽 세 명, 뒤쪽 중앙 한 명, 오른쪽 두 명의 인물이 보인다.",
        "entities": "임정현은 검은 머리를 상투와 검은 망건으로 정리한 한국인 성인 남성으로 표현되며, 얼굴과 마른 체격, 때 묻은 밝은색 옷과 검은 허리끈은 인물 참조에 대체로 부합한다. 자연스러운 크기의 금속 숟가락, 흰쌀밥과 도자기 식기가 보인다. 솥과 숨긴 탄약은 화면에서 확인되지 않는다. 허용되지 않은 다른 남성들이 등장한다. 읽을 수 있는 글자나 총기는 없다.",
        "hard_violations": [
         "임정현만 등장해야 하는 장면에 다른 사람 여섯 명을 추가했다.",
         "흙바닥에서 이어져야 할 식사 자리에 참조에 없는 식탁형 상판을 추가하고 그 위에 식기를 배치했다."
        ],
        "physics": "숟가락은 임정현의 오른손 손가락에 실제로 잡혀 있고 손목과 팔뚝이 소매에 자연스럽게 연결된다. 숟가락을 공중에서 멈추는 자세는 가능하다. 식기는 상판이나 지면에 놓여 있으며, 배경 인물들은 흙바닥에 앉아 있다. 임정현의 하체 접촉점은 프레임 아래로 잘리지만 몸이 떠 있다는 증거는 없다."
       },
       {
        "label": "A",
        "direction": "임정현은 몸통을 거의 정면으로 둔 채 고개와 눈을 화면 왼쪽 위로 들어 올렸다. 시선의 끝은 화면 밖이지만 뒤쪽 산림과 이어지는 방향으로 읽혀 숲의 기척에 반응하는 동작이 A보다 분명하다. 숟가락은 얼굴 아래 허리 높이에서 오른쪽으로 비스듬히 뻗어 멈춰 있다. 총구는 노출되지 않았다.",
        "built_space": "흙 공터 뒤로 왼쪽 움막 하나, 중앙 천막형 구조 하나, 오른쪽 움막 하나와 목책이 보인다. 중앙 오른쪽의 흙 화덕, 오른쪽의 큰 원형 금속판 하나와 주변 솥류도 이전 장소의 특징을 이어간다. 임정현은 공터 바닥에 양반다리로 앉고 밥그릇은 바로 앞 지면에 놓여 있다. 다만 화면 양끝에 별도 인물의 다리가 들어왔고, 주인공의 신발까지 포함하여 미디엄 숏보다 넓게 잡았다.",
        "entities": "임정현은 참조와 유사한 얼굴, 검은 상투와 망건, 마른 체격을 가진 한국인 성인 남성으로 표현된다. 밝은색의 해진 겉옷, 가슴의 매듭과 검은 허리끈도 대체로 일치한다. 오른손에 든 금속 숟가락에는 흰쌀밥이 있고 앞에는 온전한 밥그릇이 있다. 오른쪽 전경의 큰 용기는 가장자리가 불규칙하지만 파손 여부는 단정하기 어렵다. 양끝의 다른 사람들은 하체 일부만 보여 신원을 확인할 수 없으나 임정현의 몸은 아니다. 탄약과 읽을 수 있는 글자는 보이지 않는다.",
        "hard_violations": [
         "화면 왼쪽과 오른쪽 가장자리에 임정현이 아닌 두 사람의 하체를 추가했다."
        ],
        "physics": "임정현의 엉덩이와 접힌 다리가 흙바닥에 지지되고 신발도 지면에 닿는다. 오른손이 숟가락 손잡이를 잡고 있으며 밥은 위를 향한 숟가락 오목면에 얹혀 있다. 왼팔은 무릎 위에 걸쳐 손을 자연스럽게 늘어뜨린다. 식기와 솥류는 지면에 놓여 있고 큰 금속판은 기울여 세워져 있다. 지지 없이 떠 있는 신체나 물체는 보이지 않는다."
       }
      ]
     },
     "ok": true
    }
   ],
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "slot_winner_match": true,
   "slot_winner": {
    "gemini-pro": "A",
    "gpt-high": "A"
   },
   "route": "cross_slot_agree"
  },
  "dual": {
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "normalized": {
    "A": 2.0,
    "B": 1.167
   },
   "adjusted": {
    "A": 1.75,
    "B": 0.917
   },
   "violations": {
    "A": [
     "[gemini-pro] 추가 인물 등장: 프롬프트에 명시되지 않은 다른 사람들의 신체 일부(무릎과 다리)가 화면 좌우 가장자리에 등장함.",
     "[gpt-high] 화면 왼쪽과 오른쪽 가장자리에 임정현이 아닌 두 사람의 하체를 추가했다."
    ],
    "B": [
     "[gemini-pro] 추가 인물 등장: 프롬프트에 명시되지 않은 여러 인물들이 배경과 화면 우측 전경에 등장함.",
     "[gemini-pro] 잘못된 구조물 배치: 참조 이미지의 야영지 바닥 묘사와 달리 흙으로 된 탁자가 등장함.",
     "[gemini-pro] 물리적 오류: 숟가락의 손잡이가 두 갈래로 나뉘어 허공에 왜곡된 형태로 떠 있음.",
     "[gpt-high] 임정현만 등장해야 하는 장면에 다른 사람 여섯 명을 추가했다.",
     "[gpt-high] 흙바닥에서 이어져야 할 식사 자리에 참조에 없는 식탁형 상판을 추가하고 그 위에 식기를 배치했다."
    ]
   },
   "per_model_winner": {
    "gemini-pro": "A",
    "gpt-high": "A"
   },
   "agreed": true
  },
  "totals": {
   "A": 1750,
   "B": 917
  },
  "selected": "A",
  "ranking": [
   "A",
   "B"
  ],
  "verdicts": [
   {
    "label": "A",
    "score": 1750,
    "verdict_ko": "명시되지 않은 다른 인물들의 신체 일부가 화면 가장자리에 나타나 하드 위반에 해당하나, 참조 이미지의 야영지 배경과 바닥에 앉은 설정은 프롬프트를 비교적 잘 구현했습니다.  ★위반: [gemini-pro] 추가 인물 등장: 프롬프트에 명시되지 않은 다른 사람들의 신체 일부(무릎과 다리)가 화면 좌우 가장자리에 등장함. / [gpt-high] 화면 왼쪽과 오른쪽 가장자리에 임정현이 아닌 두 사람의 하체를 추가했다."
   },
   {
    "label": "B",
    "score": 917,
    "verdict_ko": "명시되지 않은 다수의 인물들이 등장하고 장소에 맞지 않는 흙 탁자 구조물이 나타났으며, 숟가락의 형태가 크게 왜곡되는 등 다수의 하드 위반이 발생했습니다.  ★위반: [gemini-pro] 추가 인물 등장: 프롬프트에 명시되지 않은 여러 인물들이 배경과 화면 우측 전경에 등장함. / [gemini-pro] 잘못된 구조물 배치: 참조 이미지의 야영지 바닥 묘사와 달리 흙으로 된 탁자가 등장함. / [gemini-pro] 물리적 오류: 숟가락의 손잡이가 두 갈래로 나뉘어 허공에 왜곡된 형태로 떠 있음. / [gpt-high] 임정현만 등장해야 하는 장면에 다른 사람 여섯 명을 추가했다. / [gpt-high] 흙바닥에서 이어져야 할 식사 자리에 참조에 없는 식탁형 상판을 추가하고 그 위에 식기를 배치했다."
   }
  ],
  "refs": [
   {
    "label": "PREVIOUS SHOT STILL — a visually related earlier shot of this same place: the location's look, materials, fixed features and lighting mood are LOCKED to this photo; never copy its camera framing. The people visible in that photo are NOT in this shot: never carry their faces, bodies or clothing onto anyone here. Each person in THIS shot is defined solely by the PEOPLE list, their own CHARACTER REFERENCE images and their wardrobe notes. If this photo shows a character who cannot move (dead or unconscious), that character's exact pose, position and orientation are ALSO LOCKED — treat their body as a fixed prop of the set that only the camera moves around.",
    "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/db00517b-165b-4005-a287-f28936b59b95/images/272bd7af-9294-4490-98c6-88a40e9deacc/scene/recipe/S50sh1_sel.png",
    "asset_id": "1624ad2e-c774-4af4-b6d5-eefeff78f9d3",
    "role": "prev_still"
   },
   {
    "label": "CHARACTER REFERENCE — 임정현: the exact person appearing in this shot; match face, hair and build exactly.",
    "path": "<bytes:858188>",
    "asset_id": "789c3359-d7f8-4832-9a6f-7e4ad2695c02",
    "role": "character_ref"
   }
  ],
  "critique_skipped": true,
  "needs_reshoot": true,
  "shot_run_produced": false,
  "shot_run_uid": "06aa1437-92b6-7848-bdec-0a0d4d3a7ed8",
  "ref_mode": "prev+엔티티",
  "share_plan": {
   "ref_plan": "prev",
   "prev_anchor_tag": "S50sh1"
  }
 },
 "S54sh4::cine": {
  "applied": true,
  "attempted_at": "2026-09-09T07:11:28.188226+00:00",
  "fingerprint": "c5e128516b7f62537f39e31348ed94e96048803335ff72b5f40e2a9416b33b60",
  "fingerprint_version": 2,
  "provider": "mai",
  "endpoint": "openrouter/chat-completions",
  "model": "microsoft/mai-image-2.6",
  "pack": "24.202608252115",
  "source_file": "S54sh4_sel.png",
  "source_sha256": "788b884472cad122818ec90ab350ed46aacfb998213ac803d4a533aaefd8a5d1",
  "file": "S54sh4_cine.png",
  "staged_sha256": "e82713126d6338ba0e10f95c94c9213c6c6d8004fca13e218d62a06f9c2078b3",
  "latency_ms": 28588
 },
 "S54sh7::signage": {
  "fp": "03527e90fe3a4607",
  "inscriptions": [],
  "cues": [],
  "dropped": []
 },
 "S54sh7": {
  "input_fingerprint": "33f42e9342b55ca7",
  "prompt": "Create ONE FINAL photorealistic live-action film still of the moment below — early 20th-century Korea, Jeolla region; all people are Korean unless explicitly stated otherwise. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 임정현의 등 뒤 편백나무 틈새로 여러 개의 검은 총구들이 일제히 산채 쪽을 향해 튀어나온 섬뜩한 전경.\n\nLOCATION (lock): The cypress-tree perimeter directly behind the mountain camp's eating ground, with gaps facing inward toward the clearing. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: wide shot\n- FRAME LAYOUT: Left cypress gap with a muzzle aimed into camp in the middle-left of the frame, background; Right cypress gap with a muzzle aimed into camp in the middle-right of the frame, background.\n- KEY BACKGROUND ELEMENTS: Cypress trees (Separate gaps conceal the approaching gunmen) — Trunks and openings are viewed across the camp toward the forest; used as Create separated background apertures without allowing 임정현's body to mask the threat; Black gun muzzles (Visible between trees on opposite sides, aimed toward the camp) — Project from separate gaps toward the camp foreground at differing oblique angles; used as Form small, distinct threat cues on both sides of the background; no single muzzle exceeds natural scale.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Retain the preceding daylight and muted contrast, letting the black muzzles become readable through composition rather than an artificial darkening of the scene.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The people visible in that photo are NOT in this shot: never carry their faces, bodies or clothing onto anyone here. Each person in THIS shot is defined solely by the PEOPLE list, their own CHARACTER REFERENCE images and their wardrobe notes. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The intact rice meal and upright pot remain in the encircled camp; ammunition is concealed at prepared tree-and-stone positions.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 임정현 (한국인 남성, 성인 나이대, 자연스러운 검은 머리) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.",
  "roll_prompts": {
   "A": "Create ONE FINAL photorealistic live-action film still of the moment below — early 20th-century Korea, Jeolla region; all people are Korean unless explicitly stated otherwise. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 임정현의 등 뒤 편백나무 틈새로 여러 개의 검은 총구들이 일제히 산채 쪽을 향해 튀어나온 섬뜩한 전경.\n\nLOCATION (lock): The cypress-tree perimeter directly behind the mountain camp's eating ground, with gaps facing inward toward the clearing. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: wide shot\n- FRAME LAYOUT: Left cypress gap with a muzzle aimed into camp in the middle-left of the frame, background; Right cypress gap with a muzzle aimed into camp in the middle-right of the frame, background.\n- KEY BACKGROUND ELEMENTS: Cypress trees (Separate gaps conceal the approaching gunmen) — Trunks and openings are viewed across the camp toward the forest; used as Create separated background apertures without allowing 임정현's body to mask the threat; Black gun muzzles (Visible between trees on opposite sides, aimed toward the camp) — Project from separate gaps toward the camp foreground at differing oblique angles; used as Form small, distinct threat cues on both sides of the background; no single muzzle exceeds natural scale.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Retain the preceding daylight and muted contrast, letting the black muzzles become readable through composition rather than an artificial darkening of the scene.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The people visible in that photo are NOT in this shot: never carry their faces, bodies or clothing onto anyone here. Each person in THIS shot is defined solely by the PEOPLE list, their own CHARACTER REFERENCE images and their wardrobe notes. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The intact rice meal and upright pot remain in the encircled camp; ammunition is concealed at prepared tree-and-stone positions.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 임정현 (한국인 남성, 성인 나이대, 자연스러운 검은 머리) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.",
   "B": "Create ONE FINAL photorealistic live-action film still of the moment below — early 20th-century Korea, Jeolla region; all people are Korean unless explicitly stated otherwise. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 임정현의 등 뒤 편백나무 틈새로 여러 개의 검은 총구들이 일제히 산채 쪽을 향해 튀어나온 섬뜩한 전경.\n\nLOCATION (lock): The cypress-tree perimeter directly behind the mountain camp's eating ground, with gaps facing inward toward the clearing. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: wide shot\n- FRAME LAYOUT: Left cypress gap with a muzzle aimed into camp in the middle-left of the frame, background; Right cypress gap with a muzzle aimed into camp in the middle-right of the frame, background.\n- KEY BACKGROUND ELEMENTS: Cypress trees (Separate gaps conceal the approaching gunmen) — Trunks and openings are viewed across the camp toward the forest; used as Create separated background apertures without allowing 임정현's body to mask the threat; Black gun muzzles (Visible between trees on opposite sides, aimed toward the camp) — Project from separate gaps toward the camp foreground at differing oblique angles; used as Form small, distinct threat cues on both sides of the background; no single muzzle exceeds natural scale.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Retain the preceding daylight and muted contrast, letting the black muzzles become readable through composition rather than an artificial darkening of the scene.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The people visible in that photo are NOT in this shot: never carry their faces, bodies or clothing onto anyone here. Each person in THIS shot is defined solely by the PEOPLE list, their own CHARACTER REFERENCE images and their wardrobe notes. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The intact rice meal and upright pot remain in the encircled camp; ammunition is concealed at prepared tree-and-stone positions.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 임정현 (한국인 남성, 성인 나이대, 자연스러운 검은 머리) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.\n\nCOMPOSITION VARIATION (this candidate only): keep the place, the people and the moment exactly as specified, and keep from the CAMERA & FRAME contract its subject scale, screen placement and key background elements exactly. For this candidate only, its camera angle and height are deliberately left open: choose a camera position distinctly different from the obvious one, as a film director picking a second setup on the same blocking. Stay true to any explicit framing the SHOT TEXT itself dictates."
  },
  "shot_run_spend_attempt_count": 1,
  "readings": [
   {
    "label": "A",
    "direction": "화면 좌측과 우측의 편백나무 틈새에서 각각 하나의 검은 총구가 전경(산채)을 향해 겨냥되어 있음.",
    "built_space": "편백나무 숲과 돌담이 배경에 배치되어 있으며, 지형에 맞게 나무 사이의 틈새 공간이 형성되어 있음.",
    "entities": "지침에 명시된 두 개의 총구가 나무 뒤에서 나와 있으며, 총을 든 사람의 팔 부분이 부분적으로 보임. 임정현 및 식량(밥, 솥)은 화면에 보이지 않음.",
    "hard_violations": [
     "[gpt-high] 임정현 외에는 사람이나 신체 일부를 보이지 말라는 제한과 달리, 서로 다른 두 나무 뒤에 사수 두 명의 옷과 신체 일부를 노출했다."
    ],
    "physics": "총구는 나무 뒤에 숨은 사람의 팔과 손에 의해 물리적으로 자연스럽게 지지되고 있음."
   },
   {
    "label": "B",
    "direction": "화면 중앙의 인물과 전경을 향해 좌우 편백나무에서 여러 개의 총구가 겨냥되어 있음.",
    "built_space": "편백나무들이 공간을 둘러싸고 있으며, 숲 한가운데 바위 위에 식량과 솥이 놓여 있음.",
    "entities": "지정된 복장을 한 임정현의 뒷모습이 중앙에 있으며, 바위 위에 밥과 솥이 있음. 여러 개의 총구가 나무 기둥에 박혀 있음.",
    "hard_violations": [
     "[gemini-pro] 지지되지 않은 사물 및 물리적으로 불가능한 연출 (소총들이 총을 쥔 사람의 손 없이 단단한 나무 기둥 한가운데를 관통하여 허공에 떠 있음)"
    ],
    "physics": "여러 개의 소총이 지지하는 손이나 사람 없이 단단한 나무 기둥을 뚫고 튀어나와 허공에 완전히 떠 있는 물리적으로 불가능한 상태임."
   }
  ],
  "cross_model_order": {
   "policy": "cross_model_order_v2_slot0_fwd_slot1_rev_parallel2",
   "slots": [
    {
     "model": "gemini-pro",
     "order": "forward",
     "display_to_canonical": {
      "A": "A",
      "B": "B"
     },
     "raw": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 7,
        "verdict_ko": "좌우 나무 틈새에 총구를 배치하는 프레임 지침을 잘 따랐고 물리적 현실성을 유지했으나, 임정현의 모습과 솥 등의 소품이 프레임에서 누락되었습니다."
       },
       {
        "label": "B",
        "score": 2,
        "verdict_ko": "임정현의 뒷모습과 식량 소품을 배치하여 샷 텍스트의 요소를 포함했으나, 소총들이 사람의 손 없이 단단한 나무 기둥을 뚫고 나와 허공에 떠 있는 치명적인 물리적 오류가 있습니다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "화면 좌측과 우측의 편백나무 틈새에서 각각 하나의 검은 총구가 전경(산채)을 향해 겨냥되어 있음.",
        "built_space": "편백나무 숲과 돌담이 배경에 배치되어 있으며, 지형에 맞게 나무 사이의 틈새 공간이 형성되어 있음.",
        "entities": "지침에 명시된 두 개의 총구가 나무 뒤에서 나와 있으며, 총을 든 사람의 팔 부분이 부분적으로 보임. 임정현 및 식량(밥, 솥)은 화면에 보이지 않음.",
        "hard_violations": [],
        "physics": "총구는 나무 뒤에 숨은 사람의 팔과 손에 의해 물리적으로 자연스럽게 지지되고 있음."
       },
       {
        "label": "B",
        "direction": "화면 중앙의 인물과 전경을 향해 좌우 편백나무에서 여러 개의 총구가 겨냥되어 있음.",
        "built_space": "편백나무들이 공간을 둘러싸고 있으며, 숲 한가운데 바위 위에 식량과 솥이 놓여 있음.",
        "entities": "지정된 복장을 한 임정현의 뒷모습이 중앙에 있으며, 바위 위에 밥과 솥이 있음. 여러 개의 총구가 나무 기둥에 박혀 있음.",
        "hard_violations": [
         "지지되지 않은 사물 및 물리적으로 불가능한 연출 (소총들이 총을 쥔 사람의 손 없이 단단한 나무 기둥 한가운데를 관통하여 허공에 떠 있음)"
        ],
        "physics": "여러 개의 소총이 지지하는 손이나 사람 없이 단단한 나무 기둥을 뚫고 튀어나와 허공에 완전히 떠 있는 물리적으로 불가능한 상태임."
       }
      ],
      "all_candidates_fail": false
     },
     "normalized": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 7,
        "verdict_ko": "좌우 나무 틈새에 총구를 배치하는 프레임 지침을 잘 따랐고 물리적 현실성을 유지했으나, 임정현의 모습과 솥 등의 소품이 프레임에서 누락되었습니다."
       },
       {
        "label": "B",
        "score": 2,
        "verdict_ko": "임정현의 뒷모습과 식량 소품을 배치하여 샷 텍스트의 요소를 포함했으나, 소총들이 사람의 손 없이 단단한 나무 기둥을 뚫고 나와 허공에 떠 있는 치명적인 물리적 오류가 있습니다."
       }
      ],
      "all_candidates_fail": false,
      "readings": [
       {
        "label": "A",
        "direction": "화면 좌측과 우측의 편백나무 틈새에서 각각 하나의 검은 총구가 전경(산채)을 향해 겨냥되어 있음.",
        "built_space": "편백나무 숲과 돌담이 배경에 배치되어 있으며, 지형에 맞게 나무 사이의 틈새 공간이 형성되어 있음.",
        "entities": "지침에 명시된 두 개의 총구가 나무 뒤에서 나와 있으며, 총을 든 사람의 팔 부분이 부분적으로 보임. 임정현 및 식량(밥, 솥)은 화면에 보이지 않음.",
        "hard_violations": [],
        "physics": "총구는 나무 뒤에 숨은 사람의 팔과 손에 의해 물리적으로 자연스럽게 지지되고 있음."
       },
       {
        "label": "B",
        "direction": "화면 중앙의 인물과 전경을 향해 좌우 편백나무에서 여러 개의 총구가 겨냥되어 있음.",
        "built_space": "편백나무들이 공간을 둘러싸고 있으며, 숲 한가운데 바위 위에 식량과 솥이 놓여 있음.",
        "entities": "지정된 복장을 한 임정현의 뒷모습이 중앙에 있으며, 바위 위에 밥과 솥이 있음. 여러 개의 총구가 나무 기둥에 박혀 있음.",
        "hard_violations": [
         "지지되지 않은 사물 및 물리적으로 불가능한 연출 (소총들이 총을 쥔 사람의 손 없이 단단한 나무 기둥 한가운데를 관통하여 허공에 떠 있음)"
        ],
        "physics": "여러 개의 소총이 지지하는 손이나 사람 없이 단단한 나무 기둥을 뚫고 튀어나와 허공에 완전히 떠 있는 물리적으로 불가능한 상태임."
       }
      ]
     },
     "ok": true
    },
    {
     "model": "gpt-high",
     "order": "reverse",
     "display_to_canonical": {
      "A": "B",
      "B": "A"
     },
     "raw": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 5,
        "verdict_ko": "추가 인물 없이 와이드 숏과 양쪽 총구를 구현했지만, 임정현이 위협을 등지지 않고 숲을 바라보며 총구들도 작은 배경 단서보다 크게 강조된다."
       },
       {
        "label": "B",
        "score": 3,
        "verdict_ko": "중간 좌우의 작은 총구와 빈 전경은 구도 지시에 더 충실하지만, 허용되지 않은 두 사수의 신체와 옷을 나무 양쪽에 노출한 것이 결정적 위반이다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "왼쪽의 총열 세 개는 화면 오른쪽의 빈터 안쪽으로, 오른쪽의 총열 여러 개는 왼쪽의 빈터 안쪽으로 향한다. 산채 중앙을 양쪽에서 겨누는 관계는 읽히지만, 전경을 향한 서로 다른 사선보다는 화면을 가로지르는 수평 방향이 강하다. 중앙 남성은 카메라에 등을 보이고 숲을 바라보므로, 총구들이 임정현의 등 뒤에서 나타난다는 관계는 구현되지 않았다.",
        "built_space": "흙과 돌로 된 빈터 뒤에 편백 줄기들이 늘어서고, 낮은 돌담이 왼쪽과 오른쪽 가장자리에 보인다. 중앙의 낮은 바위 하나가 식사 받침으로 쓰인다. 참고의 거친 돌쌓기와 흙바닥 재질은 이어지지만, 열린 산비탈 대신 조밀한 숲 안쪽이 강조된다. 와이드 숏은 맞으나 총구들이 중간 좌우의 작은 배경 요소에 머물지 않고 화면 양 끝까지 반복된다.",
        "entities": "검은 머리의 성인 남성 한 명이 회갈색 저고리와 바지 차림으로 서 있다. 뒷모습이라 한국인 여부나 얼굴 정체성은 확인할 수 없다. 검은 총열은 적어도 일곱 개 보이고, 별도의 사수 신체는 드러나지 않는다. 바위 위에는 흰 쌀밥 한 그릇과 반찬 접시, 작은 컵이 있으며, 옆에는 똑바로 놓인 검은 솥과 받침이 있다. 탄약은 보이지 않아 은폐 상태와 모순되지 않는다. 읽을 수 있는 글자는 없다.",
        "hard_violations": [],
        "physics": "남성은 두 발로 흙바닥을 딛고 있다. 식기는 바위에, 솥은 원통형 받침에, 받침은 지면에 놓여 있다. 총열 뒤쪽은 나무에 가려져 손이나 개머리판의 지지는 확인되지 않는다. 다만 은폐된 사수가 뒤에서 받칠 공간은 있어, 노출된 총열만으로 지지 없는 부유라고 단정할 수는 없다."
       },
       {
        "label": "B",
        "direction": "왼쪽 총구는 오른쪽 빈터 중앙으로, 오른쪽 총구는 왼쪽 빈터 중앙으로 향한다. 두 총구가 산채 안쪽을 교차해 겨누는 배치는 읽히지만, 카메라 쪽 전경으로 향하는 사선은 약하다. 얼굴과 눈은 가려져 시선은 확인되지 않는다.",
        "built_space": "흙 빈터 너머 중간 왼쪽과 중간 오른쪽에 굵은 편백 줄기가 있고, 그 사이와 뒤로 숲이 이어진다. 낮은 돌담은 왼쪽, 중앙 뒤편, 오른쪽에 나뉘어 보인다. 참고의 돌과 흙 재질을 대체로 유지하며, 두 총구를 자연스러운 작은 크기로 떨어뜨려 놓은 와이드 구도는 지시에 가깝다. 식사와 솥은 이 프레임에 보이지 않으며, 화면 밖에 있을 수 있으므로 부재 자체는 위반이 아니다.",
        "entities": "검은 총구 두 개가 보인다. 왼쪽 나무 양옆에는 회갈색 옷을 입은 사람의 몸통 일부와 팔이, 오른쪽 나무 옆에도 다른 사람의 옷 입은 신체 일부가 드러난다. 두 사람의 얼굴은 보이지 않아 나이·성별·한국인 여부는 판별할 수 없다. 이들은 단 한 명만 허용된 임정현으로 함께 성립할 수 없는 추가 사수들이다. 읽을 수 있는 글자는 없다.",
        "hard_violations": [
         "임정현 외에는 사람이나 신체 일부를 보이지 말라는 제한과 달리, 서로 다른 두 나무 뒤에 사수 두 명의 옷과 신체 일부를 노출했다."
        ],
        "physics": "왼쪽 총은 나무 뒤에서 나온 팔과 손 부근에 연결되어 지지되는 모습이다. 오른쪽 총의 손 접촉은 줄기에 가려져 있지만, 뒤에 노출된 사수의 신체와 연결 가능한 위치다. 사수들의 하체는 나무와 돌담에 가려져 있으며 지면에 설 공간이 있다. 명백히 떠 있는 몸이나 물체는 보이지 않는다."
       }
      ],
      "all_candidates_fail": false
     },
     "normalized": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 5,
        "verdict_ko": "추가 인물 없이 와이드 숏과 양쪽 총구를 구현했지만, 임정현이 위협을 등지지 않고 숲을 바라보며 총구들도 작은 배경 단서보다 크게 강조된다."
       },
       {
        "label": "A",
        "score": 3,
        "verdict_ko": "중간 좌우의 작은 총구와 빈 전경은 구도 지시에 더 충실하지만, 허용되지 않은 두 사수의 신체와 옷을 나무 양쪽에 노출한 것이 결정적 위반이다."
       }
      ],
      "all_candidates_fail": false,
      "readings": [
       {
        "label": "B",
        "direction": "왼쪽의 총열 세 개는 화면 오른쪽의 빈터 안쪽으로, 오른쪽의 총열 여러 개는 왼쪽의 빈터 안쪽으로 향한다. 산채 중앙을 양쪽에서 겨누는 관계는 읽히지만, 전경을 향한 서로 다른 사선보다는 화면을 가로지르는 수평 방향이 강하다. 중앙 남성은 카메라에 등을 보이고 숲을 바라보므로, 총구들이 임정현의 등 뒤에서 나타난다는 관계는 구현되지 않았다.",
        "built_space": "흙과 돌로 된 빈터 뒤에 편백 줄기들이 늘어서고, 낮은 돌담이 왼쪽과 오른쪽 가장자리에 보인다. 중앙의 낮은 바위 하나가 식사 받침으로 쓰인다. 참고의 거친 돌쌓기와 흙바닥 재질은 이어지지만, 열린 산비탈 대신 조밀한 숲 안쪽이 강조된다. 와이드 숏은 맞으나 총구들이 중간 좌우의 작은 배경 요소에 머물지 않고 화면 양 끝까지 반복된다.",
        "entities": "검은 머리의 성인 남성 한 명이 회갈색 저고리와 바지 차림으로 서 있다. 뒷모습이라 한국인 여부나 얼굴 정체성은 확인할 수 없다. 검은 총열은 적어도 일곱 개 보이고, 별도의 사수 신체는 드러나지 않는다. 바위 위에는 흰 쌀밥 한 그릇과 반찬 접시, 작은 컵이 있으며, 옆에는 똑바로 놓인 검은 솥과 받침이 있다. 탄약은 보이지 않아 은폐 상태와 모순되지 않는다. 읽을 수 있는 글자는 없다.",
        "hard_violations": [],
        "physics": "남성은 두 발로 흙바닥을 딛고 있다. 식기는 바위에, 솥은 원통형 받침에, 받침은 지면에 놓여 있다. 총열 뒤쪽은 나무에 가려져 손이나 개머리판의 지지는 확인되지 않는다. 다만 은폐된 사수가 뒤에서 받칠 공간은 있어, 노출된 총열만으로 지지 없는 부유라고 단정할 수는 없다."
       },
       {
        "label": "A",
        "direction": "왼쪽 총구는 오른쪽 빈터 중앙으로, 오른쪽 총구는 왼쪽 빈터 중앙으로 향한다. 두 총구가 산채 안쪽을 교차해 겨누는 배치는 읽히지만, 카메라 쪽 전경으로 향하는 사선은 약하다. 얼굴과 눈은 가려져 시선은 확인되지 않는다.",
        "built_space": "흙 빈터 너머 중간 왼쪽과 중간 오른쪽에 굵은 편백 줄기가 있고, 그 사이와 뒤로 숲이 이어진다. 낮은 돌담은 왼쪽, 중앙 뒤편, 오른쪽에 나뉘어 보인다. 참고의 돌과 흙 재질을 대체로 유지하며, 두 총구를 자연스러운 작은 크기로 떨어뜨려 놓은 와이드 구도는 지시에 가깝다. 식사와 솥은 이 프레임에 보이지 않으며, 화면 밖에 있을 수 있으므로 부재 자체는 위반이 아니다.",
        "entities": "검은 총구 두 개가 보인다. 왼쪽 나무 양옆에는 회갈색 옷을 입은 사람의 몸통 일부와 팔이, 오른쪽 나무 옆에도 다른 사람의 옷 입은 신체 일부가 드러난다. 두 사람의 얼굴은 보이지 않아 나이·성별·한국인 여부는 판별할 수 없다. 이들은 단 한 명만 허용된 임정현으로 함께 성립할 수 없는 추가 사수들이다. 읽을 수 있는 글자는 없다.",
        "hard_violations": [
         "임정현 외에는 사람이나 신체 일부를 보이지 말라는 제한과 달리, 서로 다른 두 나무 뒤에 사수 두 명의 옷과 신체 일부를 노출했다."
        ],
        "physics": "왼쪽 총은 나무 뒤에서 나온 팔과 손 부근에 연결되어 지지되는 모습이다. 오른쪽 총의 손 접촉은 줄기에 가려져 있지만, 뒤에 노출된 사수의 신체와 연결 가능한 위치다. 사수들의 하체는 나무와 돌담에 가려져 있으며 지면에 설 공간이 있다. 명백히 떠 있는 몸이나 물체는 보이지 않는다."
       }
      ]
     },
     "ok": true
    }
   ],
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "slot_winner_match": false,
   "slot_winner": {
    "gemini-pro": "A",
    "gpt-high": "B"
   },
   "route": "cross_slot_combined"
  },
  "dual": {
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "normalized": {
    "A": 1.6,
    "B": 1.286
   },
   "adjusted": {
    "A": 1.35,
    "B": 1.036
   },
   "violations": {
    "B": [
     "[gemini-pro] 지지되지 않은 사물 및 물리적으로 불가능한 연출 (소총들이 총을 쥔 사람의 손 없이 단단한 나무 기둥 한가운데를 관통하여 허공에 떠 있음)"
    ],
    "A": [
     "[gpt-high] 임정현 외에는 사람이나 신체 일부를 보이지 말라는 제한과 달리, 서로 다른 두 나무 뒤에 사수 두 명의 옷과 신체 일부를 노출했다."
    ]
   },
   "per_model_winner": {
    "gemini-pro": "A",
    "gpt-high": "B"
   },
   "agreed": false
  },
  "totals": {
   "A": 1350,
   "B": 1036
  },
  "selected": "A",
  "ranking": [
   "A",
   "B"
  ],
  "verdicts": [
   {
    "label": "A",
    "score": 1350,
    "verdict_ko": "좌우 나무 틈새에 총구를 배치하는 프레임 지침을 잘 따랐고 물리적 현실성을 유지했으나, 임정현의 모습과 솥 등의 소품이 프레임에서 누락되었습니다.  ★위반: [gpt-high] 임정현 외에는 사람이나 신체 일부를 보이지 말라는 제한과 달리, 서로 다른 두 나무 뒤에 사수 두 명의 옷과 신체 일부를 노출했다."
   },
   {
    "label": "B",
    "score": 1036,
    "verdict_ko": "임정현의 뒷모습과 식량 소품을 배치하여 샷 텍스트의 요소를 포함했으나, 소총들이 사람의 손 없이 단단한 나무 기둥을 뚫고 나와 허공에 떠 있는 치명적인 물리적 오류가 있습니다.  ★위반: [gemini-pro] 지지되지 않은 사물 및 물리적으로 불가능한 연출 (소총들이 총을 쥔 사람의 손 없이 단단한 나무 기둥 한가운데를 관통하여 허공에 떠 있음)"
   }
  ],
  "refs": [
   {
    "label": "PREVIOUS SHOT STILL — a visually related earlier shot of this same place: the location's look, materials, fixed features and lighting mood are LOCKED to this photo; never copy its camera framing. The people visible in that photo are NOT in this shot: never carry their faces, bodies or clothing onto anyone here. Each person in THIS shot is defined solely by the PEOPLE list, their own CHARACTER REFERENCE images and their wardrobe notes. If this photo shows a character who cannot move (dead or unconscious), that character's exact pose, position and orientation are ALSO LOCKED — treat their body as a fixed prop of the set that only the camera moves around.",
    "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/db00517b-165b-4005-a287-f28936b59b95/images/272bd7af-9294-4490-98c6-88a40e9deacc/scene/recipe/S50sh7_sel.png",
    "asset_id": "e43dcdeb-6c1b-489f-994b-aa1ed12960d7",
    "role": "prev_still"
   }
  ],
  "critique_skipped": true,
  "shot_run_produced": false,
  "shot_run_uid": "06aa1437-9453-7016-9563-f344f41e744f",
  "ref_mode": "prev만 (배경 전용·공유 계획)",
  "share_plan": {
   "ref_plan": "prev",
   "prev_anchor_tag": "S50sh7"
  },
  "lane_policy": "share_plan_prev_bgonly"
 },
 "S54sh7::cine": {
  "applied": true,
  "attempted_at": "2026-09-09T07:13:16.650145+00:00",
  "fingerprint": "3809184ba5288b85a75dfcf06a4356d4c979e3595b14ebf77cf0fda99c25b95d",
  "fingerprint_version": 2,
  "provider": "mai",
  "endpoint": "openrouter/chat-completions",
  "model": "microsoft/mai-image-2.6",
  "pack": "24.202608252115",
  "source_file": "S54sh7_sel.png",
  "source_sha256": "8817c96a999c0e3c2f9ea28d279b0cfc6dbeab009196fac6efe44aa079401833",
  "file": "S54sh7_cine.png",
  "staged_sha256": "f1fe7904deb3bb973ec8c47d89fd1fce3735f076b1b6d720735a62128b5c9f0b",
  "latency_ms": 29045
 },
 "S55sh11::signage": {
  "fp": "381e8024e60efbf0",
  "inscriptions": [],
  "cues": [],
  "dropped": []
 },
 "S55sh11": {
  "input_fingerprint": "818f95929cbf3e8f",
  "prompt": "Create ONE FINAL photorealistic live-action film still of the moment below — early 20th-century Korea, Jeolla region; all people are Korean unless explicitly stated otherwise. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 배에 칼이 박힌 채 눈을 크게 뜨고 몸이 뻣뻣하게 굳어버린 이관회의 상체.\n\nLOCATION (lock): The dirt market lane beside the damaged stalls, at the rear edge of the defenders' fighting position. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: medium shot\n- KEY BACKGROUND ELEMENTS: Knife (Still embedded in 이관회's abdomen before withdrawal) — The exposed portion projects obliquely from the lower torso; used as Provides a restrained injury cue at the bottom of the medium composition without graphic wound emphasis.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Maintain sober daylight, muted saturation, and realistic contrast without visual distortion, allowing bodily stillness to convey the moment's shock.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The clothing, hair and overall look of 이관회 — who appear both in that photo and in this shot — are LOCKED to that photo. Anyone else visible in that photo is NOT in this shot: never carry their face, body or clothing onto anyone here. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The disrupted stall and fallen weapons remain in the market fighting area. 이관회: A sword is embedded in his abdomen; his earlier flank slash is still bleeding, and he retains the club used in the fight.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 이관회 (한국인 남성, 청년 나이대, 자연스러운 검은 머리) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.",
  "roll_prompts": {
   "A": "Create ONE FINAL photorealistic live-action film still of the moment below — early 20th-century Korea, Jeolla region; all people are Korean unless explicitly stated otherwise. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 배에 칼이 박힌 채 눈을 크게 뜨고 몸이 뻣뻣하게 굳어버린 이관회의 상체.\n\nLOCATION (lock): The dirt market lane beside the damaged stalls, at the rear edge of the defenders' fighting position. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: medium shot\n- KEY BACKGROUND ELEMENTS: Knife (Still embedded in 이관회's abdomen before withdrawal) — The exposed portion projects obliquely from the lower torso; used as Provides a restrained injury cue at the bottom of the medium composition without graphic wound emphasis.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Maintain sober daylight, muted saturation, and realistic contrast without visual distortion, allowing bodily stillness to convey the moment's shock.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The clothing, hair and overall look of 이관회 — who appear both in that photo and in this shot — are LOCKED to that photo. Anyone else visible in that photo is NOT in this shot: never carry their face, body or clothing onto anyone here. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The disrupted stall and fallen weapons remain in the market fighting area. 이관회: A sword is embedded in his abdomen; his earlier flank slash is still bleeding, and he retains the club used in the fight.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 이관회 (한국인 남성, 청년 나이대, 자연스러운 검은 머리) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.",
   "B": "Create ONE FINAL photorealistic live-action film still of the moment below — early 20th-century Korea, Jeolla region; all people are Korean unless explicitly stated otherwise. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 배에 칼이 박힌 채 눈을 크게 뜨고 몸이 뻣뻣하게 굳어버린 이관회의 상체.\n\nLOCATION (lock): The dirt market lane beside the damaged stalls, at the rear edge of the defenders' fighting position. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: medium shot\n- KEY BACKGROUND ELEMENTS: Knife (Still embedded in 이관회's abdomen before withdrawal) — The exposed portion projects obliquely from the lower torso; used as Provides a restrained injury cue at the bottom of the medium composition without graphic wound emphasis.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Maintain sober daylight, muted saturation, and realistic contrast without visual distortion, allowing bodily stillness to convey the moment's shock.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The clothing, hair and overall look of 이관회 — who appear both in that photo and in this shot — are LOCKED to that photo. Anyone else visible in that photo is NOT in this shot: never carry their face, body or clothing onto anyone here. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The disrupted stall and fallen weapons remain in the market fighting area. 이관회: A sword is embedded in his abdomen; his earlier flank slash is still bleeding, and he retains the club used in the fight.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 이관회 (한국인 남성, 청년 나이대, 자연스러운 검은 머리) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.\n\nCOMPOSITION VARIATION (this candidate only): keep the place, the people and the moment exactly as specified, and keep from the CAMERA & FRAME contract its subject scale, screen placement and key background elements exactly. For this candidate only, its camera angle and height are deliberately left open: choose a camera position distinctly different from the obvious one, as a film director picking a second setup on the same blocking. Stay true to any explicit framing the SHOT TEXT itself dictates."
  },
  "shot_run_spend_attempt_count": 1,
  "readings": [
   {
    "label": "A",
    "direction": "시선은 화면 좌측을 향하고 있으며, 지시된 대로 눈을 크게 뜨고 있음.",
    "built_space": "흙바닥 시장 골목이나, 참조 이미지의 장소에는 존재하지 않는 모래주머니 진지와 흐릿한 인물들이 배경에 임의로 추가되어 있음.",
    "entities": "이관회의 외모와 복장은 참조 이미지와 일치함. 오른손에 몽둥이를 들고 있고 왼쪽 옆구리에 핏자국이 있으나, 지시된 '박힌 칼' 대신 칼집에 들어있는 칼이 묘사됨.",
    "hard_violations": [
     "[gemini-pro] 배에 박혀야 할 칼이 칼집에 들어있는 채로 복부 앞에 물리적 지지대 없이 허공에 떠 있음 (물리적 오류).",
     "[gemini-pro] 참조 이미지에 존재하지 않는 인물들과 모래주머니 진지가 배경에 추가됨 (장소 설정 위반 및 발명된 사람/오브젝트).",
     "[gpt-high] 방어벽 뒤에 최소 세 명의 추가 인물이 보여, 이관회 한 명만 등장해야 한다는 명시적 조건을 위반한다."
    ],
    "physics": "오른손은 몽둥이를 정상적으로 쥐고 있으나, 복부 앞에 위치한 칼은 신체나 허리띠에 제대로 꽂히지 않고 허공에 뜬 채로 부자연스럽게 위치함."
   },
   {
    "label": "B",
    "direction": "시선은 화면 좌측 전방을 향하고 있으며, 충격으로 눈을 크게 뜬 표정이 잘 드러남.",
    "built_space": "부서진 가판대가 있는 흙바닥 시장 골목. 나무 구조물, 걸려있는 짚신 등 이전 샷 스틸컷의 배경 요소와 정확하게 일치함.",
    "entities": "이관회의 외모와 찢어진 옷 등은 참조 이미지와 일치함. 오른손에 몽둥이를 쥐고 있으며, 왼쪽 옆구리의 출혈과 복부에 비스듬히 박힌 칼(프롭 참조와 일치하는 외형)이 정확히 묘사됨.",
    "hard_violations": [],
    "physics": "오른손으로 몽둥이를 안정적으로 쥐고 있으며, 칼날이 복부의 옷과 살을 뚫고 들어가 물리적으로 단단히 지지되어 있음. 몸이 뻣뻣하게 굳은 자세도 훌륭하게 구현됨."
   }
  ],
  "cross_model_order": {
   "policy": "cross_model_order_v2_slot0_fwd_slot1_rev_parallel2",
   "slots": [
    {
     "model": "gemini-pro",
     "order": "forward",
     "display_to_canonical": {
      "A": "A",
      "B": "B"
     },
     "raw": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 10,
        "verdict_ko": "크게 뜬 눈과 뻣뻣하게 굳은 자세, 배에 비스듬히 박힌 칼, 그리고 이전 샷과 완벽하게 일치하는 배경까지 프롬프트의 모든 지시사항을 훌륭하게 구현했습니다."
       },
       {
        "label": "A",
        "score": 2,
        "verdict_ko": "칼이 배에 박혀 있지 않고 칼집째로 물리적 지지 없이 떠 있으며, 이전 샷에 없던 구조물과 인물이 배경에 추가되는 치명적인 오류를 범했습니다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "시선은 화면 좌측을 향하고 있으며, 지시된 대로 눈을 크게 뜨고 있음.",
        "built_space": "흙바닥 시장 골목이나, 참조 이미지의 장소에는 존재하지 않는 모래주머니 진지와 흐릿한 인물들이 배경에 임의로 추가되어 있음.",
        "entities": "이관회의 외모와 복장은 참조 이미지와 일치함. 오른손에 몽둥이를 들고 있고 왼쪽 옆구리에 핏자국이 있으나, 지시된 '박힌 칼' 대신 칼집에 들어있는 칼이 묘사됨.",
        "hard_violations": [
         "배에 박혀야 할 칼이 칼집에 들어있는 채로 복부 앞에 물리적 지지대 없이 허공에 떠 있음 (물리적 오류).",
         "참조 이미지에 존재하지 않는 인물들과 모래주머니 진지가 배경에 추가됨 (장소 설정 위반 및 발명된 사람/오브젝트)."
        ],
        "physics": "오른손은 몽둥이를 정상적으로 쥐고 있으나, 복부 앞에 위치한 칼은 신체나 허리띠에 제대로 꽂히지 않고 허공에 뜬 채로 부자연스럽게 위치함."
       },
       {
        "label": "B",
        "direction": "시선은 화면 좌측 전방을 향하고 있으며, 충격으로 눈을 크게 뜬 표정이 잘 드러남.",
        "built_space": "부서진 가판대가 있는 흙바닥 시장 골목. 나무 구조물, 걸려있는 짚신 등 이전 샷 스틸컷의 배경 요소와 정확하게 일치함.",
        "entities": "이관회의 외모와 찢어진 옷 등은 참조 이미지와 일치함. 오른손에 몽둥이를 쥐고 있으며, 왼쪽 옆구리의 출혈과 복부에 비스듬히 박힌 칼(프롭 참조와 일치하는 외형)이 정확히 묘사됨.",
        "hard_violations": [],
        "physics": "오른손으로 몽둥이를 안정적으로 쥐고 있으며, 칼날이 복부의 옷과 살을 뚫고 들어가 물리적으로 단단히 지지되어 있음. 몸이 뻣뻣하게 굳은 자세도 훌륭하게 구현됨."
       }
      ],
      "all_candidates_fail": false
     },
     "normalized": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 10,
        "verdict_ko": "크게 뜬 눈과 뻣뻣하게 굳은 자세, 배에 비스듬히 박힌 칼, 그리고 이전 샷과 완벽하게 일치하는 배경까지 프롬프트의 모든 지시사항을 훌륭하게 구현했습니다."
       },
       {
        "label": "A",
        "score": 2,
        "verdict_ko": "칼이 배에 박혀 있지 않고 칼집째로 물리적 지지 없이 떠 있으며, 이전 샷에 없던 구조물과 인물이 배경에 추가되는 치명적인 오류를 범했습니다."
       }
      ],
      "all_candidates_fail": false,
      "readings": [
       {
        "label": "A",
        "direction": "시선은 화면 좌측을 향하고 있으며, 지시된 대로 눈을 크게 뜨고 있음.",
        "built_space": "흙바닥 시장 골목이나, 참조 이미지의 장소에는 존재하지 않는 모래주머니 진지와 흐릿한 인물들이 배경에 임의로 추가되어 있음.",
        "entities": "이관회의 외모와 복장은 참조 이미지와 일치함. 오른손에 몽둥이를 들고 있고 왼쪽 옆구리에 핏자국이 있으나, 지시된 '박힌 칼' 대신 칼집에 들어있는 칼이 묘사됨.",
        "hard_violations": [
         "배에 박혀야 할 칼이 칼집에 들어있는 채로 복부 앞에 물리적 지지대 없이 허공에 떠 있음 (물리적 오류).",
         "참조 이미지에 존재하지 않는 인물들과 모래주머니 진지가 배경에 추가됨 (장소 설정 위반 및 발명된 사람/오브젝트)."
        ],
        "physics": "오른손은 몽둥이를 정상적으로 쥐고 있으나, 복부 앞에 위치한 칼은 신체나 허리띠에 제대로 꽂히지 않고 허공에 뜬 채로 부자연스럽게 위치함."
       },
       {
        "label": "B",
        "direction": "시선은 화면 좌측 전방을 향하고 있으며, 충격으로 눈을 크게 뜬 표정이 잘 드러남.",
        "built_space": "부서진 가판대가 있는 흙바닥 시장 골목. 나무 구조물, 걸려있는 짚신 등 이전 샷 스틸컷의 배경 요소와 정확하게 일치함.",
        "entities": "이관회의 외모와 찢어진 옷 등은 참조 이미지와 일치함. 오른손에 몽둥이를 쥐고 있으며, 왼쪽 옆구리의 출혈과 복부에 비스듬히 박힌 칼(프롭 참조와 일치하는 외형)이 정확히 묘사됨.",
        "hard_violations": [],
        "physics": "오른손으로 몽둥이를 안정적으로 쥐고 있으며, 칼날이 복부의 옷과 살을 뚫고 들어가 물리적으로 단단히 지지되어 있음. 몸이 뻣뻣하게 굳은 자세도 훌륭하게 구현됨."
       }
      ]
     },
     "ok": true
    },
    {
     "model": "gpt-high",
     "order": "reverse",
     "display_to_canonical": {
      "A": "B",
      "B": "A"
     },
     "raw": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 7,
        "verdict_ko": "복부에 비스듬히 박힌 칼과 크게 뜬 눈, 이관회만 등장하는 조건을 지켰지만, 허벅지까지 담은 구도와 차렷에 가까운 자세는 요청한 상체 중심의 순간 포착보다 약하다."
       },
       {
        "label": "B",
        "score": 2,
        "verdict_ko": "상체 중심 구도와 멈칫한 동작은 적절하지만, 배경에 금지된 인물들을 추가했고 복부에 박힌 칼 대신 칼집에 든 검을 허리 앞에 배치했다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "이관회는 눈을 크게 뜨고 카메라 왼쪽의 화면 밖을 바라보며, 시선의 대상은 보이지 않는다. 칼은 화면 아래 왼쪽의 손잡이에서 위쪽 오른쪽의 하복부로 이어지고, 드러난 칼날 끝부분이 저고리 밑 복부로 들어간다. 오른손의 몽둥이는 아래로 향해 화면 밖으로 이어진다.",
        "built_space": "인물은 흙길 중앙에 서 있고 왼쪽에는 목조 점포의 긴 진열대 하나, 매달린 짚신 한 켤레, 바닥의 바구니 여러 개가 보인다. 뒤에는 천 차양과 기울어진 판재가 있고, 오른쪽 길바닥에는 긴 무기류가 놓여 있다. 이전 사진의 낡은 목재·직물·짚 재료와 낮 조명을 대체로 잇는다. 고정 시설의 명백한 중복이나 불가능한 반사는 없다. 다만 상체보다 넓게 허벅지까지 담았다.",
        "entities": "보이는 사람은 이관회 한 명으로, 청년 남성의 얼굴과 검은 가르마 머리, 거친 갈색 저고리와 밝은 바지가 참조와 대체로 일치한다. 눈은 정상적인 홍채와 동공을 유지한 채 크게 떠 있다. 칼의 감긴 손잡이와 금속 코등이는 소품 참조에 가깝고, 노출된 날이 복부에 꽂혀 있다. 화면 오른쪽 옆구리에는 혈흔이 있고 오른손에는 몽둥이 일부가 보인다. 읽을 수 있는 글자는 없다.",
        "hard_violations": [],
        "physics": "몸통은 세워져 있으며 하체가 화면 아래로 이어진다. 발은 구도 밖이므로 지면 접촉은 직접 확인할 수 없지만 공중에 떠 있는 자세는 아니다. 몽둥이는 오른손이 쥐고 있고, 칼은 복부에 삽입된 부분이 지지하는 배치다. 바구니와 떨어진 무기는 지면이나 진열대에 놓여 있다. 다만 양팔을 거의 곧게 내린 대칭적 자세여서 전투 동작 도중 굳어버린 순간보다는 차렷 자세에 가깝다."
       },
       {
        "label": "B",
        "direction": "이관회의 시선은 카메라 왼쪽의 화면 밖을 향하며 눈이 크게 떠 있다. 검 손잡이는 화면 왼쪽 위로, 검은 칼집은 오른쪽 아래로 뻗는다. 검 전체가 허리 앞을 가로지르는 방향으로 보여 칼날이 복부 안으로 향하는 삽입 관계가 확인되지 않는다. 오른손에 쥔 몽둥이의 굵은 끝은 화면 왼쪽으로 향한다.",
        "built_space": "흙길 왼쪽에 천을 덮은 판매대 하나와 낮은 목조 진열대, 바구니들이 있고, 오른쪽에는 기대어 놓은 장총류와 점포 구조물이 보인다. 길 뒤쪽에는 포대 방어벽 하나가 있으며 그 너머로 최소 세 명의 인물이 드러난다. 낡은 시장의 재료와 낮 분위기는 이어지지만 이전 사진의 짚신 진열 구역과 정확히 같은 위치인지는 확인하기 어렵다. 인물은 허리 부근까지 크게 담겨 A보다 상체 중심 구도에 가깝다.",
        "entities": "주인공의 청년 남성 외형, 검은 머리, 얼굴과 해진 갈색 저고리는 참조와 대체로 일치한다. 정상적인 눈을 크게 뜨고 입을 벌린 충격 표현이 보인다. 몽둥이는 오른손에 분명히 들려 있다. 검은 참조의 감긴 손잡이와 장식 있는 검은 칼집을 닮았지만 칼집에서 뽑혀 복부에 박힌 상태가 아니다. 옆구리의 찢어진 천과 상처, 아래로 번진 피는 A보다 강하게 강조된다. 배경 인물들은 이관회 외에는 누구도 등장시키지 말라는 조건에 어긋난다.",
        "hard_violations": [
         "방어벽 뒤에 최소 세 명의 추가 인물이 보여, 이관회 한 명만 등장해야 한다는 명시적 조건을 위반한다."
        ],
        "physics": "몸통을 약간 틀고 양팔을 벌린 자세는 동작 중 멈춘 순간으로 가능하며, 하체와 발은 화면 밖이다. 몽둥이는 오른손의 손가락이 감싸 지지한다. 칼집은 허리 앞에 밀착되어 띠와 겹치므로 떠 있다고 단정할 수는 없지만, 복부에 박힌 칼날이 검을 지지하는 요청된 상태는 보이지 않는다. 바구니와 판재는 바닥이나 판매대에 놓였고 장총류는 점포 구조물에 기대어 있다."
       }
      ],
      "all_candidates_fail": false
     },
     "normalized": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 7,
        "verdict_ko": "복부에 비스듬히 박힌 칼과 크게 뜬 눈, 이관회만 등장하는 조건을 지켰지만, 허벅지까지 담은 구도와 차렷에 가까운 자세는 요청한 상체 중심의 순간 포착보다 약하다."
       },
       {
        "label": "A",
        "score": 2,
        "verdict_ko": "상체 중심 구도와 멈칫한 동작은 적절하지만, 배경에 금지된 인물들을 추가했고 복부에 박힌 칼 대신 칼집에 든 검을 허리 앞에 배치했다."
       }
      ],
      "all_candidates_fail": false,
      "readings": [
       {
        "label": "B",
        "direction": "이관회는 눈을 크게 뜨고 카메라 왼쪽의 화면 밖을 바라보며, 시선의 대상은 보이지 않는다. 칼은 화면 아래 왼쪽의 손잡이에서 위쪽 오른쪽의 하복부로 이어지고, 드러난 칼날 끝부분이 저고리 밑 복부로 들어간다. 오른손의 몽둥이는 아래로 향해 화면 밖으로 이어진다.",
        "built_space": "인물은 흙길 중앙에 서 있고 왼쪽에는 목조 점포의 긴 진열대 하나, 매달린 짚신 한 켤레, 바닥의 바구니 여러 개가 보인다. 뒤에는 천 차양과 기울어진 판재가 있고, 오른쪽 길바닥에는 긴 무기류가 놓여 있다. 이전 사진의 낡은 목재·직물·짚 재료와 낮 조명을 대체로 잇는다. 고정 시설의 명백한 중복이나 불가능한 반사는 없다. 다만 상체보다 넓게 허벅지까지 담았다.",
        "entities": "보이는 사람은 이관회 한 명으로, 청년 남성의 얼굴과 검은 가르마 머리, 거친 갈색 저고리와 밝은 바지가 참조와 대체로 일치한다. 눈은 정상적인 홍채와 동공을 유지한 채 크게 떠 있다. 칼의 감긴 손잡이와 금속 코등이는 소품 참조에 가깝고, 노출된 날이 복부에 꽂혀 있다. 화면 오른쪽 옆구리에는 혈흔이 있고 오른손에는 몽둥이 일부가 보인다. 읽을 수 있는 글자는 없다.",
        "hard_violations": [],
        "physics": "몸통은 세워져 있으며 하체가 화면 아래로 이어진다. 발은 구도 밖이므로 지면 접촉은 직접 확인할 수 없지만 공중에 떠 있는 자세는 아니다. 몽둥이는 오른손이 쥐고 있고, 칼은 복부에 삽입된 부분이 지지하는 배치다. 바구니와 떨어진 무기는 지면이나 진열대에 놓여 있다. 다만 양팔을 거의 곧게 내린 대칭적 자세여서 전투 동작 도중 굳어버린 순간보다는 차렷 자세에 가깝다."
       },
       {
        "label": "A",
        "direction": "이관회의 시선은 카메라 왼쪽의 화면 밖을 향하며 눈이 크게 떠 있다. 검 손잡이는 화면 왼쪽 위로, 검은 칼집은 오른쪽 아래로 뻗는다. 검 전체가 허리 앞을 가로지르는 방향으로 보여 칼날이 복부 안으로 향하는 삽입 관계가 확인되지 않는다. 오른손에 쥔 몽둥이의 굵은 끝은 화면 왼쪽으로 향한다.",
        "built_space": "흙길 왼쪽에 천을 덮은 판매대 하나와 낮은 목조 진열대, 바구니들이 있고, 오른쪽에는 기대어 놓은 장총류와 점포 구조물이 보인다. 길 뒤쪽에는 포대 방어벽 하나가 있으며 그 너머로 최소 세 명의 인물이 드러난다. 낡은 시장의 재료와 낮 분위기는 이어지지만 이전 사진의 짚신 진열 구역과 정확히 같은 위치인지는 확인하기 어렵다. 인물은 허리 부근까지 크게 담겨 A보다 상체 중심 구도에 가깝다.",
        "entities": "주인공의 청년 남성 외형, 검은 머리, 얼굴과 해진 갈색 저고리는 참조와 대체로 일치한다. 정상적인 눈을 크게 뜨고 입을 벌린 충격 표현이 보인다. 몽둥이는 오른손에 분명히 들려 있다. 검은 참조의 감긴 손잡이와 장식 있는 검은 칼집을 닮았지만 칼집에서 뽑혀 복부에 박힌 상태가 아니다. 옆구리의 찢어진 천과 상처, 아래로 번진 피는 A보다 강하게 강조된다. 배경 인물들은 이관회 외에는 누구도 등장시키지 말라는 조건에 어긋난다.",
        "hard_violations": [
         "방어벽 뒤에 최소 세 명의 추가 인물이 보여, 이관회 한 명만 등장해야 한다는 명시적 조건을 위반한다."
        ],
        "physics": "몸통을 약간 틀고 양팔을 벌린 자세는 동작 중 멈춘 순간으로 가능하며, 하체와 발은 화면 밖이다. 몽둥이는 오른손의 손가락이 감싸 지지한다. 칼집은 허리 앞에 밀착되어 띠와 겹치므로 떠 있다고 단정할 수는 없지만, 복부에 박힌 칼날이 검을 지지하는 요청된 상태는 보이지 않는다. 바구니와 판재는 바닥이나 판매대에 놓였고 장총류는 점포 구조물에 기대어 있다."
       }
      ]
     },
     "ok": true
    }
   ],
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "slot_winner_match": true,
   "slot_winner": {
    "gemini-pro": "B",
    "gpt-high": "B"
   },
   "route": "cross_slot_agree"
  },
  "dual": {
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "normalized": {
    "A": 0.486,
    "B": 2.0
   },
   "adjusted": {
    "A": 0.236,
    "B": 2.0
   },
   "violations": {
    "A": [
     "[gemini-pro] 배에 박혀야 할 칼이 칼집에 들어있는 채로 복부 앞에 물리적 지지대 없이 허공에 떠 있음 (물리적 오류).",
     "[gemini-pro] 참조 이미지에 존재하지 않는 인물들과 모래주머니 진지가 배경에 추가됨 (장소 설정 위반 및 발명된 사람/오브젝트).",
     "[gpt-high] 방어벽 뒤에 최소 세 명의 추가 인물이 보여, 이관회 한 명만 등장해야 한다는 명시적 조건을 위반한다."
    ]
   },
   "per_model_winner": {
    "gemini-pro": "B",
    "gpt-high": "B"
   },
   "agreed": true
  },
  "totals": {
   "B": 2000,
   "A": 236
  },
  "selected": "B",
  "ranking": [
   "B",
   "A"
  ],
  "verdicts": [
   {
    "label": "B",
    "score": 2000,
    "verdict_ko": "크게 뜬 눈과 뻣뻣하게 굳은 자세, 배에 비스듬히 박힌 칼, 그리고 이전 샷과 완벽하게 일치하는 배경까지 프롬프트의 모든 지시사항을 훌륭하게 구현했습니다."
   },
   {
    "label": "A",
    "score": 236,
    "verdict_ko": "칼이 배에 박혀 있지 않고 칼집째로 물리적 지지 없이 떠 있으며, 이전 샷에 없던 구조물과 인물이 배경에 추가되는 치명적인 오류를 범했습니다.  ★위반: [gemini-pro] 배에 박혀야 할 칼이 칼집에 들어있는 채로 복부 앞에 물리적 지지대 없이 허공에 떠 있음 (물리적 오류). / [gemini-pro] 참조 이미지에 존재하지 않는 인물들과 모래주머니 진지가 배경에 추가됨 (장소 설정 위반 및 발명된 사람/오브젝트). / [gpt-high] 방어벽 뒤에 최소 세 명의 추가 인물이 보여, 이관회 한 명만 등장해야 한다는 명시적 조건을 위반한다."
   }
  ],
  "refs": [
   {
    "label": "PREVIOUS SHOT STILL — a visually related earlier shot of this same place: the location's look, materials, fixed features and lighting mood are LOCKED to this photo; never copy its camera framing. The clothing, hair and overall look of 이관회 — who appear both in that photo and in this shot — are LOCKED to that photo. Anyone else visible in that photo is NOT in this shot: never carry their face, body or clothing onto anyone here. If this photo shows a character who cannot move (dead or unconscious), that character's exact pose, position and orientation are ALSO LOCKED — treat their body as a fixed prop of the set that only the camera moves around.",
    "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/db00517b-165b-4005-a287-f28936b59b95/images/272bd7af-9294-4490-98c6-88a40e9deacc/scene/recipe/S53sh2_sel.png",
    "asset_id": "589ffa1a-6307-4d54-b4ad-2c769f29f91b",
    "role": "prev_still"
   },
   {
    "label": "CHARACTER REFERENCE — 이관회: the exact person appearing in this shot; match face, hair and build exactly.",
    "path": "<bytes:1150504>",
    "asset_id": "b3980731-4afc-4e00-bc4c-07458f899a53",
    "role": "character_ref"
   },
   {
    "label": "PROP REFERENCE — 칼: the exact object appearing in this shot; match its look, material and wear exactly.",
    "path": "<bytes:780122>",
    "asset_id": "82614014-6091-423c-89ca-1e15d209de72",
    "role": "prop_ref"
   }
  ],
  "critique_skipped": true,
  "shot_run_produced": false,
  "shot_run_uid": "06aa1437-95e9-7531-8273-f88cf236dd6c",
  "ref_mode": "prev+엔티티",
  "share_plan": {
   "ref_plan": "prev",
   "prev_anchor_tag": "S53sh2"
  }
 },
 "S55sh11::cine": {
  "applied": true,
  "attempted_at": "2026-09-09T07:15:13.372614+00:00",
  "fingerprint": "90a668141229beb21c872d25517fa24514b22dd6fda781c65326908f1fa0154a",
  "fingerprint_version": 2,
  "provider": "mai",
  "endpoint": "openrouter/chat-completions",
  "model": "microsoft/mai-image-2.6",
  "pack": "24.202608252115",
  "source_file": "S55sh11_sel.png",
  "source_sha256": "0c8b0ab34fdd1c863067aca48b8413e930d840d931ef713163987c94277675ea",
  "file": "S55sh11_cine.png",
  "staged_sha256": "262f23134821f0f4680f3ccacb4eb026f3d45e3603a31340a66a1db56278b409",
  "latency_ms": 26436
 },
 "S55sh34::signage": {
  "fp": "7b7bbedf028659d3",
  "inscriptions": [],
  "cues": [],
  "dropped": []
 },
 "S55sh34": {
  "input_fingerprint": "7cdb85aae71885d2",
  "prompt": "Create ONE FINAL photorealistic live-action film still of the moment below — early 20th-century Korea, Jeolla region; all people are Korean unless explicitly stated otherwise. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 숨을 거둔 이관회를 끌어안은 채 미동도 없이 굳어버린 안규홍의 슬픔에 잠긴 측면.\n\nLOCATION (lock): Ground level in the market's central dirt thoroughfare, amid the fallen attackers near the stalls. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: medium shot\n- KEY BACKGROUND ELEMENTS: Market earth floor (The two men are seated on it); used as Provides a narrow grounding strip beneath their bodies, retaining the physical weight of the collapse without widening away from their faces.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Keep the established daylight and subdued color continuous, with naturalistic contrast that preserves human warmth while giving the death sober weight.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The clothing, hair and overall look of 이관회 — who appear both in that photo and in this shot — are LOCKED to that photo. Anyone else visible in that photo is NOT in this shot: never carry their face, body or clothing onto anyone here. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nIMMOBILE CHARACTER POSE — CANONICAL (identical wherever this character appears in ANY panel; on any conflict THIS POSE WINS): His dead body is held against An Gyuhong, who is seated on the market's dirt ground and supports his collapsed upper body; his eyes remain open and unfocused, and his hand has slipped from Gyuhong's collar. The screenplay does not specify his head angle, where the released hand comes to rest, the position of his other arm, or the arrangement of his legs.\n\nIMMOBILE BODIES OBEY GRAVITY: a person who is dead or unconscious\nexerts NO muscular effort. Every part of their body — head, torso,\narms, hands, fingers, legs — rests fully on whatever supports it\n(floor, wall, furniture, their own lap) and hangs or slumps with\ngravity. NEVER show any part of an immobile person's body lifted,\nraised, held up in the air, or posed as if presenting something:\nan object in their grip stays clenched in a hand that itself lies\nfallen on a support — the hand does not hold the object up. If the\ncanonical pose leaves a body part unspecified, resolve it as the\nmost gravity-compliant, fully-supported position.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The bloodied sickle has been set down within reach on the dirt; the market fighting area contains fallen weapons and broken stall debris. 이관회: He is dead with open, unfocused eyes, an abdominal stab wound and a bleeding flank slash; his hand has fallen limp. 안규홍: He sits on the dirt in a cradling posture, his hands soaked in blood and his own back wound bleeding through his jeogori.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 안규홍 (한국인 남성, 청년 나이대, 검은 머리, 단단히 묶은 머리); 이관회 (한국인 남성, 청년 나이대, 자연스러운 검은 머리) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.",
  "roll_prompts": {
   "A": "Create ONE FINAL photorealistic live-action film still of the moment below — early 20th-century Korea, Jeolla region; all people are Korean unless explicitly stated otherwise. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 숨을 거둔 이관회를 끌어안은 채 미동도 없이 굳어버린 안규홍의 슬픔에 잠긴 측면.\n\nLOCATION (lock): Ground level in the market's central dirt thoroughfare, amid the fallen attackers near the stalls. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: medium shot\n- KEY BACKGROUND ELEMENTS: Market earth floor (The two men are seated on it); used as Provides a narrow grounding strip beneath their bodies, retaining the physical weight of the collapse without widening away from their faces.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Keep the established daylight and subdued color continuous, with naturalistic contrast that preserves human warmth while giving the death sober weight.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The clothing, hair and overall look of 이관회 — who appear both in that photo and in this shot — are LOCKED to that photo. Anyone else visible in that photo is NOT in this shot: never carry their face, body or clothing onto anyone here. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nIMMOBILE CHARACTER POSE — CANONICAL (identical wherever this character appears in ANY panel; on any conflict THIS POSE WINS): His dead body is held against An Gyuhong, who is seated on the market's dirt ground and supports his collapsed upper body; his eyes remain open and unfocused, and his hand has slipped from Gyuhong's collar. The screenplay does not specify his head angle, where the released hand comes to rest, the position of his other arm, or the arrangement of his legs.\n\nIMMOBILE BODIES OBEY GRAVITY: a person who is dead or unconscious\nexerts NO muscular effort. Every part of their body — head, torso,\narms, hands, fingers, legs — rests fully on whatever supports it\n(floor, wall, furniture, their own lap) and hangs or slumps with\ngravity. NEVER show any part of an immobile person's body lifted,\nraised, held up in the air, or posed as if presenting something:\nan object in their grip stays clenched in a hand that itself lies\nfallen on a support — the hand does not hold the object up. If the\ncanonical pose leaves a body part unspecified, resolve it as the\nmost gravity-compliant, fully-supported position.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The bloodied sickle has been set down within reach on the dirt; the market fighting area contains fallen weapons and broken stall debris. 이관회: He is dead with open, unfocused eyes, an abdominal stab wound and a bleeding flank slash; his hand has fallen limp. 안규홍: He sits on the dirt in a cradling posture, his hands soaked in blood and his own back wound bleeding through his jeogori.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 안규홍 (한국인 남성, 청년 나이대, 검은 머리, 단단히 묶은 머리); 이관회 (한국인 남성, 청년 나이대, 자연스러운 검은 머리) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.",
   "B": "Create ONE FINAL photorealistic live-action film still of the moment below — early 20th-century Korea, Jeolla region; all people are Korean unless explicitly stated otherwise. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 숨을 거둔 이관회를 끌어안은 채 미동도 없이 굳어버린 안규홍의 슬픔에 잠긴 측면.\n\nLOCATION (lock): Ground level in the market's central dirt thoroughfare, amid the fallen attackers near the stalls. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: medium shot\n- KEY BACKGROUND ELEMENTS: Market earth floor (The two men are seated on it); used as Provides a narrow grounding strip beneath their bodies, retaining the physical weight of the collapse without widening away from their faces.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Keep the established daylight and subdued color continuous, with naturalistic contrast that preserves human warmth while giving the death sober weight.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The clothing, hair and overall look of 이관회 — who appear both in that photo and in this shot — are LOCKED to that photo. Anyone else visible in that photo is NOT in this shot: never carry their face, body or clothing onto anyone here. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nIMMOBILE CHARACTER POSE — CANONICAL (identical wherever this character appears in ANY panel; on any conflict THIS POSE WINS): His dead body is held against An Gyuhong, who is seated on the market's dirt ground and supports his collapsed upper body; his eyes remain open and unfocused, and his hand has slipped from Gyuhong's collar. The screenplay does not specify his head angle, where the released hand comes to rest, the position of his other arm, or the arrangement of his legs.\n\nIMMOBILE BODIES OBEY GRAVITY: a person who is dead or unconscious\nexerts NO muscular effort. Every part of their body — head, torso,\narms, hands, fingers, legs — rests fully on whatever supports it\n(floor, wall, furniture, their own lap) and hangs or slumps with\ngravity. NEVER show any part of an immobile person's body lifted,\nraised, held up in the air, or posed as if presenting something:\nan object in their grip stays clenched in a hand that itself lies\nfallen on a support — the hand does not hold the object up. If the\ncanonical pose leaves a body part unspecified, resolve it as the\nmost gravity-compliant, fully-supported position.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The bloodied sickle has been set down within reach on the dirt; the market fighting area contains fallen weapons and broken stall debris. 이관회: He is dead with open, unfocused eyes, an abdominal stab wound and a bleeding flank slash; his hand has fallen limp. 안규홍: He sits on the dirt in a cradling posture, his hands soaked in blood and his own back wound bleeding through his jeogori.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 안규홍 (한국인 남성, 청년 나이대, 검은 머리, 단단히 묶은 머리); 이관회 (한국인 남성, 청년 나이대, 자연스러운 검은 머리) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.\n\nCOMPOSITION VARIATION (this candidate only): keep the place, the people and the moment exactly as specified, and keep from the CAMERA & FRAME contract its subject scale, screen placement and key background elements exactly. For this candidate only, its camera angle and height are deliberately left open: choose a camera position distinctly different from the obvious one, as a film director picking a second setup on the same blocking. Stay true to any explicit framing the SHOT TEXT itself dictates."
  },
  "shot_run_spend_attempt_count": 1,
  "readings": [
   {
    "label": "A",
    "direction": "안아주는 인물은 화면 오른쪽 먼 곳을 응시하고, 죽은 인물은 눈을 뜬 채 초점 없이 위쪽을 향하고 있음 (텍스트의 'open, unfocused eyes' 지시 충족).",
    "built_space": "시장 흙바닥. 양옆으로 목조 난전 구조물이 있으며, 왼쪽 기둥에는 짚신이 걸려 있음. 배경에는 쓰러진 공격자들의 시신 2구가 배치되어 장소 묘사와 완벽히 일치함.",
    "entities": "안아주는 인물: 짧은 머리에 갈색 저고리 착용 (안규홍의 묶은 머리와 흰옷 지시를 실패하고 이관회와 유사함). 죽은 인물: 짧은 머리에 베이지색 저고리 착용. 피 묻은 낫: 왼쪽 바닥에 놓여 있음. 배경의 시신들: 명확히 존재함.",
    "hard_violations": [
     "[gpt-high] 두 주인공 외에는 사람을 보이지 말라는 명시적 제한과 달리, 배경 통로에 추가 시신 두 명을 배치했다."
    ],
    "physics": "안아주는 인물이 흙바닥에 앉아 죽은 인물을 지탱함. 죽은 인물의 체중은 안아주는 인물의 무릎과 가슴에 온전히 실려 있으며, 두 손 역시 가슴과 다리 위에 자연스럽게 얹혀 중력을 따름. 공중에 뜬 요소 없이 낫도 바닥에 안정적으로 놓여 있음."
   },
   {
    "label": "B",
    "direction": "안아주는 인물은 죽은 인물의 얼굴을 내려다보고 있으며, 죽은 인물은 두 눈을 감고 있음 ('open, unfocused eyes' 지시 정면 위반).",
    "built_space": "시장 흙바닥. 양옆으로 목조 구조물과 바구니 등이 보이나, 배경에 텍스트가 요구한 '쓰러진 공격자들'의 모습은 명확하게 보이지 않음.",
    "entities": "안아주는 인물: 짧은 머리에 흰색 저고리 착용 (안규홍의 묶은 머리 지시 실패). 죽은 인물: 뒤로 묶은 머리에 흰색 저고리 착용 (이관회의 짧은 머리와 갈색 의상 지시 실패, 오히려 안규홍의 특징을 가짐). 피 묻은 낫: 바닥에 놓여 있음.",
    "hard_violations": [],
    "physics": "두 인물 모두 흙바닥에 앉아 서로 기대고 있으며, 죽은 인물의 손과 팔은 본인의 신체 위에 얹혀 중력을 따름. 낫을 비롯해 부자연스럽게 허공에 떠 있는 요소는 없음."
   }
  ],
  "cross_model_order": {
   "policy": "cross_model_order_v2_slot0_fwd_slot1_rev_parallel2",
   "slots": [
    {
     "model": "gemini-pro",
     "order": "forward",
     "display_to_canonical": {
      "A": "A",
      "B": "B"
     },
     "raw": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 6,
        "verdict_ko": "두 인물의 외형 특징(헤어스타일 및 의상)이 참조 이미지와 어긋나는 아쉬움이 있으나, 죽은 인물의 '초점 없이 뜬 눈', 배경의 쓰러진 자들, 등의 상처 등 텍스트의 핵심 연출 지시를 충실히 따랐습니다."
       },
       {
        "label": "B",
        "score": 4,
        "verdict_ko": "인물들의 외형과 역할이 뒤바뀌었을 뿐만 아니라, 죽은 인물이 눈을 뜨고 있어야 한다는 명시적 지시를 위반하고 눈을 감고 있어 연출 의도에서 크게 벗어납니다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "안아주는 인물은 화면 오른쪽 먼 곳을 응시하고, 죽은 인물은 눈을 뜬 채 초점 없이 위쪽을 향하고 있음 (텍스트의 'open, unfocused eyes' 지시 충족).",
        "built_space": "시장 흙바닥. 양옆으로 목조 난전 구조물이 있으며, 왼쪽 기둥에는 짚신이 걸려 있음. 배경에는 쓰러진 공격자들의 시신 2구가 배치되어 장소 묘사와 완벽히 일치함.",
        "entities": "안아주는 인물: 짧은 머리에 갈색 저고리 착용 (안규홍의 묶은 머리와 흰옷 지시를 실패하고 이관회와 유사함). 죽은 인물: 짧은 머리에 베이지색 저고리 착용. 피 묻은 낫: 왼쪽 바닥에 놓여 있음. 배경의 시신들: 명확히 존재함.",
        "hard_violations": [],
        "physics": "안아주는 인물이 흙바닥에 앉아 죽은 인물을 지탱함. 죽은 인물의 체중은 안아주는 인물의 무릎과 가슴에 온전히 실려 있으며, 두 손 역시 가슴과 다리 위에 자연스럽게 얹혀 중력을 따름. 공중에 뜬 요소 없이 낫도 바닥에 안정적으로 놓여 있음."
       },
       {
        "label": "B",
        "direction": "안아주는 인물은 죽은 인물의 얼굴을 내려다보고 있으며, 죽은 인물은 두 눈을 감고 있음 ('open, unfocused eyes' 지시 정면 위반).",
        "built_space": "시장 흙바닥. 양옆으로 목조 구조물과 바구니 등이 보이나, 배경에 텍스트가 요구한 '쓰러진 공격자들'의 모습은 명확하게 보이지 않음.",
        "entities": "안아주는 인물: 짧은 머리에 흰색 저고리 착용 (안규홍의 묶은 머리 지시 실패). 죽은 인물: 뒤로 묶은 머리에 흰색 저고리 착용 (이관회의 짧은 머리와 갈색 의상 지시 실패, 오히려 안규홍의 특징을 가짐). 피 묻은 낫: 바닥에 놓여 있음.",
        "hard_violations": [],
        "physics": "두 인물 모두 흙바닥에 앉아 서로 기대고 있으며, 죽은 인물의 손과 팔은 본인의 신체 위에 얹혀 중력을 따름. 낫을 비롯해 부자연스럽게 허공에 떠 있는 요소는 없음."
       }
      ],
      "all_candidates_fail": false
     },
     "normalized": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 6,
        "verdict_ko": "두 인물의 외형 특징(헤어스타일 및 의상)이 참조 이미지와 어긋나는 아쉬움이 있으나, 죽은 인물의 '초점 없이 뜬 눈', 배경의 쓰러진 자들, 등의 상처 등 텍스트의 핵심 연출 지시를 충실히 따랐습니다."
       },
       {
        "label": "B",
        "score": 4,
        "verdict_ko": "인물들의 외형과 역할이 뒤바뀌었을 뿐만 아니라, 죽은 인물이 눈을 뜨고 있어야 한다는 명시적 지시를 위반하고 눈을 감고 있어 연출 의도에서 크게 벗어납니다."
       }
      ],
      "all_candidates_fail": false,
      "readings": [
       {
        "label": "A",
        "direction": "안아주는 인물은 화면 오른쪽 먼 곳을 응시하고, 죽은 인물은 눈을 뜬 채 초점 없이 위쪽을 향하고 있음 (텍스트의 'open, unfocused eyes' 지시 충족).",
        "built_space": "시장 흙바닥. 양옆으로 목조 난전 구조물이 있으며, 왼쪽 기둥에는 짚신이 걸려 있음. 배경에는 쓰러진 공격자들의 시신 2구가 배치되어 장소 묘사와 완벽히 일치함.",
        "entities": "안아주는 인물: 짧은 머리에 갈색 저고리 착용 (안규홍의 묶은 머리와 흰옷 지시를 실패하고 이관회와 유사함). 죽은 인물: 짧은 머리에 베이지색 저고리 착용. 피 묻은 낫: 왼쪽 바닥에 놓여 있음. 배경의 시신들: 명확히 존재함.",
        "hard_violations": [],
        "physics": "안아주는 인물이 흙바닥에 앉아 죽은 인물을 지탱함. 죽은 인물의 체중은 안아주는 인물의 무릎과 가슴에 온전히 실려 있으며, 두 손 역시 가슴과 다리 위에 자연스럽게 얹혀 중력을 따름. 공중에 뜬 요소 없이 낫도 바닥에 안정적으로 놓여 있음."
       },
       {
        "label": "B",
        "direction": "안아주는 인물은 죽은 인물의 얼굴을 내려다보고 있으며, 죽은 인물은 두 눈을 감고 있음 ('open, unfocused eyes' 지시 정면 위반).",
        "built_space": "시장 흙바닥. 양옆으로 목조 구조물과 바구니 등이 보이나, 배경에 텍스트가 요구한 '쓰러진 공격자들'의 모습은 명확하게 보이지 않음.",
        "entities": "안아주는 인물: 짧은 머리에 흰색 저고리 착용 (안규홍의 묶은 머리 지시 실패). 죽은 인물: 뒤로 묶은 머리에 흰색 저고리 착용 (이관회의 짧은 머리와 갈색 의상 지시 실패, 오히려 안규홍의 특징을 가짐). 피 묻은 낫: 바닥에 놓여 있음.",
        "hard_violations": [],
        "physics": "두 인물 모두 흙바닥에 앉아 서로 기대고 있으며, 죽은 인물의 손과 팔은 본인의 신체 위에 얹혀 중력을 따름. 낫을 비롯해 부자연스럽게 허공에 떠 있는 요소는 없음."
       }
      ]
     },
     "ok": true
    },
    {
     "model": "gpt-high",
     "order": "reverse",
     "display_to_canonical": {
      "A": "B",
      "B": "A"
     },
     "raw": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 4,
        "verdict_ko": "두 사람만의 정지된 포옹과 신체 지지는 성립하지만, 안규홍과 이관회의 얼굴·머리 배치가 뒤바뀌고 이관회의 갈색 저고리도 사라졌다."
       },
       {
        "label": "B",
        "score": 2,
        "verdict_ko": "측면 포옹은 보이나 배경에 금지된 추가 시신들이 등장하며, 안규홍에게 이관회의 외모와 갈색 옷을 부여해 인물 연속성도 어긋난다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "안고 있는 오른쪽 남자는 왼쪽 아래, 품에 안긴 남자의 머리 쪽을 내려다본다. 안긴 남자는 눈을 뜬 채 아래쪽을 멍하게 보며 카메라와 눈을 맞추지 않는다. 겨누거나 이동하는 물체는 없고, 낫은 왼쪽 흙바닥에 누워 있다.",
        "built_space": "두 사람은 목조 좌판 앞 흙길에 앉아 있다. 왼쪽의 긴 좌판 하나, 중앙 뒤의 천 덮인 판매대 하나, 오른쪽 가장자리의 나무 상자 하나가 보이며, 벽에는 짚신과 짚으로 만든 물건들이 걸려 있다. 낡은 판벽·천막·바구니·부서진 목재는 장소 참고와 대체로 이어진다. 다만 바닥이 하단의 좁은 접지 띠에 그치지 않고 좌우와 전경에 넓게 드러나 얼굴 중심의 지정 구도보다 공간 비중이 크다.",
        "entities": "보이는 인물은 청년 동아시아계 남성 두 명뿐이다. 그러나 안규홍 역할로 안고 있는 남자는 이관회 참고의 짧게 가른 검은 머리와 얼굴에 가깝고, 죽은 이관회 역할의 남자는 안규홍 참고의 뒤로 단단히 넘긴 머리와 얼굴에 가깝다. 둘 다 밝은 저고리여서 이관회에게 고정된 거친 갈색 저고리가 맞지 않는다. 안긴 남자의 눈은 열려 있고 손은 옷깃에서 떨어져 있다. 안는 남자의 손과 등에는 피가 보인다. 피 묻은 낫 한 자루, 떨어진 장검과 목재 잔해가 보이며 읽을 수 있는 글자는 없다.",
        "hard_violations": [],
        "physics": "안는 남자는 흙바닥에 엉덩이와 접은 다리를 두고 앉아 있다. 그의 두 팔과 손이 상대의 어깨와 위팔을 감싸며 상체를 받친다. 안긴 남자의 머리는 상대의 가슴과 목 아래에 기대고, 내려놓은 손들은 무릎 위에 포개져 지지된다. 보이는 신체에 무지지 부유는 없으며 낫과 잔해도 바닥에 닿아 있다."
       },
       {
        "label": "B",
        "direction": "안고 있는 왼쪽 남자는 화면 오른쪽의 먼 곳을 향해 옆얼굴을 보인다. 안긴 남자는 눈을 뜨고 위쪽 허공을 바라보며, 서로 시선을 맞추지 않는다. 허공을 보는 눈은 보이지만 초점이 완전히 풀린 죽음의 연기는 다소 약하다. 낫과 장검은 바닥에 놓였고 겨누는 동작은 없다.",
        "built_space": "두 사람은 중앙 흙길에 앉아 있으며, 왼쪽 긴 목조 좌판 하나와 그 뒤 천 덮인 판매대 하나, 오른쪽 큰 나무 상자 하나 및 비스듬히 넘어진 판재가 보인다. 바구니와 낡은 목재의 재질은 참고 시장과 유사하다. 그러나 화면 위쪽 통로에 추가로 쓰러진 사람 두 명이 보인다. 바닥과 통로가 넓게 포함되어 지정된 좁은 접지 띠보다 배경 설명에 많은 면적을 할애한다.",
        "entities": "전경의 두 남자는 청년 동아시아계 외모이며, 안긴 이관회는 자연스러운 검은 머리와 거친 갈색 저고리, 밝은 바지 및 복부·옆구리 혈흔을 갖춘다. 반면 안규홍 역할의 남자도 이관회 참고에 가까운 짧게 가른 머리와 갈색 저고리여서 안규홍의 묶은 머리와 밝은 복장이 맞지 않는다. 안규홍 역할의 손과 등에 피가 있고, 피 묻은 낫 한 자루와 떨어진 무기 및 좌판 잔해가 보인다. 배경의 추가 시신 두 명은 허용된 등장인물이 아니다. 읽을 수 있는 글자는 없다.",
        "hard_violations": [
         "두 주인공 외에는 사람을 보이지 말라는 명시적 제한과 달리, 배경 통로에 추가 시신 두 명을 배치했다."
        ],
        "physics": "안는 남자의 엉덩이와 접은 다리는 흙바닥에 놓인다. 이관회의 몸통은 안규홍의 품과 무릎에 기대며, 머리 옆과 뒤는 상대의 가슴·어깨 부근에 닿아 지지되는 것으로 보인다. 이관회의 양손은 무릎과 허벅지 위에 내려와 있어 떠 있지 않다. 낫과 무기, 배경 시신들도 바닥에 놓여 있으며 명백히 지지 없이 공중에 뜬 물체는 없다."
       }
      ],
      "all_candidates_fail": true
     },
     "normalized": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 4,
        "verdict_ko": "두 사람만의 정지된 포옹과 신체 지지는 성립하지만, 안규홍과 이관회의 얼굴·머리 배치가 뒤바뀌고 이관회의 갈색 저고리도 사라졌다."
       },
       {
        "label": "A",
        "score": 2,
        "verdict_ko": "측면 포옹은 보이나 배경에 금지된 추가 시신들이 등장하며, 안규홍에게 이관회의 외모와 갈색 옷을 부여해 인물 연속성도 어긋난다."
       }
      ],
      "all_candidates_fail": true,
      "readings": [
       {
        "label": "B",
        "direction": "안고 있는 오른쪽 남자는 왼쪽 아래, 품에 안긴 남자의 머리 쪽을 내려다본다. 안긴 남자는 눈을 뜬 채 아래쪽을 멍하게 보며 카메라와 눈을 맞추지 않는다. 겨누거나 이동하는 물체는 없고, 낫은 왼쪽 흙바닥에 누워 있다.",
        "built_space": "두 사람은 목조 좌판 앞 흙길에 앉아 있다. 왼쪽의 긴 좌판 하나, 중앙 뒤의 천 덮인 판매대 하나, 오른쪽 가장자리의 나무 상자 하나가 보이며, 벽에는 짚신과 짚으로 만든 물건들이 걸려 있다. 낡은 판벽·천막·바구니·부서진 목재는 장소 참고와 대체로 이어진다. 다만 바닥이 하단의 좁은 접지 띠에 그치지 않고 좌우와 전경에 넓게 드러나 얼굴 중심의 지정 구도보다 공간 비중이 크다.",
        "entities": "보이는 인물은 청년 동아시아계 남성 두 명뿐이다. 그러나 안규홍 역할로 안고 있는 남자는 이관회 참고의 짧게 가른 검은 머리와 얼굴에 가깝고, 죽은 이관회 역할의 남자는 안규홍 참고의 뒤로 단단히 넘긴 머리와 얼굴에 가깝다. 둘 다 밝은 저고리여서 이관회에게 고정된 거친 갈색 저고리가 맞지 않는다. 안긴 남자의 눈은 열려 있고 손은 옷깃에서 떨어져 있다. 안는 남자의 손과 등에는 피가 보인다. 피 묻은 낫 한 자루, 떨어진 장검과 목재 잔해가 보이며 읽을 수 있는 글자는 없다.",
        "hard_violations": [],
        "physics": "안는 남자는 흙바닥에 엉덩이와 접은 다리를 두고 앉아 있다. 그의 두 팔과 손이 상대의 어깨와 위팔을 감싸며 상체를 받친다. 안긴 남자의 머리는 상대의 가슴과 목 아래에 기대고, 내려놓은 손들은 무릎 위에 포개져 지지된다. 보이는 신체에 무지지 부유는 없으며 낫과 잔해도 바닥에 닿아 있다."
       },
       {
        "label": "A",
        "direction": "안고 있는 왼쪽 남자는 화면 오른쪽의 먼 곳을 향해 옆얼굴을 보인다. 안긴 남자는 눈을 뜨고 위쪽 허공을 바라보며, 서로 시선을 맞추지 않는다. 허공을 보는 눈은 보이지만 초점이 완전히 풀린 죽음의 연기는 다소 약하다. 낫과 장검은 바닥에 놓였고 겨누는 동작은 없다.",
        "built_space": "두 사람은 중앙 흙길에 앉아 있으며, 왼쪽 긴 목조 좌판 하나와 그 뒤 천 덮인 판매대 하나, 오른쪽 큰 나무 상자 하나 및 비스듬히 넘어진 판재가 보인다. 바구니와 낡은 목재의 재질은 참고 시장과 유사하다. 그러나 화면 위쪽 통로에 추가로 쓰러진 사람 두 명이 보인다. 바닥과 통로가 넓게 포함되어 지정된 좁은 접지 띠보다 배경 설명에 많은 면적을 할애한다.",
        "entities": "전경의 두 남자는 청년 동아시아계 외모이며, 안긴 이관회는 자연스러운 검은 머리와 거친 갈색 저고리, 밝은 바지 및 복부·옆구리 혈흔을 갖춘다. 반면 안규홍 역할의 남자도 이관회 참고에 가까운 짧게 가른 머리와 갈색 저고리여서 안규홍의 묶은 머리와 밝은 복장이 맞지 않는다. 안규홍 역할의 손과 등에 피가 있고, 피 묻은 낫 한 자루와 떨어진 무기 및 좌판 잔해가 보인다. 배경의 추가 시신 두 명은 허용된 등장인물이 아니다. 읽을 수 있는 글자는 없다.",
        "hard_violations": [
         "두 주인공 외에는 사람을 보이지 말라는 명시적 제한과 달리, 배경 통로에 추가 시신 두 명을 배치했다."
        ],
        "physics": "안는 남자의 엉덩이와 접은 다리는 흙바닥에 놓인다. 이관회의 몸통은 안규홍의 품과 무릎에 기대며, 머리 옆과 뒤는 상대의 가슴·어깨 부근에 닿아 지지되는 것으로 보인다. 이관회의 양손은 무릎과 허벅지 위에 내려와 있어 떠 있지 않다. 낫과 무기, 배경 시신들도 바닥에 놓여 있으며 명백히 지지 없이 공중에 뜬 물체는 없다."
       }
      ]
     },
     "ok": true
    }
   ],
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "slot_winner_match": false,
   "slot_winner": {
    "gemini-pro": "A",
    "gpt-high": "B"
   },
   "route": "cross_slot_combined"
  },
  "dual": {
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "normalized": {
    "A": 1.5,
    "B": 1.667
   },
   "adjusted": {
    "A": 1.25,
    "B": 1.667
   },
   "violations": {
    "A": [
     "[gpt-high] 두 주인공 외에는 사람을 보이지 말라는 명시적 제한과 달리, 배경 통로에 추가 시신 두 명을 배치했다."
    ]
   },
   "per_model_winner": {
    "gemini-pro": "A",
    "gpt-high": "B"
   },
   "agreed": false
  },
  "totals": {
   "A": 1250,
   "B": 1667
  },
  "selected": "B",
  "ranking": [
   "B",
   "A"
  ],
  "verdicts": [
   {
    "label": "A",
    "score": 1250,
    "verdict_ko": "두 인물의 외형 특징(헤어스타일 및 의상)이 참조 이미지와 어긋나는 아쉬움이 있으나, 죽은 인물의 '초점 없이 뜬 눈', 배경의 쓰러진 자들, 등의 상처 등 텍스트의 핵심 연출 지시를 충실히 따랐습니다.  ★위반: [gpt-high] 두 주인공 외에는 사람을 보이지 말라는 명시적 제한과 달리, 배경 통로에 추가 시신 두 명을 배치했다."
   },
   {
    "label": "B",
    "score": 1667,
    "verdict_ko": "인물들의 외형과 역할이 뒤바뀌었을 뿐만 아니라, 죽은 인물이 눈을 뜨고 있어야 한다는 명시적 지시를 위반하고 눈을 감고 있어 연출 의도에서 크게 벗어납니다."
   }
  ],
  "refs": [
   {
    "label": "PREVIOUS SHOT STILL — a visually related earlier shot of this same place: the location's look, materials, fixed features and lighting mood are LOCKED to this photo; never copy its camera framing. The clothing, hair and overall look of 이관회 — who appear both in that photo and in this shot — are LOCKED to that photo. Anyone else visible in that photo is NOT in this shot: never carry their face, body or clothing onto anyone here. If this photo shows a character who cannot move (dead or unconscious), that character's exact pose, position and orientation are ALSO LOCKED — treat their body as a fixed prop of the set that only the camera moves around.",
    "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/db00517b-165b-4005-a287-f28936b59b95/images/272bd7af-9294-4490-98c6-88a40e9deacc/scene/recipe/S55sh11_sel.png",
    "asset_id": "58844d9a-2f38-431a-8669-3fb3fcabf2d4",
    "role": "prev_still"
   },
   {
    "label": "CHARACTER REFERENCE — 안규홍: the exact person appearing in this shot; match face, hair and build exactly.",
    "path": "<bytes:1028384>",
    "asset_id": "3b5894db-ce55-4f34-b966-5e725c3aa952",
    "role": "character_ref"
   },
   {
    "label": "CHARACTER REFERENCE — 이관회: the exact person appearing in this shot; match face, hair and build exactly.",
    "path": "<bytes:1358419>",
    "asset_id": "3c1163c2-5756-4932-b1bf-b080bec87144",
    "role": "character_ref"
   }
  ],
  "critique_skipped": true,
  "shot_run_produced": false,
  "shot_run_uid": "06aa1437-978d-7c01-9b17-d4883563f3de",
  "ref_mode": "prev+엔티티",
  "share_plan": {
   "ref_plan": "prev",
   "prev_anchor_tag": "S55sh11"
  }
 },
 "S55sh34::cine": {
  "applied": false,
  "attempted_at": "2026-09-09T09:32:43.464484+00:00",
  "fingerprint": "9b54c376166b4e240a1756c293918cdfa9e38a6f12e93cf0ac3b01a88d038def",
  "fingerprint_version": 2,
  "provider": "mai",
  "endpoint": "openrouter/chat-completions",
  "model": "microsoft/mai-image-2.6",
  "multiroll_tag": "still_S55sh34_cine",
  "slot": "",
  "pack": "24.202608252115",
  "source_file": "S55sh34_sel.png",
  "source_sha256": "a940cb9d728311a5720aff64b8e0ee656771b5a7c80b26e0b0e491695d007c80",
  "error": "RuntimeError: moderation blocked: {\"message\": \"Output content violated imagegen safety policies.\", \"code\": 400}",
  "moderation_refusals": 2,
  "declined": true,
  "declined_reason": "moderation"
 },
 "S55sh40::signage": {
  "fp": "d32b05af606e78d0",
  "inscriptions": [],
  "cues": [],
  "dropped": []
 },
 "S55sh40": {
  "input_fingerprint": "04258cb936eadbf3",
  "prompt": "Create ONE FINAL photorealistic live-action film still of the moment below — early 20th-century Korea, Jeolla region; all people are Korean unless explicitly stated otherwise. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 낫을 거머쥔 채 바닥을 밀어내며 반쯤 일어선 안규홍의 거대하고 위협적인 뒷모습.\n\nLOCATION (lock): The outdoor dirt space beside the fallen companion in the market lane, facing the fleeing man beyond the collapsed stalls. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: medium shot\n- KEY BACKGROUND ELEMENTS: 흙바닥 (Visible beneath and beside his rising body); used as A narrow lower-frame strip grounds the low viewpoint and establishes human scale; 피 묻은 낫 (Gripped in 안규홍's hand) — Seen obliquely beside his body, with the blade separated from his outline; used as A small lower-edge detail connects his rising posture to the impending pursuit.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Natural daylight with muted saturation and restrained contrast preserves the physical weight of grief turning into violence.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The clothing, hair and overall look of 안규홍 — who appear both in that photo and in this shot — are LOCKED to that photo. Anyone else visible in that photo is NOT in this shot: never carry their face, body or clothing onto anyone here. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nIMMOBILE CHARACTER POSE — CANONICAL (identical wherever this character appears in ANY panel; on any conflict THIS POSE WINS): His dead body has been lowered onto the market's dirt ground and remains motionless among the Iljinhoe corpses, no longer supported by An Gyuhong. His eyes remain open, but the screenplay does not specify his torso orientation, head angle, or the resting positions of his arms and legs.\n\nIMMOBILE BODIES OBEY GRAVITY: a person who is dead or unconscious\nexerts NO muscular effort. Every part of their body — head, torso,\narms, hands, fingers, legs — rests fully on whatever supports it\n(floor, wall, furniture, their own lap) and hangs or slumps with\ngravity. NEVER show any part of an immobile person's body lifted,\nraised, held up in the air, or posed as if presenting something:\nan object in their grip stays clenched in a hand that itself lies\nfallen on a support — the hand does not hold the object up. If the\ncanonical pose leaves a body part unspecified, resolve it as the\nmost gravity-compliant, fully-supported position.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): Blood and broken stall debris remain on the dirt; the bloodied sickle has been picked up again. 안규홍: He rises with the bloodied sickle, blood covering his hands and soaking his jeogori from the back wound; his neck remains bare.\n\nTHE HAND THAT IS DOING THIS: the object at the centre of this shot is being held, operated, read, aimed or handed over by 안규홍 right now, so 안규홍's hand — and as much of the wrist and forearm as the framing reaches — is in the frame, gripping or resting on that object exactly the way its real use requires. Match that hand to 안규홍: its size, build, skin, age, grooming, sleeve and anything worn on it belong to that person and to no one else.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 안규홍 (한국인 남성, 청년 나이대, 검은 머리, 단단히 묶은 머리) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.",
  "roll_prompts": {
   "A": "Create ONE FINAL photorealistic live-action film still of the moment below — early 20th-century Korea, Jeolla region; all people are Korean unless explicitly stated otherwise. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 낫을 거머쥔 채 바닥을 밀어내며 반쯤 일어선 안규홍의 거대하고 위협적인 뒷모습.\n\nLOCATION (lock): The outdoor dirt space beside the fallen companion in the market lane, facing the fleeing man beyond the collapsed stalls. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: medium shot\n- KEY BACKGROUND ELEMENTS: 흙바닥 (Visible beneath and beside his rising body); used as A narrow lower-frame strip grounds the low viewpoint and establishes human scale; 피 묻은 낫 (Gripped in 안규홍's hand) — Seen obliquely beside his body, with the blade separated from his outline; used as A small lower-edge detail connects his rising posture to the impending pursuit.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Natural daylight with muted saturation and restrained contrast preserves the physical weight of grief turning into violence.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The clothing, hair and overall look of 안규홍 — who appear both in that photo and in this shot — are LOCKED to that photo. Anyone else visible in that photo is NOT in this shot: never carry their face, body or clothing onto anyone here. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nIMMOBILE CHARACTER POSE — CANONICAL (identical wherever this character appears in ANY panel; on any conflict THIS POSE WINS): His dead body has been lowered onto the market's dirt ground and remains motionless among the Iljinhoe corpses, no longer supported by An Gyuhong. His eyes remain open, but the screenplay does not specify his torso orientation, head angle, or the resting positions of his arms and legs.\n\nIMMOBILE BODIES OBEY GRAVITY: a person who is dead or unconscious\nexerts NO muscular effort. Every part of their body — head, torso,\narms, hands, fingers, legs — rests fully on whatever supports it\n(floor, wall, furniture, their own lap) and hangs or slumps with\ngravity. NEVER show any part of an immobile person's body lifted,\nraised, held up in the air, or posed as if presenting something:\nan object in their grip stays clenched in a hand that itself lies\nfallen on a support — the hand does not hold the object up. If the\ncanonical pose leaves a body part unspecified, resolve it as the\nmost gravity-compliant, fully-supported position.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): Blood and broken stall debris remain on the dirt; the bloodied sickle has been picked up again. 안규홍: He rises with the bloodied sickle, blood covering his hands and soaking his jeogori from the back wound; his neck remains bare.\n\nTHE HAND THAT IS DOING THIS: the object at the centre of this shot is being held, operated, read, aimed or handed over by 안규홍 right now, so 안규홍's hand — and as much of the wrist and forearm as the framing reaches — is in the frame, gripping or resting on that object exactly the way its real use requires. Match that hand to 안규홍: its size, build, skin, age, grooming, sleeve and anything worn on it belong to that person and to no one else.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 안규홍 (한국인 남성, 청년 나이대, 검은 머리, 단단히 묶은 머리) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.",
   "B": "Create ONE FINAL photorealistic live-action film still of the moment below — early 20th-century Korea, Jeolla region; all people are Korean unless explicitly stated otherwise. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 낫을 거머쥔 채 바닥을 밀어내며 반쯤 일어선 안규홍의 거대하고 위협적인 뒷모습.\n\nLOCATION (lock): The outdoor dirt space beside the fallen companion in the market lane, facing the fleeing man beyond the collapsed stalls. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: medium shot\n- KEY BACKGROUND ELEMENTS: 흙바닥 (Visible beneath and beside his rising body); used as A narrow lower-frame strip grounds the low viewpoint and establishes human scale; 피 묻은 낫 (Gripped in 안규홍's hand) — Seen obliquely beside his body, with the blade separated from his outline; used as A small lower-edge detail connects his rising posture to the impending pursuit.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Natural daylight with muted saturation and restrained contrast preserves the physical weight of grief turning into violence.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The clothing, hair and overall look of 안규홍 — who appear both in that photo and in this shot — are LOCKED to that photo. Anyone else visible in that photo is NOT in this shot: never carry their face, body or clothing onto anyone here. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nIMMOBILE CHARACTER POSE — CANONICAL (identical wherever this character appears in ANY panel; on any conflict THIS POSE WINS): His dead body has been lowered onto the market's dirt ground and remains motionless among the Iljinhoe corpses, no longer supported by An Gyuhong. His eyes remain open, but the screenplay does not specify his torso orientation, head angle, or the resting positions of his arms and legs.\n\nIMMOBILE BODIES OBEY GRAVITY: a person who is dead or unconscious\nexerts NO muscular effort. Every part of their body — head, torso,\narms, hands, fingers, legs — rests fully on whatever supports it\n(floor, wall, furniture, their own lap) and hangs or slumps with\ngravity. NEVER show any part of an immobile person's body lifted,\nraised, held up in the air, or posed as if presenting something:\nan object in their grip stays clenched in a hand that itself lies\nfallen on a support — the hand does not hold the object up. If the\ncanonical pose leaves a body part unspecified, resolve it as the\nmost gravity-compliant, fully-supported position.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): Blood and broken stall debris remain on the dirt; the bloodied sickle has been picked up again. 안규홍: He rises with the bloodied sickle, blood covering his hands and soaking his jeogori from the back wound; his neck remains bare.\n\nTHE HAND THAT IS DOING THIS: the object at the centre of this shot is being held, operated, read, aimed or handed over by 안규홍 right now, so 안규홍's hand — and as much of the wrist and forearm as the framing reaches — is in the frame, gripping or resting on that object exactly the way its real use requires. Match that hand to 안규홍: its size, build, skin, age, grooming, sleeve and anything worn on it belong to that person and to no one else.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 안규홍 (한국인 남성, 청년 나이대, 검은 머리, 단단히 묶은 머리) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.\n\nCOMPOSITION VARIATION (this candidate only): keep the place, the people and the moment exactly as specified, and keep from the CAMERA & FRAME contract its subject scale, screen placement and key background elements exactly. For this candidate only, its camera angle and height are deliberately left open: choose a camera position distinctly different from the obvious one, as a film director picking a second setup on the same blocking. Stay true to any explicit framing the SHOT TEXT itself dictates."
  },
  "shot_run_spend_attempt_count": 1,
  "readings": [
   {
    "label": "A",
    "direction": "안규홍은 멀리 도망치는 사내를 향해 몸을 두고 있으며, 오른손에 쥔 낫은 바닥을 향해 있음.",
    "built_space": "난전이 늘어선 시장 골목의 흙바닥이 구현되었고, 인물들이 공간 안에 배치되어 있음.",
    "entities": "안규홍(검은 바지를 입고 짧은 머리를 하고 있어 레퍼런스와 불일치), 쓰러진 동료(바닥), 도망치는 사내(배경). 낫이 안규홍의 손과 바닥에 각각 하나씩 있음.",
    "hard_violations": [
     "[gemini-pro] 물리적으로 불가능한 해부학적 구조 (바닥에 쓰러진 동료의 머리와 하체가 안규홍의 다리를 관통하듯 비정상적으로 길게 분리 및 왜곡됨)",
     "[gemini-pro] 소품의 중복 생성 (손에 쥔 낫 외에 바닥에 또 다른 낫이 존재함)",
     "[gpt-high] 안규홍만 보여야 하는 숏에 배경의 달리는 남자를 추가했다.",
     "[gpt-high] 안규홍 이외 인물의 신체를 보이지 말라는 지시와 달리 누운 사람의 얼굴과 몸을 포함했다.",
     "[gpt-high] 다시 집어 든 낫 외에 전경 바닥에 두 번째 낫을 추가했다."
    ],
    "physics": "안규홍은 한쪽 무릎을 꿇고 몸을 일으키는 중이나, 바닥에 깔린 동료의 몸은 해부학적으로 연결되지 않은 채 바닥에 흩어져 있음."
   },
   {
    "label": "B",
    "direction": "안규홍은 배경에서 도망치는 사내를 정면으로 바라보고 서 있으며, 오른손에 든 낫의 날은 바닥 쪽을 향함.",
    "built_space": "무너진 난전과 시장 골목의 흙바닥이 이전 샷의 환경 및 레퍼런스와 일치하게 묘사됨.",
    "entities": "안규홍(상투를 틀고 피 묻은 흰옷을 입어 레퍼런스 일치), 낫(프롭 레퍼런스와 형태 일치), 쓰러진 동료(우측 하단 바닥), 도망치는 사내(배경).",
    "hard_violations": [
     "[gpt-high] 안규홍만 보여야 하는 숏에 배경의 달리는 남자를 추가했다.",
     "[gpt-high] 안규홍 이외 인물의 신체를 보이지 말라는 지시와 달리 오른쪽에 누운 사람의 몸을 포함했다."
    ],
    "physics": "안규홍은 중력을 지탱하며 두 발로 온전히 기립해 있고, 배경의 사내는 달리고 있으며, 쓰러진 동료는 바닥에 자연스럽게 지탱되어 있음."
   }
  ],
  "cross_model_order": {
   "policy": "cross_model_order_v2_slot0_fwd_slot1_rev_parallel2",
   "slots": [
    {
     "model": "gemini-pro",
     "order": "forward",
     "display_to_canonical": {
      "A": "A",
      "B": "B"
     },
     "raw": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 2,
        "verdict_ko": "동료의 신체 구조가 물리적으로 불가능하게 왜곡되었고 낫이 중복 생성되는 치명적 오류가 있으며, 프레이밍 척도(풀샷)와 캐릭터 복장마저 지시를 위반했습니다."
       },
       {
        "label": "B",
        "score": 8,
        "verdict_ko": "바닥을 밀어내며 반쯤 일어선 자세를 완벽히 구현하지 못하고 기립한 상태인 점은 아쉽지만, 지정된 미디엄 샷 프레이밍과 캐릭터의 복장, 소품의 형태를 레퍼런스에 맞게 충실히 재현했습니다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "안규홍은 멀리 도망치는 사내를 향해 몸을 두고 있으며, 오른손에 쥔 낫은 바닥을 향해 있음.",
        "built_space": "난전이 늘어선 시장 골목의 흙바닥이 구현되었고, 인물들이 공간 안에 배치되어 있음.",
        "entities": "안규홍(검은 바지를 입고 짧은 머리를 하고 있어 레퍼런스와 불일치), 쓰러진 동료(바닥), 도망치는 사내(배경). 낫이 안규홍의 손과 바닥에 각각 하나씩 있음.",
        "hard_violations": [
         "물리적으로 불가능한 해부학적 구조 (바닥에 쓰러진 동료의 머리와 하체가 안규홍의 다리를 관통하듯 비정상적으로 길게 분리 및 왜곡됨)",
         "소품의 중복 생성 (손에 쥔 낫 외에 바닥에 또 다른 낫이 존재함)"
        ],
        "physics": "안규홍은 한쪽 무릎을 꿇고 몸을 일으키는 중이나, 바닥에 깔린 동료의 몸은 해부학적으로 연결되지 않은 채 바닥에 흩어져 있음."
       },
       {
        "label": "B",
        "direction": "안규홍은 배경에서 도망치는 사내를 정면으로 바라보고 서 있으며, 오른손에 든 낫의 날은 바닥 쪽을 향함.",
        "built_space": "무너진 난전과 시장 골목의 흙바닥이 이전 샷의 환경 및 레퍼런스와 일치하게 묘사됨.",
        "entities": "안규홍(상투를 틀고 피 묻은 흰옷을 입어 레퍼런스 일치), 낫(프롭 레퍼런스와 형태 일치), 쓰러진 동료(우측 하단 바닥), 도망치는 사내(배경).",
        "hard_violations": [],
        "physics": "안규홍은 중력을 지탱하며 두 발로 온전히 기립해 있고, 배경의 사내는 달리고 있으며, 쓰러진 동료는 바닥에 자연스럽게 지탱되어 있음."
       }
      ],
      "all_candidates_fail": false
     },
     "normalized": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 2,
        "verdict_ko": "동료의 신체 구조가 물리적으로 불가능하게 왜곡되었고 낫이 중복 생성되는 치명적 오류가 있으며, 프레이밍 척도(풀샷)와 캐릭터 복장마저 지시를 위반했습니다."
       },
       {
        "label": "B",
        "score": 8,
        "verdict_ko": "바닥을 밀어내며 반쯤 일어선 자세를 완벽히 구현하지 못하고 기립한 상태인 점은 아쉽지만, 지정된 미디엄 샷 프레이밍과 캐릭터의 복장, 소품의 형태를 레퍼런스에 맞게 충실히 재현했습니다."
       }
      ],
      "all_candidates_fail": false,
      "readings": [
       {
        "label": "A",
        "direction": "안규홍은 멀리 도망치는 사내를 향해 몸을 두고 있으며, 오른손에 쥔 낫은 바닥을 향해 있음.",
        "built_space": "난전이 늘어선 시장 골목의 흙바닥이 구현되었고, 인물들이 공간 안에 배치되어 있음.",
        "entities": "안규홍(검은 바지를 입고 짧은 머리를 하고 있어 레퍼런스와 불일치), 쓰러진 동료(바닥), 도망치는 사내(배경). 낫이 안규홍의 손과 바닥에 각각 하나씩 있음.",
        "hard_violations": [
         "물리적으로 불가능한 해부학적 구조 (바닥에 쓰러진 동료의 머리와 하체가 안규홍의 다리를 관통하듯 비정상적으로 길게 분리 및 왜곡됨)",
         "소품의 중복 생성 (손에 쥔 낫 외에 바닥에 또 다른 낫이 존재함)"
        ],
        "physics": "안규홍은 한쪽 무릎을 꿇고 몸을 일으키는 중이나, 바닥에 깔린 동료의 몸은 해부학적으로 연결되지 않은 채 바닥에 흩어져 있음."
       },
       {
        "label": "B",
        "direction": "안규홍은 배경에서 도망치는 사내를 정면으로 바라보고 서 있으며, 오른손에 든 낫의 날은 바닥 쪽을 향함.",
        "built_space": "무너진 난전과 시장 골목의 흙바닥이 이전 샷의 환경 및 레퍼런스와 일치하게 묘사됨.",
        "entities": "안규홍(상투를 틀고 피 묻은 흰옷을 입어 레퍼런스 일치), 낫(프롭 레퍼런스와 형태 일치), 쓰러진 동료(우측 하단 바닥), 도망치는 사내(배경).",
        "hard_violations": [],
        "physics": "안규홍은 중력을 지탱하며 두 발로 온전히 기립해 있고, 배경의 사내는 달리고 있으며, 쓰러진 동료는 바닥에 자연스럽게 지탱되어 있음."
       }
      ]
     },
     "ok": true
    },
    {
     "model": "gpt-high",
     "order": "reverse",
     "display_to_canonical": {
      "A": "B",
      "B": "A"
     },
     "raw": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 3,
        "verdict_ko": "불필요한 인물들 때문에 부적합하지만, B보다 미디엄 뒷모습 구도와 이전 장면의 밝은 복장을 잘 유지하며, 반쯤 일어나는 동작 대신 이미 선 모습인 점은 큰 약점이다."
       },
       {
        "label": "B",
        "score": 2,
        "verdict_ko": "낮은 시점의 반기립 동작은 더 정확하지만, 추가 인물들과 두 번째 낫을 넣었고 전신에 가까운 구도와 검은 바지가 프레이밍·연속성 지시를 어긴다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "안규홍은 카메라에 등을 보이고 골목 중앙의 남자를 향한다. 눈은 보이지 않아 정확한 시선은 확인할 수 없다. 배경 남자는 등을 돌려 달아나는 것이 아니라 안규홍 쪽으로 달려온다. 오른손의 낫은 몸 오른쪽으로 떨어져 있으며 날끝은 오른쪽 아래로 굽어 있어, 배경 남자를 직접 겨누는 자세는 아니다.",
        "built_space": "왼쪽에 목조 점포와 진열대 두 단, 매달린 짚 제품과 바구니들이 있고, 오른쪽에는 큰 나무 상자 하나와 기울어진 판재·무너진 가판이 있다. 뒤에는 기와를 얹은 출입구 하나가 보인다. 낡은 목재와 흙길의 재료는 참고 장소와 대체로 맞지만, 흙바닥이 하단의 좁은 띠보다 넓게 노출된다. 안규홍은 중앙 통로에 있고 오른쪽 바닥에는 다른 사람의 몸이 놓여 있다.",
        "entities": "중앙에는 청년 남성 한 명이 보이며 얼굴이 가려져 정확한 한국인 정체성이나 얼굴 일치는 확인할 수 없다. 짧은 검은 머리와 밝은 저고리·바지는 이전 장면에 가깝고, 목에는 장신구가 없다. 등에 큰 혈흔이 있고 오른손도 피에 젖어 있다. 나무 손잡이와 굽은 금속 날을 가진 낫 하나를 쥐고 있으나, 참고보다 날이 가늘고 손잡이가 짧게 보인다. 배경의 달리는 남자와 오른쪽에 누운 사람은 이 숏에서 허용한 안규홍 이외의 인물이다. 읽을 수 있는 글자는 없다.",
        "hard_violations": [
         "안규홍만 보여야 하는 숏에 배경의 달리는 남자를 추가했다.",
         "안규홍 이외 인물의 신체를 보이지 말라는 지시와 달리 오른쪽에 누운 사람의 몸을 포함했다."
        ],
        "physics": "안규홍의 하체는 화면 아래로 이어지고 발 접점은 프레임 밖이므로 부유한다고 볼 근거는 없다. 다만 상체가 곧게 서 있고 왼손이 내려와 있어 바닥을 밀며 반쯤 일어나는 순간은 드러나지 않는다. 낫 손잡이는 오른손이 실제로 감싸 지지한다. 누운 사람의 보이는 팔과 다리는 흙바닥에 놓여 있다. 배경 남자는 한쪽 다리로 지면을 딛고 다른 다리를 들어 달리는 자세여서 무지지 부유는 아니다."
       },
       {
        "label": "B",
        "direction": "안규홍은 몸을 앞으로 숙이고 골목 안쪽을 향하며, 얼굴과 눈은 거의 보이지 않는다. 배경 남자는 안규홍 쪽을 보고 앞으로 달려오므로 도주 방향과 반대다. 오른손의 낫은 아래로 내려가 날이 땅 가까이 있고, 끝은 화면 왼쪽 위로 굽는다. 바닥에 누운 사람의 얼굴은 카메라 쪽으로 돌아가 있다.",
        "built_space": "양쪽에 목조 점포가 늘어서고, 왼쪽에는 진열대 두 구간과 바구니들, 오른쪽에는 큰 나무 상자 하나와 그 뒤에 비스듬히 놓인 문짝 같은 판재가 보인다. 통로 위에는 여러 천 차양이 걸려 있다. 시장의 목재·짚·흙 재료는 참고와 유사하다. 카메라는 지면 가까이에 있지만 안규홍의 양발까지 포함해 거의 전신을 보여주며, 흙바닥도 좁은 하단 띠가 아니라 넓은 전경을 차지한다. 누운 사람은 안규홍의 다리 뒤를 가로질러 놓여 있다.",
        "entities": "안규홍은 검은 머리의 청년 남성으로 보이고, 피 묻은 밝은 저고리와 맨목은 유지된다. 얼굴이 가려져 정확한 얼굴 일치는 판단할 수 없다. 바지는 참고의 밝은색과 달리 짙은 검정 계열이다. 피 묻은 오른손에 나무 손잡이 낫 하나가 있으며, 화면 맨 아래에도 별도의 낫이 하나 더 놓여 있다. 손에 든 낫의 재료와 굽은 형태는 참고와 유사하다. 누운 청년 남성과 배경의 달리는 남성이 추가로 보인다. 읽을 수 있는 글자는 없다.",
        "hard_violations": [
         "안규홍만 보여야 하는 숏에 배경의 달리는 남자를 추가했다.",
         "안규홍 이외 인물의 신체를 보이지 말라는 지시와 달리 누운 사람의 얼굴과 몸을 포함했다.",
         "다시 집어 든 낫 외에 전경 바닥에 두 번째 낫을 추가했다."
        ],
        "physics": "왼발은 흙바닥을 딛고 오른쪽 무릎과 접힌 다리는 지면 가까이에 놓여 있어, 다리로 체중을 지지하며 일어나는 중간 자세로 읽힌다. 앞으로 기울어진 몸에도 확인 가능한 지지가 있으며 부유하지 않는다. 오른손은 낫 손잡이를 확실히 잡고 있다. 누운 사람의 머리·몸통·팔·다리는 지면에 기대어 있고, 전경의 두 번째 낫도 바닥에 놓여 있다. 배경 남자는 한 발의 착지와 다른 발의 들림으로 달리는 동작을 보인다."
       }
      ],
      "all_candidates_fail": true
     },
     "normalized": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 3,
        "verdict_ko": "불필요한 인물들 때문에 부적합하지만, B보다 미디엄 뒷모습 구도와 이전 장면의 밝은 복장을 잘 유지하며, 반쯤 일어나는 동작 대신 이미 선 모습인 점은 큰 약점이다."
       },
       {
        "label": "A",
        "score": 2,
        "verdict_ko": "낮은 시점의 반기립 동작은 더 정확하지만, 추가 인물들과 두 번째 낫을 넣었고 전신에 가까운 구도와 검은 바지가 프레이밍·연속성 지시를 어긴다."
       }
      ],
      "all_candidates_fail": true,
      "readings": [
       {
        "label": "B",
        "direction": "안규홍은 카메라에 등을 보이고 골목 중앙의 남자를 향한다. 눈은 보이지 않아 정확한 시선은 확인할 수 없다. 배경 남자는 등을 돌려 달아나는 것이 아니라 안규홍 쪽으로 달려온다. 오른손의 낫은 몸 오른쪽으로 떨어져 있으며 날끝은 오른쪽 아래로 굽어 있어, 배경 남자를 직접 겨누는 자세는 아니다.",
        "built_space": "왼쪽에 목조 점포와 진열대 두 단, 매달린 짚 제품과 바구니들이 있고, 오른쪽에는 큰 나무 상자 하나와 기울어진 판재·무너진 가판이 있다. 뒤에는 기와를 얹은 출입구 하나가 보인다. 낡은 목재와 흙길의 재료는 참고 장소와 대체로 맞지만, 흙바닥이 하단의 좁은 띠보다 넓게 노출된다. 안규홍은 중앙 통로에 있고 오른쪽 바닥에는 다른 사람의 몸이 놓여 있다.",
        "entities": "중앙에는 청년 남성 한 명이 보이며 얼굴이 가려져 정확한 한국인 정체성이나 얼굴 일치는 확인할 수 없다. 짧은 검은 머리와 밝은 저고리·바지는 이전 장면에 가깝고, 목에는 장신구가 없다. 등에 큰 혈흔이 있고 오른손도 피에 젖어 있다. 나무 손잡이와 굽은 금속 날을 가진 낫 하나를 쥐고 있으나, 참고보다 날이 가늘고 손잡이가 짧게 보인다. 배경의 달리는 남자와 오른쪽에 누운 사람은 이 숏에서 허용한 안규홍 이외의 인물이다. 읽을 수 있는 글자는 없다.",
        "hard_violations": [
         "안규홍만 보여야 하는 숏에 배경의 달리는 남자를 추가했다.",
         "안규홍 이외 인물의 신체를 보이지 말라는 지시와 달리 오른쪽에 누운 사람의 몸을 포함했다."
        ],
        "physics": "안규홍의 하체는 화면 아래로 이어지고 발 접점은 프레임 밖이므로 부유한다고 볼 근거는 없다. 다만 상체가 곧게 서 있고 왼손이 내려와 있어 바닥을 밀며 반쯤 일어나는 순간은 드러나지 않는다. 낫 손잡이는 오른손이 실제로 감싸 지지한다. 누운 사람의 보이는 팔과 다리는 흙바닥에 놓여 있다. 배경 남자는 한쪽 다리로 지면을 딛고 다른 다리를 들어 달리는 자세여서 무지지 부유는 아니다."
       },
       {
        "label": "A",
        "direction": "안규홍은 몸을 앞으로 숙이고 골목 안쪽을 향하며, 얼굴과 눈은 거의 보이지 않는다. 배경 남자는 안규홍 쪽을 보고 앞으로 달려오므로 도주 방향과 반대다. 오른손의 낫은 아래로 내려가 날이 땅 가까이 있고, 끝은 화면 왼쪽 위로 굽는다. 바닥에 누운 사람의 얼굴은 카메라 쪽으로 돌아가 있다.",
        "built_space": "양쪽에 목조 점포가 늘어서고, 왼쪽에는 진열대 두 구간과 바구니들, 오른쪽에는 큰 나무 상자 하나와 그 뒤에 비스듬히 놓인 문짝 같은 판재가 보인다. 통로 위에는 여러 천 차양이 걸려 있다. 시장의 목재·짚·흙 재료는 참고와 유사하다. 카메라는 지면 가까이에 있지만 안규홍의 양발까지 포함해 거의 전신을 보여주며, 흙바닥도 좁은 하단 띠가 아니라 넓은 전경을 차지한다. 누운 사람은 안규홍의 다리 뒤를 가로질러 놓여 있다.",
        "entities": "안규홍은 검은 머리의 청년 남성으로 보이고, 피 묻은 밝은 저고리와 맨목은 유지된다. 얼굴이 가려져 정확한 얼굴 일치는 판단할 수 없다. 바지는 참고의 밝은색과 달리 짙은 검정 계열이다. 피 묻은 오른손에 나무 손잡이 낫 하나가 있으며, 화면 맨 아래에도 별도의 낫이 하나 더 놓여 있다. 손에 든 낫의 재료와 굽은 형태는 참고와 유사하다. 누운 청년 남성과 배경의 달리는 남성이 추가로 보인다. 읽을 수 있는 글자는 없다.",
        "hard_violations": [
         "안규홍만 보여야 하는 숏에 배경의 달리는 남자를 추가했다.",
         "안규홍 이외 인물의 신체를 보이지 말라는 지시와 달리 누운 사람의 얼굴과 몸을 포함했다.",
         "다시 집어 든 낫 외에 전경 바닥에 두 번째 낫을 추가했다."
        ],
        "physics": "왼발은 흙바닥을 딛고 오른쪽 무릎과 접힌 다리는 지면 가까이에 놓여 있어, 다리로 체중을 지지하며 일어나는 중간 자세로 읽힌다. 앞으로 기울어진 몸에도 확인 가능한 지지가 있으며 부유하지 않는다. 오른손은 낫 손잡이를 확실히 잡고 있다. 누운 사람의 머리·몸통·팔·다리는 지면에 기대어 있고, 전경의 두 번째 낫도 바닥에 놓여 있다. 배경 남자는 한 발의 착지와 다른 발의 들림으로 달리는 동작을 보인다."
       }
      ]
     },
     "ok": true
    }
   ],
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "slot_winner_match": true,
   "slot_winner": {
    "gemini-pro": "B",
    "gpt-high": "B"
   },
   "route": "cross_slot_agree"
  },
  "dual": {
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "normalized": {
    "A": 0.917,
    "B": 2.0
   },
   "adjusted": {
    "A": 0.667,
    "B": 1.75
   },
   "violations": {
    "A": [
     "[gemini-pro] 물리적으로 불가능한 해부학적 구조 (바닥에 쓰러진 동료의 머리와 하체가 안규홍의 다리를 관통하듯 비정상적으로 길게 분리 및 왜곡됨)",
     "[gemini-pro] 소품의 중복 생성 (손에 쥔 낫 외에 바닥에 또 다른 낫이 존재함)",
     "[gpt-high] 안규홍만 보여야 하는 숏에 배경의 달리는 남자를 추가했다.",
     "[gpt-high] 안규홍 이외 인물의 신체를 보이지 말라는 지시와 달리 누운 사람의 얼굴과 몸을 포함했다.",
     "[gpt-high] 다시 집어 든 낫 외에 전경 바닥에 두 번째 낫을 추가했다."
    ],
    "B": [
     "[gpt-high] 안규홍만 보여야 하는 숏에 배경의 달리는 남자를 추가했다.",
     "[gpt-high] 안규홍 이외 인물의 신체를 보이지 말라는 지시와 달리 오른쪽에 누운 사람의 몸을 포함했다."
    ]
   },
   "per_model_winner": {
    "gemini-pro": "B",
    "gpt-high": "B"
   },
   "agreed": true
  },
  "totals": {
   "A": 667,
   "B": 1750
  },
  "selected": "B",
  "ranking": [
   "B",
   "A"
  ],
  "verdicts": [
   {
    "label": "A",
    "score": 667,
    "verdict_ko": "동료의 신체 구조가 물리적으로 불가능하게 왜곡되었고 낫이 중복 생성되는 치명적 오류가 있으며, 프레이밍 척도(풀샷)와 캐릭터 복장마저 지시를 위반했습니다.  ★위반: [gemini-pro] 물리적으로 불가능한 해부학적 구조 (바닥에 쓰러진 동료의 머리와 하체가 안규홍의 다리를 관통하듯 비정상적으로 길게 분리 및 왜곡됨) / [gemini-pro] 소품의 중복 생성 (손에 쥔 낫 외에 바닥에 또 다른 낫이 존재함) / [gpt-high] 안규홍만 보여야 하는 숏에 배경의 달리는 남자를 추가했다. / [gpt-high] 안규홍 이외 인물의 신체를 보이지 말라는 지시와 달리 누운 사람의 얼굴과 몸을 포함했다. / [gpt-high] 다시 집어 든 낫 외에 전경 바닥에 두 번째 낫을 추가했다."
   },
   {
    "label": "B",
    "score": 1750,
    "verdict_ko": "바닥을 밀어내며 반쯤 일어선 자세를 완벽히 구현하지 못하고 기립한 상태인 점은 아쉽지만, 지정된 미디엄 샷 프레이밍과 캐릭터의 복장, 소품의 형태를 레퍼런스에 맞게 충실히 재현했습니다.  ★위반: [gpt-high] 안규홍만 보여야 하는 숏에 배경의 달리는 남자를 추가했다. / [gpt-high] 안규홍 이외 인물의 신체를 보이지 말라는 지시와 달리 오른쪽에 누운 사람의 몸을 포함했다."
   }
  ],
  "refs": [
   {
    "label": "PREVIOUS SHOT STILL — a visually related earlier shot of this same place: the location's look, materials, fixed features and lighting mood are LOCKED to this photo; never copy its camera framing. The clothing, hair and overall look of 안규홍 — who appear both in that photo and in this shot — are LOCKED to that photo. Anyone else visible in that photo is NOT in this shot: never carry their face, body or clothing onto anyone here. If this photo shows a character who cannot move (dead or unconscious), that character's exact pose, position and orientation are ALSO LOCKED — treat their body as a fixed prop of the set that only the camera moves around.",
    "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/db00517b-165b-4005-a287-f28936b59b95/images/272bd7af-9294-4490-98c6-88a40e9deacc/scene/recipe/S55sh34_sel.png",
    "asset_id": "17233d44-bfa0-4e91-a62a-67f3afdfec79",
    "role": "prev_still"
   },
   {
    "label": "CHARACTER REFERENCE — 안규홍: the exact person appearing in this shot; match face, hair and build exactly.",
    "path": "<bytes:1028384>",
    "asset_id": "3b5894db-ce55-4f34-b966-5e725c3aa952",
    "role": "character_ref"
   },
   {
    "label": "PROP REFERENCE — 낫: the exact object appearing in this shot; match its look, material and wear exactly.",
    "path": "<bytes:977465>",
    "asset_id": "e9884693-b2aa-4769-9ee0-4c7b50db40c2",
    "role": "prop_ref"
   }
  ],
  "critique_skipped": true,
  "shot_run_produced": false,
  "shot_run_uid": "06aa1438-de10-7310-badb-1b5f169bbc7d",
  "ref_mode": "prev+엔티티",
  "share_plan": {
   "ref_plan": "prev",
   "prev_anchor_tag": "S55sh34"
  }
 },
 "S55sh40::cine": {
  "applied": true,
  "attempted_at": "2026-09-09T07:19:36.780547+00:00",
  "fingerprint": "7b2a039b54e02a5ce94d6059ccb31cb4c6630a01fc7881b4dabf5a700b68cac9",
  "fingerprint_version": 2,
  "provider": "mai",
  "endpoint": "openrouter/chat-completions",
  "model": "microsoft/mai-image-2.6",
  "pack": "24.202608252115",
  "source_file": "S55sh40_sel.png",
  "source_sha256": "2ead01633489c6b1fd098cb978d0596bf91b540ec3c3b44a994a67cf4c3cf29c",
  "file": "S55sh40_cine.png",
  "staged_sha256": "38f4e36e010a706f85a0ed4ec1caf16836b82521a2cd6ff18afbcf8b5abb3212",
  "latency_ms": 26682
 },
 "S56sh1::signage": {
  "fp": "7bb8b1425b3c2f48",
  "inscriptions": [],
  "cues": [],
  "dropped": []
 },
 "S56sh1": {
  "input_fingerprint": "2794c9b55f498817",
  "prompt": "Create ONE FINAL photorealistic live-action film still of the moment below — early 20th-century Korea, Jeolla region; all people are Korean unless explicitly stated otherwise. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 오른팔을 허공으로 치켜든 채 산채 쪽을 뚫어져라 내려다보는 토미이시의 서늘한 정면.\n\nLOCATION (lock): The ridge overlooking the mountain camp's clearing, above the surrounding cypress forest in daylight. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: medium shot\n- KEY BACKGROUND ELEMENTS: 능선 (Occupied by 토미이시 above the camp); used as A restrained background contour locates the command position without introducing details from inside the camp.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Natural daylight, muted saturation, and controlled facial contrast give the held command a sober, unsentimental chill.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The clothing, hair and overall look of 토미이시 — who appear both in that photo and in this shot — are LOCKED to that photo. Anyone else visible in that photo is NOT in this shot: never carry their face, body or clothing onto anyone here. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The rice meal and upright cauldron remain intact below the ridge; cloth-wrapped ammunition is already hidden beneath stones near trees. 토미이시: He stands on the ridge with his right arm raised, retaining his sword.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 토미이시 (일본인 남성, 성인 나이대, 자연스러운 검은 머리) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.",
  "roll_prompts": {
   "A": "Create ONE FINAL photorealistic live-action film still of the moment below — early 20th-century Korea, Jeolla region; all people are Korean unless explicitly stated otherwise. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 오른팔을 허공으로 치켜든 채 산채 쪽을 뚫어져라 내려다보는 토미이시의 서늘한 정면.\n\nLOCATION (lock): The ridge overlooking the mountain camp's clearing, above the surrounding cypress forest in daylight. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: medium shot\n- KEY BACKGROUND ELEMENTS: 능선 (Occupied by 토미이시 above the camp); used as A restrained background contour locates the command position without introducing details from inside the camp.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Natural daylight, muted saturation, and controlled facial contrast give the held command a sober, unsentimental chill.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The clothing, hair and overall look of 토미이시 — who appear both in that photo and in this shot — are LOCKED to that photo. Anyone else visible in that photo is NOT in this shot: never carry their face, body or clothing onto anyone here. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The rice meal and upright cauldron remain intact below the ridge; cloth-wrapped ammunition is already hidden beneath stones near trees. 토미이시: He stands on the ridge with his right arm raised, retaining his sword.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 토미이시 (일본인 남성, 성인 나이대, 자연스러운 검은 머리) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.",
   "B": "Create ONE FINAL photorealistic live-action film still of the moment below — early 20th-century Korea, Jeolla region; all people are Korean unless explicitly stated otherwise. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 오른팔을 허공으로 치켜든 채 산채 쪽을 뚫어져라 내려다보는 토미이시의 서늘한 정면.\n\nLOCATION (lock): The ridge overlooking the mountain camp's clearing, above the surrounding cypress forest in daylight. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: medium shot\n- KEY BACKGROUND ELEMENTS: 능선 (Occupied by 토미이시 above the camp); used as A restrained background contour locates the command position without introducing details from inside the camp.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Natural daylight, muted saturation, and controlled facial contrast give the held command a sober, unsentimental chill.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The clothing, hair and overall look of 토미이시 — who appear both in that photo and in this shot — are LOCKED to that photo. Anyone else visible in that photo is NOT in this shot: never carry their face, body or clothing onto anyone here. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The rice meal and upright cauldron remain intact below the ridge; cloth-wrapped ammunition is already hidden beneath stones near trees. 토미이시: He stands on the ridge with his right arm raised, retaining his sword.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 토미이시 (일본인 남성, 성인 나이대, 자연스러운 검은 머리) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.\n\nCOMPOSITION VARIATION (this candidate only): keep the place, the people and the moment exactly as specified, and keep from the CAMERA & FRAME contract its subject scale, screen placement and key background elements exactly. For this candidate only, its camera angle and height are deliberately left open: choose a camera position distinctly different from the obvious one, as a film director picking a second setup on the same blocking. Stay true to any explicit framing the SHOT TEXT itself dictates."
  },
  "shot_run_spend_attempt_count": 1,
  "readings": [
   {
    "label": "A",
    "direction": "피사체는 정면(카메라)을 응시하고 있으며, 오른팔을 허공으로 치켜든 채 손바닥을 앞을 향하고 있다. 산채 쪽을 내려다봐야 하나 산채가 등 뒤에 위치해 시선 방향과 공간 구조상 모순이 있음.",
    "built_space": "야외 능선 위로, 등 뒤 아래쪽 배경에 초가집과 나무 울타리가 있는 산채가 뚜렷하게 보이며, 참조 이미지의 지형 요소가 유사하게 배치됨.",
    "entities": "토미이시(일본인 성인 남성)가 참조 이미지와 일치하는 군복, 군모, 훈장을 착용하고 있으며, 명시된 대로 오른팔을 들고 있음.",
    "hard_violations": [],
    "physics": "피사체는 지면에 안정적으로 서 있고, 허리에 찬 칼은 중력에 맞게 자연스럽게 매달려 있으며, 치켜든 팔의 자세와 지탱도 자연스러움."
   },
   {
    "label": "B",
    "direction": "피사체는 카메라 쪽(정면)을 바라보고 있으며, 오른팔을 치켜들고 손바닥은 측면을 향함. 산채가 피사체 뒤편 배경에 있어 내려다보는 시선 방향이 성립하지 않음.",
    "built_space": "능선 위에 서 있으며, 뒤쪽 아래로 산채의 구조물(초가집, 울타리)이 선명하게 나타남.",
    "entities": "토미이시가 지정된 군복을 입고 있으나, 허리에 찬 칼의 길이가 비정상적으로 길게 왜곡되어 나타남.",
    "hard_violations": [
     "[gemini-pro] 물리적으로 불가능한 비율로 길어져 지면까지 닿아 있는 기형적인 칼집 (소품 및 해부학적 물리 오류)"
    ],
    "physics": "피사체 자체는 서 있으나, 왼쪽 허리에 매달린 칼집이 마치 지팡이처럼 바닥까지 늘어져 지면에 닿아 있어 물리적 형태와 비례가 완전히 붕괴됨."
   }
  ],
  "cross_model_order": {
   "policy": "cross_model_order_v2_slot0_fwd_slot1_rev_parallel2",
   "slots": [
    {
     "model": "gemini-pro",
     "order": "forward",
     "display_to_canonical": {
      "A": "A",
      "B": "B"
     },
     "raw": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 7,
        "verdict_ko": "미디엄 샷 규격을 비교적 잘 따랐으며 복장과 오른팔을 든 자세를 충실히 구현했으나, 산채를 내려다보는 시선 논리와 배경 디테일 제한 요청을 완전히 소화하지는 못함."
       },
       {
        "label": "B",
        "score": 3,
        "verdict_ko": "지면까지 비정상적으로 늘어난 칼집으로 인해 심각한 소품 및 물리적 오류가 발생했으며, 미디엄 샷 프레이밍 지시도 어김."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "피사체는 정면(카메라)을 응시하고 있으며, 오른팔을 허공으로 치켜든 채 손바닥을 앞을 향하고 있다. 산채 쪽을 내려다봐야 하나 산채가 등 뒤에 위치해 시선 방향과 공간 구조상 모순이 있음.",
        "built_space": "야외 능선 위로, 등 뒤 아래쪽 배경에 초가집과 나무 울타리가 있는 산채가 뚜렷하게 보이며, 참조 이미지의 지형 요소가 유사하게 배치됨.",
        "entities": "토미이시(일본인 성인 남성)가 참조 이미지와 일치하는 군복, 군모, 훈장을 착용하고 있으며, 명시된 대로 오른팔을 들고 있음.",
        "hard_violations": [],
        "physics": "피사체는 지면에 안정적으로 서 있고, 허리에 찬 칼은 중력에 맞게 자연스럽게 매달려 있으며, 치켜든 팔의 자세와 지탱도 자연스러움."
       },
       {
        "label": "B",
        "direction": "피사체는 카메라 쪽(정면)을 바라보고 있으며, 오른팔을 치켜들고 손바닥은 측면을 향함. 산채가 피사체 뒤편 배경에 있어 내려다보는 시선 방향이 성립하지 않음.",
        "built_space": "능선 위에 서 있으며, 뒤쪽 아래로 산채의 구조물(초가집, 울타리)이 선명하게 나타남.",
        "entities": "토미이시가 지정된 군복을 입고 있으나, 허리에 찬 칼의 길이가 비정상적으로 길게 왜곡되어 나타남.",
        "hard_violations": [
         "물리적으로 불가능한 비율로 길어져 지면까지 닿아 있는 기형적인 칼집 (소품 및 해부학적 물리 오류)"
        ],
        "physics": "피사체 자체는 서 있으나, 왼쪽 허리에 매달린 칼집이 마치 지팡이처럼 바닥까지 늘어져 지면에 닿아 있어 물리적 형태와 비례가 완전히 붕괴됨."
       }
      ],
      "all_candidates_fail": false
     },
     "normalized": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 7,
        "verdict_ko": "미디엄 샷 규격을 비교적 잘 따랐으며 복장과 오른팔을 든 자세를 충실히 구현했으나, 산채를 내려다보는 시선 논리와 배경 디테일 제한 요청을 완전히 소화하지는 못함."
       },
       {
        "label": "B",
        "score": 3,
        "verdict_ko": "지면까지 비정상적으로 늘어난 칼집으로 인해 심각한 소품 및 물리적 오류가 발생했으며, 미디엄 샷 프레이밍 지시도 어김."
       }
      ],
      "all_candidates_fail": false,
      "readings": [
       {
        "label": "A",
        "direction": "피사체는 정면(카메라)을 응시하고 있으며, 오른팔을 허공으로 치켜든 채 손바닥을 앞을 향하고 있다. 산채 쪽을 내려다봐야 하나 산채가 등 뒤에 위치해 시선 방향과 공간 구조상 모순이 있음.",
        "built_space": "야외 능선 위로, 등 뒤 아래쪽 배경에 초가집과 나무 울타리가 있는 산채가 뚜렷하게 보이며, 참조 이미지의 지형 요소가 유사하게 배치됨.",
        "entities": "토미이시(일본인 성인 남성)가 참조 이미지와 일치하는 군복, 군모, 훈장을 착용하고 있으며, 명시된 대로 오른팔을 들고 있음.",
        "hard_violations": [],
        "physics": "피사체는 지면에 안정적으로 서 있고, 허리에 찬 칼은 중력에 맞게 자연스럽게 매달려 있으며, 치켜든 팔의 자세와 지탱도 자연스러움."
       },
       {
        "label": "B",
        "direction": "피사체는 카메라 쪽(정면)을 바라보고 있으며, 오른팔을 치켜들고 손바닥은 측면을 향함. 산채가 피사체 뒤편 배경에 있어 내려다보는 시선 방향이 성립하지 않음.",
        "built_space": "능선 위에 서 있으며, 뒤쪽 아래로 산채의 구조물(초가집, 울타리)이 선명하게 나타남.",
        "entities": "토미이시가 지정된 군복을 입고 있으나, 허리에 찬 칼의 길이가 비정상적으로 길게 왜곡되어 나타남.",
        "hard_violations": [
         "물리적으로 불가능한 비율로 길어져 지면까지 닿아 있는 기형적인 칼집 (소품 및 해부학적 물리 오류)"
        ],
        "physics": "피사체 자체는 서 있으나, 왼쪽 허리에 매달린 칼집이 마치 지팡이처럼 바닥까지 늘어져 지면에 닿아 있어 물리적 형태와 비례가 완전히 붕괴됨."
       }
      ]
     },
     "ok": true
    },
    {
     "model": "gpt-high",
     "order": "reverse",
     "display_to_canonical": {
      "A": "B",
      "B": "A"
     },
     "raw": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 3,
        "verdict_ko": "오른팔을 든 정면과 복장은 맞지만, 산채를 등지고 렌즈를 바라보며 미디엄 숏 대신 장화까지 보이는 넓은 구도로 핵심 지시를 놓쳤다."
       },
       {
        "label": "B",
        "score": 5,
        "verdict_ko": "A보다 인물 크기가 커 미디엄 숏에 가깝고 서늘한 정면도 분명하지만, 등 뒤 산채가 아닌 카메라를 응시하며 산채 내부까지 드러낸다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "인물의 오른팔은 머리 위로 올라가고 손끝은 위를 향한다. 얼굴과 눈은 거의 수평으로 카메라를 향한다. 산채는 인물 뒤 화면 왼쪽 아래에 있으므로, 보이는 시선은 산채에 도달하지 않는다. 왼쪽 허리에 찬 검은 칼집에 들어간 채 아래로 향한다.",
        "built_space": "흙과 돌, 노출된 뿌리가 있는 능선 위에 인물 한 명이 서 있다. 뒤쪽 아래에는 초가 건물 두 동, 그 둘을 둘러싼 목책 한 둘레와 앞쪽 출입구 한 곳이 보인다. 숲과 능선의 재질 및 배치는 이전 숏과 대체로 일치하지만, 절제된 능선 윤곽만 요구한 배경에 산채 마당과 건물을 상세히 노출한다.",
        "entities": "성인 동아시아계 남성 한 명만 보이며, 콧수염과 턱수염, 모자 아래 검은 머리, 얼굴 윤곽은 토미이시 참조와 대체로 부합한다. 일본인이라는 국적 자체는 외관만으로 확인할 수 없다. 갈색 군복과 군모, 붉은 견장, 훈장, 가죽 교차 띠와 허리띠, 장화 및 왼쪽 허리의 검이 보인다. 밥과 가마솥, 숨긴 탄약은 식별되지 않으며 이 구도에서 노출할 필요가 없다. 추가 인물이나 읽을 수 있는 글자는 없다.",
        "hard_violations": [],
        "physics": "든 오른팔은 어깨와 팔꿈치로 자연스럽게 연결되어 있으며 실제로 유지 가능한 자세다. 몸은 곧게 선 다리로 이어지고 발의 접지점은 화면 아래에 잘려 있다. 공중에 떠 있다는 증거는 없다. 검과 칼집은 허리띠에 매달려 있고, 가죽 띠와 군복도 몸에 지지되어 있다."
       },
       {
        "label": "B",
        "direction": "오른팔을 위로 치켜들고 손바닥을 카메라 쪽으로 펼친다. 턱은 약간 내려가 있지만 눈은 카메라 부근을 응시한다. 산채는 등 뒤 화면 왼쪽에 있으므로 산채를 내려다보는 방향은 아니다. 왼쪽 허리의 검과 칼집은 아래로 향한다.",
        "built_space": "인물 한 명이 뿌리와 돌이 드러난 흙 능선에 서 있고 뒤로 울창한 침엽수림이 이어진다. 아래 산채에는 초가 두 동과 목책 한 둘레, 앞쪽 출입구 한 곳이 보인다. 참조 장소의 주요 구성은 유지되지만, 산채 내부를 보여주지 말라는 배경 지시에는 어긋난다. 인물을 허벅지 부근에서 자르는 구도로 A보다 미디엄 숏에 가깝지만 여전히 다소 넓다.",
        "entities": "토미이시 참조와 유사한 얼굴, 콧수염과 턱수염을 가진 성인 동아시아계 남성 한 명이다. 모자 아래 검은 머리가 일부 보이며 국적은 시각적으로 확정할 수 없다. 군모, 갈색 군복, 붉은 견장, 훈장, 가죽 띠와 허리띠, 왼쪽 허리의 검이 참조와 대체로 일치한다. 장화는 프레임 밖이므로 평가하지 않는다. 밥과 가마솥 및 숨긴 탄약도 식별되지 않지만 이를 보여주기 위해 구도를 넓힐 필요는 없다. 추가 인물이나 읽을 수 있는 글자는 없다.",
        "hard_violations": [],
        "physics": "오른손과 팔은 어깨에 정상적으로 연결되고, 팔꿈치를 약간 굽힌 채 들어 올린 자세는 가능하다. 왼팔은 몸 옆으로 내려와 있다. 하체와 발의 접지점은 잘려 있지만 몸통은 아래 다리로 자연스럽게 이어져 서 있는 자세로 읽힌다. 부유하는 신체나 물체는 보이지 않으며 검과 칼집은 허리 장구에 지지된다."
       }
      ],
      "all_candidates_fail": true
     },
     "normalized": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 3,
        "verdict_ko": "오른팔을 든 정면과 복장은 맞지만, 산채를 등지고 렌즈를 바라보며 미디엄 숏 대신 장화까지 보이는 넓은 구도로 핵심 지시를 놓쳤다."
       },
       {
        "label": "A",
        "score": 5,
        "verdict_ko": "A보다 인물 크기가 커 미디엄 숏에 가깝고 서늘한 정면도 분명하지만, 등 뒤 산채가 아닌 카메라를 응시하며 산채 내부까지 드러낸다."
       }
      ],
      "all_candidates_fail": true,
      "readings": [
       {
        "label": "B",
        "direction": "인물의 오른팔은 머리 위로 올라가고 손끝은 위를 향한다. 얼굴과 눈은 거의 수평으로 카메라를 향한다. 산채는 인물 뒤 화면 왼쪽 아래에 있으므로, 보이는 시선은 산채에 도달하지 않는다. 왼쪽 허리에 찬 검은 칼집에 들어간 채 아래로 향한다.",
        "built_space": "흙과 돌, 노출된 뿌리가 있는 능선 위에 인물 한 명이 서 있다. 뒤쪽 아래에는 초가 건물 두 동, 그 둘을 둘러싼 목책 한 둘레와 앞쪽 출입구 한 곳이 보인다. 숲과 능선의 재질 및 배치는 이전 숏과 대체로 일치하지만, 절제된 능선 윤곽만 요구한 배경에 산채 마당과 건물을 상세히 노출한다.",
        "entities": "성인 동아시아계 남성 한 명만 보이며, 콧수염과 턱수염, 모자 아래 검은 머리, 얼굴 윤곽은 토미이시 참조와 대체로 부합한다. 일본인이라는 국적 자체는 외관만으로 확인할 수 없다. 갈색 군복과 군모, 붉은 견장, 훈장, 가죽 교차 띠와 허리띠, 장화 및 왼쪽 허리의 검이 보인다. 밥과 가마솥, 숨긴 탄약은 식별되지 않으며 이 구도에서 노출할 필요가 없다. 추가 인물이나 읽을 수 있는 글자는 없다.",
        "hard_violations": [],
        "physics": "든 오른팔은 어깨와 팔꿈치로 자연스럽게 연결되어 있으며 실제로 유지 가능한 자세다. 몸은 곧게 선 다리로 이어지고 발의 접지점은 화면 아래에 잘려 있다. 공중에 떠 있다는 증거는 없다. 검과 칼집은 허리띠에 매달려 있고, 가죽 띠와 군복도 몸에 지지되어 있다."
       },
       {
        "label": "A",
        "direction": "오른팔을 위로 치켜들고 손바닥을 카메라 쪽으로 펼친다. 턱은 약간 내려가 있지만 눈은 카메라 부근을 응시한다. 산채는 등 뒤 화면 왼쪽에 있으므로 산채를 내려다보는 방향은 아니다. 왼쪽 허리의 검과 칼집은 아래로 향한다.",
        "built_space": "인물 한 명이 뿌리와 돌이 드러난 흙 능선에 서 있고 뒤로 울창한 침엽수림이 이어진다. 아래 산채에는 초가 두 동과 목책 한 둘레, 앞쪽 출입구 한 곳이 보인다. 참조 장소의 주요 구성은 유지되지만, 산채 내부를 보여주지 말라는 배경 지시에는 어긋난다. 인물을 허벅지 부근에서 자르는 구도로 A보다 미디엄 숏에 가깝지만 여전히 다소 넓다.",
        "entities": "토미이시 참조와 유사한 얼굴, 콧수염과 턱수염을 가진 성인 동아시아계 남성 한 명이다. 모자 아래 검은 머리가 일부 보이며 국적은 시각적으로 확정할 수 없다. 군모, 갈색 군복, 붉은 견장, 훈장, 가죽 띠와 허리띠, 왼쪽 허리의 검이 참조와 대체로 일치한다. 장화는 프레임 밖이므로 평가하지 않는다. 밥과 가마솥 및 숨긴 탄약도 식별되지 않지만 이를 보여주기 위해 구도를 넓힐 필요는 없다. 추가 인물이나 읽을 수 있는 글자는 없다.",
        "hard_violations": [],
        "physics": "오른손과 팔은 어깨에 정상적으로 연결되고, 팔꿈치를 약간 굽힌 채 들어 올린 자세는 가능하다. 왼팔은 몸 옆으로 내려와 있다. 하체와 발의 접지점은 잘려 있지만 몸통은 아래 다리로 자연스럽게 이어져 서 있는 자세로 읽힌다. 부유하는 신체나 물체는 보이지 않으며 검과 칼집은 허리 장구에 지지된다."
       }
      ]
     },
     "ok": true
    }
   ],
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "slot_winner_match": true,
   "slot_winner": {
    "gemini-pro": "A",
    "gpt-high": "A"
   },
   "route": "cross_slot_agree"
  },
  "dual": {
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "normalized": {
    "A": 2.0,
    "B": 1.029
   },
   "adjusted": {
    "A": 2.0,
    "B": 0.779
   },
   "violations": {
    "B": [
     "[gemini-pro] 물리적으로 불가능한 비율로 길어져 지면까지 닿아 있는 기형적인 칼집 (소품 및 해부학적 물리 오류)"
    ]
   },
   "per_model_winner": {
    "gemini-pro": "A",
    "gpt-high": "A"
   },
   "agreed": true
  },
  "totals": {
   "A": 2000,
   "B": 779
  },
  "selected": "A",
  "ranking": [
   "A",
   "B"
  ],
  "verdicts": [
   {
    "label": "A",
    "score": 2000,
    "verdict_ko": "미디엄 샷 규격을 비교적 잘 따랐으며 복장과 오른팔을 든 자세를 충실히 구현했으나, 산채를 내려다보는 시선 논리와 배경 디테일 제한 요청을 완전히 소화하지는 못함."
   },
   {
    "label": "B",
    "score": 779,
    "verdict_ko": "지면까지 비정상적으로 늘어난 칼집으로 인해 심각한 소품 및 물리적 오류가 발생했으며, 미디엄 샷 프레이밍 지시도 어김.  ★위반: [gemini-pro] 물리적으로 불가능한 비율로 길어져 지면까지 닿아 있는 기형적인 칼집 (소품 및 해부학적 물리 오류)"
   }
  ],
  "refs": [
   {
    "label": "PREVIOUS SHOT STILL — a visually related earlier shot of this same place: the location's look, materials, fixed features and lighting mood are LOCKED to this photo; never copy its camera framing. The clothing, hair and overall look of 토미이시 — who appear both in that photo and in this shot — are LOCKED to that photo. Anyone else visible in that photo is NOT in this shot: never carry their face, body or clothing onto anyone here. If this photo shows a character who cannot move (dead or unconscious), that character's exact pose, position and orientation are ALSO LOCKED — treat their body as a fixed prop of the set that only the camera moves around.",
    "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/db00517b-165b-4005-a287-f28936b59b95/images/272bd7af-9294-4490-98c6-88a40e9deacc/scene/recipe/S52sh1_sel.png",
    "asset_id": "959a963f-9d07-435d-9add-87cd97f56e93",
    "role": "prev_still"
   },
   {
    "label": "CHARACTER REFERENCE — 토미이시: the exact person appearing in this shot; match face, hair and build exactly.",
    "path": "<bytes:797699>",
    "asset_id": "c861ba0f-8750-4262-8324-25cf5c6c9dcd",
    "role": "character_ref"
   }
  ],
  "critique_skipped": true,
  "shot_run_produced": false,
  "shot_run_uid": "06aa1438-dfba-7754-8e0d-cd29f85d6236",
  "ref_mode": "prev+엔티티",
  "share_plan": {
   "ref_plan": "prev",
   "prev_anchor_tag": "S52sh1"
  }
 },
 "S56sh1::cine": {
  "applied": true,
  "attempted_at": "2026-09-09T07:21:10.479046+00:00",
  "fingerprint": "165aa426598428ea5bc736b5c2b14a105903b03508c7ac162d833b20b2509da6",
  "fingerprint_version": 2,
  "provider": "mai",
  "endpoint": "openrouter/chat-completions",
  "model": "microsoft/mai-image-2.6",
  "pack": "24.202608252115",
  "source_file": "S56sh1_sel.png",
  "source_sha256": "f67adc9e48c5c76af100b38c8842e43eaeeecd043a813a8e5ce0fca29c33e00a",
  "file": "S56sh1_cine.png",
  "staged_sha256": "f1793be5f50148c4fc70ceb2fde5cb09a6eb6cd483f0e10441e3de6132039dd6",
  "latency_ms": 29545
 },
 "S56sh31::signage": {
  "fp": "0b4ef091b39f154b",
  "inscriptions": [],
  "cues": [],
  "dropped": []
 },
 "S56sh31::bgfirst_bg": {
  "input_fingerprint": "098c662a3d6ce048",
  "prompt": "Create the EMPTY BACKGROUND PLATE for one film shot — NO PEOPLE, no figures, no body parts, no sketch lines, no arrows anywhere.\n\"Empty\" means no people only: KEEP the location's inherent occupants and stock that define the place — animals in an animal shelter, pen or farm, goods in a market, moored boats in a harbour — unless the shot text explicitly removes them.\nThe FIRST attached image is a thin-line storyboard sketch: use ONLY its camera angle, horizon, perspective and the placement/size of buildings and set masses — ignore the sketched people and arrows entirely. The SECOND attached image (LOCATION PHOTOGRAPH) is the real place: take its architecture, materials, signage and fixed features, and RE-PROJECT them into the sketch's camera. If the photograph's camera differs from the sketch's, the sketch's camera wins.\nHUMAN-SCALE CALIBRATION: derive every structure's true size from human-scale elements — a door ≈ 2m, a window ≈ 1–1.5m wide, one storey ≈ 2.5–3m; never inflate a small structure or shrink a large one.\n\nSHOT TEXT this background must serve (Korean): 가슴에 피가 번진 채 두 눈을 부릅뜨고 몸이 뻣뻣하게 굳어버린 임정현의 충격받은 정면.\n\nLOCATION (lock): Against the exterior wall of a mountain-camp hut, beside rock cover and the exposed firing lanes across the clearing.\n\nTIME OF DAY (lock): day.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: close-up\n- KEY BACKGROUND ELEMENTS: 산채 움막 (Immediately behind 임정현 as his wounded body stops) — Its rear exterior is seen obliquely beside his shoulder; used as A narrow edge identifies his shelter and anticipates the surface against which he will sink; 바위 틈 (Open gap beyond his position) — The gap is viewed obliquely past 임정현; used as A small background opening preserves the established observation axis without competing with his face and chest.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Natural daylight and restrained contrast keep his eyes and the spreading blood readable without theatrical lighting or heightened saturation.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nRender ONE photorealistic empty location photograph, 16:9, that this shot can be staged inside later. No readable writing anywhere: surfaces that would carry writing may be present, but stage any wording out of legibility — an oblique angle, distance, shallow focus. No captions, watermarks or overlay text.",
  "effective_prompt": "Create the EMPTY BACKGROUND PLATE for one film shot — NO PEOPLE, no figures, no body parts, no sketch lines, no arrows anywhere.\n\"Empty\" means no people only: KEEP the location's inherent occupants and stock that define the place — animals in an animal shelter, pen or farm, goods in a market, moored boats in a harbour — unless the shot text explicitly removes them.\nThe FIRST attached image is a thin-line storyboard sketch: use ONLY its camera angle, horizon, perspective and the placement/size of buildings and set masses — ignore the sketched people and arrows entirely. The SECOND attached image (LOCATION PHOTOGRAPH) is the real place: take its architecture, materials, signage and fixed features, and RE-PROJECT them into the sketch's camera. If the photograph's camera differs from the sketch's, the sketch's camera wins.\nHUMAN-SCALE CALIBRATION: derive every structure's true size from human-scale elements — a door ≈ 2m, a window ≈ 1–1.5m wide, one storey ≈ 2.5–3m; never inflate a small structure or shrink a large one.\n\nSHOT TEXT this background must serve (Korean): 가슴에 피가 번진 채 두 눈을 부릅뜨고 몸이 뻣뻣하게 굳어버린 임정현의 충격받은 정면.\n\nLOCATION (lock): Against the exterior wall of a mountain-camp hut, beside rock cover and the exposed firing lanes across the clearing.\n\nTIME OF DAY (lock): day.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: close-up\n- KEY BACKGROUND ELEMENTS: 산채 움막 (Immediately behind 임정현 as his wounded body stops) — Its rear exterior is seen obliquely beside his shoulder; used as A narrow edge identifies his shelter and anticipates the surface against which he will sink; 바위 틈 (Open gap beyond his position) — The gap is viewed obliquely past 임정현; used as A small background opening preserves the established observation axis without competing with his face and chest.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Natural daylight and restrained contrast keep his eyes and the spreading blood readable without theatrical lighting or heightened saturation.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nRender ONE photorealistic empty location photograph, 16:9, that this shot can be staged inside later. No readable writing anywhere: surfaces that would carry writing may be present, but stage any wording out of legibility — an oblique angle, distance, shallow focus. No captions, watermarks or overlay text.",
  "bg_path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/db00517b-165b-4005-a287-f28936b59b95/images/272bd7af-9294-4490-98c6-88a40e9deacc/scene/recipe/S56sh31__bgfirst_bg.png",
  "asset_id": "beccce91-60f6-4c00-9897-46000932dd0c",
  "input_asset_ids": [
   "bc76a30b-f151-4a0a-a91e-af39885e2129",
   "8c17c7ae-efff-4bb5-b88c-7c5528f00716"
  ]
 },
 "S56sh31": {
  "input_fingerprint": "15f44b65f6db2cd0",
  "prompt": "Create ONE FINAL photorealistic live-action film still of the moment below — early 20th-century Korea, Jeolla region; all people are Korean unless explicitly stated otherwise. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 가슴에 피가 번진 채 두 눈을 부릅뜨고 몸이 뻣뻣하게 굳어버린 임정현의 충격받은 정면.\n\nLOCATION (lock): Against the exterior wall of a mountain-camp hut, beside rock cover and the exposed firing lanes across the clearing. The shot takes place here — the attached LOCATION STRUCTURE PHOTOGRAPH is the single authority for this exact place — its fixed structure and permanent site details are LOCKED to it. No separate location photograph exists for this place. Build everything else strictly from the location text above and the shot text; the layout sketch (when attached) governs framing and placement only, and the shot text governs time of day, lighting and action.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: close-up\n- KEY BACKGROUND ELEMENTS: 산채 움막 (Immediately behind 임정현 as his wounded body stops) — Its rear exterior is seen obliquely beside his shoulder; used as A narrow edge identifies his shelter and anticipates the surface against which he will sink; 바위 틈 (Open gap beyond his position) — The gap is viewed obliquely past 임정현; used as A small background opening preserves the established observation axis without competing with his face and chest.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Natural daylight and restrained contrast keep his eyes and the spreading blood readable without theatrical lighting or heightened saturation.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The cauldron is overturned, rice bowls are shattered and rice is scattered in the dirt; a stone has been lifted to expose a cloth-wrapped ammunition bundle. 임정현: He is against the hut with dirt on his face and a fresh bleeding chest gunshot wound.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 임정현 (한국인 남성, 성인 나이대, 자연스러운 검은 머리) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.",
  "roll_prompts": {
   "A": "Stage the shot. The FIRST attached image (SHOT BACKGROUND) is the finished empty background of this shot — keep it EXACTLY: its camera, perspective, architecture, lighting and every fixed feature stay untouched. The SECOND attached image (LAYOUT SKETCH) tells you ONLY where the people go: each sketched person's position, screen size, pose and the gaze/motion arrows. Ignore the sketch's background lines. The CHARACTER REFERENCE photographs show the real people.\nPlace the real people into the background at exactly the sketched positions, sizes and poses, following the arrow directions. No sketch lines or arrows may remain.\n\nCreate ONE FINAL photorealistic live-action film still of the moment below — early 20th-century Korea, Jeolla region; all people are Korean unless explicitly stated otherwise. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 가슴에 피가 번진 채 두 눈을 부릅뜨고 몸이 뻣뻣하게 굳어버린 임정현의 충격받은 정면.\n\nLOCATION (lock): Against the exterior wall of a mountain-camp hut, beside rock cover and the exposed firing lanes across the clearing. The shot takes place here — the FIRST attached image (SHOT BACKGROUND) is this exact place, already built: its ground, structures, horizon, materials and lighting are the finished truth of this location and must not be redesigned or replaced. No location photograph is attached — read the place from that image alone, and add no scenery, structure, vehicle or fixture that it does not already show. This lock governs the place only; the figures in the shot follow the staging and pose instructions.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: close-up\n- KEY BACKGROUND ELEMENTS: 산채 움막 (Immediately behind 임정현 as his wounded body stops) — Its rear exterior is seen obliquely beside his shoulder; used as A narrow edge identifies his shelter and anticipates the surface against which he will sink; 바위 틈 (Open gap beyond his position) — The gap is viewed obliquely past 임정현; used as A small background opening preserves the established observation axis without competing with his face and chest.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Natural daylight and restrained contrast keep his eyes and the spreading blood readable without theatrical lighting or heightened saturation.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The cauldron is overturned, rice bowls are shattered and rice is scattered in the dirt; a stone has been lifted to expose a cloth-wrapped ammunition bundle. 임정현: He is against the hut with dirt on his face and a fresh bleeding chest gunshot wound.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 임정현 (한국인 남성, 성인 나이대, 자연스러운 검은 머리) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.",
   "B": "Create ONE FINAL photorealistic live-action film still of the moment below — early 20th-century Korea, Jeolla region; all people are Korean unless explicitly stated otherwise. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 가슴에 피가 번진 채 두 눈을 부릅뜨고 몸이 뻣뻣하게 굳어버린 임정현의 충격받은 정면.\n\nLOCATION (lock): Against the exterior wall of a mountain-camp hut, beside rock cover and the exposed firing lanes across the clearing. The shot takes place here — the attached LOCATION STRUCTURE PHOTOGRAPH is the single authority for this exact place — its fixed structure and permanent site details are LOCKED to it. No separate location photograph exists for this place. Build everything else strictly from the location text above and the shot text; the layout sketch (when attached) governs framing and placement only, and the shot text governs time of day, lighting and action.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: close-up\n- KEY BACKGROUND ELEMENTS: 산채 움막 (Immediately behind 임정현 as his wounded body stops) — Its rear exterior is seen obliquely beside his shoulder; used as A narrow edge identifies his shelter and anticipates the surface against which he will sink; 바위 틈 (Open gap beyond his position) — The gap is viewed obliquely past 임정현; used as A small background opening preserves the established observation axis without competing with his face and chest.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Natural daylight and restrained contrast keep his eyes and the spreading blood readable without theatrical lighting or heightened saturation.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The cauldron is overturned, rice bowls are shattered and rice is scattered in the dirt; a stone has been lifted to expose a cloth-wrapped ammunition bundle. 임정현: He is against the hut with dirt on his face and a fresh bleeding chest gunshot wound.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 임정현 (한국인 남성, 성인 나이대, 자연스러운 검은 머리) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.\n\nCOMPOSITION VARIATION (this candidate only): keep the place, the people and the moment exactly as specified, and keep from the CAMERA & FRAME contract its subject scale, screen placement and key background elements exactly. For this candidate only, its camera angle and height are deliberately left open: choose a camera position distinctly different from the obvious one, as a film director picking a second setup on the same blocking. Stay true to any explicit framing the SHOT TEXT itself dictates."
  },
  "roll_refs": {
   "A": [
    {
     "label": "SHOT BACKGROUND",
     "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/db00517b-165b-4005-a287-f28936b59b95/images/272bd7af-9294-4490-98c6-88a40e9deacc/scene/recipe/S56sh31__bgfirst_bg.png",
     "asset_id": "beccce91-60f6-4c00-9897-46000932dd0c",
     "role": "bgfirst_bg"
    },
    {
     "label": "LAYOUT SKETCH (people placement only)",
     "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/db00517b-165b-4005-a287-f28936b59b95/images/272bd7af-9294-4490-98c6-88a40e9deacc/conti/conti_S56sh31.png",
     "asset_id": "bc76a30b-f151-4a0a-a91e-af39885e2129",
     "role": "conti_light"
    },
    {
     "label": "CHARACTER REFERENCE — 임정현: the exact person appearing in this shot; match face, hair and build exactly.",
     "path": "<bytes:1434550>",
     "asset_id": "e79b9a4a-9b27-47f3-ba1e-89801d92223a",
     "role": "character_ref"
    }
   ],
   "B": [
    {
     "label": "LOCATION STRUCTURE PHOTOGRAPH — the confirmed photograph of this exact place and its fixed structure: it is the SINGLE authority for the location, the structure's shape, proportions, materials, colors, openings and every permanent site detail. Never copy its camera framing, time of day or lighting — the shot text is the authority for those.",
     "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/db00517b-165b-4005-a287-f28936b59b95/images/background_chain/seed_bg_mountain_camp_sel.png",
     "asset_id": "8c17c7ae-efff-4bb5-b88c-7c5528f00716",
     "role": "location_seed_bg"
    },
    {
     "label": "CHARACTER REFERENCE — 임정현: the exact person appearing in this shot; match face, hair and build exactly.",
     "path": "<bytes:1434550>",
     "asset_id": "e79b9a4a-9b27-47f3-ba1e-89801d92223a",
     "role": "character_ref"
    }
   ]
  },
  "shot_run_spend_attempt_count": 1,
  "judge_flip_skipped": "judge_fn_owns_order",
  "readings": [
   {
    "label": "A",
    "direction": "시선이 화면 우측 밖을 향하고 있어 프롬프트가 요구한 충격받은 정면 구도가 아님.",
    "built_space": "산채 움막 벽을 등지고 앉아 있는 각도이며, 배경에 바위 틈과 엎어진 가마솥, 쌀, 보따리 등이 보임.",
    "entities": "임정현의 얼굴과 의상은 레퍼런스와 유사하나, 화면 우측 하단의 손과 탄약 보따리, 바위, 흙바닥이 실사가 아닌 2D 일러스트로 그려져 있음.",
    "hard_violations": [
     "[gemini-pro] 평면적인 그래픽 오버레이/비실사 요소 포함 (우측의 손, 바위, 탄약 보따리, 바닥이 2D 애니메이션 스타일로 그려짐)",
     "[gpt-high] 배경의 온전한 갈색 그릇이 바닥 위에 떠 있으며 손·받침·낙하를 유발한 동작 등 지지나 운동의 근거가 없다.",
     "[gpt-high] 오른쪽 아래의 팔·손과 돌·꾸러미가 사진 속에 삽화를 붙인 듯한 선과 면으로 표현되어, 하나의 실사 영화 스틸이라는 필수 조건을 위반한다."
    ],
    "physics": "화면 우측의 그려진 손과 팔이 인체의 실제적인 연결 없이 허공에 덧씌워진 것처럼 어색하게 떠 있음."
   },
   {
    "label": "B",
    "direction": "두 눈을 부릅뜨고 카메라 렌즈 방향(정면)을 똑바로 응시하고 있음.",
    "built_space": "인물 바로 뒤에 진흙으로 된 산채 움막 벽이 위치하며, 인물의 우측 어깨 너머 사선으로 바위 틈과 오솔길이 정확히 묘사됨.",
    "entities": "임정현의 인상착의가 레퍼런스와 일치하며, 흙먼지가 묻은 얼굴과 가슴에 피가 번진 총상이 사실적으로 묘사됨.",
    "hard_violations": [],
    "physics": "충격을 받아 몸이 뻣뻣하게 굳은 채 움막 벽에 등을 기대고 있는 자세가 물리적으로 자연스럽게 지탱됨."
   }
  ],
  "cross_model_order": {
   "policy": "cross_model_order_v2_slot0_fwd_slot1_rev_parallel2",
   "slots": [
    {
     "model": "gemini-pro",
     "order": "forward",
     "display_to_canonical": {
      "A": "A",
      "B": "B"
     },
     "raw": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 10,
        "verdict_ko": "카메라 정면을 향해 두 눈을 부릅뜨고 굳어버린 임정현의 표정과 가슴의 총상, 지정된 배경(움막 벽, 바위 틈)을 클로즈업 앵글로 완벽하고 사실적으로 구현했습니다."
       },
       {
        "label": "A",
        "score": 0,
        "verdict_ko": "정면 구도를 벗어났으며, 화면 우측의 손과 사물들이 실사가 아닌 2D 그래픽으로 그려져 있어 심각한 하드 위반입니다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "시선이 화면 우측 밖을 향하고 있어 프롬프트가 요구한 충격받은 정면 구도가 아님.",
        "built_space": "산채 움막 벽을 등지고 앉아 있는 각도이며, 배경에 바위 틈과 엎어진 가마솥, 쌀, 보따리 등이 보임.",
        "entities": "임정현의 얼굴과 의상은 레퍼런스와 유사하나, 화면 우측 하단의 손과 탄약 보따리, 바위, 흙바닥이 실사가 아닌 2D 일러스트로 그려져 있음.",
        "hard_violations": [
         "평면적인 그래픽 오버레이/비실사 요소 포함 (우측의 손, 바위, 탄약 보따리, 바닥이 2D 애니메이션 스타일로 그려짐)"
        ],
        "physics": "화면 우측의 그려진 손과 팔이 인체의 실제적인 연결 없이 허공에 덧씌워진 것처럼 어색하게 떠 있음."
       },
       {
        "label": "B",
        "direction": "두 눈을 부릅뜨고 카메라 렌즈 방향(정면)을 똑바로 응시하고 있음.",
        "built_space": "인물 바로 뒤에 진흙으로 된 산채 움막 벽이 위치하며, 인물의 우측 어깨 너머 사선으로 바위 틈과 오솔길이 정확히 묘사됨.",
        "entities": "임정현의 인상착의가 레퍼런스와 일치하며, 흙먼지가 묻은 얼굴과 가슴에 피가 번진 총상이 사실적으로 묘사됨.",
        "hard_violations": [],
        "physics": "충격을 받아 몸이 뻣뻣하게 굳은 채 움막 벽에 등을 기대고 있는 자세가 물리적으로 자연스럽게 지탱됨."
       }
      ],
      "all_candidates_fail": false
     },
     "normalized": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 10,
        "verdict_ko": "카메라 정면을 향해 두 눈을 부릅뜨고 굳어버린 임정현의 표정과 가슴의 총상, 지정된 배경(움막 벽, 바위 틈)을 클로즈업 앵글로 완벽하고 사실적으로 구현했습니다."
       },
       {
        "label": "A",
        "score": 0,
        "verdict_ko": "정면 구도를 벗어났으며, 화면 우측의 손과 사물들이 실사가 아닌 2D 그래픽으로 그려져 있어 심각한 하드 위반입니다."
       }
      ],
      "all_candidates_fail": false,
      "readings": [
       {
        "label": "A",
        "direction": "시선이 화면 우측 밖을 향하고 있어 프롬프트가 요구한 충격받은 정면 구도가 아님.",
        "built_space": "산채 움막 벽을 등지고 앉아 있는 각도이며, 배경에 바위 틈과 엎어진 가마솥, 쌀, 보따리 등이 보임.",
        "entities": "임정현의 얼굴과 의상은 레퍼런스와 유사하나, 화면 우측 하단의 손과 탄약 보따리, 바위, 흙바닥이 실사가 아닌 2D 일러스트로 그려져 있음.",
        "hard_violations": [
         "평면적인 그래픽 오버레이/비실사 요소 포함 (우측의 손, 바위, 탄약 보따리, 바닥이 2D 애니메이션 스타일로 그려짐)"
        ],
        "physics": "화면 우측의 그려진 손과 팔이 인체의 실제적인 연결 없이 허공에 덧씌워진 것처럼 어색하게 떠 있음."
       },
       {
        "label": "B",
        "direction": "두 눈을 부릅뜨고 카메라 렌즈 방향(정면)을 똑바로 응시하고 있음.",
        "built_space": "인물 바로 뒤에 진흙으로 된 산채 움막 벽이 위치하며, 인물의 우측 어깨 너머 사선으로 바위 틈과 오솔길이 정확히 묘사됨.",
        "entities": "임정현의 인상착의가 레퍼런스와 일치하며, 흙먼지가 묻은 얼굴과 가슴에 피가 번진 총상이 사실적으로 묘사됨.",
        "hard_violations": [],
        "physics": "충격을 받아 몸이 뻣뻣하게 굳은 채 움막 벽에 등을 기대고 있는 자세가 물리적으로 자연스럽게 지탱됨."
       }
      ]
     },
     "ok": true
    },
    {
     "model": "gpt-high",
     "order": "reverse",
     "display_to_canonical": {
      "A": "B",
      "B": "A"
     },
     "raw": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 6,
        "verdict_ko": "정면의 부릅뜬 눈과 굳은 몸, 가슴의 혈흔은 맞지만, 클로즈업보다 넓은 구도와 과도한 벽 노출, 망건 누락 및 부족한 얼굴 오염이 아쉽다."
       },
       {
        "label": "B",
        "score": 1,
        "verdict_ko": "정면 클로즈업 대신 비스듬한 상반신과 소품을 넓게 보여주며, 공중에 뜬 그릇과 삽화처럼 합성된 팔·소품 때문에 실사 장면으로 사용할 수 없다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "얼굴과 가슴이 거의 정면을 향하고 두 눈은 카메라 부근의 화면 밖을 응시한다. 특정 상대는 보이지 않으며, 충격받은 정면이라는 지시와 맞는다. 무기나 겨누는 물체는 없다.",
        "built_space": "인물 뒤로 흙벽 한 면이 화면 대부분을 차지하고, 오른쪽에 수직 목재 기둥 하나와 위쪽의 짧은 처마가 보인다. 벽에는 가로로 묶인 짚 다발이 있다. 오른쪽 바위 사이 통로에는 가로 목재들이 보인다. 인물은 벽 바로 앞에 있지만, 움막을 어깨 옆의 좁고 비스듬한 가장자리로 보여달라는 지시와 달리 벽이 넓고 정면으로 드러난다. 바위 틈도 작은 배경 개구부보다 크게 보인다. 참조의 흙·목재·초가 재료는 대체로 이어지지만 정확한 후면 구조의 일치는 확인하기 어렵다.",
        "entities": "성인 한국인 남성으로 설정과 부합하는 인물 한 명만 보인다. 검은 머리와 얼굴 윤곽, 낡은 담색 저고리는 참조와 대체로 유사하지만 검은 망건과 그 장식이 없고 머리도 더 단정하다. 얼굴의 흙과 상처는 참조 및 지시보다 훨씬 약하다. 왼쪽 가슴에는 상처와 아래로 번진 붉은 혈흔이 있으나 젖은 피의 광택은 약하다. 눈은 정상적인 홍채와 동공을 유지한 채 크게 떠 있다. 가마솥·그릇·쌀·탄약 꾸러미는 프레임 밖이므로 누락으로 판단하지 않는다. 읽을 수 있는 글자는 없다.",
        "hard_violations": [],
        "physics": "몸통은 수직으로 서 있고 등과 어깨가 바로 뒤 벽에 기대거나 닿을 수 있는 위치다. 발은 프레임 밖이지만 공중에 떠 있다는 증거는 없다. 팔은 아래로 내려와 있으며 몸은 다소 대칭적으로 굳어 있어, 움직임 중 멈춘 순간보다는 정자세에 가깝다. 혈흔은 옷 표면을 따라 아래로 번지고, 벽의 짚 다발은 묶여 고정되어 있다."
       },
       {
        "label": "B",
        "direction": "얼굴과 시선은 화면 오른쪽의 보이지 않는 대상을 향한다. 눈은 크게 떠 있지만 얼굴이 비스듬해 요청한 충격받은 정면이 아니다. 피 묻은 손은 가슴 앞에서 손바닥을 카메라 쪽으로 들고 있고, 오른쪽 아래의 별도 팔 형태는 화면 오른쪽으로 뻗는다. 무기나 조준 대상은 없다.",
        "built_space": "왼쪽에는 목재 틀과 가로대가 있는 어두운 개구부 하나, 인물 뒤에는 흙벽, 위에는 크게 드러난 초가 처마가 있다. 오른쪽 배경에는 바위 무리 두 쪽 사이의 틈과 목책이 보인다. 인물의 등은 벽에 붙어 있어 위치 자체는 가능하지만, 움막 가장자리와 바위 틈뿐 아니라 넓은 마당과 소품까지 보여 주어 지정된 클로즈업을 벗어난다. 참조와 재료는 유사하나 후면을 좁게 보여주는 구도는 아니다.",
        "entities": "검은 머리의 성인 한국인 남성 한 명의 얼굴과 몸통이 보이며, 담색 저고리와 얼굴의 흙은 지시에 부합한다. 참조의 검은 망건은 없고 얼굴도 비스듬하게 제시된다. 가슴에는 젖은 피와 큰 상처가 보인다. 배경에는 온전한 갈색 그릇 하나, 여러 그릇 조각과 흩어진 쌀이 보이지만 뒤집힌 가마솥은 식별되지 않는다. 오른쪽 아래에는 돌과 묶인 꾸러미가 있으나 내용물이 탄약인지는 확인되지 않는다. 그 돌·꾸러미와 전경의 팔은 실사보다 윤곽선이 뚜렷한 삽화처럼 보인다. 읽을 수 있는 글자는 없다.",
        "hard_violations": [
         "배경의 온전한 갈색 그릇이 바닥 위에 떠 있으며 손·받침·낙하를 유발한 동작 등 지지나 운동의 근거가 없다.",
         "오른쪽 아래의 팔·손과 돌·꾸러미가 사진 속에 삽화를 붙인 듯한 선과 면으로 표현되어, 하나의 실사 영화 스틸이라는 필수 조건을 위반한다."
        ],
        "physics": "인물의 등은 흙벽에 지지되고, 가슴 앞에 든 피 묻은 손은 소매와 팔에 연결되어 있어 유지 가능한 자세다. 오른쪽 아래 전경의 팔은 인물 몸통까지의 연결 경로가 보이지 않아 귀속과 자세가 불명확하다. 그릇 조각과 쌀은 땅에 놓여 있고 큰 돌과 꾸러미도 바닥에 닿아 있다. 반면 온전한 갈색 그릇은 아무것도 받치지 않으며, 이미 그릇이 깨져 흩어진 상태라는 지시만으로 이 공중 위치를 설명할 수 없다."
       }
      ],
      "all_candidates_fail": true
     },
     "normalized": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 6,
        "verdict_ko": "정면의 부릅뜬 눈과 굳은 몸, 가슴의 혈흔은 맞지만, 클로즈업보다 넓은 구도와 과도한 벽 노출, 망건 누락 및 부족한 얼굴 오염이 아쉽다."
       },
       {
        "label": "A",
        "score": 1,
        "verdict_ko": "정면 클로즈업 대신 비스듬한 상반신과 소품을 넓게 보여주며, 공중에 뜬 그릇과 삽화처럼 합성된 팔·소품 때문에 실사 장면으로 사용할 수 없다."
       }
      ],
      "all_candidates_fail": true,
      "readings": [
       {
        "label": "B",
        "direction": "얼굴과 가슴이 거의 정면을 향하고 두 눈은 카메라 부근의 화면 밖을 응시한다. 특정 상대는 보이지 않으며, 충격받은 정면이라는 지시와 맞는다. 무기나 겨누는 물체는 없다.",
        "built_space": "인물 뒤로 흙벽 한 면이 화면 대부분을 차지하고, 오른쪽에 수직 목재 기둥 하나와 위쪽의 짧은 처마가 보인다. 벽에는 가로로 묶인 짚 다발이 있다. 오른쪽 바위 사이 통로에는 가로 목재들이 보인다. 인물은 벽 바로 앞에 있지만, 움막을 어깨 옆의 좁고 비스듬한 가장자리로 보여달라는 지시와 달리 벽이 넓고 정면으로 드러난다. 바위 틈도 작은 배경 개구부보다 크게 보인다. 참조의 흙·목재·초가 재료는 대체로 이어지지만 정확한 후면 구조의 일치는 확인하기 어렵다.",
        "entities": "성인 한국인 남성으로 설정과 부합하는 인물 한 명만 보인다. 검은 머리와 얼굴 윤곽, 낡은 담색 저고리는 참조와 대체로 유사하지만 검은 망건과 그 장식이 없고 머리도 더 단정하다. 얼굴의 흙과 상처는 참조 및 지시보다 훨씬 약하다. 왼쪽 가슴에는 상처와 아래로 번진 붉은 혈흔이 있으나 젖은 피의 광택은 약하다. 눈은 정상적인 홍채와 동공을 유지한 채 크게 떠 있다. 가마솥·그릇·쌀·탄약 꾸러미는 프레임 밖이므로 누락으로 판단하지 않는다. 읽을 수 있는 글자는 없다.",
        "hard_violations": [],
        "physics": "몸통은 수직으로 서 있고 등과 어깨가 바로 뒤 벽에 기대거나 닿을 수 있는 위치다. 발은 프레임 밖이지만 공중에 떠 있다는 증거는 없다. 팔은 아래로 내려와 있으며 몸은 다소 대칭적으로 굳어 있어, 움직임 중 멈춘 순간보다는 정자세에 가깝다. 혈흔은 옷 표면을 따라 아래로 번지고, 벽의 짚 다발은 묶여 고정되어 있다."
       },
       {
        "label": "A",
        "direction": "얼굴과 시선은 화면 오른쪽의 보이지 않는 대상을 향한다. 눈은 크게 떠 있지만 얼굴이 비스듬해 요청한 충격받은 정면이 아니다. 피 묻은 손은 가슴 앞에서 손바닥을 카메라 쪽으로 들고 있고, 오른쪽 아래의 별도 팔 형태는 화면 오른쪽으로 뻗는다. 무기나 조준 대상은 없다.",
        "built_space": "왼쪽에는 목재 틀과 가로대가 있는 어두운 개구부 하나, 인물 뒤에는 흙벽, 위에는 크게 드러난 초가 처마가 있다. 오른쪽 배경에는 바위 무리 두 쪽 사이의 틈과 목책이 보인다. 인물의 등은 벽에 붙어 있어 위치 자체는 가능하지만, 움막 가장자리와 바위 틈뿐 아니라 넓은 마당과 소품까지 보여 주어 지정된 클로즈업을 벗어난다. 참조와 재료는 유사하나 후면을 좁게 보여주는 구도는 아니다.",
        "entities": "검은 머리의 성인 한국인 남성 한 명의 얼굴과 몸통이 보이며, 담색 저고리와 얼굴의 흙은 지시에 부합한다. 참조의 검은 망건은 없고 얼굴도 비스듬하게 제시된다. 가슴에는 젖은 피와 큰 상처가 보인다. 배경에는 온전한 갈색 그릇 하나, 여러 그릇 조각과 흩어진 쌀이 보이지만 뒤집힌 가마솥은 식별되지 않는다. 오른쪽 아래에는 돌과 묶인 꾸러미가 있으나 내용물이 탄약인지는 확인되지 않는다. 그 돌·꾸러미와 전경의 팔은 실사보다 윤곽선이 뚜렷한 삽화처럼 보인다. 읽을 수 있는 글자는 없다.",
        "hard_violations": [
         "배경의 온전한 갈색 그릇이 바닥 위에 떠 있으며 손·받침·낙하를 유발한 동작 등 지지나 운동의 근거가 없다.",
         "오른쪽 아래의 팔·손과 돌·꾸러미가 사진 속에 삽화를 붙인 듯한 선과 면으로 표현되어, 하나의 실사 영화 스틸이라는 필수 조건을 위반한다."
        ],
        "physics": "인물의 등은 흙벽에 지지되고, 가슴 앞에 든 피 묻은 손은 소매와 팔에 연결되어 있어 유지 가능한 자세다. 오른쪽 아래 전경의 팔은 인물 몸통까지의 연결 경로가 보이지 않아 귀속과 자세가 불명확하다. 그릇 조각과 쌀은 땅에 놓여 있고 큰 돌과 꾸러미도 바닥에 닿아 있다. 반면 온전한 갈색 그릇은 아무것도 받치지 않으며, 이미 그릇이 깨져 흩어진 상태라는 지시만으로 이 공중 위치를 설명할 수 없다."
       }
      ]
     },
     "ok": true
    }
   ],
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "slot_winner_match": true,
   "slot_winner": {
    "gemini-pro": "B",
    "gpt-high": "B"
   },
   "route": "cross_slot_agree"
  },
  "dual": {
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "normalized": {
    "A": 0.167,
    "B": 2.0
   },
   "adjusted": {
    "A": -0.083,
    "B": 2.0
   },
   "violations": {
    "A": [
     "[gemini-pro] 평면적인 그래픽 오버레이/비실사 요소 포함 (우측의 손, 바위, 탄약 보따리, 바닥이 2D 애니메이션 스타일로 그려짐)",
     "[gpt-high] 배경의 온전한 갈색 그릇이 바닥 위에 떠 있으며 손·받침·낙하를 유발한 동작 등 지지나 운동의 근거가 없다.",
     "[gpt-high] 오른쪽 아래의 팔·손과 돌·꾸러미가 사진 속에 삽화를 붙인 듯한 선과 면으로 표현되어, 하나의 실사 영화 스틸이라는 필수 조건을 위반한다."
    ]
   },
   "per_model_winner": {
    "gemini-pro": "B",
    "gpt-high": "B"
   },
   "agreed": true
  },
  "totals": {
   "B": 2000,
   "A": -83
  },
  "selected": "B",
  "ranking": [
   "B",
   "A"
  ],
  "verdicts": [
   {
    "label": "B",
    "score": 2000,
    "verdict_ko": "카메라 정면을 향해 두 눈을 부릅뜨고 굳어버린 임정현의 표정과 가슴의 총상, 지정된 배경(움막 벽, 바위 틈)을 클로즈업 앵글로 완벽하고 사실적으로 구현했습니다."
   },
   {
    "label": "A",
    "score": -83,
    "verdict_ko": "정면 구도를 벗어났으며, 화면 우측의 손과 사물들이 실사가 아닌 2D 그래픽으로 그려져 있어 심각한 하드 위반입니다.  ★위반: [gemini-pro] 평면적인 그래픽 오버레이/비실사 요소 포함 (우측의 손, 바위, 탄약 보따리, 바닥이 2D 애니메이션 스타일로 그려짐) / [gpt-high] 배경의 온전한 갈색 그릇이 바닥 위에 떠 있으며 손·받침·낙하를 유발한 동작 등 지지나 운동의 근거가 없다. / [gpt-high] 오른쪽 아래의 팔·손과 돌·꾸러미가 사진 속에 삽화를 붙인 듯한 선과 면으로 표현되어, 하나의 실사 영화 스틸이라는 필수 조건을 위반한다."
   }
  ],
  "refs": [
   {
    "label": "LOCATION STRUCTURE PHOTOGRAPH — the confirmed photograph of this exact place and its fixed structure: it is the SINGLE authority for the location, the structure's shape, proportions, materials, colors, openings and every permanent site detail. Never copy its camera framing, time of day or lighting — the shot text is the authority for those.",
    "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/db00517b-165b-4005-a287-f28936b59b95/images/background_chain/seed_bg_mountain_camp_sel.png",
    "asset_id": "8c17c7ae-efff-4bb5-b88c-7c5528f00716",
    "role": "location_seed_bg"
   },
   {
    "label": "CHARACTER REFERENCE — 임정현: the exact person appearing in this shot; match face, hair and build exactly.",
    "path": "<bytes:1434550>",
    "asset_id": "e79b9a4a-9b27-47f3-ba1e-89801d92223a",
    "role": "character_ref"
   }
  ],
  "critique_skipped": true,
  "shot_run_produced": false,
  "shot_run_uid": "06aa1438-e15a-7869-8469-cc65d7fe97ad",
  "bgfirst": {
   "bg_path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/db00517b-165b-4005-a287-f28936b59b95/images/272bd7af-9294-4490-98c6-88a40e9deacc/scene/recipe/S56sh31__bgfirst_bg.png",
   "bg_asset_id": "beccce91-60f6-4c00-9897-46000932dd0c",
   "bg_record_key": "S56sh31::bgfirst_bg",
   "chain_winner": false,
   "authority": "seed_bg"
  },
  "ref_mode": "seed-bg+엔티티 (2택1: 무콘티 승)",
  "share_plan": {
   "ref_plan": "background"
  },
  "lane_policy": "ab_select_ready"
 },
 "S56sh31::cine": {
  "applied": true,
  "attempted_at": "2026-09-09T07:24:42.107155+00:00",
  "fingerprint": "5ee75ea3575ed1ae4b00b621ec34c11536cd1bbce7307875f904d8926970153d",
  "fingerprint_version": 2,
  "provider": "mai",
  "endpoint": "openrouter/chat-completions",
  "model": "microsoft/mai-image-2.6",
  "pack": "24.202608252115",
  "source_file": "S56sh31_sel.png",
  "source_sha256": "a98c6ca2ad9d40b3bc8c7f90a9e171f25faca6c30d722bd024275e235aa6920b",
  "file": "S56sh31_cine.png",
  "staged_sha256": "a013cc6f5a8e71b3693b62d8579d616eac022adcceab2b79a2bb3254207bffe2",
  "latency_ms": 33394
 },
 "S56sh40::signage": {
  "fp": "e2f754438033e6eb",
  "inscriptions": [],
  "cues": [],
  "dropped": []
 },
 "S56sh40": {
  "input_fingerprint": "c81fe874007c1019",
  "prompt": "Create ONE FINAL photorealistic live-action film still of the moment below — early 20th-century Korea, Jeolla region; all people are Korean unless explicitly stated otherwise. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 살아남은 의병들을 향해 입을 크게 벌린 채 매섭게 호통치는 오주일의 얼굴.\n\nLOCATION (lock): The low defensive position behind rocks in the mountain camp's outdoor clearing, facing the surviving men. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: close-up\n- KEY BACKGROUND ELEMENTS: 엄폐 바위 (Between 오주일's lowered position and the exposed interval) — The sheltered side and a short upper edge are visible beside his shoulder; used as A modest lower-edge presence explains his compressed posture without obscuring the command.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Natural daylight with muted saturation and naturalistic facial contrast retains the grief beneath the force of his command.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The people visible in that photo are NOT in this shot: never carry their faces, bodies or clothing onto anyone here. Each person in THIS shot is defined solely by the PEOPLE list, their own CHARACTER REFERENCE images and their wardrobe notes. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The overturned cauldron, broken rice bowls and scattered rice remain amid bullet-struck dirt; one ammunition cache is uncovered. 오주일: He remains low behind cover with reddened eyes.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 오주일 (한국인 남성, 성인 나이대, 자연스러운 검은 머리) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.",
  "roll_prompts": {
   "A": "Create ONE FINAL photorealistic live-action film still of the moment below — early 20th-century Korea, Jeolla region; all people are Korean unless explicitly stated otherwise. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 살아남은 의병들을 향해 입을 크게 벌린 채 매섭게 호통치는 오주일의 얼굴.\n\nLOCATION (lock): The low defensive position behind rocks in the mountain camp's outdoor clearing, facing the surviving men. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: close-up\n- KEY BACKGROUND ELEMENTS: 엄폐 바위 (Between 오주일's lowered position and the exposed interval) — The sheltered side and a short upper edge are visible beside his shoulder; used as A modest lower-edge presence explains his compressed posture without obscuring the command.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Natural daylight with muted saturation and naturalistic facial contrast retains the grief beneath the force of his command.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The people visible in that photo are NOT in this shot: never carry their faces, bodies or clothing onto anyone here. Each person in THIS shot is defined solely by the PEOPLE list, their own CHARACTER REFERENCE images and their wardrobe notes. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The overturned cauldron, broken rice bowls and scattered rice remain amid bullet-struck dirt; one ammunition cache is uncovered. 오주일: He remains low behind cover with reddened eyes.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 오주일 (한국인 남성, 성인 나이대, 자연스러운 검은 머리) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.",
   "B": "Create ONE FINAL photorealistic live-action film still of the moment below — early 20th-century Korea, Jeolla region; all people are Korean unless explicitly stated otherwise. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 살아남은 의병들을 향해 입을 크게 벌린 채 매섭게 호통치는 오주일의 얼굴.\n\nLOCATION (lock): The low defensive position behind rocks in the mountain camp's outdoor clearing, facing the surviving men. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: close-up\n- KEY BACKGROUND ELEMENTS: 엄폐 바위 (Between 오주일's lowered position and the exposed interval) — The sheltered side and a short upper edge are visible beside his shoulder; used as A modest lower-edge presence explains his compressed posture without obscuring the command.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Natural daylight with muted saturation and naturalistic facial contrast retains the grief beneath the force of his command.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The people visible in that photo are NOT in this shot: never carry their faces, bodies or clothing onto anyone here. Each person in THIS shot is defined solely by the PEOPLE list, their own CHARACTER REFERENCE images and their wardrobe notes. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The overturned cauldron, broken rice bowls and scattered rice remain amid bullet-struck dirt; one ammunition cache is uncovered. 오주일: He remains low behind cover with reddened eyes.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 오주일 (한국인 남성, 성인 나이대, 자연스러운 검은 머리) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.\n\nCOMPOSITION VARIATION (this candidate only): keep the place, the people and the moment exactly as specified, and keep from the CAMERA & FRAME contract its subject scale, screen placement and key background elements exactly. For this candidate only, its camera angle and height are deliberately left open: choose a camera position distinctly different from the obvious one, as a film director picking a second setup on the same blocking. Stay true to any explicit framing the SHOT TEXT itself dictates."
  },
  "shot_run_spend_attempt_count": 1,
  "readings": [
   {
    "label": "A",
    "direction": "화면 왼쪽 밖의 대상(의병들)을 향해 시선을 고정하고 입을 크게 벌려 호통을 치고 있습니다.",
    "built_space": "이전 샷과 동일한 산채 야외 공터의 엄폐 바위 뒤에 위치하며, 배경에 목책과 움집이 보이고 왼쪽 하단 참호 쪽에 열린 탄약 상자가 배치되어 있습니다.",
    "entities": "레퍼런스와 일치하는 외모와 복장을 한 오주일 단 한 명만 등장하며, 프롬프트에 지시된 대로 눈이 붉게 충혈되어 있습니다.",
    "hard_violations": [],
    "physics": "바위 뒤에 몸을 낮추고 상체를 앞으로 내민 채 지면에 의해 안정적으로 지탱되는 자연스러운 자세를 취하고 있습니다."
   },
   {
    "label": "B",
    "direction": "화면 왼쪽 밖의 대상을 향해 시선을 두고 입을 벌려 소리치고 있습니다.",
    "built_space": "엄폐 바위 사이 공간에 위치하며 배경으로 산채의 목책과 구조물들이 보입니다.",
    "entities": "오주일이 등장하나, 가슴에 '오주일'이라는 읽을 수 있는 글씨가 적힌 명찰이 있으며, 배경에 프롬프트에 없는 3명의 추가 인물(의병들)이 등장합니다.",
    "hard_violations": [
     "[gemini-pro] 지정된 인물 외에 프롬프트에 명시되지 않은 3명의 인물 추가 (PEOPLE 규정 위반)",
     "[gemini-pro] 가슴에 읽을 수 있는 텍스트(명찰) 포함 (No readable writing 규정 위반)",
     "[gpt-high] 화면에 허용된 오주일 외에 남성 세 명을 추가했다.",
     "[gpt-high] 군복 가슴 표찰에 읽을 수 있는 한글 일부가 노출돼 무문자 지시를 위반했다."
    ],
    "physics": "바위에 팔과 상체를 기대어 몸을 지탱하고 있는 자세입니다."
   }
  ],
  "cross_model_order": {
   "policy": "cross_model_order_v2_slot0_fwd_slot1_rev_parallel2",
   "slots": [
    {
     "model": "gemini-pro",
     "order": "forward",
     "display_to_canonical": {
      "A": "A",
      "B": "B"
     },
     "raw": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 10,
        "verdict_ko": "샷 텍스트와 모든 제약 조건(오주일 단독 등장, 붉어진 눈, 텍스트 없음, 열린 탄약 상자 배치 등)을 완벽하게 준수하여 캐릭터의 감정과 지정된 배경을 훌륭하게 구현했습니다."
       },
       {
        "label": "B",
        "score": 2,
        "verdict_ko": "프롬프트에서 엄격히 금지한 추가 인물들의 등장과 읽을 수 있는 텍스트(가슴의 명찰)가 포함되어 치명적인 규정 위반이 발생했습니다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "화면 왼쪽 밖의 대상(의병들)을 향해 시선을 고정하고 입을 크게 벌려 호통을 치고 있습니다.",
        "built_space": "이전 샷과 동일한 산채 야외 공터의 엄폐 바위 뒤에 위치하며, 배경에 목책과 움집이 보이고 왼쪽 하단 참호 쪽에 열린 탄약 상자가 배치되어 있습니다.",
        "entities": "레퍼런스와 일치하는 외모와 복장을 한 오주일 단 한 명만 등장하며, 프롬프트에 지시된 대로 눈이 붉게 충혈되어 있습니다.",
        "hard_violations": [],
        "physics": "바위 뒤에 몸을 낮추고 상체를 앞으로 내민 채 지면에 의해 안정적으로 지탱되는 자연스러운 자세를 취하고 있습니다."
       },
       {
        "label": "B",
        "direction": "화면 왼쪽 밖의 대상을 향해 시선을 두고 입을 벌려 소리치고 있습니다.",
        "built_space": "엄폐 바위 사이 공간에 위치하며 배경으로 산채의 목책과 구조물들이 보입니다.",
        "entities": "오주일이 등장하나, 가슴에 '오주일'이라는 읽을 수 있는 글씨가 적힌 명찰이 있으며, 배경에 프롬프트에 없는 3명의 추가 인물(의병들)이 등장합니다.",
        "hard_violations": [
         "지정된 인물 외에 프롬프트에 명시되지 않은 3명의 인물 추가 (PEOPLE 규정 위반)",
         "가슴에 읽을 수 있는 텍스트(명찰) 포함 (No readable writing 규정 위반)"
        ],
        "physics": "바위에 팔과 상체를 기대어 몸을 지탱하고 있는 자세입니다."
       }
      ],
      "all_candidates_fail": false
     },
     "normalized": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 10,
        "verdict_ko": "샷 텍스트와 모든 제약 조건(오주일 단독 등장, 붉어진 눈, 텍스트 없음, 열린 탄약 상자 배치 등)을 완벽하게 준수하여 캐릭터의 감정과 지정된 배경을 훌륭하게 구현했습니다."
       },
       {
        "label": "B",
        "score": 2,
        "verdict_ko": "프롬프트에서 엄격히 금지한 추가 인물들의 등장과 읽을 수 있는 텍스트(가슴의 명찰)가 포함되어 치명적인 규정 위반이 발생했습니다."
       }
      ],
      "all_candidates_fail": false,
      "readings": [
       {
        "label": "A",
        "direction": "화면 왼쪽 밖의 대상(의병들)을 향해 시선을 고정하고 입을 크게 벌려 호통을 치고 있습니다.",
        "built_space": "이전 샷과 동일한 산채 야외 공터의 엄폐 바위 뒤에 위치하며, 배경에 목책과 움집이 보이고 왼쪽 하단 참호 쪽에 열린 탄약 상자가 배치되어 있습니다.",
        "entities": "레퍼런스와 일치하는 외모와 복장을 한 오주일 단 한 명만 등장하며, 프롬프트에 지시된 대로 눈이 붉게 충혈되어 있습니다.",
        "hard_violations": [],
        "physics": "바위 뒤에 몸을 낮추고 상체를 앞으로 내민 채 지면에 의해 안정적으로 지탱되는 자연스러운 자세를 취하고 있습니다."
       },
       {
        "label": "B",
        "direction": "화면 왼쪽 밖의 대상을 향해 시선을 두고 입을 벌려 소리치고 있습니다.",
        "built_space": "엄폐 바위 사이 공간에 위치하며 배경으로 산채의 목책과 구조물들이 보입니다.",
        "entities": "오주일이 등장하나, 가슴에 '오주일'이라는 읽을 수 있는 글씨가 적힌 명찰이 있으며, 배경에 프롬프트에 없는 3명의 추가 인물(의병들)이 등장합니다.",
        "hard_violations": [
         "지정된 인물 외에 프롬프트에 명시되지 않은 3명의 인물 추가 (PEOPLE 규정 위반)",
         "가슴에 읽을 수 있는 텍스트(명찰) 포함 (No readable writing 규정 위반)"
        ],
        "physics": "바위에 팔과 상체를 기대어 몸을 지탱하고 있는 자세입니다."
       }
      ]
     },
     "ok": true
    },
    {
     "model": "gpt-high",
     "order": "reverse",
     "display_to_canonical": {
      "A": "B",
      "B": "A"
     },
     "raw": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 2,
        "verdict_ko": "허용되지 않은 남성 세 명과 읽히는 가슴 표찰이 추가됐고, 호통의 방향도 뒤에 있는 의병들과 어긋나며 큰 양옆 바위가 소박한 엄폐 구도를 벗어난다."
       },
       {
        "label": "B",
        "score": 8,
        "verdict_ko": "오주일만 낮은 엄폐 자세로 화면 밖 의병들을 향해 호통치는 순간과 붉어진 눈을 잘 구현하지만, 참조의 군모가 빠지고 바위와 상체의 비중이 다소 크다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "오주일의 얼굴과 눈은 화면 왼쪽 앞쪽을 향하고 입을 크게 벌리고 있다. 보이는 의병 세 명은 그의 왼쪽 뒤에 있어 호통이 그들에게 향하지 않는다. 세 남성은 대체로 카메라 쪽을 바라본다. 조준하는 무기나 이동하는 물체는 없다.",
        "built_space": "뒤쪽에 초가 건물 세 채의 일부와 목책, 숲이 보인다. 전경 아래에는 낮은 바위 가장자리가 있지만 양옆에도 머리 높이 이상의 거대한 바위가 있어 좁은 바위 틈처럼 구성됐다. 어깨 옆과 하단에만 소박하게 엄폐 바위를 두라는 요구보다 바위의 비중이 훨씬 크다. 오주일은 앞쪽, 추가 남성 세 명은 그 뒤에 배치돼 있다.",
        "entities": "주인공은 검은 머리의 성인 동아시아계 남성으로, 참조와 유사한 군모와 기운 흔적이 있는 어두운 군복을 입었다. 얼굴은 참조보다 나이 들고 거칠게 보인다. 붉어진 눈과 크게 열린 입은 맞는다. 뒤에는 밝은 한복과 검은 머리쓰개를 착용한 성인 남성 세 명이 추가됐다. 가슴 표찰에는 한글 일부가 읽히게 노출돼 있다. 가마솥, 깨진 밥그릇, 흩어진 쌀, 탄약 은닉처는 이 크롭에서 확인되지 않으며, 제외 자체는 감점 사유가 아니다.",
        "hard_violations": [
         "화면에 허용된 오주일 외에 남성 세 명을 추가했다.",
         "군복 가슴 표찰에 읽을 수 있는 한글 일부가 노출돼 무문자 지시를 위반했다."
        ],
        "physics": "오주일은 바위 뒤에서 상체를 앞으로 숙이고 있으며 하체와 지면 접촉점은 프레임 밖이다. 보이는 상체는 낮게 웅크린 자세로 설명 가능하고 공중에 떠 있다는 징후는 없다. 뒤의 남성들도 하체가 가려져 있으며 부유한 자세는 아니다. 바위들은 지면 또는 화면 밖으로 이어지는 바위 덩어리에 지지돼 있다."
       },
       {
        "label": "B",
        "direction": "오주일의 얼굴과 시선, 열린 입은 화면 오른쪽 바깥을 향한다. 호통의 대상인 생존 의병은 보이지 않지만, 화면 밖 상대에게 명령하는 방향으로 자연스럽게 읽힌다. 렌즈를 정면으로 응시하지 않으며 조준하는 무기나 이동 물체는 없다.",
        "built_space": "배경에 초가 건물 세 채, 이어진 목책과 산림이 보이고 흙마당의 재질도 이전 장면과 유사하다. 인물 뒤 왼쪽 아래에는 뚜껑이 열린 탄약 상자 한 개와 파인 흙 지형이 있다. 오주일은 전경 엄폐 바위의 안쪽에 낮게 자리한다. 하단 바위 가장자리와 오른쪽 어깨 옆 바위가 자세를 설명하지만, 오른쪽 바위의 화면 점유율은 요구된 소박한 존재감보다 크다.",
        "entities": "검은 머리의 성인 동아시아계 남성 한 명만 보인다. 오주일 참조와 유사한 얼굴 구조와 어두운 수선 군복을 갖췄지만 얼굴이 다소 더 나이 들어 보이고 참조의 군모는 없다. 눈가가 붉고 입을 크게 벌린 매서운 표정은 명시된 행동에 맞는다. 열린 상자와 내부의 길쭉한 물체들이 탄약 은닉처를 시사하지만 내용물의 종류는 선명하지 않다. 가마솥, 밥그릇, 쌀과 참조의 소총은 이 크롭에서 확인되지 않으며, 이를 보여주기 위해 구도를 넓힐 필요는 없다. 읽히는 문자는 없다.",
        "hard_violations": [],
        "physics": "오주일은 어깨를 낮추고 몸통을 앞으로 기울인 채 바위 뒤에 있다. 무릎과 발은 잘려 있어 직접적인 지면 접촉은 확인할 수 없지만, 보이는 자세는 지면에 웅크린 몸으로 자연스럽게 설명된다. 공중에 뜬 신체나 물체는 없다. 탄약 상자는 흙바닥에 놓여 있고 열린 뚜껑은 상자 뒤로 기대어 있으며, 전경 바위들도 지면과 다른 바위에 이어져 있다."
       }
      ],
      "all_candidates_fail": false
     },
     "normalized": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 2,
        "verdict_ko": "허용되지 않은 남성 세 명과 읽히는 가슴 표찰이 추가됐고, 호통의 방향도 뒤에 있는 의병들과 어긋나며 큰 양옆 바위가 소박한 엄폐 구도를 벗어난다."
       },
       {
        "label": "A",
        "score": 8,
        "verdict_ko": "오주일만 낮은 엄폐 자세로 화면 밖 의병들을 향해 호통치는 순간과 붉어진 눈을 잘 구현하지만, 참조의 군모가 빠지고 바위와 상체의 비중이 다소 크다."
       }
      ],
      "all_candidates_fail": false,
      "readings": [
       {
        "label": "B",
        "direction": "오주일의 얼굴과 눈은 화면 왼쪽 앞쪽을 향하고 입을 크게 벌리고 있다. 보이는 의병 세 명은 그의 왼쪽 뒤에 있어 호통이 그들에게 향하지 않는다. 세 남성은 대체로 카메라 쪽을 바라본다. 조준하는 무기나 이동하는 물체는 없다.",
        "built_space": "뒤쪽에 초가 건물 세 채의 일부와 목책, 숲이 보인다. 전경 아래에는 낮은 바위 가장자리가 있지만 양옆에도 머리 높이 이상의 거대한 바위가 있어 좁은 바위 틈처럼 구성됐다. 어깨 옆과 하단에만 소박하게 엄폐 바위를 두라는 요구보다 바위의 비중이 훨씬 크다. 오주일은 앞쪽, 추가 남성 세 명은 그 뒤에 배치돼 있다.",
        "entities": "주인공은 검은 머리의 성인 동아시아계 남성으로, 참조와 유사한 군모와 기운 흔적이 있는 어두운 군복을 입었다. 얼굴은 참조보다 나이 들고 거칠게 보인다. 붉어진 눈과 크게 열린 입은 맞는다. 뒤에는 밝은 한복과 검은 머리쓰개를 착용한 성인 남성 세 명이 추가됐다. 가슴 표찰에는 한글 일부가 읽히게 노출돼 있다. 가마솥, 깨진 밥그릇, 흩어진 쌀, 탄약 은닉처는 이 크롭에서 확인되지 않으며, 제외 자체는 감점 사유가 아니다.",
        "hard_violations": [
         "화면에 허용된 오주일 외에 남성 세 명을 추가했다.",
         "군복 가슴 표찰에 읽을 수 있는 한글 일부가 노출돼 무문자 지시를 위반했다."
        ],
        "physics": "오주일은 바위 뒤에서 상체를 앞으로 숙이고 있으며 하체와 지면 접촉점은 프레임 밖이다. 보이는 상체는 낮게 웅크린 자세로 설명 가능하고 공중에 떠 있다는 징후는 없다. 뒤의 남성들도 하체가 가려져 있으며 부유한 자세는 아니다. 바위들은 지면 또는 화면 밖으로 이어지는 바위 덩어리에 지지돼 있다."
       },
       {
        "label": "A",
        "direction": "오주일의 얼굴과 시선, 열린 입은 화면 오른쪽 바깥을 향한다. 호통의 대상인 생존 의병은 보이지 않지만, 화면 밖 상대에게 명령하는 방향으로 자연스럽게 읽힌다. 렌즈를 정면으로 응시하지 않으며 조준하는 무기나 이동 물체는 없다.",
        "built_space": "배경에 초가 건물 세 채, 이어진 목책과 산림이 보이고 흙마당의 재질도 이전 장면과 유사하다. 인물 뒤 왼쪽 아래에는 뚜껑이 열린 탄약 상자 한 개와 파인 흙 지형이 있다. 오주일은 전경 엄폐 바위의 안쪽에 낮게 자리한다. 하단 바위 가장자리와 오른쪽 어깨 옆 바위가 자세를 설명하지만, 오른쪽 바위의 화면 점유율은 요구된 소박한 존재감보다 크다.",
        "entities": "검은 머리의 성인 동아시아계 남성 한 명만 보인다. 오주일 참조와 유사한 얼굴 구조와 어두운 수선 군복을 갖췄지만 얼굴이 다소 더 나이 들어 보이고 참조의 군모는 없다. 눈가가 붉고 입을 크게 벌린 매서운 표정은 명시된 행동에 맞는다. 열린 상자와 내부의 길쭉한 물체들이 탄약 은닉처를 시사하지만 내용물의 종류는 선명하지 않다. 가마솥, 밥그릇, 쌀과 참조의 소총은 이 크롭에서 확인되지 않으며, 이를 보여주기 위해 구도를 넓힐 필요는 없다. 읽히는 문자는 없다.",
        "hard_violations": [],
        "physics": "오주일은 어깨를 낮추고 몸통을 앞으로 기울인 채 바위 뒤에 있다. 무릎과 발은 잘려 있어 직접적인 지면 접촉은 확인할 수 없지만, 보이는 자세는 지면에 웅크린 몸으로 자연스럽게 설명된다. 공중에 뜬 신체나 물체는 없다. 탄약 상자는 흙바닥에 놓여 있고 열린 뚜껑은 상자 뒤로 기대어 있으며, 전경 바위들도 지면과 다른 바위에 이어져 있다."
       }
      ]
     },
     "ok": true
    }
   ],
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "slot_winner_match": true,
   "slot_winner": {
    "gemini-pro": "A",
    "gpt-high": "A"
   },
   "route": "cross_slot_agree"
  },
  "dual": {
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "normalized": {
    "A": 2.0,
    "B": 0.45
   },
   "adjusted": {
    "A": 2.0,
    "B": 0.2
   },
   "violations": {
    "B": [
     "[gemini-pro] 지정된 인물 외에 프롬프트에 명시되지 않은 3명의 인물 추가 (PEOPLE 규정 위반)",
     "[gemini-pro] 가슴에 읽을 수 있는 텍스트(명찰) 포함 (No readable writing 규정 위반)",
     "[gpt-high] 화면에 허용된 오주일 외에 남성 세 명을 추가했다.",
     "[gpt-high] 군복 가슴 표찰에 읽을 수 있는 한글 일부가 노출돼 무문자 지시를 위반했다."
    ]
   },
   "per_model_winner": {
    "gemini-pro": "A",
    "gpt-high": "A"
   },
   "agreed": true
  },
  "totals": {
   "A": 2000,
   "B": 200
  },
  "selected": "A",
  "ranking": [
   "A",
   "B"
  ],
  "verdicts": [
   {
    "label": "A",
    "score": 2000,
    "verdict_ko": "샷 텍스트와 모든 제약 조건(오주일 단독 등장, 붉어진 눈, 텍스트 없음, 열린 탄약 상자 배치 등)을 완벽하게 준수하여 캐릭터의 감정과 지정된 배경을 훌륭하게 구현했습니다."
   },
   {
    "label": "B",
    "score": 200,
    "verdict_ko": "프롬프트에서 엄격히 금지한 추가 인물들의 등장과 읽을 수 있는 텍스트(가슴의 명찰)가 포함되어 치명적인 규정 위반이 발생했습니다.  ★위반: [gemini-pro] 지정된 인물 외에 프롬프트에 명시되지 않은 3명의 인물 추가 (PEOPLE 규정 위반) / [gemini-pro] 가슴에 읽을 수 있는 텍스트(명찰) 포함 (No readable writing 규정 위반) / [gpt-high] 화면에 허용된 오주일 외에 남성 세 명을 추가했다. / [gpt-high] 군복 가슴 표찰에 읽을 수 있는 한글 일부가 노출돼 무문자 지시를 위반했다."
   }
  ],
  "refs": [
   {
    "label": "PREVIOUS SHOT STILL — a visually related earlier shot of this same place: the location's look, materials, fixed features and lighting mood are LOCKED to this photo; never copy its camera framing. The people visible in that photo are NOT in this shot: never carry their faces, bodies or clothing onto anyone here. Each person in THIS shot is defined solely by the PEOPLE list, their own CHARACTER REFERENCE images and their wardrobe notes. If this photo shows a character who cannot move (dead or unconscious), that character's exact pose, position and orientation are ALSO LOCKED — treat their body as a fixed prop of the set that only the camera moves around.",
    "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/db00517b-165b-4005-a287-f28936b59b95/images/272bd7af-9294-4490-98c6-88a40e9deacc/scene/recipe/S54sh4_sel.png",
    "asset_id": "95f71b3c-2b48-45d0-9a62-6cd4f9036cb6",
    "role": "prev_still"
   },
   {
    "label": "CHARACTER REFERENCE — 오주일: the exact person appearing in this shot; match face, hair and build exactly.",
    "path": "<bytes:809716>",
    "asset_id": "7c31e825-ceef-4208-915f-71c4f0e15dd6",
    "role": "character_ref"
   }
  ],
  "critique_skipped": true,
  "shot_run_produced": false,
  "shot_run_uid": "06aa1438-e45d-7abe-82fe-7c50c86d84f5",
  "ref_mode": "prev+엔티티",
  "share_plan": {
   "ref_plan": "prev",
   "prev_anchor_tag": "S54sh4"
  }
 },
 "S56sh40::cine": {
  "applied": true,
  "attempted_at": "2026-09-09T07:26:12.327608+00:00",
  "fingerprint": "d0bb87b79070dff43e95b74fc7db4fd93248f9bec51577f0f639aa47c75890a9",
  "fingerprint_version": 2,
  "provider": "mai",
  "endpoint": "openrouter/chat-completions",
  "model": "microsoft/mai-image-2.6",
  "pack": "24.202608252115",
  "source_file": "S56sh40_sel.png",
  "source_sha256": "fe8f843cc8bf3ca44d9293e60ff1d95a92249bfc1f9ae2c6357fbecc5b18104d",
  "file": "S56sh40_cine.png",
  "staged_sha256": "aa50938c1b2006dbb4036c57b8ca679f1375aaf802486369097427251b145ef8",
  "latency_ms": 24769
 },
 "S57sh7::signage": {
  "fp": "ea8849a2e0b9a36d",
  "inscriptions": [],
  "cues": [],
  "dropped": []
 },
 "S57sh7::bgfirst_bg": {
  "input_fingerprint": "09a653ef5821b51d",
  "prompt": "Create the EMPTY BACKGROUND PLATE for one film shot — NO PEOPLE, no figures, no body parts, no sketch lines, no arrows anywhere.\n\"Empty\" means no people only: KEEP the location's inherent occupants and stock that define the place — animals in an animal shelter, pen or farm, goods in a market, moored boats in a harbour — unless the shot text explicitly removes them.\nThe FIRST attached image is a thin-line storyboard sketch: use ONLY its camera angle, horizon, perspective and the placement/size of buildings and set masses — ignore the sketched people and arrows entirely. The SECOND attached image (LOCATION PHOTOGRAPH) is the real place: take its architecture, materials, signage and fixed features, and RE-PROJECT them into the sketch's camera. If the photograph's camera differs from the sketch's, the sketch's camera wins.\nHUMAN-SCALE CALIBRATION: derive every structure's true size from human-scale elements — a door ≈ 2m, a window ≈ 1–1.5m wide, one storey ≈ 2.5–3m; never inflate a small structure or shrink a large one.\n\nSHOT TEXT this background must serve (Korean): 허공을 날아온 낫이 도망치는 박상현의 등 한가운데 깊숙이 푹 박혀 상체가 충격으로 앞으로 쏠린 미드임팩트 순간.\n\nLOCATION (lock): The outdoor market alley beside an exterior wall, along the fleeing man's route away from the central fighting area.\n\nTIME OF DAY (lock): day.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: medium shot\n- KEY BACKGROUND ELEMENTS: 벽 (Ahead of 박상현, before his body strikes it) — Seen obliquely along the direction of pursuit; used as A narrow right-side plane makes the imminent collision legible without blocking the back wound; 등에 박힌 낫 (Embedded in the middle of 박상현's back) — The exposed portion is viewed from behind at an oblique angle; used as A realistically scaled detail against his back establishes the cause of the forward pitch.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Natural daylight and subdued saturation render the impact with blunt physical clarity rather than stylized spectacle.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nRender ONE photorealistic empty location photograph, 16:9, that this shot can be staged inside later. No readable writing anywhere: surfaces that would carry writing may be present, but stage any wording out of legibility — an oblique angle, distance, shallow focus. No captions, watermarks or overlay text.",
  "effective_prompt": "Create the EMPTY BACKGROUND PLATE for one film shot — NO PEOPLE, no figures, no body parts, no sketch lines, no arrows anywhere.\n\"Empty\" means no people only: KEEP the location's inherent occupants and stock that define the place — animals in an animal shelter, pen or farm, goods in a market, moored boats in a harbour — unless the shot text explicitly removes them.\nThe FIRST attached image is a thin-line storyboard sketch: use ONLY its camera angle, horizon, perspective and the placement/size of buildings and set masses — ignore the sketched people and arrows entirely. The SECOND attached image (LOCATION PHOTOGRAPH) is the real place: take its architecture, materials, signage and fixed features, and RE-PROJECT them into the sketch's camera. If the photograph's camera differs from the sketch's, the sketch's camera wins.\nHUMAN-SCALE CALIBRATION: derive every structure's true size from human-scale elements — a door ≈ 2m, a window ≈ 1–1.5m wide, one storey ≈ 2.5–3m; never inflate a small structure or shrink a large one.\n\nSHOT TEXT this background must serve (Korean): 허공을 날아온 낫이 도망치는 박상현의 등 한가운데 깊숙이 푹 박혀 상체가 충격으로 앞으로 쏠린 미드임팩트 순간.\n\nLOCATION (lock): The outdoor market alley beside an exterior wall, along the fleeing man's route away from the central fighting area.\n\nTIME OF DAY (lock): day.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: medium shot\n- KEY BACKGROUND ELEMENTS: 벽 (Ahead of 박상현, before his body strikes it) — Seen obliquely along the direction of pursuit; used as A narrow right-side plane makes the imminent collision legible without blocking the back wound; 등에 박힌 낫 (Embedded in the middle of 박상현's back) — The exposed portion is viewed from behind at an oblique angle; used as A realistically scaled detail against his back establishes the cause of the forward pitch.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Natural daylight and subdued saturation render the impact with blunt physical clarity rather than stylized spectacle.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nRender ONE photorealistic empty location photograph, 16:9, that this shot can be staged inside later. No readable writing anywhere: surfaces that would carry writing may be present, but stage any wording out of legibility — an oblique angle, distance, shallow focus. No captions, watermarks or overlay text.",
  "bg_path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/db00517b-165b-4005-a287-f28936b59b95/images/272bd7af-9294-4490-98c6-88a40e9deacc/scene/recipe/S57sh7__bgfirst_bg.png",
  "asset_id": "7609081c-7d2f-4cd9-83f7-fa5f622edfa3",
  "input_asset_ids": [
   "32ba9ee3-5ee8-4bf5-b51d-628359c1bf61",
   "32b543ac-1086-4282-b907-3fcd465b580a"
  ]
 },
 "S57sh7": {
  "input_fingerprint": "d5f3119d2b2f923a",
  "prompt": "Create ONE FINAL photorealistic live-action film still of the moment below — early 20th-century Korea, Jeolla region; all people are Korean unless explicitly stated otherwise. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 허공을 날아온 낫이 도망치는 박상현의 등 한가운데 깊숙이 푹 박혀 상체가 충격으로 앞으로 쏠린 미드임팩트 순간.\n\nLOCATION (lock): The outdoor market alley beside an exterior wall, along the fleeing man's route away from the central fighting area. The shot takes place here — the attached LOCATION STRUCTURE PHOTOGRAPH is the single authority for this exact place — its fixed structure and permanent site details are LOCKED to it. No separate location photograph exists for this place. Build everything else strictly from the location text above and the shot text; the layout sketch (when attached) governs framing and placement only, and the shot text governs time of day, lighting and action.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: medium shot\n- KEY BACKGROUND ELEMENTS: 벽 (Ahead of 박상현, before his body strikes it) — Seen obliquely along the direction of pursuit; used as A narrow right-side plane makes the imminent collision legible without blocking the back wound; 등에 박힌 낫 (Embedded in the middle of 박상현's back) — The exposed portion is viewed from behind at an oblique angle; used as A realistically scaled detail against his back establishes the cause of the forward pitch.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Natural daylight and subdued saturation render the impact with blunt physical clarity rather than stylized spectacle.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): Broken stalls, fallen weapons and blood remain along the emptied market lane. 박상현: A thrown sickle is embedded in his back as he strikes the wall.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 박상현 (한국인 남성, 성인 나이대, 자연스러운 검은 머리) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.",
  "roll_prompts": {
   "A": "Stage the shot. The FIRST attached image (SHOT BACKGROUND) is the finished empty background of this shot — keep it EXACTLY: its camera, perspective, architecture, lighting and every fixed feature stay untouched. The SECOND attached image (LAYOUT SKETCH) tells you ONLY where the people go: each sketched person's position, screen size, pose and the gaze/motion arrows. Ignore the sketch's background lines. The CHARACTER REFERENCE photographs show the real people.\nPlace the real people into the background at exactly the sketched positions, sizes and poses, following the arrow directions. No sketch lines or arrows may remain.\n\nCreate ONE FINAL photorealistic live-action film still of the moment below — early 20th-century Korea, Jeolla region; all people are Korean unless explicitly stated otherwise. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 허공을 날아온 낫이 도망치는 박상현의 등 한가운데 깊숙이 푹 박혀 상체가 충격으로 앞으로 쏠린 미드임팩트 순간.\n\nLOCATION (lock): The outdoor market alley beside an exterior wall, along the fleeing man's route away from the central fighting area. The shot takes place here — the FIRST attached image (SHOT BACKGROUND) is this exact place, already built: its ground, structures, horizon, materials and lighting are the finished truth of this location and must not be redesigned or replaced. No location photograph is attached — read the place from that image alone, and add no scenery, structure, vehicle or fixture that it does not already show. This lock governs the place only; the figures in the shot follow the staging and pose instructions.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: medium shot\n- KEY BACKGROUND ELEMENTS: 벽 (Ahead of 박상현, before his body strikes it) — Seen obliquely along the direction of pursuit; used as A narrow right-side plane makes the imminent collision legible without blocking the back wound; 등에 박힌 낫 (Embedded in the middle of 박상현's back) — The exposed portion is viewed from behind at an oblique angle; used as A realistically scaled detail against his back establishes the cause of the forward pitch.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Natural daylight and subdued saturation render the impact with blunt physical clarity rather than stylized spectacle.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): Broken stalls, fallen weapons and blood remain along the emptied market lane. 박상현: A thrown sickle is embedded in his back as he strikes the wall.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 박상현 (한국인 남성, 성인 나이대, 자연스러운 검은 머리) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.",
   "B": "Create ONE FINAL photorealistic live-action film still of the moment below — early 20th-century Korea, Jeolla region; all people are Korean unless explicitly stated otherwise. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 허공을 날아온 낫이 도망치는 박상현의 등 한가운데 깊숙이 푹 박혀 상체가 충격으로 앞으로 쏠린 미드임팩트 순간.\n\nLOCATION (lock): The outdoor market alley beside an exterior wall, along the fleeing man's route away from the central fighting area. The shot takes place here — the attached LOCATION STRUCTURE PHOTOGRAPH is the single authority for this exact place — its fixed structure and permanent site details are LOCKED to it. No separate location photograph exists for this place. Build everything else strictly from the location text above and the shot text; the layout sketch (when attached) governs framing and placement only, and the shot text governs time of day, lighting and action.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: medium shot\n- KEY BACKGROUND ELEMENTS: 벽 (Ahead of 박상현, before his body strikes it) — Seen obliquely along the direction of pursuit; used as A narrow right-side plane makes the imminent collision legible without blocking the back wound; 등에 박힌 낫 (Embedded in the middle of 박상현's back) — The exposed portion is viewed from behind at an oblique angle; used as A realistically scaled detail against his back establishes the cause of the forward pitch.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Natural daylight and subdued saturation render the impact with blunt physical clarity rather than stylized spectacle.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): Broken stalls, fallen weapons and blood remain along the emptied market lane. 박상현: A thrown sickle is embedded in his back as he strikes the wall.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 박상현 (한국인 남성, 성인 나이대, 자연스러운 검은 머리) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.\n\nCOMPOSITION VARIATION (this candidate only): keep the place, the people and the moment exactly as specified, and keep from the CAMERA & FRAME contract its subject scale, screen placement and key background elements exactly. For this candidate only, its camera angle and height are deliberately left open: choose a camera position distinctly different from the obvious one, as a film director picking a second setup on the same blocking. Stay true to any explicit framing the SHOT TEXT itself dictates."
  },
  "roll_refs": {
   "A": [
    {
     "label": "SHOT BACKGROUND",
     "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/db00517b-165b-4005-a287-f28936b59b95/images/272bd7af-9294-4490-98c6-88a40e9deacc/scene/recipe/S57sh7__bgfirst_bg.png",
     "asset_id": "7609081c-7d2f-4cd9-83f7-fa5f622edfa3",
     "role": "bgfirst_bg"
    },
    {
     "label": "LAYOUT SKETCH (people placement only)",
     "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/db00517b-165b-4005-a287-f28936b59b95/images/272bd7af-9294-4490-98c6-88a40e9deacc/conti/conti_S57sh7.png",
     "asset_id": "32ba9ee3-5ee8-4bf5-b51d-628359c1bf61",
     "role": "conti_light"
    },
    {
     "label": "CHARACTER REFERENCE — 박상현: the exact person appearing in this shot; match face, hair and build exactly.",
     "path": "<bytes:1167996>",
     "asset_id": "6e1bedfc-d3f9-4391-a49e-1ca6e50bf276",
     "role": "character_ref"
    },
    {
     "label": "PROP REFERENCE — 낫: the exact object appearing in this shot; match its look, material and wear exactly.",
     "path": "<bytes:977465>",
     "asset_id": "e9884693-b2aa-4769-9ee0-4c7b50db40c2",
     "role": "prop_ref"
    }
   ],
   "B": [
    {
     "label": "LOCATION STRUCTURE PHOTOGRAPH — the confirmed photograph of this exact place and its fixed structure: it is the SINGLE authority for the location, the structure's shape, proportions, materials, colors, openings and every permanent site detail. Never copy its camera framing, time of day or lighting — the shot text is the authority for those.",
     "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/db00517b-165b-4005-a287-f28936b59b95/images/background_chain/seed_bg_market_sel.png",
     "asset_id": "32b543ac-1086-4282-b907-3fcd465b580a",
     "role": "location_seed_bg"
    },
    {
     "label": "CHARACTER REFERENCE — 박상현: the exact person appearing in this shot; match face, hair and build exactly.",
     "path": "<bytes:1167996>",
     "asset_id": "6e1bedfc-d3f9-4391-a49e-1ca6e50bf276",
     "role": "character_ref"
    },
    {
     "label": "PROP REFERENCE — 낫: the exact object appearing in this shot; match its look, material and wear exactly.",
     "path": "<bytes:977465>",
     "asset_id": "e9884693-b2aa-4769-9ee0-4c7b50db40c2",
     "role": "prop_ref"
    }
   ]
  },
  "shot_run_spend_attempt_count": 1,
  "judge_flip_skipped": "judge_fn_owns_order",
  "readings": [
   {
    "label": "A",
    "direction": "시선과 몸의 방향이 앞으로 강하게 쏠려 있으며, 뻗은 오른손이 우측 흙벽을 향해 부딪히고 있다. 등에 박힌 낫의 손잡이는 좌측 아래를 향한다.",
    "built_space": "왼편에 늘어선 초가집들과 오른편의 지붕 없는 높은 흙벽이 레퍼런스 사진의 공간 구조와 일치한다. 길바닥에는 부서진 무기류가 흩어져 있다. 인물은 우측 벽에 밀착하여 충돌하는 위치에 있다.",
    "entities": "피 묻은 회색 한복과 갓을 착용한 박상현의 모습이 레퍼런스와 일치하게 묘사되었다. 등 한가운데에 레퍼런스 이미지와 일치하는 형태의 낫이 꽂혀 있다.",
    "hard_violations": [],
    "physics": "앞으로 넘어지는 상체의 무게중심이 벽을 짚은 오른손을 통해 지탱되고 있어 충격을 받은 물리적 묘사가 자연스럽다. 낫은 몸의 근육과 옷에 의해 지지되어 등에 깊숙이 박혀 있으며, 바닥의 잔해물들은 땅에 안정적으로 놓여 있다."
   },
   {
    "label": "B",
    "direction": "몸이 앞을 향해 있으나 벽으로 쏠리거나 부딪히는 방향성이 없고 시선은 바닥을 향해 걷는 모습이다. 등에 박힌 낫은 손잡이가 수평으로 튀어나와 있다.",
    "built_space": "좌측에 초가집이 있고 우측에도 초가지붕이 덮인 벽이 묘사되어, 우측이 지붕 없는 맨 흙벽이어야 하는 위치 레퍼런스의 고정된 물리적 구조를 어겼다. 인물은 벽과 거리를 두고 길 한가운데를 걷고 있다.",
    "entities": "회색 한복을 입은 박상현이나, 레퍼런스와 달리 갓을 쓰지 않고 상투만 튼 모습이다. 등에는 낫이 박혀 있다.",
    "hard_violations": [
     "[gemini-pro] 우측 벽의 구조가 레퍼런스(지붕 없는 흙벽)와 달리 초가지붕이 덮인 형태로 변형됨"
    ],
    "physics": "다리로 체중을 지탱하며 걷는 안정적인 자세를 유지하고 있어 벽에 부딪히는 물리적 충격이 보이지 않는다. 등에 박힌 낫은 몸에 의해 지지되고 있다."
   }
  ],
  "cross_model_order": {
   "policy": "cross_model_order_v2_slot0_fwd_slot1_rev_parallel2",
   "slots": [
    {
     "model": "gemini-pro",
     "order": "forward",
     "display_to_canonical": {
      "A": "A",
      "B": "B"
     },
     "raw": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 9,
        "verdict_ko": "벽에 부딪히며 상체가 쏠리는 '미드임팩트'의 순간을 역동적으로 잘 포착했으며, 레퍼런스의 장소 구조와 소품을 충실히 구현했습니다."
       },
       {
        "label": "B",
        "score": 4,
        "verdict_ko": "벽에 부딪히는 핵심 액션이 누락된 채 단순히 걷고 있으며, 우측 흙벽의 구조를 임의로 변형하고 캐릭터의 갓을 빠뜨렸습니다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "시선과 몸의 방향이 앞으로 강하게 쏠려 있으며, 뻗은 오른손이 우측 흙벽을 향해 부딪히고 있다. 등에 박힌 낫의 손잡이는 좌측 아래를 향한다.",
        "built_space": "왼편에 늘어선 초가집들과 오른편의 지붕 없는 높은 흙벽이 레퍼런스 사진의 공간 구조와 일치한다. 길바닥에는 부서진 무기류가 흩어져 있다. 인물은 우측 벽에 밀착하여 충돌하는 위치에 있다.",
        "entities": "피 묻은 회색 한복과 갓을 착용한 박상현의 모습이 레퍼런스와 일치하게 묘사되었다. 등 한가운데에 레퍼런스 이미지와 일치하는 형태의 낫이 꽂혀 있다.",
        "hard_violations": [],
        "physics": "앞으로 넘어지는 상체의 무게중심이 벽을 짚은 오른손을 통해 지탱되고 있어 충격을 받은 물리적 묘사가 자연스럽다. 낫은 몸의 근육과 옷에 의해 지지되어 등에 깊숙이 박혀 있으며, 바닥의 잔해물들은 땅에 안정적으로 놓여 있다."
       },
       {
        "label": "B",
        "direction": "몸이 앞을 향해 있으나 벽으로 쏠리거나 부딪히는 방향성이 없고 시선은 바닥을 향해 걷는 모습이다. 등에 박힌 낫은 손잡이가 수평으로 튀어나와 있다.",
        "built_space": "좌측에 초가집이 있고 우측에도 초가지붕이 덮인 벽이 묘사되어, 우측이 지붕 없는 맨 흙벽이어야 하는 위치 레퍼런스의 고정된 물리적 구조를 어겼다. 인물은 벽과 거리를 두고 길 한가운데를 걷고 있다.",
        "entities": "회색 한복을 입은 박상현이나, 레퍼런스와 달리 갓을 쓰지 않고 상투만 튼 모습이다. 등에는 낫이 박혀 있다.",
        "hard_violations": [
         "우측 벽의 구조가 레퍼런스(지붕 없는 흙벽)와 달리 초가지붕이 덮인 형태로 변형됨"
        ],
        "physics": "다리로 체중을 지탱하며 걷는 안정적인 자세를 유지하고 있어 벽에 부딪히는 물리적 충격이 보이지 않는다. 등에 박힌 낫은 몸에 의해 지지되고 있다."
       }
      ],
      "all_candidates_fail": false
     },
     "normalized": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 9,
        "verdict_ko": "벽에 부딪히며 상체가 쏠리는 '미드임팩트'의 순간을 역동적으로 잘 포착했으며, 레퍼런스의 장소 구조와 소품을 충실히 구현했습니다."
       },
       {
        "label": "B",
        "score": 4,
        "verdict_ko": "벽에 부딪히는 핵심 액션이 누락된 채 단순히 걷고 있으며, 우측 흙벽의 구조를 임의로 변형하고 캐릭터의 갓을 빠뜨렸습니다."
       }
      ],
      "all_candidates_fail": false,
      "readings": [
       {
        "label": "A",
        "direction": "시선과 몸의 방향이 앞으로 강하게 쏠려 있으며, 뻗은 오른손이 우측 흙벽을 향해 부딪히고 있다. 등에 박힌 낫의 손잡이는 좌측 아래를 향한다.",
        "built_space": "왼편에 늘어선 초가집들과 오른편의 지붕 없는 높은 흙벽이 레퍼런스 사진의 공간 구조와 일치한다. 길바닥에는 부서진 무기류가 흩어져 있다. 인물은 우측 벽에 밀착하여 충돌하는 위치에 있다.",
        "entities": "피 묻은 회색 한복과 갓을 착용한 박상현의 모습이 레퍼런스와 일치하게 묘사되었다. 등 한가운데에 레퍼런스 이미지와 일치하는 형태의 낫이 꽂혀 있다.",
        "hard_violations": [],
        "physics": "앞으로 넘어지는 상체의 무게중심이 벽을 짚은 오른손을 통해 지탱되고 있어 충격을 받은 물리적 묘사가 자연스럽다. 낫은 몸의 근육과 옷에 의해 지지되어 등에 깊숙이 박혀 있으며, 바닥의 잔해물들은 땅에 안정적으로 놓여 있다."
       },
       {
        "label": "B",
        "direction": "몸이 앞을 향해 있으나 벽으로 쏠리거나 부딪히는 방향성이 없고 시선은 바닥을 향해 걷는 모습이다. 등에 박힌 낫은 손잡이가 수평으로 튀어나와 있다.",
        "built_space": "좌측에 초가집이 있고 우측에도 초가지붕이 덮인 벽이 묘사되어, 우측이 지붕 없는 맨 흙벽이어야 하는 위치 레퍼런스의 고정된 물리적 구조를 어겼다. 인물은 벽과 거리를 두고 길 한가운데를 걷고 있다.",
        "entities": "회색 한복을 입은 박상현이나, 레퍼런스와 달리 갓을 쓰지 않고 상투만 튼 모습이다. 등에는 낫이 박혀 있다.",
        "hard_violations": [
         "우측 벽의 구조가 레퍼런스(지붕 없는 흙벽)와 달리 초가지붕이 덮인 형태로 변형됨"
        ],
        "physics": "다리로 체중을 지탱하며 걷는 안정적인 자세를 유지하고 있어 벽에 부딪히는 물리적 충격이 보이지 않는다. 등에 박힌 낫은 몸에 의해 지지되고 있다."
       }
      ]
     },
     "ok": true
    },
    {
     "model": "gpt-high",
     "order": "reverse",
     "display_to_canonical": {
      "A": "B",
      "B": "A"
     },
     "raw": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 6,
        "verdict_ko": "앞으로 숙여지는 중간 샷은 맞지만 충격보다 느린 굽힘처럼 보이며, 갓과 등 상처의 표현이 빠지고 오른쪽 외벽 구조도 참조와 덜 맞습니다."
       },
       {
        "label": "B",
        "score": 8,
        "verdict_ko": "등 중앙에 박힌 낫과 상처, 벽 쪽으로 무너지는 동작, 갓과 돌기단 흙벽이 핵심 지시를 더 충실히 구현하지만 의복 길이와 낫의 비례는 참조와 다릅니다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "남자는 화면 오른쪽 앞의 흙벽을 향해 고개와 상체를 숙이고 있으며 카메라를 보지 않습니다. 낫 손잡이는 왼쪽 뒤로 돌출되고 굽은 날 끝은 등 중앙보다 조금 높은 부위로 들어갑니다. 몸의 진행 방향과 충돌할 벽은 연결되지만 팔이 아래로 처져 있어 급격한 피격 순간의 방향성은 약합니다.",
        "built_space": "오른쪽에는 초가 처마 아래 흙벽 한 면과 그 왼쪽 목재 문 한 곳이 크게 보입니다. 왼쪽과 중앙 후경에는 초가 건물 여러 채, 목재 기둥과 무너진 좌판이 있는 흙길이 이어집니다. 참조의 흙·목재·짚 재료는 따르지만, 오른쪽의 돌기단을 둔 긴 담장 대신 건물 처마와 문이 붙은 벽이 강조됩니다. 벽은 좁은 우측 면보다는 넓은 전경 면을 차지합니다. 인물은 벽 바로 앞 골목에 있고 반사면은 없습니다.",
        "entities": "보이는 사람은 검은 머리를 묶은 성인 남성 한 명뿐이며 한국인 설정과 외관상 충돌하지 않습니다. 얼굴은 옆면 일부만 보여 정확한 동일인 여부를 판단하기 어렵습니다. 회색 긴 도포와 허리띠는 참조와 유사하지만 검은 갓은 없고 옷의 혈흔과 손상도 훨씬 적습니다. 낫 한 자루는 목재 손잡이와 어두운 굽은 금속 날을 갖추지만 참조보다 날이 짧고 둥글게 보입니다. 등에는 뚜렷한 출혈이나 깊은 상처가 보이지 않습니다. 후경의 부서진 좌판은 보이나 떨어진 무기와 바닥 혈흔은 명확하게 식별되지 않습니다. 읽을 수 있는 글자는 없습니다.",
        "hard_violations": [],
        "physics": "하체와 발은 프레임 밖으로 이어지며, 몸 전체가 공중에 떠 있는 모습은 아닙니다. 골반 위에서 상체를 앞으로 굽힌 자세 자체는 가능합니다. 낫은 날 끝이 옷과 등으로 들어간 접점에 의해 지지되는 것으로 보이고 손잡이만 뒤로 노출됩니다. 다만 옷의 당김이나 상처 반응이 약해 깊숙한 관입과 갑작스러운 충격의 물리적 증거가 부족합니다. 손이나 몸통이 벽에 닿은 모습은 아직 없습니다."
       },
       {
        "label": "B",
        "direction": "남자는 카메라에서 멀어지는 방향으로 몸을 틀면서 오른쪽 벽으로 기울고, 오른손을 그 벽에 뻗어 접촉합니다. 얼굴은 진행 방향으로 돌아가 눈의 정확한 시선은 보이지 않습니다. 낫 손잡이는 왼쪽 뒤·아래로 돌출되고 굽은 날 끝은 혈흔이 있는 등 중앙으로 들어갑니다. 등 상처를 가리지 않으면서 벽과의 임박한 몸통 충돌을 보여줍니다.",
        "built_space": "오른쪽에는 돌기단 위의 긴 흙담 한 면이 골목을 따라 이어지고, 왼쪽에는 목재 기둥으로 처마를 받친 초가 건물들이 연속합니다. 흙길은 중앙 후경의 초가집들 사이로 좁아집니다. 담장 재료, 기단, 초가지붕과 골목의 관계가 참조에 더 가깝습니다. 왼쪽 전경에는 부서진 좌판 목재가 놓여 있습니다. 인물은 담장 옆 통행 공간에 있고 오른손만 벽에 닿아 몸통이 부딪히기 전의 위치로 성립합니다. 다만 담장이 화면 오른쪽에서 차지하는 면적은 지시한 좁은 면보다 큽니다. 반사면은 없습니다.",
        "entities": "인물은 한 명이며 검은 머리, 검은 갓과 머리띠를 착용한 성인 남성으로 참조의 주요 식별 요소를 유지합니다. 얼굴 대부분이 가려져 얼굴 일치 여부는 확인할 수 없습니다. 회색 옷과 혈흔은 맞지만 상의가 참조의 긴 도포보다 짧아 바지가 크게 드러납니다. 등에 박힌 낫은 닳은 목재 손잡이와 굽은 어두운 금속 날을 갖추며 혈흔이 관입 위치에 모여 있습니다. 다만 날의 곡률과 손잡이 비례는 참조와 완전히 같지는 않습니다. 바닥에는 여러 떨어진 날붙이와 부서진 좌판이 보이지만 바닥 혈흔은 뚜렷하지 않습니다. 추가 인물이나 읽을 수 있는 글자는 없습니다.",
        "hard_violations": [],
        "physics": "양쪽 허벅지는 프레임 아래로 이어지고 발은 잘려 있어 접지 상태를 직접 확인할 수 없지만, 몸이 통째로 공중에 떠 있지는 않습니다. 벌어진 하체와 기운 몸통은 도주 중 균형이 깨진 자세로 가능하며 오른손은 벽에 실제로 닿아 추가 지지점을 만듭니다. 왼팔은 균형을 잡듯 벌어져 있습니다. 낫은 혈흔과 옷 주름이 모이는 등 관입부에 고정되어 노출된 손잡이를 지지합니다. 순수하게 앞으로 꺾이는 동작보다는 오른쪽으로 넘어지는 성분이 크지만 피격 직후 벽으로 쏠리는 순간으로 읽힙니다."
       }
      ],
      "all_candidates_fail": false
     },
     "normalized": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 6,
        "verdict_ko": "앞으로 숙여지는 중간 샷은 맞지만 충격보다 느린 굽힘처럼 보이며, 갓과 등 상처의 표현이 빠지고 오른쪽 외벽 구조도 참조와 덜 맞습니다."
       },
       {
        "label": "A",
        "score": 8,
        "verdict_ko": "등 중앙에 박힌 낫과 상처, 벽 쪽으로 무너지는 동작, 갓과 돌기단 흙벽이 핵심 지시를 더 충실히 구현하지만 의복 길이와 낫의 비례는 참조와 다릅니다."
       }
      ],
      "all_candidates_fail": false,
      "readings": [
       {
        "label": "B",
        "direction": "남자는 화면 오른쪽 앞의 흙벽을 향해 고개와 상체를 숙이고 있으며 카메라를 보지 않습니다. 낫 손잡이는 왼쪽 뒤로 돌출되고 굽은 날 끝은 등 중앙보다 조금 높은 부위로 들어갑니다. 몸의 진행 방향과 충돌할 벽은 연결되지만 팔이 아래로 처져 있어 급격한 피격 순간의 방향성은 약합니다.",
        "built_space": "오른쪽에는 초가 처마 아래 흙벽 한 면과 그 왼쪽 목재 문 한 곳이 크게 보입니다. 왼쪽과 중앙 후경에는 초가 건물 여러 채, 목재 기둥과 무너진 좌판이 있는 흙길이 이어집니다. 참조의 흙·목재·짚 재료는 따르지만, 오른쪽의 돌기단을 둔 긴 담장 대신 건물 처마와 문이 붙은 벽이 강조됩니다. 벽은 좁은 우측 면보다는 넓은 전경 면을 차지합니다. 인물은 벽 바로 앞 골목에 있고 반사면은 없습니다.",
        "entities": "보이는 사람은 검은 머리를 묶은 성인 남성 한 명뿐이며 한국인 설정과 외관상 충돌하지 않습니다. 얼굴은 옆면 일부만 보여 정확한 동일인 여부를 판단하기 어렵습니다. 회색 긴 도포와 허리띠는 참조와 유사하지만 검은 갓은 없고 옷의 혈흔과 손상도 훨씬 적습니다. 낫 한 자루는 목재 손잡이와 어두운 굽은 금속 날을 갖추지만 참조보다 날이 짧고 둥글게 보입니다. 등에는 뚜렷한 출혈이나 깊은 상처가 보이지 않습니다. 후경의 부서진 좌판은 보이나 떨어진 무기와 바닥 혈흔은 명확하게 식별되지 않습니다. 읽을 수 있는 글자는 없습니다.",
        "hard_violations": [],
        "physics": "하체와 발은 프레임 밖으로 이어지며, 몸 전체가 공중에 떠 있는 모습은 아닙니다. 골반 위에서 상체를 앞으로 굽힌 자세 자체는 가능합니다. 낫은 날 끝이 옷과 등으로 들어간 접점에 의해 지지되는 것으로 보이고 손잡이만 뒤로 노출됩니다. 다만 옷의 당김이나 상처 반응이 약해 깊숙한 관입과 갑작스러운 충격의 물리적 증거가 부족합니다. 손이나 몸통이 벽에 닿은 모습은 아직 없습니다."
       },
       {
        "label": "A",
        "direction": "남자는 카메라에서 멀어지는 방향으로 몸을 틀면서 오른쪽 벽으로 기울고, 오른손을 그 벽에 뻗어 접촉합니다. 얼굴은 진행 방향으로 돌아가 눈의 정확한 시선은 보이지 않습니다. 낫 손잡이는 왼쪽 뒤·아래로 돌출되고 굽은 날 끝은 혈흔이 있는 등 중앙으로 들어갑니다. 등 상처를 가리지 않으면서 벽과의 임박한 몸통 충돌을 보여줍니다.",
        "built_space": "오른쪽에는 돌기단 위의 긴 흙담 한 면이 골목을 따라 이어지고, 왼쪽에는 목재 기둥으로 처마를 받친 초가 건물들이 연속합니다. 흙길은 중앙 후경의 초가집들 사이로 좁아집니다. 담장 재료, 기단, 초가지붕과 골목의 관계가 참조에 더 가깝습니다. 왼쪽 전경에는 부서진 좌판 목재가 놓여 있습니다. 인물은 담장 옆 통행 공간에 있고 오른손만 벽에 닿아 몸통이 부딪히기 전의 위치로 성립합니다. 다만 담장이 화면 오른쪽에서 차지하는 면적은 지시한 좁은 면보다 큽니다. 반사면은 없습니다.",
        "entities": "인물은 한 명이며 검은 머리, 검은 갓과 머리띠를 착용한 성인 남성으로 참조의 주요 식별 요소를 유지합니다. 얼굴 대부분이 가려져 얼굴 일치 여부는 확인할 수 없습니다. 회색 옷과 혈흔은 맞지만 상의가 참조의 긴 도포보다 짧아 바지가 크게 드러납니다. 등에 박힌 낫은 닳은 목재 손잡이와 굽은 어두운 금속 날을 갖추며 혈흔이 관입 위치에 모여 있습니다. 다만 날의 곡률과 손잡이 비례는 참조와 완전히 같지는 않습니다. 바닥에는 여러 떨어진 날붙이와 부서진 좌판이 보이지만 바닥 혈흔은 뚜렷하지 않습니다. 추가 인물이나 읽을 수 있는 글자는 없습니다.",
        "hard_violations": [],
        "physics": "양쪽 허벅지는 프레임 아래로 이어지고 발은 잘려 있어 접지 상태를 직접 확인할 수 없지만, 몸이 통째로 공중에 떠 있지는 않습니다. 벌어진 하체와 기운 몸통은 도주 중 균형이 깨진 자세로 가능하며 오른손은 벽에 실제로 닿아 추가 지지점을 만듭니다. 왼팔은 균형을 잡듯 벌어져 있습니다. 낫은 혈흔과 옷 주름이 모이는 등 관입부에 고정되어 노출된 손잡이를 지지합니다. 순수하게 앞으로 꺾이는 동작보다는 오른쪽으로 넘어지는 성분이 크지만 피격 직후 벽으로 쏠리는 순간으로 읽힙니다."
       }
      ]
     },
     "ok": true
    }
   ],
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "slot_winner_match": true,
   "slot_winner": {
    "gemini-pro": "A",
    "gpt-high": "A"
   },
   "route": "cross_slot_agree"
  },
  "dual": {
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "normalized": {
    "A": 2.0,
    "B": 1.194
   },
   "adjusted": {
    "A": 2.0,
    "B": 0.944
   },
   "violations": {
    "B": [
     "[gemini-pro] 우측 벽의 구조가 레퍼런스(지붕 없는 흙벽)와 달리 초가지붕이 덮인 형태로 변형됨"
    ]
   },
   "per_model_winner": {
    "gemini-pro": "A",
    "gpt-high": "A"
   },
   "agreed": true
  },
  "totals": {
   "A": 2000,
   "B": 944
  },
  "selected": "A",
  "ranking": [
   "A",
   "B"
  ],
  "verdicts": [
   {
    "label": "A",
    "score": 2000,
    "verdict_ko": "벽에 부딪히며 상체가 쏠리는 '미드임팩트'의 순간을 역동적으로 잘 포착했으며, 레퍼런스의 장소 구조와 소품을 충실히 구현했습니다."
   },
   {
    "label": "B",
    "score": 944,
    "verdict_ko": "벽에 부딪히는 핵심 액션이 누락된 채 단순히 걷고 있으며, 우측 흙벽의 구조를 임의로 변형하고 캐릭터의 갓을 빠뜨렸습니다.  ★위반: [gemini-pro] 우측 벽의 구조가 레퍼런스(지붕 없는 흙벽)와 달리 초가지붕이 덮인 형태로 변형됨"
   }
  ],
  "refs": [
   {
    "label": "LOCATION STRUCTURE PHOTOGRAPH — the confirmed photograph of this exact place and its fixed structure: it is the SINGLE authority for the location, the structure's shape, proportions, materials, colors, openings and every permanent site detail. Never copy its camera framing, time of day or lighting — the shot text is the authority for those.",
    "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/db00517b-165b-4005-a287-f28936b59b95/images/background_chain/seed_bg_market_sel.png",
    "asset_id": "32b543ac-1086-4282-b907-3fcd465b580a",
    "role": "location_seed_bg"
   },
   {
    "label": "CHARACTER REFERENCE — 박상현: the exact person appearing in this shot; match face, hair and build exactly.",
    "path": "<bytes:1167996>",
    "asset_id": "6e1bedfc-d3f9-4391-a49e-1ca6e50bf276",
    "role": "character_ref"
   },
   {
    "label": "PROP REFERENCE — 낫: the exact object appearing in this shot; match its look, material and wear exactly.",
    "path": "<bytes:977465>",
    "asset_id": "e9884693-b2aa-4769-9ee0-4c7b50db40c2",
    "role": "prop_ref"
   }
  ],
  "critique_skipped": true,
  "shot_run_produced": false,
  "shot_run_uid": "06aa1438-e5f6-75a9-b270-a97b42a4d8df",
  "bgfirst": {
   "bg_path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/db00517b-165b-4005-a287-f28936b59b95/images/272bd7af-9294-4490-98c6-88a40e9deacc/scene/recipe/S57sh7__bgfirst_bg.png",
   "bg_asset_id": "7609081c-7d2f-4cd9-83f7-fa5f622edfa3",
   "bg_record_key": "S57sh7::bgfirst_bg",
   "chain_winner": true,
   "authority": "seed_bg"
  },
  "ref_mode": "재투영 배경+콘티+엔티티 (2택1: 체인 승)",
  "share_plan": {
   "ref_plan": "background"
  },
  "lane_policy": "ab_select_ready"
 },
 "S57sh7::cine": {
  "applied": true,
  "attempted_at": "2026-09-09T07:30:06.140526+00:00",
  "fingerprint": "9f2c3575e304b000d08a40302054312a6d85e9b59c7234cc6156945c7021378b",
  "fingerprint_version": 2,
  "provider": "mai",
  "endpoint": "openrouter/chat-completions",
  "model": "microsoft/mai-image-2.6",
  "pack": "24.202608252115",
  "source_file": "S57sh7_sel.png",
  "source_sha256": "4b408fbf4e389d2d36c7bbf3518350395d9ccde2076c82dba4f90eaa734bf6d0",
  "file": "S57sh7_cine.png",
  "staged_sha256": "5b503ee0a3716d98831e05b9f2907c337f126dbc38c93d3da82c6ab316278623",
  "latency_ms": 28035
 },
 "S57sh12::signage": {
  "fp": "1c34a0239b43ed3f",
  "inscriptions": [],
  "cues": [],
  "dropped": []
 },
 "S57sh12": {
  "input_fingerprint": "f609a1c98f6af415",
  "prompt": "Create ONE FINAL photorealistic live-action film still of the moment below — early 20th-century Korea, Jeolla region; all people are Korean unless explicitly stated otherwise. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 텅 빈 저잣거리 한가운데, 조선인 일진회 시체들과 뒤섞여 쓰러져 있는 이관회의 널브러진 전신.\n\nLOCATION (lock): The deserted center of the market's dirt thoroughfare, where bodies remain among the abandoned stalls. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: wide shot\n- KEY BACKGROUND ELEMENTS: 저잣거리 흙바닥 (Deserted of moving people, with bodies lying in its middle); used as Open foreground and unobstructed ground around 이관회 create separation; surrounding bodies lie at uneven intervals with distinct shoulder, hip, and limb orientations.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Unembellished daylight, muted saturation, and sober contrast leave the deserted space and still bodies exposed without romanticizing the loss.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The people visible in that photo are NOT in this shot: never carry their faces, bodies or clothing onto anyone here. Each person in THIS shot is defined solely by the PEOPLE list, their own CHARACTER REFERENCE images and their wardrobe notes. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The market is deserted and littered with broken stall debris, dropped weapons and blood.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 이관회 (한국인 남성, 청년 나이대, 자연스러운 검은 머리) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.",
  "roll_prompts": {
   "A": "Create ONE FINAL photorealistic live-action film still of the moment below — early 20th-century Korea, Jeolla region; all people are Korean unless explicitly stated otherwise. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 텅 빈 저잣거리 한가운데, 조선인 일진회 시체들과 뒤섞여 쓰러져 있는 이관회의 널브러진 전신.\n\nLOCATION (lock): The deserted center of the market's dirt thoroughfare, where bodies remain among the abandoned stalls. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: wide shot\n- KEY BACKGROUND ELEMENTS: 저잣거리 흙바닥 (Deserted of moving people, with bodies lying in its middle); used as Open foreground and unobstructed ground around 이관회 create separation; surrounding bodies lie at uneven intervals with distinct shoulder, hip, and limb orientations.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Unembellished daylight, muted saturation, and sober contrast leave the deserted space and still bodies exposed without romanticizing the loss.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The people visible in that photo are NOT in this shot: never carry their faces, bodies or clothing onto anyone here. Each person in THIS shot is defined solely by the PEOPLE list, their own CHARACTER REFERENCE images and their wardrobe notes. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The market is deserted and littered with broken stall debris, dropped weapons and blood.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 이관회 (한국인 남성, 청년 나이대, 자연스러운 검은 머리) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.",
   "B": "Create ONE FINAL photorealistic live-action film still of the moment below — early 20th-century Korea, Jeolla region; all people are Korean unless explicitly stated otherwise. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 텅 빈 저잣거리 한가운데, 조선인 일진회 시체들과 뒤섞여 쓰러져 있는 이관회의 널브러진 전신.\n\nLOCATION (lock): The deserted center of the market's dirt thoroughfare, where bodies remain among the abandoned stalls. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: wide shot\n- KEY BACKGROUND ELEMENTS: 저잣거리 흙바닥 (Deserted of moving people, with bodies lying in its middle); used as Open foreground and unobstructed ground around 이관회 create separation; surrounding bodies lie at uneven intervals with distinct shoulder, hip, and limb orientations.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Unembellished daylight, muted saturation, and sober contrast leave the deserted space and still bodies exposed without romanticizing the loss.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The people visible in that photo are NOT in this shot: never carry their faces, bodies or clothing onto anyone here. Each person in THIS shot is defined solely by the PEOPLE list, their own CHARACTER REFERENCE images and their wardrobe notes. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The market is deserted and littered with broken stall debris, dropped weapons and blood.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 이관회 (한국인 남성, 청년 나이대, 자연스러운 검은 머리) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.\n\nCOMPOSITION VARIATION (this candidate only): keep the place, the people and the moment exactly as specified, and keep from the CAMERA & FRAME contract its subject scale, screen placement and key background elements exactly. For this candidate only, its camera angle and height are deliberately left open: choose a camera position distinctly different from the obvious one, as a film director picking a second setup on the same blocking. Stay true to any explicit framing the SHOT TEXT itself dictates."
  },
  "shot_run_spend_attempt_count": 1,
  "readings": [
   {
    "label": "A",
    "direction": "시체들이 흙바닥에 널브러져 있으며, 중앙 인물의 시선은 위쪽 하늘을 향함.",
    "built_space": "양옆의 나무 가판대와 배경의 문 등 레퍼런스 이미지에 나타난 저잣거리 구조를 재현함.",
    "entities": "이관회로 보이는 청년과 주변 시체들이 등장하나, 다수 인물의 해부학적 형태가 무너짐.",
    "hard_violations": [
     "[gemini-pro] 물리적으로 불가능한 해부학: 우측 하단에 엎드린 시체의 머리가 목 부분에서 완전히 소실됨.",
     "[gemini-pro] 물리적으로 불가능한 해부학: 중앙 인물의 왼쪽 어깨 뒤에서 몸통이 들어갈 공간 없이 여분의 얼굴과 팔이 튀어나와 있음."
    ],
    "physics": "시체들이 바닥에 누워 지탱되고 있으나, 일부 신체가 물리적으로 불가능하게 절단되거나 융합되어 있음."
   },
   {
    "label": "B",
    "direction": "시체들이 흙바닥에 널브러져 있으며, 중앙 인물은 지면을 바라보며 좌측 하단을 향해 누워 있음.",
    "built_space": "가판대, 부서진 잔해, 배경의 문 등 레퍼런스 이미지의 공간적 특성과 고정된 요소들을 충실히 재현함.",
    "entities": "이관회와 여러 시체들이 등장함. 그러나 프롬프트에서 명시적으로 금지한 이전 샷 인물의 특징(등에 피 묻은 마름모 자국이 있는 옷, 들고 있는 낫)을 중앙 인물에게 그대로 적용하는 오류를 범함.",
    "hard_violations": [
     "[gpt-high] 후면 출입문 위 간판에 읽히는 한글 글자가 노출되어, 어디에도 읽을 수 있는 문자가 없어야 한다는 조건을 위반한다."
    ],
    "physics": "모든 시체와 바닥에 놓인 낫이 지면에 자연스럽게 밀착되어 지탱되고 있음."
   }
  ],
  "cross_model_order": {
   "policy": "cross_model_order_v2_slot0_fwd_slot1_rev_parallel2",
   "slots": [
    {
     "model": "gemini-pro",
     "order": "forward",
     "display_to_canonical": {
      "A": "A",
      "B": "B"
     },
     "raw": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 7,
        "verdict_ko": "이전 샷 인물의 특징(피 묻은 옷, 낫)을 반영하지 말라는 지시를 어겼으나, 해부학적 결함 없이 와이드 샷의 구도와 널브러진 시체들을 충실히 구현함."
       },
       {
        "label": "A",
        "score": 2,
        "verdict_ko": "우측 하단의 머리가 없는 시체와 중앙 인물 어깨 뒤에 몸통 없이 붙어 있는 얼굴 등 심각한 해부학적 오류로 인해 실격됨."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "시체들이 흙바닥에 널브러져 있으며, 중앙 인물의 시선은 위쪽 하늘을 향함.",
        "built_space": "양옆의 나무 가판대와 배경의 문 등 레퍼런스 이미지에 나타난 저잣거리 구조를 재현함.",
        "entities": "이관회로 보이는 청년과 주변 시체들이 등장하나, 다수 인물의 해부학적 형태가 무너짐.",
        "hard_violations": [
         "물리적으로 불가능한 해부학: 우측 하단에 엎드린 시체의 머리가 목 부분에서 완전히 소실됨.",
         "물리적으로 불가능한 해부학: 중앙 인물의 왼쪽 어깨 뒤에서 몸통이 들어갈 공간 없이 여분의 얼굴과 팔이 튀어나와 있음."
        ],
        "physics": "시체들이 바닥에 누워 지탱되고 있으나, 일부 신체가 물리적으로 불가능하게 절단되거나 융합되어 있음."
       },
       {
        "label": "B",
        "direction": "시체들이 흙바닥에 널브러져 있으며, 중앙 인물은 지면을 바라보며 좌측 하단을 향해 누워 있음.",
        "built_space": "가판대, 부서진 잔해, 배경의 문 등 레퍼런스 이미지의 공간적 특성과 고정된 요소들을 충실히 재현함.",
        "entities": "이관회와 여러 시체들이 등장함. 그러나 프롬프트에서 명시적으로 금지한 이전 샷 인물의 특징(등에 피 묻은 마름모 자국이 있는 옷, 들고 있는 낫)을 중앙 인물에게 그대로 적용하는 오류를 범함.",
        "hard_violations": [],
        "physics": "모든 시체와 바닥에 놓인 낫이 지면에 자연스럽게 밀착되어 지탱되고 있음."
       }
      ],
      "all_candidates_fail": false
     },
     "normalized": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 7,
        "verdict_ko": "이전 샷 인물의 특징(피 묻은 옷, 낫)을 반영하지 말라는 지시를 어겼으나, 해부학적 결함 없이 와이드 샷의 구도와 널브러진 시체들을 충실히 구현함."
       },
       {
        "label": "A",
        "score": 2,
        "verdict_ko": "우측 하단의 머리가 없는 시체와 중앙 인물 어깨 뒤에 몸통 없이 붙어 있는 얼굴 등 심각한 해부학적 오류로 인해 실격됨."
       }
      ],
      "all_candidates_fail": false,
      "readings": [
       {
        "label": "A",
        "direction": "시체들이 흙바닥에 널브러져 있으며, 중앙 인물의 시선은 위쪽 하늘을 향함.",
        "built_space": "양옆의 나무 가판대와 배경의 문 등 레퍼런스 이미지에 나타난 저잣거리 구조를 재현함.",
        "entities": "이관회로 보이는 청년과 주변 시체들이 등장하나, 다수 인물의 해부학적 형태가 무너짐.",
        "hard_violations": [
         "물리적으로 불가능한 해부학: 우측 하단에 엎드린 시체의 머리가 목 부분에서 완전히 소실됨.",
         "물리적으로 불가능한 해부학: 중앙 인물의 왼쪽 어깨 뒤에서 몸통이 들어갈 공간 없이 여분의 얼굴과 팔이 튀어나와 있음."
        ],
        "physics": "시체들이 바닥에 누워 지탱되고 있으나, 일부 신체가 물리적으로 불가능하게 절단되거나 융합되어 있음."
       },
       {
        "label": "B",
        "direction": "시체들이 흙바닥에 널브러져 있으며, 중앙 인물은 지면을 바라보며 좌측 하단을 향해 누워 있음.",
        "built_space": "가판대, 부서진 잔해, 배경의 문 등 레퍼런스 이미지의 공간적 특성과 고정된 요소들을 충실히 재현함.",
        "entities": "이관회와 여러 시체들이 등장함. 그러나 프롬프트에서 명시적으로 금지한 이전 샷 인물의 특징(등에 피 묻은 마름모 자국이 있는 옷, 들고 있는 낫)을 중앙 인물에게 그대로 적용하는 오류를 범함.",
        "hard_violations": [],
        "physics": "모든 시체와 바닥에 놓인 낫이 지면에 자연스럽게 밀착되어 지탱되고 있음."
       }
      ]
     },
     "ok": true
    },
    {
     "model": "gpt-high",
     "order": "reverse",
     "display_to_canonical": {
      "A": "B",
      "B": "A"
     },
     "raw": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 3,
        "verdict_ko": "이관회의 전신과 불규칙하게 쓰러진 시신들을 와이드 숏으로 담았지만, 후면 간판의 읽히는 글자가 명시적인 무문자 조건을 위반한다."
       },
       {
        "label": "B",
        "score": 8,
        "verdict_ko": "같은 목조 저잣거리의 낮 풍경에서 이관회의 널브러진 전신과 서로 다른 방향의 시신들을 보여주며, 읽히는 글자 없이 핵심 장면을 충실히 구현한다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "중앙 청년은 머리를 화면 왼쪽, 발을 오른쪽으로 두고 옆으로 쓰러져 있으며 얼굴은 카메라 쪽 바닥을 향한다. 주변 시신들은 앞쪽 양 모서리, 중경 좌우, 후경에 서로 다른 방향으로 놓여 있다. 중앙 손 가까이의 낫은 날이 왼쪽으로 굽어 바닥에 놓였으며, 누구를 겨누거나 이동하는 인물은 없다.",
        "built_space": "좌우 목조 점포 사이 흙길 끝에 닫힌 문 하나와 기와 처마가 보인다. 왼쪽에는 천을 덮은 진열대와 바구니들, 오른쪽에는 기울어진 좌판 더미와 큰 나무 상자 하나, 화면 가장자리의 다른 상자 하나가 있다. 기준 사진의 낡은 목재·천막·무너진 좌판 배치는 대체로 유지되지만, 후면 문 위에 글자가 보이는 간판이 노출되어 있다. 중앙 인물 주변에는 흙바닥 여백이 있으나 전경 양쪽의 시신이 크게 잘려 들어온다.",
        "entities": "총 일곱 명의 쓰러진 인물이 보이며, 중앙 인물은 자연스러운 검은 머리의 동아시아계 청년 남성으로 요청된 이관회의 외형 범주에 부합한다. 주변 인물들도 당시 조선인의 복식으로 읽히지만 일진회 소속 자체를 확인할 표지는 없다. 중앙 인물의 밝은 옷에는 혈흔이 있고 낫, 부서진 목재, 버려진 시장 물품이 보인다. 배경 간판에는 식별 가능한 한글 글자가 있어 글자 금지 조건과 맞지 않는다.",
        "hard_violations": [
         "후면 출입문 위 간판에 읽히는 한글 글자가 노출되어, 어디에도 읽을 수 있는 문자가 없어야 한다는 조건을 위반한다."
        ],
        "physics": "중앙 인물은 머리·어깨·옆구리·다리가 흙바닥에 닿아 지지되고, 손 가까이의 낫도 바닥에 놓여 있다. 주변 시신들의 몸통과 팔다리 역시 지면에 닿아 있으며, 오른쪽 인물의 상체는 뒤의 짐에 기대어 있다. 기울어진 좌판은 하단이 지면에 닿고 주변 구조물에 기대어 있어, 뚜렷하게 떠 있는 몸이나 물체는 없다."
       },
       {
        "label": "B",
        "direction": "중앙 청년은 머리를 골목 안쪽, 벌어진 두 발을 카메라 쪽으로 두고 바로 누워 있으며 눈을 감은 얼굴은 위를 향한다. 앞쪽 좌우 시신은 각각 바깥쪽으로 얼굴을 돌리고 웅크렸고, 뒤쪽 두 시신은 길을 가로지르는 방향으로 누워 있다. 바닥의 긴 막대형 물체들은 사람을 겨누고 있지 않으며 움직이는 인물도 없다.",
        "built_space": "좌우 목조 점포, 왼쪽의 천 덮인 진열대와 바구니들, 오른쪽의 무너진 좌판 더미와 큰 나무 상자 하나, 골목 끝 닫힌 문 하나가 보인다. 기준 사진의 흙길과 낡은 시장 재료 및 주요 배치를 잘 따른다. 중앙 인물의 전신은 가려지지 않고 손과 발 주변에 흙바닥이 드러나지만, 머리 바로 뒤에 다른 시신이 가까이 놓여 분리는 다소 약하다. 읽히는 간판이나 불가능한 반사는 보이지 않는다.",
        "entities": "총 다섯 명의 쓰러진 남성이 보인다. 중앙 인물은 검은 머리의 동아시아계 청년 남성이고 밝은 전통 복식을 입어 요청된 이관회의 외형과 시대에 부합한다. 주변 네 명은 구체적인 신체를 가진 시신으로 표현되며 일진회 소속은 외형만으로 확정할 수 없다. 흙길, 파손된 좌판과 목재, 여러 혈흔은 보이지만 떨어진 긴 물체들이 무기인지는 선명하지 않다.",
        "hard_violations": [],
        "physics": "중앙 인물은 등·골반·팔·다리가 지면에 놓인 자세이며, 머리 뒤는 가까이 누운 다른 시신과 일부 겹쳐 보인다. 뒤쪽 왼쪽 인물의 세운 무릎은 바닥에 닿은 발과 누운 골반으로 지지되고, 나머지 시신들도 옆구리나 몸통과 팔다리가 땅에 닿아 있다. 오른쪽 좌판 더미는 지면과 인접 구조물에 기대며, 혈액은 흙 표면에 고여 있어 지지 없이 떠 있는 요소는 보이지 않는다."
       }
      ],
      "all_candidates_fail": false
     },
     "normalized": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 3,
        "verdict_ko": "이관회의 전신과 불규칙하게 쓰러진 시신들을 와이드 숏으로 담았지만, 후면 간판의 읽히는 글자가 명시적인 무문자 조건을 위반한다."
       },
       {
        "label": "A",
        "score": 8,
        "verdict_ko": "같은 목조 저잣거리의 낮 풍경에서 이관회의 널브러진 전신과 서로 다른 방향의 시신들을 보여주며, 읽히는 글자 없이 핵심 장면을 충실히 구현한다."
       }
      ],
      "all_candidates_fail": false,
      "readings": [
       {
        "label": "B",
        "direction": "중앙 청년은 머리를 화면 왼쪽, 발을 오른쪽으로 두고 옆으로 쓰러져 있으며 얼굴은 카메라 쪽 바닥을 향한다. 주변 시신들은 앞쪽 양 모서리, 중경 좌우, 후경에 서로 다른 방향으로 놓여 있다. 중앙 손 가까이의 낫은 날이 왼쪽으로 굽어 바닥에 놓였으며, 누구를 겨누거나 이동하는 인물은 없다.",
        "built_space": "좌우 목조 점포 사이 흙길 끝에 닫힌 문 하나와 기와 처마가 보인다. 왼쪽에는 천을 덮은 진열대와 바구니들, 오른쪽에는 기울어진 좌판 더미와 큰 나무 상자 하나, 화면 가장자리의 다른 상자 하나가 있다. 기준 사진의 낡은 목재·천막·무너진 좌판 배치는 대체로 유지되지만, 후면 문 위에 글자가 보이는 간판이 노출되어 있다. 중앙 인물 주변에는 흙바닥 여백이 있으나 전경 양쪽의 시신이 크게 잘려 들어온다.",
        "entities": "총 일곱 명의 쓰러진 인물이 보이며, 중앙 인물은 자연스러운 검은 머리의 동아시아계 청년 남성으로 요청된 이관회의 외형 범주에 부합한다. 주변 인물들도 당시 조선인의 복식으로 읽히지만 일진회 소속 자체를 확인할 표지는 없다. 중앙 인물의 밝은 옷에는 혈흔이 있고 낫, 부서진 목재, 버려진 시장 물품이 보인다. 배경 간판에는 식별 가능한 한글 글자가 있어 글자 금지 조건과 맞지 않는다.",
        "hard_violations": [
         "후면 출입문 위 간판에 읽히는 한글 글자가 노출되어, 어디에도 읽을 수 있는 문자가 없어야 한다는 조건을 위반한다."
        ],
        "physics": "중앙 인물은 머리·어깨·옆구리·다리가 흙바닥에 닿아 지지되고, 손 가까이의 낫도 바닥에 놓여 있다. 주변 시신들의 몸통과 팔다리 역시 지면에 닿아 있으며, 오른쪽 인물의 상체는 뒤의 짐에 기대어 있다. 기울어진 좌판은 하단이 지면에 닿고 주변 구조물에 기대어 있어, 뚜렷하게 떠 있는 몸이나 물체는 없다."
       },
       {
        "label": "A",
        "direction": "중앙 청년은 머리를 골목 안쪽, 벌어진 두 발을 카메라 쪽으로 두고 바로 누워 있으며 눈을 감은 얼굴은 위를 향한다. 앞쪽 좌우 시신은 각각 바깥쪽으로 얼굴을 돌리고 웅크렸고, 뒤쪽 두 시신은 길을 가로지르는 방향으로 누워 있다. 바닥의 긴 막대형 물체들은 사람을 겨누고 있지 않으며 움직이는 인물도 없다.",
        "built_space": "좌우 목조 점포, 왼쪽의 천 덮인 진열대와 바구니들, 오른쪽의 무너진 좌판 더미와 큰 나무 상자 하나, 골목 끝 닫힌 문 하나가 보인다. 기준 사진의 흙길과 낡은 시장 재료 및 주요 배치를 잘 따른다. 중앙 인물의 전신은 가려지지 않고 손과 발 주변에 흙바닥이 드러나지만, 머리 바로 뒤에 다른 시신이 가까이 놓여 분리는 다소 약하다. 읽히는 간판이나 불가능한 반사는 보이지 않는다.",
        "entities": "총 다섯 명의 쓰러진 남성이 보인다. 중앙 인물은 검은 머리의 동아시아계 청년 남성이고 밝은 전통 복식을 입어 요청된 이관회의 외형과 시대에 부합한다. 주변 네 명은 구체적인 신체를 가진 시신으로 표현되며 일진회 소속은 외형만으로 확정할 수 없다. 흙길, 파손된 좌판과 목재, 여러 혈흔은 보이지만 떨어진 긴 물체들이 무기인지는 선명하지 않다.",
        "hard_violations": [],
        "physics": "중앙 인물은 등·골반·팔·다리가 지면에 놓인 자세이며, 머리 뒤는 가까이 누운 다른 시신과 일부 겹쳐 보인다. 뒤쪽 왼쪽 인물의 세운 무릎은 바닥에 닿은 발과 누운 골반으로 지지되고, 나머지 시신들도 옆구리나 몸통과 팔다리가 땅에 닿아 있다. 오른쪽 좌판 더미는 지면과 인접 구조물에 기대며, 혈액은 흙 표면에 고여 있어 지지 없이 떠 있는 요소는 보이지 않는다."
       }
      ]
     },
     "ok": true
    }
   ],
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "slot_winner_match": false,
   "slot_winner": {
    "gemini-pro": "B",
    "gpt-high": "A"
   },
   "route": "cross_slot_combined"
  },
  "dual": {
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "normalized": {
    "A": 1.286,
    "B": 1.375
   },
   "adjusted": {
    "A": 1.036,
    "B": 1.125
   },
   "violations": {
    "A": [
     "[gemini-pro] 물리적으로 불가능한 해부학: 우측 하단에 엎드린 시체의 머리가 목 부분에서 완전히 소실됨.",
     "[gemini-pro] 물리적으로 불가능한 해부학: 중앙 인물의 왼쪽 어깨 뒤에서 몸통이 들어갈 공간 없이 여분의 얼굴과 팔이 튀어나와 있음."
    ],
    "B": [
     "[gpt-high] 후면 출입문 위 간판에 읽히는 한글 글자가 노출되어, 어디에도 읽을 수 있는 문자가 없어야 한다는 조건을 위반한다."
    ]
   },
   "per_model_winner": {
    "gemini-pro": "B",
    "gpt-high": "A"
   },
   "agreed": false
  },
  "totals": {
   "B": 1125,
   "A": 1036
  },
  "selected": "B",
  "ranking": [
   "B",
   "A"
  ],
  "verdicts": [
   {
    "label": "B",
    "score": 1125,
    "verdict_ko": "이전 샷 인물의 특징(피 묻은 옷, 낫)을 반영하지 말라는 지시를 어겼으나, 해부학적 결함 없이 와이드 샷의 구도와 널브러진 시체들을 충실히 구현함.  ★위반: [gpt-high] 후면 출입문 위 간판에 읽히는 한글 글자가 노출되어, 어디에도 읽을 수 있는 문자가 없어야 한다는 조건을 위반한다."
   },
   {
    "label": "A",
    "score": 1036,
    "verdict_ko": "우측 하단의 머리가 없는 시체와 중앙 인물 어깨 뒤에 몸통 없이 붙어 있는 얼굴 등 심각한 해부학적 오류로 인해 실격됨.  ★위반: [gemini-pro] 물리적으로 불가능한 해부학: 우측 하단에 엎드린 시체의 머리가 목 부분에서 완전히 소실됨. / [gemini-pro] 물리적으로 불가능한 해부학: 중앙 인물의 왼쪽 어깨 뒤에서 몸통이 들어갈 공간 없이 여분의 얼굴과 팔이 튀어나와 있음."
   }
  ],
  "refs": [
   {
    "label": "PREVIOUS SHOT STILL — a visually related earlier shot of this same place: the location's look, materials, fixed features and lighting mood are LOCKED to this photo; never copy its camera framing. The people visible in that photo are NOT in this shot: never carry their faces, bodies or clothing onto anyone here. Each person in THIS shot is defined solely by the PEOPLE list, their own CHARACTER REFERENCE images and their wardrobe notes. If this photo shows a character who cannot move (dead or unconscious), that character's exact pose, position and orientation are ALSO LOCKED — treat their body as a fixed prop of the set that only the camera moves around.",
    "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/db00517b-165b-4005-a287-f28936b59b95/images/272bd7af-9294-4490-98c6-88a40e9deacc/scene/recipe/S55sh40_sel.png",
    "asset_id": "02e2a7a0-5081-4722-ad1d-54610e67f78e",
    "role": "prev_still"
   }
  ],
  "critique_skipped": true,
  "shot_run_produced": false,
  "shot_run_uid": "06aa1438-e8fb-7f93-baa1-8be5472c0a79",
  "ref_mode": "prev만 (배경 전용·공유 계획)",
  "share_plan": {
   "ref_plan": "prev",
   "prev_anchor_tag": "S55sh40"
  },
  "lane_policy": "ab_select_bypass:bg_only:share_plan_prev_bgonly",
  "locked_char_refs_excluded": [
   "이관회(C02)"
  ]
 },
 "S57sh12::cine": {
  "applied": true,
  "attempted_at": "2026-09-09T07:32:05.522961+00:00",
  "fingerprint": "e3ddf36dea9eb1af18f799bf1c379485ca8e27c524237f6bf55bf87ef5e3e9e7",
  "fingerprint_version": 2,
  "provider": "mai",
  "endpoint": "openrouter/chat-completions",
  "model": "microsoft/mai-image-2.6",
  "pack": "24.202608252115",
  "source_file": "S57sh12_sel.png",
  "source_sha256": "1d6cfa9066fa79c979284b917603d5ed5c21e93ab6ac43af25af205671207eb2",
  "file": "S57sh12_cine.png",
  "staged_sha256": "f0ba45ef5ed99856a8eca32212ea90833e663c0ae0e7f4c568c863daacb74168",
  "latency_ms": 31821
 },
 "S57sh19::signage": {
  "fp": "63f9478369885472",
  "inscriptions": [],
  "cues": [],
  "dropped": []
 },
 "S57sh19": {
  "input_fingerprint": "4291978c3d0800eb",
  "prompt": "Create ONE FINAL photorealistic live-action film still of the moment below — early 20th-century Korea, Jeolla region; all people are Korean unless explicitly stated otherwise. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 총성이 들리는 산을 향해 맹렬하게 튀어나가며 뒷발로 바닥을 박차고 몸이 막 떠오른 안규홍의 역동적인 뒷모습.\n\nLOCATION (lock): The outdoor market lane leading away from the fallen bodies toward the mountain-bound exit. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: wide shot\n- FRAME LAYOUT: mountain in the direction of departure in the upper-right of the frame, background.\n- KEY BACKGROUND ELEMENTS: 장터 흙바닥 (Beneath 안규홍's departing stride); used as Visible beneath both feet to distinguish a running push-off from a suspended pose; 산 (The destination toward which 안규홍 runs); used as A distant upper-right anchor makes his departure direction readable.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Natural daylight with restrained saturation and consistent contrast preserves the exhausted physical urgency of his departure.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The people visible in that photo are NOT in this shot: never carry their faces, bodies or clothing onto anyone here. Each person in THIS shot is defined solely by the PEOPLE list, their own CHARACTER REFERENCE images and their wardrobe notes. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): Blood, discarded weapons and broken stall debris remain in the deserted market. 안규홍: He runs toward the mountain with two bloodied sickles, blood-soaked hands and a bleeding back wound; his neck remains bare.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 안규홍 (한국인 남성, 청년 나이대, 검은 머리, 단단히 묶은 머리) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.",
  "roll_prompts": {
   "A": "Create ONE FINAL photorealistic live-action film still of the moment below — early 20th-century Korea, Jeolla region; all people are Korean unless explicitly stated otherwise. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 총성이 들리는 산을 향해 맹렬하게 튀어나가며 뒷발로 바닥을 박차고 몸이 막 떠오른 안규홍의 역동적인 뒷모습.\n\nLOCATION (lock): The outdoor market lane leading away from the fallen bodies toward the mountain-bound exit. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: wide shot\n- FRAME LAYOUT: mountain in the direction of departure in the upper-right of the frame, background.\n- KEY BACKGROUND ELEMENTS: 장터 흙바닥 (Beneath 안규홍's departing stride); used as Visible beneath both feet to distinguish a running push-off from a suspended pose; 산 (The destination toward which 안규홍 runs); used as A distant upper-right anchor makes his departure direction readable.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Natural daylight with restrained saturation and consistent contrast preserves the exhausted physical urgency of his departure.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The people visible in that photo are NOT in this shot: never carry their faces, bodies or clothing onto anyone here. Each person in THIS shot is defined solely by the PEOPLE list, their own CHARACTER REFERENCE images and their wardrobe notes. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): Blood, discarded weapons and broken stall debris remain in the deserted market. 안규홍: He runs toward the mountain with two bloodied sickles, blood-soaked hands and a bleeding back wound; his neck remains bare.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 안규홍 (한국인 남성, 청년 나이대, 검은 머리, 단단히 묶은 머리) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.",
   "B": "Create ONE FINAL photorealistic live-action film still of the moment below — early 20th-century Korea, Jeolla region; all people are Korean unless explicitly stated otherwise. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 총성이 들리는 산을 향해 맹렬하게 튀어나가며 뒷발로 바닥을 박차고 몸이 막 떠오른 안규홍의 역동적인 뒷모습.\n\nLOCATION (lock): The outdoor market lane leading away from the fallen bodies toward the mountain-bound exit. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: wide shot\n- FRAME LAYOUT: mountain in the direction of departure in the upper-right of the frame, background.\n- KEY BACKGROUND ELEMENTS: 장터 흙바닥 (Beneath 안규홍's departing stride); used as Visible beneath both feet to distinguish a running push-off from a suspended pose; 산 (The destination toward which 안규홍 runs); used as A distant upper-right anchor makes his departure direction readable.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Natural daylight with restrained saturation and consistent contrast preserves the exhausted physical urgency of his departure.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The people visible in that photo are NOT in this shot: never carry their faces, bodies or clothing onto anyone here. Each person in THIS shot is defined solely by the PEOPLE list, their own CHARACTER REFERENCE images and their wardrobe notes. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): Blood, discarded weapons and broken stall debris remain in the deserted market. 안규홍: He runs toward the mountain with two bloodied sickles, blood-soaked hands and a bleeding back wound; his neck remains bare.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 안규홍 (한국인 남성, 청년 나이대, 검은 머리, 단단히 묶은 머리) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.\n\nCOMPOSITION VARIATION (this candidate only): keep the place, the people and the moment exactly as specified, and keep from the CAMERA & FRAME contract its subject scale, screen placement and key background elements exactly. For this candidate only, its camera angle and height are deliberately left open: choose a camera position distinctly different from the obvious one, as a film director picking a second setup on the same blocking. Stay true to any explicit framing the SHOT TEXT itself dictates."
  },
  "shot_run_spend_attempt_count": 1,
  "readings": [
   {
    "label": "A",
    "direction": "안규홍의 방향과 시선이 시장 골목길을 지나 정면 멀리 있는 산을 향해 있음.",
    "built_space": "흙길 양옆으로 초가집과 우측 흙담이 있으며, 바닥에 부러진 나무 파편과 낫이 흩어져 있음. 이전 샷의 환경 구조와 잘 일치함.",
    "entities": "안규홍(단단히 묶은 검은 머리, 등에 난 핏자국, 붉은 피로 덮인 두 손, 피 묻은 낫 두 자루, 맨 목). 그 외 다른 인물은 등장하지 않음.",
    "hard_violations": [
     "[gemini-pro] 물리적으로 불가능한 연출 (도약의 역학이나 추진력 없이 허공에 꼿꼿하게 서서 부양하고 있음)"
    ],
    "physics": "두 발이 땅에서 떨어져 공중에 떠 있으나, 달리는 도약(launch)을 위한 다리의 구부러짐이나 근육의 움직임이 전혀 없이 똑바로 서 있음. 몸을 지탱하거나 밀어내는 힘이 전혀 보이지 않는 부양 상태임."
   },
   {
    "label": "B",
    "direction": "안규홍이 우측 상단 배경에 솟아 있는 산을 향해 골목길을 마주하고 있음.",
    "built_space": "초가집과 흙담이 늘어선 흙길이며 바닥에 낫과 무기들이 떨어져 있음.",
    "entities": "안규홍(단단히 묶은 머리, 등 상처, 피 묻은 두 손, 낫 두 자루). 샷 텍스트에 없는 누워있는 시체 두 구가 배경 바닥에 추가로 존재함.",
    "hard_violations": [
     "[gemini-pro] 없는 인물 추가 (프롬프트 텍스트에 없는 두 명의 시체가 배경에 등장함)",
     "[gemini-pro] 물리적으로 불가능한 연출 (달리기 추진력이 아닌 기이한 대칭형 점프로 허공에 떠 있음)",
     "[gpt-high] 안규홍만 허용된 샷에 쓰러진 인물 두 명을 새로 추가했다."
    ],
    "physics": "두 발이 땅에서 떨어져 공중에 떠 있으나, 맹렬히 달려나가는 폼이 아니라 양 다리가 무릎에서 뒤로 대칭되게 꺾인 어색한 제자리 점프 상태임. 신체를 자연스럽게 지탱하거나 앞으로 뻗어나가는 역학이 없음."
   }
  ],
  "cross_model_order": {
   "policy": "cross_model_order_v2_slot0_fwd_slot1_rev_parallel2",
   "slots": [
    {
     "model": "gemini-pro",
     "order": "forward",
     "display_to_canonical": {
      "A": "A",
      "B": "B"
     },
     "raw": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 3,
        "verdict_ko": "환경과 인물 외형은 잘 따랐으나, '역동적인 뒷모습'과 '도약'이라는 묘사와 달리 추진력 없이 허공에 꼿꼿이 떠 있는 자세로 렌더링되어 물리적 연출에서 실패함."
       },
       {
        "label": "B",
        "score": 1,
        "verdict_ko": "샷 텍스트에 명시되지 않은 인물(시체)을 배경에 임의로 추가하는 치명적인 오류를 범했으며, 달리기 역학에 맞지 않는 기이한 점프 자세를 보임."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "안규홍의 방향과 시선이 시장 골목길을 지나 정면 멀리 있는 산을 향해 있음.",
        "built_space": "흙길 양옆으로 초가집과 우측 흙담이 있으며, 바닥에 부러진 나무 파편과 낫이 흩어져 있음. 이전 샷의 환경 구조와 잘 일치함.",
        "entities": "안규홍(단단히 묶은 검은 머리, 등에 난 핏자국, 붉은 피로 덮인 두 손, 피 묻은 낫 두 자루, 맨 목). 그 외 다른 인물은 등장하지 않음.",
        "hard_violations": [
         "물리적으로 불가능한 연출 (도약의 역학이나 추진력 없이 허공에 꼿꼿하게 서서 부양하고 있음)"
        ],
        "physics": "두 발이 땅에서 떨어져 공중에 떠 있으나, 달리는 도약(launch)을 위한 다리의 구부러짐이나 근육의 움직임이 전혀 없이 똑바로 서 있음. 몸을 지탱하거나 밀어내는 힘이 전혀 보이지 않는 부양 상태임."
       },
       {
        "label": "B",
        "direction": "안규홍이 우측 상단 배경에 솟아 있는 산을 향해 골목길을 마주하고 있음.",
        "built_space": "초가집과 흙담이 늘어선 흙길이며 바닥에 낫과 무기들이 떨어져 있음.",
        "entities": "안규홍(단단히 묶은 머리, 등 상처, 피 묻은 두 손, 낫 두 자루). 샷 텍스트에 없는 누워있는 시체 두 구가 배경 바닥에 추가로 존재함.",
        "hard_violations": [
         "없는 인물 추가 (프롬프트 텍스트에 없는 두 명의 시체가 배경에 등장함)",
         "물리적으로 불가능한 연출 (달리기 추진력이 아닌 기이한 대칭형 점프로 허공에 떠 있음)"
        ],
        "physics": "두 발이 땅에서 떨어져 공중에 떠 있으나, 맹렬히 달려나가는 폼이 아니라 양 다리가 무릎에서 뒤로 대칭되게 꺾인 어색한 제자리 점프 상태임. 신체를 자연스럽게 지탱하거나 앞으로 뻗어나가는 역학이 없음."
       }
      ],
      "all_candidates_fail": true
     },
     "normalized": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 3,
        "verdict_ko": "환경과 인물 외형은 잘 따랐으나, '역동적인 뒷모습'과 '도약'이라는 묘사와 달리 추진력 없이 허공에 꼿꼿이 떠 있는 자세로 렌더링되어 물리적 연출에서 실패함."
       },
       {
        "label": "B",
        "score": 1,
        "verdict_ko": "샷 텍스트에 명시되지 않은 인물(시체)을 배경에 임의로 추가하는 치명적인 오류를 범했으며, 달리기 역학에 맞지 않는 기이한 점프 자세를 보임."
       }
      ],
      "all_candidates_fail": true,
      "readings": [
       {
        "label": "A",
        "direction": "안규홍의 방향과 시선이 시장 골목길을 지나 정면 멀리 있는 산을 향해 있음.",
        "built_space": "흙길 양옆으로 초가집과 우측 흙담이 있으며, 바닥에 부러진 나무 파편과 낫이 흩어져 있음. 이전 샷의 환경 구조와 잘 일치함.",
        "entities": "안규홍(단단히 묶은 검은 머리, 등에 난 핏자국, 붉은 피로 덮인 두 손, 피 묻은 낫 두 자루, 맨 목). 그 외 다른 인물은 등장하지 않음.",
        "hard_violations": [
         "물리적으로 불가능한 연출 (도약의 역학이나 추진력 없이 허공에 꼿꼿하게 서서 부양하고 있음)"
        ],
        "physics": "두 발이 땅에서 떨어져 공중에 떠 있으나, 달리는 도약(launch)을 위한 다리의 구부러짐이나 근육의 움직임이 전혀 없이 똑바로 서 있음. 몸을 지탱하거나 밀어내는 힘이 전혀 보이지 않는 부양 상태임."
       },
       {
        "label": "B",
        "direction": "안규홍이 우측 상단 배경에 솟아 있는 산을 향해 골목길을 마주하고 있음.",
        "built_space": "초가집과 흙담이 늘어선 흙길이며 바닥에 낫과 무기들이 떨어져 있음.",
        "entities": "안규홍(단단히 묶은 머리, 등 상처, 피 묻은 두 손, 낫 두 자루). 샷 텍스트에 없는 누워있는 시체 두 구가 배경 바닥에 추가로 존재함.",
        "hard_violations": [
         "없는 인물 추가 (프롬프트 텍스트에 없는 두 명의 시체가 배경에 등장함)",
         "물리적으로 불가능한 연출 (달리기 추진력이 아닌 기이한 대칭형 점프로 허공에 떠 있음)"
        ],
        "physics": "두 발이 땅에서 떨어져 공중에 떠 있으나, 맹렬히 달려나가는 폼이 아니라 양 다리가 무릎에서 뒤로 대칭되게 꺾인 어색한 제자리 점프 상태임. 신체를 자연스럽게 지탱하거나 앞으로 뻗어나가는 역학이 없음."
       }
      ]
     },
     "ok": true
    },
    {
     "model": "gpt-high",
     "order": "reverse",
     "display_to_canonical": {
      "A": "B",
      "B": "A"
     },
     "raw": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 3,
        "verdict_ko": "양손의 낫과 도약은 보이지만, 허용되지 않은 쓰러진 인물 두 명을 추가했으며 산도 참조보다 지나치게 가깝고 크게 보인다."
       },
       {
        "label": "B",
        "score": 8,
        "verdict_ko": "추가 인물 없이 기존 장터와 우측 상단의 먼 산을 유지하고 막 지면을 떠난 뒷모습을 구현했으나, 맹렬한 추진 동작은 다소 약하다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "안규홍의 얼굴은 보이지 않고 머리와 몸통은 골목 안쪽을 향한다. 진행축은 화면 중앙에 가깝고 산봉우리는 그보다 오른쪽 위에 있어 목적지는 대체로 읽힌다. 양손의 낫날은 아래로 늘어져 있으며 다른 사람을 겨누지는 않는다.",
        "built_space": "전신과 양발 아래 흙바닥이 포함된 넓은 구도다. 오른쪽에는 돌기단 위 흙담 한 줄이 있고, 왼쪽에는 목재 기둥으로 받친 초가 점포가 적어도 네 채 이어진다. 바닥의 버려진 무기와 부서진 목재는 참조의 재료와 분위기에 가깝다. 다만 골목이 앞쪽에서 넓어지고 참조의 먼 산줄기가 훨씬 크고 가까운 단일 봉우리처럼 바뀌었다. 중경 좌우에는 쓰러진 인물이 한 명씩 추가되어 있다.",
        "entities": "주인공은 검은 머리를 단단히 묶은 젊은 남성의 뒷모습이며 체격은 인물 참조와 대체로 맞는다. 얼굴이 가려져 정확한 얼굴 동일성이나 한국인 정체성은 직접 확인할 수 없다. 밝은색 저고리와 바지, 짚신은 맞지만 참조의 굵은 허리끈은 보이지 않는다. 목에는 장신구가 없고 등에 큰 혈흔, 양손에 피, 손마다 피 묻은 낫 한 자루가 보인다. 추가된 쓰러진 두 사람은 이 샷에서 허용된 인물이 아니다. 읽을 수 있는 글자는 없다.",
        "hard_violations": [
         "안규홍만 허용된 샷에 쓰러진 인물 두 명을 새로 추가했다."
        ],
        "physics": "양발은 모두 지면에서 떨어져 있고 오른쪽 무릎이 크게 굽혀져 있다. 뒤쪽 흙바닥의 발자국과 솟은 먼지가 발을 박찬 출발점을 제시하고, 진행 방향의 빈 골목이 착지 공간을 제공하므로 근거 없는 공중부양으로 보지는 않는다. 다만 몸통이 곧고 양다리가 뒤로 접혀 있어 맹렬한 달리기보다는 위로 뛰는 동작에 가깝다. 두 낫은 각각 손으로 손잡이를 쥐고 있으며, 쓰러진 두 몸과 버려진 무기는 지면에 놓여 있다."
       },
       {
        "label": "B",
        "direction": "머리와 등은 카메라 반대쪽 골목 출구를 향한다. 골목은 화면 오른쪽 위의 먼 산줄기 쪽으로 이어져 이동 목적지가 읽힌다. 눈은 보이지 않아 실제 시선은 확인할 수 없다. 두 낫은 손 아래로 향하고 있어 공격 대상을 겨누는 자세가 아니라 들고 달아나는 상태다.",
        "built_space": "전신과 양발 아래 흙바닥을 보여 주는 넓은 구도다. 오른쪽의 긴 흙담 한 줄과 돌기단, 왼쪽의 목조 초가 점포열, 길 끝의 초가 건물들이 참조의 골목 구조를 유지한다. 왼쪽에는 적어도 네 채의 점포가 구분되며 전경 점포의 목재 지지기둥과 무너진 판재가 보인다. 버려진 무기와 잔해가 흙길에 남아 있고 산줄기는 우측 상단의 먼 배경 크기로 유지된다. 추가 인물이나 반사상은 없다.",
        "entities": "젊은 남성 한 명만 등장하며 묶은 검은 머리, 체격, 밝은색 낡은 저고리와 바지, 짚신은 안규홍 참조와 대체로 맞는다. 뒷모습이므로 얼굴 동일성과 한국인 정체성은 직접 확인할 수 없다. 참조의 굵은 허리끈은 드러나지 않는다. 목에는 장신구가 없고 등에 출혈 흔적이 있으며 피 묻은 양손이 각각 낫 한 자루를 들고 있다. 바닥에는 혈흔, 버려진 무기와 부서진 점포 목재가 보인다. 읽을 수 있는 글자는 없다.",
        "hard_violations": [],
        "physics": "두 발이 흙바닥 바로 위에 떠 있고 한쪽 다리는 뒤로 더 뻗어 있어 달리기의 짧은 체공 구간으로 읽힌다. 발뒤의 낮은 먼지구름이 직전 지면 차기를 뒷받침하며 앞쪽 골목에는 착지할 공간이 있다. 따라서 지지나 발사 근거 없는 부유 상태는 아니다. 다만 상체의 전방 기울기와 무릎 추진이 작아 폭발적인 출발의 강도는 부족하다. 낫 두 자루는 각각 손의 손잡이 파지로 지지되고, 바닥의 무기와 잔해는 지면에 닿아 있다."
       }
      ],
      "all_candidates_fail": false
     },
     "normalized": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 3,
        "verdict_ko": "양손의 낫과 도약은 보이지만, 허용되지 않은 쓰러진 인물 두 명을 추가했으며 산도 참조보다 지나치게 가깝고 크게 보인다."
       },
       {
        "label": "A",
        "score": 8,
        "verdict_ko": "추가 인물 없이 기존 장터와 우측 상단의 먼 산을 유지하고 막 지면을 떠난 뒷모습을 구현했으나, 맹렬한 추진 동작은 다소 약하다."
       }
      ],
      "all_candidates_fail": false,
      "readings": [
       {
        "label": "B",
        "direction": "안규홍의 얼굴은 보이지 않고 머리와 몸통은 골목 안쪽을 향한다. 진행축은 화면 중앙에 가깝고 산봉우리는 그보다 오른쪽 위에 있어 목적지는 대체로 읽힌다. 양손의 낫날은 아래로 늘어져 있으며 다른 사람을 겨누지는 않는다.",
        "built_space": "전신과 양발 아래 흙바닥이 포함된 넓은 구도다. 오른쪽에는 돌기단 위 흙담 한 줄이 있고, 왼쪽에는 목재 기둥으로 받친 초가 점포가 적어도 네 채 이어진다. 바닥의 버려진 무기와 부서진 목재는 참조의 재료와 분위기에 가깝다. 다만 골목이 앞쪽에서 넓어지고 참조의 먼 산줄기가 훨씬 크고 가까운 단일 봉우리처럼 바뀌었다. 중경 좌우에는 쓰러진 인물이 한 명씩 추가되어 있다.",
        "entities": "주인공은 검은 머리를 단단히 묶은 젊은 남성의 뒷모습이며 체격은 인물 참조와 대체로 맞는다. 얼굴이 가려져 정확한 얼굴 동일성이나 한국인 정체성은 직접 확인할 수 없다. 밝은색 저고리와 바지, 짚신은 맞지만 참조의 굵은 허리끈은 보이지 않는다. 목에는 장신구가 없고 등에 큰 혈흔, 양손에 피, 손마다 피 묻은 낫 한 자루가 보인다. 추가된 쓰러진 두 사람은 이 샷에서 허용된 인물이 아니다. 읽을 수 있는 글자는 없다.",
        "hard_violations": [
         "안규홍만 허용된 샷에 쓰러진 인물 두 명을 새로 추가했다."
        ],
        "physics": "양발은 모두 지면에서 떨어져 있고 오른쪽 무릎이 크게 굽혀져 있다. 뒤쪽 흙바닥의 발자국과 솟은 먼지가 발을 박찬 출발점을 제시하고, 진행 방향의 빈 골목이 착지 공간을 제공하므로 근거 없는 공중부양으로 보지는 않는다. 다만 몸통이 곧고 양다리가 뒤로 접혀 있어 맹렬한 달리기보다는 위로 뛰는 동작에 가깝다. 두 낫은 각각 손으로 손잡이를 쥐고 있으며, 쓰러진 두 몸과 버려진 무기는 지면에 놓여 있다."
       },
       {
        "label": "A",
        "direction": "머리와 등은 카메라 반대쪽 골목 출구를 향한다. 골목은 화면 오른쪽 위의 먼 산줄기 쪽으로 이어져 이동 목적지가 읽힌다. 눈은 보이지 않아 실제 시선은 확인할 수 없다. 두 낫은 손 아래로 향하고 있어 공격 대상을 겨누는 자세가 아니라 들고 달아나는 상태다.",
        "built_space": "전신과 양발 아래 흙바닥을 보여 주는 넓은 구도다. 오른쪽의 긴 흙담 한 줄과 돌기단, 왼쪽의 목조 초가 점포열, 길 끝의 초가 건물들이 참조의 골목 구조를 유지한다. 왼쪽에는 적어도 네 채의 점포가 구분되며 전경 점포의 목재 지지기둥과 무너진 판재가 보인다. 버려진 무기와 잔해가 흙길에 남아 있고 산줄기는 우측 상단의 먼 배경 크기로 유지된다. 추가 인물이나 반사상은 없다.",
        "entities": "젊은 남성 한 명만 등장하며 묶은 검은 머리, 체격, 밝은색 낡은 저고리와 바지, 짚신은 안규홍 참조와 대체로 맞는다. 뒷모습이므로 얼굴 동일성과 한국인 정체성은 직접 확인할 수 없다. 참조의 굵은 허리끈은 드러나지 않는다. 목에는 장신구가 없고 등에 출혈 흔적이 있으며 피 묻은 양손이 각각 낫 한 자루를 들고 있다. 바닥에는 혈흔, 버려진 무기와 부서진 점포 목재가 보인다. 읽을 수 있는 글자는 없다.",
        "hard_violations": [],
        "physics": "두 발이 흙바닥 바로 위에 떠 있고 한쪽 다리는 뒤로 더 뻗어 있어 달리기의 짧은 체공 구간으로 읽힌다. 발뒤의 낮은 먼지구름이 직전 지면 차기를 뒷받침하며 앞쪽 골목에는 착지할 공간이 있다. 따라서 지지나 발사 근거 없는 부유 상태는 아니다. 다만 상체의 전방 기울기와 무릎 추진이 작아 폭발적인 출발의 강도는 부족하다. 낫 두 자루는 각각 손의 손잡이 파지로 지지되고, 바닥의 무기와 잔해는 지면에 닿아 있다."
       }
      ]
     },
     "ok": true
    }
   ],
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "slot_winner_match": true,
   "slot_winner": {
    "gemini-pro": "A",
    "gpt-high": "A"
   },
   "route": "cross_slot_agree"
  },
  "dual": {
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "normalized": {
    "A": 2.0,
    "B": 0.708
   },
   "adjusted": {
    "A": 1.75,
    "B": 0.458
   },
   "violations": {
    "A": [
     "[gemini-pro] 물리적으로 불가능한 연출 (도약의 역학이나 추진력 없이 허공에 꼿꼿하게 서서 부양하고 있음)"
    ],
    "B": [
     "[gemini-pro] 없는 인물 추가 (프롬프트 텍스트에 없는 두 명의 시체가 배경에 등장함)",
     "[gemini-pro] 물리적으로 불가능한 연출 (달리기 추진력이 아닌 기이한 대칭형 점프로 허공에 떠 있음)",
     "[gpt-high] 안규홍만 허용된 샷에 쓰러진 인물 두 명을 새로 추가했다."
    ]
   },
   "per_model_winner": {
    "gemini-pro": "A",
    "gpt-high": "A"
   },
   "agreed": true
  },
  "totals": {
   "A": 1750,
   "B": 458
  },
  "selected": "A",
  "ranking": [
   "A",
   "B"
  ],
  "verdicts": [
   {
    "label": "A",
    "score": 1750,
    "verdict_ko": "환경과 인물 외형은 잘 따랐으나, '역동적인 뒷모습'과 '도약'이라는 묘사와 달리 추진력 없이 허공에 꼿꼿이 떠 있는 자세로 렌더링되어 물리적 연출에서 실패함.  ★위반: [gemini-pro] 물리적으로 불가능한 연출 (도약의 역학이나 추진력 없이 허공에 꼿꼿하게 서서 부양하고 있음)"
   },
   {
    "label": "B",
    "score": 458,
    "verdict_ko": "샷 텍스트에 명시되지 않은 인물(시체)을 배경에 임의로 추가하는 치명적인 오류를 범했으며, 달리기 역학에 맞지 않는 기이한 점프 자세를 보임.  ★위반: [gemini-pro] 없는 인물 추가 (프롬프트 텍스트에 없는 두 명의 시체가 배경에 등장함) / [gemini-pro] 물리적으로 불가능한 연출 (달리기 추진력이 아닌 기이한 대칭형 점프로 허공에 떠 있음) / [gpt-high] 안규홍만 허용된 샷에 쓰러진 인물 두 명을 새로 추가했다."
   }
  ],
  "refs": [
   {
    "label": "PREVIOUS SHOT STILL — a visually related earlier shot of this same place: the location's look, materials, fixed features and lighting mood are LOCKED to this photo; never copy its camera framing. The people visible in that photo are NOT in this shot: never carry their faces, bodies or clothing onto anyone here. Each person in THIS shot is defined solely by the PEOPLE list, their own CHARACTER REFERENCE images and their wardrobe notes. If this photo shows a character who cannot move (dead or unconscious), that character's exact pose, position and orientation are ALSO LOCKED — treat their body as a fixed prop of the set that only the camera moves around.",
    "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/db00517b-165b-4005-a287-f28936b59b95/images/272bd7af-9294-4490-98c6-88a40e9deacc/scene/recipe/S57sh7_sel.png",
    "asset_id": "44f4615c-846d-4890-a055-4101081e1efc",
    "role": "prev_still"
   },
   {
    "label": "CHARACTER REFERENCE — 안규홍: the exact person appearing in this shot; match face, hair and build exactly.",
    "path": "<bytes:1028384>",
    "asset_id": "3b5894db-ce55-4f34-b966-5e725c3aa952",
    "role": "character_ref"
   }
  ],
  "critique_skipped": true,
  "shot_run_produced": false,
  "shot_run_uid": "06aa1438-eaa3-7810-85e3-20f0107f9f6b",
  "ref_mode": "prev+엔티티",
  "share_plan": {
   "ref_plan": "prev",
   "prev_anchor_tag": "S57sh7"
  }
 },
 "S57sh19::cine": {
  "applied": true,
  "attempted_at": "2026-09-09T07:33:59.831508+00:00",
  "fingerprint": "b380e1da662309d2cd8a49ae0aba1fc17d9c3551a5a8f9c35c2a0e353d9ab31f",
  "fingerprint_version": 2,
  "provider": "mai",
  "endpoint": "openrouter/chat-completions",
  "model": "microsoft/mai-image-2.6",
  "pack": "24.202608252115",
  "source_file": "S57sh19_sel.png",
  "source_sha256": "90c623c15a9858bb54ce922ec1bcd2a4c32a983b337b3e4af0bb6d68876e1233",
  "file": "S57sh19_cine.png",
  "staged_sha256": "a33959372444223049469b62529994e2663f15c7bd09a42b260ef0fb8fa8f0c0",
  "latency_ms": 27720
 },
 "S58sh3::signage": {
  "fp": "45b678462cd07c12",
  "inscriptions": [],
  "cues": [],
  "dropped": []
 },
 "S58sh3": {
  "input_fingerprint": "59e755146f3394eb",
  "prompt": "Create ONE FINAL photorealistic live-action film still of the moment below — early 20th-century Korea, Jeolla region; all people are Korean unless explicitly stated otherwise. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 두 눈을 부릅뜬 채 흙바닥에 널브러져 있는 임정현의 굳은 시신.\n\nLOCATION (lock): The dirt ground at the exterior base of a hut in the devastated mountain camp, amid smoke and scattered meal debris. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: close-up\n- KEY BACKGROUND ELEMENTS: 산채 흙바닥 (Supporting 임정현's motionless body); used as Visible around his head and shoulder, locating the body physically rather than isolating it against an abstract background.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Daylight softened by the screenplay's gunpowder smoke retains readable eyes and subdued contrast without introducing a dreamlike treatment.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The people visible in that photo are NOT in this shot: never carry their faces, bodies or clothing onto anyone here. Each person in THIS shot is defined solely by the PEOPLE list, their own CHARACTER REFERENCE images and their wardrobe notes. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): Powder smoke hangs over the overturned pot, shattered bowls and rice mixed with dirt and blood; huts and timber have been damaged.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 임정현 (한국인 남성, 성인 나이대, 자연스러운 검은 머리) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.",
  "roll_prompts": {
   "A": "Create ONE FINAL photorealistic live-action film still of the moment below — early 20th-century Korea, Jeolla region; all people are Korean unless explicitly stated otherwise. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 두 눈을 부릅뜬 채 흙바닥에 널브러져 있는 임정현의 굳은 시신.\n\nLOCATION (lock): The dirt ground at the exterior base of a hut in the devastated mountain camp, amid smoke and scattered meal debris. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: close-up\n- KEY BACKGROUND ELEMENTS: 산채 흙바닥 (Supporting 임정현's motionless body); used as Visible around his head and shoulder, locating the body physically rather than isolating it against an abstract background.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Daylight softened by the screenplay's gunpowder smoke retains readable eyes and subdued contrast without introducing a dreamlike treatment.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The people visible in that photo are NOT in this shot: never carry their faces, bodies or clothing onto anyone here. Each person in THIS shot is defined solely by the PEOPLE list, their own CHARACTER REFERENCE images and their wardrobe notes. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): Powder smoke hangs over the overturned pot, shattered bowls and rice mixed with dirt and blood; huts and timber have been damaged.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 임정현 (한국인 남성, 성인 나이대, 자연스러운 검은 머리) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.",
   "B": "Create ONE FINAL photorealistic live-action film still of the moment below — early 20th-century Korea, Jeolla region; all people are Korean unless explicitly stated otherwise. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 두 눈을 부릅뜬 채 흙바닥에 널브러져 있는 임정현의 굳은 시신.\n\nLOCATION (lock): The dirt ground at the exterior base of a hut in the devastated mountain camp, amid smoke and scattered meal debris. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: close-up\n- KEY BACKGROUND ELEMENTS: 산채 흙바닥 (Supporting 임정현's motionless body); used as Visible around his head and shoulder, locating the body physically rather than isolating it against an abstract background.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Daylight softened by the screenplay's gunpowder smoke retains readable eyes and subdued contrast without introducing a dreamlike treatment.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The people visible in that photo are NOT in this shot: never carry their faces, bodies or clothing onto anyone here. Each person in THIS shot is defined solely by the PEOPLE list, their own CHARACTER REFERENCE images and their wardrobe notes. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): Powder smoke hangs over the overturned pot, shattered bowls and rice mixed with dirt and blood; huts and timber have been damaged.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 임정현 (한국인 남성, 성인 나이대, 자연스러운 검은 머리) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.\n\nCOMPOSITION VARIATION (this candidate only): keep the place, the people and the moment exactly as specified, and keep from the CAMERA & FRAME contract its subject scale, screen placement and key background elements exactly. For this candidate only, its camera angle and height are deliberately left open: choose a camera position distinctly different from the obvious one, as a film director picking a second setup on the same blocking. Stay true to any explicit framing the SHOT TEXT itself dictates."
  },
  "shot_run_spend_attempt_count": 1,
  "readings": [
   {
    "label": "A",
    "direction": "남성의 두 눈은 크게 떠진 채 초점 없이 허공을 향하고 있어 굳은 시신이라는 텍스트 묘사에 부합합니다.",
    "built_space": "배경에 참조 이미지와 동일한 엮은 나뭇가지 벽면과 나무 기둥을 가진 오두막 구조물이 정확한 위치에 있으며, 남성은 그 앞 흙바닥에 위치해 있습니다.",
    "entities": "한국인 남성이 흙바닥에 누워 있으며, 지시와 달리 참조 이미지의 인물과 동일한 얼굴과 남색 군복 형태의 옷을 입고 있습니다. 바닥에는 깨진 그릇 조각, 흰 쌀알, 그리고 피가 섞인 흙이 명확히 존재합니다.",
    "hard_violations": [],
    "physics": "남성의 머리와 등은 흙바닥에 자연스럽게 지탱되어 있으며, 흩어진 쌀알과 깨진 그릇 조각들도 바닥에 안정적으로 놓여 물리적 어색함이 없습니다."
   },
   {
    "label": "B",
    "direction": "남성의 두 눈은 크게 떠져 있으며 초점 없이 위쪽을 응시하여 시신의 느낌을 줍니다.",
    "built_space": "배경에 흙벽으로 된 오두막과 출입구가 보이나, 이는 참조 이미지에 나타난 엮은 나뭇가지 형태의 벽면 재질과 일치하지 않습니다.",
    "entities": "한국인 남성이 밝은 색의 한복 스타일 옷을 입고 누워 있어 참조 이미지와 다른 외형을 보여줍니다. 바닥에 깨진 그릇과 쌀알이 있으나 피가 섞인 흙의 묘사는 보이지 않습니다.",
    "hard_violations": [],
    "physics": "남성의 몸은 바닥에 잘 지탱되어 있으나, 옷의 가파른 수직 주름 위에 얹혀 있는 일부 쌀알들은 중력을 거스른 듯 접착된 것처럼 보여 다소 부자연스럽습니다."
   }
  ],
  "cross_model_order": {
   "policy": "cross_model_order_v2_slot0_fwd_slot1_rev_parallel2",
   "slots": [
    {
     "model": "gemini-pro",
     "order": "forward",
     "display_to_canonical": {
      "A": "A",
      "B": "B"
     },
     "raw": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 7,
        "verdict_ko": "참조 이미지의 오두막 재질(엮은 나뭇가지)과 바닥의 파편(쌀, 깨진 그릇, 피)을 우선순위 3에 맞게 훌륭히 구현했으나, 인물의 외형과 복장을 참조에서 그대로 가져오지 말라는 지시(우선순위 4)를 위반한 점이 감점 요인입니다."
       },
       {
        "label": "B",
        "score": 5,
        "verdict_ko": "새로운 복장을 적용해 인물 관련 부정 지시어는 잘 따랐으나, 우선순위가 더 높은 장소 고정 조건(오두막의 재질)을 어기고 임의의 흙벽을 생성했으며 바닥에 섞인 피의 묘사가 부족합니다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "남성의 두 눈은 크게 떠진 채 초점 없이 허공을 향하고 있어 굳은 시신이라는 텍스트 묘사에 부합합니다.",
        "built_space": "배경에 참조 이미지와 동일한 엮은 나뭇가지 벽면과 나무 기둥을 가진 오두막 구조물이 정확한 위치에 있으며, 남성은 그 앞 흙바닥에 위치해 있습니다.",
        "entities": "한국인 남성이 흙바닥에 누워 있으며, 지시와 달리 참조 이미지의 인물과 동일한 얼굴과 남색 군복 형태의 옷을 입고 있습니다. 바닥에는 깨진 그릇 조각, 흰 쌀알, 그리고 피가 섞인 흙이 명확히 존재합니다.",
        "hard_violations": [],
        "physics": "남성의 머리와 등은 흙바닥에 자연스럽게 지탱되어 있으며, 흩어진 쌀알과 깨진 그릇 조각들도 바닥에 안정적으로 놓여 물리적 어색함이 없습니다."
       },
       {
        "label": "B",
        "direction": "남성의 두 눈은 크게 떠져 있으며 초점 없이 위쪽을 응시하여 시신의 느낌을 줍니다.",
        "built_space": "배경에 흙벽으로 된 오두막과 출입구가 보이나, 이는 참조 이미지에 나타난 엮은 나뭇가지 형태의 벽면 재질과 일치하지 않습니다.",
        "entities": "한국인 남성이 밝은 색의 한복 스타일 옷을 입고 누워 있어 참조 이미지와 다른 외형을 보여줍니다. 바닥에 깨진 그릇과 쌀알이 있으나 피가 섞인 흙의 묘사는 보이지 않습니다.",
        "hard_violations": [],
        "physics": "남성의 몸은 바닥에 잘 지탱되어 있으나, 옷의 가파른 수직 주름 위에 얹혀 있는 일부 쌀알들은 중력을 거스른 듯 접착된 것처럼 보여 다소 부자연스럽습니다."
       }
      ],
      "all_candidates_fail": false
     },
     "normalized": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 7,
        "verdict_ko": "참조 이미지의 오두막 재질(엮은 나뭇가지)과 바닥의 파편(쌀, 깨진 그릇, 피)을 우선순위 3에 맞게 훌륭히 구현했으나, 인물의 외형과 복장을 참조에서 그대로 가져오지 말라는 지시(우선순위 4)를 위반한 점이 감점 요인입니다."
       },
       {
        "label": "B",
        "score": 5,
        "verdict_ko": "새로운 복장을 적용해 인물 관련 부정 지시어는 잘 따랐으나, 우선순위가 더 높은 장소 고정 조건(오두막의 재질)을 어기고 임의의 흙벽을 생성했으며 바닥에 섞인 피의 묘사가 부족합니다."
       }
      ],
      "all_candidates_fail": false,
      "readings": [
       {
        "label": "A",
        "direction": "남성의 두 눈은 크게 떠진 채 초점 없이 허공을 향하고 있어 굳은 시신이라는 텍스트 묘사에 부합합니다.",
        "built_space": "배경에 참조 이미지와 동일한 엮은 나뭇가지 벽면과 나무 기둥을 가진 오두막 구조물이 정확한 위치에 있으며, 남성은 그 앞 흙바닥에 위치해 있습니다.",
        "entities": "한국인 남성이 흙바닥에 누워 있으며, 지시와 달리 참조 이미지의 인물과 동일한 얼굴과 남색 군복 형태의 옷을 입고 있습니다. 바닥에는 깨진 그릇 조각, 흰 쌀알, 그리고 피가 섞인 흙이 명확히 존재합니다.",
        "hard_violations": [],
        "physics": "남성의 머리와 등은 흙바닥에 자연스럽게 지탱되어 있으며, 흩어진 쌀알과 깨진 그릇 조각들도 바닥에 안정적으로 놓여 물리적 어색함이 없습니다."
       },
       {
        "label": "B",
        "direction": "남성의 두 눈은 크게 떠져 있으며 초점 없이 위쪽을 응시하여 시신의 느낌을 줍니다.",
        "built_space": "배경에 흙벽으로 된 오두막과 출입구가 보이나, 이는 참조 이미지에 나타난 엮은 나뭇가지 형태의 벽면 재질과 일치하지 않습니다.",
        "entities": "한국인 남성이 밝은 색의 한복 스타일 옷을 입고 누워 있어 참조 이미지와 다른 외형을 보여줍니다. 바닥에 깨진 그릇과 쌀알이 있으나 피가 섞인 흙의 묘사는 보이지 않습니다.",
        "hard_violations": [],
        "physics": "남성의 몸은 바닥에 잘 지탱되어 있으나, 옷의 가파른 수직 주름 위에 얹혀 있는 일부 쌀알들은 중력을 거스른 듯 접착된 것처럼 보여 다소 부자연스럽습니다."
       }
      ]
     },
     "ok": true
    },
    {
     "model": "gpt-high",
     "order": "reverse",
     "display_to_canonical": {
      "A": "B",
      "B": "A"
     },
     "raw": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 9,
        "verdict_ko": "흙바닥에 밀착된 머리와 어깨의 클로즈업, 부릅뜬 눈과 굳은 표정이 정확하며 이전 숏 인물의 의상도 옮겨오지 않았다."
       },
       {
        "label": "B",
        "score": 7,
        "verdict_ko": "시신의 접지와 클로즈업은 맞지만, 이전 숏 인물의 얼굴과 특히 검은 누더기 의상을 재사용한 모습이 명시적 인물 분리 지시와 충돌한다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "얼굴이 카메라 쪽으로 돌아가 있고 두 눈은 크게 열린 채 렌즈 부근의 허공을 응시한다. 특정 인물을 바라보거나 움직이는 모습은 아니어서 시신의 고정된 시선으로 읽힌다. 무기나 방향을 확인할 손 소품은 없다.",
        "built_space": "머리와 어깨를 크게 잡은 지면 높이의 클로즈업이다. 배경에는 흙벽과 목재로 된 움막의 어두운 출입구 하나, 왼쪽의 목책과 무너진 목재가 보인다. 머리 둘레와 어깨 아래에 실제 흙바닥이 이어진다. 참고 장소의 거친 재료와 산채 분위기는 유지되지만 동일한 움막인지는 이 구도만으로 확정하기 어렵다. 중복된 고정 설비나 반사는 없다.",
        "entities": "자연스러운 검은 머리의 성인 동아시아계 남성 한 명으로, 한국인 남성 설정에 어긋나는 외형은 없다. 별도의 임정현 얼굴 참고가 없어 정확한 얼굴 동일성은 확인할 수 없다. 흙 묻은 회베이지색 전통 의복은 이전 숏의 검은 단추·덧댐 의상과 구별된다. 깨진 그릇 여러 조각, 그릇의 남은 몸체 하나, 흩어진 쌀과 작은 혈흔, 연기, 손상된 목재가 보인다. 뒤집힌 솥은 이 클로즈업 안에서 확인되지 않는다. 읽을 수 있는 글자나 추가 인물은 없다.",
        "hard_violations": [],
        "physics": "뒤통수와 머리카락이 흙에 직접 닿고 목은 누운 몸통으로 자연스럽게 이어진다. 오른쪽 아래의 어깨와 몸통도 지면에 놓여 있어 머리나 몸이 떠 있지 않다. 쌀과 그릇 조각은 흙 위에, 옷에 붙은 쌀은 천의 주름 위에 놓여 있다. 움직임이나 비현실적인 관절 변형 없이 정지한 시신으로 성립한다."
       },
       {
        "label": "B",
        "direction": "얼굴은 카메라 쪽으로 기울고 열린 두 눈은 렌즈보다 약간 위쪽의 허공을 향한다. 특정 대상과 시선을 주고받는 장면은 아니다. 시선은 고정되어 있지만 A보다 눈을 부릅뜬 정도가 약하다. 무기나 지향성 소품은 없다.",
        "built_space": "머리와 어깨 중심의 지면 높이 클로즈업이다. 왼쪽에는 어두운 출입구 하나와 목재 기둥, 엮은 벽체가 있고 오른쪽 뒤에는 돌무더기가 있다. 중앙 후경에는 연기와 목책이 보인다. 몸은 움막 바깥 흙바닥에 놓였으며 머리 주변의 지면도 충분히 드러난다. 참고 장소의 목재·흙·돌 재료와 잘 이어지며, 설비 중복이나 불가능한 반사는 보이지 않는다.",
        "entities": "자연스러운 검은 머리의 성인 동아시아계 남성 한 명이다. 얼굴이 이전 숏 남성과 매우 유사하고, 검은 상의의 단추와 거친 직사각형 덧댐까지 이전 숏 의상을 강하게 반복한다. 이는 그 인물의 얼굴과 의상을 옮기지 말라는 지시에 맞지 않는다. 흙 위에 쌀, 큰 그릇 파편 두 개와 작은 파편들, 검붉은 혈흔이 보이며 배경에는 연기가 있다. 뒤집힌 솥은 화면에서 확인되지 않는다. 추가 인물과 읽을 수 있는 글자는 없다.",
        "hard_violations": [],
        "physics": "뒤통수와 귀 주변이 흙바닥에 닿고 어깨와 몸통도 지면에 누워 있다. 목의 연결과 옆으로 돌아간 머리 자세가 물리적으로 가능하다. 그릇 파편과 쌀은 땅이 받치며 혈액은 지면에 고여 있다. 떠 있는 신체나 지지 없는 물체는 없고, 정지한 시신의 자세로 성립한다."
       }
      ],
      "all_candidates_fail": false
     },
     "normalized": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 9,
        "verdict_ko": "흙바닥에 밀착된 머리와 어깨의 클로즈업, 부릅뜬 눈과 굳은 표정이 정확하며 이전 숏 인물의 의상도 옮겨오지 않았다."
       },
       {
        "label": "A",
        "score": 7,
        "verdict_ko": "시신의 접지와 클로즈업은 맞지만, 이전 숏 인물의 얼굴과 특히 검은 누더기 의상을 재사용한 모습이 명시적 인물 분리 지시와 충돌한다."
       }
      ],
      "all_candidates_fail": false,
      "readings": [
       {
        "label": "B",
        "direction": "얼굴이 카메라 쪽으로 돌아가 있고 두 눈은 크게 열린 채 렌즈 부근의 허공을 응시한다. 특정 인물을 바라보거나 움직이는 모습은 아니어서 시신의 고정된 시선으로 읽힌다. 무기나 방향을 확인할 손 소품은 없다.",
        "built_space": "머리와 어깨를 크게 잡은 지면 높이의 클로즈업이다. 배경에는 흙벽과 목재로 된 움막의 어두운 출입구 하나, 왼쪽의 목책과 무너진 목재가 보인다. 머리 둘레와 어깨 아래에 실제 흙바닥이 이어진다. 참고 장소의 거친 재료와 산채 분위기는 유지되지만 동일한 움막인지는 이 구도만으로 확정하기 어렵다. 중복된 고정 설비나 반사는 없다.",
        "entities": "자연스러운 검은 머리의 성인 동아시아계 남성 한 명으로, 한국인 남성 설정에 어긋나는 외형은 없다. 별도의 임정현 얼굴 참고가 없어 정확한 얼굴 동일성은 확인할 수 없다. 흙 묻은 회베이지색 전통 의복은 이전 숏의 검은 단추·덧댐 의상과 구별된다. 깨진 그릇 여러 조각, 그릇의 남은 몸체 하나, 흩어진 쌀과 작은 혈흔, 연기, 손상된 목재가 보인다. 뒤집힌 솥은 이 클로즈업 안에서 확인되지 않는다. 읽을 수 있는 글자나 추가 인물은 없다.",
        "hard_violations": [],
        "physics": "뒤통수와 머리카락이 흙에 직접 닿고 목은 누운 몸통으로 자연스럽게 이어진다. 오른쪽 아래의 어깨와 몸통도 지면에 놓여 있어 머리나 몸이 떠 있지 않다. 쌀과 그릇 조각은 흙 위에, 옷에 붙은 쌀은 천의 주름 위에 놓여 있다. 움직임이나 비현실적인 관절 변형 없이 정지한 시신으로 성립한다."
       },
       {
        "label": "A",
        "direction": "얼굴은 카메라 쪽으로 기울고 열린 두 눈은 렌즈보다 약간 위쪽의 허공을 향한다. 특정 대상과 시선을 주고받는 장면은 아니다. 시선은 고정되어 있지만 A보다 눈을 부릅뜬 정도가 약하다. 무기나 지향성 소품은 없다.",
        "built_space": "머리와 어깨 중심의 지면 높이 클로즈업이다. 왼쪽에는 어두운 출입구 하나와 목재 기둥, 엮은 벽체가 있고 오른쪽 뒤에는 돌무더기가 있다. 중앙 후경에는 연기와 목책이 보인다. 몸은 움막 바깥 흙바닥에 놓였으며 머리 주변의 지면도 충분히 드러난다. 참고 장소의 목재·흙·돌 재료와 잘 이어지며, 설비 중복이나 불가능한 반사는 보이지 않는다.",
        "entities": "자연스러운 검은 머리의 성인 동아시아계 남성 한 명이다. 얼굴이 이전 숏 남성과 매우 유사하고, 검은 상의의 단추와 거친 직사각형 덧댐까지 이전 숏 의상을 강하게 반복한다. 이는 그 인물의 얼굴과 의상을 옮기지 말라는 지시에 맞지 않는다. 흙 위에 쌀, 큰 그릇 파편 두 개와 작은 파편들, 검붉은 혈흔이 보이며 배경에는 연기가 있다. 뒤집힌 솥은 화면에서 확인되지 않는다. 추가 인물과 읽을 수 있는 글자는 없다.",
        "hard_violations": [],
        "physics": "뒤통수와 귀 주변이 흙바닥에 닿고 어깨와 몸통도 지면에 누워 있다. 목의 연결과 옆으로 돌아간 머리 자세가 물리적으로 가능하다. 그릇 파편과 쌀은 땅이 받치며 혈액은 지면에 고여 있다. 떠 있는 신체나 지지 없는 물체는 없고, 정지한 시신의 자세로 성립한다."
       }
      ]
     },
     "ok": true
    }
   ],
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "slot_winner_match": false,
   "slot_winner": {
    "gemini-pro": "A",
    "gpt-high": "B"
   },
   "route": "cross_slot_combined"
  },
  "dual": {
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "normalized": {
    "A": 1.778,
    "B": 1.714
   },
   "adjusted": {
    "A": 1.778,
    "B": 1.714
   },
   "violations": {},
   "per_model_winner": {
    "gemini-pro": "A",
    "gpt-high": "B"
   },
   "agreed": false
  },
  "totals": {
   "A": 1778,
   "B": 1714
  },
  "selected": "A",
  "ranking": [
   "A",
   "B"
  ],
  "verdicts": [
   {
    "label": "A",
    "score": 1778,
    "verdict_ko": "참조 이미지의 오두막 재질(엮은 나뭇가지)과 바닥의 파편(쌀, 깨진 그릇, 피)을 우선순위 3에 맞게 훌륭히 구현했으나, 인물의 외형과 복장을 참조에서 그대로 가져오지 말라는 지시(우선순위 4)를 위반한 점이 감점 요인입니다."
   },
   {
    "label": "B",
    "score": 1714,
    "verdict_ko": "새로운 복장을 적용해 인물 관련 부정 지시어는 잘 따랐으나, 우선순위가 더 높은 장소 고정 조건(오두막의 재질)을 어기고 임의의 흙벽을 생성했으며 바닥에 섞인 피의 묘사가 부족합니다."
   }
  ],
  "refs": [
   {
    "label": "PREVIOUS SHOT STILL — a visually related earlier shot of this same place: the location's look, materials, fixed features and lighting mood are LOCKED to this photo; never copy its camera framing. The people visible in that photo are NOT in this shot: never carry their faces, bodies or clothing onto anyone here. Each person in THIS shot is defined solely by the PEOPLE list, their own CHARACTER REFERENCE images and their wardrobe notes. If this photo shows a character who cannot move (dead or unconscious), that character's exact pose, position and orientation are ALSO LOCKED — treat their body as a fixed prop of the set that only the camera moves around.",
    "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/db00517b-165b-4005-a287-f28936b59b95/images/272bd7af-9294-4490-98c6-88a40e9deacc/scene/recipe/S56sh40_sel.png",
    "asset_id": "9d1cd195-667b-40af-ad5e-5755f6a59694",
    "role": "prev_still"
   }
  ],
  "critique_skipped": true,
  "shot_run_produced": false,
  "shot_run_uid": "06aa1438-ec4c-773a-8ff0-ea28c900afce",
  "ref_mode": "prev만 (배경 전용·공유 계획)",
  "share_plan": {
   "ref_plan": "prev",
   "prev_anchor_tag": "S56sh40"
  },
  "lane_policy": "share_plan_prev_bgonly"
 },
 "S58sh3::cine": {
  "applied": true,
  "attempted_at": "2026-09-09T07:35:49.911242+00:00",
  "fingerprint": "997017cbb9d4173b2a83158f828c9ceb4d9d5ecf36085aaf78fbd43db79013f9",
  "fingerprint_version": 2,
  "provider": "mai",
  "endpoint": "openrouter/chat-completions",
  "model": "microsoft/mai-image-2.6",
  "pack": "24.202608252115",
  "source_file": "S58sh3_sel.png",
  "source_sha256": "ff796d1797b6b9b76547432a75cff606c648aa901ad97e3eccc117e600a736ac",
  "file": "S58sh3_cine.png",
  "staged_sha256": "1fc7238bf193c5e41a667a8b2a36b384925832aecbdcc44ef834654aa1ec9560",
  "latency_ms": 26698
 },
 "S58sh9::signage": {
  "fp": "81998ccac89f53ff",
  "inscriptions": [],
  "cues": [],
  "dropped": []
 },
 "S58sh9": {
  "input_fingerprint": "03e27d4dad352bb0",
  "prompt": "Create ONE FINAL photorealistic live-action film still of the moment below — early 20th-century Korea, Jeolla region; all people are Korean unless explicitly stated otherwise. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 능선 위에서 참혹한 산채를 팔짱 낀 채 서늘하게 내려다보는 토미이시의 전신.\n\nLOCATION (lock): The ridge above the ruined mountain camp, overlooking the smoke-filled clearing after the assault. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: wide shot\n- KEY BACKGROUND ELEMENTS: 능선 (토미이시's observation position above the camp); used as A visible strip beneath his feet establishes stable footing and separates his full figure from the surrounding open space.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Natural daylight with muted saturation and controlled contrast maintains the ridge's sober realism without extending the camp's smoke into this position.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The clothing, hair and overall look of 토미이시 — who appear both in that photo and in this shot — are LOCKED to that photo. Anyone else visible in that photo is NOT in this shot: never carry their face, body or clothing onto anyone here. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): Powder smoke covers the ruined camp, with an overturned pot, broken bowls and bloodied rice on the ground. 토미이시: He stands on the ridge retaining his sword.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 토미이시 (일본인 남성, 성인 나이대, 자연스러운 검은 머리) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.",
  "roll_prompts": {
   "A": "Create ONE FINAL photorealistic live-action film still of the moment below — early 20th-century Korea, Jeolla region; all people are Korean unless explicitly stated otherwise. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 능선 위에서 참혹한 산채를 팔짱 낀 채 서늘하게 내려다보는 토미이시의 전신.\n\nLOCATION (lock): The ridge above the ruined mountain camp, overlooking the smoke-filled clearing after the assault. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: wide shot\n- KEY BACKGROUND ELEMENTS: 능선 (토미이시's observation position above the camp); used as A visible strip beneath his feet establishes stable footing and separates his full figure from the surrounding open space.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Natural daylight with muted saturation and controlled contrast maintains the ridge's sober realism without extending the camp's smoke into this position.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The clothing, hair and overall look of 토미이시 — who appear both in that photo and in this shot — are LOCKED to that photo. Anyone else visible in that photo is NOT in this shot: never carry their face, body or clothing onto anyone here. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): Powder smoke covers the ruined camp, with an overturned pot, broken bowls and bloodied rice on the ground. 토미이시: He stands on the ridge retaining his sword.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 토미이시 (일본인 남성, 성인 나이대, 자연스러운 검은 머리) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.",
   "B": "Create ONE FINAL photorealistic live-action film still of the moment below — early 20th-century Korea, Jeolla region; all people are Korean unless explicitly stated otherwise. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 능선 위에서 참혹한 산채를 팔짱 낀 채 서늘하게 내려다보는 토미이시의 전신.\n\nLOCATION (lock): The ridge above the ruined mountain camp, overlooking the smoke-filled clearing after the assault. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: wide shot\n- KEY BACKGROUND ELEMENTS: 능선 (토미이시's observation position above the camp); used as A visible strip beneath his feet establishes stable footing and separates his full figure from the surrounding open space.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Natural daylight with muted saturation and controlled contrast maintains the ridge's sober realism without extending the camp's smoke into this position.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The clothing, hair and overall look of 토미이시 — who appear both in that photo and in this shot — are LOCKED to that photo. Anyone else visible in that photo is NOT in this shot: never carry their face, body or clothing onto anyone here. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): Powder smoke covers the ruined camp, with an overturned pot, broken bowls and bloodied rice on the ground. 토미이시: He stands on the ridge retaining his sword.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 토미이시 (일본인 남성, 성인 나이대, 자연스러운 검은 머리) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.\n\nCOMPOSITION VARIATION (this candidate only): keep the place, the people and the moment exactly as specified, and keep from the CAMERA & FRAME contract its subject scale, screen placement and key background elements exactly. For this candidate only, its camera angle and height are deliberately left open: choose a camera position distinctly different from the obvious one, as a film director picking a second setup on the same blocking. Stay true to any explicit framing the SHOT TEXT itself dictates."
  },
  "shot_run_spend_attempt_count": 1,
  "readings": [
   {
    "label": "A",
    "direction": "시선이 화면 왼쪽 아래의 산채를 향해 내려다보고 있음.",
    "built_space": "오른쪽 능선에 인물이 서 있고 왼쪽 아래에 산채가 위치함. 그러나 산채 마당에 엎어진 가마솥과 그릇들이 뒤의 초가집과 맞먹거나 더 큰 비정상적인 크기로 렌더링되어 공간의 물리적 비율이 완전히 붕괴됨.",
    "entities": "토미이시의 얼굴, 복장, 체형이 레퍼런스와 일치함. 팔짱을 낀 자세를 취하고 있음.",
    "hard_violations": [
     "[gemini-pro] 물리적으로 불가능한 스테이징 및 스케일 (배경의 가마솥과 식기류가 건물 크기로 거대하게 렌더링됨)"
    ],
    "physics": "두 발로 능선 흙바닥을 딛고 서 있으며 체중이 자연스럽게 실려 있음."
   },
   {
    "label": "B",
    "direction": "시선이 아래쪽 산채 방향을 서늘하게 내려다보고 있음.",
    "built_space": "이전 샷의 위치와 완벽히 호환되는 능선과 아래쪽 산채의 구도. 배경의 구조물, 나무 울타리, 가마솥 등의 크기가 실제 물리적 스케일에 맞게 적절히 묘사됨.",
    "entities": "토미이시의 외모와 군복이 레퍼런스와 일치함. 지시된 대로 팔짱을 낀 채 칼을 차고 있음.",
    "hard_violations": [],
    "physics": "능선 위에 두 발로 단단히 서 있으며, 자연스러운 자세로 바닥의 지지를 받고 있음."
   }
  ],
  "cross_model_order": {
   "policy": "cross_model_order_v2_slot0_fwd_slot1_rev_parallel2",
   "slots": [
    {
     "model": "gemini-pro",
     "order": "forward",
     "display_to_canonical": {
      "A": "A",
      "B": "B"
     },
     "raw": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 9,
        "verdict_ko": "프롬프트의 요구사항(팔짱을 끼고 내려다보는 전신)을 정확히 구현하였으며, 이전 샷과 동일한 배경의 물리적 비율을 사실적으로 유지했습니다."
       },
       {
        "label": "A",
        "score": 2,
        "verdict_ko": "인물의 자세는 맞으나, 산채 배경에 있는 가마솥과 사발 등의 사물이 집채만 한 크기로 과장되어 물리적 스케일을 완전히 위반했습니다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "시선이 화면 왼쪽 아래의 산채를 향해 내려다보고 있음.",
        "built_space": "오른쪽 능선에 인물이 서 있고 왼쪽 아래에 산채가 위치함. 그러나 산채 마당에 엎어진 가마솥과 그릇들이 뒤의 초가집과 맞먹거나 더 큰 비정상적인 크기로 렌더링되어 공간의 물리적 비율이 완전히 붕괴됨.",
        "entities": "토미이시의 얼굴, 복장, 체형이 레퍼런스와 일치함. 팔짱을 낀 자세를 취하고 있음.",
        "hard_violations": [
         "물리적으로 불가능한 스테이징 및 스케일 (배경의 가마솥과 식기류가 건물 크기로 거대하게 렌더링됨)"
        ],
        "physics": "두 발로 능선 흙바닥을 딛고 서 있으며 체중이 자연스럽게 실려 있음."
       },
       {
        "label": "B",
        "direction": "시선이 아래쪽 산채 방향을 서늘하게 내려다보고 있음.",
        "built_space": "이전 샷의 위치와 완벽히 호환되는 능선과 아래쪽 산채의 구도. 배경의 구조물, 나무 울타리, 가마솥 등의 크기가 실제 물리적 스케일에 맞게 적절히 묘사됨.",
        "entities": "토미이시의 외모와 군복이 레퍼런스와 일치함. 지시된 대로 팔짱을 낀 채 칼을 차고 있음.",
        "hard_violations": [],
        "physics": "능선 위에 두 발로 단단히 서 있으며, 자연스러운 자세로 바닥의 지지를 받고 있음."
       }
      ],
      "all_candidates_fail": false
     },
     "normalized": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 9,
        "verdict_ko": "프롬프트의 요구사항(팔짱을 끼고 내려다보는 전신)을 정확히 구현하였으며, 이전 샷과 동일한 배경의 물리적 비율을 사실적으로 유지했습니다."
       },
       {
        "label": "A",
        "score": 2,
        "verdict_ko": "인물의 자세는 맞으나, 산채 배경에 있는 가마솥과 사발 등의 사물이 집채만 한 크기로 과장되어 물리적 스케일을 완전히 위반했습니다."
       }
      ],
      "all_candidates_fail": false,
      "readings": [
       {
        "label": "A",
        "direction": "시선이 화면 왼쪽 아래의 산채를 향해 내려다보고 있음.",
        "built_space": "오른쪽 능선에 인물이 서 있고 왼쪽 아래에 산채가 위치함. 그러나 산채 마당에 엎어진 가마솥과 그릇들이 뒤의 초가집과 맞먹거나 더 큰 비정상적인 크기로 렌더링되어 공간의 물리적 비율이 완전히 붕괴됨.",
        "entities": "토미이시의 얼굴, 복장, 체형이 레퍼런스와 일치함. 팔짱을 낀 자세를 취하고 있음.",
        "hard_violations": [
         "물리적으로 불가능한 스테이징 및 스케일 (배경의 가마솥과 식기류가 건물 크기로 거대하게 렌더링됨)"
        ],
        "physics": "두 발로 능선 흙바닥을 딛고 서 있으며 체중이 자연스럽게 실려 있음."
       },
       {
        "label": "B",
        "direction": "시선이 아래쪽 산채 방향을 서늘하게 내려다보고 있음.",
        "built_space": "이전 샷의 위치와 완벽히 호환되는 능선과 아래쪽 산채의 구도. 배경의 구조물, 나무 울타리, 가마솥 등의 크기가 실제 물리적 스케일에 맞게 적절히 묘사됨.",
        "entities": "토미이시의 외모와 군복이 레퍼런스와 일치함. 지시된 대로 팔짱을 낀 채 칼을 차고 있음.",
        "hard_violations": [],
        "physics": "능선 위에 두 발로 단단히 서 있으며, 자연스러운 자세로 바닥의 지지를 받고 있음."
       }
      ]
     },
     "ok": true
    },
    {
     "model": "gpt-high",
     "order": "reverse",
     "display_to_canonical": {
      "A": "B",
      "B": "A"
     },
     "raw": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 6,
        "verdict_ko": "전신 와이드숏과 안정적인 능선 발판, 참조 장소의 배치는 잘 지켰지만, 시선이 등 뒤 산채가 아닌 앞쪽 땅을 향한다."
       },
       {
        "label": "B",
        "score": 4,
        "verdict_ko": "팔짱 낀 전신과 군도는 충실하지만 산채를 내려다보는 시선이 불명확하고, 배경의 솥과 그릇이 건물 대비 지나치게 커 실제 거리감과 축척을 훼손한다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "토미이시는 몸을 거의 정면으로 세우고 눈을 카메라 쪽 아래로 내리깐다. 실제 시선의 대상은 앞쪽 능선 바닥으로 읽히며, 화면 왼쪽의 등 뒤 아래에 있는 산채를 향하지 않는다. 군도는 칼집에 들어가 허리에서 아래로 늘어져 있고 겨냥하는 동작은 없다.",
        "built_space": "전경에는 흙과 바위, 뿌리로 된 능선 발판이 있으며 인물은 그 위 중앙에 서 있다. 아래 산채에는 입구 하나가 난 목책과 뒤편 초가지붕 두 채, 인물 오른쪽 아래로 일부 드러난 지붕이 보인다. 왼쪽 외곽길, 바위군, 오른쪽 수목은 참조 장소의 배치와 대체로 이어진다. 인물의 머리부터 부츠까지 모두 포함하며, 발밑 지면이 선명하다. 반사는 없다.",
        "entities": "성인 동아시아계 남성 한 명만 보이며, 일본인이라는 국적 자체는 외관으로 확정할 수 없다. 검은 머리는 군모에 대부분 가려지고 콧수염과 턱수염, 얼굴과 체격은 참조와 대체로 일치한다. 갈색빛 군복, 붉은 깃장, 훈장, 가죽 교차 끈과 허리띠, 장화, 허리의 군도가 유지된다. 산채에는 연기와 작은 불길, 옆으로 넘어진 검은 솥 하나, 흩어진 그릇과 밝은 쌀 같은 물질이 보이나 피 묻은 쌀은 명확히 식별되지 않는다. 솥은 주변 건물에 비해 다소 크게 보인다. 낮의 저채도 자연광이며 능선의 인물은 연기에 덮이지 않는다. 읽을 수 있는 글자는 없다.",
        "hard_violations": [],
        "physics": "양쪽 부츠 밑창이 능선의 흙바닥에 닿아 체중을 받친다. 팔짱은 양팔을 몸 앞에서 교차해 서로 지지하는 가능한 자세다. 군도와 가죽 장비는 허리띠에 매달려 있고, 산채의 솥과 파편은 지면에 놓여 있다. 연기는 산채의 연소 지점에서 올라오며, 지지 없이 떠 있는 사람이나 물체는 보이지 않는다."
       },
       {
        "label": "B",
        "direction": "토미이시는 정면에서 약간 화면 왼쪽으로 몸과 얼굴을 틀고 아래를 본다. 그러나 눈길은 왼쪽 앞 능선 방향으로 읽히며, 그의 뒤쪽 아래에 펼쳐진 산채 내부에 정확히 닿는 시선은 아니다. 군도는 허리 옆에서 칼집째 아래로 향하며 사용하거나 겨냥하지 않는다.",
        "built_space": "토미이시는 화면 오른쪽의 흙과 뿌리가 드러난 능선에 서 있고 전신과 발밑 지면이 모두 보인다. 왼쪽 아래에는 입구 하나가 있는 목책, 뒤편의 초가 건물 두 채와 연기에 가려진 추가 지붕 부분이 보인다. 수목과 바위, 초가와 목책이라는 장소 요소는 유지되지만 산채가 화면에서 훨씬 크게 차지한다. 특히 솥 두 개와 여러 그릇이 건물 기둥 및 출입구에 비해 비정상적으로 커 보인다. 반사는 없다.",
        "entities": "성인 동아시아계 남성 한 명이며 참조와 유사한 콧수염, 턱수염과 체격을 갖는다. 국적은 외관만으로 판별할 수 없고 검은 머리는 군모에 대부분 가려져 있다. 군모, 갈색빛 군복, 붉은 깃장, 훈장, 가죽 끈과 허리띠, 장화, 군도를 유지한다. 산채에는 옆으로 누운 대형 솥 두 개, 흩어진 그릇, 흰 쌀 같은 덩어리와 붉은 얼룩이 보인다. 요청된 참상은 드러나지만 생활용품의 축척이 과장되어 있다. 낮의 저채도 조명이고 연기는 주로 산채와 배경을 채우며 인물 앞은 비교적 맑다. 추가 인물이나 읽을 수 있는 글자는 없다.",
        "hard_violations": [],
        "physics": "두 부츠가 능선 바닥에 닿아 서 있는 몸을 지지하며 팔짱도 해부학적으로 가능한 자세다. 군도는 허리 장비에 매달려 있다. 큰 솥들은 옆면으로 땅에 기대고 그릇과 쌀 같은 물질도 지면 위에 있어 공중 부유는 없다. 산채에서 퍼지는 연기도 물리적으로 가능한 모습이지만, 솥과 식기의 과도한 크기는 실제 생활용품의 축척과 맞지 않는다."
       }
      ],
      "all_candidates_fail": true
     },
     "normalized": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 6,
        "verdict_ko": "전신 와이드숏과 안정적인 능선 발판, 참조 장소의 배치는 잘 지켰지만, 시선이 등 뒤 산채가 아닌 앞쪽 땅을 향한다."
       },
       {
        "label": "A",
        "score": 4,
        "verdict_ko": "팔짱 낀 전신과 군도는 충실하지만 산채를 내려다보는 시선이 불명확하고, 배경의 솥과 그릇이 건물 대비 지나치게 커 실제 거리감과 축척을 훼손한다."
       }
      ],
      "all_candidates_fail": true,
      "readings": [
       {
        "label": "B",
        "direction": "토미이시는 몸을 거의 정면으로 세우고 눈을 카메라 쪽 아래로 내리깐다. 실제 시선의 대상은 앞쪽 능선 바닥으로 읽히며, 화면 왼쪽의 등 뒤 아래에 있는 산채를 향하지 않는다. 군도는 칼집에 들어가 허리에서 아래로 늘어져 있고 겨냥하는 동작은 없다.",
        "built_space": "전경에는 흙과 바위, 뿌리로 된 능선 발판이 있으며 인물은 그 위 중앙에 서 있다. 아래 산채에는 입구 하나가 난 목책과 뒤편 초가지붕 두 채, 인물 오른쪽 아래로 일부 드러난 지붕이 보인다. 왼쪽 외곽길, 바위군, 오른쪽 수목은 참조 장소의 배치와 대체로 이어진다. 인물의 머리부터 부츠까지 모두 포함하며, 발밑 지면이 선명하다. 반사는 없다.",
        "entities": "성인 동아시아계 남성 한 명만 보이며, 일본인이라는 국적 자체는 외관으로 확정할 수 없다. 검은 머리는 군모에 대부분 가려지고 콧수염과 턱수염, 얼굴과 체격은 참조와 대체로 일치한다. 갈색빛 군복, 붉은 깃장, 훈장, 가죽 교차 끈과 허리띠, 장화, 허리의 군도가 유지된다. 산채에는 연기와 작은 불길, 옆으로 넘어진 검은 솥 하나, 흩어진 그릇과 밝은 쌀 같은 물질이 보이나 피 묻은 쌀은 명확히 식별되지 않는다. 솥은 주변 건물에 비해 다소 크게 보인다. 낮의 저채도 자연광이며 능선의 인물은 연기에 덮이지 않는다. 읽을 수 있는 글자는 없다.",
        "hard_violations": [],
        "physics": "양쪽 부츠 밑창이 능선의 흙바닥에 닿아 체중을 받친다. 팔짱은 양팔을 몸 앞에서 교차해 서로 지지하는 가능한 자세다. 군도와 가죽 장비는 허리띠에 매달려 있고, 산채의 솥과 파편은 지면에 놓여 있다. 연기는 산채의 연소 지점에서 올라오며, 지지 없이 떠 있는 사람이나 물체는 보이지 않는다."
       },
       {
        "label": "A",
        "direction": "토미이시는 정면에서 약간 화면 왼쪽으로 몸과 얼굴을 틀고 아래를 본다. 그러나 눈길은 왼쪽 앞 능선 방향으로 읽히며, 그의 뒤쪽 아래에 펼쳐진 산채 내부에 정확히 닿는 시선은 아니다. 군도는 허리 옆에서 칼집째 아래로 향하며 사용하거나 겨냥하지 않는다.",
        "built_space": "토미이시는 화면 오른쪽의 흙과 뿌리가 드러난 능선에 서 있고 전신과 발밑 지면이 모두 보인다. 왼쪽 아래에는 입구 하나가 있는 목책, 뒤편의 초가 건물 두 채와 연기에 가려진 추가 지붕 부분이 보인다. 수목과 바위, 초가와 목책이라는 장소 요소는 유지되지만 산채가 화면에서 훨씬 크게 차지한다. 특히 솥 두 개와 여러 그릇이 건물 기둥 및 출입구에 비해 비정상적으로 커 보인다. 반사는 없다.",
        "entities": "성인 동아시아계 남성 한 명이며 참조와 유사한 콧수염, 턱수염과 체격을 갖는다. 국적은 외관만으로 판별할 수 없고 검은 머리는 군모에 대부분 가려져 있다. 군모, 갈색빛 군복, 붉은 깃장, 훈장, 가죽 끈과 허리띠, 장화, 군도를 유지한다. 산채에는 옆으로 누운 대형 솥 두 개, 흩어진 그릇, 흰 쌀 같은 덩어리와 붉은 얼룩이 보인다. 요청된 참상은 드러나지만 생활용품의 축척이 과장되어 있다. 낮의 저채도 조명이고 연기는 주로 산채와 배경을 채우며 인물 앞은 비교적 맑다. 추가 인물이나 읽을 수 있는 글자는 없다.",
        "hard_violations": [],
        "physics": "두 부츠가 능선 바닥에 닿아 서 있는 몸을 지지하며 팔짱도 해부학적으로 가능한 자세다. 군도는 허리 장비에 매달려 있다. 큰 솥들은 옆면으로 땅에 기대고 그릇과 쌀 같은 물질도 지면 위에 있어 공중 부유는 없다. 산채에서 퍼지는 연기도 물리적으로 가능한 모습이지만, 솥과 식기의 과도한 크기는 실제 생활용품의 축척과 맞지 않는다."
       }
      ]
     },
     "ok": true
    }
   ],
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "slot_winner_match": true,
   "slot_winner": {
    "gemini-pro": "B",
    "gpt-high": "B"
   },
   "route": "cross_slot_agree"
  },
  "dual": {
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "normalized": {
    "A": 0.889,
    "B": 2.0
   },
   "adjusted": {
    "A": 0.639,
    "B": 2.0
   },
   "violations": {
    "A": [
     "[gemini-pro] 물리적으로 불가능한 스테이징 및 스케일 (배경의 가마솥과 식기류가 건물 크기로 거대하게 렌더링됨)"
    ]
   },
   "per_model_winner": {
    "gemini-pro": "B",
    "gpt-high": "B"
   },
   "agreed": true
  },
  "totals": {
   "B": 2000,
   "A": 639
  },
  "selected": "B",
  "ranking": [
   "B",
   "A"
  ],
  "verdicts": [
   {
    "label": "B",
    "score": 2000,
    "verdict_ko": "프롬프트의 요구사항(팔짱을 끼고 내려다보는 전신)을 정확히 구현하였으며, 이전 샷과 동일한 배경의 물리적 비율을 사실적으로 유지했습니다."
   },
   {
    "label": "A",
    "score": 639,
    "verdict_ko": "인물의 자세는 맞으나, 산채 배경에 있는 가마솥과 사발 등의 사물이 집채만 한 크기로 과장되어 물리적 스케일을 완전히 위반했습니다.  ★위반: [gemini-pro] 물리적으로 불가능한 스테이징 및 스케일 (배경의 가마솥과 식기류가 건물 크기로 거대하게 렌더링됨)"
   }
  ],
  "refs": [
   {
    "label": "PREVIOUS SHOT STILL — a visually related earlier shot of this same place: the location's look, materials, fixed features and lighting mood are LOCKED to this photo; never copy its camera framing. The clothing, hair and overall look of 토미이시 — who appear both in that photo and in this shot — are LOCKED to that photo. Anyone else visible in that photo is NOT in this shot: never carry their face, body or clothing onto anyone here. If this photo shows a character who cannot move (dead or unconscious), that character's exact pose, position and orientation are ALSO LOCKED — treat their body as a fixed prop of the set that only the camera moves around.",
    "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/db00517b-165b-4005-a287-f28936b59b95/images/272bd7af-9294-4490-98c6-88a40e9deacc/scene/recipe/S56sh1_sel.png",
    "asset_id": "cb123cb1-3106-49d4-808b-250870ccfc1f",
    "role": "prev_still"
   },
   {
    "label": "CHARACTER REFERENCE — 토미이시: the exact person appearing in this shot; match face, hair and build exactly.",
    "path": "<bytes:797699>",
    "asset_id": "c861ba0f-8750-4262-8324-25cf5c6c9dcd",
    "role": "character_ref"
   }
  ],
  "critique_skipped": true,
  "shot_run_produced": false,
  "shot_run_uid": "06aa1438-edec-7679-bbf2-73f13038f5ff",
  "ref_mode": "prev+엔티티",
  "share_plan": {
   "ref_plan": "prev",
   "prev_anchor_tag": "S56sh1"
  }
 },
 "S58sh9::cine": {
  "applied": true,
  "attempted_at": "2026-09-09T07:37:13.994652+00:00",
  "fingerprint": "49a847c259d1e41d526c4d5a7b707aef5aebbcf1cf234531942663baef5d8a71",
  "fingerprint_version": 2,
  "provider": "mai",
  "endpoint": "openrouter/chat-completions",
  "model": "microsoft/mai-image-2.6",
  "pack": "24.202608252115",
  "source_file": "S58sh9_sel.png",
  "source_sha256": "81d71b0a4f332a159fb2d04f85b160b37ec70c2ab9346bc54bfe76059cd397f6",
  "file": "S58sh9_cine.png",
  "staged_sha256": "2615723aa886571fd8ba1012f8b9abd8d18d22d335df5e0d131d48ae59d8f886",
  "latency_ms": 30541
 },
 "S58sh12::signage": {
  "fp": "87c07e5c0ebcdec1",
  "inscriptions": [],
  "cues": [],
  "dropped": []
 },
 "S58sh12": {
  "input_fingerprint": "4a96482d67b37cde",
  "prompt": "Create ONE FINAL photorealistic live-action film still of the moment below — early 20th-century Korea, Jeolla region; all people are Korean unless explicitly stated otherwise. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 몸은 산 아래를 향한 채 고개만 피바다가 된 산채 쪽을 향해 홱 꺾여 있는 배상도의 측면.\n\nLOCATION (lock): The departure edge of the overlooking ridge, where the downhill forest route leads away from the devastated camp. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: medium shot\n- KEY BACKGROUND ELEMENTS: 하산 동선 (The route 배상도 follows after the departing soldiers); used as The visible ground slopes through the lower background toward screen right, supporting the opposition between his departing body and backward glance.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Natural daylight, subdued saturation, and restrained contrast hold the backward glance in the same sober aftermath without adding a new atmospheric effect.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The people visible in that photo are NOT in this shot: never carry their faces, bodies or clothing onto anyone here. Each person in THIS shot is defined solely by the PEOPLE list, their own CHARACTER REFERENCE images and their wardrobe notes. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The smoke-filled camp retains its overturned pot, broken dishes, bloodied rice and collapsed timber. 배상도: He turns back while beginning the descent.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 배상도 (한국인 남성, 성인 나이대, 자연스러운 검은 머리) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.",
  "roll_prompts": {
   "A": "Create ONE FINAL photorealistic live-action film still of the moment below — early 20th-century Korea, Jeolla region; all people are Korean unless explicitly stated otherwise. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 몸은 산 아래를 향한 채 고개만 피바다가 된 산채 쪽을 향해 홱 꺾여 있는 배상도의 측면.\n\nLOCATION (lock): The departure edge of the overlooking ridge, where the downhill forest route leads away from the devastated camp. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: medium shot\n- KEY BACKGROUND ELEMENTS: 하산 동선 (The route 배상도 follows after the departing soldiers); used as The visible ground slopes through the lower background toward screen right, supporting the opposition between his departing body and backward glance.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Natural daylight, subdued saturation, and restrained contrast hold the backward glance in the same sober aftermath without adding a new atmospheric effect.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The people visible in that photo are NOT in this shot: never carry their faces, bodies or clothing onto anyone here. Each person in THIS shot is defined solely by the PEOPLE list, their own CHARACTER REFERENCE images and their wardrobe notes. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The smoke-filled camp retains its overturned pot, broken dishes, bloodied rice and collapsed timber. 배상도: He turns back while beginning the descent.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 배상도 (한국인 남성, 성인 나이대, 자연스러운 검은 머리) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.",
   "B": "Create ONE FINAL photorealistic live-action film still of the moment below — early 20th-century Korea, Jeolla region; all people are Korean unless explicitly stated otherwise. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 몸은 산 아래를 향한 채 고개만 피바다가 된 산채 쪽을 향해 홱 꺾여 있는 배상도의 측면.\n\nLOCATION (lock): The departure edge of the overlooking ridge, where the downhill forest route leads away from the devastated camp. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: medium shot\n- KEY BACKGROUND ELEMENTS: 하산 동선 (The route 배상도 follows after the departing soldiers); used as The visible ground slopes through the lower background toward screen right, supporting the opposition between his departing body and backward glance.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Natural daylight, subdued saturation, and restrained contrast hold the backward glance in the same sober aftermath without adding a new atmospheric effect.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The people visible in that photo are NOT in this shot: never carry their faces, bodies or clothing onto anyone here. Each person in THIS shot is defined solely by the PEOPLE list, their own CHARACTER REFERENCE images and their wardrobe notes. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The smoke-filled camp retains its overturned pot, broken dishes, bloodied rice and collapsed timber. 배상도: He turns back while beginning the descent.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 배상도 (한국인 남성, 성인 나이대, 자연스러운 검은 머리) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.\n\nCOMPOSITION VARIATION (this candidate only): keep the place, the people and the moment exactly as specified, and keep from the CAMERA & FRAME contract its subject scale, screen placement and key background elements exactly. For this candidate only, its camera angle and height are deliberately left open: choose a camera position distinctly different from the obvious one, as a film director picking a second setup on the same blocking. Stay true to any explicit framing the SHOT TEXT itself dictates."
  },
  "shot_run_spend_attempt_count": 1,
  "readings": [
   {
    "label": "A",
    "direction": "몸통과 얼굴 모두 화면 좌측(산채 방향)을 향하고 있어 몸과 고개의 방향이 엇갈리는 묘사가 전혀 이루어지지 않음.",
    "built_space": "이전 샷과 유사한 산채 배경과 능선이 존재하나, 산채 바닥의 핏자국 묘사가 부족함.",
    "entities": "배상도의 얼굴과 한복, 모자 등 레퍼런스의 외형 요소는 일치함.",
    "hard_violations": [
     "[gemini-pro] 지시된 신체 방향과 시선의 분리가 전혀 이루어지지 않은 완전히 잘못된 포즈",
     "[gemini-pro] 인물 머리 뒤편 배경 허공에 아무런 지지대 없이 일본도 손잡이 모양의 막대가 떠 있음",
     "[gpt-high] 기준 장소에 없고 현재 장면의 소품으로도 지정되지 않은, 손잡이와 원형 코등이를 가진 긴 칼 형태의 수직 물체를 배경에 추가했다."
    ],
    "physics": "인물은 땅에 서 있으나, 배경 허공에 아무것도 쥐고 있지 않은 칼자루 막대가 중력을 거스르고 떠 있음."
   },
   {
    "label": "B",
    "direction": "몸은 산 아래(화면 우측)를 향하고 있으며, 고개를 왼쪽 어깨너머로 홱 꺾어 피바다가 된 산채(화면 좌측)를 바라보고 있음.",
    "built_space": "이전 샷의 흙길과 능선, 연기 나는 산채의 구조가 유지되었으며, 가마솥과 핏자국이 바닥에 배치되어 있음.",
    "entities": "배상도의 얼굴, 체형, 한복 의상이 레퍼런스와 일치하며, 프롬프트에 명시된 '자연스러운 검은 머리'가 잘 표현됨.",
    "hard_violations": [],
    "physics": "지면에 안정적으로 서 있으며, 뒤돌아보는 자세의 무게 중심과 옷의 주름이 자연스럽게 지탱됨."
   }
  ],
  "cross_model_order": {
   "policy": "cross_model_order_v2_slot0_fwd_slot1_rev_parallel2",
   "slots": [
    {
     "model": "gemini-pro",
     "order": "forward",
     "display_to_canonical": {
      "A": "A",
      "B": "B"
     },
     "raw": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 10,
        "verdict_ko": "프롬프트가 요구한 '몸은 하산 방향, 고개만 산채 쪽으로 꺾인' 구도와 피바다가 된 산채의 모습을 완벽하게 구현했습니다."
       },
       {
        "label": "A",
        "score": 2,
        "verdict_ko": "몸과 시선이 모두 같은 방향을 향해 있어 요구된 자세를 실패했으며, 배경에 공중에 뜬 물체가 존재하는 치명적인 오류가 있습니다."
       }
      ],
      "readings": [
       {
        "label": "B",
        "direction": "몸은 산 아래(화면 우측)를 향하고 있으며, 고개를 왼쪽 어깨너머로 홱 꺾어 피바다가 된 산채(화면 좌측)를 바라보고 있음.",
        "built_space": "이전 샷의 흙길과 능선, 연기 나는 산채의 구조가 유지되었으며, 가마솥과 핏자국이 바닥에 배치되어 있음.",
        "entities": "배상도의 얼굴, 체형, 한복 의상이 레퍼런스와 일치하며, 프롬프트에 명시된 '자연스러운 검은 머리'가 잘 표현됨.",
        "hard_violations": [],
        "physics": "지면에 안정적으로 서 있으며, 뒤돌아보는 자세의 무게 중심과 옷의 주름이 자연스럽게 지탱됨."
       },
       {
        "label": "A",
        "direction": "몸통과 얼굴 모두 화면 좌측(산채 방향)을 향하고 있어 몸과 고개의 방향이 엇갈리는 묘사가 전혀 이루어지지 않음.",
        "built_space": "이전 샷과 유사한 산채 배경과 능선이 존재하나, 산채 바닥의 핏자국 묘사가 부족함.",
        "entities": "배상도의 얼굴과 한복, 모자 등 레퍼런스의 외형 요소는 일치함.",
        "hard_violations": [
         "지시된 신체 방향과 시선의 분리가 전혀 이루어지지 않은 완전히 잘못된 포즈",
         "인물 머리 뒤편 배경 허공에 아무런 지지대 없이 일본도 손잡이 모양의 막대가 떠 있음"
        ],
        "physics": "인물은 땅에 서 있으나, 배경 허공에 아무것도 쥐고 있지 않은 칼자루 막대가 중력을 거스르고 떠 있음."
       }
      ],
      "all_candidates_fail": false
     },
     "normalized": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 10,
        "verdict_ko": "프롬프트가 요구한 '몸은 하산 방향, 고개만 산채 쪽으로 꺾인' 구도와 피바다가 된 산채의 모습을 완벽하게 구현했습니다."
       },
       {
        "label": "A",
        "score": 2,
        "verdict_ko": "몸과 시선이 모두 같은 방향을 향해 있어 요구된 자세를 실패했으며, 배경에 공중에 뜬 물체가 존재하는 치명적인 오류가 있습니다."
       }
      ],
      "all_candidates_fail": false,
      "readings": [
       {
        "label": "B",
        "direction": "몸은 산 아래(화면 우측)를 향하고 있으며, 고개를 왼쪽 어깨너머로 홱 꺾어 피바다가 된 산채(화면 좌측)를 바라보고 있음.",
        "built_space": "이전 샷의 흙길과 능선, 연기 나는 산채의 구조가 유지되었으며, 가마솥과 핏자국이 바닥에 배치되어 있음.",
        "entities": "배상도의 얼굴, 체형, 한복 의상이 레퍼런스와 일치하며, 프롬프트에 명시된 '자연스러운 검은 머리'가 잘 표현됨.",
        "hard_violations": [],
        "physics": "지면에 안정적으로 서 있으며, 뒤돌아보는 자세의 무게 중심과 옷의 주름이 자연스럽게 지탱됨."
       },
       {
        "label": "A",
        "direction": "몸통과 얼굴 모두 화면 좌측(산채 방향)을 향하고 있어 몸과 고개의 방향이 엇갈리는 묘사가 전혀 이루어지지 않음.",
        "built_space": "이전 샷과 유사한 산채 배경과 능선이 존재하나, 산채 바닥의 핏자국 묘사가 부족함.",
        "entities": "배상도의 얼굴과 한복, 모자 등 레퍼런스의 외형 요소는 일치함.",
        "hard_violations": [
         "지시된 신체 방향과 시선의 분리가 전혀 이루어지지 않은 완전히 잘못된 포즈",
         "인물 머리 뒤편 배경 허공에 아무런 지지대 없이 일본도 손잡이 모양의 막대가 떠 있음"
        ],
        "physics": "인물은 땅에 서 있으나, 배경 허공에 아무것도 쥐고 있지 않은 칼자루 막대가 중력을 거스르고 떠 있음."
       }
      ]
     },
     "ok": true
    },
    {
     "model": "gpt-high",
     "order": "reverse",
     "display_to_canonical": {
      "A": "B",
      "B": "A"
     },
     "raw": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 4,
        "verdict_ko": "몸이 오른쪽 하산로를 향하는 관계는 낫지만, 고개와 시선이 왼쪽 뒤 산채가 아닌 카메라로 향하고 기준 의상의 검은 모자도 빠졌다."
       },
       {
        "label": "B",
        "score": 2,
        "verdict_ko": "허리 위 측면 구도와 모자는 맞지만 몸과 얼굴이 모두 왼쪽을 향해 돌아보는 동작이 없고, 기준 장소에 없던 칼 형태의 긴 물체가 추가됐다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "몸통과 옷 앞쪽은 화면 오른쪽 길을 향한다. 얼굴은 카메라 쪽으로 돌아와 있고 눈도 관객 쪽을 본다. 산채는 인물의 왼쪽 뒤 아래에 있으므로 시선이 요구된 산채에 닿지 않는다. 겨누거나 들고 있는 물건은 없다.",
        "built_space": "전경의 흙 능선과 오른쪽 숲길, 드러난 뿌리와 바위가 보인다. 아래 산채에는 둘러친 목책 한 줄, 전면 출입구 하나, 뒤쪽 초가 건물과 오른쪽 무너진 목구조가 있다. 뒤집힌 큰 솥 하나가 마당에 놓여 있다. 기준 장소의 주요 재료와 배치는 대체로 이어지지만, 오른쪽 길이 아래로 이어지는 경사는 뚜렷하지 않다. 인물은 능선 길 가장자리에 서 있다.",
        "entities": "뚜렷하게 식별되는 인물은 검은 머리의 한국인 성인 남성 한 명이다. 흰 전통 복식과 검은 조끼는 배상도 기준과 맞지만 검은 모자는 없다. 얼굴의 동일성은 기준 사진과 완전히 일치한다고 보기 어렵다. 산채에는 솥, 흩어진 그릇과 밝은 곡물 같은 내용물, 붉은 얼룩, 붕괴한 목재와 연기가 보인다. 곡물에 피가 묻었는지까지는 판별하기 어렵다. 읽을 수 있는 글자는 없다.",
        "hard_violations": [],
        "physics": "인물은 허벅지 부근에서 잘려 발 접점은 보이지 않지만, 몸통과 팔은 지면에 서 있는 사람의 정상적인 자세로 이어진다. 공중에 떠 있다는 증거는 없다. 손은 자연스럽게 내려와 있으며 물건을 들지 않는다. 솥과 잔해는 마당에 놓이고 연기는 불에서 위로 올라간다. 다만 하산을 막 시작한 체중 이동이나 급하게 고개를 꺾는 동작은 약하다."
       },
       {
        "label": "B",
        "direction": "코와 시선은 화면 왼쪽을 향하지만, 조끼 앞섶과 가슴도 같은 왼쪽을 향한다. 왼쪽 배경에 산채가 있으나 눈은 그보다 높은 수평 방향을 보는 인상이다. 오른쪽 하산로를 향한 몸과 뒤돌아본 머리의 반대 방향 관계가 없다. 배경의 칼 형태 물체는 수직으로 서 있다.",
        "built_space": "인물은 오른쪽 전경의 흙 능선에 있고, 허리 위 중심의 측면 구도다. 아래에는 목책 한 줄과 전면 출입구 하나, 뒤쪽 초가 건물, 오른쪽 붕괴한 지붕 골조, 마당의 솥 하나가 보인다. 오른쪽 가장자리에 길이 남아 있으나 하산 경사는 명확하지 않다. 산채와 능선 사이에 손잡이와 원형 코등이처럼 보이는 부분을 가진 긴 수직 물체 하나가 새로 서 있어 기준 장소의 고정 요소와 맞지 않는다.",
        "entities": "한국인 성인 남성 한 명이 보이며 검은 모자, 검은 조끼, 흰 옷과 목 부분의 매듭은 인물 기준에 가깝다. 검은 머리는 모자 아래로 조금 보인다. 솥, 깨진 그릇, 흩어진 밝은 내용물, 붉은 얼룩, 무너진 목재와 연기가 있다. 인물 뒤의 긴 칼 형태 물체는 이 장면에 지정되지 않았다. 읽을 수 있는 글자는 없다.",
        "hard_violations": [
         "기준 장소에 없고 현재 장면의 소품으로도 지정되지 않은, 손잡이와 원형 코등이를 가진 긴 칼 형태의 수직 물체를 배경에 추가했다."
        ],
        "physics": "발은 프레임 밖이므로 실제 접지는 확인되지 않지만 상체는 정상적인 직립 자세이고 부유 징후는 없다. 팔은 아래로 이어지며 손에 든 물건은 보이지 않는다. 산채의 솥과 파편은 지면에 놓이고 무너진 골조는 남은 구조물과 바닥에 걸쳐 있다. 긴 수직 물체는 아래쪽이 지면 부근의 식생에 가려져 고정 방식은 확인할 수 없다. 목과 몸통은 거의 같은 방향이라 급히 뒤돌아보는 신체 동작이 성립하지 않는다."
       }
      ],
      "all_candidates_fail": true
     },
     "normalized": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 4,
        "verdict_ko": "몸이 오른쪽 하산로를 향하는 관계는 낫지만, 고개와 시선이 왼쪽 뒤 산채가 아닌 카메라로 향하고 기준 의상의 검은 모자도 빠졌다."
       },
       {
        "label": "A",
        "score": 2,
        "verdict_ko": "허리 위 측면 구도와 모자는 맞지만 몸과 얼굴이 모두 왼쪽을 향해 돌아보는 동작이 없고, 기준 장소에 없던 칼 형태의 긴 물체가 추가됐다."
       }
      ],
      "all_candidates_fail": true,
      "readings": [
       {
        "label": "B",
        "direction": "몸통과 옷 앞쪽은 화면 오른쪽 길을 향한다. 얼굴은 카메라 쪽으로 돌아와 있고 눈도 관객 쪽을 본다. 산채는 인물의 왼쪽 뒤 아래에 있으므로 시선이 요구된 산채에 닿지 않는다. 겨누거나 들고 있는 물건은 없다.",
        "built_space": "전경의 흙 능선과 오른쪽 숲길, 드러난 뿌리와 바위가 보인다. 아래 산채에는 둘러친 목책 한 줄, 전면 출입구 하나, 뒤쪽 초가 건물과 오른쪽 무너진 목구조가 있다. 뒤집힌 큰 솥 하나가 마당에 놓여 있다. 기준 장소의 주요 재료와 배치는 대체로 이어지지만, 오른쪽 길이 아래로 이어지는 경사는 뚜렷하지 않다. 인물은 능선 길 가장자리에 서 있다.",
        "entities": "뚜렷하게 식별되는 인물은 검은 머리의 한국인 성인 남성 한 명이다. 흰 전통 복식과 검은 조끼는 배상도 기준과 맞지만 검은 모자는 없다. 얼굴의 동일성은 기준 사진과 완전히 일치한다고 보기 어렵다. 산채에는 솥, 흩어진 그릇과 밝은 곡물 같은 내용물, 붉은 얼룩, 붕괴한 목재와 연기가 보인다. 곡물에 피가 묻었는지까지는 판별하기 어렵다. 읽을 수 있는 글자는 없다.",
        "hard_violations": [],
        "physics": "인물은 허벅지 부근에서 잘려 발 접점은 보이지 않지만, 몸통과 팔은 지면에 서 있는 사람의 정상적인 자세로 이어진다. 공중에 떠 있다는 증거는 없다. 손은 자연스럽게 내려와 있으며 물건을 들지 않는다. 솥과 잔해는 마당에 놓이고 연기는 불에서 위로 올라간다. 다만 하산을 막 시작한 체중 이동이나 급하게 고개를 꺾는 동작은 약하다."
       },
       {
        "label": "A",
        "direction": "코와 시선은 화면 왼쪽을 향하지만, 조끼 앞섶과 가슴도 같은 왼쪽을 향한다. 왼쪽 배경에 산채가 있으나 눈은 그보다 높은 수평 방향을 보는 인상이다. 오른쪽 하산로를 향한 몸과 뒤돌아본 머리의 반대 방향 관계가 없다. 배경의 칼 형태 물체는 수직으로 서 있다.",
        "built_space": "인물은 오른쪽 전경의 흙 능선에 있고, 허리 위 중심의 측면 구도다. 아래에는 목책 한 줄과 전면 출입구 하나, 뒤쪽 초가 건물, 오른쪽 붕괴한 지붕 골조, 마당의 솥 하나가 보인다. 오른쪽 가장자리에 길이 남아 있으나 하산 경사는 명확하지 않다. 산채와 능선 사이에 손잡이와 원형 코등이처럼 보이는 부분을 가진 긴 수직 물체 하나가 새로 서 있어 기준 장소의 고정 요소와 맞지 않는다.",
        "entities": "한국인 성인 남성 한 명이 보이며 검은 모자, 검은 조끼, 흰 옷과 목 부분의 매듭은 인물 기준에 가깝다. 검은 머리는 모자 아래로 조금 보인다. 솥, 깨진 그릇, 흩어진 밝은 내용물, 붉은 얼룩, 무너진 목재와 연기가 있다. 인물 뒤의 긴 칼 형태 물체는 이 장면에 지정되지 않았다. 읽을 수 있는 글자는 없다.",
        "hard_violations": [
         "기준 장소에 없고 현재 장면의 소품으로도 지정되지 않은, 손잡이와 원형 코등이를 가진 긴 칼 형태의 수직 물체를 배경에 추가했다."
        ],
        "physics": "발은 프레임 밖이므로 실제 접지는 확인되지 않지만 상체는 정상적인 직립 자세이고 부유 징후는 없다. 팔은 아래로 이어지며 손에 든 물건은 보이지 않는다. 산채의 솥과 파편은 지면에 놓이고 무너진 골조는 남은 구조물과 바닥에 걸쳐 있다. 긴 수직 물체는 아래쪽이 지면 부근의 식생에 가려져 고정 방식은 확인할 수 없다. 목과 몸통은 거의 같은 방향이라 급히 뒤돌아보는 신체 동작이 성립하지 않는다."
       }
      ]
     },
     "ok": true
    }
   ],
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "slot_winner_match": true,
   "slot_winner": {
    "gemini-pro": "B",
    "gpt-high": "B"
   },
   "route": "cross_slot_agree"
  },
  "dual": {
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "normalized": {
    "A": 0.7,
    "B": 2.0
   },
   "adjusted": {
    "A": 0.45,
    "B": 2.0
   },
   "violations": {
    "A": [
     "[gemini-pro] 지시된 신체 방향과 시선의 분리가 전혀 이루어지지 않은 완전히 잘못된 포즈",
     "[gemini-pro] 인물 머리 뒤편 배경 허공에 아무런 지지대 없이 일본도 손잡이 모양의 막대가 떠 있음",
     "[gpt-high] 기준 장소에 없고 현재 장면의 소품으로도 지정되지 않은, 손잡이와 원형 코등이를 가진 긴 칼 형태의 수직 물체를 배경에 추가했다."
    ]
   },
   "per_model_winner": {
    "gemini-pro": "B",
    "gpt-high": "B"
   },
   "agreed": true
  },
  "totals": {
   "B": 2000,
   "A": 450
  },
  "selected": "B",
  "ranking": [
   "B",
   "A"
  ],
  "verdicts": [
   {
    "label": "B",
    "score": 2000,
    "verdict_ko": "프롬프트가 요구한 '몸은 하산 방향, 고개만 산채 쪽으로 꺾인' 구도와 피바다가 된 산채의 모습을 완벽하게 구현했습니다."
   },
   {
    "label": "A",
    "score": 450,
    "verdict_ko": "몸과 시선이 모두 같은 방향을 향해 있어 요구된 자세를 실패했으며, 배경에 공중에 뜬 물체가 존재하는 치명적인 오류가 있습니다.  ★위반: [gemini-pro] 지시된 신체 방향과 시선의 분리가 전혀 이루어지지 않은 완전히 잘못된 포즈 / [gemini-pro] 인물 머리 뒤편 배경 허공에 아무런 지지대 없이 일본도 손잡이 모양의 막대가 떠 있음 / [gpt-high] 기준 장소에 없고 현재 장면의 소품으로도 지정되지 않은, 손잡이와 원형 코등이를 가진 긴 칼 형태의 수직 물체를 배경에 추가했다."
   }
  ],
  "refs": [
   {
    "label": "PREVIOUS SHOT STILL — a visually related earlier shot of this same place: the location's look, materials, fixed features and lighting mood are LOCKED to this photo; never copy its camera framing. The people visible in that photo are NOT in this shot: never carry their faces, bodies or clothing onto anyone here. Each person in THIS shot is defined solely by the PEOPLE list, their own CHARACTER REFERENCE images and their wardrobe notes. If this photo shows a character who cannot move (dead or unconscious), that character's exact pose, position and orientation are ALSO LOCKED — treat their body as a fixed prop of the set that only the camera moves around.",
    "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/db00517b-165b-4005-a287-f28936b59b95/images/272bd7af-9294-4490-98c6-88a40e9deacc/scene/recipe/S58sh9_sel.png",
    "asset_id": "8483955c-17e5-4bc9-99be-21ee46136900",
    "role": "prev_still"
   },
   {
    "label": "CHARACTER REFERENCE — 배상도: the exact person appearing in this shot; match face, hair and build exactly.",
    "path": "<bytes:851718>",
    "asset_id": "034f5325-975e-4474-854b-e48359710bf8",
    "role": "character_ref"
   }
  ],
  "critique_skipped": true,
  "shot_run_produced": false,
  "shot_run_uid": "06aa1438-ef90-7071-b905-635c750c1f3d",
  "ref_mode": "prev+엔티티",
  "share_plan": {
   "ref_plan": "prev",
   "prev_anchor_tag": "S58sh9"
  }
 },
 "S58sh12::cine": {
  "applied": true,
  "attempted_at": "2026-09-09T07:38:42.529781+00:00",
  "fingerprint": "f50a747d5fdec4ee0a715eddd4c8c638f78b1e188d661bc83633fcfda0a1c22c",
  "fingerprint_version": 2,
  "provider": "mai",
  "endpoint": "openrouter/chat-completions",
  "model": "microsoft/mai-image-2.6",
  "pack": "24.202608252115",
  "source_file": "S58sh12_sel.png",
  "source_sha256": "2a952b601af325920024309d4c4994a289e599f9513f23449c9c51020a58d353",
  "file": "S58sh12_cine.png",
  "staged_sha256": "abcd01c2635e11a11e758624f30c7373e13bb7a7e1d95def7d8a73f030e60aaf",
  "latency_ms": 25002
 },
 "S59sh2::signage": {
  "fp": "3c4eef35b3bdd1c8",
  "inscriptions": [],
  "cues": [],
  "dropped": []
 },
 "era_assess::4be972a49c647ce0": {
  "subjects": [],
  "subject_text": "파청 산채 진입 오르막 숲길\n나무 사이로 산채까지 이어지는 가파른 흙길. 비탈과 굽은 길이 겹치며, 아래쪽 마을이 수목 틈으로 멀리 내려다보인다.",
  "identity": "canonical",
  "scope_id": "L25",
  "scope_role": "location_exterior",
  "scope_sha": "f0c39be78791dd59"
 },
 "groupbg::wounded_ascent": {
  "input_fingerprint": "bfa9a574e00687bc",
  "meta": {
   "model": "gpt-image-2",
   "size": "1536x864",
   "pack": "11.202607220237",
   "contract": "bgfirst_full_v3",
   "group_sig": {
    "key": "wounded_ascent",
    "tags": [
     "S59sh2",
     "S59sh9"
    ]
   },
   "context_sig": "e4cc80365f37f3d1"
  },
  "prompt": "Create ONE empty live-action location background photograph — NO PEOPLE, no figures, no body parts, no silhouettes, no shadows or reflections of people anywhere.\n\"Empty\" means no people only: KEEP the location's inherent occupants and stock that define the place — animals in an animal shelter, pen or farm, goods and produce in a market, moored boats in a harbour — unless the shot text explicitly removes them.\nThe FIRST attached image is a thin-line storyboard sketch of a shot that happens at this location: use it ONLY as spatial evidence — what this place contains, how its ground, structures and landmarks are arranged and proportioned. Ignore the sketched people and arrows entirely, and do NOT copy its line style: render a fully photographic, physically plausible real place that fits THE LOCATION text below.\nHUMAN-SCALE CALIBRATION: derive every structure's true size from human-scale elements — a door ≈ 2m, a window ≈ 1–1.5m wide, one storey ≈ 2.5–3m; never inflate a small structure or shrink a large one.\n\nTHE LOCATION — early 20th-century Korea, Jeolla region; all people are Korean unless explicitly stated otherwise: The steep wooded uphill path leading from the market side toward the mountain camp.\n\nLOCATION DETAIL (descriptive evidence about this place from the production's location records — evidence, not a staging order): use it to understand what kind of place this is and what it permanently contains. Carry over ONLY its enduring physical features — layout, structures, ground and materials, fixed equipment and installations. Do NOT treat any lighting, weather or time-of-day wording, momentary object states, depicted events or subjective impressions in it as fixed facts of the place; the TIME OF DAY lock below is the sole authority for time and lighting.\n파청 산채 진입 오르막 숲길: 산채로 연결되는 경사가 가파른 흙비탈 숲길. 나무 뿌리와 잡목이 길가로 돌출되어 있다. (특징: 경사가 심한 흙바닥 비탈; 우거진 수목과 돌출된 나무 뿌리; 군화나 짚신 자국이 남는 흙길)\n\nSCENE EVIDENCE (verbatim quotes from the screenplay about this place — treat them as evidence of what the location physically contains and looks like; stage the PLACE those moments happen in, but do NOT depict the momentary actions, people or staged props themselves):\n- 규홍이 혼자 산을 오른다.\n- 비탈에서 발을 헛디딘다.\n- 뒤를 돌아본다. 나무 사이로 저잣거리 쪽이 아득하게 내려다보인다.\n\nTIME OF DAY (lock): day.\n\nRender ONE photorealistic empty location photograph, 16:9, neutral enough that every shot of this place can be staged from it later. No readable writing anywhere: surfaces that would carry writing may be present, but stage any wording out of legibility — an oblique angle, distance, shallow focus. No captions, watermarks or overlay text.",
  "effective_prompt": "Create ONE empty live-action location background photograph — NO PEOPLE, no figures, no body parts, no silhouettes, no shadows or reflections of people anywhere.\n\"Empty\" means no people only: KEEP the location's inherent occupants and stock that define the place — animals in an animal shelter, pen or farm, goods and produce in a market, moored boats in a harbour — unless the shot text explicitly removes them.\nThe FIRST attached image is a thin-line storyboard sketch of a shot that happens at this location: use it ONLY as spatial evidence — what this place contains, how its ground, structures and landmarks are arranged and proportioned. Ignore the sketched people and arrows entirely, and do NOT copy its line style: render a fully photographic, physically plausible real place that fits THE LOCATION text below.\nHUMAN-SCALE CALIBRATION: derive every structure's true size from human-scale elements — a door ≈ 2m, a window ≈ 1–1.5m wide, one storey ≈ 2.5–3m; never inflate a small structure or shrink a large one.\n\nTHE LOCATION — early 20th-century Korea, Jeolla region; all people are Korean unless explicitly stated otherwise: The steep wooded uphill path leading from the market side toward the mountain camp.\n\nLOCATION DETAIL (descriptive evidence about this place from the production's location records — evidence, not a staging order): use it to understand what kind of place this is and what it permanently contains. Carry over ONLY its enduring physical features — layout, structures, ground and materials, fixed equipment and installations. Do NOT treat any lighting, weather or time-of-day wording, momentary object states, depicted events or subjective impressions in it as fixed facts of the place; the TIME OF DAY lock below is the sole authority for time and lighting.\n파청 산채 진입 오르막 숲길: 산채로 연결되는 경사가 가파른 흙비탈 숲길. 나무 뿌리와 잡목이 길가로 돌출되어 있다. (특징: 경사가 심한 흙바닥 비탈; 우거진 수목과 돌출된 나무 뿌리; 군화나 짚신 자국이 남는 흙길)\n\nSCENE EVIDENCE (verbatim quotes from the screenplay about this place — treat them as evidence of what the location physically contains and looks like; stage the PLACE those moments happen in, but do NOT depict the momentary actions, people or staged props themselves):\n- 규홍이 혼자 산을 오른다.\n- 비탈에서 발을 헛디딘다.\n- 뒤를 돌아본다. 나무 사이로 저잣거리 쪽이 아득하게 내려다보인다.\n\nTIME OF DAY (lock): day.\n\nRender ONE photorealistic empty location photograph, 16:9, neutral enough that every shot of this place can be staged from it later. No readable writing anywhere: surfaces that would carry writing may be present, but stage any wording out of legibility — an oblique angle, distance, shallow focus. No captions, watermarks or overlay text.",
  "bg_path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/db00517b-165b-4005-a287-f28936b59b95/images/272bd7af-9294-4490-98c6-88a40e9deacc/scene/recipe/groupbg_wounded_ascent_f8055c.png",
  "asset_id": "5fa31c1e-0d16-4a07-9d51-1572b715fb98",
  "input_asset_ids": [
   "2e08c992-7e18-41ad-8a14-57af8366804d"
  ],
  "origin_tag": "S59sh2",
  "place_text": "The steep wooded uphill path leading from the market side toward the mountain camp.",
  "origin_inputs": {
   "place_text": "The steep wooded uphill path leading from the market side toward the mountain camp.",
   "time_of_day_en": "day",
   "conti_asset_id": "2e08c992-7e18-41ad-8a14-57af8366804d"
  }
 },
 "S59sh2::bgfirst_bg": {
  "input_fingerprint": "6cf9e3f7c1173f16",
  "prompt": "Create the EMPTY BACKGROUND PLATE for one film shot — NO PEOPLE, no figures, no body parts, no sketch lines, no arrows anywhere.\n\"Empty\" means no people only: KEEP the location's inherent occupants and stock that define the place — animals in an animal shelter, pen or farm, goods in a market, moored boats in a harbour — unless the shot text explicitly removes them.\nThe FIRST attached image is a thin-line storyboard sketch: use ONLY its camera angle, horizon, perspective and the placement/size of buildings and set masses — ignore the sketched people and arrows entirely. The SECOND attached image (LOCATION PHOTOGRAPH) is the real place: take its architecture, materials, signage and fixed features, and RE-PROJECT them into the sketch's camera. If the photograph's camera differs from the sketch's, the sketch's camera wins.\nHUMAN-SCALE CALIBRATION: derive every structure's true size from human-scale elements — a door ≈ 2m, a window ≈ 1–1.5m wide, one storey ≈ 2.5–3m; never inflate a small structure or shrink a large one.\n\nSHOT TEXT this background must serve (Korean): 안규홍의 등 한가운데 난 깊은 상처에서 붉은 피가 흘러 저고리를 짙게 적신 클로즈업.\n\nLOCATION (lock): The steep wooded uphill path leading from the market side toward the mountain camp.\n\nTIME OF DAY (lock): day.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: close-up\n- KEY BACKGROUND ELEMENTS: Uphill forest path (Visible only at the margins of the climbing torso) — The path continues uphill beyond the subject; used as Provides a restrained spatial reference without competing with the wound.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Daylight keeps the red blood and darkened clothing legible with restrained saturation and naturalistic contrast.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nRender ONE photorealistic empty location photograph, 16:9, that this shot can be staged inside later. No readable writing anywhere: surfaces that would carry writing may be present, but stage any wording out of legibility — an oblique angle, distance, shallow focus. No captions, watermarks or overlay text.",
  "effective_prompt": "Create the EMPTY BACKGROUND PLATE for one film shot — NO PEOPLE, no figures, no body parts, no sketch lines, no arrows anywhere.\n\"Empty\" means no people only: KEEP the location's inherent occupants and stock that define the place — animals in an animal shelter, pen or farm, goods in a market, moored boats in a harbour — unless the shot text explicitly removes them.\nThe FIRST attached image is a thin-line storyboard sketch: use ONLY its camera angle, horizon, perspective and the placement/size of buildings and set masses — ignore the sketched people and arrows entirely. The SECOND attached image (LOCATION PHOTOGRAPH) is the real place: take its architecture, materials, signage and fixed features, and RE-PROJECT them into the sketch's camera. If the photograph's camera differs from the sketch's, the sketch's camera wins.\nHUMAN-SCALE CALIBRATION: derive every structure's true size from human-scale elements — a door ≈ 2m, a window ≈ 1–1.5m wide, one storey ≈ 2.5–3m; never inflate a small structure or shrink a large one.\n\nSHOT TEXT this background must serve (Korean): 안규홍의 등 한가운데 난 깊은 상처에서 붉은 피가 흘러 저고리를 짙게 적신 클로즈업.\n\nLOCATION (lock): The steep wooded uphill path leading from the market side toward the mountain camp.\n\nTIME OF DAY (lock): day.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: close-up\n- KEY BACKGROUND ELEMENTS: Uphill forest path (Visible only at the margins of the climbing torso) — The path continues uphill beyond the subject; used as Provides a restrained spatial reference without competing with the wound.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Daylight keeps the red blood and darkened clothing legible with restrained saturation and naturalistic contrast.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nRender ONE photorealistic empty location photograph, 16:9, that this shot can be staged inside later. No readable writing anywhere: surfaces that would carry writing may be present, but stage any wording out of legibility — an oblique angle, distance, shallow focus. No captions, watermarks or overlay text.",
  "bg_path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/db00517b-165b-4005-a287-f28936b59b95/images/272bd7af-9294-4490-98c6-88a40e9deacc/scene/recipe/S59sh2__bgfirst_bg.png",
  "asset_id": "d8f016ac-7892-423c-b067-811955d8cad7",
  "input_asset_ids": [
   "2e08c992-7e18-41ad-8a14-57af8366804d",
   "5fa31c1e-0d16-4a07-9d51-1572b715fb98"
  ]
 },
 "S59sh2": {
  "input_fingerprint": "669a086417e5fe25",
  "prompt": "Create ONE FINAL photorealistic live-action film still of the moment below — early 20th-century Korea, Jeolla region; all people are Korean unless explicitly stated otherwise. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 안규홍의 등 한가운데 난 깊은 상처에서 붉은 피가 흘러 저고리를 짙게 적신 클로즈업.\n\nLOCATION (lock): The steep wooded uphill path leading from the market side toward the mountain camp. The shot takes place here — the attached LOCATION PHOTOGRAPH shows the exact spot.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: close-up\n- KEY BACKGROUND ELEMENTS: Uphill forest path (Visible only at the margins of the climbing torso) — The path continues uphill beyond the subject; used as Provides a restrained spatial reference without competing with the wound.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Daylight keeps the red blood and darkened clothing legible with restrained saturation and naturalistic contrast.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The steep wooded approach rises above the distant market. 안규홍: He climbs alone carrying two bloodied sickles; blood from his deep back slash soaks his jeogori, and his hands remain blood-covered.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 안규홍 (한국인 남성, 청년 나이대, 검은 머리, 단단히 묶은 머리) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.",
  "roll_prompts": {
   "A": "Stage the shot. The FIRST attached image (SHOT BACKGROUND) is the finished empty background of this shot — keep it EXACTLY: its camera, perspective, architecture, lighting and every fixed feature stay untouched. The SECOND attached image (LAYOUT SKETCH) tells you ONLY where the people go: each sketched person's position, screen size, pose and the gaze/motion arrows. Ignore the sketch's background lines. The CHARACTER REFERENCE photographs show the real people.\nPlace the real people into the background at exactly the sketched positions, sizes and poses, following the arrow directions. No sketch lines or arrows may remain.\n\nCreate ONE FINAL photorealistic live-action film still of the moment below — early 20th-century Korea, Jeolla region; all people are Korean unless explicitly stated otherwise. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 안규홍의 등 한가운데 난 깊은 상처에서 붉은 피가 흘러 저고리를 짙게 적신 클로즈업.\n\nLOCATION (lock): The steep wooded uphill path leading from the market side toward the mountain camp. The shot takes place here — the FIRST attached image (SHOT BACKGROUND) is this exact place, already built: its ground, structures, horizon, materials and lighting are the finished truth of this location and must not be redesigned or replaced. No location photograph is attached — read the place from that image alone, and add no scenery, structure, vehicle or fixture that it does not already show. This lock governs the place only; the figures in the shot follow the staging and pose instructions.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: close-up\n- KEY BACKGROUND ELEMENTS: Uphill forest path (Visible only at the margins of the climbing torso) — The path continues uphill beyond the subject; used as Provides a restrained spatial reference without competing with the wound.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Daylight keeps the red blood and darkened clothing legible with restrained saturation and naturalistic contrast.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The steep wooded approach rises above the distant market. 안규홍: He climbs alone carrying two bloodied sickles; blood from his deep back slash soaks his jeogori, and his hands remain blood-covered.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 안규홍 (한국인 남성, 청년 나이대, 검은 머리, 단단히 묶은 머리) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.",
   "B": "Create ONE FINAL photorealistic live-action film still of the moment below — early 20th-century Korea, Jeolla region; all people are Korean unless explicitly stated otherwise. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 안규홍의 등 한가운데 난 깊은 상처에서 붉은 피가 흘러 저고리를 짙게 적신 클로즈업.\n\nLOCATION (lock): The steep wooded uphill path leading from the market side toward the mountain camp. The shot takes place here — the attached LOCATION PHOTOGRAPH shows the exact spot.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: close-up\n- KEY BACKGROUND ELEMENTS: Uphill forest path (Visible only at the margins of the climbing torso) — The path continues uphill beyond the subject; used as Provides a restrained spatial reference without competing with the wound.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Daylight keeps the red blood and darkened clothing legible with restrained saturation and naturalistic contrast.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The steep wooded approach rises above the distant market. 안규홍: He climbs alone carrying two bloodied sickles; blood from his deep back slash soaks his jeogori, and his hands remain blood-covered.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 안규홍 (한국인 남성, 청년 나이대, 검은 머리, 단단히 묶은 머리) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.\n\nCOMPOSITION VARIATION (this candidate only): keep the place, the people and the moment exactly as specified, and keep from the CAMERA & FRAME contract its subject scale, screen placement and key background elements exactly. For this candidate only, its camera angle and height are deliberately left open: choose a camera position distinctly different from the obvious one, as a film director picking a second setup on the same blocking. Stay true to any explicit framing the SHOT TEXT itself dictates."
  },
  "roll_refs": {
   "A": [
    {
     "label": "SHOT BACKGROUND",
     "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/db00517b-165b-4005-a287-f28936b59b95/images/272bd7af-9294-4490-98c6-88a40e9deacc/scene/recipe/S59sh2__bgfirst_bg.png",
     "asset_id": "d8f016ac-7892-423c-b067-811955d8cad7",
     "role": "bgfirst_bg"
    },
    {
     "label": "LAYOUT SKETCH (people placement only)",
     "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/db00517b-165b-4005-a287-f28936b59b95/images/272bd7af-9294-4490-98c6-88a40e9deacc/conti/conti_S59sh2.png",
     "asset_id": "2e08c992-7e18-41ad-8a14-57af8366804d",
     "role": "conti_light"
    },
    {
     "label": "CHARACTER REFERENCE — 안규홍: the exact person appearing in this shot; match face, hair and build exactly.",
     "path": "<bytes:1483645>",
     "asset_id": "b52dff85-b663-44f9-aa9c-46b828990f55",
     "role": "character_ref"
    }
   ],
   "B": [
    {
     "label": "LOCATION PHOTOGRAPH — the exact place of this shot: its architecture, materials, fixed features and lighting mood are spatial truth; stage the moment inside this place. Never copy its camera framing.",
     "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/db00517b-165b-4005-a287-f28936b59b95/images/272bd7af-9294-4490-98c6-88a40e9deacc/scene/recipe/groupbg_wounded_ascent_f8055c.png",
     "asset_id": "5fa31c1e-0d16-4a07-9d51-1572b715fb98",
     "role": "bgfirst_group_bg"
    },
    {
     "label": "CHARACTER REFERENCE — 안규홍: the exact person appearing in this shot; match face, hair and build exactly.",
     "path": "<bytes:1483645>",
     "asset_id": "b52dff85-b663-44f9-aa9c-46b828990f55",
     "role": "character_ref"
    }
   ]
  },
  "shot_run_spend_attempt_count": 1,
  "judge_flip_skipped": "judge_fn_owns_order",
  "readings": [
   {
    "label": "A",
    "direction": "카메라를 등지고 오르막길 위쪽을 향해 서 있음.",
    "built_space": "레퍼런스와 일치하는 숲속 흙길 오르막이 인물의 양옆과 너머로 보임.",
    "entities": "검은 머리를 묶은 남성, 피로 붉게 물든 찢어진 저고리, 등에 난 깊은 상처, 양손에 쥔 피 묻은 낫 2개, 피 묻은 손 모두 프롬프트와 일치함.",
    "hard_violations": [],
    "physics": "양손이 낫의 나무 손잡이를 단단히 쥐고 있으며, 발은 보이지 않으나 지면에 안정적으로 서 있는 자세임."
   },
   {
    "label": "B",
    "direction": "카메라를 등지고 오르막길 위쪽을 향해 서 있음.",
    "built_space": "레퍼런스와 일치하는 숲속 흙길 오르막이 인물의 우측과 너머로 보임.",
    "entities": "검은 머리를 묶은 남성, 흙먼지만 묻은 흰 저고리. 상처와 피, 낫은 존재하지 않음.",
    "hard_violations": [],
    "physics": "자연스럽게 서 있는 뒷모습이나 손에 들고 있어야 할 물건(낫)이 보이지 않음."
   }
  ],
  "cross_model_order": {
   "policy": "cross_model_order_v2_slot0_fwd_slot1_rev_parallel2",
   "slots": [
    {
     "model": "gemini-pro",
     "order": "forward",
     "display_to_canonical": {
      "A": "A",
      "B": "B"
     },
     "raw": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 10,
        "verdict_ko": "프롬프트가 요구한 등 한가운데의 깊은 상처, 피에 젖은 저고리, 양손에 든 피 묻은 낫과 피 묻은 손을 프레임 내에 완벽하고 사실적으로 구현했습니다."
       },
       {
        "label": "B",
        "score": 2,
        "verdict_ko": "프롬프트의 가장 핵심적인 지시인 '등 한가운데의 깊은 상처와 피에 젖은 저고리'가 완전히 누락되었으며, 낫을 든 상태도 확인되지 않습니다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "카메라를 등지고 오르막길 위쪽을 향해 서 있음.",
        "built_space": "레퍼런스와 일치하는 숲속 흙길 오르막이 인물의 양옆과 너머로 보임.",
        "entities": "검은 머리를 묶은 남성, 피로 붉게 물든 찢어진 저고리, 등에 난 깊은 상처, 양손에 쥔 피 묻은 낫 2개, 피 묻은 손 모두 프롬프트와 일치함.",
        "hard_violations": [],
        "physics": "양손이 낫의 나무 손잡이를 단단히 쥐고 있으며, 발은 보이지 않으나 지면에 안정적으로 서 있는 자세임."
       },
       {
        "label": "B",
        "direction": "카메라를 등지고 오르막길 위쪽을 향해 서 있음.",
        "built_space": "레퍼런스와 일치하는 숲속 흙길 오르막이 인물의 우측과 너머로 보임.",
        "entities": "검은 머리를 묶은 남성, 흙먼지만 묻은 흰 저고리. 상처와 피, 낫은 존재하지 않음.",
        "hard_violations": [],
        "physics": "자연스럽게 서 있는 뒷모습이나 손에 들고 있어야 할 물건(낫)이 보이지 않음."
       }
      ],
      "all_candidates_fail": false
     },
     "normalized": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 10,
        "verdict_ko": "프롬프트가 요구한 등 한가운데의 깊은 상처, 피에 젖은 저고리, 양손에 든 피 묻은 낫과 피 묻은 손을 프레임 내에 완벽하고 사실적으로 구현했습니다."
       },
       {
        "label": "B",
        "score": 2,
        "verdict_ko": "프롬프트의 가장 핵심적인 지시인 '등 한가운데의 깊은 상처와 피에 젖은 저고리'가 완전히 누락되었으며, 낫을 든 상태도 확인되지 않습니다."
       }
      ],
      "all_candidates_fail": false,
      "readings": [
       {
        "label": "A",
        "direction": "카메라를 등지고 오르막길 위쪽을 향해 서 있음.",
        "built_space": "레퍼런스와 일치하는 숲속 흙길 오르막이 인물의 양옆과 너머로 보임.",
        "entities": "검은 머리를 묶은 남성, 피로 붉게 물든 찢어진 저고리, 등에 난 깊은 상처, 양손에 쥔 피 묻은 낫 2개, 피 묻은 손 모두 프롬프트와 일치함.",
        "hard_violations": [],
        "physics": "양손이 낫의 나무 손잡이를 단단히 쥐고 있으며, 발은 보이지 않으나 지면에 안정적으로 서 있는 자세임."
       },
       {
        "label": "B",
        "direction": "카메라를 등지고 오르막길 위쪽을 향해 서 있음.",
        "built_space": "레퍼런스와 일치하는 숲속 흙길 오르막이 인물의 우측과 너머로 보임.",
        "entities": "검은 머리를 묶은 남성, 흙먼지만 묻은 흰 저고리. 상처와 피, 낫은 존재하지 않음.",
        "hard_violations": [],
        "physics": "자연스럽게 서 있는 뒷모습이나 손에 들고 있어야 할 물건(낫)이 보이지 않음."
       }
      ]
     },
     "ok": true
    },
    {
     "model": "gpt-high",
     "order": "reverse",
     "display_to_canonical": {
      "A": "B",
      "B": "A"
     },
     "raw": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 3,
        "verdict_ko": "등을 크게 잡은 클로즈업과 숲길은 맞지만, 핵심 순간인 등 중앙의 깊은 상처와 붉은 출혈, 피에 짙게 젖은 저고리가 전혀 보이지 않는다."
       },
       {
        "label": "B",
        "score": 8,
        "verdict_ko": "손과 낫까지 포함해 요구보다 다소 넓지만, 등 중앙의 깊은 상처에서 피가 흘러 저고리를 짙게 적시는 핵심 순간을 명확히 구현한다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "인물은 카메라에 등을 보이고 화면 오른쪽 위로 이어지는 오르막을 향해 상체를 숙인다. 얼굴과 눈이 가려져 정확한 시선은 확인할 수 없다. 손과 낫은 프레임 밖이므로 방향을 판단할 수 없다.",
        "built_space": "인공 구조물이나 고정 설비는 없다. 인물 오른쪽 가장자리에 흙과 뿌리가 드러난 산길 하나가 위쪽으로 이어지고, 왼쪽에는 나무 사이로 낮은 지대가 보인다. 참조의 오른쪽 오르막과 왼쪽 내리막 지형에 부합하며, 등과 어깨가 화면 대부분을 차지한다.",
        "entities": "한 사람의 뒷머리, 목, 등과 양팔이 보인다. 검은 머리를 묶은 모습과 때 묻은 미색 저고리는 참조에 부합한다. 얼굴이 보이지 않아 정확한 인물 동일성이나 한국인 청년이라는 설정은 직접 확인하기 어렵다. 등에는 회갈색 오염만 있고, 요구된 깊은 상처와 붉은 피, 짙은 혈액 침윤은 보이지 않는다. 피 묻은 손과 낫 두 자루는 구도 밖이므로 누락으로 벌점 주지 않는다. 추가 인물이나 읽을 수 있는 문자는 없다.",
        "hard_violations": [],
        "physics": "몸통은 앞으로 기울고 왼팔은 굽혀져 있어 오르막을 걷는 자세로 가능하다. 발과 지면 접촉은 클로즈업 밖이라 확인할 수 없지만, 몸이 공중에 떠 있다는 징후는 없다. 저고리는 어깨와 몸통에 걸려 자연스럽게 접힌다."
       },
       {
        "label": "B",
        "direction": "인물은 카메라 반대쪽, 오른쪽 위로 이어지는 산길을 향해 몸을 기울인다. 눈은 보이지 않는다. 양손의 낫은 허리 양옆으로 낮게 들려 있고 날은 위로 올라가 옆으로 굽으며, 특정 대상을 공격하거나 겨누지 않는 운반 상태다. 상처 아래의 핏줄기는 등에서 허리 쪽으로 내려간다.",
        "built_space": "인공 구조물이나 고정 설비는 없다. 오른쪽의 뿌리와 돌이 있는 흙길 하나가 인물 앞쪽으로 올라가며, 왼쪽 나무 사이에는 낮은 골짜기와 먼 마을이 보인다. 참조의 지형과 식생 배치에 부합한다. 다만 등뿐 아니라 허리와 양손까지 담아 상처 클로즈업 요구보다 넓고, 배경 면적도 더 크다.",
        "entities": "인물은 한 명이며, 묶은 검은 머리와 미색의 낡은 저고리가 참조와 부합한다. 얼굴은 가려져 정확한 얼굴 동일성은 검증할 수 없다. 등 중앙에는 길게 벌어진 깊은 상처와 붉은 피가 보이고, 주변 천은 짙은 적갈색으로 젖어 있다. 양손에는 피가 묻어 있으며, 각각 나무 손잡이와 굽은 금속 날을 가진 낫 한 자루씩, 총 두 자루를 쥐고 있다. 날에도 어두운 붉은 얼룩이 보인다. 추가 인물이나 읽을 수 있는 문자는 없다.",
        "hard_violations": [],
        "physics": "양손이 각각 낫 손잡이를 감싸 쥐고 있어 두 낫의 지지가 명확하다. 팔꿈치를 굽히고 상체를 앞으로 기울인 자세는 낫을 들고 오르는 동작으로 가능하다. 발은 프레임 밖이지만 부유나 비정상적인 신체 배치는 보이지 않는다. 피는 상처에서 아래로 흐르고 천으로 번지며, 저고리와 허리의 천 띠는 몸에 밀착해 주름진다."
       }
      ],
      "all_candidates_fail": false
     },
     "normalized": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 3,
        "verdict_ko": "등을 크게 잡은 클로즈업과 숲길은 맞지만, 핵심 순간인 등 중앙의 깊은 상처와 붉은 출혈, 피에 짙게 젖은 저고리가 전혀 보이지 않는다."
       },
       {
        "label": "A",
        "score": 8,
        "verdict_ko": "손과 낫까지 포함해 요구보다 다소 넓지만, 등 중앙의 깊은 상처에서 피가 흘러 저고리를 짙게 적시는 핵심 순간을 명확히 구현한다."
       }
      ],
      "all_candidates_fail": false,
      "readings": [
       {
        "label": "B",
        "direction": "인물은 카메라에 등을 보이고 화면 오른쪽 위로 이어지는 오르막을 향해 상체를 숙인다. 얼굴과 눈이 가려져 정확한 시선은 확인할 수 없다. 손과 낫은 프레임 밖이므로 방향을 판단할 수 없다.",
        "built_space": "인공 구조물이나 고정 설비는 없다. 인물 오른쪽 가장자리에 흙과 뿌리가 드러난 산길 하나가 위쪽으로 이어지고, 왼쪽에는 나무 사이로 낮은 지대가 보인다. 참조의 오른쪽 오르막과 왼쪽 내리막 지형에 부합하며, 등과 어깨가 화면 대부분을 차지한다.",
        "entities": "한 사람의 뒷머리, 목, 등과 양팔이 보인다. 검은 머리를 묶은 모습과 때 묻은 미색 저고리는 참조에 부합한다. 얼굴이 보이지 않아 정확한 인물 동일성이나 한국인 청년이라는 설정은 직접 확인하기 어렵다. 등에는 회갈색 오염만 있고, 요구된 깊은 상처와 붉은 피, 짙은 혈액 침윤은 보이지 않는다. 피 묻은 손과 낫 두 자루는 구도 밖이므로 누락으로 벌점 주지 않는다. 추가 인물이나 읽을 수 있는 문자는 없다.",
        "hard_violations": [],
        "physics": "몸통은 앞으로 기울고 왼팔은 굽혀져 있어 오르막을 걷는 자세로 가능하다. 발과 지면 접촉은 클로즈업 밖이라 확인할 수 없지만, 몸이 공중에 떠 있다는 징후는 없다. 저고리는 어깨와 몸통에 걸려 자연스럽게 접힌다."
       },
       {
        "label": "A",
        "direction": "인물은 카메라 반대쪽, 오른쪽 위로 이어지는 산길을 향해 몸을 기울인다. 눈은 보이지 않는다. 양손의 낫은 허리 양옆으로 낮게 들려 있고 날은 위로 올라가 옆으로 굽으며, 특정 대상을 공격하거나 겨누지 않는 운반 상태다. 상처 아래의 핏줄기는 등에서 허리 쪽으로 내려간다.",
        "built_space": "인공 구조물이나 고정 설비는 없다. 오른쪽의 뿌리와 돌이 있는 흙길 하나가 인물 앞쪽으로 올라가며, 왼쪽 나무 사이에는 낮은 골짜기와 먼 마을이 보인다. 참조의 지형과 식생 배치에 부합한다. 다만 등뿐 아니라 허리와 양손까지 담아 상처 클로즈업 요구보다 넓고, 배경 면적도 더 크다.",
        "entities": "인물은 한 명이며, 묶은 검은 머리와 미색의 낡은 저고리가 참조와 부합한다. 얼굴은 가려져 정확한 얼굴 동일성은 검증할 수 없다. 등 중앙에는 길게 벌어진 깊은 상처와 붉은 피가 보이고, 주변 천은 짙은 적갈색으로 젖어 있다. 양손에는 피가 묻어 있으며, 각각 나무 손잡이와 굽은 금속 날을 가진 낫 한 자루씩, 총 두 자루를 쥐고 있다. 날에도 어두운 붉은 얼룩이 보인다. 추가 인물이나 읽을 수 있는 문자는 없다.",
        "hard_violations": [],
        "physics": "양손이 각각 낫 손잡이를 감싸 쥐고 있어 두 낫의 지지가 명확하다. 팔꿈치를 굽히고 상체를 앞으로 기울인 자세는 낫을 들고 오르는 동작으로 가능하다. 발은 프레임 밖이지만 부유나 비정상적인 신체 배치는 보이지 않는다. 피는 상처에서 아래로 흐르고 천으로 번지며, 저고리와 허리의 천 띠는 몸에 밀착해 주름진다."
       }
      ]
     },
     "ok": true
    }
   ],
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "slot_winner_match": true,
   "slot_winner": {
    "gemini-pro": "A",
    "gpt-high": "A"
   },
   "route": "cross_slot_agree"
  },
  "dual": {
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "normalized": {
    "A": 2.0,
    "B": 0.575
   },
   "adjusted": {
    "A": 2.0,
    "B": 0.575
   },
   "violations": {},
   "per_model_winner": {
    "gemini-pro": "A",
    "gpt-high": "A"
   },
   "agreed": true
  },
  "totals": {
   "A": 2000,
   "B": 575
  },
  "selected": "A",
  "ranking": [
   "A",
   "B"
  ],
  "verdicts": [
   {
    "label": "A",
    "score": 2000,
    "verdict_ko": "프롬프트가 요구한 등 한가운데의 깊은 상처, 피에 젖은 저고리, 양손에 든 피 묻은 낫과 피 묻은 손을 프레임 내에 완벽하고 사실적으로 구현했습니다."
   },
   {
    "label": "B",
    "score": 575,
    "verdict_ko": "프롬프트의 가장 핵심적인 지시인 '등 한가운데의 깊은 상처와 피에 젖은 저고리'가 완전히 누락되었으며, 낫을 든 상태도 확인되지 않습니다."
   }
  ],
  "refs": [
   {
    "label": "LOCATION PHOTOGRAPH — the exact place of this shot: its architecture, materials, fixed features and lighting mood are spatial truth; stage the moment inside this place. Never copy its camera framing.",
    "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/db00517b-165b-4005-a287-f28936b59b95/images/272bd7af-9294-4490-98c6-88a40e9deacc/scene/recipe/groupbg_wounded_ascent_f8055c.png",
    "asset_id": "5fa31c1e-0d16-4a07-9d51-1572b715fb98",
    "role": "bgfirst_group_bg"
   },
   {
    "label": "CHARACTER REFERENCE — 안규홍: the exact person appearing in this shot; match face, hair and build exactly.",
    "path": "<bytes:1483645>",
    "asset_id": "b52dff85-b663-44f9-aa9c-46b828990f55",
    "role": "character_ref"
   }
  ],
  "critique_skipped": true,
  "shot_run_produced": false,
  "shot_run_uid": "06aa1438-f13e-761c-a059-7aa080932e02",
  "bgfirst": {
   "bg_path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/db00517b-165b-4005-a287-f28936b59b95/images/272bd7af-9294-4490-98c6-88a40e9deacc/scene/recipe/S59sh2__bgfirst_bg.png",
   "bg_asset_id": "d8f016ac-7892-423c-b067-811955d8cad7",
   "bg_record_key": "S59sh2::bgfirst_bg",
   "chain_winner": true,
   "authority": "groupbg",
   "group_key": "wounded_ascent",
   "groupbg_asset_id": "5fa31c1e-0d16-4a07-9d51-1572b715fb98"
  },
  "ref_mode": "재투영 배경+콘티+엔티티 (2택1: 체인 승)",
  "share_plan": {
   "ref_plan": "background"
  }
 },
 "S59sh2::cine": {
  "applied": false,
  "attempted_at": "2026-09-09T09:33:34.337005+00:00",
  "fingerprint": "2fbff212cc28680b03684504d6f6f6e1a2ba9cd0b9e4768dcec0d741e0951921",
  "fingerprint_version": 2,
  "provider": "mai",
  "endpoint": "openrouter/chat-completions",
  "model": "microsoft/mai-image-2.6",
  "multiroll_tag": "still_S59sh2_cine",
  "slot": "",
  "pack": "24.202608252115",
  "source_file": "S59sh2_sel.png",
  "source_sha256": "c05391a9e15768a293b2998aa2df14dd38aefa6d9a894ef717eefef8ed324047",
  "error": "RuntimeError: moderation blocked: {\"message\": \"Output content violated imagegen safety policies.\", \"code\": 400}",
  "moderation_refusals": 2,
  "declined": true,
  "declined_reason": "moderation"
 },
 "S59sh9::signage": {
  "fp": "d10b2e4db22ccf8c",
  "inscriptions": [],
  "cues": [],
  "dropped": []
 },
 "S59sh9": {
  "input_fingerprint": "5afe2c75d7f028b8",
  "prompt": "Create ONE FINAL photorealistic live-action film still of the moment below — early 20th-century Korea, Jeolla region; all people are Korean unless explicitly stated otherwise. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 이를 꽉 악물고 흙바닥을 짚어 몸을 반쯤 일으켜 세운 도중의 역동적인 안규홍의 전신.\n\nLOCATION (lock): The loose dirt surface of a steep section of the mountain camp's forest approach, where a traveler has slipped. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: wide shot\n- KEY BACKGROUND ELEMENTS: Sloping earth path (Beneath his supporting hands and feet) — The incline crosses the composition obliquely; used as Makes the difficulty of regaining his footing visible; Forest trees (Present along the path); used as Retain the forest setting behind the full-body action.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Consistent daylight and sober naturalistic contrast preserve the effort in his face without glamorizing the injury.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The clothing, hair and overall look of 안규홍 — who appear both in that photo and in this shot — are LOCKED to that photo. Anyone else visible in that photo is NOT in this shot: never carry their face, body or clothing onto anyone here. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The steep forest slope remains dry dirt beneath the trees. 안규홍: He pushes up after falling face-first, newly coated in dirt over the blood; his back wound still bleeds, and both sickles remain retained.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 안규홍 (한국인 남성, 청년 나이대, 검은 머리, 단단히 묶은 머리) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.",
  "roll_prompts": {
   "A": "Create ONE FINAL photorealistic live-action film still of the moment below — early 20th-century Korea, Jeolla region; all people are Korean unless explicitly stated otherwise. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 이를 꽉 악물고 흙바닥을 짚어 몸을 반쯤 일으켜 세운 도중의 역동적인 안규홍의 전신.\n\nLOCATION (lock): The loose dirt surface of a steep section of the mountain camp's forest approach, where a traveler has slipped. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: wide shot\n- KEY BACKGROUND ELEMENTS: Sloping earth path (Beneath his supporting hands and feet) — The incline crosses the composition obliquely; used as Makes the difficulty of regaining his footing visible; Forest trees (Present along the path); used as Retain the forest setting behind the full-body action.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Consistent daylight and sober naturalistic contrast preserve the effort in his face without glamorizing the injury.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The clothing, hair and overall look of 안규홍 — who appear both in that photo and in this shot — are LOCKED to that photo. Anyone else visible in that photo is NOT in this shot: never carry their face, body or clothing onto anyone here. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The steep forest slope remains dry dirt beneath the trees. 안규홍: He pushes up after falling face-first, newly coated in dirt over the blood; his back wound still bleeds, and both sickles remain retained.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 안규홍 (한국인 남성, 청년 나이대, 검은 머리, 단단히 묶은 머리) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.",
   "B": "Create ONE FINAL photorealistic live-action film still of the moment below — early 20th-century Korea, Jeolla region; all people are Korean unless explicitly stated otherwise. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 이를 꽉 악물고 흙바닥을 짚어 몸을 반쯤 일으켜 세운 도중의 역동적인 안규홍의 전신.\n\nLOCATION (lock): The loose dirt surface of a steep section of the mountain camp's forest approach, where a traveler has slipped. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: wide shot\n- KEY BACKGROUND ELEMENTS: Sloping earth path (Beneath his supporting hands and feet) — The incline crosses the composition obliquely; used as Makes the difficulty of regaining his footing visible; Forest trees (Present along the path); used as Retain the forest setting behind the full-body action.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Consistent daylight and sober naturalistic contrast preserve the effort in his face without glamorizing the injury.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The clothing, hair and overall look of 안규홍 — who appear both in that photo and in this shot — are LOCKED to that photo. Anyone else visible in that photo is NOT in this shot: never carry their face, body or clothing onto anyone here. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The steep forest slope remains dry dirt beneath the trees. 안규홍: He pushes up after falling face-first, newly coated in dirt over the blood; his back wound still bleeds, and both sickles remain retained.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 안규홍 (한국인 남성, 청년 나이대, 검은 머리, 단단히 묶은 머리) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.\n\nCOMPOSITION VARIATION (this candidate only): keep the place, the people and the moment exactly as specified, and keep from the CAMERA & FRAME contract its subject scale, screen placement and key background elements exactly. For this candidate only, its camera angle and height are deliberately left open: choose a camera position distinctly different from the obvious one, as a film director picking a second setup on the same blocking. Stay true to any explicit framing the SHOT TEXT itself dictates."
  },
  "shot_run_spend_attempt_count": 1,
  "readings": [
   {
    "label": "A",
    "direction": "시선은 앞쪽 바닥을 향하며 이를 꽉 악물고 목표한 표정을 보여줌.",
    "built_space": "가파른 경사의 흙길과 나무가 있는 숲 배경이 이전 샷 레퍼런스와 자연스럽게 연결됨.",
    "entities": "안규홍의 외모, 묶은 머리, 흙과 피가 묻은 의상이 레퍼런스와 일치함.",
    "hard_violations": [
     "[gemini-pro] 낫의 칼날과 자루가 복부와 등을 관통하여 물리적으로 불가능하게 튀어나와 있음 (physically impossible staging)"
    ],
    "physics": "오른손으로 낫을 쥐고 왼손으로 땅을 짚고 있으나, 몸통 부근에서 정체불명의 낫 부품들이 신체를 뚫고 나온 채 지지점 없이 존재함."
   },
   {
    "label": "B",
    "direction": "시선은 앞쪽 아래를 향하고 있으며 이를 꽉 악물고 힘을 쓰는 표정을 보여줌.",
    "built_space": "경사진 흙길과 주변의 나무들이 레퍼런스의 환경과 완벽히 일치하게 묘사됨.",
    "entities": "안규홍의 외모와 의상, 특히 등에 베인 상처와 피가 흐르는 묘사가 레퍼런스의 특징을 정확히 반영함.",
    "hard_violations": [],
    "physics": "양손바닥과 발로 단단히 흙바닥을 짚고 체중을 지탱하며 몸을 반쯤 일으키는 동작이 매우 자연스럽고 사실적임."
   }
  ],
  "cross_model_order": {
   "policy": "cross_model_order_v2_slot0_fwd_slot1_rev_parallel2",
   "slots": [
    {
     "model": "gemini-pro",
     "order": "forward",
     "display_to_canonical": {
      "A": "A",
      "B": "B"
     },
     "raw": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 8,
        "verdict_ko": "두 개의 낫을 유지해야 한다는 지시는 누락되었지만, 신체를 관통하는 형태적 오류 없이 흙바닥을 짚고 반쯤 일어나는 안규홍의 역동적인 전신과 피 묻은 등 상처를 사실적으로 잘 묘사하여 우위를 점함."
       },
       {
        "label": "A",
        "score": 2,
        "verdict_ko": "오른손에 낫을 쥐고 있으나 배와 등 부위에서 낫의 칼날과 자루가 몸을 관통하여 튀어나오는 물리적으로 불가능한 하드 위반(Hard Violation) 오류가 발생하여 탈락함."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "시선은 앞쪽 바닥을 향하며 이를 꽉 악물고 목표한 표정을 보여줌.",
        "built_space": "가파른 경사의 흙길과 나무가 있는 숲 배경이 이전 샷 레퍼런스와 자연스럽게 연결됨.",
        "entities": "안규홍의 외모, 묶은 머리, 흙과 피가 묻은 의상이 레퍼런스와 일치함.",
        "hard_violations": [
         "낫의 칼날과 자루가 복부와 등을 관통하여 물리적으로 불가능하게 튀어나와 있음 (physically impossible staging)"
        ],
        "physics": "오른손으로 낫을 쥐고 왼손으로 땅을 짚고 있으나, 몸통 부근에서 정체불명의 낫 부품들이 신체를 뚫고 나온 채 지지점 없이 존재함."
       },
       {
        "label": "B",
        "direction": "시선은 앞쪽 아래를 향하고 있으며 이를 꽉 악물고 힘을 쓰는 표정을 보여줌.",
        "built_space": "경사진 흙길과 주변의 나무들이 레퍼런스의 환경과 완벽히 일치하게 묘사됨.",
        "entities": "안규홍의 외모와 의상, 특히 등에 베인 상처와 피가 흐르는 묘사가 레퍼런스의 특징을 정확히 반영함.",
        "hard_violations": [],
        "physics": "양손바닥과 발로 단단히 흙바닥을 짚고 체중을 지탱하며 몸을 반쯤 일으키는 동작이 매우 자연스럽고 사실적임."
       }
      ],
      "all_candidates_fail": false
     },
     "normalized": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 8,
        "verdict_ko": "두 개의 낫을 유지해야 한다는 지시는 누락되었지만, 신체를 관통하는 형태적 오류 없이 흙바닥을 짚고 반쯤 일어나는 안규홍의 역동적인 전신과 피 묻은 등 상처를 사실적으로 잘 묘사하여 우위를 점함."
       },
       {
        "label": "A",
        "score": 2,
        "verdict_ko": "오른손에 낫을 쥐고 있으나 배와 등 부위에서 낫의 칼날과 자루가 몸을 관통하여 튀어나오는 물리적으로 불가능한 하드 위반(Hard Violation) 오류가 발생하여 탈락함."
       }
      ],
      "all_candidates_fail": false,
      "readings": [
       {
        "label": "A",
        "direction": "시선은 앞쪽 바닥을 향하며 이를 꽉 악물고 목표한 표정을 보여줌.",
        "built_space": "가파른 경사의 흙길과 나무가 있는 숲 배경이 이전 샷 레퍼런스와 자연스럽게 연결됨.",
        "entities": "안규홍의 외모, 묶은 머리, 흙과 피가 묻은 의상이 레퍼런스와 일치함.",
        "hard_violations": [
         "낫의 칼날과 자루가 복부와 등을 관통하여 물리적으로 불가능하게 튀어나와 있음 (physically impossible staging)"
        ],
        "physics": "오른손으로 낫을 쥐고 왼손으로 땅을 짚고 있으나, 몸통 부근에서 정체불명의 낫 부품들이 신체를 뚫고 나온 채 지지점 없이 존재함."
       },
       {
        "label": "B",
        "direction": "시선은 앞쪽 아래를 향하고 있으며 이를 꽉 악물고 힘을 쓰는 표정을 보여줌.",
        "built_space": "경사진 흙길과 주변의 나무들이 레퍼런스의 환경과 완벽히 일치하게 묘사됨.",
        "entities": "안규홍의 외모와 의상, 특히 등에 베인 상처와 피가 흐르는 묘사가 레퍼런스의 특징을 정확히 반영함.",
        "hard_violations": [],
        "physics": "양손바닥과 발로 단단히 흙바닥을 짚고 체중을 지탱하며 몸을 반쯤 일으키는 동작이 매우 자연스럽고 사실적임."
       }
      ]
     },
     "ok": true
    },
    {
     "model": "gpt-high",
     "order": "reverse",
     "display_to_canonical": {
      "A": "B",
      "B": "A"
     },
     "raw": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 8,
        "verdict_ko": "두 낫의 유지 상태는 확인되지 않지만, 전신을 온전히 담고 이를 악문 채 흙을 짚어 반쯤 일어나는 순간을 보여 주어 최우선 촬영 지시에 더 충실하다."
       },
       {
        "label": "B",
        "score": 7,
        "verdict_ko": "낫 두 자루와 참고 장소의 연속성은 더 분명하지만, 아래쪽 발이 프레임에 잘려 명시된 전신 구도에서는 A보다 부족하다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "얼굴과 시선은 화면 오른쪽 위의 오르막 흙면을 향한다. 양손은 그 경사면을 짚고 몸통은 앞쪽으로 들려 있어, 넘어진 뒤 오르막에서 몸을 일으키려는 방향과 맞는다. 확인 가능한 낫날은 없다.",
        "built_space": "인공 구조물이나 고정 설비는 없다. 마른 흙과 잔돌의 경사면이 왼쪽 아래에서 오른쪽 위로 비스듬히 가로지르고, 뒤편 왼쪽에는 숲길이 이어진다. 길 양쪽에 나무와 드러난 뿌리가 있다. 참고 사진의 숲과 흙길 재질은 유지되지만 개별 나무의 정확한 대응은 확인하기 어렵다.",
        "entities": "인물은 한 명이며, 참고 인물과 유사한 젊은 한국인 남성의 외형, 묶은 검은 머리, 얼굴 상처, 때 묻은 미색 옷과 허리끈이 보인다. 이를 드러내며 악문 표정이다. 등 위쪽에 혈흔이 보이고 손과 소매에도 피가 묻어 있다. 머리부터 두 발까지 프레임 안에 들어온다. 양손은 비어 있으며 낫 두 자루가 계속 몸에 보유되어 있는지는 확인되지 않는다. 읽을 수 있는 글자는 없다.",
        "hard_violations": [],
        "physics": "양 손바닥이 흙면에 닿고, 화면 왼쪽으로 뻗은 다리의 발과 몸 아래 접힌 다리의 무릎·발이 지면을 받친다. 팔로 상체를 밀어 올리면서 하체를 몸 아래로 끌어오는 자세가 가능하다. 몸이 공중에 떠 있거나 지지 없이 놓인 물체는 보이지 않는다."
       },
       {
        "label": "B",
        "direction": "고개와 시선은 화면 오른쪽 아래, 지지 손 주변의 흙바닥을 향한다. 몸통은 그 손을 축으로 위로 들리는 방향이다. 손에 든 낫날은 아래로 향하며 사람을 겨누지 않는다. 허리 쪽 다른 낫날도 아래로 굽어 있다.",
        "built_space": "인공 구조물이나 고정 설비는 없다. 흙 경사면은 왼쪽 아래에서 오른쪽 위로 올라가고, 뒤편 오른쪽에 뿌리가 드러난 숲길이 이어진다. 왼쪽 나무 사이로 계곡과 먼 취락이 보여 참고 사진의 지형적 배치가 A보다 분명하게 이어진다. 낮의 자연광과 건조한 흙 표면도 맞는다.",
        "entities": "인물은 한 명이며, 젊은 한국인 남성의 외형, 묶은 검은 머리, 상처 난 얼굴과 미색 전통 복장이 참고 인물과 대체로 일치한다. 이를 악문 고통스러운 표정과 옷의 흙먼지·혈흔이 보인다. 낫 한 자루는 손에 들려 있고 다른 한 자루는 허리 뒤에 끼워져 보인다. 등 상처는 카메라 반대편이라 판단할 수 없다. 화면 왼쪽 아래의 발 끝이 잘려 전신이 완전히 들어오지는 않는다. 읽을 수 있는 글자는 없다.",
        "hard_violations": [],
        "physics": "화면 오른쪽 손바닥이 경사면을 짚고, 뻗은 다리의 발과 접힌 다리의 무릎·발이 하중을 나누어 받는다. 한 손으로 몸을 밀어 올리며 다른 손에 낫을 유지하는 동작은 가능하다. 손의 낫은 손잡이를 쥔 손이 지지하고, 다른 낫은 허리끈에 걸쳐 지지되는 것으로 보인다. 지지 없이 떠 있는 신체나 물체는 없다."
       }
      ],
      "all_candidates_fail": false
     },
     "normalized": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 8,
        "verdict_ko": "두 낫의 유지 상태는 확인되지 않지만, 전신을 온전히 담고 이를 악문 채 흙을 짚어 반쯤 일어나는 순간을 보여 주어 최우선 촬영 지시에 더 충실하다."
       },
       {
        "label": "A",
        "score": 7,
        "verdict_ko": "낫 두 자루와 참고 장소의 연속성은 더 분명하지만, 아래쪽 발이 프레임에 잘려 명시된 전신 구도에서는 A보다 부족하다."
       }
      ],
      "all_candidates_fail": false,
      "readings": [
       {
        "label": "B",
        "direction": "얼굴과 시선은 화면 오른쪽 위의 오르막 흙면을 향한다. 양손은 그 경사면을 짚고 몸통은 앞쪽으로 들려 있어, 넘어진 뒤 오르막에서 몸을 일으키려는 방향과 맞는다. 확인 가능한 낫날은 없다.",
        "built_space": "인공 구조물이나 고정 설비는 없다. 마른 흙과 잔돌의 경사면이 왼쪽 아래에서 오른쪽 위로 비스듬히 가로지르고, 뒤편 왼쪽에는 숲길이 이어진다. 길 양쪽에 나무와 드러난 뿌리가 있다. 참고 사진의 숲과 흙길 재질은 유지되지만 개별 나무의 정확한 대응은 확인하기 어렵다.",
        "entities": "인물은 한 명이며, 참고 인물과 유사한 젊은 한국인 남성의 외형, 묶은 검은 머리, 얼굴 상처, 때 묻은 미색 옷과 허리끈이 보인다. 이를 드러내며 악문 표정이다. 등 위쪽에 혈흔이 보이고 손과 소매에도 피가 묻어 있다. 머리부터 두 발까지 프레임 안에 들어온다. 양손은 비어 있으며 낫 두 자루가 계속 몸에 보유되어 있는지는 확인되지 않는다. 읽을 수 있는 글자는 없다.",
        "hard_violations": [],
        "physics": "양 손바닥이 흙면에 닿고, 화면 왼쪽으로 뻗은 다리의 발과 몸 아래 접힌 다리의 무릎·발이 지면을 받친다. 팔로 상체를 밀어 올리면서 하체를 몸 아래로 끌어오는 자세가 가능하다. 몸이 공중에 떠 있거나 지지 없이 놓인 물체는 보이지 않는다."
       },
       {
        "label": "A",
        "direction": "고개와 시선은 화면 오른쪽 아래, 지지 손 주변의 흙바닥을 향한다. 몸통은 그 손을 축으로 위로 들리는 방향이다. 손에 든 낫날은 아래로 향하며 사람을 겨누지 않는다. 허리 쪽 다른 낫날도 아래로 굽어 있다.",
        "built_space": "인공 구조물이나 고정 설비는 없다. 흙 경사면은 왼쪽 아래에서 오른쪽 위로 올라가고, 뒤편 오른쪽에 뿌리가 드러난 숲길이 이어진다. 왼쪽 나무 사이로 계곡과 먼 취락이 보여 참고 사진의 지형적 배치가 A보다 분명하게 이어진다. 낮의 자연광과 건조한 흙 표면도 맞는다.",
        "entities": "인물은 한 명이며, 젊은 한국인 남성의 외형, 묶은 검은 머리, 상처 난 얼굴과 미색 전통 복장이 참고 인물과 대체로 일치한다. 이를 악문 고통스러운 표정과 옷의 흙먼지·혈흔이 보인다. 낫 한 자루는 손에 들려 있고 다른 한 자루는 허리 뒤에 끼워져 보인다. 등 상처는 카메라 반대편이라 판단할 수 없다. 화면 왼쪽 아래의 발 끝이 잘려 전신이 완전히 들어오지는 않는다. 읽을 수 있는 글자는 없다.",
        "hard_violations": [],
        "physics": "화면 오른쪽 손바닥이 경사면을 짚고, 뻗은 다리의 발과 접힌 다리의 무릎·발이 하중을 나누어 받는다. 한 손으로 몸을 밀어 올리며 다른 손에 낫을 유지하는 동작은 가능하다. 손의 낫은 손잡이를 쥔 손이 지지하고, 다른 낫은 허리끈에 걸쳐 지지되는 것으로 보인다. 지지 없이 떠 있는 신체나 물체는 없다."
       }
      ]
     },
     "ok": true
    }
   ],
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "slot_winner_match": true,
   "slot_winner": {
    "gemini-pro": "B",
    "gpt-high": "B"
   },
   "route": "cross_slot_agree"
  },
  "dual": {
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "normalized": {
    "A": 1.125,
    "B": 2.0
   },
   "adjusted": {
    "A": 0.875,
    "B": 2.0
   },
   "violations": {
    "A": [
     "[gemini-pro] 낫의 칼날과 자루가 복부와 등을 관통하여 물리적으로 불가능하게 튀어나와 있음 (physically impossible staging)"
    ]
   },
   "per_model_winner": {
    "gemini-pro": "B",
    "gpt-high": "B"
   },
   "agreed": true
  },
  "totals": {
   "B": 2000,
   "A": 875
  },
  "selected": "B",
  "ranking": [
   "B",
   "A"
  ],
  "verdicts": [
   {
    "label": "B",
    "score": 2000,
    "verdict_ko": "두 개의 낫을 유지해야 한다는 지시는 누락되었지만, 신체를 관통하는 형태적 오류 없이 흙바닥을 짚고 반쯤 일어나는 안규홍의 역동적인 전신과 피 묻은 등 상처를 사실적으로 잘 묘사하여 우위를 점함."
   },
   {
    "label": "A",
    "score": 875,
    "verdict_ko": "오른손에 낫을 쥐고 있으나 배와 등 부위에서 낫의 칼날과 자루가 몸을 관통하여 튀어나오는 물리적으로 불가능한 하드 위반(Hard Violation) 오류가 발생하여 탈락함.  ★위반: [gemini-pro] 낫의 칼날과 자루가 복부와 등을 관통하여 물리적으로 불가능하게 튀어나와 있음 (physically impossible staging)"
   }
  ],
  "refs": [
   {
    "label": "PREVIOUS SHOT STILL — a visually related earlier shot of this same place: the location's look, materials, fixed features and lighting mood are LOCKED to this photo; never copy its camera framing. The clothing, hair and overall look of 안규홍 — who appear both in that photo and in this shot — are LOCKED to that photo. Anyone else visible in that photo is NOT in this shot: never carry their face, body or clothing onto anyone here. If this photo shows a character who cannot move (dead or unconscious), that character's exact pose, position and orientation are ALSO LOCKED — treat their body as a fixed prop of the set that only the camera moves around.",
    "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/db00517b-165b-4005-a287-f28936b59b95/images/272bd7af-9294-4490-98c6-88a40e9deacc/scene/recipe/S59sh2_sel.png",
    "asset_id": "48db48bd-6ffc-4b6a-9c7f-4f90e1b756ac",
    "role": "prev_still"
   },
   {
    "label": "CHARACTER REFERENCE — 안규홍: the exact person appearing in this shot; match face, hair and build exactly.",
    "path": "<bytes:1483645>",
    "asset_id": "b52dff85-b663-44f9-aa9c-46b828990f55",
    "role": "character_ref"
   }
  ],
  "critique_skipped": true,
  "shot_run_produced": false,
  "shot_run_uid": "06aa1441-21e2-775a-9b12-6c161b205528",
  "ref_mode": "prev+엔티티",
  "share_plan": {
   "ref_plan": "prev",
   "prev_anchor_tag": "S59sh2"
  }
 },
 "S59sh9::cine": {
  "applied": true,
  "attempted_at": "2026-09-09T09:34:19.579460+00:00",
  "fingerprint": "63274e1931c79aff2abbe96a5d7450da484b78c6057b35136e71e70cfd6b7839",
  "fingerprint_version": 2,
  "provider": "mai",
  "endpoint": "openrouter/chat-completions",
  "model": "microsoft/mai-image-2.6",
  "multiroll_tag": "still_S59sh9_cine",
  "slot": "",
  "pack": "24.202608252115",
  "source_file": "S59sh9_sel.png",
  "source_sha256": "9fd6c1c9ac44890e5c893bdbf6293b5518413c2b5ae61ba86ae85094e81a295a",
  "file": "S59sh9_cine.png",
  "staged_sha256": "12581315a9ab769e4b6b15287fd0d4d90d6a213cb23f4c02872bcd2c7bcbc9ce",
  "latency_ms": 28432
 },
 "S59sh11::signage": {
  "fp": "29d6d4277f82ad66",
  "inscriptions": [],
  "cues": [],
  "dropped": []
 },
 "groupbg::opposite_descent": {
  "input_fingerprint": "c13251927d32a7f6",
  "meta": {
   "model": "gpt-image-2",
   "size": "1536x864",
   "pack": "11.202607220237",
   "contract": "bgfirst_full_v3",
   "group_sig": {
    "key": "opposite_descent",
    "tags": [
     "S59sh11",
     "S62sh12",
     "S62sh19",
     "S62sh22"
    ]
   },
   "context_sig": "9e5d0ee4fe052c92"
  },
  "prompt": "Create ONE empty live-action location background photograph — NO PEOPLE, no figures, no body parts, no silhouettes, no shadows or reflections of people anywhere.\n\"Empty\" means no people only: KEEP the location's inherent occupants and stock that define the place — animals in an animal shelter, pen or farm, goods and produce in a market, moored boats in a harbour — unless the shot text explicitly removes them.\nThe FIRST attached image is a thin-line storyboard sketch of a shot that happens at this location: use it ONLY as spatial evidence — what this place contains, how its ground, structures and landmarks are arranged and proportioned. Ignore the sketched people and arrows entirely, and do NOT copy its line style: render a fully photographic, physically plausible real place that fits THE LOCATION text below.\nHUMAN-SCALE CALIBRATION: derive every structure's true size from human-scale elements — a door ≈ 2m, a window ≈ 1–1.5m wide, one storey ≈ 2.5–3m; never inflate a small structure or shrink a large one.\n\nTHE LOCATION — early 20th-century Korea, Jeolla region; all people are Korean unless explicitly stated otherwise: The separate forest descent on the opposite side of the mountain camp, away from the approaching lone climber.\n\nLOCATION DETAIL (descriptive evidence about this place from the production's location records — evidence, not a staging order): use it to understand what kind of place this is and what it permanently contains. Carry over ONLY its enduring physical features — layout, structures, ground and materials, fixed equipment and installations. Do NOT treat any lighting, weather or time-of-day wording, momentary object states, depicted events or subjective impressions in it as fixed facts of the place; the TIME OF DAY lock below is the sole authority for time and lighting.\n파청 산채 진입 오르막 숲길: 산채로 연결되는 경사가 가파른 흙비탈 숲길. 나무 뿌리와 잡목이 길가로 돌출되어 있다. (특징: 경사가 심한 흙바닥 비탈; 우거진 수목과 돌출된 나무 뿌리; 군화나 짚신 자국이 남는 흙길) / 파청 산채 하산 숲길·비탈: 산채에서 하부 장터 쪽으로 길게 뻗어 내리는 험한 내리막 숲길. 나무 기둥 사이로 아득히 저잣거리의 풍경이 부분적으로 보인다. (특징: 가파른 내리막 흙길 비탈; 시야를 가리는 나무 기둥들; 멀리 아득하게 보이는 마을/장터 전경)\n\nSCENE EVIDENCE (verbatim quotes from the screenplay about this place — treat them as evidence of what the location physically contains and looks like; stage the PLACE those moments happen in, but do NOT depict the momentary actions, people or staged props themselves):\n- 반대편 산길.\n토미이시를 비롯한 일본군 일부가 하산하고 있다. 두 길은 서로 어긋난다.\n- S#62. 파청 산채 하산 숲길 / 낮 - 계속\n- 하산하던 토미이시가 걸음을 멈춘다.\n- 일본군들이 몸을 돌려 다시 산을 오른다.\n- 몇 걸음 오르지 않아—\n\nTIME OF DAY (lock): day.\n\nRender ONE photorealistic empty location photograph, 16:9, neutral enough that every shot of this place can be staged from it later. No readable writing anywhere: surfaces that would carry writing may be present, but stage any wording out of legibility — an oblique angle, distance, shallow focus. No captions, watermarks or overlay text.",
  "effective_prompt": "Create ONE empty live-action location background photograph — NO PEOPLE, no figures, no body parts, no silhouettes, no shadows or reflections of people anywhere.\n\"Empty\" means no people only: KEEP the location's inherent occupants and stock that define the place — animals in an animal shelter, pen or farm, goods and produce in a market, moored boats in a harbour — unless the shot text explicitly removes them.\nThe FIRST attached image is a thin-line storyboard sketch of a shot that happens at this location: use it ONLY as spatial evidence — what this place contains, how its ground, structures and landmarks are arranged and proportioned. Ignore the sketched people and arrows entirely, and do NOT copy its line style: render a fully photographic, physically plausible real place that fits THE LOCATION text below.\nHUMAN-SCALE CALIBRATION: derive every structure's true size from human-scale elements — a door ≈ 2m, a window ≈ 1–1.5m wide, one storey ≈ 2.5–3m; never inflate a small structure or shrink a large one.\n\nTHE LOCATION — early 20th-century Korea, Jeolla region; all people are Korean unless explicitly stated otherwise: The separate forest descent on the opposite side of the mountain camp, away from the approaching lone climber.\n\nLOCATION DETAIL (descriptive evidence about this place from the production's location records — evidence, not a staging order): use it to understand what kind of place this is and what it permanently contains. Carry over ONLY its enduring physical features — layout, structures, ground and materials, fixed equipment and installations. Do NOT treat any lighting, weather or time-of-day wording, momentary object states, depicted events or subjective impressions in it as fixed facts of the place; the TIME OF DAY lock below is the sole authority for time and lighting.\n파청 산채 진입 오르막 숲길: 산채로 연결되는 경사가 가파른 흙비탈 숲길. 나무 뿌리와 잡목이 길가로 돌출되어 있다. (특징: 경사가 심한 흙바닥 비탈; 우거진 수목과 돌출된 나무 뿌리; 군화나 짚신 자국이 남는 흙길) / 파청 산채 하산 숲길·비탈: 산채에서 하부 장터 쪽으로 길게 뻗어 내리는 험한 내리막 숲길. 나무 기둥 사이로 아득히 저잣거리의 풍경이 부분적으로 보인다. (특징: 가파른 내리막 흙길 비탈; 시야를 가리는 나무 기둥들; 멀리 아득하게 보이는 마을/장터 전경)\n\nSCENE EVIDENCE (verbatim quotes from the screenplay about this place — treat them as evidence of what the location physically contains and looks like; stage the PLACE those moments happen in, but do NOT depict the momentary actions, people or staged props themselves):\n- 반대편 산길.\n토미이시를 비롯한 일본군 일부가 하산하고 있다. 두 길은 서로 어긋난다.\n- S#62. 파청 산채 하산 숲길 / 낮 - 계속\n- 하산하던 토미이시가 걸음을 멈춘다.\n- 일본군들이 몸을 돌려 다시 산을 오른다.\n- 몇 걸음 오르지 않아—\n\nTIME OF DAY (lock): day.\n\nRender ONE photorealistic empty location photograph, 16:9, neutral enough that every shot of this place can be staged from it later. No readable writing anywhere: surfaces that would carry writing may be present, but stage any wording out of legibility — an oblique angle, distance, shallow focus. No captions, watermarks or overlay text.",
  "bg_path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/db00517b-165b-4005-a287-f28936b59b95/images/272bd7af-9294-4490-98c6-88a40e9deacc/scene/recipe/groupbg_opposite_descent_da96a5.png",
  "asset_id": "952ff4b1-07e5-4cbd-a68d-f12e2d4e9856",
  "input_asset_ids": [
   "a9955fc4-6842-40a5-a557-ac83ab7a0d93"
  ],
  "origin_tag": "S59sh11",
  "place_text": "The separate forest descent on the opposite side of the mountain camp, away from the approaching lone climber.",
  "origin_inputs": {
   "place_text": "The separate forest descent on the opposite side of the mountain camp, away from the approaching lone climber.",
   "time_of_day_en": "day",
   "conti_asset_id": "a9955fc4-6842-40a5-a557-ac83ab7a0d93"
  }
 },
 "S59sh11::bgfirst_bg": {
  "input_fingerprint": "6da8b48813944f37",
  "prompt": "Create the EMPTY BACKGROUND PLATE for one film shot — NO PEOPLE, no figures, no body parts, no sketch lines, no arrows anywhere.\n\"Empty\" means no people only: KEEP the location's inherent occupants and stock that define the place — animals in an animal shelter, pen or farm, goods in a market, moored boats in a harbour — unless the shot text explicitly removes them.\nThe FIRST attached image is a thin-line storyboard sketch: use ONLY its camera angle, horizon, perspective and the placement/size of buildings and set masses — ignore the sketched people and arrows entirely. The SECOND attached image (LOCATION PHOTOGRAPH) is the real place: take its architecture, materials, signage and fixed features, and RE-PROJECT them into the sketch's camera. If the photograph's camera differs from the sketch's, the sketch's camera wins.\nHUMAN-SCALE CALIBRATION: derive every structure's true size from human-scale elements — a door ≈ 2m, a window ≈ 1–1.5m wide, one storey ≈ 2.5–3m; never inflate a small structure or shrink a large one.\n\nSHOT TEXT this background must serve (Korean): 반대편 숲길 비탈, 토미이시와 일본인 군인들이 총을 멘 채 아래로 발을 내디딘 mid-stride 순간의 뒷모습.\n\nLOCATION (lock): The separate forest descent on the opposite side of the mountain camp, away from the approaching lone climber.\n\nTIME OF DAY (lock): day.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: wide shot\n- FRAME LAYOUT: Downhill continuation of the opposite forest path in the lower-right of the frame, midground.\n- KEY BACKGROUND ELEMENTS: Opposite descending forest path (Occupied by the descending party) — Runs from the upper-left depth toward the lower-right frame edge; used as Establishes a separate route and a clearly descending movement; Slung rifles (Carried by the soldiers) — Seen obliquely beside their backs, at ordinary body-relative scale; used as Clarify the armed party without enlarging any single weapon.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Daylight remains naturalistic and restrained, maintaining the sober tonal continuity across the cut to the other path.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nRender ONE photorealistic empty location photograph, 16:9, that this shot can be staged inside later. No readable writing anywhere: surfaces that would carry writing may be present, but stage any wording out of legibility — an oblique angle, distance, shallow focus. No captions, watermarks or overlay text.",
  "effective_prompt": "Create the EMPTY BACKGROUND PLATE for one film shot — NO PEOPLE, no figures, no body parts, no sketch lines, no arrows anywhere.\n\"Empty\" means no people only: KEEP the location's inherent occupants and stock that define the place — animals in an animal shelter, pen or farm, goods in a market, moored boats in a harbour — unless the shot text explicitly removes them.\nThe FIRST attached image is a thin-line storyboard sketch: use ONLY its camera angle, horizon, perspective and the placement/size of buildings and set masses — ignore the sketched people and arrows entirely. The SECOND attached image (LOCATION PHOTOGRAPH) is the real place: take its architecture, materials, signage and fixed features, and RE-PROJECT them into the sketch's camera. If the photograph's camera differs from the sketch's, the sketch's camera wins.\nHUMAN-SCALE CALIBRATION: derive every structure's true size from human-scale elements — a door ≈ 2m, a window ≈ 1–1.5m wide, one storey ≈ 2.5–3m; never inflate a small structure or shrink a large one.\n\nSHOT TEXT this background must serve (Korean): 반대편 숲길 비탈, 토미이시와 일본인 군인들이 총을 멘 채 아래로 발을 내디딘 mid-stride 순간의 뒷모습.\n\nLOCATION (lock): The separate forest descent on the opposite side of the mountain camp, away from the approaching lone climber.\n\nTIME OF DAY (lock): day.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: wide shot\n- FRAME LAYOUT: Downhill continuation of the opposite forest path in the lower-right of the frame, midground.\n- KEY BACKGROUND ELEMENTS: Opposite descending forest path (Occupied by the descending party) — Runs from the upper-left depth toward the lower-right frame edge; used as Establishes a separate route and a clearly descending movement; Slung rifles (Carried by the soldiers) — Seen obliquely beside their backs, at ordinary body-relative scale; used as Clarify the armed party without enlarging any single weapon.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Daylight remains naturalistic and restrained, maintaining the sober tonal continuity across the cut to the other path.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nRender ONE photorealistic empty location photograph, 16:9, that this shot can be staged inside later. No readable writing anywhere: surfaces that would carry writing may be present, but stage any wording out of legibility — an oblique angle, distance, shallow focus. No captions, watermarks or overlay text.",
  "bg_path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/db00517b-165b-4005-a287-f28936b59b95/images/272bd7af-9294-4490-98c6-88a40e9deacc/scene/recipe/S59sh11__bgfirst_bg.png",
  "asset_id": "4b90ccc2-3840-44cc-be9c-7539502dfaf6",
  "input_asset_ids": [
   "a9955fc4-6842-40a5-a557-ac83ab7a0d93",
   "952ff4b1-07e5-4cbd-a68d-f12e2d4e9856"
  ]
 },
 "S59sh11": {
  "input_fingerprint": "5b6c47e9e1e88280",
  "prompt": "Create ONE FINAL photorealistic live-action film still of the moment below — early 20th-century Korea, Jeolla region; all people are Korean unless explicitly stated otherwise. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 반대편 숲길 비탈, 토미이시와 일본인 군인들이 총을 멘 채 아래로 발을 내디딘 mid-stride 순간의 뒷모습.\n\nLOCATION (lock): The separate forest descent on the opposite side of the mountain camp, away from the approaching lone climber. The shot takes place here — the attached LOCATION PHOTOGRAPH shows the exact spot.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: wide shot\n- FRAME LAYOUT: Downhill continuation of the opposite forest path in the lower-right of the frame, midground.\n- KEY BACKGROUND ELEMENTS: Opposite descending forest path (Occupied by the descending party) — Runs from the upper-left depth toward the lower-right frame edge; used as Establishes a separate route and a clearly descending movement; Slung rifles (Carried by the soldiers) — Seen obliquely beside their backs, at ordinary body-relative scale; used as Clarify the armed party without enlarging any single weapon.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Daylight remains naturalistic and restrained, maintaining the sober tonal continuity across the cut to the other path.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): Separate wooded paths carry the ascent and descent without intersecting. 토미이시: He walks downhill retaining his sword.\n\nTHE HAND THAT IS DOING THIS: the object at the centre of this shot is being held, operated, read, aimed or handed over by 토미이시와 일본인 군인들 right now, so 토미이시와 일본인 군인들's hand — and as much of the wrist and forearm as the framing reaches — is in the frame, gripping or resting on that object exactly the way its real use requires. Match that hand to 토미이시와 일본인 군인들: its size, build, skin, age, grooming, sleeve and anything worn on it belong to that person and to no one else.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 토미이시 (일본인 남성, 성인 나이대, 자연스러운 검은 머리) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.",
  "roll_prompts": {
   "A": "Stage the shot. The FIRST attached image (SHOT BACKGROUND) is the finished empty background of this shot — keep it EXACTLY: its camera, perspective, architecture, lighting and every fixed feature stay untouched. The SECOND attached image (LAYOUT SKETCH) tells you ONLY where the people go: each sketched person's position, screen size, pose and the gaze/motion arrows. Ignore the sketch's background lines. The CHARACTER REFERENCE photographs show the real people.\nPlace the real people into the background at exactly the sketched positions, sizes and poses, following the arrow directions. No sketch lines or arrows may remain.\n\nCreate ONE FINAL photorealistic live-action film still of the moment below — early 20th-century Korea, Jeolla region; all people are Korean unless explicitly stated otherwise. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 반대편 숲길 비탈, 토미이시와 일본인 군인들이 총을 멘 채 아래로 발을 내디딘 mid-stride 순간의 뒷모습.\n\nLOCATION (lock): The separate forest descent on the opposite side of the mountain camp, away from the approaching lone climber. The shot takes place here — the FIRST attached image (SHOT BACKGROUND) is this exact place, already built: its ground, structures, horizon, materials and lighting are the finished truth of this location and must not be redesigned or replaced. No location photograph is attached — read the place from that image alone, and add no scenery, structure, vehicle or fixture that it does not already show. This lock governs the place only; the figures in the shot follow the staging and pose instructions.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: wide shot\n- FRAME LAYOUT: Downhill continuation of the opposite forest path in the lower-right of the frame, midground.\n- KEY BACKGROUND ELEMENTS: Opposite descending forest path (Occupied by the descending party) — Runs from the upper-left depth toward the lower-right frame edge; used as Establishes a separate route and a clearly descending movement; Slung rifles (Carried by the soldiers) — Seen obliquely beside their backs, at ordinary body-relative scale; used as Clarify the armed party without enlarging any single weapon.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Daylight remains naturalistic and restrained, maintaining the sober tonal continuity across the cut to the other path.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): Separate wooded paths carry the ascent and descent without intersecting. 토미이시: He walks downhill retaining his sword.\n\nTHE HAND THAT IS DOING THIS: the object at the centre of this shot is being held, operated, read, aimed or handed over by 토미이시와 일본인 군인들 right now, so 토미이시와 일본인 군인들's hand — and as much of the wrist and forearm as the framing reaches — is in the frame, gripping or resting on that object exactly the way its real use requires. Match that hand to 토미이시와 일본인 군인들: its size, build, skin, age, grooming, sleeve and anything worn on it belong to that person and to no one else.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 토미이시 (일본인 남성, 성인 나이대, 자연스러운 검은 머리) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.",
   "B": "Create ONE FINAL photorealistic live-action film still of the moment below — early 20th-century Korea, Jeolla region; all people are Korean unless explicitly stated otherwise. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 반대편 숲길 비탈, 토미이시와 일본인 군인들이 총을 멘 채 아래로 발을 내디딘 mid-stride 순간의 뒷모습.\n\nLOCATION (lock): The separate forest descent on the opposite side of the mountain camp, away from the approaching lone climber. The shot takes place here — the attached LOCATION PHOTOGRAPH shows the exact spot.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: wide shot\n- FRAME LAYOUT: Downhill continuation of the opposite forest path in the lower-right of the frame, midground.\n- KEY BACKGROUND ELEMENTS: Opposite descending forest path (Occupied by the descending party) — Runs from the upper-left depth toward the lower-right frame edge; used as Establishes a separate route and a clearly descending movement; Slung rifles (Carried by the soldiers) — Seen obliquely beside their backs, at ordinary body-relative scale; used as Clarify the armed party without enlarging any single weapon.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Daylight remains naturalistic and restrained, maintaining the sober tonal continuity across the cut to the other path.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): Separate wooded paths carry the ascent and descent without intersecting. 토미이시: He walks downhill retaining his sword.\n\nTHE HAND THAT IS DOING THIS: the object at the centre of this shot is being held, operated, read, aimed or handed over by 토미이시와 일본인 군인들 right now, so 토미이시와 일본인 군인들's hand — and as much of the wrist and forearm as the framing reaches — is in the frame, gripping or resting on that object exactly the way its real use requires. Match that hand to 토미이시와 일본인 군인들: its size, build, skin, age, grooming, sleeve and anything worn on it belong to that person and to no one else.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 토미이시 (일본인 남성, 성인 나이대, 자연스러운 검은 머리) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.\n\nCOMPOSITION VARIATION (this candidate only): keep the place, the people and the moment exactly as specified, and keep from the CAMERA & FRAME contract its subject scale, screen placement and key background elements exactly. For this candidate only, its camera angle and height are deliberately left open: choose a camera position distinctly different from the obvious one, as a film director picking a second setup on the same blocking. Stay true to any explicit framing the SHOT TEXT itself dictates."
  },
  "roll_refs": {
   "A": [
    {
     "label": "SHOT BACKGROUND",
     "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/db00517b-165b-4005-a287-f28936b59b95/images/272bd7af-9294-4490-98c6-88a40e9deacc/scene/recipe/S59sh11__bgfirst_bg.png",
     "asset_id": "4b90ccc2-3840-44cc-be9c-7539502dfaf6",
     "role": "bgfirst_bg"
    },
    {
     "label": "LAYOUT SKETCH (people placement only)",
     "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/db00517b-165b-4005-a287-f28936b59b95/images/272bd7af-9294-4490-98c6-88a40e9deacc/conti/conti_S59sh11.png",
     "asset_id": "a9955fc4-6842-40a5-a557-ac83ab7a0d93",
     "role": "conti_light"
    },
    {
     "label": "CHARACTER REFERENCE — 토미이시: the exact person appearing in this shot; match face, hair and build exactly.",
     "path": "<bytes:797699>",
     "asset_id": "c861ba0f-8750-4262-8324-25cf5c6c9dcd",
     "role": "character_ref"
    },
    {
     "label": "PROP REFERENCE — 근대식 소총: the exact object appearing in this shot; match its look, material and wear exactly.",
     "path": "<bytes:828568>",
     "asset_id": "7543806c-8087-4caa-8d7b-56d10864fd3c",
     "role": "prop_ref"
    }
   ],
   "B": [
    {
     "label": "LOCATION PHOTOGRAPH — the exact place of this shot: its architecture, materials, fixed features and lighting mood are spatial truth; stage the moment inside this place. Never copy its camera framing.",
     "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/db00517b-165b-4005-a287-f28936b59b95/images/272bd7af-9294-4490-98c6-88a40e9deacc/scene/recipe/groupbg_opposite_descent_da96a5.png",
     "asset_id": "952ff4b1-07e5-4cbd-a68d-f12e2d4e9856",
     "role": "bgfirst_group_bg"
    },
    {
     "label": "CHARACTER REFERENCE — 토미이시: the exact person appearing in this shot; match face, hair and build exactly.",
     "path": "<bytes:797699>",
     "asset_id": "c861ba0f-8750-4262-8324-25cf5c6c9dcd",
     "role": "character_ref"
    },
    {
     "label": "PROP REFERENCE — 근대식 소총: the exact object appearing in this shot; match its look, material and wear exactly.",
     "path": "<bytes:828568>",
     "asset_id": "7543806c-8087-4caa-8d7b-56d10864fd3c",
     "role": "prop_ref"
    }
   ]
  },
  "shot_run_spend_attempt_count": 1,
  "judge_flip_skipped": "judge_fn_owns_order",
  "readings": [
   {
    "label": "A",
    "direction": "일행은 카메라를 등지고 우측 하단 숲길 방향을 향해 이동 중임.",
    "built_space": "레퍼런스와 동일한 질감과 구조를 가진 숲길 비탈이 우측 하단으로 뻗어 있음.",
    "entities": "일본군 병사들과 소총은 확인되나, 캐릭터 레퍼런스에 명시된 토미이시(정모, 가죽 부츠 착용)와 일치하는 인물이 없음. 전원 전투모와 각반 착용.",
    "hard_violations": [
     "[gemini-pro] 좌측 인물의 허리에 찬 검(검집)이 끈이나 지지대 없이 엉덩이 뒤쪽 허공을 향해 수평으로 뻗어 있어 물리적으로 불가능한 형태임."
    ],
    "physics": "발이 땅을 딛고 걷는 동작 자체는 성립하나, 좌측 인물의 검집이 중력을 무시하고 허공에 뜬 채 신체에 고정되어 있음."
   },
   {
    "label": "B",
    "direction": "일행은 카메라를 등지고 우측 하단으로 이어지는 숲길 내리막을 향해 시선과 방향을 두고 있음.",
    "built_space": "레퍼런스와 일치하는 숲속 비탈길. 좌측 상단에서 우측 하단으로 자연스럽게 이어지는 동선이 프레임에 맞게 배치됨.",
    "entities": "정모와 가죽 부츠, 장교복을 착용한 토미이시가 명확히 식별됨. 주변 일본군 병사들도 구현되었으며, 지시된 형태의 근대식 소총을 메고 있고 토미이시의 검도 확인됨.",
    "hard_violations": [],
    "physics": "내리막을 걷는 mid-stride 동작에 맞춰 체중이 실린 발이 땅을 안정적으로 딛고 있으며, 등에 멘 총과 허리에 찬 검이 중력에 맞게 늘어져 있음."
   }
  ],
  "cross_model_order": {
   "policy": "cross_model_order_v2_slot0_fwd_slot1_rev_parallel2",
   "slots": [
    {
     "model": "gemini-pro",
     "order": "forward",
     "display_to_canonical": {
      "A": "A",
      "B": "B"
     },
     "raw": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 9,
        "verdict_ko": "토미이시의 캐릭터 레퍼런스(정모, 가죽 부츠)를 정확히 반영하였고, 샷 텍스트에서 요구한 내리막길을 걷는 뒷모습과 무기 지참 상태를 훌륭하게 구현했습니다."
       },
       {
        "label": "A",
        "score": 4,
        "verdict_ko": "모든 인물이 전투모와 각반을 착용하여 토미이시 캐릭터 레퍼런스를 누락했으며, 좌측 인물의 검집이 비현실적으로 허공에 뻗어 있습니다."
       }
      ],
      "readings": [
       {
        "label": "B",
        "direction": "일행은 카메라를 등지고 우측 하단으로 이어지는 숲길 내리막을 향해 시선과 방향을 두고 있음.",
        "built_space": "레퍼런스와 일치하는 숲속 비탈길. 좌측 상단에서 우측 하단으로 자연스럽게 이어지는 동선이 프레임에 맞게 배치됨.",
        "entities": "정모와 가죽 부츠, 장교복을 착용한 토미이시가 명확히 식별됨. 주변 일본군 병사들도 구현되었으며, 지시된 형태의 근대식 소총을 메고 있고 토미이시의 검도 확인됨.",
        "hard_violations": [],
        "physics": "내리막을 걷는 mid-stride 동작에 맞춰 체중이 실린 발이 땅을 안정적으로 딛고 있으며, 등에 멘 총과 허리에 찬 검이 중력에 맞게 늘어져 있음."
       },
       {
        "label": "A",
        "direction": "일행은 카메라를 등지고 우측 하단 숲길 방향을 향해 이동 중임.",
        "built_space": "레퍼런스와 동일한 질감과 구조를 가진 숲길 비탈이 우측 하단으로 뻗어 있음.",
        "entities": "일본군 병사들과 소총은 확인되나, 캐릭터 레퍼런스에 명시된 토미이시(정모, 가죽 부츠 착용)와 일치하는 인물이 없음. 전원 전투모와 각반 착용.",
        "hard_violations": [
         "좌측 인물의 허리에 찬 검(검집)이 끈이나 지지대 없이 엉덩이 뒤쪽 허공을 향해 수평으로 뻗어 있어 물리적으로 불가능한 형태임."
        ],
        "physics": "발이 땅을 딛고 걷는 동작 자체는 성립하나, 좌측 인물의 검집이 중력을 무시하고 허공에 뜬 채 신체에 고정되어 있음."
       }
      ],
      "all_candidates_fail": false
     },
     "normalized": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 9,
        "verdict_ko": "토미이시의 캐릭터 레퍼런스(정모, 가죽 부츠)를 정확히 반영하였고, 샷 텍스트에서 요구한 내리막길을 걷는 뒷모습과 무기 지참 상태를 훌륭하게 구현했습니다."
       },
       {
        "label": "A",
        "score": 4,
        "verdict_ko": "모든 인물이 전투모와 각반을 착용하여 토미이시 캐릭터 레퍼런스를 누락했으며, 좌측 인물의 검집이 비현실적으로 허공에 뻗어 있습니다."
       }
      ],
      "all_candidates_fail": false,
      "readings": [
       {
        "label": "B",
        "direction": "일행은 카메라를 등지고 우측 하단으로 이어지는 숲길 내리막을 향해 시선과 방향을 두고 있음.",
        "built_space": "레퍼런스와 일치하는 숲속 비탈길. 좌측 상단에서 우측 하단으로 자연스럽게 이어지는 동선이 프레임에 맞게 배치됨.",
        "entities": "정모와 가죽 부츠, 장교복을 착용한 토미이시가 명확히 식별됨. 주변 일본군 병사들도 구현되었으며, 지시된 형태의 근대식 소총을 메고 있고 토미이시의 검도 확인됨.",
        "hard_violations": [],
        "physics": "내리막을 걷는 mid-stride 동작에 맞춰 체중이 실린 발이 땅을 안정적으로 딛고 있으며, 등에 멘 총과 허리에 찬 검이 중력에 맞게 늘어져 있음."
       },
       {
        "label": "A",
        "direction": "일행은 카메라를 등지고 우측 하단 숲길 방향을 향해 이동 중임.",
        "built_space": "레퍼런스와 동일한 질감과 구조를 가진 숲길 비탈이 우측 하단으로 뻗어 있음.",
        "entities": "일본군 병사들과 소총은 확인되나, 캐릭터 레퍼런스에 명시된 토미이시(정모, 가죽 부츠 착용)와 일치하는 인물이 없음. 전원 전투모와 각반 착용.",
        "hard_violations": [
         "좌측 인물의 허리에 찬 검(검집)이 끈이나 지지대 없이 엉덩이 뒤쪽 허공을 향해 수평으로 뻗어 있어 물리적으로 불가능한 형태임."
        ],
        "physics": "발이 땅을 딛고 걷는 동작 자체는 성립하나, 좌측 인물의 검집이 중력을 무시하고 허공에 뜬 채 신체에 고정되어 있음."
       }
      ]
     },
     "ok": true
    },
    {
     "model": "gpt-high",
     "order": "reverse",
     "display_to_canonical": {
      "A": "B",
      "B": "A"
     },
     "raw": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 6,
        "verdict_ko": "토미이시의 복장과 장소는 더 가깝지만, 일행이 오른쪽 아래에서 중앙 안쪽으로 멀어져 내려가므로 지정된 왼쪽 위→오른쪽 아래 하행 구도를 따르지 않는다."
       },
       {
        "label": "B",
        "score": 8,
        "verdict_ko": "왼쪽 위에서 오른쪽 아래로 이어지는 내리막과 일행의 뒷모습 보행 순간을 더 충실히 구현하지만, 토미이시의 모자와 각반은 인물 참조와 다르다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "네 사람이 등을 보인 채 오른쪽 아래의 가까운 지점에서 화면 중앙의 먼 길로 내려간다. 머리도 대체로 앞선 일행과 진행할 길을 향하며 눈 자체는 보이지 않는다. 가까운 두 사람의 소총은 등 뒤에 비스듬히 걸려 총구가 왼쪽 위를 향하고, 특정 대상을 겨누지 않는다. 실제 하행은 보이지만 이동 방향은 지정된 왼쪽 위 깊은 곳에서 오른쪽 아래 가장자리로 향하는 흐름과 반대다.",
        "built_space": "인공 구조물이나 고정 설비는 없다. 왼쪽의 흙 절개면과 노출된 뿌리, 오른쪽의 높은 나무들, 나무 사이로 보이는 계곡과 산은 장소 참조와 매우 가깝다. 네 사람은 하나의 좁은 흙길 위에 차례로 놓였고 다른 등산객이나 교차로는 보이지 않는다. 다만 오른쪽 아래 길이 하행의 도착 방향이 아니라 가까운 출발 지점이 되었다.",
        "entities": "장교 한 명과 병사 세 명으로 보이는 성인 남성 네 명이 있다. 뒷모습이라 일본인이라는 설정과 얼굴의 정확한 동일성은 확인할 수 없다. 중앙 장교는 검은 머리, 정모, 갈색 계열 군복, 교차 가죽띠와 긴 가죽 장화를 갖춰 토미이시 참조에 가깝다. 왼쪽 허리 옆에는 칼집으로 보이는 긴 물체가 일부 보인다. 가까운 소총들은 목재 개머리판과 금속 총열을 갖추고 신체 대비 크기도 자연스럽지만, 참조의 발화 장치와 총검 형태까지 일치하는지는 확인되지 않는다. 읽을 수 있는 글자는 없다.",
        "hard_violations": [],
        "physics": "장교는 한쪽 장화로 비탈을 지지하면서 다른 발을 옮기며, 가까운 병사도 지면에 둔 다리와 뒤로 들린 발로 보행 중 체중 이동을 보인다. 앞선 두 사람 역시 길 위에서 걷는 자세다. 소총은 어깨끈과 몸통에 지지되어 있어 공중에 떠 있지 않는다. 다만 가까운 인물들의 보이는 손은 아래로 내려가 있어, 무기나 그 끈에 손을 대라는 별도 지시는 명확히 구현되지 않았다."
       },
       {
        "label": "B",
        "direction": "네 사람의 몸과 머리가 오른쪽 아래로 이어지는 내리막 진행 방향을 향하고, 카메라에는 뒤와 약간의 옆모습이 보인다. 왼쪽의 가까운 인물부터 오른쪽의 먼 인물까지 낮아지는 대각선 행렬이 지정된 하행 흐름을 만든다. 병사들의 소총은 등 옆에 걸려 총구가 어깨 위쪽을 향하며 누구도 조준하지 않는다. 토미이시로 보이는 왼쪽 인물의 칼집은 허리에서 왼쪽 아래로 뻗는다.",
        "built_space": "인공 구조물이나 고정 설비는 없다. 왼쪽의 뿌리 드러난 흙비탈, 오른쪽의 수직 나무줄기들, 뒤편 계곡과 산이 장소 참조의 지형과 재질을 유지한다. 네 사람 모두 같은 좁은 길에 서 있으며 길은 왼쪽 높은 쪽에서 오른쪽 아래 중경으로 내려간다. 다른 접근로와의 교차나 별도의 등산객은 보이지 않는다. 와이드 숏 안에서 오른쪽 아래의 하행 연장부가 읽힌다.",
        "entities": "칼을 찬 장교로 보이는 인물 한 명과 소총을 멘 병사 세 명, 총 네 명의 성인 남성이 보인다. 일부 검은 머리와 옆얼굴은 보이지만 국적이나 참조 얼굴의 정확한 일치 여부를 확정할 수는 없다. 토미이시로 보이는 왼쪽 인물은 갈색 군복과 교차 가죽띠, 허리에 찬 칼을 유지한다. 그러나 참조의 정모 대신 병사형 모자를 쓰고, 긴 가죽 장화 대신 각반을 착용했다. 병사들의 소총은 목재와 금속으로 된 정상 크기의 장총이지만 참조의 세부 기구와 총검 일치는 확인하기 어렵다. 읽을 수 있는 글자는 없다.",
        "hard_violations": [],
        "physics": "가까운 인물들은 한쪽 발을 비탈에 두고 반대쪽 발꿈치나 발 전체를 들어 다음 보폭으로 옮기는 모습이다. 뒤로 들린 발바닥과 굽힌 무릎이 내려가는 보행에서 나올 수 있는 자세이며, 먼 인물들도 길에 발을 딛고 있다. 소총은 몸을 가로지르는 끈에, 칼집은 허리띠에 지지된다. 지지 없이 떠 있는 몸이나 물체는 없다. 다만 보이는 손들은 주로 보행에 따라 내려가 있어 무기나 끈을 직접 잡으라는 별도 지시는 명확히 충족하지 않는다."
       }
      ],
      "all_candidates_fail": false
     },
     "normalized": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 6,
        "verdict_ko": "토미이시의 복장과 장소는 더 가깝지만, 일행이 오른쪽 아래에서 중앙 안쪽으로 멀어져 내려가므로 지정된 왼쪽 위→오른쪽 아래 하행 구도를 따르지 않는다."
       },
       {
        "label": "A",
        "score": 8,
        "verdict_ko": "왼쪽 위에서 오른쪽 아래로 이어지는 내리막과 일행의 뒷모습 보행 순간을 더 충실히 구현하지만, 토미이시의 모자와 각반은 인물 참조와 다르다."
       }
      ],
      "all_candidates_fail": false,
      "readings": [
       {
        "label": "B",
        "direction": "네 사람이 등을 보인 채 오른쪽 아래의 가까운 지점에서 화면 중앙의 먼 길로 내려간다. 머리도 대체로 앞선 일행과 진행할 길을 향하며 눈 자체는 보이지 않는다. 가까운 두 사람의 소총은 등 뒤에 비스듬히 걸려 총구가 왼쪽 위를 향하고, 특정 대상을 겨누지 않는다. 실제 하행은 보이지만 이동 방향은 지정된 왼쪽 위 깊은 곳에서 오른쪽 아래 가장자리로 향하는 흐름과 반대다.",
        "built_space": "인공 구조물이나 고정 설비는 없다. 왼쪽의 흙 절개면과 노출된 뿌리, 오른쪽의 높은 나무들, 나무 사이로 보이는 계곡과 산은 장소 참조와 매우 가깝다. 네 사람은 하나의 좁은 흙길 위에 차례로 놓였고 다른 등산객이나 교차로는 보이지 않는다. 다만 오른쪽 아래 길이 하행의 도착 방향이 아니라 가까운 출발 지점이 되었다.",
        "entities": "장교 한 명과 병사 세 명으로 보이는 성인 남성 네 명이 있다. 뒷모습이라 일본인이라는 설정과 얼굴의 정확한 동일성은 확인할 수 없다. 중앙 장교는 검은 머리, 정모, 갈색 계열 군복, 교차 가죽띠와 긴 가죽 장화를 갖춰 토미이시 참조에 가깝다. 왼쪽 허리 옆에는 칼집으로 보이는 긴 물체가 일부 보인다. 가까운 소총들은 목재 개머리판과 금속 총열을 갖추고 신체 대비 크기도 자연스럽지만, 참조의 발화 장치와 총검 형태까지 일치하는지는 확인되지 않는다. 읽을 수 있는 글자는 없다.",
        "hard_violations": [],
        "physics": "장교는 한쪽 장화로 비탈을 지지하면서 다른 발을 옮기며, 가까운 병사도 지면에 둔 다리와 뒤로 들린 발로 보행 중 체중 이동을 보인다. 앞선 두 사람 역시 길 위에서 걷는 자세다. 소총은 어깨끈과 몸통에 지지되어 있어 공중에 떠 있지 않는다. 다만 가까운 인물들의 보이는 손은 아래로 내려가 있어, 무기나 그 끈에 손을 대라는 별도 지시는 명확히 구현되지 않았다."
       },
       {
        "label": "A",
        "direction": "네 사람의 몸과 머리가 오른쪽 아래로 이어지는 내리막 진행 방향을 향하고, 카메라에는 뒤와 약간의 옆모습이 보인다. 왼쪽의 가까운 인물부터 오른쪽의 먼 인물까지 낮아지는 대각선 행렬이 지정된 하행 흐름을 만든다. 병사들의 소총은 등 옆에 걸려 총구가 어깨 위쪽을 향하며 누구도 조준하지 않는다. 토미이시로 보이는 왼쪽 인물의 칼집은 허리에서 왼쪽 아래로 뻗는다.",
        "built_space": "인공 구조물이나 고정 설비는 없다. 왼쪽의 뿌리 드러난 흙비탈, 오른쪽의 수직 나무줄기들, 뒤편 계곡과 산이 장소 참조의 지형과 재질을 유지한다. 네 사람 모두 같은 좁은 길에 서 있으며 길은 왼쪽 높은 쪽에서 오른쪽 아래 중경으로 내려간다. 다른 접근로와의 교차나 별도의 등산객은 보이지 않는다. 와이드 숏 안에서 오른쪽 아래의 하행 연장부가 읽힌다.",
        "entities": "칼을 찬 장교로 보이는 인물 한 명과 소총을 멘 병사 세 명, 총 네 명의 성인 남성이 보인다. 일부 검은 머리와 옆얼굴은 보이지만 국적이나 참조 얼굴의 정확한 일치 여부를 확정할 수는 없다. 토미이시로 보이는 왼쪽 인물은 갈색 군복과 교차 가죽띠, 허리에 찬 칼을 유지한다. 그러나 참조의 정모 대신 병사형 모자를 쓰고, 긴 가죽 장화 대신 각반을 착용했다. 병사들의 소총은 목재와 금속으로 된 정상 크기의 장총이지만 참조의 세부 기구와 총검 일치는 확인하기 어렵다. 읽을 수 있는 글자는 없다.",
        "hard_violations": [],
        "physics": "가까운 인물들은 한쪽 발을 비탈에 두고 반대쪽 발꿈치나 발 전체를 들어 다음 보폭으로 옮기는 모습이다. 뒤로 들린 발바닥과 굽힌 무릎이 내려가는 보행에서 나올 수 있는 자세이며, 먼 인물들도 길에 발을 딛고 있다. 소총은 몸을 가로지르는 끈에, 칼집은 허리띠에 지지된다. 지지 없이 떠 있는 몸이나 물체는 없다. 다만 보이는 손들은 주로 보행에 따라 내려가 있어 무기나 끈을 직접 잡으라는 별도 지시는 명확히 충족하지 않는다."
       }
      ]
     },
     "ok": true
    }
   ],
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "slot_winner_match": false,
   "slot_winner": {
    "gemini-pro": "B",
    "gpt-high": "A"
   },
   "route": "cross_slot_combined"
  },
  "dual": {
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "normalized": {
    "A": 1.444,
    "B": 1.75
   },
   "adjusted": {
    "A": 1.194,
    "B": 1.75
   },
   "violations": {
    "A": [
     "[gemini-pro] 좌측 인물의 허리에 찬 검(검집)이 끈이나 지지대 없이 엉덩이 뒤쪽 허공을 향해 수평으로 뻗어 있어 물리적으로 불가능한 형태임."
    ]
   },
   "per_model_winner": {
    "gemini-pro": "B",
    "gpt-high": "A"
   },
   "agreed": false
  },
  "totals": {
   "B": 1750,
   "A": 1194
  },
  "selected": "B",
  "ranking": [
   "B",
   "A"
  ],
  "verdicts": [
   {
    "label": "B",
    "score": 1750,
    "verdict_ko": "토미이시의 캐릭터 레퍼런스(정모, 가죽 부츠)를 정확히 반영하였고, 샷 텍스트에서 요구한 내리막길을 걷는 뒷모습과 무기 지참 상태를 훌륭하게 구현했습니다."
   },
   {
    "label": "A",
    "score": 1194,
    "verdict_ko": "모든 인물이 전투모와 각반을 착용하여 토미이시 캐릭터 레퍼런스를 누락했으며, 좌측 인물의 검집이 비현실적으로 허공에 뻗어 있습니다.  ★위반: [gemini-pro] 좌측 인물의 허리에 찬 검(검집)이 끈이나 지지대 없이 엉덩이 뒤쪽 허공을 향해 수평으로 뻗어 있어 물리적으로 불가능한 형태임."
   }
  ],
  "refs": [
   {
    "label": "LOCATION PHOTOGRAPH — the exact place of this shot: its architecture, materials, fixed features and lighting mood are spatial truth; stage the moment inside this place. Never copy its camera framing.",
    "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/db00517b-165b-4005-a287-f28936b59b95/images/272bd7af-9294-4490-98c6-88a40e9deacc/scene/recipe/groupbg_opposite_descent_da96a5.png",
    "asset_id": "952ff4b1-07e5-4cbd-a68d-f12e2d4e9856",
    "role": "bgfirst_group_bg"
   },
   {
    "label": "CHARACTER REFERENCE — 토미이시: the exact person appearing in this shot; match face, hair and build exactly.",
    "path": "<bytes:797699>",
    "asset_id": "c861ba0f-8750-4262-8324-25cf5c6c9dcd",
    "role": "character_ref"
   },
   {
    "label": "PROP REFERENCE — 근대식 소총: the exact object appearing in this shot; match its look, material and wear exactly.",
    "path": "<bytes:828568>",
    "asset_id": "7543806c-8087-4caa-8d7b-56d10864fd3c",
    "role": "prop_ref"
   }
  ],
  "critique_skipped": true,
  "shot_run_produced": false,
  "shot_run_uid": "06aa1441-238f-75d3-a6e1-00514bc5e692",
  "bgfirst": {
   "bg_path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/db00517b-165b-4005-a287-f28936b59b95/images/272bd7af-9294-4490-98c6-88a40e9deacc/scene/recipe/S59sh11__bgfirst_bg.png",
   "bg_asset_id": "4b90ccc2-3840-44cc-be9c-7539502dfaf6",
   "bg_record_key": "S59sh11::bgfirst_bg",
   "chain_winner": false,
   "authority": "groupbg",
   "group_key": "opposite_descent",
   "groupbg_asset_id": "952ff4b1-07e5-4cbd-a68d-f12e2d4e9856"
  },
  "ref_mode": "그룹 배경+엔티티 (2택1: 무콘티 승)",
  "share_plan": {
   "ref_plan": "background"
  }
 },
 "S59sh11::cine": {
  "applied": true,
  "attempted_at": "2026-09-09T07:52:44.275821+00:00",
  "fingerprint": "78d834becc2d1995ec7b2c37d5a2b0cef2a14a888248320bbdfe6015e186176c",
  "fingerprint_version": 2,
  "provider": "mai",
  "endpoint": "openrouter/chat-completions",
  "model": "microsoft/mai-image-2.6",
  "pack": "24.202608252115",
  "source_file": "S59sh11_sel.png",
  "source_sha256": "ecf3d4ee3f56ccfcd031b2becfd42504da9c8c5e63e87e3ef082ebcb5b39879a",
  "file": "S59sh11_cine.png",
  "staged_sha256": "454d8545b5335ed397fbccc92bfc2eaf2f90c7d1b131d722bac41cfb9506e9ec",
  "latency_ms": 29028
 },
 "S60sh2::signage": {
  "fp": "254e3c008b00faa0",
  "inscriptions": [],
  "cues": [],
  "dropped": []
 },
 "S60sh2": {
  "input_fingerprint": "c6f43121a1a58a10",
  "prompt": "Create ONE FINAL photorealistic live-action film still of the moment below — early 20th-century Korea, Jeolla region; all people are Korean unless explicitly stated otherwise. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 자욱한 연기 속, 피투성이로 쓰러진 조선인 의병들의 시신들이 널브러진 참혹한 전경.\n\nLOCATION (lock): The ruined outdoor clearing of the mountain camp below the ridge, filled with lingering smoke and fallen men. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: wide shot\n- KEY BACKGROUND ELEMENTS: Camp ground (Scattered with motionless fallen fighters) — Seen diagonally from the ridge above; used as Supplies a continuous spatial field for the dispersed bodies rather than isolating a single casualty; Smoke (Hanging between the fallen bodies); used as Partially separates depth layers without concealing the overall aftermath.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Daylight filtered through the scene's smoke softens separation while restrained saturation preserves the gravity of the bloodied aftermath.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The people visible in that photo are NOT in this shot: never carry their faces, bodies or clothing onto anyone here. Each person in THIS shot is defined solely by the PEOPLE list, their own CHARACTER REFERENCE images and their wardrobe notes. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): Smoke hangs over the ruined camp; the overturned cauldron, broken bowls and white rice scattered in blood and dirt remain in place.\n\nNO PEOPLE IN THIS SHOT: the shot text shows only the place and its state — no living person appears in frame, and no face, and no figure standing, sitting or passing through. Do not invent someone to occupy the space.\n\nOne thing this rule does NOT forbid: the hand or arm of whoever is doing what the shot text describes. If the moment is an object being held, carried, operated, read, aimed, worn or handed over, the hand doing it belongs in frame, resting on or gripping that object exactly as its real use requires. An object photographed floating unsupported, with the hand that must be holding it erased, is a worse failure than the person you were avoiding — render the hand.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.",
  "roll_prompts": {
   "A": "Create ONE FINAL photorealistic live-action film still of the moment below — early 20th-century Korea, Jeolla region; all people are Korean unless explicitly stated otherwise. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 자욱한 연기 속, 피투성이로 쓰러진 조선인 의병들의 시신들이 널브러진 참혹한 전경.\n\nLOCATION (lock): The ruined outdoor clearing of the mountain camp below the ridge, filled with lingering smoke and fallen men. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: wide shot\n- KEY BACKGROUND ELEMENTS: Camp ground (Scattered with motionless fallen fighters) — Seen diagonally from the ridge above; used as Supplies a continuous spatial field for the dispersed bodies rather than isolating a single casualty; Smoke (Hanging between the fallen bodies); used as Partially separates depth layers without concealing the overall aftermath.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Daylight filtered through the scene's smoke softens separation while restrained saturation preserves the gravity of the bloodied aftermath.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The people visible in that photo are NOT in this shot: never carry their faces, bodies or clothing onto anyone here. Each person in THIS shot is defined solely by the PEOPLE list, their own CHARACTER REFERENCE images and their wardrobe notes. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): Smoke hangs over the ruined camp; the overturned cauldron, broken bowls and white rice scattered in blood and dirt remain in place.\n\nNO PEOPLE IN THIS SHOT: the shot text shows only the place and its state — no living person appears in frame, and no face, and no figure standing, sitting or passing through. Do not invent someone to occupy the space.\n\nOne thing this rule does NOT forbid: the hand or arm of whoever is doing what the shot text describes. If the moment is an object being held, carried, operated, read, aimed, worn or handed over, the hand doing it belongs in frame, resting on or gripping that object exactly as its real use requires. An object photographed floating unsupported, with the hand that must be holding it erased, is a worse failure than the person you were avoiding — render the hand.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.",
   "B": "Create ONE FINAL photorealistic live-action film still of the moment below — early 20th-century Korea, Jeolla region; all people are Korean unless explicitly stated otherwise. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 자욱한 연기 속, 피투성이로 쓰러진 조선인 의병들의 시신들이 널브러진 참혹한 전경.\n\nLOCATION (lock): The ruined outdoor clearing of the mountain camp below the ridge, filled with lingering smoke and fallen men. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: wide shot\n- KEY BACKGROUND ELEMENTS: Camp ground (Scattered with motionless fallen fighters) — Seen diagonally from the ridge above; used as Supplies a continuous spatial field for the dispersed bodies rather than isolating a single casualty; Smoke (Hanging between the fallen bodies); used as Partially separates depth layers without concealing the overall aftermath.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Daylight filtered through the scene's smoke softens separation while restrained saturation preserves the gravity of the bloodied aftermath.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The people visible in that photo are NOT in this shot: never carry their faces, bodies or clothing onto anyone here. Each person in THIS shot is defined solely by the PEOPLE list, their own CHARACTER REFERENCE images and their wardrobe notes. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): Smoke hangs over the ruined camp; the overturned cauldron, broken bowls and white rice scattered in blood and dirt remain in place.\n\nNO PEOPLE IN THIS SHOT: the shot text shows only the place and its state — no living person appears in frame, and no face, and no figure standing, sitting or passing through. Do not invent someone to occupy the space.\n\nOne thing this rule does NOT forbid: the hand or arm of whoever is doing what the shot text describes. If the moment is an object being held, carried, operated, read, aimed, worn or handed over, the hand doing it belongs in frame, resting on or gripping that object exactly as its real use requires. An object photographed floating unsupported, with the hand that must be holding it erased, is a worse failure than the person you were avoiding — render the hand.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.\n\nCOMPOSITION VARIATION (this candidate only): keep the place, the people and the moment exactly as specified, and keep from the CAMERA & FRAME contract its subject scale, screen placement and key background elements exactly. For this candidate only, its camera angle and height are deliberately left open: choose a camera position distinctly different from the obvious one, as a film director picking a second setup on the same blocking. Stay true to any explicit framing the SHOT TEXT itself dictates."
  },
  "shot_run_spend_attempt_count": 1,
  "readings": [
   {
    "label": "A",
    "direction": "카메라는 능선 위에서 캠프장을 비스듬히 내려다보는 앵글이며, 바닥에 쓰러진 시신들 사이로 연기가 자연스럽게 깔려 있습니다.",
    "built_space": "흙바닥으로 된 야영장이며, 배경에 레퍼런스 이미지와 동일한 형태의 수직 통나무 울타리와 거친 나무 기둥 및 천으로 엮은 막사들이 정확하게 배치되어 있습니다.",
    "entities": "피투성이가 된 20세기 초 조선인 의병들의 시신들이 다수 널브러져 있으며, 화면 좌측 하단에 뒤집힌 무쇠 가마솥, 깨진 사발들, 그리고 흙과 피에 섞여 쏟아진 흰 쌀이 명확하게 묘사되었습니다.",
    "hard_violations": [
     "[gpt-high] 얼굴을 보이지 말라는 명시적 지시와 달리 전경과 중경의 여러 시신 얼굴이 노출됩니다."
    ],
    "physics": "모든 시신과 총기, 가마솥, 사발 등의 사물들이 지면의 굴곡에 맞춰 흙바닥에 안정적으로 놓여 지지를 받고 있습니다."
   },
   {
    "label": "B",
    "direction": "카메라는 야영장을 대각선으로 바라보고 있으며, 전체적으로 연기가 흩뿌려져 있습니다.",
    "built_space": "흙바닥 야영장으로, 화면 우측에는 돌담이 있고 좌측에는 레퍼런스에서 볼 수 없었던 단단한 흙벽 구조의 초가집이 눈에 띄게 자리 잡고 있습니다.",
    "entities": "피투성이의 조선인 의병 시신들이 흩어져 있고, 중앙 부근에 엎어진 가마솥과 사발들, 그리고 핏물 위에 흰 쌀알들이 흩뿌려져 있습니다.",
    "hard_violations": [
     "[gpt-high] 얼굴을 보이지 말라는 명시적 지시와 달리 전경과 중경의 여러 시신 얼굴이 선명하게 노출됩니다.",
     "[gpt-high] 오른쪽 전경 시신은 돌담에 등을 대고 상체를 세운 앉은 자세로 보여, 앉아 있는 형상을 넣지 말라는 명시적 제한을 위반합니다."
    ],
    "physics": "시신과 주요 물건들은 바닥에 놓여 있으나, 바닥에 흩어진 일부 사발과 쌀알들이 핏물 표면 위에 다소 평면적으로 떠 있는 듯한 이질적인 질감을 줍니다."
   }
  ],
  "cross_model_order": {
   "policy": "cross_model_order_v2_slot0_fwd_slot1_rev_parallel2",
   "slots": [
    {
     "model": "gemini-pro",
     "order": "forward",
     "display_to_canonical": {
      "A": "A",
      "B": "B"
     },
     "raw": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 10,
        "verdict_ko": "레퍼런스에 제시된 목책과 임시 막사의 형태를 정확히 유지하여 장소의 연속성을 살렸으며, 뒤집힌 가마솥과 쏟아진 쌀 등의 디테일을 지시된 와이드 샷 구도 안에 완벽하게 구현했습니다."
       },
       {
        "label": "B",
        "score": 6,
        "verdict_ko": "쓰러진 의병들과 참혹한 전경을 잘 표현했으나, 레퍼런스에 존재하지 않던 흙벽 건물이 화면 좌측에 크게 등장하여 이전 샷과의 장소 일관성을 훼손했습니다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "카메라는 능선 위에서 캠프장을 비스듬히 내려다보는 앵글이며, 바닥에 쓰러진 시신들 사이로 연기가 자연스럽게 깔려 있습니다.",
        "built_space": "흙바닥으로 된 야영장이며, 배경에 레퍼런스 이미지와 동일한 형태의 수직 통나무 울타리와 거친 나무 기둥 및 천으로 엮은 막사들이 정확하게 배치되어 있습니다.",
        "entities": "피투성이가 된 20세기 초 조선인 의병들의 시신들이 다수 널브러져 있으며, 화면 좌측 하단에 뒤집힌 무쇠 가마솥, 깨진 사발들, 그리고 흙과 피에 섞여 쏟아진 흰 쌀이 명확하게 묘사되었습니다.",
        "hard_violations": [],
        "physics": "모든 시신과 총기, 가마솥, 사발 등의 사물들이 지면의 굴곡에 맞춰 흙바닥에 안정적으로 놓여 지지를 받고 있습니다."
       },
       {
        "label": "B",
        "direction": "카메라는 야영장을 대각선으로 바라보고 있으며, 전체적으로 연기가 흩뿌려져 있습니다.",
        "built_space": "흙바닥 야영장으로, 화면 우측에는 돌담이 있고 좌측에는 레퍼런스에서 볼 수 없었던 단단한 흙벽 구조의 초가집이 눈에 띄게 자리 잡고 있습니다.",
        "entities": "피투성이의 조선인 의병 시신들이 흩어져 있고, 중앙 부근에 엎어진 가마솥과 사발들, 그리고 핏물 위에 흰 쌀알들이 흩뿌려져 있습니다.",
        "hard_violations": [],
        "physics": "시신과 주요 물건들은 바닥에 놓여 있으나, 바닥에 흩어진 일부 사발과 쌀알들이 핏물 표면 위에 다소 평면적으로 떠 있는 듯한 이질적인 질감을 줍니다."
       }
      ],
      "all_candidates_fail": false
     },
     "normalized": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 10,
        "verdict_ko": "레퍼런스에 제시된 목책과 임시 막사의 형태를 정확히 유지하여 장소의 연속성을 살렸으며, 뒤집힌 가마솥과 쏟아진 쌀 등의 디테일을 지시된 와이드 샷 구도 안에 완벽하게 구현했습니다."
       },
       {
        "label": "B",
        "score": 6,
        "verdict_ko": "쓰러진 의병들과 참혹한 전경을 잘 표현했으나, 레퍼런스에 존재하지 않던 흙벽 건물이 화면 좌측에 크게 등장하여 이전 샷과의 장소 일관성을 훼손했습니다."
       }
      ],
      "all_candidates_fail": false,
      "readings": [
       {
        "label": "A",
        "direction": "카메라는 능선 위에서 캠프장을 비스듬히 내려다보는 앵글이며, 바닥에 쓰러진 시신들 사이로 연기가 자연스럽게 깔려 있습니다.",
        "built_space": "흙바닥으로 된 야영장이며, 배경에 레퍼런스 이미지와 동일한 형태의 수직 통나무 울타리와 거친 나무 기둥 및 천으로 엮은 막사들이 정확하게 배치되어 있습니다.",
        "entities": "피투성이가 된 20세기 초 조선인 의병들의 시신들이 다수 널브러져 있으며, 화면 좌측 하단에 뒤집힌 무쇠 가마솥, 깨진 사발들, 그리고 흙과 피에 섞여 쏟아진 흰 쌀이 명확하게 묘사되었습니다.",
        "hard_violations": [],
        "physics": "모든 시신과 총기, 가마솥, 사발 등의 사물들이 지면의 굴곡에 맞춰 흙바닥에 안정적으로 놓여 지지를 받고 있습니다."
       },
       {
        "label": "B",
        "direction": "카메라는 야영장을 대각선으로 바라보고 있으며, 전체적으로 연기가 흩뿌려져 있습니다.",
        "built_space": "흙바닥 야영장으로, 화면 우측에는 돌담이 있고 좌측에는 레퍼런스에서 볼 수 없었던 단단한 흙벽 구조의 초가집이 눈에 띄게 자리 잡고 있습니다.",
        "entities": "피투성이의 조선인 의병 시신들이 흩어져 있고, 중앙 부근에 엎어진 가마솥과 사발들, 그리고 핏물 위에 흰 쌀알들이 흩뿌려져 있습니다.",
        "hard_violations": [],
        "physics": "시신과 주요 물건들은 바닥에 놓여 있으나, 바닥에 흩어진 일부 사발과 쌀알들이 핏물 표면 위에 다소 평면적으로 떠 있는 듯한 이질적인 질감을 줍니다."
       }
      ]
     },
     "ok": true
    },
    {
     "model": "gpt-high",
     "order": "reverse",
     "display_to_canonical": {
      "A": "B",
      "B": "A"
     },
     "raw": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 3,
        "verdict_ko": "연기와 피투성이 시신의 전경은 구현했지만 얼굴과 앉은 듯한 시신을 노출하고, 능선 위 시점 및 참조 야영지의 구조에서도 벗어납니다."
       },
       {
        "label": "B",
        "score": 5,
        "verdict_ko": "세 채의 임시 막사와 목책, 지면에 흩어진 시신으로 장소와 참상을 더 충실히 잇지만, 얼굴 금지와 능선 위 사선 시점은 충족하지 못합니다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "시신들은 화면 깊이와 좌우로 서로 다른 방향으로 누워 있으며, 앞쪽 여러 얼굴이 카메라 쪽으로 드러납니다. 살아 있는 사람의 응시나 조준은 보이지 않습니다. 오른쪽 돌담에 기대 놓인 소총들의 총구는 위쪽을 향하고, 바닥의 총들은 지면과 거의 평행합니다. 카메라는 야영지 안에서 완만하게 내려다보며, 능선 위에서 전체 공터를 비스듬히 내려다보는 높이는 아닙니다.",
        "built_space": "왼쪽 전경에 창이 있는 흙벽 초가 한 채, 그 뒤로 연기에 가린 낮은 지붕 구조물들이 보이고, 오른쪽에는 길게 이어지는 돌담이 있습니다. 참조의 세 임시 막사와 뒤쪽 목책 배치보다 흙벽 건물과 돌담 사이의 통로처럼 보입니다. 시신은 통로 바닥과 벽 가장자리에 놓여 있으며 오른쪽 전경 한 구는 돌담에 상체를 기대고 있습니다. 가마솥 한 개와 분리된 둥근 뚜껑으로 보이는 물체가 중앙 전경에 있습니다. 반사는 없습니다.",
        "entities": "피 묻은 헐렁한 전통 복식과 머리띠를 착용한 성인 남성 시신 다수가 보이며 조선 의병이라는 설정에 어울립니다. 정확한 국적이나 개별 신원은 외형만으로 확정할 수 없습니다. 여러 얼굴이 뚜렷하게 노출됩니다. 연기, 쓰러진 남성들, 넘어진 검은 솥, 깨진 그릇과 온전한 그릇들, 흰 쌀, 흙 위 핏물이 모두 보입니다. 낮의 산악 배경과 낮은 채도도 구현되어 있습니다. 읽을 수 있는 글자는 없습니다.",
        "hard_violations": [
         "얼굴을 보이지 말라는 명시적 지시와 달리 전경과 중경의 여러 시신 얼굴이 선명하게 노출됩니다.",
         "오른쪽 전경 시신은 돌담에 등을 대고 상체를 세운 앉은 자세로 보여, 앉아 있는 형상을 넣지 말라는 명시적 제한을 위반합니다."
        ],
        "physics": "시신들의 몸통과 팔다리는 흙바닥, 돌 또는 인접한 몸에 닿아 있습니다. 오른쪽 반좌위 시신은 등 뒤 돌담과 바닥에 놓인 골반·다리로 지지됩니다. 넘어진 솥은 옆면과 돌출부가 지면에 닿고, 그릇·쌀·소총도 바닥이나 돌담에 놓여 있습니다. 떠 있는 몸이나 물체는 보이지 않으며, 핏물은 지면에 고여 있습니다."
       },
       {
        "label": "B",
        "direction": "시신들은 옆으로 웅크리거나 엎드린 채 여러 방향으로 분산되어 있습니다. 고개 대부분은 바닥을 향하지만 전경과 중경의 일부 얼굴은 카메라 방향으로 노출됩니다. 전경 소총의 총구는 화면 오른쪽을 향하고 오른쪽 가장자리의 기대 놓인 총들은 위쪽을 향하며, 누구도 조준하지 않습니다. 카메라는 공터 가장자리에서 약간 내려다보는 넓은 시야로, 능선 위의 뚜렷한 사선 조망은 아닙니다.",
        "built_space": "왼쪽 가장자리, 중앙 뒤쪽, 오른쪽 뒤쪽에 임시 막사 세 채가 있고 그 뒤를 뾰족한 목책이 연결합니다. 거친 목재 기둥, 늘어진 천, 낡은 초가지붕은 참조 야영지와 잘 이어집니다. 왼쪽 막사 앞에는 작은 작업대 하나, 왼쪽 전경에는 나무 상자와 넘어진 솥 하나가 보입니다. 시신들은 막사 사이의 연속된 흙 공터에 분산되어 있고 연기가 중앙 깊이를 나눕니다. 참조 전경의 큰 바위와 파인 지면은 이 구도에서 확인되지 않습니다. 반사는 없습니다.",
        "entities": "흰색·베이지색·짙은색의 낡은 전통 복식을 입은 성인 남성 시신 다수가 보이며 의병 집단 설정에 부합합니다. 얼굴과 머리 모양은 시대 배경에 어울리지만 개별 신원과 국적은 확정할 수 없습니다. 일부 얼굴이 식별 가능한 정도로 드러납니다. 넘어진 가마솥, 그릇과 파편, 흰 쌀, 흙과 피, 시신 사이의 연기가 모두 있습니다. 살아 움직이거나 서 있는 사람은 보이지 않습니다. 연기로 부드러워진 낮빛이며 읽을 수 있는 글자는 없습니다.",
        "hard_violations": [
         "얼굴을 보이지 말라는 명시적 지시와 달리 전경과 중경의 여러 시신 얼굴이 노출됩니다."
        ],
        "physics": "전경 시신들은 몸통·골반·다리와 팔을 바닥에 대고 있으며, 겹쳐진 일부 신체는 다른 시신 위에 놓여 있습니다. 중경의 높아 보이는 몸도 웅크린 몸통이나 주변 시신과 접촉합니다. 솥은 아래쪽 테두리와 옆면으로 지면에 기대고, 쌀은 솥 안과 바닥에 쌓여 있습니다. 소총·그릇·상자는 땅이나 구조물에 지지되며, 근거 없이 공중에 떠 있는 물체나 몸은 보이지 않습니다."
       }
      ],
      "all_candidates_fail": true
     },
     "normalized": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 3,
        "verdict_ko": "연기와 피투성이 시신의 전경은 구현했지만 얼굴과 앉은 듯한 시신을 노출하고, 능선 위 시점 및 참조 야영지의 구조에서도 벗어납니다."
       },
       {
        "label": "A",
        "score": 5,
        "verdict_ko": "세 채의 임시 막사와 목책, 지면에 흩어진 시신으로 장소와 참상을 더 충실히 잇지만, 얼굴 금지와 능선 위 사선 시점은 충족하지 못합니다."
       }
      ],
      "all_candidates_fail": true,
      "readings": [
       {
        "label": "B",
        "direction": "시신들은 화면 깊이와 좌우로 서로 다른 방향으로 누워 있으며, 앞쪽 여러 얼굴이 카메라 쪽으로 드러납니다. 살아 있는 사람의 응시나 조준은 보이지 않습니다. 오른쪽 돌담에 기대 놓인 소총들의 총구는 위쪽을 향하고, 바닥의 총들은 지면과 거의 평행합니다. 카메라는 야영지 안에서 완만하게 내려다보며, 능선 위에서 전체 공터를 비스듬히 내려다보는 높이는 아닙니다.",
        "built_space": "왼쪽 전경에 창이 있는 흙벽 초가 한 채, 그 뒤로 연기에 가린 낮은 지붕 구조물들이 보이고, 오른쪽에는 길게 이어지는 돌담이 있습니다. 참조의 세 임시 막사와 뒤쪽 목책 배치보다 흙벽 건물과 돌담 사이의 통로처럼 보입니다. 시신은 통로 바닥과 벽 가장자리에 놓여 있으며 오른쪽 전경 한 구는 돌담에 상체를 기대고 있습니다. 가마솥 한 개와 분리된 둥근 뚜껑으로 보이는 물체가 중앙 전경에 있습니다. 반사는 없습니다.",
        "entities": "피 묻은 헐렁한 전통 복식과 머리띠를 착용한 성인 남성 시신 다수가 보이며 조선 의병이라는 설정에 어울립니다. 정확한 국적이나 개별 신원은 외형만으로 확정할 수 없습니다. 여러 얼굴이 뚜렷하게 노출됩니다. 연기, 쓰러진 남성들, 넘어진 검은 솥, 깨진 그릇과 온전한 그릇들, 흰 쌀, 흙 위 핏물이 모두 보입니다. 낮의 산악 배경과 낮은 채도도 구현되어 있습니다. 읽을 수 있는 글자는 없습니다.",
        "hard_violations": [
         "얼굴을 보이지 말라는 명시적 지시와 달리 전경과 중경의 여러 시신 얼굴이 선명하게 노출됩니다.",
         "오른쪽 전경 시신은 돌담에 등을 대고 상체를 세운 앉은 자세로 보여, 앉아 있는 형상을 넣지 말라는 명시적 제한을 위반합니다."
        ],
        "physics": "시신들의 몸통과 팔다리는 흙바닥, 돌 또는 인접한 몸에 닿아 있습니다. 오른쪽 반좌위 시신은 등 뒤 돌담과 바닥에 놓인 골반·다리로 지지됩니다. 넘어진 솥은 옆면과 돌출부가 지면에 닿고, 그릇·쌀·소총도 바닥이나 돌담에 놓여 있습니다. 떠 있는 몸이나 물체는 보이지 않으며, 핏물은 지면에 고여 있습니다."
       },
       {
        "label": "A",
        "direction": "시신들은 옆으로 웅크리거나 엎드린 채 여러 방향으로 분산되어 있습니다. 고개 대부분은 바닥을 향하지만 전경과 중경의 일부 얼굴은 카메라 방향으로 노출됩니다. 전경 소총의 총구는 화면 오른쪽을 향하고 오른쪽 가장자리의 기대 놓인 총들은 위쪽을 향하며, 누구도 조준하지 않습니다. 카메라는 공터 가장자리에서 약간 내려다보는 넓은 시야로, 능선 위의 뚜렷한 사선 조망은 아닙니다.",
        "built_space": "왼쪽 가장자리, 중앙 뒤쪽, 오른쪽 뒤쪽에 임시 막사 세 채가 있고 그 뒤를 뾰족한 목책이 연결합니다. 거친 목재 기둥, 늘어진 천, 낡은 초가지붕은 참조 야영지와 잘 이어집니다. 왼쪽 막사 앞에는 작은 작업대 하나, 왼쪽 전경에는 나무 상자와 넘어진 솥 하나가 보입니다. 시신들은 막사 사이의 연속된 흙 공터에 분산되어 있고 연기가 중앙 깊이를 나눕니다. 참조 전경의 큰 바위와 파인 지면은 이 구도에서 확인되지 않습니다. 반사는 없습니다.",
        "entities": "흰색·베이지색·짙은색의 낡은 전통 복식을 입은 성인 남성 시신 다수가 보이며 의병 집단 설정에 부합합니다. 얼굴과 머리 모양은 시대 배경에 어울리지만 개별 신원과 국적은 확정할 수 없습니다. 일부 얼굴이 식별 가능한 정도로 드러납니다. 넘어진 가마솥, 그릇과 파편, 흰 쌀, 흙과 피, 시신 사이의 연기가 모두 있습니다. 살아 움직이거나 서 있는 사람은 보이지 않습니다. 연기로 부드러워진 낮빛이며 읽을 수 있는 글자는 없습니다.",
        "hard_violations": [
         "얼굴을 보이지 말라는 명시적 지시와 달리 전경과 중경의 여러 시신 얼굴이 노출됩니다."
        ],
        "physics": "전경 시신들은 몸통·골반·다리와 팔을 바닥에 대고 있으며, 겹쳐진 일부 신체는 다른 시신 위에 놓여 있습니다. 중경의 높아 보이는 몸도 웅크린 몸통이나 주변 시신과 접촉합니다. 솥은 아래쪽 테두리와 옆면으로 지면에 기대고, 쌀은 솥 안과 바닥에 쌓여 있습니다. 소총·그릇·상자는 땅이나 구조물에 지지되며, 근거 없이 공중에 떠 있는 물체나 몸은 보이지 않습니다."
       }
      ]
     },
     "ok": true
    }
   ],
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "slot_winner_match": true,
   "slot_winner": {
    "gemini-pro": "A",
    "gpt-high": "A"
   },
   "route": "cross_slot_agree"
  },
  "dual": {
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "normalized": {
    "A": 2.0,
    "B": 1.2
   },
   "adjusted": {
    "A": 1.75,
    "B": 0.95
   },
   "violations": {
    "B": [
     "[gpt-high] 얼굴을 보이지 말라는 명시적 지시와 달리 전경과 중경의 여러 시신 얼굴이 선명하게 노출됩니다.",
     "[gpt-high] 오른쪽 전경 시신은 돌담에 등을 대고 상체를 세운 앉은 자세로 보여, 앉아 있는 형상을 넣지 말라는 명시적 제한을 위반합니다."
    ],
    "A": [
     "[gpt-high] 얼굴을 보이지 말라는 명시적 지시와 달리 전경과 중경의 여러 시신 얼굴이 노출됩니다."
    ]
   },
   "per_model_winner": {
    "gemini-pro": "A",
    "gpt-high": "A"
   },
   "agreed": true
  },
  "totals": {
   "A": 1750,
   "B": 950
  },
  "selected": "A",
  "ranking": [
   "A",
   "B"
  ],
  "verdicts": [
   {
    "label": "A",
    "score": 1750,
    "verdict_ko": "레퍼런스에 제시된 목책과 임시 막사의 형태를 정확히 유지하여 장소의 연속성을 살렸으며, 뒤집힌 가마솥과 쏟아진 쌀 등의 디테일을 지시된 와이드 샷 구도 안에 완벽하게 구현했습니다.  ★위반: [gpt-high] 얼굴을 보이지 말라는 명시적 지시와 달리 전경과 중경의 여러 시신 얼굴이 노출됩니다."
   },
   {
    "label": "B",
    "score": 950,
    "verdict_ko": "쓰러진 의병들과 참혹한 전경을 잘 표현했으나, 레퍼런스에 존재하지 않던 흙벽 건물이 화면 좌측에 크게 등장하여 이전 샷과의 장소 일관성을 훼손했습니다.  ★위반: [gpt-high] 얼굴을 보이지 말라는 명시적 지시와 달리 전경과 중경의 여러 시신 얼굴이 선명하게 노출됩니다. / [gpt-high] 오른쪽 전경 시신은 돌담에 등을 대고 상체를 세운 앉은 자세로 보여, 앉아 있는 형상을 넣지 말라는 명시적 제한을 위반합니다."
   }
  ],
  "refs": [
   {
    "label": "PREVIOUS SHOT STILL — a visually related earlier shot of this same place: the location's look, materials, fixed features and lighting mood are LOCKED to this photo; never copy its camera framing. The people visible in that photo are NOT in this shot: never carry their faces, bodies or clothing onto anyone here. Each person in THIS shot is defined solely by the PEOPLE list, their own CHARACTER REFERENCE images and their wardrobe notes. If this photo shows a character who cannot move (dead or unconscious), that character's exact pose, position and orientation are ALSO LOCKED — treat their body as a fixed prop of the set that only the camera moves around.",
    "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/db00517b-165b-4005-a287-f28936b59b95/images/272bd7af-9294-4490-98c6-88a40e9deacc/scene/recipe/S56sh40_sel.png",
    "asset_id": "9d1cd195-667b-40af-ad5e-5755f6a59694",
    "role": "prev_still"
   }
  ],
  "critique_skipped": true,
  "shot_run_produced": false,
  "shot_run_uid": "06aa1441-27d6-73ba-a07d-ed3bbea7ec59",
  "ref_mode": "prev만 (배경 전용·공유 계획)",
  "share_plan": {
   "ref_plan": "prev",
   "prev_anchor_tag": "S56sh40"
  },
  "lane_policy": "ab_select_bypass:bg_only:share_plan_prev_bgonly"
 },
 "S60sh2::cine": {
  "applied": false,
  "attempted_at": "2026-09-09T09:34:48.557004+00:00",
  "fingerprint": "7424268e030f6c487a6419de0bd24d512e14064d856df3c5d3be01c6a42002a6",
  "fingerprint_version": 2,
  "provider": "mai",
  "endpoint": "openrouter/chat-completions",
  "model": "microsoft/mai-image-2.6",
  "multiroll_tag": "still_S60sh2_cine",
  "slot": "",
  "pack": "24.202608252115",
  "source_file": "S60sh2_sel.png",
  "source_sha256": "978fbbcbd84fd11dd411f6d34fe361a8e06949b63164f53e1b733b8ac130f669",
  "error": "RuntimeError: moderation blocked: {\"message\": \"Response content blocked by label 'MultiSeverity_ViolenceScore'.\", \"code\": 400}",
  "moderation_refusals": 2,
  "declined": true,
  "declined_reason": "moderation"
 },
 "S60sh11::signage": {
  "fp": "67cde80c9561edc1",
  "inscriptions": [],
  "cues": [],
  "dropped": []
 },
 "S60sh11": {
  "input_fingerprint": "4b848c8f29d76f24",
  "prompt": "Create ONE FINAL photorealistic live-action film still of the moment below — early 20th-century Korea, Jeolla region; all people are Korean unless explicitly stated otherwise. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 일본인 군인이 뒤집힌 시신의 가슴팍에 소총 끝 총검을 무자비하게 푹 찌른 찰나.\n\nLOCATION (lock): The corpse-strewn dirt clearing among the mountain camp's damaged huts and surrounding trees. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: wide shot\n- FRAME LAYOUT: Concealing tree trunk kept clear of bayonet contact in the middle-left of the frame, foreground.\n- KEY BACKGROUND ELEMENTS: Concealing tree trunk (Separating the hidden camera position from the soldier) — Its edge occupies a narrow strip at frame left; used as Establishes concealment without blocking the chest or weapon contact; Rifle and fixed bayonet (Driven into the overturned fighter's chest) — Runs diagonally downward from the soldier's hands toward the body; used as Connects the soldier's force to the unresponsive body at a legible distance; Camp ground (Supports the overturned body); used as Maintains the shared physical space of the soldier and casualty.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Restrained daylight and naturalistic contrast keep the action readable without theatrical emphasis on the wound.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The people visible in that photo are NOT in this shot: never carry their faces, bodies or clothing onto anyone here. Each person in THIS shot is defined solely by the PEOPLE list, their own CHARACTER REFERENCE images and their wardrobe notes. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nPREVIOUS STILL USAGE (follow exactly — what to take from the attached still and what to exclude; it governs place and objects only, never who is in this shot or how the camera sees them): Take the damaged camp structures, broken bowls, scattered rice, and bloodied dirt visible in the reference. Exclude intact meal arrangements and any soldiers who have already withdrawn.\n\nIMMOBILE CHARACTER POSE — CANONICAL (identical wherever this character appears in ANY panel; on any conflict THIS POSE WINS): The unidentified militia corpse rests motionless on the camp ground in the position into which a Japanese soldier has just turned it. The screenplay does not specify the resulting torso orientation, head direction, or arrangement of the arms and legs.\n\nIMMOBILE BODIES OBEY GRAVITY: a person who is dead or unconscious\nexerts NO muscular effort. Every part of their body — head, torso,\narms, hands, fingers, legs — rests fully on whatever supports it\n(floor, wall, furniture, their own lap) and hangs or slumps with\ngravity. NEVER show any part of an immobile person's body lifted,\nraised, held up in the air, or posed as if presenting something:\nan object in their grip stays clenched in a hand that itself lies\nfallen on a support — the hand does not hold the object up. If the\ncanonical pose leaves a body part unspecified, resolve it as the\nmost gravity-compliant, fully-supported position.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The smoky ruins retain the overturned pot, shattered bowls, bloodied rice and fallen timber.\n\nTHE HAND THAT IS DOING THIS: the object at the centre of this shot is being held, operated, read, aimed or handed over by 일본인 군인 right now, so 일본인 군인's hand — and as much of the wrist and forearm as the framing reaches — is in the frame, gripping or resting on that object exactly the way its real use requires. Match that hand to 일본인 군인: its size, build, skin, age, grooming, sleeve and anything worn on it belong to that person and to no one else.\n\nNo people appear unless the shot text itself says so.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.",
  "roll_prompts": {
   "A": "Create ONE FINAL photorealistic live-action film still of the moment below — early 20th-century Korea, Jeolla region; all people are Korean unless explicitly stated otherwise. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 일본인 군인이 뒤집힌 시신의 가슴팍에 소총 끝 총검을 무자비하게 푹 찌른 찰나.\n\nLOCATION (lock): The corpse-strewn dirt clearing among the mountain camp's damaged huts and surrounding trees. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: wide shot\n- FRAME LAYOUT: Concealing tree trunk kept clear of bayonet contact in the middle-left of the frame, foreground.\n- KEY BACKGROUND ELEMENTS: Concealing tree trunk (Separating the hidden camera position from the soldier) — Its edge occupies a narrow strip at frame left; used as Establishes concealment without blocking the chest or weapon contact; Rifle and fixed bayonet (Driven into the overturned fighter's chest) — Runs diagonally downward from the soldier's hands toward the body; used as Connects the soldier's force to the unresponsive body at a legible distance; Camp ground (Supports the overturned body); used as Maintains the shared physical space of the soldier and casualty.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Restrained daylight and naturalistic contrast keep the action readable without theatrical emphasis on the wound.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The people visible in that photo are NOT in this shot: never carry their faces, bodies or clothing onto anyone here. Each person in THIS shot is defined solely by the PEOPLE list, their own CHARACTER REFERENCE images and their wardrobe notes. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nPREVIOUS STILL USAGE (follow exactly — what to take from the attached still and what to exclude; it governs place and objects only, never who is in this shot or how the camera sees them): Take the damaged camp structures, broken bowls, scattered rice, and bloodied dirt visible in the reference. Exclude intact meal arrangements and any soldiers who have already withdrawn.\n\nIMMOBILE CHARACTER POSE — CANONICAL (identical wherever this character appears in ANY panel; on any conflict THIS POSE WINS): The unidentified militia corpse rests motionless on the camp ground in the position into which a Japanese soldier has just turned it. The screenplay does not specify the resulting torso orientation, head direction, or arrangement of the arms and legs.\n\nIMMOBILE BODIES OBEY GRAVITY: a person who is dead or unconscious\nexerts NO muscular effort. Every part of their body — head, torso,\narms, hands, fingers, legs — rests fully on whatever supports it\n(floor, wall, furniture, their own lap) and hangs or slumps with\ngravity. NEVER show any part of an immobile person's body lifted,\nraised, held up in the air, or posed as if presenting something:\nan object in their grip stays clenched in a hand that itself lies\nfallen on a support — the hand does not hold the object up. If the\ncanonical pose leaves a body part unspecified, resolve it as the\nmost gravity-compliant, fully-supported position.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The smoky ruins retain the overturned pot, shattered bowls, bloodied rice and fallen timber.\n\nTHE HAND THAT IS DOING THIS: the object at the centre of this shot is being held, operated, read, aimed or handed over by 일본인 군인 right now, so 일본인 군인's hand — and as much of the wrist and forearm as the framing reaches — is in the frame, gripping or resting on that object exactly the way its real use requires. Match that hand to 일본인 군인: its size, build, skin, age, grooming, sleeve and anything worn on it belong to that person and to no one else.\n\nNo people appear unless the shot text itself says so.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.",
   "B": "Create ONE FINAL photorealistic live-action film still of the moment below — early 20th-century Korea, Jeolla region; all people are Korean unless explicitly stated otherwise. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 일본인 군인이 뒤집힌 시신의 가슴팍에 소총 끝 총검을 무자비하게 푹 찌른 찰나.\n\nLOCATION (lock): The corpse-strewn dirt clearing among the mountain camp's damaged huts and surrounding trees. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: wide shot\n- FRAME LAYOUT: Concealing tree trunk kept clear of bayonet contact in the middle-left of the frame, foreground.\n- KEY BACKGROUND ELEMENTS: Concealing tree trunk (Separating the hidden camera position from the soldier) — Its edge occupies a narrow strip at frame left; used as Establishes concealment without blocking the chest or weapon contact; Rifle and fixed bayonet (Driven into the overturned fighter's chest) — Runs diagonally downward from the soldier's hands toward the body; used as Connects the soldier's force to the unresponsive body at a legible distance; Camp ground (Supports the overturned body); used as Maintains the shared physical space of the soldier and casualty.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Restrained daylight and naturalistic contrast keep the action readable without theatrical emphasis on the wound.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The people visible in that photo are NOT in this shot: never carry their faces, bodies or clothing onto anyone here. Each person in THIS shot is defined solely by the PEOPLE list, their own CHARACTER REFERENCE images and their wardrobe notes. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nPREVIOUS STILL USAGE (follow exactly — what to take from the attached still and what to exclude; it governs place and objects only, never who is in this shot or how the camera sees them): Take the damaged camp structures, broken bowls, scattered rice, and bloodied dirt visible in the reference. Exclude intact meal arrangements and any soldiers who have already withdrawn.\n\nIMMOBILE CHARACTER POSE — CANONICAL (identical wherever this character appears in ANY panel; on any conflict THIS POSE WINS): The unidentified militia corpse rests motionless on the camp ground in the position into which a Japanese soldier has just turned it. The screenplay does not specify the resulting torso orientation, head direction, or arrangement of the arms and legs.\n\nIMMOBILE BODIES OBEY GRAVITY: a person who is dead or unconscious\nexerts NO muscular effort. Every part of their body — head, torso,\narms, hands, fingers, legs — rests fully on whatever supports it\n(floor, wall, furniture, their own lap) and hangs or slumps with\ngravity. NEVER show any part of an immobile person's body lifted,\nraised, held up in the air, or posed as if presenting something:\nan object in their grip stays clenched in a hand that itself lies\nfallen on a support — the hand does not hold the object up. If the\ncanonical pose leaves a body part unspecified, resolve it as the\nmost gravity-compliant, fully-supported position.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The smoky ruins retain the overturned pot, shattered bowls, bloodied rice and fallen timber.\n\nTHE HAND THAT IS DOING THIS: the object at the centre of this shot is being held, operated, read, aimed or handed over by 일본인 군인 right now, so 일본인 군인's hand — and as much of the wrist and forearm as the framing reaches — is in the frame, gripping or resting on that object exactly the way its real use requires. Match that hand to 일본인 군인: its size, build, skin, age, grooming, sleeve and anything worn on it belong to that person and to no one else.\n\nNo people appear unless the shot text itself says so.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.\n\nCOMPOSITION VARIATION (this candidate only): keep the place, the people and the moment exactly as specified, and keep from the CAMERA & FRAME contract its subject scale, screen placement and key background elements exactly. For this candidate only, its camera angle and height are deliberately left open: choose a camera position distinctly different from the obvious one, as a film director picking a second setup on the same blocking. Stay true to any explicit framing the SHOT TEXT itself dictates."
  },
  "shot_run_spend_attempt_count": 1,
  "initial_roll_all_fail": true,
  "readings": [
   {
    "label": "A",
    "direction": "일본인 군인은 소총을 옆으로 들고 서 있으며, 바닥에 쓰러진 사람을 찌르지 않고 있다. 바닥에 있는 남자는 고개를 들고 군인을 바라보고 있다.",
    "built_space": "왼쪽 전경에 나무 기둥이 배치되어 있고, 뒤집힌 솥과 쏟아진 쌀이 재현되어 있다. 배경의 파괴된 초가집과 야영지 구조물도 레퍼런스와 부합한다.",
    "entities": "일본군 복장의 군인 1명, 바닥에 쓰러진 한복 차림의 남성 1명이 있다. 하지만 바닥의 남성은 시신이어야 함에도 살아있는 모습이다.",
    "hard_violations": [
     "[gemini-pro] 명시된 핵심 행동(총검으로 가슴을 찌름)이 전혀 나타나지 않음",
     "[gemini-pro] 시신이 고개를 들고 팔로 버티며 살아 움직임 (중력/부동 상태 위반)",
     "[gpt-high] 시신이어야 할 인물이 팔로 상체를 받치고 머리를 들어 군인을 바라보도록 배치되어, 근육을 쓰지 않고 완전히 지지되어야 하는 사망 자세를 위반한다."
    ],
    "physics": "시신이어야 할 남성이 스스로의 근육을 사용해 팔로 상체를 지탱하고 고개를 들고 있어 '무생물적 중력 순응' 규칙을 위반했다."
   },
   {
    "label": "B",
    "direction": "두 명의 일본인 군인이 서 있으며 아무도 바닥의 남자를 찌르지 않고 있다. 오른쪽 군인의 소총은 어깨에 메어져 있다. 바닥의 남자는 고개를 들고 위를 향해 시선을 두고 있다.",
    "built_space": "왼쪽 전경에 비스듬한 나무 기둥이 있고, 쏟아진 쌀과 가마솥이 있다. 야영지 배경은 지시와 대체로 일치한다.",
    "entities": "일본군 2명이 등장하여 인물이 중복/추가되었으며, 바닥에 쓰러진 남성 1명 역시 시신이 아닌 살아있는 모습이다.",
    "hard_violations": [
     "[gemini-pro] 지시에 없는 불필요한 인물 추가 (일본군이 2명임)",
     "[gemini-pro] 명시된 핵심 행동(총검으로 가슴을 찌름) 누락",
     "[gemini-pro] 시신이 고개를 들고 버티며 살아 있음 (중력/부동 상태 위반)",
     "[gpt-high] 장면에 지정되지 않은 군인 한 명이 추가되어 총 세 사람이 등장한다.",
     "[gpt-high] 시신이어야 할 인물이 머리와 상체를 능동적으로 세우고 군인을 올려다보는 자세로 배치되어, 완전히 지지된 무반응 사망 자세를 위반한다."
    ],
    "physics": "바닥에 누운 남성이 상체를 일으키고 고개를 들고 있어 죽은 사람의 중력 지침을 완전히 위반했다."
   }
  ],
  "cross_model_order": {
   "policy": "cross_model_order_v2_slot0_fwd_slot1_rev_parallel2",
   "slots": [
    {
     "model": "gemini-pro",
     "order": "forward",
     "display_to_canonical": {
      "A": "A",
      "B": "B"
     },
     "raw": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 1,
        "verdict_ko": "시신을 향해 총검을 찌르는 핵심 행동이 전혀 묘사되지 않았으며, 시신이어야 할 인물이 고개를 들고 살아 있어 완전히 실패한 결과물입니다."
       },
       {
        "label": "B",
        "score": 0,
        "verdict_ko": "핵심 행동인 총검 찌르기가 누락된 데다 지시에 없는 일본군이 한 명 더 추가되었고, 시신이 살아 움직여 사용할 수 없는 이미지입니다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "일본인 군인은 소총을 옆으로 들고 서 있으며, 바닥에 쓰러진 사람을 찌르지 않고 있다. 바닥에 있는 남자는 고개를 들고 군인을 바라보고 있다.",
        "built_space": "왼쪽 전경에 나무 기둥이 배치되어 있고, 뒤집힌 솥과 쏟아진 쌀이 재현되어 있다. 배경의 파괴된 초가집과 야영지 구조물도 레퍼런스와 부합한다.",
        "entities": "일본군 복장의 군인 1명, 바닥에 쓰러진 한복 차림의 남성 1명이 있다. 하지만 바닥의 남성은 시신이어야 함에도 살아있는 모습이다.",
        "hard_violations": [
         "명시된 핵심 행동(총검으로 가슴을 찌름)이 전혀 나타나지 않음",
         "시신이 고개를 들고 팔로 버티며 살아 움직임 (중력/부동 상태 위반)"
        ],
        "physics": "시신이어야 할 남성이 스스로의 근육을 사용해 팔로 상체를 지탱하고 고개를 들고 있어 '무생물적 중력 순응' 규칙을 위반했다."
       },
       {
        "label": "B",
        "direction": "두 명의 일본인 군인이 서 있으며 아무도 바닥의 남자를 찌르지 않고 있다. 오른쪽 군인의 소총은 어깨에 메어져 있다. 바닥의 남자는 고개를 들고 위를 향해 시선을 두고 있다.",
        "built_space": "왼쪽 전경에 비스듬한 나무 기둥이 있고, 쏟아진 쌀과 가마솥이 있다. 야영지 배경은 지시와 대체로 일치한다.",
        "entities": "일본군 2명이 등장하여 인물이 중복/추가되었으며, 바닥에 쓰러진 남성 1명 역시 시신이 아닌 살아있는 모습이다.",
        "hard_violations": [
         "지시에 없는 불필요한 인물 추가 (일본군이 2명임)",
         "명시된 핵심 행동(총검으로 가슴을 찌름) 누락",
         "시신이 고개를 들고 버티며 살아 있음 (중력/부동 상태 위반)"
        ],
        "physics": "바닥에 누운 남성이 상체를 일으키고 고개를 들고 있어 죽은 사람의 중력 지침을 완전히 위반했다."
       }
      ],
      "all_candidates_fail": true
     },
     "normalized": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 1,
        "verdict_ko": "시신을 향해 총검을 찌르는 핵심 행동이 전혀 묘사되지 않았으며, 시신이어야 할 인물이 고개를 들고 살아 있어 완전히 실패한 결과물입니다."
       },
       {
        "label": "B",
        "score": 0,
        "verdict_ko": "핵심 행동인 총검 찌르기가 누락된 데다 지시에 없는 일본군이 한 명 더 추가되었고, 시신이 살아 움직여 사용할 수 없는 이미지입니다."
       }
      ],
      "all_candidates_fail": true,
      "readings": [
       {
        "label": "A",
        "direction": "일본인 군인은 소총을 옆으로 들고 서 있으며, 바닥에 쓰러진 사람을 찌르지 않고 있다. 바닥에 있는 남자는 고개를 들고 군인을 바라보고 있다.",
        "built_space": "왼쪽 전경에 나무 기둥이 배치되어 있고, 뒤집힌 솥과 쏟아진 쌀이 재현되어 있다. 배경의 파괴된 초가집과 야영지 구조물도 레퍼런스와 부합한다.",
        "entities": "일본군 복장의 군인 1명, 바닥에 쓰러진 한복 차림의 남성 1명이 있다. 하지만 바닥의 남성은 시신이어야 함에도 살아있는 모습이다.",
        "hard_violations": [
         "명시된 핵심 행동(총검으로 가슴을 찌름)이 전혀 나타나지 않음",
         "시신이 고개를 들고 팔로 버티며 살아 움직임 (중력/부동 상태 위반)"
        ],
        "physics": "시신이어야 할 남성이 스스로의 근육을 사용해 팔로 상체를 지탱하고 고개를 들고 있어 '무생물적 중력 순응' 규칙을 위반했다."
       },
       {
        "label": "B",
        "direction": "두 명의 일본인 군인이 서 있으며 아무도 바닥의 남자를 찌르지 않고 있다. 오른쪽 군인의 소총은 어깨에 메어져 있다. 바닥의 남자는 고개를 들고 위를 향해 시선을 두고 있다.",
        "built_space": "왼쪽 전경에 비스듬한 나무 기둥이 있고, 쏟아진 쌀과 가마솥이 있다. 야영지 배경은 지시와 대체로 일치한다.",
        "entities": "일본군 2명이 등장하여 인물이 중복/추가되었으며, 바닥에 쓰러진 남성 1명 역시 시신이 아닌 살아있는 모습이다.",
        "hard_violations": [
         "지시에 없는 불필요한 인물 추가 (일본군이 2명임)",
         "명시된 핵심 행동(총검으로 가슴을 찌름) 누락",
         "시신이 고개를 들고 버티며 살아 있음 (중력/부동 상태 위반)"
        ],
        "physics": "바닥에 누운 남성이 상체를 일으키고 고개를 들고 있어 죽은 사람의 중력 지침을 완전히 위반했다."
       }
      ]
     },
     "ok": true
    },
    {
     "model": "gpt-high",
     "order": "reverse",
     "display_to_canonical": {
      "A": "B",
      "B": "A"
     },
     "raw": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 0,
        "verdict_ko": "군인이 한 명 추가되었고 시신이 살아 있는 듯 상체를 세우며, 총검으로 가슴을 찌르는 핵심 동작이 전혀 없다."
       },
       {
        "label": "B",
        "score": 2,
        "verdict_ko": "두 사람의 인원 구성과 왼쪽 나무를 둔 와이드 구도는 더 가깝지만, 소총은 등에 걸려 있고 시신이 고개를 들어 찌르는 순간과 사망 자세를 모두 어긴다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "왼쪽 군인은 바닥의 남자 쪽으로 내려다보고, 오른쪽 군인은 왼쪽 인물들 쪽을 본다. 누운 남자는 고개를 들어 왼쪽 군인을 올려다본다. 오른쪽 군인의 소총 총구는 등 뒤에서 위를 향하며, 시신의 가슴을 향하는 총검이나 무기 접촉은 없다.",
        "built_space": "왼쪽·중앙·오른쪽에 초가 구조물 세 채가 보이고 뒤에는 목책과 산림이 있다. 앞쪽 왼편에는 기울어진 큰 솥 하나, 여러 그릇과 파편, 쏟아진 쌀이 있어 참조 장소의 재료와 배치를 대체로 따른다. 군인 둘은 바닥의 남자 양옆에 서 있다. 전경 나무는 왼쪽의 좁은 띠가 아니라 굵은 갈래가 화면 중앙 위쪽까지 뻗어 은폐 구도를 지나치게 크게 차지한다.",
        "entities": "성인 동아시아계 외형의 남성 세 명으로, 군복 차림 둘과 흰옷·검은 조끼 차림 하나가 보인다. 일본군 역할에 부합하는 군복은 있지만 국적은 외형만으로 확인할 수 없다. 요구된 군인 한 명보다 한 명 많다. 바닥의 남자는 눈을 뜨고 상대를 올려다보아 무반응 시신으로 읽히지 않는다. 소총은 보이지만 가슴에 박힌 고정 총검은 없다. 피 묻은 흙과 쌀, 솥, 그릇 파편, 연기 낀 초가 배경은 보이며 읽을 수 있는 글자는 없다.",
        "hard_violations": [
         "장면에 지정되지 않은 군인 한 명이 추가되어 총 세 사람이 등장한다.",
         "시신이어야 할 인물이 머리와 상체를 능동적으로 세우고 군인을 올려다보는 자세로 배치되어, 완전히 지지된 무반응 사망 자세를 위반한다."
        ],
        "physics": "군인들의 발은 땅에 닿고 소총은 오른쪽 군인의 어깨끈으로 지지된다. 바닥 남자의 다리와 골반은 땅에 놓이고 굽힌 팔 주변이 상체를 받치지만, 들린 머리에는 외부 받침이 없고 목과 몸통의 능동적인 힘이 필요하다. 이는 시신의 이완된 자세가 아니다. 누구도 소총을 양손으로 잡아 내려찌르지 않으므로 가슴으로 전달되는 힘이나 충격도 보이지 않는다."
       },
       {
        "label": "B",
        "direction": "군인은 화면 오른쪽 아래의 누운 남자를 바라보고, 그 남자는 왼쪽 위의 군인을 올려다본다. 소총은 군인의 등 뒤에서 오른쪽 위로 향한다. 총구와 총검이 남자의 가슴을 겨누지 않으며, 둘 사이에 무기 접촉도 없다.",
        "built_space": "초가 구조물 세 채가 왼쪽 가장자리·중앙 뒤·오른쪽에 있고, 뒤로 목책과 산림이 이어진다. 큰 솥 하나와 여러 그릇·파편, 흩어진 쌀, 목재 잔해가 흙바닥에 있다. 왼쪽 전경 나무는 가장자리의 세로 띠로 남아 두 사람을 가리지 않는다. 군인은 중앙 왼쪽, 누운 남자는 오른쪽 아래에 있어 와이드 숏과 은폐 시점에는 A보다 가깝지만, 가슴을 찌를 수 있는 근접 배치는 아니다.",
        "entities": "성인 동아시아계 외형의 남성 두 명으로, 군복 차림 한 명과 때 묻은 전통 복식 차림 한 명이 보여 요구 인원은 맞는다. 군복은 일본군 역할을 표현하지만 국적 자체는 외형으로 확정할 수 없다. 누운 남자는 눈을 뜨고 고개와 상체를 세워 시신보다는 살아 있는 부상자로 보인다. 소총은 등에 걸려 있으며 가슴에 꽂힌 총검은 없다. 솥, 깨진 그릇, 쌀, 혈흔성 얼룩과 연기 낀 진영은 보이고 읽을 수 있는 글자는 없다.",
        "hard_violations": [
         "시신이어야 할 인물이 팔로 상체를 받치고 머리를 들어 군인을 바라보도록 배치되어, 근육을 쓰지 않고 완전히 지지되어야 하는 사망 자세를 위반한다."
        ],
        "physics": "군인은 양발로 땅을 딛고 있다. 소총은 몸을 가로지르는 멜빵으로 지지되고 한 손이 아래쪽 개머리판 부근에 닿아 있어 부유하지 않는다. 누운 남자의 다리와 골반, 아래쪽 팔은 땅과 천에 지지되지만 들린 머리를 받치는 물체는 없으며 목의 근력이 필요하다. 시신에게 허용되지 않는 능동 자세다. 군인의 몸과 소총에는 아래로 찌르는 힘이나 가슴 접촉이 전혀 나타나지 않는다."
       }
      ],
      "all_candidates_fail": true
     },
     "normalized": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 0,
        "verdict_ko": "군인이 한 명 추가되었고 시신이 살아 있는 듯 상체를 세우며, 총검으로 가슴을 찌르는 핵심 동작이 전혀 없다."
       },
       {
        "label": "A",
        "score": 2,
        "verdict_ko": "두 사람의 인원 구성과 왼쪽 나무를 둔 와이드 구도는 더 가깝지만, 소총은 등에 걸려 있고 시신이 고개를 들어 찌르는 순간과 사망 자세를 모두 어긴다."
       }
      ],
      "all_candidates_fail": true,
      "readings": [
       {
        "label": "B",
        "direction": "왼쪽 군인은 바닥의 남자 쪽으로 내려다보고, 오른쪽 군인은 왼쪽 인물들 쪽을 본다. 누운 남자는 고개를 들어 왼쪽 군인을 올려다본다. 오른쪽 군인의 소총 총구는 등 뒤에서 위를 향하며, 시신의 가슴을 향하는 총검이나 무기 접촉은 없다.",
        "built_space": "왼쪽·중앙·오른쪽에 초가 구조물 세 채가 보이고 뒤에는 목책과 산림이 있다. 앞쪽 왼편에는 기울어진 큰 솥 하나, 여러 그릇과 파편, 쏟아진 쌀이 있어 참조 장소의 재료와 배치를 대체로 따른다. 군인 둘은 바닥의 남자 양옆에 서 있다. 전경 나무는 왼쪽의 좁은 띠가 아니라 굵은 갈래가 화면 중앙 위쪽까지 뻗어 은폐 구도를 지나치게 크게 차지한다.",
        "entities": "성인 동아시아계 외형의 남성 세 명으로, 군복 차림 둘과 흰옷·검은 조끼 차림 하나가 보인다. 일본군 역할에 부합하는 군복은 있지만 국적은 외형만으로 확인할 수 없다. 요구된 군인 한 명보다 한 명 많다. 바닥의 남자는 눈을 뜨고 상대를 올려다보아 무반응 시신으로 읽히지 않는다. 소총은 보이지만 가슴에 박힌 고정 총검은 없다. 피 묻은 흙과 쌀, 솥, 그릇 파편, 연기 낀 초가 배경은 보이며 읽을 수 있는 글자는 없다.",
        "hard_violations": [
         "장면에 지정되지 않은 군인 한 명이 추가되어 총 세 사람이 등장한다.",
         "시신이어야 할 인물이 머리와 상체를 능동적으로 세우고 군인을 올려다보는 자세로 배치되어, 완전히 지지된 무반응 사망 자세를 위반한다."
        ],
        "physics": "군인들의 발은 땅에 닿고 소총은 오른쪽 군인의 어깨끈으로 지지된다. 바닥 남자의 다리와 골반은 땅에 놓이고 굽힌 팔 주변이 상체를 받치지만, 들린 머리에는 외부 받침이 없고 목과 몸통의 능동적인 힘이 필요하다. 이는 시신의 이완된 자세가 아니다. 누구도 소총을 양손으로 잡아 내려찌르지 않으므로 가슴으로 전달되는 힘이나 충격도 보이지 않는다."
       },
       {
        "label": "A",
        "direction": "군인은 화면 오른쪽 아래의 누운 남자를 바라보고, 그 남자는 왼쪽 위의 군인을 올려다본다. 소총은 군인의 등 뒤에서 오른쪽 위로 향한다. 총구와 총검이 남자의 가슴을 겨누지 않으며, 둘 사이에 무기 접촉도 없다.",
        "built_space": "초가 구조물 세 채가 왼쪽 가장자리·중앙 뒤·오른쪽에 있고, 뒤로 목책과 산림이 이어진다. 큰 솥 하나와 여러 그릇·파편, 흩어진 쌀, 목재 잔해가 흙바닥에 있다. 왼쪽 전경 나무는 가장자리의 세로 띠로 남아 두 사람을 가리지 않는다. 군인은 중앙 왼쪽, 누운 남자는 오른쪽 아래에 있어 와이드 숏과 은폐 시점에는 A보다 가깝지만, 가슴을 찌를 수 있는 근접 배치는 아니다.",
        "entities": "성인 동아시아계 외형의 남성 두 명으로, 군복 차림 한 명과 때 묻은 전통 복식 차림 한 명이 보여 요구 인원은 맞는다. 군복은 일본군 역할을 표현하지만 국적 자체는 외형으로 확정할 수 없다. 누운 남자는 눈을 뜨고 고개와 상체를 세워 시신보다는 살아 있는 부상자로 보인다. 소총은 등에 걸려 있으며 가슴에 꽂힌 총검은 없다. 솥, 깨진 그릇, 쌀, 혈흔성 얼룩과 연기 낀 진영은 보이고 읽을 수 있는 글자는 없다.",
        "hard_violations": [
         "시신이어야 할 인물이 팔로 상체를 받치고 머리를 들어 군인을 바라보도록 배치되어, 근육을 쓰지 않고 완전히 지지되어야 하는 사망 자세를 위반한다."
        ],
        "physics": "군인은 양발로 땅을 딛고 있다. 소총은 몸을 가로지르는 멜빵으로 지지되고 한 손이 아래쪽 개머리판 부근에 닿아 있어 부유하지 않는다. 누운 남자의 다리와 골반, 아래쪽 팔은 땅과 천에 지지되지만 들린 머리를 받치는 물체는 없으며 목의 근력이 필요하다. 시신에게 허용되지 않는 능동 자세다. 군인의 몸과 소총에는 아래로 찌르는 힘이나 가슴 접촉이 전혀 나타나지 않는다."
       }
      ]
     },
     "ok": true
    }
   ],
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "slot_winner_match": true,
   "slot_winner": {
    "gemini-pro": "A",
    "gpt-high": "A"
   },
   "route": "cross_slot_agree"
  },
  "dual": {
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "normalized": {
    "A": 2.0,
    "B": 0.0
   },
   "adjusted": {
    "A": 1.75,
    "B": -0.25
   },
   "violations": {
    "A": [
     "[gemini-pro] 명시된 핵심 행동(총검으로 가슴을 찌름)이 전혀 나타나지 않음",
     "[gemini-pro] 시신이 고개를 들고 팔로 버티며 살아 움직임 (중력/부동 상태 위반)",
     "[gpt-high] 시신이어야 할 인물이 팔로 상체를 받치고 머리를 들어 군인을 바라보도록 배치되어, 근육을 쓰지 않고 완전히 지지되어야 하는 사망 자세를 위반한다."
    ],
    "B": [
     "[gemini-pro] 지시에 없는 불필요한 인물 추가 (일본군이 2명임)",
     "[gemini-pro] 명시된 핵심 행동(총검으로 가슴을 찌름) 누락",
     "[gemini-pro] 시신이 고개를 들고 버티며 살아 있음 (중력/부동 상태 위반)",
     "[gpt-high] 장면에 지정되지 않은 군인 한 명이 추가되어 총 세 사람이 등장한다.",
     "[gpt-high] 시신이어야 할 인물이 머리와 상체를 능동적으로 세우고 군인을 올려다보는 자세로 배치되어, 완전히 지지된 무반응 사망 자세를 위반한다."
    ]
   },
   "per_model_winner": {
    "gemini-pro": "A",
    "gpt-high": "A"
   },
   "agreed": true
  },
  "totals": {
   "A": 1750,
   "B": -250
  },
  "selected": "A",
  "ranking": [
   "A",
   "B"
  ],
  "verdicts": [
   {
    "label": "A",
    "score": 1750,
    "verdict_ko": "시신을 향해 총검을 찌르는 핵심 행동이 전혀 묘사되지 않았으며, 시신이어야 할 인물이 고개를 들고 살아 있어 완전히 실패한 결과물입니다.  ★위반: [gemini-pro] 명시된 핵심 행동(총검으로 가슴을 찌름)이 전혀 나타나지 않음 / [gemini-pro] 시신이 고개를 들고 팔로 버티며 살아 움직임 (중력/부동 상태 위반) / [gpt-high] 시신이어야 할 인물이 팔로 상체를 받치고 머리를 들어 군인을 바라보도록 배치되어, 근육을 쓰지 않고 완전히 지지되어야 하는 사망 자세를 위반한다."
   },
   {
    "label": "B",
    "score": -250,
    "verdict_ko": "핵심 행동인 총검 찌르기가 누락된 데다 지시에 없는 일본군이 한 명 더 추가되었고, 시신이 살아 움직여 사용할 수 없는 이미지입니다.  ★위반: [gemini-pro] 지시에 없는 불필요한 인물 추가 (일본군이 2명임) / [gemini-pro] 명시된 핵심 행동(총검으로 가슴을 찌름) 누락 / [gemini-pro] 시신이 고개를 들고 버티며 살아 있음 (중력/부동 상태 위반) / [gpt-high] 장면에 지정되지 않은 군인 한 명이 추가되어 총 세 사람이 등장한다. / [gpt-high] 시신이어야 할 인물이 머리와 상체를 능동적으로 세우고 군인을 올려다보는 자세로 배치되어, 완전히 지지된 무반응 사망 자세를 위반한다."
   }
  ],
  "refs": [
   {
    "label": "PREVIOUS SHOT STILL — a visually related earlier shot of this same place: the location's look, materials, fixed features and lighting mood are LOCKED to this photo; never copy its camera framing. The people visible in that photo are NOT in this shot: never carry their faces, bodies or clothing onto anyone here. Each person in THIS shot is defined solely by the PEOPLE list, their own CHARACTER REFERENCE images and their wardrobe notes. If this photo shows a character who cannot move (dead or unconscious), that character's exact pose, position and orientation are ALSO LOCKED — treat their body as a fixed prop of the set that only the camera moves around.",
    "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/db00517b-165b-4005-a287-f28936b59b95/images/272bd7af-9294-4490-98c6-88a40e9deacc/scene/recipe/S60sh2_sel.png",
    "asset_id": "9a8a1465-bc86-433f-b821-e073cfb330f4",
    "role": "prev_still"
   }
  ],
  "critique_skipped": true,
  "needs_reshoot": true,
  "shot_run_produced": false,
  "shot_run_uid": "06aa1442-35a0-7b6f-b2c1-0ec75664e643",
  "ref_mode": "prev+엔티티",
  "share_plan": {
   "ref_plan": "prev",
   "prev_anchor_tag": "S60sh2"
  }
 },
 "S60sh11::cine": {
  "applied": false,
  "attempted_at": "2026-09-09T09:35:26.365848+00:00",
  "fingerprint": "7da0d54b1ff96dd2423ea83e211b1f8d6c9da79368e1769999c8aec6d7df310e",
  "fingerprint_version": 2,
  "provider": "mai",
  "endpoint": "openrouter/chat-completions",
  "model": "microsoft/mai-image-2.6",
  "multiroll_tag": "still_S60sh11_cine",
  "slot": "",
  "pack": "24.202608252115",
  "source_file": "S60sh11_sel.png",
  "source_sha256": "eca46a50e26d507273049bf846984e3340d970caef5f6ad7f1a60ede82304a40",
  "error": "RuntimeError: moderation blocked: {\"message\": \"Response content blocked by label 'MultiSeverity_ViolenceScore'.\", \"code\": 400}",
  "moderation_refusals": 2,
  "declined": true,
  "declined_reason": "moderation"
 },
 "S60sh12::signage": {
  "fp": "a5f1b8a2f307d354",
  "inscriptions": [],
  "cues": [],
  "dropped": []
 },
 "groupbg::ruined_camp_treeline": {
  "input_fingerprint": "1016244ba73261b8",
  "meta": {
   "model": "gpt-image-2",
   "size": "1536x864",
   "pack": "11.202607220237",
   "contract": "bgfirst_full_v3",
   "group_sig": {
    "key": "ruined_camp_treeline",
    "tags": [
     "S60sh12"
    ]
   },
   "context_sig": "bfad95abc362d943"
  },
  "prompt": "Create ONE empty live-action location background photograph — NO PEOPLE, no figures, no body parts, no silhouettes, no shadows or reflections of people anywhere.\n\"Empty\" means no people only: KEEP the location's inherent occupants and stock that define the place — animals in an animal shelter, pen or farm, goods and produce in a market, moored boats in a harbour — unless the shot text explicitly removes them.\nThe FIRST attached image is a thin-line storyboard sketch of a shot that happens at this location: use it ONLY as spatial evidence — what this place contains, how its ground, structures and landmarks are arranged and proportioned. Ignore the sketched people and arrows entirely, and do NOT copy its line style: render a fully photographic, physically plausible real place that fits THE LOCATION text below.\nHUMAN-SCALE CALIBRATION: derive every structure's true size from human-scale elements — a door ≈ 2m, a window ≈ 1–1.5m wide, one storey ≈ 2.5–3m; never inflate a small structure or shrink a large one.\n\nTHE LOCATION — early 20th-century Korea, Jeolla region; all people are Korean unless explicitly stated otherwise: Behind a tree at the edge of the devastated mountain camp, adjoining the forest cover outside the clearing.\n\nLOCATION DETAIL (descriptive evidence about this place from the production's location records — evidence, not a staging order): use it to understand what kind of place this is and what it permanently contains. Carry over ONLY its enduring physical features — layout, structures, ground and materials, fixed equipment and installations. Do NOT treat any lighting, weather or time-of-day wording, momentary object states, depicted events or subjective impressions in it as fixed facts of the place; the TIME OF DAY lock below is the sole authority for time and lighting.\n파청 산채 공터·훈련장·취사장·막사 주변: 산 중턱에 통나무를 세워 만든 조잡한 목책 안의 넓은 흙바닥 공터. 나무와 거적으로 만든 움막과 막사가 흩어져 있고, 돌과 진흙으로 만든 취사용 아궁이와 무쇠 솥, 모닥불 터가 있다. 전투 후에는 솥이 뒤집히고 막사가 불타거나 무너진다. (특징: 통나무를 박아 세운 목책; 넓게 다져진 흙바닥 (공터/훈련장); 돌과 진흙으로 만든 다수의 아궁이와 무쇠솥; 나무와 거적으로 엮어 만든 움막과 막사; 부서진 토기 밥그릇과 뒤집힌 무쇠솥 (전투 후); 불에 탄 나무 잔해와 흩어진 쌀밥 (전투 후))\n\nSCENE EVIDENCE (verbatim quotes from the screenplay about this place — treat them as evidence of what the location physically contains and looks like; stage the PLACE those moments happen in, but do NOT depict the momentary actions, people or staged props themselves):\n- 저편에서 인기척. 규홍이 나무 뒤로 몸을 숨긴다.\n- 규홍이 피 묻은 쌍낫을 움켜쥔다. 말없이 숲속으로 사라진다.\n\nTIME OF DAY (lock): day.\n\nRender ONE photorealistic empty location photograph, 16:9, neutral enough that every shot of this place can be staged from it later. No readable writing anywhere: surfaces that would carry writing may be present, but stage any wording out of legibility — an oblique angle, distance, shallow focus. No captions, watermarks or overlay text.",
  "effective_prompt": "Create ONE empty live-action location background photograph — NO PEOPLE, no figures, no body parts, no silhouettes, no shadows or reflections of people anywhere.\n\"Empty\" means no people only: KEEP the location's inherent occupants and stock that define the place — animals in an animal shelter, pen or farm, goods and produce in a market, moored boats in a harbour — unless the shot text explicitly removes them.\nThe FIRST attached image is a thin-line storyboard sketch of a shot that happens at this location: use it ONLY as spatial evidence — what this place contains, how its ground, structures and landmarks are arranged and proportioned. Ignore the sketched people and arrows entirely, and do NOT copy its line style: render a fully photographic, physically plausible real place that fits THE LOCATION text below.\nHUMAN-SCALE CALIBRATION: derive every structure's true size from human-scale elements — a door ≈ 2m, a window ≈ 1–1.5m wide, one storey ≈ 2.5–3m; never inflate a small structure or shrink a large one.\n\nTHE LOCATION — early 20th-century Korea, Jeolla region; all people are Korean unless explicitly stated otherwise: Behind a tree at the edge of the devastated mountain camp, adjoining the forest cover outside the clearing.\n\nLOCATION DETAIL (descriptive evidence about this place from the production's location records — evidence, not a staging order): use it to understand what kind of place this is and what it permanently contains. Carry over ONLY its enduring physical features — layout, structures, ground and materials, fixed equipment and installations. Do NOT treat any lighting, weather or time-of-day wording, momentary object states, depicted events or subjective impressions in it as fixed facts of the place; the TIME OF DAY lock below is the sole authority for time and lighting.\n파청 산채 공터·훈련장·취사장·막사 주변: 산 중턱에 통나무를 세워 만든 조잡한 목책 안의 넓은 흙바닥 공터. 나무와 거적으로 만든 움막과 막사가 흩어져 있고, 돌과 진흙으로 만든 취사용 아궁이와 무쇠 솥, 모닥불 터가 있다. 전투 후에는 솥이 뒤집히고 막사가 불타거나 무너진다. (특징: 통나무를 박아 세운 목책; 넓게 다져진 흙바닥 (공터/훈련장); 돌과 진흙으로 만든 다수의 아궁이와 무쇠솥; 나무와 거적으로 엮어 만든 움막과 막사; 부서진 토기 밥그릇과 뒤집힌 무쇠솥 (전투 후); 불에 탄 나무 잔해와 흩어진 쌀밥 (전투 후))\n\nSCENE EVIDENCE (verbatim quotes from the screenplay about this place — treat them as evidence of what the location physically contains and looks like; stage the PLACE those moments happen in, but do NOT depict the momentary actions, people or staged props themselves):\n- 저편에서 인기척. 규홍이 나무 뒤로 몸을 숨긴다.\n- 규홍이 피 묻은 쌍낫을 움켜쥔다. 말없이 숲속으로 사라진다.\n\nTIME OF DAY (lock): day.\n\nRender ONE photorealistic empty location photograph, 16:9, neutral enough that every shot of this place can be staged from it later. No readable writing anywhere: surfaces that would carry writing may be present, but stage any wording out of legibility — an oblique angle, distance, shallow focus. No captions, watermarks or overlay text.",
  "bg_path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/db00517b-165b-4005-a287-f28936b59b95/images/272bd7af-9294-4490-98c6-88a40e9deacc/scene/recipe/groupbg_ruined_camp_treeline_e39a9f.png",
  "asset_id": "ce860305-92d9-4331-bf03-df567841a4aa",
  "input_asset_ids": [
   "46742370-5b9f-4d7a-808e-0ece1e4b21c2"
  ],
  "origin_tag": "S60sh12",
  "place_text": "Behind a tree at the edge of the devastated mountain camp, adjoining the forest cover outside the clearing.",
  "origin_inputs": {
   "place_text": "Behind a tree at the edge of the devastated mountain camp, adjoining the forest cover outside the clearing.",
   "time_of_day_en": "day",
   "conti_asset_id": "46742370-5b9f-4d7a-808e-0ece1e4b21c2"
  }
 },
 "S60sh12::bgfirst_bg": {
  "input_fingerprint": "43bbd55658ce0214",
  "prompt": "Create the EMPTY BACKGROUND PLATE for one film shot — NO PEOPLE, no figures, no body parts, no sketch lines, no arrows anywhere.\n\"Empty\" means no people only: KEEP the location's inherent occupants and stock that define the place — animals in an animal shelter, pen or farm, goods in a market, moored boats in a harbour — unless the shot text explicitly removes them.\nThe FIRST attached image is a thin-line storyboard sketch: use ONLY its camera angle, horizon, perspective and the placement/size of buildings and set masses — ignore the sketched people and arrows entirely. The SECOND attached image (LOCATION PHOTOGRAPH) is the real place: take its architecture, materials, signage and fixed features, and RE-PROJECT them into the sketch's camera. If the photograph's camera differs from the sketch's, the sketch's camera wins.\nHUMAN-SCALE CALIBRATION: derive every structure's true size from human-scale elements — a door ≈ 2m, a window ≈ 1–1.5m wide, one storey ≈ 2.5–3m; never inflate a small structure or shrink a large one.\n\nSHOT TEXT this background must serve (Korean): 피 묻은 쌍낫의 자루를 꽉 움켜쥐어 손가락 마디가 하얘진 안규홍의 주먹 클로즈업.\n\nLOCATION (lock): Behind a tree at the edge of the devastated mountain camp, adjoining the forest cover outside the clearing.\n\nTIME OF DAY (lock): day.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: close-up\n- KEY BACKGROUND ELEMENTS: Bloodstained paired sickles (Held tightly in 안규홍's hands) — Handles lie obliquely across the lower frame, with only portions of the blades visible; used as Give the tightening grip its narrative object while leaving the hands as the primary subject; Concealing tree trunk (Still beside 안규홍's hidden position) — A partial edge remains behind his near flank; used as Maintains continuity with the concealed observation.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Naturalistic daylight preserves the contrast between whitening knuckles and blood with muted saturation and sober restraint.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nRender ONE photorealistic empty location photograph, 16:9, that this shot can be staged inside later. No readable writing anywhere: surfaces that would carry writing may be present, but stage any wording out of legibility — an oblique angle, distance, shallow focus. No captions, watermarks or overlay text.",
  "effective_prompt": "Create the EMPTY BACKGROUND PLATE for one film shot — NO PEOPLE, no figures, no body parts, no sketch lines, no arrows anywhere.\n\"Empty\" means no people only: KEEP the location's inherent occupants and stock that define the place — animals in an animal shelter, pen or farm, goods in a market, moored boats in a harbour — unless the shot text explicitly removes them.\nThe FIRST attached image is a thin-line storyboard sketch: use ONLY its camera angle, horizon, perspective and the placement/size of buildings and set masses — ignore the sketched people and arrows entirely. The SECOND attached image (LOCATION PHOTOGRAPH) is the real place: take its architecture, materials, signage and fixed features, and RE-PROJECT them into the sketch's camera. If the photograph's camera differs from the sketch's, the sketch's camera wins.\nHUMAN-SCALE CALIBRATION: derive every structure's true size from human-scale elements — a door ≈ 2m, a window ≈ 1–1.5m wide, one storey ≈ 2.5–3m; never inflate a small structure or shrink a large one.\n\nSHOT TEXT this background must serve (Korean): 피 묻은 쌍낫의 자루를 꽉 움켜쥐어 손가락 마디가 하얘진 안규홍의 주먹 클로즈업.\n\nLOCATION (lock): Behind a tree at the edge of the devastated mountain camp, adjoining the forest cover outside the clearing.\n\nTIME OF DAY (lock): day.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: close-up\n- KEY BACKGROUND ELEMENTS: Bloodstained paired sickles (Held tightly in 안규홍's hands) — Handles lie obliquely across the lower frame, with only portions of the blades visible; used as Give the tightening grip its narrative object while leaving the hands as the primary subject; Concealing tree trunk (Still beside 안규홍's hidden position) — A partial edge remains behind his near flank; used as Maintains continuity with the concealed observation.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Naturalistic daylight preserves the contrast between whitening knuckles and blood with muted saturation and sober restraint.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nRender ONE photorealistic empty location photograph, 16:9, that this shot can be staged inside later. No readable writing anywhere: surfaces that would carry writing may be present, but stage any wording out of legibility — an oblique angle, distance, shallow focus. No captions, watermarks or overlay text.",
  "bg_path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/db00517b-165b-4005-a287-f28936b59b95/images/272bd7af-9294-4490-98c6-88a40e9deacc/scene/recipe/S60sh12__bgfirst_bg.png",
  "asset_id": "4c578bd0-065c-4d45-85cd-3aeb86ee8abb",
  "input_asset_ids": [
   "46742370-5b9f-4d7a-808e-0ece1e4b21c2",
   "ce860305-92d9-4331-bf03-df567841a4aa"
  ]
 },
 "S60sh12": {
  "input_fingerprint": "26180f58583fbcc6",
  "prompt": "Create ONE FINAL photorealistic live-action film still of the moment below — early 20th-century Korea, Jeolla region; all people are Korean unless explicitly stated otherwise. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 피 묻은 쌍낫의 자루를 꽉 움켜쥐어 손가락 마디가 하얘진 안규홍의 주먹 클로즈업.\n\nLOCATION (lock): Behind a tree at the edge of the devastated mountain camp, adjoining the forest cover outside the clearing. The shot takes place here — the attached LOCATION PHOTOGRAPH shows the exact spot.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: close-up\n- KEY BACKGROUND ELEMENTS: Bloodstained paired sickles (Held tightly in 안규홍's hands) — Handles lie obliquely across the lower frame, with only portions of the blades visible; used as Give the tightening grip its narrative object while leaving the hands as the primary subject; Concealing tree trunk (Still beside 안규홍's hidden position) — A partial edge remains behind his near flank; used as Maintains continuity with the concealed observation.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Naturalistic daylight preserves the contrast between whitening knuckles and blood with muted saturation and sober restraint.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): Smoke, broken bowls, overturned cookware, fallen timber and bloodied rice persist throughout the camp. 안규홍: He grips both bloodied sickles and moves into the forest; his face and hands are smeared with blood and dirt, and his back wound remains open.\n\nTHE HAND THAT IS DOING THIS: the object at the centre of this shot is being held, operated, read, aimed or handed over by 안규홍 right now, so 안규홍's hand — and as much of the wrist and forearm as the framing reaches — is in the frame, gripping or resting on that object exactly the way its real use requires. Match that hand to 안규홍: its size, build, skin, age, grooming, sleeve and anything worn on it belong to that person and to no one else.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 안규홍 (한국인 남성, 청년 나이대, 검은 머리, 단단히 묶은 머리) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.",
  "roll_prompts": {
   "A": "Stage the shot. The FIRST attached image (SHOT BACKGROUND) is the finished empty background of this shot — keep it EXACTLY: its camera, perspective, architecture, lighting and every fixed feature stay untouched. The SECOND attached image (LAYOUT SKETCH) tells you ONLY where the people go: each sketched person's position, screen size, pose and the gaze/motion arrows. Ignore the sketch's background lines. The CHARACTER REFERENCE photographs show the real people.\nPlace the real people into the background at exactly the sketched positions, sizes and poses, following the arrow directions. No sketch lines or arrows may remain.\n\nCreate ONE FINAL photorealistic live-action film still of the moment below — early 20th-century Korea, Jeolla region; all people are Korean unless explicitly stated otherwise. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 피 묻은 쌍낫의 자루를 꽉 움켜쥐어 손가락 마디가 하얘진 안규홍의 주먹 클로즈업.\n\nLOCATION (lock): Behind a tree at the edge of the devastated mountain camp, adjoining the forest cover outside the clearing. The shot takes place here — the FIRST attached image (SHOT BACKGROUND) is this exact place, already built: its ground, structures, horizon, materials and lighting are the finished truth of this location and must not be redesigned or replaced. No location photograph is attached — read the place from that image alone, and add no scenery, structure, vehicle or fixture that it does not already show. This lock governs the place only; the figures in the shot follow the staging and pose instructions.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: close-up\n- KEY BACKGROUND ELEMENTS: Bloodstained paired sickles (Held tightly in 안규홍's hands) — Handles lie obliquely across the lower frame, with only portions of the blades visible; used as Give the tightening grip its narrative object while leaving the hands as the primary subject; Concealing tree trunk (Still beside 안규홍's hidden position) — A partial edge remains behind his near flank; used as Maintains continuity with the concealed observation.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Naturalistic daylight preserves the contrast between whitening knuckles and blood with muted saturation and sober restraint.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): Smoke, broken bowls, overturned cookware, fallen timber and bloodied rice persist throughout the camp. 안규홍: He grips both bloodied sickles and moves into the forest; his face and hands are smeared with blood and dirt, and his back wound remains open.\n\nTHE HAND THAT IS DOING THIS: the object at the centre of this shot is being held, operated, read, aimed or handed over by 안규홍 right now, so 안규홍's hand — and as much of the wrist and forearm as the framing reaches — is in the frame, gripping or resting on that object exactly the way its real use requires. Match that hand to 안규홍: its size, build, skin, age, grooming, sleeve and anything worn on it belong to that person and to no one else.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 안규홍 (한국인 남성, 청년 나이대, 검은 머리, 단단히 묶은 머리) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.",
   "B": "Create ONE FINAL photorealistic live-action film still of the moment below — early 20th-century Korea, Jeolla region; all people are Korean unless explicitly stated otherwise. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 피 묻은 쌍낫의 자루를 꽉 움켜쥐어 손가락 마디가 하얘진 안규홍의 주먹 클로즈업.\n\nLOCATION (lock): Behind a tree at the edge of the devastated mountain camp, adjoining the forest cover outside the clearing. The shot takes place here — the attached LOCATION PHOTOGRAPH shows the exact spot.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: close-up\n- KEY BACKGROUND ELEMENTS: Bloodstained paired sickles (Held tightly in 안규홍's hands) — Handles lie obliquely across the lower frame, with only portions of the blades visible; used as Give the tightening grip its narrative object while leaving the hands as the primary subject; Concealing tree trunk (Still beside 안규홍's hidden position) — A partial edge remains behind his near flank; used as Maintains continuity with the concealed observation.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Naturalistic daylight preserves the contrast between whitening knuckles and blood with muted saturation and sober restraint.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): Smoke, broken bowls, overturned cookware, fallen timber and bloodied rice persist throughout the camp. 안규홍: He grips both bloodied sickles and moves into the forest; his face and hands are smeared with blood and dirt, and his back wound remains open.\n\nTHE HAND THAT IS DOING THIS: the object at the centre of this shot is being held, operated, read, aimed or handed over by 안규홍 right now, so 안규홍's hand — and as much of the wrist and forearm as the framing reaches — is in the frame, gripping or resting on that object exactly the way its real use requires. Match that hand to 안규홍: its size, build, skin, age, grooming, sleeve and anything worn on it belong to that person and to no one else.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 안규홍 (한국인 남성, 청년 나이대, 검은 머리, 단단히 묶은 머리) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.\n\nCOMPOSITION VARIATION (this candidate only): keep the place, the people and the moment exactly as specified, and keep from the CAMERA & FRAME contract its subject scale, screen placement and key background elements exactly. For this candidate only, its camera angle and height are deliberately left open: choose a camera position distinctly different from the obvious one, as a film director picking a second setup on the same blocking. Stay true to any explicit framing the SHOT TEXT itself dictates."
  },
  "roll_refs": {
   "A": [
    {
     "label": "SHOT BACKGROUND",
     "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/db00517b-165b-4005-a287-f28936b59b95/images/272bd7af-9294-4490-98c6-88a40e9deacc/scene/recipe/S60sh12__bgfirst_bg.png",
     "asset_id": "4c578bd0-065c-4d45-85cd-3aeb86ee8abb",
     "role": "bgfirst_bg"
    },
    {
     "label": "LAYOUT SKETCH (people placement only)",
     "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/db00517b-165b-4005-a287-f28936b59b95/images/272bd7af-9294-4490-98c6-88a40e9deacc/conti/conti_S60sh12.png",
     "asset_id": "46742370-5b9f-4d7a-808e-0ece1e4b21c2",
     "role": "conti_light"
    },
    {
     "label": "CHARACTER REFERENCE — 안규홍: the exact person appearing in this shot; match face, hair and build exactly.",
     "path": "<bytes:1483645>",
     "asset_id": "b52dff85-b663-44f9-aa9c-46b828990f55",
     "role": "character_ref"
    },
    {
     "label": "PROP REFERENCE — 낫: the exact object appearing in this shot; match its look, material and wear exactly.",
     "path": "<bytes:977465>",
     "asset_id": "e9884693-b2aa-4769-9ee0-4c7b50db40c2",
     "role": "prop_ref"
    }
   ],
   "B": [
    {
     "label": "LOCATION PHOTOGRAPH — the exact place of this shot: its architecture, materials, fixed features and lighting mood are spatial truth; stage the moment inside this place. Never copy its camera framing.",
     "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/db00517b-165b-4005-a287-f28936b59b95/images/272bd7af-9294-4490-98c6-88a40e9deacc/scene/recipe/groupbg_ruined_camp_treeline_e39a9f.png",
     "asset_id": "ce860305-92d9-4331-bf03-df567841a4aa",
     "role": "bgfirst_group_bg"
    },
    {
     "label": "CHARACTER REFERENCE — 안규홍: the exact person appearing in this shot; match face, hair and build exactly.",
     "path": "<bytes:1483645>",
     "asset_id": "b52dff85-b663-44f9-aa9c-46b828990f55",
     "role": "character_ref"
    },
    {
     "label": "PROP REFERENCE — 낫: the exact object appearing in this shot; match its look, material and wear exactly.",
     "path": "<bytes:977465>",
     "asset_id": "e9884693-b2aa-4769-9ee0-4c7b50db40c2",
     "role": "prop_ref"
    }
   ]
  },
  "shot_run_spend_attempt_count": 1,
  "judge_flip_skipped": "judge_fn_owns_order",
  "readings": [
   {
    "label": "A",
    "direction": "인물의 주먹이 두 개의 낫 자루를 꽉 쥐고 있으며, 낫의 금속 부위는 화면 오른쪽을 향함.",
    "built_space": "배경에 폐허가 된 산채가 보이며, 화면 좌측 인물의 팔 뒤로 나무 기둥의 가장자리가 위치함.",
    "entities": "핏자국과 흙이 묻은 안규홍의 소매와 팔, 하얗게 마디가 드러날 정도로 꽉 쥔 주먹, 두 개의 낫 자루(위쪽 낫은 날이 없거나 부러진 형태로 보임).",
    "hard_violations": [],
    "physics": "오른손이 두 개의 낫 자루 중간을 단단히 쥐어 지탱하고 있음."
   },
   {
    "label": "B",
    "direction": "두 손이 각각 낫 자루를 쥐고 있으며, 낫의 날이 교차하며 화면 아래쪽을 향함.",
    "built_space": "배경에 산채가 위치하고, 나무 기둥이 화면 중앙 인물의 손 바로 뒤에 배치됨.",
    "entities": "피 묻은 의상을 입은 두 팔과 손, 두 개의 낫.",
    "hard_violations": [
     "[gemini-pro] 물리적으로 불가능한 사물 융합: 왼손이 쥔 낫 자루가 오른손이 쥔 낫의 금속 깃 부분과 한 덩어리로 융합되어 물리적으로 불가능한 형태를 만듦."
    ],
    "physics": "두 손이 자루를 쥐고 있으나, 낫 구조 자체가 융합되어 물리적 성립이 불가능함."
   }
  ],
  "cross_model_order": {
   "policy": "cross_model_order_v2_slot0_fwd_slot1_rev_parallel2",
   "slots": [
    {
     "model": "gemini-pro",
     "order": "forward",
     "display_to_canonical": {
      "A": "A",
      "B": "B"
     },
     "raw": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 8,
        "verdict_ko": "주먹을 꽉 쥐어 하얗게 질린 손가락 마디의 디테일을 잘 살렸고 나무 뒤의 구도를 훌륭히 구현했으나, 위쪽 낫의 날이 부러진 것처럼 묘사된 점이 아쉽습니다."
       },
       {
        "label": "B",
        "score": 2,
        "verdict_ko": "두 개의 낫이 한 덩어리로 엉겨 붙어 물리적으로 불가능한 형태를 띠는 치명적인 사물 융합 오류가 있습니다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "인물의 주먹이 두 개의 낫 자루를 꽉 쥐고 있으며, 낫의 금속 부위는 화면 오른쪽을 향함.",
        "built_space": "배경에 폐허가 된 산채가 보이며, 화면 좌측 인물의 팔 뒤로 나무 기둥의 가장자리가 위치함.",
        "entities": "핏자국과 흙이 묻은 안규홍의 소매와 팔, 하얗게 마디가 드러날 정도로 꽉 쥔 주먹, 두 개의 낫 자루(위쪽 낫은 날이 없거나 부러진 형태로 보임).",
        "hard_violations": [],
        "physics": "오른손이 두 개의 낫 자루 중간을 단단히 쥐어 지탱하고 있음."
       },
       {
        "label": "B",
        "direction": "두 손이 각각 낫 자루를 쥐고 있으며, 낫의 날이 교차하며 화면 아래쪽을 향함.",
        "built_space": "배경에 산채가 위치하고, 나무 기둥이 화면 중앙 인물의 손 바로 뒤에 배치됨.",
        "entities": "피 묻은 의상을 입은 두 팔과 손, 두 개의 낫.",
        "hard_violations": [
         "물리적으로 불가능한 사물 융합: 왼손이 쥔 낫 자루가 오른손이 쥔 낫의 금속 깃 부분과 한 덩어리로 융합되어 물리적으로 불가능한 형태를 만듦."
        ],
        "physics": "두 손이 자루를 쥐고 있으나, 낫 구조 자체가 융합되어 물리적 성립이 불가능함."
       }
      ],
      "all_candidates_fail": false
     },
     "normalized": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 8,
        "verdict_ko": "주먹을 꽉 쥐어 하얗게 질린 손가락 마디의 디테일을 잘 살렸고 나무 뒤의 구도를 훌륭히 구현했으나, 위쪽 낫의 날이 부러진 것처럼 묘사된 점이 아쉽습니다."
       },
       {
        "label": "B",
        "score": 2,
        "verdict_ko": "두 개의 낫이 한 덩어리로 엉겨 붙어 물리적으로 불가능한 형태를 띠는 치명적인 사물 융합 오류가 있습니다."
       }
      ],
      "all_candidates_fail": false,
      "readings": [
       {
        "label": "A",
        "direction": "인물의 주먹이 두 개의 낫 자루를 꽉 쥐고 있으며, 낫의 금속 부위는 화면 오른쪽을 향함.",
        "built_space": "배경에 폐허가 된 산채가 보이며, 화면 좌측 인물의 팔 뒤로 나무 기둥의 가장자리가 위치함.",
        "entities": "핏자국과 흙이 묻은 안규홍의 소매와 팔, 하얗게 마디가 드러날 정도로 꽉 쥔 주먹, 두 개의 낫 자루(위쪽 낫은 날이 없거나 부러진 형태로 보임).",
        "hard_violations": [],
        "physics": "오른손이 두 개의 낫 자루 중간을 단단히 쥐어 지탱하고 있음."
       },
       {
        "label": "B",
        "direction": "두 손이 각각 낫 자루를 쥐고 있으며, 낫의 날이 교차하며 화면 아래쪽을 향함.",
        "built_space": "배경에 산채가 위치하고, 나무 기둥이 화면 중앙 인물의 손 바로 뒤에 배치됨.",
        "entities": "피 묻은 의상을 입은 두 팔과 손, 두 개의 낫.",
        "hard_violations": [
         "물리적으로 불가능한 사물 융합: 왼손이 쥔 낫 자루가 오른손이 쥔 낫의 금속 깃 부분과 한 덩어리로 융합되어 물리적으로 불가능한 형태를 만듦."
        ],
        "physics": "두 손이 자루를 쥐고 있으나, 낫 구조 자체가 융합되어 물리적 성립이 불가능함."
       }
      ]
     },
     "ok": true
    },
    {
     "model": "gpt-high",
     "order": "reverse",
     "display_to_canonical": {
      "A": "B",
      "B": "A"
     },
     "raw": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 7,
        "verdict_ko": "양손으로 피 묻은 낫을 쥔 동작과 사선 자루는 맞지만, 크게 드러난 두 낫날이 주먹과 경쟁하고 하얗게 질린 손가락 마디의 강조가 약하다."
       },
       {
        "label": "B",
        "score": 8,
        "verdict_ko": "하얗게 도드라진 마디와 힘줄을 가진 주먹에 시선이 모이고 쌍낫 자루가 하단을 지나 핵심 순간에 더 충실하나, 두 자루를 한 손에 모아 쥔 점은 양손 지시와 다르다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "얼굴과 시선은 프레임 밖이다. 앞쪽 손의 자루는 왼쪽 위에서 오른쪽 아래로 놓이고, 뒤쪽 손의 자루는 반대 사선으로 교차한다. 두 낫날 끝은 각각 화면 왼쪽과 오른쪽으로 굽어 있으며 특정 대상을 겨누지는 않는다. 공격이 아니라 꽉 움켜쥐는 순간으로 읽힌다.",
        "built_space": "인물의 몸통 일부와 두 팔이 왼쪽, 굵은 나무줄기 하나가 바로 뒤 중앙에 있다. 오른쪽 배경에는 목책, 작은 움막과 천막, 낮은 평상 하나, 무너진 취사 자리와 여러 그릇이 보인다. 참조의 나무 옆 야영지와 공간 관계가 맞고 배경 시설도 작게 보이지만, 나무가 부분 가장자리보다 넓게 노출되고 야영지도 상당히 많이 들어온다.",
        "entities": "한 사람의 두 손과 팔, 오염된 미색 소매가 보이며 다른 사람은 없다. 손과 팔은 성인 남성의 체격에 부합하지만 얼굴이 없어 한국인 청년이라는 신원과 머리는 직접 확인할 수 없다. 낡은 나무 자루와 금속 고리를 가진 피 묻은 낫 두 자루가 있고 참조의 재질은 대체로 맞는다. 다만 날의 곡선이 참조보다 길고 크게 벌어진 인상이다. 손과 팔에 때와 혈흔은 있으나 주먹 마디의 창백함은 뚜렷하지 않다. 배경에는 연기와 쓰러진 목재, 취사도구가 보이지만 피 묻은 쌀은 구별되지 않는다. 읽을 수 있는 글자는 없다.",
        "hard_violations": [],
        "physics": "각 자루를 별도의 손이 감싸 쥐고 있고 손목과 팔이 소매 쪽으로 자연스럽게 이어진다. 두 낫 모두 손의 지지가 확인되며 공중에 따로 떠 있는 물체는 없다. 낫을 교차해 유지하는 자세는 가능하지만, 보이는 힘의 표현은 극도로 조인 주먹보다는 단단히 들고 있는 상태에 가깝다."
       },
       {
        "label": "B",
        "direction": "얼굴과 시선은 보이지 않는다. 두 자루가 화면 하단에서 왼쪽 아래부터 오른쪽 위로 완만한 사선을 이루며, 한 주먹이 이를 함께 감싼다. 잘 보이는 낫날은 오른쪽으로 뻗어 아래로 굽고, 다른 날은 겹쳐 옆면이 좁게 보인다. 겨누는 대상이나 휘두르는 동작 없이 움켜쥔 상태다.",
        "built_space": "몸통과 팔 일부가 왼쪽에 있고 나무줄기 하나가 바로 뒤 중앙 왼쪽을 차지한다. 오른쪽에는 목책, 중앙의 작은 움막, 평상 하나, 오른쪽 천막과 낮은 취사 덮개, 흩어진 그릇과 솥이 보인다. 참조 장소의 주요 배치와 재료가 유지되며 나무 뒤 은폐 위치도 성립한다. 다만 주먹 클로즈업에 비해 야영지 배경과 나무의 노출 면적은 넓다.",
        "entities": "한 사람의 손 하나와 팔, 피와 때가 묻은 미색 소매가 보인다. 성인 남성에게 맞는 손 크기와 체격이며 얼굴과 머리는 구도 밖이라 신원을 직접 확인할 수 없다. 주먹 마디와 손등 힘줄이 밝게 도드라지고 혈흔과 대비되어 지시된 긴장을 잘 전달한다. 나무 자루 두 개와 피 묻은 금속 날이 보이지만, 두 번째 날의 전체 형태는 겹침 때문에 확인하기 어렵다. 보이는 날은 참조보다 가늘고 길게 느껴진다. 배경에 전복된 취사도구와 혈흔은 있으나 연기와 피 묻은 쌀은 명확하지 않다. 읽을 수 있는 글자는 없다.",
        "hard_violations": [],
        "physics": "보이는 한 손의 손가락이 두 자루를 함께 감싸 지지하고 손목은 팔과 자연스럽게 연결된다. 두 자루를 한 손으로 모아 쥐는 것은 물리적으로 가능하며 지지 없는 낫은 없다. 팽팽한 손등과 돌출된 마디는 강하게 움켜쥐는 동작에 부합한다. 다만 명시된 양손 보유 대신 한 손에 두 낫이 모여 있다."
       }
      ],
      "all_candidates_fail": false
     },
     "normalized": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 7,
        "verdict_ko": "양손으로 피 묻은 낫을 쥔 동작과 사선 자루는 맞지만, 크게 드러난 두 낫날이 주먹과 경쟁하고 하얗게 질린 손가락 마디의 강조가 약하다."
       },
       {
        "label": "A",
        "score": 8,
        "verdict_ko": "하얗게 도드라진 마디와 힘줄을 가진 주먹에 시선이 모이고 쌍낫 자루가 하단을 지나 핵심 순간에 더 충실하나, 두 자루를 한 손에 모아 쥔 점은 양손 지시와 다르다."
       }
      ],
      "all_candidates_fail": false,
      "readings": [
       {
        "label": "B",
        "direction": "얼굴과 시선은 프레임 밖이다. 앞쪽 손의 자루는 왼쪽 위에서 오른쪽 아래로 놓이고, 뒤쪽 손의 자루는 반대 사선으로 교차한다. 두 낫날 끝은 각각 화면 왼쪽과 오른쪽으로 굽어 있으며 특정 대상을 겨누지는 않는다. 공격이 아니라 꽉 움켜쥐는 순간으로 읽힌다.",
        "built_space": "인물의 몸통 일부와 두 팔이 왼쪽, 굵은 나무줄기 하나가 바로 뒤 중앙에 있다. 오른쪽 배경에는 목책, 작은 움막과 천막, 낮은 평상 하나, 무너진 취사 자리와 여러 그릇이 보인다. 참조의 나무 옆 야영지와 공간 관계가 맞고 배경 시설도 작게 보이지만, 나무가 부분 가장자리보다 넓게 노출되고 야영지도 상당히 많이 들어온다.",
        "entities": "한 사람의 두 손과 팔, 오염된 미색 소매가 보이며 다른 사람은 없다. 손과 팔은 성인 남성의 체격에 부합하지만 얼굴이 없어 한국인 청년이라는 신원과 머리는 직접 확인할 수 없다. 낡은 나무 자루와 금속 고리를 가진 피 묻은 낫 두 자루가 있고 참조의 재질은 대체로 맞는다. 다만 날의 곡선이 참조보다 길고 크게 벌어진 인상이다. 손과 팔에 때와 혈흔은 있으나 주먹 마디의 창백함은 뚜렷하지 않다. 배경에는 연기와 쓰러진 목재, 취사도구가 보이지만 피 묻은 쌀은 구별되지 않는다. 읽을 수 있는 글자는 없다.",
        "hard_violations": [],
        "physics": "각 자루를 별도의 손이 감싸 쥐고 있고 손목과 팔이 소매 쪽으로 자연스럽게 이어진다. 두 낫 모두 손의 지지가 확인되며 공중에 따로 떠 있는 물체는 없다. 낫을 교차해 유지하는 자세는 가능하지만, 보이는 힘의 표현은 극도로 조인 주먹보다는 단단히 들고 있는 상태에 가깝다."
       },
       {
        "label": "A",
        "direction": "얼굴과 시선은 보이지 않는다. 두 자루가 화면 하단에서 왼쪽 아래부터 오른쪽 위로 완만한 사선을 이루며, 한 주먹이 이를 함께 감싼다. 잘 보이는 낫날은 오른쪽으로 뻗어 아래로 굽고, 다른 날은 겹쳐 옆면이 좁게 보인다. 겨누는 대상이나 휘두르는 동작 없이 움켜쥔 상태다.",
        "built_space": "몸통과 팔 일부가 왼쪽에 있고 나무줄기 하나가 바로 뒤 중앙 왼쪽을 차지한다. 오른쪽에는 목책, 중앙의 작은 움막, 평상 하나, 오른쪽 천막과 낮은 취사 덮개, 흩어진 그릇과 솥이 보인다. 참조 장소의 주요 배치와 재료가 유지되며 나무 뒤 은폐 위치도 성립한다. 다만 주먹 클로즈업에 비해 야영지 배경과 나무의 노출 면적은 넓다.",
        "entities": "한 사람의 손 하나와 팔, 피와 때가 묻은 미색 소매가 보인다. 성인 남성에게 맞는 손 크기와 체격이며 얼굴과 머리는 구도 밖이라 신원을 직접 확인할 수 없다. 주먹 마디와 손등 힘줄이 밝게 도드라지고 혈흔과 대비되어 지시된 긴장을 잘 전달한다. 나무 자루 두 개와 피 묻은 금속 날이 보이지만, 두 번째 날의 전체 형태는 겹침 때문에 확인하기 어렵다. 보이는 날은 참조보다 가늘고 길게 느껴진다. 배경에 전복된 취사도구와 혈흔은 있으나 연기와 피 묻은 쌀은 명확하지 않다. 읽을 수 있는 글자는 없다.",
        "hard_violations": [],
        "physics": "보이는 한 손의 손가락이 두 자루를 함께 감싸 지지하고 손목은 팔과 자연스럽게 연결된다. 두 자루를 한 손으로 모아 쥐는 것은 물리적으로 가능하며 지지 없는 낫은 없다. 팽팽한 손등과 돌출된 마디는 강하게 움켜쥐는 동작에 부합한다. 다만 명시된 양손 보유 대신 한 손에 두 낫이 모여 있다."
       }
      ]
     },
     "ok": true
    }
   ],
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "slot_winner_match": true,
   "slot_winner": {
    "gemini-pro": "A",
    "gpt-high": "A"
   },
   "route": "cross_slot_agree"
  },
  "dual": {
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "normalized": {
    "A": 2.0,
    "B": 1.125
   },
   "adjusted": {
    "A": 2.0,
    "B": 0.875
   },
   "violations": {
    "B": [
     "[gemini-pro] 물리적으로 불가능한 사물 융합: 왼손이 쥔 낫 자루가 오른손이 쥔 낫의 금속 깃 부분과 한 덩어리로 융합되어 물리적으로 불가능한 형태를 만듦."
    ]
   },
   "per_model_winner": {
    "gemini-pro": "A",
    "gpt-high": "A"
   },
   "agreed": true
  },
  "totals": {
   "A": 2000,
   "B": 875
  },
  "selected": "A",
  "ranking": [
   "A",
   "B"
  ],
  "verdicts": [
   {
    "label": "A",
    "score": 2000,
    "verdict_ko": "주먹을 꽉 쥐어 하얗게 질린 손가락 마디의 디테일을 잘 살렸고 나무 뒤의 구도를 훌륭히 구현했으나, 위쪽 낫의 날이 부러진 것처럼 묘사된 점이 아쉽습니다."
   },
   {
    "label": "B",
    "score": 875,
    "verdict_ko": "두 개의 낫이 한 덩어리로 엉겨 붙어 물리적으로 불가능한 형태를 띠는 치명적인 사물 융합 오류가 있습니다.  ★위반: [gemini-pro] 물리적으로 불가능한 사물 융합: 왼손이 쥔 낫 자루가 오른손이 쥔 낫의 금속 깃 부분과 한 덩어리로 융합되어 물리적으로 불가능한 형태를 만듦."
   }
  ],
  "refs": [
   {
    "label": "LOCATION PHOTOGRAPH — the exact place of this shot: its architecture, materials, fixed features and lighting mood are spatial truth; stage the moment inside this place. Never copy its camera framing.",
    "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/db00517b-165b-4005-a287-f28936b59b95/images/272bd7af-9294-4490-98c6-88a40e9deacc/scene/recipe/groupbg_ruined_camp_treeline_e39a9f.png",
    "asset_id": "ce860305-92d9-4331-bf03-df567841a4aa",
    "role": "bgfirst_group_bg"
   },
   {
    "label": "CHARACTER REFERENCE — 안규홍: the exact person appearing in this shot; match face, hair and build exactly.",
    "path": "<bytes:1483645>",
    "asset_id": "b52dff85-b663-44f9-aa9c-46b828990f55",
    "role": "character_ref"
   },
   {
    "label": "PROP REFERENCE — 낫: the exact object appearing in this shot; match its look, material and wear exactly.",
    "path": "<bytes:977465>",
    "asset_id": "e9884693-b2aa-4769-9ee0-4c7b50db40c2",
    "role": "prop_ref"
   }
  ],
  "critique_skipped": true,
  "shot_run_produced": false,
  "shot_run_uid": "06aa1443-2e9e-715c-9bce-4e652937034b",
  "bgfirst": {
   "bg_path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/db00517b-165b-4005-a287-f28936b59b95/images/272bd7af-9294-4490-98c6-88a40e9deacc/scene/recipe/S60sh12__bgfirst_bg.png",
   "bg_asset_id": "4c578bd0-065c-4d45-85cd-3aeb86ee8abb",
   "bg_record_key": "S60sh12::bgfirst_bg",
   "chain_winner": true,
   "authority": "groupbg",
   "group_key": "ruined_camp_treeline",
   "groupbg_asset_id": "ce860305-92d9-4331-bf03-df567841a4aa"
  },
  "ref_mode": "재투영 배경+콘티+엔티티 (2택1: 체인 승)",
  "share_plan": {
   "ref_plan": "background"
  }
 },
 "S60sh12::cine": {
  "applied": false,
  "attempted_at": "2026-09-09T09:36:02.836885+00:00",
  "fingerprint": "c4966dae516911eb9e7fff00ea863f3baa6aaa14b05dcfda25fcf4422aa8aa04",
  "fingerprint_version": 2,
  "provider": "mai",
  "endpoint": "openrouter/chat-completions",
  "model": "microsoft/mai-image-2.6",
  "multiroll_tag": "still_S60sh12_cine",
  "slot": "",
  "pack": "24.202608252115",
  "source_file": "S60sh12_sel.png",
  "source_sha256": "97a4a2fdd1c34a68c50e496eff39c8e5c428ab56df863333e0626cf21c83418f",
  "error": "RuntimeError: moderation blocked: {\"message\": \"Output content violated imagegen safety policies.\", \"code\": 400}",
  "moderation_refusals": 2,
  "declined": true,
  "declined_reason": "moderation"
 },
 "S61sh4::signage": {
  "fp": "a87fae6f9087d309",
  "inscriptions": [],
  "cues": [],
  "dropped": []
 },
 "S61sh4": {
  "input_fingerprint": "c2b01baa277afa67",
  "prompt": "Create ONE FINAL photorealistic live-action film still of the moment below — early 20th-century Korea, Jeolla region; all people are Korean unless explicitly stated otherwise. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 저항하며 몸을 뒤틀고 있는 일본인 군인의 입을 손으로 꽉 틀어막은 안규홍의 역동적인 상체.\n\nLOCATION (lock): The tree-shadowed exterior behind a collapsed mountain-camp hut, immediately adjoining dense forest cover. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: medium shot\n- KEY BACKGROUND ELEMENTS: Collapsed hut (Collapsed, with the struggle occurring behind it) — Only a partial rear aspect remains at the frame margin; used as Anchors the ambush location without competing with the joined upper bodies; Forest trees (Along the route into which the soldier is being pulled); used as Maintain continuity with the backward pull into the forest.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Daylight and the established tree shadows maintain grounded contrast across the struggling bodies without stylized highlights.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The people visible in that photo are NOT in this shot: never carry their faces, bodies or clothing onto anyone here. Each person in THIS shot is defined solely by the PEOPLE list, their own CHARACTER REFERENCE images and their wardrobe notes. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The collapsed hut and timber remain among smoke, shattered dishes and bloodied rice. 안규홍: He holds a close restraining stance with both arms engaged, retaining his two sickles; blood and dirt cover him and his back wound continues bleeding.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 안규홍 (한국인 남성, 청년 나이대, 검은 머리, 단단히 묶은 머리) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.",
  "roll_prompts": {
   "A": "Create ONE FINAL photorealistic live-action film still of the moment below — early 20th-century Korea, Jeolla region; all people are Korean unless explicitly stated otherwise. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 저항하며 몸을 뒤틀고 있는 일본인 군인의 입을 손으로 꽉 틀어막은 안규홍의 역동적인 상체.\n\nLOCATION (lock): The tree-shadowed exterior behind a collapsed mountain-camp hut, immediately adjoining dense forest cover. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: medium shot\n- KEY BACKGROUND ELEMENTS: Collapsed hut (Collapsed, with the struggle occurring behind it) — Only a partial rear aspect remains at the frame margin; used as Anchors the ambush location without competing with the joined upper bodies; Forest trees (Along the route into which the soldier is being pulled); used as Maintain continuity with the backward pull into the forest.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Daylight and the established tree shadows maintain grounded contrast across the struggling bodies without stylized highlights.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The people visible in that photo are NOT in this shot: never carry their faces, bodies or clothing onto anyone here. Each person in THIS shot is defined solely by the PEOPLE list, their own CHARACTER REFERENCE images and their wardrobe notes. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The collapsed hut and timber remain among smoke, shattered dishes and bloodied rice. 안규홍: He holds a close restraining stance with both arms engaged, retaining his two sickles; blood and dirt cover him and his back wound continues bleeding.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 안규홍 (한국인 남성, 청년 나이대, 검은 머리, 단단히 묶은 머리) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.",
   "B": "Create ONE FINAL photorealistic live-action film still of the moment below — early 20th-century Korea, Jeolla region; all people are Korean unless explicitly stated otherwise. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 저항하며 몸을 뒤틀고 있는 일본인 군인의 입을 손으로 꽉 틀어막은 안규홍의 역동적인 상체.\n\nLOCATION (lock): The tree-shadowed exterior behind a collapsed mountain-camp hut, immediately adjoining dense forest cover. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: medium shot\n- KEY BACKGROUND ELEMENTS: Collapsed hut (Collapsed, with the struggle occurring behind it) — Only a partial rear aspect remains at the frame margin; used as Anchors the ambush location without competing with the joined upper bodies; Forest trees (Along the route into which the soldier is being pulled); used as Maintain continuity with the backward pull into the forest.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Daylight and the established tree shadows maintain grounded contrast across the struggling bodies without stylized highlights.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The people visible in that photo are NOT in this shot: never carry their faces, bodies or clothing onto anyone here. Each person in THIS shot is defined solely by the PEOPLE list, their own CHARACTER REFERENCE images and their wardrobe notes. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The collapsed hut and timber remain among smoke, shattered dishes and bloodied rice. 안규홍: He holds a close restraining stance with both arms engaged, retaining his two sickles; blood and dirt cover him and his back wound continues bleeding.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 안규홍 (한국인 남성, 청년 나이대, 검은 머리, 단단히 묶은 머리) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.\n\nCOMPOSITION VARIATION (this candidate only): keep the place, the people and the moment exactly as specified, and keep from the CAMERA & FRAME contract its subject scale, screen placement and key background elements exactly. For this candidate only, its camera angle and height are deliberately left open: choose a camera position distinctly different from the obvious one, as a film director picking a second setup on the same blocking. Stay true to any explicit framing the SHOT TEXT itself dictates."
  },
  "shot_run_spend_attempt_count": 1,
  "readings": [
   {
    "label": "A",
    "direction": "안규홍의 시선은 일본군을 향하고 있으며, 그의 왼손은 목표인 일본군의 입을 정확히 꽉 틀어막고 있음.",
    "built_space": "왼쪽에 짚단이 묶인 흙벽이 이전 샷 스틸과 동일하게 온전한 형태로 나타나며, 프롬프트가 요구한 '붕괴된 오두막'의 모습은 반영되지 않음. 배경으로 숲이 보임.",
    "entities": "안규홍은 흰 옷을 입고 머리를 묶은 모습이나 지시된 2개의 낫을 들고 있지 않음. 일본군은 군복을 입고 있음.",
    "hard_violations": [],
    "physics": "안규홍은 지면에 발을 딛고 서서 힘을 주고 있으며, 일본군은 한쪽 다리를 공중에 치켜들고 몸부림치는 자세가 지면에 닿은 왼발과 안규홍의 제압에 의해 자연스럽게 지지되고 있음."
   },
   {
    "label": "B",
    "direction": "오른쪽 인물이 시선을 왼쪽 인물에게 고정한 채, 오른손으로 왼쪽 인물의 입을 막고 있음.",
    "built_space": "배경에 무너진 흙집 잔해, 나무 기둥, 깨진 도자기, 붉은 얼룩, 연기 등 프롬프트의 지시가 충실히 구현됨.",
    "entities": "지시된 2개의 낫이 보이지 않음. 일본군 복장을 한 인물이 안규홍의 레퍼런스 얼굴을 하고 있으며, 안규홍 역할의 인물이 이전 샷 스틸 인물의 얼굴을 하고 있어 신원이 완전히 뒤바뀜.",
    "hard_violations": [
     "[gemini-pro] 이전 샷 스틸에 등장한 인물의 얼굴을 이 샷에 절대 사용하지 말라는 지시(never carry their faces)를 명백히 위반하여 오른쪽 인물에 적용함.",
     "[gemini-pro] 일본군(왼쪽 인물)에게 안규홍의 캐릭터 레퍼런스 얼굴을 잘못 적용하여 신원과 역할을 뒤바꿈."
    ],
    "physics": "오른쪽 인물이 두 발로 서서 체중을 지탱하며 왼쪽 인물을 붙잡고 있고, 왼쪽 인물은 뒤로 밀리며 오른쪽 인물의 팔에 의해 지지되고 있음."
   }
  ],
  "cross_model_order": {
   "policy": "cross_model_order_v2_slot0_fwd_slot1_rev_parallel2",
   "slots": [
    {
     "model": "gemini-pro",
     "order": "forward",
     "display_to_canonical": {
      "A": "A",
      "B": "B"
     },
     "raw": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 4,
        "verdict_ko": "붕괴된 오두막 배경과 두 개의 낫 등 디테일을 누락했으나, B와 같은 치명적인 신원 오류 없이 인물들의 역동적인 격투 자세를 무난하게 구현함."
       },
       {
        "label": "B",
        "score": 1,
        "verdict_ko": "무너진 오두막 배경 묘사는 우수하나, 금지된 이전 샷 인물의 얼굴을 사용하고 일본군에게 안규홍의 얼굴을 적용하는 치명적인 신원 위반을 범함."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "안규홍의 시선은 일본군을 향하고 있으며, 그의 왼손은 목표인 일본군의 입을 정확히 꽉 틀어막고 있음.",
        "built_space": "왼쪽에 짚단이 묶인 흙벽이 이전 샷 스틸과 동일하게 온전한 형태로 나타나며, 프롬프트가 요구한 '붕괴된 오두막'의 모습은 반영되지 않음. 배경으로 숲이 보임.",
        "entities": "안규홍은 흰 옷을 입고 머리를 묶은 모습이나 지시된 2개의 낫을 들고 있지 않음. 일본군은 군복을 입고 있음.",
        "hard_violations": [],
        "physics": "안규홍은 지면에 발을 딛고 서서 힘을 주고 있으며, 일본군은 한쪽 다리를 공중에 치켜들고 몸부림치는 자세가 지면에 닿은 왼발과 안규홍의 제압에 의해 자연스럽게 지지되고 있음."
       },
       {
        "label": "B",
        "direction": "오른쪽 인물이 시선을 왼쪽 인물에게 고정한 채, 오른손으로 왼쪽 인물의 입을 막고 있음.",
        "built_space": "배경에 무너진 흙집 잔해, 나무 기둥, 깨진 도자기, 붉은 얼룩, 연기 등 프롬프트의 지시가 충실히 구현됨.",
        "entities": "지시된 2개의 낫이 보이지 않음. 일본군 복장을 한 인물이 안규홍의 레퍼런스 얼굴을 하고 있으며, 안규홍 역할의 인물이 이전 샷 스틸 인물의 얼굴을 하고 있어 신원이 완전히 뒤바뀜.",
        "hard_violations": [
         "이전 샷 스틸에 등장한 인물의 얼굴을 이 샷에 절대 사용하지 말라는 지시(never carry their faces)를 명백히 위반하여 오른쪽 인물에 적용함.",
         "일본군(왼쪽 인물)에게 안규홍의 캐릭터 레퍼런스 얼굴을 잘못 적용하여 신원과 역할을 뒤바꿈."
        ],
        "physics": "오른쪽 인물이 두 발로 서서 체중을 지탱하며 왼쪽 인물을 붙잡고 있고, 왼쪽 인물은 뒤로 밀리며 오른쪽 인물의 팔에 의해 지지되고 있음."
       }
      ],
      "all_candidates_fail": false
     },
     "normalized": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 4,
        "verdict_ko": "붕괴된 오두막 배경과 두 개의 낫 등 디테일을 누락했으나, B와 같은 치명적인 신원 오류 없이 인물들의 역동적인 격투 자세를 무난하게 구현함."
       },
       {
        "label": "B",
        "score": 1,
        "verdict_ko": "무너진 오두막 배경 묘사는 우수하나, 금지된 이전 샷 인물의 얼굴을 사용하고 일본군에게 안규홍의 얼굴을 적용하는 치명적인 신원 위반을 범함."
       }
      ],
      "all_candidates_fail": false,
      "readings": [
       {
        "label": "A",
        "direction": "안규홍의 시선은 일본군을 향하고 있으며, 그의 왼손은 목표인 일본군의 입을 정확히 꽉 틀어막고 있음.",
        "built_space": "왼쪽에 짚단이 묶인 흙벽이 이전 샷 스틸과 동일하게 온전한 형태로 나타나며, 프롬프트가 요구한 '붕괴된 오두막'의 모습은 반영되지 않음. 배경으로 숲이 보임.",
        "entities": "안규홍은 흰 옷을 입고 머리를 묶은 모습이나 지시된 2개의 낫을 들고 있지 않음. 일본군은 군복을 입고 있음.",
        "hard_violations": [],
        "physics": "안규홍은 지면에 발을 딛고 서서 힘을 주고 있으며, 일본군은 한쪽 다리를 공중에 치켜들고 몸부림치는 자세가 지면에 닿은 왼발과 안규홍의 제압에 의해 자연스럽게 지지되고 있음."
       },
       {
        "label": "B",
        "direction": "오른쪽 인물이 시선을 왼쪽 인물에게 고정한 채, 오른손으로 왼쪽 인물의 입을 막고 있음.",
        "built_space": "배경에 무너진 흙집 잔해, 나무 기둥, 깨진 도자기, 붉은 얼룩, 연기 등 프롬프트의 지시가 충실히 구현됨.",
        "entities": "지시된 2개의 낫이 보이지 않음. 일본군 복장을 한 인물이 안규홍의 레퍼런스 얼굴을 하고 있으며, 안규홍 역할의 인물이 이전 샷 스틸 인물의 얼굴을 하고 있어 신원이 완전히 뒤바뀜.",
        "hard_violations": [
         "이전 샷 스틸에 등장한 인물의 얼굴을 이 샷에 절대 사용하지 말라는 지시(never carry their faces)를 명백히 위반하여 오른쪽 인물에 적용함.",
         "일본군(왼쪽 인물)에게 안규홍의 캐릭터 레퍼런스 얼굴을 잘못 적용하여 신원과 역할을 뒤바꿈."
        ],
        "physics": "오른쪽 인물이 두 발로 서서 체중을 지탱하며 왼쪽 인물을 붙잡고 있고, 왼쪽 인물은 뒤로 밀리며 오른쪽 인물의 팔에 의해 지지되고 있음."
       }
      ]
     },
     "ok": true
    },
    {
     "model": "gpt-high",
     "order": "reverse",
     "display_to_canonical": {
      "A": "B",
      "B": "A"
     },
     "raw": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 6,
        "verdict_ko": "입을 틀어막으며 밀착해 저항을 제압하는 상체 중심 미디엄 숏이 우세하지만, 안규홍의 얼굴과 묶은 머리 재현은 크게 어긋난다."
       },
       {
        "label": "B",
        "score": 5,
        "verdict_ko": "안규홍의 묶은 머리와 피 묻은 의상은 충실하지만, 다리까지 확장한 구도와 팔을 뻗어 미는 듯한 동작이 밀착된 상체 중심 제압 장면에서 벗어난다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "오른쪽 안규홍 역할 인물은 왼쪽 군인의 얼굴을 바라보며 손바닥으로 입을 직접 덮는다. 군인은 고개를 뒤로 젖혀 위쪽을 보고 상대의 옷깃을 붙잡는다. 두 상체가 맞물려 있고 안규홍이 숲이 있는 오른쪽에서 버티므로 숲 쪽으로 끌어당기는 관계는 가능하지만, 실제 후퇴 방향까지 확정되지는 않는다. 겨누는 무기는 보이지 않는다.",
        "built_space": "왼쪽 가장자리에 흙벽 잔해 하나, 중앙 뒤에 낮게 남은 흙벽 구간 하나가 있으며, 그 사이와 전경에 무너진 목재가 놓여 있다. 연기와 깨진 그릇도 보이고 오른쪽과 뒤는 빽빽한 숲이다. 흙벽과 목재 재질은 장소 참고와 연결되지만, 무너진 오두막이 왼쪽 배경의 상당 부분을 차지해 가장자리의 작은 앵커라는 요구보다 두드러진다. 두 인물은 잔해 앞의 숲 인접 공간에서 허리 부근까지 잡힌다.",
        "entities": "성인 동아시아계 남성 두 명이 보이며, 군인의 국적은 외모만으로 확인할 수 없지만 군복과 계급장으로 일본군 역할이 표현된다. 안규홍 역할 인물의 짧고 묶이지 않은 머리와 얼굴은 지정 인물 참고와 다르고, 오히려 이전 숏 인물에 가깝다. 군인 쪽에 긴 검은 머리가 배치되어 정체성 구분도 약하다. 안규홍의 때 묻은 밝은 저고리와 등에 넓게 번진 피는 맞는다. 낫 두 자루와 피 묻은 쌀은 확인되지 않으며, 허리 아래와 일부 손은 가려져 있다. 읽을 수 있는 글자는 없다.",
        "hard_violations": [],
        "physics": "입을 덮은 손과 팔이 안규홍의 어깨로 이어지고, 다른 팔은 군인의 몸통을 감싸는 위치에 있다. 군인의 손은 안규홍의 옷깃을 실제로 잡고 있어 저항의 접촉점이 분명하다. 군인의 젖혀진 상체는 상대의 팔과 몸에 받쳐져 있으며, 하체와 발은 프레임 밖이므로 지면 접촉은 확인되지 않는다. 공중에 떠 있는 자세로 보이지 않으며 목재와 그릇은 지면 또는 잔해에 놓여 있다."
       },
       {
        "label": "B",
        "direction": "왼쪽 안규홍은 오른쪽 군인의 얼굴을 보며 한 손으로 입을 덮고 다른 팔을 목 뒤쪽으로 뻗는다. 군인은 눈을 감거나 아래로 향한 채 한 손으로 상대 팔을 붙잡고, 다른 팔은 오른쪽으로 벌린다. 군인의 몸과 들린 무릎은 오른쪽 숲 쪽으로 향하지만, 안규홍의 뻗은 팔 때문에 뒤로 끌어당기기보다 앞쪽으로 밀어내는 동작에 가깝다. 무기는 보이지 않는다.",
        "built_space": "왼쪽에 흙벽 한 면, 수직 목재 기둥 하나, 이어진 초가 처마 한 구간이 보인다. 벽에는 수평으로 묶인 짚이 있어 장소 참고의 재료와 고정 요소를 잘 잇는다. 다만 남은 벽과 지붕이 비교적 온전하고 무너진 목재와 연기는 뚜렷하지 않아 붕괴한 오두막 뒤라는 상태가 약하다. 오른쪽과 뒤에는 숲이 펼쳐진다. 안규홍의 허벅지와 군인의 들린 다리까지 보여 요구된 상체 중심 미디엄 숏보다 넓다.",
        "entities": "성인 동아시아계 남성 두 명이 보인다. 안규홍의 검은 묶은 머리, 옆얼굴, 체격, 낡은 밝은 저고리와 등의 피는 인물 참고에 더 가깝다. 군인은 군모와 황갈색 군복, 가죽 장구를 착용해 일본군 역할로 읽히지만 국적 자체는 외모로 확인할 수 없다. 안규홍의 두 팔은 모두 제압에 쓰이나 보이는 손에 낫은 없으며, 다른 위치에 두 자루를 보유했는지도 확인되지 않는다. 깨진 그릇과 피 묻은 쌀, 연기는 식별되지 않는다. 읽을 수 있는 글자는 없다.",
        "hard_violations": [],
        "physics": "입을 덮는 손, 목 뒤에 닿는 다른 손, 이를 붙잡는 군인의 손이 각각 접촉점을 만든다. 군인은 한쪽 무릎을 들고 반대쪽 다리를 화면 아래로 내려 몸을 지탱하는 자세이며, 발의 지면 접촉은 프레임 밖이다. 안규홍도 하체를 낮추고 버티는 자세다. 들린 다리는 고관절과 무릎의 굽힘으로 가능한 저항 동작이며, 근거 없이 몸 전체가 떠 있는 장면은 아니다."
       }
      ],
      "all_candidates_fail": false
     },
     "normalized": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 6,
        "verdict_ko": "입을 틀어막으며 밀착해 저항을 제압하는 상체 중심 미디엄 숏이 우세하지만, 안규홍의 얼굴과 묶은 머리 재현은 크게 어긋난다."
       },
       {
        "label": "A",
        "score": 5,
        "verdict_ko": "안규홍의 묶은 머리와 피 묻은 의상은 충실하지만, 다리까지 확장한 구도와 팔을 뻗어 미는 듯한 동작이 밀착된 상체 중심 제압 장면에서 벗어난다."
       }
      ],
      "all_candidates_fail": false,
      "readings": [
       {
        "label": "B",
        "direction": "오른쪽 안규홍 역할 인물은 왼쪽 군인의 얼굴을 바라보며 손바닥으로 입을 직접 덮는다. 군인은 고개를 뒤로 젖혀 위쪽을 보고 상대의 옷깃을 붙잡는다. 두 상체가 맞물려 있고 안규홍이 숲이 있는 오른쪽에서 버티므로 숲 쪽으로 끌어당기는 관계는 가능하지만, 실제 후퇴 방향까지 확정되지는 않는다. 겨누는 무기는 보이지 않는다.",
        "built_space": "왼쪽 가장자리에 흙벽 잔해 하나, 중앙 뒤에 낮게 남은 흙벽 구간 하나가 있으며, 그 사이와 전경에 무너진 목재가 놓여 있다. 연기와 깨진 그릇도 보이고 오른쪽과 뒤는 빽빽한 숲이다. 흙벽과 목재 재질은 장소 참고와 연결되지만, 무너진 오두막이 왼쪽 배경의 상당 부분을 차지해 가장자리의 작은 앵커라는 요구보다 두드러진다. 두 인물은 잔해 앞의 숲 인접 공간에서 허리 부근까지 잡힌다.",
        "entities": "성인 동아시아계 남성 두 명이 보이며, 군인의 국적은 외모만으로 확인할 수 없지만 군복과 계급장으로 일본군 역할이 표현된다. 안규홍 역할 인물의 짧고 묶이지 않은 머리와 얼굴은 지정 인물 참고와 다르고, 오히려 이전 숏 인물에 가깝다. 군인 쪽에 긴 검은 머리가 배치되어 정체성 구분도 약하다. 안규홍의 때 묻은 밝은 저고리와 등에 넓게 번진 피는 맞는다. 낫 두 자루와 피 묻은 쌀은 확인되지 않으며, 허리 아래와 일부 손은 가려져 있다. 읽을 수 있는 글자는 없다.",
        "hard_violations": [],
        "physics": "입을 덮은 손과 팔이 안규홍의 어깨로 이어지고, 다른 팔은 군인의 몸통을 감싸는 위치에 있다. 군인의 손은 안규홍의 옷깃을 실제로 잡고 있어 저항의 접촉점이 분명하다. 군인의 젖혀진 상체는 상대의 팔과 몸에 받쳐져 있으며, 하체와 발은 프레임 밖이므로 지면 접촉은 확인되지 않는다. 공중에 떠 있는 자세로 보이지 않으며 목재와 그릇은 지면 또는 잔해에 놓여 있다."
       },
       {
        "label": "A",
        "direction": "왼쪽 안규홍은 오른쪽 군인의 얼굴을 보며 한 손으로 입을 덮고 다른 팔을 목 뒤쪽으로 뻗는다. 군인은 눈을 감거나 아래로 향한 채 한 손으로 상대 팔을 붙잡고, 다른 팔은 오른쪽으로 벌린다. 군인의 몸과 들린 무릎은 오른쪽 숲 쪽으로 향하지만, 안규홍의 뻗은 팔 때문에 뒤로 끌어당기기보다 앞쪽으로 밀어내는 동작에 가깝다. 무기는 보이지 않는다.",
        "built_space": "왼쪽에 흙벽 한 면, 수직 목재 기둥 하나, 이어진 초가 처마 한 구간이 보인다. 벽에는 수평으로 묶인 짚이 있어 장소 참고의 재료와 고정 요소를 잘 잇는다. 다만 남은 벽과 지붕이 비교적 온전하고 무너진 목재와 연기는 뚜렷하지 않아 붕괴한 오두막 뒤라는 상태가 약하다. 오른쪽과 뒤에는 숲이 펼쳐진다. 안규홍의 허벅지와 군인의 들린 다리까지 보여 요구된 상체 중심 미디엄 숏보다 넓다.",
        "entities": "성인 동아시아계 남성 두 명이 보인다. 안규홍의 검은 묶은 머리, 옆얼굴, 체격, 낡은 밝은 저고리와 등의 피는 인물 참고에 더 가깝다. 군인은 군모와 황갈색 군복, 가죽 장구를 착용해 일본군 역할로 읽히지만 국적 자체는 외모로 확인할 수 없다. 안규홍의 두 팔은 모두 제압에 쓰이나 보이는 손에 낫은 없으며, 다른 위치에 두 자루를 보유했는지도 확인되지 않는다. 깨진 그릇과 피 묻은 쌀, 연기는 식별되지 않는다. 읽을 수 있는 글자는 없다.",
        "hard_violations": [],
        "physics": "입을 덮는 손, 목 뒤에 닿는 다른 손, 이를 붙잡는 군인의 손이 각각 접촉점을 만든다. 군인은 한쪽 무릎을 들고 반대쪽 다리를 화면 아래로 내려 몸을 지탱하는 자세이며, 발의 지면 접촉은 프레임 밖이다. 안규홍도 하체를 낮추고 버티는 자세다. 들린 다리는 고관절과 무릎의 굽힘으로 가능한 저항 동작이며, 근거 없이 몸 전체가 떠 있는 장면은 아니다."
       }
      ]
     },
     "ok": true
    }
   ],
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "slot_winner_match": false,
   "slot_winner": {
    "gemini-pro": "A",
    "gpt-high": "B"
   },
   "route": "cross_slot_combined"
  },
  "dual": {
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "normalized": {
    "A": 1.833,
    "B": 1.25
   },
   "adjusted": {
    "A": 1.833,
    "B": 1.0
   },
   "violations": {
    "B": [
     "[gemini-pro] 이전 샷 스틸에 등장한 인물의 얼굴을 이 샷에 절대 사용하지 말라는 지시(never carry their faces)를 명백히 위반하여 오른쪽 인물에 적용함.",
     "[gemini-pro] 일본군(왼쪽 인물)에게 안규홍의 캐릭터 레퍼런스 얼굴을 잘못 적용하여 신원과 역할을 뒤바꿈."
    ]
   },
   "per_model_winner": {
    "gemini-pro": "A",
    "gpt-high": "B"
   },
   "agreed": false
  },
  "totals": {
   "A": 1833,
   "B": 1000
  },
  "selected": "A",
  "ranking": [
   "A",
   "B"
  ],
  "verdicts": [
   {
    "label": "A",
    "score": 1833,
    "verdict_ko": "붕괴된 오두막 배경과 두 개의 낫 등 디테일을 누락했으나, B와 같은 치명적인 신원 오류 없이 인물들의 역동적인 격투 자세를 무난하게 구현함."
   },
   {
    "label": "B",
    "score": 1000,
    "verdict_ko": "무너진 오두막 배경 묘사는 우수하나, 금지된 이전 샷 인물의 얼굴을 사용하고 일본군에게 안규홍의 얼굴을 적용하는 치명적인 신원 위반을 범함.  ★위반: [gemini-pro] 이전 샷 스틸에 등장한 인물의 얼굴을 이 샷에 절대 사용하지 말라는 지시(never carry their faces)를 명백히 위반하여 오른쪽 인물에 적용함. / [gemini-pro] 일본군(왼쪽 인물)에게 안규홍의 캐릭터 레퍼런스 얼굴을 잘못 적용하여 신원과 역할을 뒤바꿈."
   }
  ],
  "refs": [
   {
    "label": "PREVIOUS SHOT STILL — a visually related earlier shot of this same place: the location's look, materials, fixed features and lighting mood are LOCKED to this photo; never copy its camera framing. The people visible in that photo are NOT in this shot: never carry their faces, bodies or clothing onto anyone here. Each person in THIS shot is defined solely by the PEOPLE list, their own CHARACTER REFERENCE images and their wardrobe notes. If this photo shows a character who cannot move (dead or unconscious), that character's exact pose, position and orientation are ALSO LOCKED — treat their body as a fixed prop of the set that only the camera moves around.",
    "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/db00517b-165b-4005-a287-f28936b59b95/images/272bd7af-9294-4490-98c6-88a40e9deacc/scene/recipe/S56sh31_sel.png",
    "asset_id": "53376905-56aa-46ec-abe7-e033c9597cdf",
    "role": "prev_still"
   },
   {
    "label": "CHARACTER REFERENCE — 안규홍: the exact person appearing in this shot; match face, hair and build exactly.",
    "path": "<bytes:1483645>",
    "asset_id": "b52dff85-b663-44f9-aa9c-46b828990f55",
    "role": "character_ref"
   }
  ],
  "critique_skipped": true,
  "shot_run_produced": false,
  "shot_run_uid": "06aa1444-09f6-7b1a-831a-f4e2b5291514",
  "ref_mode": "prev+엔티티",
  "share_plan": {
   "ref_plan": "prev",
   "prev_anchor_tag": "S56sh31"
  },
  "lane_policy": "ab_select_bypass:prev"
 },
 "S61sh4::cine": {
  "applied": true,
  "attempted_at": "2026-09-09T08:04:24.208892+00:00",
  "fingerprint": "5584718ee095f9facf9a5b0122413acd316a64650d367ed98a24e6c0238d9bef",
  "fingerprint_version": 2,
  "provider": "mai",
  "endpoint": "openrouter/chat-completions",
  "model": "microsoft/mai-image-2.6",
  "pack": "24.202608252115",
  "source_file": "S61sh4_sel.png",
  "source_sha256": "fe0662410fe7cd413ff762d8e8d1508b05e173b57def0607e8631ba82f457ce5",
  "file": "S61sh4_cine.png",
  "staged_sha256": "44946a15e3be59ed66359057a99a4ba29a747d9a3cd8342f5677bcffd5f83b38",
  "latency_ms": 24736
 },
 "S61sh15::signage": {
  "fp": "3fce52e505041fb4",
  "inscriptions": [],
  "cues": [],
  "dropped": []
 },
 "S61sh15": {
  "input_fingerprint": "48b82b4d5cd463ec",
  "prompt": "Create ONE FINAL photorealistic live-action film still of the moment below — early 20th-century Korea, Jeolla region; all people are Korean unless explicitly stated otherwise. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 어깨에 붉은 피가 흥건한 채 시신들 사이에 엎드린 염재보의 상체.\n\nLOCATION (lock): The dirt ground among fallen men near the wooded edge of the ruined mountain camp. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: medium shot\n- KEY BACKGROUND ELEMENTS: Matchlock (Gripped by 염재보's bloodied hand) — A partial side view lies low beside his prone torso; used as Connects his recognition to his continued readiness without obscuring the injured shoulder; Ground between fallen fighters (Occupied by the prone survivor among motionless bodies); used as Preserves visible intervals between the differently oriented bodies and 염재보.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Restrained daylight renders the blood-soaked shoulder clearly while maintaining quiet, naturalistic contrast around the surviving face.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The people visible in that photo are NOT in this shot: never carry their faces, bodies or clothing onto anyone here. Each person in THIS shot is defined solely by the PEOPLE list, their own CHARACTER REFERENCE images and their wardrobe notes. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nPREVIOUS STILL USAGE (follow exactly — what to take from the attached still and what to exclude; it governs place and objects only, never who is in this shot or how the camera sees them): Take the damaged camp materials, bloodied dirt, broken bowls, and spilled rice. Exclude the earlier corpse-inspection action and do not add intact dining arrangements.\n\nIMMOBILE CHARACTER POSE — CANONICAL (identical wherever this character appears in ANY panel; on any conflict THIS POSE WINS): The unidentified dead militia members are collapsed motionless across the camp ground. Their individual torso orientations, head directions, arm and leg arrangements, and contacts with nearby objects are not specified in the screenplay.\n\nIMMOBILE BODIES OBEY GRAVITY: a person who is dead or unconscious\nexerts NO muscular effort. Every part of their body — head, torso,\narms, hands, fingers, legs — rests fully on whatever supports it\n(floor, wall, furniture, their own lap) and hangs or slumps with\ngravity. NEVER show any part of an immobile person's body lifted,\nraised, held up in the air, or posed as if presenting something:\nan object in their grip stays clenched in a hand that itself lies\nfallen on a support — the hand does not hold the object up. If the\ncanonical pose leaves a body part unspecified, resolve it as the\nmost gravity-compliant, fully-supported position.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The camp's ruined cookware, shattered bowls, fallen timber and bloodied rice remain undisturbed. 염재보: He lies prone among bodies gripping his matchlock, his wounded shoulder saturated with blood.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 염재보 (한국인 남성, 성인 나이대, 자연스러운 검은 머리) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.",
  "roll_prompts": {
   "A": "Create ONE FINAL photorealistic live-action film still of the moment below — early 20th-century Korea, Jeolla region; all people are Korean unless explicitly stated otherwise. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 어깨에 붉은 피가 흥건한 채 시신들 사이에 엎드린 염재보의 상체.\n\nLOCATION (lock): The dirt ground among fallen men near the wooded edge of the ruined mountain camp. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: medium shot\n- KEY BACKGROUND ELEMENTS: Matchlock (Gripped by 염재보's bloodied hand) — A partial side view lies low beside his prone torso; used as Connects his recognition to his continued readiness without obscuring the injured shoulder; Ground between fallen fighters (Occupied by the prone survivor among motionless bodies); used as Preserves visible intervals between the differently oriented bodies and 염재보.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Restrained daylight renders the blood-soaked shoulder clearly while maintaining quiet, naturalistic contrast around the surviving face.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The people visible in that photo are NOT in this shot: never carry their faces, bodies or clothing onto anyone here. Each person in THIS shot is defined solely by the PEOPLE list, their own CHARACTER REFERENCE images and their wardrobe notes. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nPREVIOUS STILL USAGE (follow exactly — what to take from the attached still and what to exclude; it governs place and objects only, never who is in this shot or how the camera sees them): Take the damaged camp materials, bloodied dirt, broken bowls, and spilled rice. Exclude the earlier corpse-inspection action and do not add intact dining arrangements.\n\nIMMOBILE CHARACTER POSE — CANONICAL (identical wherever this character appears in ANY panel; on any conflict THIS POSE WINS): The unidentified dead militia members are collapsed motionless across the camp ground. Their individual torso orientations, head directions, arm and leg arrangements, and contacts with nearby objects are not specified in the screenplay.\n\nIMMOBILE BODIES OBEY GRAVITY: a person who is dead or unconscious\nexerts NO muscular effort. Every part of their body — head, torso,\narms, hands, fingers, legs — rests fully on whatever supports it\n(floor, wall, furniture, their own lap) and hangs or slumps with\ngravity. NEVER show any part of an immobile person's body lifted,\nraised, held up in the air, or posed as if presenting something:\nan object in their grip stays clenched in a hand that itself lies\nfallen on a support — the hand does not hold the object up. If the\ncanonical pose leaves a body part unspecified, resolve it as the\nmost gravity-compliant, fully-supported position.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The camp's ruined cookware, shattered bowls, fallen timber and bloodied rice remain undisturbed. 염재보: He lies prone among bodies gripping his matchlock, his wounded shoulder saturated with blood.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 염재보 (한국인 남성, 성인 나이대, 자연스러운 검은 머리) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.",
   "B": "Create ONE FINAL photorealistic live-action film still of the moment below — early 20th-century Korea, Jeolla region; all people are Korean unless explicitly stated otherwise. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 어깨에 붉은 피가 흥건한 채 시신들 사이에 엎드린 염재보의 상체.\n\nLOCATION (lock): The dirt ground among fallen men near the wooded edge of the ruined mountain camp. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: medium shot\n- KEY BACKGROUND ELEMENTS: Matchlock (Gripped by 염재보's bloodied hand) — A partial side view lies low beside his prone torso; used as Connects his recognition to his continued readiness without obscuring the injured shoulder; Ground between fallen fighters (Occupied by the prone survivor among motionless bodies); used as Preserves visible intervals between the differently oriented bodies and 염재보.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Restrained daylight renders the blood-soaked shoulder clearly while maintaining quiet, naturalistic contrast around the surviving face.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The people visible in that photo are NOT in this shot: never carry their faces, bodies or clothing onto anyone here. Each person in THIS shot is defined solely by the PEOPLE list, their own CHARACTER REFERENCE images and their wardrobe notes. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nPREVIOUS STILL USAGE (follow exactly — what to take from the attached still and what to exclude; it governs place and objects only, never who is in this shot or how the camera sees them): Take the damaged camp materials, bloodied dirt, broken bowls, and spilled rice. Exclude the earlier corpse-inspection action and do not add intact dining arrangements.\n\nIMMOBILE CHARACTER POSE — CANONICAL (identical wherever this character appears in ANY panel; on any conflict THIS POSE WINS): The unidentified dead militia members are collapsed motionless across the camp ground. Their individual torso orientations, head directions, arm and leg arrangements, and contacts with nearby objects are not specified in the screenplay.\n\nIMMOBILE BODIES OBEY GRAVITY: a person who is dead or unconscious\nexerts NO muscular effort. Every part of their body — head, torso,\narms, hands, fingers, legs — rests fully on whatever supports it\n(floor, wall, furniture, their own lap) and hangs or slumps with\ngravity. NEVER show any part of an immobile person's body lifted,\nraised, held up in the air, or posed as if presenting something:\nan object in their grip stays clenched in a hand that itself lies\nfallen on a support — the hand does not hold the object up. If the\ncanonical pose leaves a body part unspecified, resolve it as the\nmost gravity-compliant, fully-supported position.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The camp's ruined cookware, shattered bowls, fallen timber and bloodied rice remain undisturbed. 염재보: He lies prone among bodies gripping his matchlock, his wounded shoulder saturated with blood.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 염재보 (한국인 남성, 성인 나이대, 자연스러운 검은 머리) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.\n\nCOMPOSITION VARIATION (this candidate only): keep the place, the people and the moment exactly as specified, and keep from the CAMERA & FRAME contract its subject scale, screen placement and key background elements exactly. For this candidate only, its camera angle and height are deliberately left open: choose a camera position distinctly different from the obvious one, as a film director picking a second setup on the same blocking. Stay true to any explicit framing the SHOT TEXT itself dictates."
  },
  "shot_run_spend_attempt_count": 1,
  "readings": [
   {
    "label": "A",
    "direction": "인물의 시선은 화면 좌측 하단을 향하고 있으며, 쥐고 있는 소총의 총구도 좌측을 가리키고 있음.",
    "built_space": "파괴된 산채의 흙바닥, 나무 울타리, 초가 구조물, 가마솥과 쏟아진 쌀 등 배경 요소가 이전 샷의 환경과 일치하게 배치됨.",
    "entities": "염재보의 얼굴과 복장은 참조와 일치하나, 화승총(Matchlock)이 아닌 현대식 볼트액션 소총을 쥐고 있음. 요구된 '상체'보다 넓은 전신 위주의 프레이밍임.",
    "hard_violations": [],
    "physics": "바닥에 엎드린 자세로 지면에 완전히 닿아 무게가 지지되어 있으며, 오른손으로 총을 자연스럽게 쥐고 있음."
   },
   {
    "label": "B",
    "direction": "인물의 시선은 화면 정면과 약간 우측을 향하며, 쥐고 있는 화승총의 총구는 화면 좌측을 향함.",
    "built_space": "파괴된 산채의 흙바닥, 엎드린 시신들, 좌측의 가마솥과 쏟아진 쌀 등 이전 샷의 고정된 배경 요소와 시신들의 위치가 정확히 일치함.",
    "entities": "염재보의 인물 참조와 훌륭하게 일치하며, 붉은 피로 젖은 우측 어깨와 피 묻은 손에 쥔 화승총이 명확히 묘사됨. 샷 스케일 또한 지시된 미디엄 샷(상체)을 정확히 따름.",
    "hard_violations": [],
    "physics": "지면에 엎드린 상체의 무게감이 바닥에 자연스럽게 실려 있으며, 양손으로 화승총을 바닥과 밀착하여 지지하고 있음."
   }
  ],
  "cross_model_order": {
   "policy": "cross_model_order_v2_slot0_fwd_slot1_rev_parallel2",
   "slots": [
    {
     "model": "gemini-pro",
     "order": "forward",
     "display_to_canonical": {
      "A": "A",
      "B": "B"
     },
     "raw": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 10,
        "verdict_ko": "지정된 미디엄 샷과 상체 위주의 구도를 완벽하게 지켰으며, 붉은 피가 흥건한 어깨의 질감과 화승총의 디테일을 프롬프트에 맞게 훌륭하게 구현했습니다."
       },
       {
        "label": "A",
        "score": 6,
        "verdict_ko": "얼굴과 배경의 재현은 우수하나, 미디엄 샷과 상체라는 구도 지시를 벗어나 샷이 지나치게 넓어졌으며 요구된 화승총 대신 볼트액션 소총이 등장해 지시와 어긋납니다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "인물의 시선은 화면 좌측 하단을 향하고 있으며, 쥐고 있는 소총의 총구도 좌측을 가리키고 있음.",
        "built_space": "파괴된 산채의 흙바닥, 나무 울타리, 초가 구조물, 가마솥과 쏟아진 쌀 등 배경 요소가 이전 샷의 환경과 일치하게 배치됨.",
        "entities": "염재보의 얼굴과 복장은 참조와 일치하나, 화승총(Matchlock)이 아닌 현대식 볼트액션 소총을 쥐고 있음. 요구된 '상체'보다 넓은 전신 위주의 프레이밍임.",
        "hard_violations": [],
        "physics": "바닥에 엎드린 자세로 지면에 완전히 닿아 무게가 지지되어 있으며, 오른손으로 총을 자연스럽게 쥐고 있음."
       },
       {
        "label": "B",
        "direction": "인물의 시선은 화면 정면과 약간 우측을 향하며, 쥐고 있는 화승총의 총구는 화면 좌측을 향함.",
        "built_space": "파괴된 산채의 흙바닥, 엎드린 시신들, 좌측의 가마솥과 쏟아진 쌀 등 이전 샷의 고정된 배경 요소와 시신들의 위치가 정확히 일치함.",
        "entities": "염재보의 인물 참조와 훌륭하게 일치하며, 붉은 피로 젖은 우측 어깨와 피 묻은 손에 쥔 화승총이 명확히 묘사됨. 샷 스케일 또한 지시된 미디엄 샷(상체)을 정확히 따름.",
        "hard_violations": [],
        "physics": "지면에 엎드린 상체의 무게감이 바닥에 자연스럽게 실려 있으며, 양손으로 화승총을 바닥과 밀착하여 지지하고 있음."
       }
      ],
      "all_candidates_fail": false
     },
     "normalized": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 10,
        "verdict_ko": "지정된 미디엄 샷과 상체 위주의 구도를 완벽하게 지켰으며, 붉은 피가 흥건한 어깨의 질감과 화승총의 디테일을 프롬프트에 맞게 훌륭하게 구현했습니다."
       },
       {
        "label": "A",
        "score": 6,
        "verdict_ko": "얼굴과 배경의 재현은 우수하나, 미디엄 샷과 상체라는 구도 지시를 벗어나 샷이 지나치게 넓어졌으며 요구된 화승총 대신 볼트액션 소총이 등장해 지시와 어긋납니다."
       }
      ],
      "all_candidates_fail": false,
      "readings": [
       {
        "label": "A",
        "direction": "인물의 시선은 화면 좌측 하단을 향하고 있으며, 쥐고 있는 소총의 총구도 좌측을 가리키고 있음.",
        "built_space": "파괴된 산채의 흙바닥, 나무 울타리, 초가 구조물, 가마솥과 쏟아진 쌀 등 배경 요소가 이전 샷의 환경과 일치하게 배치됨.",
        "entities": "염재보의 얼굴과 복장은 참조와 일치하나, 화승총(Matchlock)이 아닌 현대식 볼트액션 소총을 쥐고 있음. 요구된 '상체'보다 넓은 전신 위주의 프레이밍임.",
        "hard_violations": [],
        "physics": "바닥에 엎드린 자세로 지면에 완전히 닿아 무게가 지지되어 있으며, 오른손으로 총을 자연스럽게 쥐고 있음."
       },
       {
        "label": "B",
        "direction": "인물의 시선은 화면 정면과 약간 우측을 향하며, 쥐고 있는 화승총의 총구는 화면 좌측을 향함.",
        "built_space": "파괴된 산채의 흙바닥, 엎드린 시신들, 좌측의 가마솥과 쏟아진 쌀 등 이전 샷의 고정된 배경 요소와 시신들의 위치가 정확히 일치함.",
        "entities": "염재보의 인물 참조와 훌륭하게 일치하며, 붉은 피로 젖은 우측 어깨와 피 묻은 손에 쥔 화승총이 명확히 묘사됨. 샷 스케일 또한 지시된 미디엄 샷(상체)을 정확히 따름.",
        "hard_violations": [],
        "physics": "지면에 엎드린 상체의 무게감이 바닥에 자연스럽게 실려 있으며, 양손으로 화승총을 바닥과 밀착하여 지지하고 있음."
       }
      ]
     },
     "ok": true
    },
    {
     "model": "gpt-high",
     "order": "reverse",
     "display_to_canonical": {
      "A": "B",
      "B": "A"
     },
     "raw": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 6,
        "verdict_ko": "피에 젖은 어깨와 상체 중심 크기는 맞지만, 팔로 상체를 높이 세우고 정면을 응시해 시신들 사이에 낮게 엎드린 순간보다 경계 자세가 강조된다."
       },
       {
        "label": "B",
        "score": 8,
        "verdict_ko": "상체와 얼굴을 땅 가까이 둔 엎드린 자세, 드러난 피투성이 어깨, 몸 옆에 낮게 놓인 총이 핵심 연출에 더 충실하지만 화승총 재현은 부족하다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "염재보는 고개를 들고 카메라 쪽을 거의 정면으로 바라본다. 총은 몸 앞을 가로지르며 총구가 화면 오른쪽 밖을 향하고, 특정 인물을 조준하지는 않는다. 시신들의 얼굴은 흙이나 옆쪽을 향한다.",
        "built_space": "뒤쪽에 왼편 일부가 잘린 움막 하나, 중앙 움막 하나, 오른편 움막 하나와 목책·숲이 보인다. 왼쪽 전경에는 넘어진 큰 솥 하나, 여러 그릇과 파편, 흩어진 쌀이 있다. 염재보는 시신 무리 앞의 빈 흙바닥에 있으며 주변 몸들과 분리되어 있다. 이전 장소의 재료와 배치는 대체로 이어지지만 배경까지 비교적 선명하게 드러난다.",
        "entities": "염재보는 검은 머리의 성인 한국인 남성으로 표현되며 얼굴과 낡은 짙은 회색 누비옷이 인물 참조와 대체로 맞는다. 화면 왼쪽 어깨는 붉은 피에 넓게 젖었고 총을 쥔 손에도 피가 보인다. 주변에는 쓰러진 성인 남성들이 있다. 총에는 부싯돌식으로 보이는 점화장치가 크게 드러나며, 요구된 화승총의 화승 구조는 확인되지 않는다. 읽을 수 있는 글자는 없다.",
        "hard_violations": [],
        "physics": "배와 하체는 땅에 놓이고 양팔의 팔꿈치와 전완이 상체를 받친다. 살아 있는 부상자가 취할 수 있는 자세이지만 가슴과 머리를 상당히 높이 들었다. 총은 양손이 잡고 개머리판은 지면에 닿아 있어 떠 있지 않다. 식별 가능한 주변 시신들의 몸과 팔다리는 땅이나 겹친 몸에 기대어 있다."
       },
       {
        "label": "B",
        "direction": "염재보는 얼굴을 낮춘 채 화면 왼쪽 앞의 지면 쪽을 비스듬히 바라본다. 총은 몸 옆에서 왼쪽 개머리판부터 오른쪽 총구로 이어지며 특정 사람을 겨누지 않는다. 주변 시신들은 서로 다른 방향으로 쓰러져 있다.",
        "built_space": "왼쪽·중앙·오른쪽의 움막 세 채와 뒤쪽 목책, 숲이 보인다. 왼쪽 가장자리에는 큰 솥 일부와 작은 조리용기, 바닥에는 깨진 그릇과 피 묻은 쌀이 있다. 염재보는 시신들 사이의 흙바닥에 낮게 누워 있고 앞쪽 총과 주변 시신 사이에도 지면 간격이 보인다. 온전한 식사 상차림이나 추가 건축물은 보이지 않는다.",
        "entities": "검은 머리, 젊은 성인 남성의 얼굴과 체격, 해진 회색 누비옷이 염재보 참조에 대체로 부합한다. 어깨에는 찢어진 천과 넓게 젖은 붉은 피가 드러난다. 총을 쥔 손의 혈흔은 A보다 약하게 보인다. 주변의 쓰러진 남성들, 그릇 파편, 쌀, 나무 잔해가 표현되어 있다. 총은 작은 노리쇠 돌기가 있는 소총처럼 보이며 화승총의 점화장치와 화승은 확인되지 않는다. 읽을 수 있는 글자는 없다.",
        "hard_violations": [],
        "physics": "가슴과 배가 흙바닥에 낮게 놓이고, 앞으로 뻗은 팔과 손도 총 및 지면 가까이에서 지지된다. 머리를 조금 드는 것은 살아 있는 인물의 목으로 가능한 동작이다. 총은 지면에 놓인 상태에서 손이 실제로 감싸 쥐고, 멜빵도 바닥에 늘어져 있다. 보이는 시신들은 지면이나 인접한 몸 위에 무게를 내려놓고 있다."
       }
      ],
      "all_candidates_fail": true
     },
     "normalized": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 6,
        "verdict_ko": "피에 젖은 어깨와 상체 중심 크기는 맞지만, 팔로 상체를 높이 세우고 정면을 응시해 시신들 사이에 낮게 엎드린 순간보다 경계 자세가 강조된다."
       },
       {
        "label": "A",
        "score": 8,
        "verdict_ko": "상체와 얼굴을 땅 가까이 둔 엎드린 자세, 드러난 피투성이 어깨, 몸 옆에 낮게 놓인 총이 핵심 연출에 더 충실하지만 화승총 재현은 부족하다."
       }
      ],
      "all_candidates_fail": true,
      "readings": [
       {
        "label": "B",
        "direction": "염재보는 고개를 들고 카메라 쪽을 거의 정면으로 바라본다. 총은 몸 앞을 가로지르며 총구가 화면 오른쪽 밖을 향하고, 특정 인물을 조준하지는 않는다. 시신들의 얼굴은 흙이나 옆쪽을 향한다.",
        "built_space": "뒤쪽에 왼편 일부가 잘린 움막 하나, 중앙 움막 하나, 오른편 움막 하나와 목책·숲이 보인다. 왼쪽 전경에는 넘어진 큰 솥 하나, 여러 그릇과 파편, 흩어진 쌀이 있다. 염재보는 시신 무리 앞의 빈 흙바닥에 있으며 주변 몸들과 분리되어 있다. 이전 장소의 재료와 배치는 대체로 이어지지만 배경까지 비교적 선명하게 드러난다.",
        "entities": "염재보는 검은 머리의 성인 한국인 남성으로 표현되며 얼굴과 낡은 짙은 회색 누비옷이 인물 참조와 대체로 맞는다. 화면 왼쪽 어깨는 붉은 피에 넓게 젖었고 총을 쥔 손에도 피가 보인다. 주변에는 쓰러진 성인 남성들이 있다. 총에는 부싯돌식으로 보이는 점화장치가 크게 드러나며, 요구된 화승총의 화승 구조는 확인되지 않는다. 읽을 수 있는 글자는 없다.",
        "hard_violations": [],
        "physics": "배와 하체는 땅에 놓이고 양팔의 팔꿈치와 전완이 상체를 받친다. 살아 있는 부상자가 취할 수 있는 자세이지만 가슴과 머리를 상당히 높이 들었다. 총은 양손이 잡고 개머리판은 지면에 닿아 있어 떠 있지 않다. 식별 가능한 주변 시신들의 몸과 팔다리는 땅이나 겹친 몸에 기대어 있다."
       },
       {
        "label": "A",
        "direction": "염재보는 얼굴을 낮춘 채 화면 왼쪽 앞의 지면 쪽을 비스듬히 바라본다. 총은 몸 옆에서 왼쪽 개머리판부터 오른쪽 총구로 이어지며 특정 사람을 겨누지 않는다. 주변 시신들은 서로 다른 방향으로 쓰러져 있다.",
        "built_space": "왼쪽·중앙·오른쪽의 움막 세 채와 뒤쪽 목책, 숲이 보인다. 왼쪽 가장자리에는 큰 솥 일부와 작은 조리용기, 바닥에는 깨진 그릇과 피 묻은 쌀이 있다. 염재보는 시신들 사이의 흙바닥에 낮게 누워 있고 앞쪽 총과 주변 시신 사이에도 지면 간격이 보인다. 온전한 식사 상차림이나 추가 건축물은 보이지 않는다.",
        "entities": "검은 머리, 젊은 성인 남성의 얼굴과 체격, 해진 회색 누비옷이 염재보 참조에 대체로 부합한다. 어깨에는 찢어진 천과 넓게 젖은 붉은 피가 드러난다. 총을 쥔 손의 혈흔은 A보다 약하게 보인다. 주변의 쓰러진 남성들, 그릇 파편, 쌀, 나무 잔해가 표현되어 있다. 총은 작은 노리쇠 돌기가 있는 소총처럼 보이며 화승총의 점화장치와 화승은 확인되지 않는다. 읽을 수 있는 글자는 없다.",
        "hard_violations": [],
        "physics": "가슴과 배가 흙바닥에 낮게 놓이고, 앞으로 뻗은 팔과 손도 총 및 지면 가까이에서 지지된다. 머리를 조금 드는 것은 살아 있는 인물의 목으로 가능한 동작이다. 총은 지면에 놓인 상태에서 손이 실제로 감싸 쥐고, 멜빵도 바닥에 늘어져 있다. 보이는 시신들은 지면이나 인접한 몸 위에 무게를 내려놓고 있다."
       }
      ]
     },
     "ok": true
    }
   ],
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "slot_winner_match": false,
   "slot_winner": {
    "gemini-pro": "B",
    "gpt-high": "A"
   },
   "route": "cross_slot_combined"
  },
  "dual": {
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "normalized": {
    "A": 1.6,
    "B": 1.75
   },
   "adjusted": {
    "A": 1.6,
    "B": 1.75
   },
   "violations": {},
   "per_model_winner": {
    "gemini-pro": "B",
    "gpt-high": "A"
   },
   "agreed": false
  },
  "totals": {
   "B": 1750,
   "A": 1600
  },
  "selected": "B",
  "ranking": [
   "B",
   "A"
  ],
  "verdicts": [
   {
    "label": "B",
    "score": 1750,
    "verdict_ko": "지정된 미디엄 샷과 상체 위주의 구도를 완벽하게 지켰으며, 붉은 피가 흥건한 어깨의 질감과 화승총의 디테일을 프롬프트에 맞게 훌륭하게 구현했습니다."
   },
   {
    "label": "A",
    "score": 1600,
    "verdict_ko": "얼굴과 배경의 재현은 우수하나, 미디엄 샷과 상체라는 구도 지시를 벗어나 샷이 지나치게 넓어졌으며 요구된 화승총 대신 볼트액션 소총이 등장해 지시와 어긋납니다."
   }
  ],
  "refs": [
   {
    "label": "PREVIOUS SHOT STILL — a visually related earlier shot of this same place: the location's look, materials, fixed features and lighting mood are LOCKED to this photo; never copy its camera framing. The people visible in that photo are NOT in this shot: never carry their faces, bodies or clothing onto anyone here. Each person in THIS shot is defined solely by the PEOPLE list, their own CHARACTER REFERENCE images and their wardrobe notes. If this photo shows a character who cannot move (dead or unconscious), that character's exact pose, position and orientation are ALSO LOCKED — treat their body as a fixed prop of the set that only the camera moves around.",
    "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/db00517b-165b-4005-a287-f28936b59b95/images/272bd7af-9294-4490-98c6-88a40e9deacc/scene/recipe/S60sh2_sel.png",
    "asset_id": "9a8a1465-bc86-433f-b821-e073cfb330f4",
    "role": "prev_still"
   },
   {
    "label": "CHARACTER REFERENCE — 염재보: the exact person appearing in this shot; match face, hair and build exactly.",
    "path": "<bytes:1331451>",
    "asset_id": "bad3b41f-75ca-4d66-8fce-161d96d99096",
    "role": "character_ref"
   }
  ],
  "critique_skipped": true,
  "shot_run_produced": false,
  "shot_run_uid": "06aa1444-0b9b-7888-89a6-b5ec023d2766",
  "ref_mode": "prev+엔티티",
  "share_plan": {
   "ref_plan": "prev",
   "prev_anchor_tag": "S60sh2"
  }
 },
 "S61sh15::cine": {
  "applied": false,
  "attempted_at": "2026-09-09T09:36:40.482949+00:00",
  "fingerprint": "93cb6d1ae3c65a85d7bff45328a98b944167de01bc9e3e59999451e93023bfc0",
  "fingerprint_version": 2,
  "provider": "mai",
  "endpoint": "openrouter/chat-completions",
  "model": "microsoft/mai-image-2.6",
  "multiroll_tag": "still_S61sh15_cine",
  "slot": "",
  "pack": "24.202608252115",
  "source_file": "S61sh15_sel.png",
  "source_sha256": "513ba4cbe189a2eaf0fe2ecb6481c107194717a5bd63aedd59ecd69e6fc0e195",
  "error": "RuntimeError: moderation blocked: {\"message\": \"Response content blocked by label 'MultiSeverity_ViolenceScore'.\", \"code\": 400}",
  "moderation_refusals": 2,
  "declined": true,
  "declined_reason": "moderation"
 },
 "S61sh32::signage": {
  "fp": "f21db3bcaa104115",
  "inscriptions": [],
  "cues": [],
  "dropped": []
 },
 "S61sh32": {
  "input_fingerprint": "2811b575b56a0239",
  "prompt": "Create ONE FINAL photorealistic live-action film still of the moment below — early 20th-century Korea, Jeolla region; all people are Korean unless explicitly stated otherwise. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 바닥에 엎드린 염재보의 화승총 총구에서 붉은 섬광이 뿜어져 나오는 찰나.\n\nLOCATION (lock): A prone firing position among bodies and trees at the ruined mountain camp's edge, opposite the decoy's position. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: close-up\n- KEY BACKGROUND ELEMENTS: Matchlock (Discharging while supported by 염재보) — Viewed from the side and slightly behind the muzzle, firing toward frame right; used as Separates muzzle, supporting hands, and prone body without exaggerated foreground scale; Ground beneath the shooter (Visible below the supported weapon and prone body); used as Supplies a continuous low-level spatial reference.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: The brief red muzzle flash punctuates otherwise restrained daylight and naturalistic contrast.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The clothing, hair and overall look of 염재보 — who appear both in that photo and in this shot — are LOCKED to that photo. Anyone else visible in that photo is NOT in this shot: never carry their face, body or clothing onto anyone here. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The ruined camp remains littered with broken bowls, overturned cookware and bloodied rice; the firing matchlock emits a brief muzzle flash.\n\nTHE HAND THAT IS DOING THIS: the object at the centre of this shot is being held, operated, read, aimed or handed over by 염재보 right now, so 염재보's hand — and as much of the wrist and forearm as the framing reaches — is in the frame, gripping or resting on that object exactly the way its real use requires. Match that hand to 염재보: its size, build, skin, age, grooming, sleeve and anything worn on it belong to that person and to no one else.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 염재보 (한국인 남성, 성인 나이대, 자연스러운 검은 머리) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.",
  "roll_prompts": {
   "A": "Create ONE FINAL photorealistic live-action film still of the moment below — early 20th-century Korea, Jeolla region; all people are Korean unless explicitly stated otherwise. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 바닥에 엎드린 염재보의 화승총 총구에서 붉은 섬광이 뿜어져 나오는 찰나.\n\nLOCATION (lock): A prone firing position among bodies and trees at the ruined mountain camp's edge, opposite the decoy's position. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: close-up\n- KEY BACKGROUND ELEMENTS: Matchlock (Discharging while supported by 염재보) — Viewed from the side and slightly behind the muzzle, firing toward frame right; used as Separates muzzle, supporting hands, and prone body without exaggerated foreground scale; Ground beneath the shooter (Visible below the supported weapon and prone body); used as Supplies a continuous low-level spatial reference.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: The brief red muzzle flash punctuates otherwise restrained daylight and naturalistic contrast.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The clothing, hair and overall look of 염재보 — who appear both in that photo and in this shot — are LOCKED to that photo. Anyone else visible in that photo is NOT in this shot: never carry their face, body or clothing onto anyone here. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The ruined camp remains littered with broken bowls, overturned cookware and bloodied rice; the firing matchlock emits a brief muzzle flash.\n\nTHE HAND THAT IS DOING THIS: the object at the centre of this shot is being held, operated, read, aimed or handed over by 염재보 right now, so 염재보's hand — and as much of the wrist and forearm as the framing reaches — is in the frame, gripping or resting on that object exactly the way its real use requires. Match that hand to 염재보: its size, build, skin, age, grooming, sleeve and anything worn on it belong to that person and to no one else.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 염재보 (한국인 남성, 성인 나이대, 자연스러운 검은 머리) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.",
   "B": "Create ONE FINAL photorealistic live-action film still of the moment below — early 20th-century Korea, Jeolla region; all people are Korean unless explicitly stated otherwise. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 바닥에 엎드린 염재보의 화승총 총구에서 붉은 섬광이 뿜어져 나오는 찰나.\n\nLOCATION (lock): A prone firing position among bodies and trees at the ruined mountain camp's edge, opposite the decoy's position. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: close-up\n- KEY BACKGROUND ELEMENTS: Matchlock (Discharging while supported by 염재보) — Viewed from the side and slightly behind the muzzle, firing toward frame right; used as Separates muzzle, supporting hands, and prone body without exaggerated foreground scale; Ground beneath the shooter (Visible below the supported weapon and prone body); used as Supplies a continuous low-level spatial reference.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: The brief red muzzle flash punctuates otherwise restrained daylight and naturalistic contrast.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The clothing, hair and overall look of 염재보 — who appear both in that photo and in this shot — are LOCKED to that photo. Anyone else visible in that photo is NOT in this shot: never carry their face, body or clothing onto anyone here. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The ruined camp remains littered with broken bowls, overturned cookware and bloodied rice; the firing matchlock emits a brief muzzle flash.\n\nTHE HAND THAT IS DOING THIS: the object at the centre of this shot is being held, operated, read, aimed or handed over by 염재보 right now, so 염재보's hand — and as much of the wrist and forearm as the framing reaches — is in the frame, gripping or resting on that object exactly the way its real use requires. Match that hand to 염재보: its size, build, skin, age, grooming, sleeve and anything worn on it belong to that person and to no one else.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 염재보 (한국인 남성, 성인 나이대, 자연스러운 검은 머리) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.\n\nCOMPOSITION VARIATION (this candidate only): keep the place, the people and the moment exactly as specified, and keep from the CAMERA & FRAME contract its subject scale, screen placement and key background elements exactly. For this candidate only, its camera angle and height are deliberately left open: choose a camera position distinctly different from the obvious one, as a film director picking a second setup on the same blocking. Stay true to any explicit framing the SHOT TEXT itself dictates."
  },
  "shot_run_spend_attempt_count": 1,
  "readings": [
   {
    "label": "A",
    "direction": "거대한 공중 부양 화승총 총열이 화면 왼쪽으로 붉은 광선을 발사함. 염재보는 정면을 향해 엎드려 있음.",
    "built_space": "부서진 초가집, 시체, 흩어진 그릇과 밥이 있는 폐허가 된 산채의 배경이 보임.",
    "entities": "바닥에 엎드려 화승총을 쥔 염재보가 있으나, 그의 얼굴 앞에 두 번째 거대한 화승총 총열이 떠 있음.",
    "hard_violations": [
     "[gemini-pro] 공중에 떠 있는 지지되지 않은 물체 (거대한 총열)",
     "[gemini-pro] 중복된 사물 (두 개의 화승총)",
     "[gemini-pro] 물리적으로 불가능한 화면 구성 (얼굴과 겹쳐 허공에 뜬 총열)",
     "[gpt-high] 얼굴 앞의 발사 중인 총열은 손이나 받침, 아래쪽 총과의 연결 없이 공중에 떠 있어 아무것도 지지하지 않는다.",
     "[gpt-high] 염재보가 잡고 있는 총과 별개로 발사 총열을 추가해 중심 소품을 중복 생성했다.",
     "[gpt-high] 염재보 외에는 등장시키지 말라는 명시적 제한에도 이전 장면의 다른 사람들과 시신을 다수 포함했다."
    ],
    "physics": "화면 중앙의 거대한 화승총 총열은 손이나 어떤 지지대도 없이 공중에 완전히 떠 있음."
   },
   {
    "label": "B",
    "direction": "화승총이 프롬프트의 지시대로 화면 오른쪽을 향해 조준되어 붉은 섬광을 뿜으며 발사됨.",
    "built_space": "배경에 초가집, 쓰러진 시체들, 쏟아진 밥과 깨진 그릇들이 레퍼런스와 동일하게 배치되어 있음.",
    "entities": "레퍼런스의 의상과 핏자국이 일치하는 염재보의 손과 팔이 화승총을 쥐고 있으며, 총구에서 붉은 섬광이 발생함.",
    "hard_violations": [
     "[gpt-high] 염재보 외에는 등장시키지 말라는 명시적 제한에도 배경에 다른 사람들의 시신을 다수 포함했다."
    ],
    "physics": "인물의 두 손이 화승총을 기능적인 위치에서 물리적으로 단단히 쥐고 지지하고 있음."
   }
  ],
  "cross_model_order": {
   "policy": "cross_model_order_v2_slot0_fwd_slot1_rev_parallel2",
   "slots": [
    {
     "model": "gemini-pro",
     "order": "forward",
     "display_to_canonical": {
      "A": "A",
      "B": "B"
     },
     "raw": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 0,
        "verdict_ko": "공중에 떠 있는 거대한 두 번째 총열과 잘못된 발사 방향으로 인해 프롬프트의 요구사항을 심각하게 위반했습니다."
       },
       {
        "label": "B",
        "score": 10,
        "verdict_ko": "지정된 클로즈업 구도, 카메라 각도, 오른쪽을 향한 붉은 섬광, 그리고 레퍼런스와 일치하는 손의 묘사를 완벽하게 구현했습니다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "거대한 공중 부양 화승총 총열이 화면 왼쪽으로 붉은 광선을 발사함. 염재보는 정면을 향해 엎드려 있음.",
        "built_space": "부서진 초가집, 시체, 흩어진 그릇과 밥이 있는 폐허가 된 산채의 배경이 보임.",
        "entities": "바닥에 엎드려 화승총을 쥔 염재보가 있으나, 그의 얼굴 앞에 두 번째 거대한 화승총 총열이 떠 있음.",
        "hard_violations": [
         "공중에 떠 있는 지지되지 않은 물체 (거대한 총열)",
         "중복된 사물 (두 개의 화승총)",
         "물리적으로 불가능한 화면 구성 (얼굴과 겹쳐 허공에 뜬 총열)"
        ],
        "physics": "화면 중앙의 거대한 화승총 총열은 손이나 어떤 지지대도 없이 공중에 완전히 떠 있음."
       },
       {
        "label": "B",
        "direction": "화승총이 프롬프트의 지시대로 화면 오른쪽을 향해 조준되어 붉은 섬광을 뿜으며 발사됨.",
        "built_space": "배경에 초가집, 쓰러진 시체들, 쏟아진 밥과 깨진 그릇들이 레퍼런스와 동일하게 배치되어 있음.",
        "entities": "레퍼런스의 의상과 핏자국이 일치하는 염재보의 손과 팔이 화승총을 쥐고 있으며, 총구에서 붉은 섬광이 발생함.",
        "hard_violations": [],
        "physics": "인물의 두 손이 화승총을 기능적인 위치에서 물리적으로 단단히 쥐고 지지하고 있음."
       }
      ],
      "all_candidates_fail": false
     },
     "normalized": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 0,
        "verdict_ko": "공중에 떠 있는 거대한 두 번째 총열과 잘못된 발사 방향으로 인해 프롬프트의 요구사항을 심각하게 위반했습니다."
       },
       {
        "label": "B",
        "score": 10,
        "verdict_ko": "지정된 클로즈업 구도, 카메라 각도, 오른쪽을 향한 붉은 섬광, 그리고 레퍼런스와 일치하는 손의 묘사를 완벽하게 구현했습니다."
       }
      ],
      "all_candidates_fail": false,
      "readings": [
       {
        "label": "A",
        "direction": "거대한 공중 부양 화승총 총열이 화면 왼쪽으로 붉은 광선을 발사함. 염재보는 정면을 향해 엎드려 있음.",
        "built_space": "부서진 초가집, 시체, 흩어진 그릇과 밥이 있는 폐허가 된 산채의 배경이 보임.",
        "entities": "바닥에 엎드려 화승총을 쥔 염재보가 있으나, 그의 얼굴 앞에 두 번째 거대한 화승총 총열이 떠 있음.",
        "hard_violations": [
         "공중에 떠 있는 지지되지 않은 물체 (거대한 총열)",
         "중복된 사물 (두 개의 화승총)",
         "물리적으로 불가능한 화면 구성 (얼굴과 겹쳐 허공에 뜬 총열)"
        ],
        "physics": "화면 중앙의 거대한 화승총 총열은 손이나 어떤 지지대도 없이 공중에 완전히 떠 있음."
       },
       {
        "label": "B",
        "direction": "화승총이 프롬프트의 지시대로 화면 오른쪽을 향해 조준되어 붉은 섬광을 뿜으며 발사됨.",
        "built_space": "배경에 초가집, 쓰러진 시체들, 쏟아진 밥과 깨진 그릇들이 레퍼런스와 동일하게 배치되어 있음.",
        "entities": "레퍼런스의 의상과 핏자국이 일치하는 염재보의 손과 팔이 화승총을 쥐고 있으며, 총구에서 붉은 섬광이 발생함.",
        "hard_violations": [],
        "physics": "인물의 두 손이 화승총을 기능적인 위치에서 물리적으로 단단히 쥐고 지지하고 있음."
       }
      ]
     },
     "ok": true
    },
    {
     "model": "gpt-high",
     "order": "reverse",
     "display_to_canonical": {
      "A": "B",
      "B": "A"
     },
     "raw": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 3,
        "verdict_ko": "손으로 지지한 총의 우향 발사와 총·손 중심 근접 구도는 맞지만, 금지된 다른 인물들의 시신을 다수 포함해 부적격이다."
       },
       {
        "label": "B",
        "score": 0,
        "verdict_ko": "지지 없이 떠 있는 별도 총구가 왼쪽으로 발사하며, 이전 장면의 넓은 인물 구도와 다른 시신들까지 복제해 핵심 지시를 위반한다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "총열은 화면 왼쪽 아래에서 오른쪽 위로 향하고 붉은 섬광도 총구에서 오른쪽으로 뻗는다. 화면 오른쪽으로 발사하라는 지시와 맞으며, 구체적인 표적은 화면에 없다. 얼굴이 잘려 사수의 시선은 확인할 수 없다.",
        "built_space": "왼쪽·중앙·오른쪽 가장자리에 움막 세 채가 보이고 뒤쪽에는 목책이 있다. 흙바닥에는 넘어진 검은 솥 하나, 여러 흰 그릇과 파편, 흩어진 쌀이 있다. 재료와 낮의 분위기는 이전 장면과 유사하다. 다만 솥과 시신들의 상대 배치가 달라 보이며, 총 아래 지면은 보이지만 지면에 밀착한 엎드린 사격 자세를 입증하는 낮은 시점은 약하다.",
        "entities": "염재보의 것으로 읽히는 성인 남성의 피 묻은 두 손과 회흑색 누비 소매가 보이며 손 참고사진의 피부색·체격·소매 질감에 대체로 맞는다. 얼굴과 머리는 범위 밖이므로 신원 세부는 확인할 수 없다. 목제 개머리판과 금속 총열·격발부가 있는 구식 총 한 정, 붉은 화염과 흰 연기가 보인다. 배경에는 염재보 외 여러 사람의 시신이 명확히 들어와 인물 제한을 어긴다. 읽을 수 있는 글자는 없다.",
        "hard_violations": [
         "염재보 외에는 등장시키지 말라는 명시적 제한에도 배경에 다른 사람들의 시신을 다수 포함했다."
        ],
        "physics": "한 손은 방아쇠 부근을 잡고 다른 손은 총의 앞부분을 아래에서 받쳐 총을 지지한다. 소매와 팔은 화면 밖으로 이어지며, 보이는 총에 지지 없는 부분은 없다. 화염과 연기는 총구에서 발생한다. 배경 시신과 그릇은 지면에 놓여 있다. 사수의 몸통이 잘려 엎드린 자세 자체는 확인할 수 없지만, 이를 부유한 몸으로 판단할 근거는 없다."
       },
       {
        "label": "B",
        "direction": "발사 중인 전경 총구는 화면 왼쪽을 향하고 붉은 빛도 왼쪽으로 나가 우향 발사 지시와 정반대다. 염재보가 실제로 손에 든 아래쪽 총은 오른쪽으로 뻗지만 발사 섬광과 연결되지 않는다. 염재보는 전경 총열 뒤에서 앞쪽을 보고 있으나 해당 총을 조준하는 시선 관계는 성립하지 않는다.",
        "built_space": "이전 장면처럼 움막 세 채와 뒤쪽 목책, 왼쪽의 넘어진 솥 하나, 흰 그릇들과 쌀이 흩어진 흙바닥이 보인다. 염재보는 중앙에서 팔과 몸통을 지면에 대고 있고 주변 시신 배치도 이전 장면을 거의 그대로 따른다. 장소의 외형은 가깝지만 총과 손의 근접 촬영 대신 이전 장면의 얼굴·상체 중심 구도를 유지했다.",
        "entities": "중앙에는 참고사진과 유사한 검은 머리의 한국인 성인 남성 염재보가 있으며, 얼굴 상처와 피 묻은 회흑색 누비옷도 유사하다. 두 손이 잡은 구식 총 한 정 외에 얼굴 앞에 별도의 거대한 총열 부분이 추가되어 있다. 붉은 발사광은 그 추가 총열에서 나온다. 주변에는 등장 금지된 다른 사람들의 시신이 다수 보인다. 읽을 수 있는 글자는 없다.",
        "hard_violations": [
         "얼굴 앞의 발사 중인 총열은 손이나 받침, 아래쪽 총과의 연결 없이 공중에 떠 있어 아무것도 지지하지 않는다.",
         "염재보가 잡고 있는 총과 별개로 발사 총열을 추가해 중심 소품을 중복 생성했다.",
         "염재보 외에는 등장시키지 말라는 명시적 제한에도 이전 장면의 다른 사람들과 시신을 다수 포함했다."
        ],
        "physics": "염재보의 팔과 몸통은 땅에 닿고, 아래쪽 총은 두 손과 지면 가까운 개머리판으로 지지된다. 그러나 발사 중인 위쪽 총열에는 손이 없고 개머리판이나 다른 지지물도 이어지지 않아 아무것도 지지하지 않는다. 아래쪽 총을 잡은 동작으로는 위쪽 총열의 발사를 물리적으로 설명할 수 없다. 주변 시신과 식기는 지면에 놓여 있다."
       }
      ],
      "all_candidates_fail": true
     },
     "normalized": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 3,
        "verdict_ko": "손으로 지지한 총의 우향 발사와 총·손 중심 근접 구도는 맞지만, 금지된 다른 인물들의 시신을 다수 포함해 부적격이다."
       },
       {
        "label": "A",
        "score": 0,
        "verdict_ko": "지지 없이 떠 있는 별도 총구가 왼쪽으로 발사하며, 이전 장면의 넓은 인물 구도와 다른 시신들까지 복제해 핵심 지시를 위반한다."
       }
      ],
      "all_candidates_fail": true,
      "readings": [
       {
        "label": "B",
        "direction": "총열은 화면 왼쪽 아래에서 오른쪽 위로 향하고 붉은 섬광도 총구에서 오른쪽으로 뻗는다. 화면 오른쪽으로 발사하라는 지시와 맞으며, 구체적인 표적은 화면에 없다. 얼굴이 잘려 사수의 시선은 확인할 수 없다.",
        "built_space": "왼쪽·중앙·오른쪽 가장자리에 움막 세 채가 보이고 뒤쪽에는 목책이 있다. 흙바닥에는 넘어진 검은 솥 하나, 여러 흰 그릇과 파편, 흩어진 쌀이 있다. 재료와 낮의 분위기는 이전 장면과 유사하다. 다만 솥과 시신들의 상대 배치가 달라 보이며, 총 아래 지면은 보이지만 지면에 밀착한 엎드린 사격 자세를 입증하는 낮은 시점은 약하다.",
        "entities": "염재보의 것으로 읽히는 성인 남성의 피 묻은 두 손과 회흑색 누비 소매가 보이며 손 참고사진의 피부색·체격·소매 질감에 대체로 맞는다. 얼굴과 머리는 범위 밖이므로 신원 세부는 확인할 수 없다. 목제 개머리판과 금속 총열·격발부가 있는 구식 총 한 정, 붉은 화염과 흰 연기가 보인다. 배경에는 염재보 외 여러 사람의 시신이 명확히 들어와 인물 제한을 어긴다. 읽을 수 있는 글자는 없다.",
        "hard_violations": [
         "염재보 외에는 등장시키지 말라는 명시적 제한에도 배경에 다른 사람들의 시신을 다수 포함했다."
        ],
        "physics": "한 손은 방아쇠 부근을 잡고 다른 손은 총의 앞부분을 아래에서 받쳐 총을 지지한다. 소매와 팔은 화면 밖으로 이어지며, 보이는 총에 지지 없는 부분은 없다. 화염과 연기는 총구에서 발생한다. 배경 시신과 그릇은 지면에 놓여 있다. 사수의 몸통이 잘려 엎드린 자세 자체는 확인할 수 없지만, 이를 부유한 몸으로 판단할 근거는 없다."
       },
       {
        "label": "A",
        "direction": "발사 중인 전경 총구는 화면 왼쪽을 향하고 붉은 빛도 왼쪽으로 나가 우향 발사 지시와 정반대다. 염재보가 실제로 손에 든 아래쪽 총은 오른쪽으로 뻗지만 발사 섬광과 연결되지 않는다. 염재보는 전경 총열 뒤에서 앞쪽을 보고 있으나 해당 총을 조준하는 시선 관계는 성립하지 않는다.",
        "built_space": "이전 장면처럼 움막 세 채와 뒤쪽 목책, 왼쪽의 넘어진 솥 하나, 흰 그릇들과 쌀이 흩어진 흙바닥이 보인다. 염재보는 중앙에서 팔과 몸통을 지면에 대고 있고 주변 시신 배치도 이전 장면을 거의 그대로 따른다. 장소의 외형은 가깝지만 총과 손의 근접 촬영 대신 이전 장면의 얼굴·상체 중심 구도를 유지했다.",
        "entities": "중앙에는 참고사진과 유사한 검은 머리의 한국인 성인 남성 염재보가 있으며, 얼굴 상처와 피 묻은 회흑색 누비옷도 유사하다. 두 손이 잡은 구식 총 한 정 외에 얼굴 앞에 별도의 거대한 총열 부분이 추가되어 있다. 붉은 발사광은 그 추가 총열에서 나온다. 주변에는 등장 금지된 다른 사람들의 시신이 다수 보인다. 읽을 수 있는 글자는 없다.",
        "hard_violations": [
         "얼굴 앞의 발사 중인 총열은 손이나 받침, 아래쪽 총과의 연결 없이 공중에 떠 있어 아무것도 지지하지 않는다.",
         "염재보가 잡고 있는 총과 별개로 발사 총열을 추가해 중심 소품을 중복 생성했다.",
         "염재보 외에는 등장시키지 말라는 명시적 제한에도 이전 장면의 다른 사람들과 시신을 다수 포함했다."
        ],
        "physics": "염재보의 팔과 몸통은 땅에 닿고, 아래쪽 총은 두 손과 지면 가까운 개머리판으로 지지된다. 그러나 발사 중인 위쪽 총열에는 손이 없고 개머리판이나 다른 지지물도 이어지지 않아 아무것도 지지하지 않는다. 아래쪽 총을 잡은 동작으로는 위쪽 총열의 발사를 물리적으로 설명할 수 없다. 주변 시신과 식기는 지면에 놓여 있다."
       }
      ]
     },
     "ok": true
    }
   ],
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "slot_winner_match": true,
   "slot_winner": {
    "gemini-pro": "B",
    "gpt-high": "B"
   },
   "route": "cross_slot_agree"
  },
  "dual": {
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "normalized": {
    "A": 0.0,
    "B": 2.0
   },
   "adjusted": {
    "A": -0.25,
    "B": 1.75
   },
   "violations": {
    "A": [
     "[gemini-pro] 공중에 떠 있는 지지되지 않은 물체 (거대한 총열)",
     "[gemini-pro] 중복된 사물 (두 개의 화승총)",
     "[gemini-pro] 물리적으로 불가능한 화면 구성 (얼굴과 겹쳐 허공에 뜬 총열)",
     "[gpt-high] 얼굴 앞의 발사 중인 총열은 손이나 받침, 아래쪽 총과의 연결 없이 공중에 떠 있어 아무것도 지지하지 않는다.",
     "[gpt-high] 염재보가 잡고 있는 총과 별개로 발사 총열을 추가해 중심 소품을 중복 생성했다.",
     "[gpt-high] 염재보 외에는 등장시키지 말라는 명시적 제한에도 이전 장면의 다른 사람들과 시신을 다수 포함했다."
    ],
    "B": [
     "[gpt-high] 염재보 외에는 등장시키지 말라는 명시적 제한에도 배경에 다른 사람들의 시신을 다수 포함했다."
    ]
   },
   "per_model_winner": {
    "gemini-pro": "B",
    "gpt-high": "B"
   },
   "agreed": true
  },
  "totals": {
   "A": -250,
   "B": 1750
  },
  "selected": "B",
  "ranking": [
   "B",
   "A"
  ],
  "verdicts": [
   {
    "label": "A",
    "score": -250,
    "verdict_ko": "공중에 떠 있는 거대한 두 번째 총열과 잘못된 발사 방향으로 인해 프롬프트의 요구사항을 심각하게 위반했습니다.  ★위반: [gemini-pro] 공중에 떠 있는 지지되지 않은 물체 (거대한 총열) / [gemini-pro] 중복된 사물 (두 개의 화승총) / [gemini-pro] 물리적으로 불가능한 화면 구성 (얼굴과 겹쳐 허공에 뜬 총열) / [gpt-high] 얼굴 앞의 발사 중인 총열은 손이나 받침, 아래쪽 총과의 연결 없이 공중에 떠 있어 아무것도 지지하지 않는다. / [gpt-high] 염재보가 잡고 있는 총과 별개로 발사 총열을 추가해 중심 소품을 중복 생성했다. / [gpt-high] 염재보 외에는 등장시키지 말라는 명시적 제한에도 이전 장면의 다른 사람들과 시신을 다수 포함했다."
   },
   {
    "label": "B",
    "score": 1750,
    "verdict_ko": "지정된 클로즈업 구도, 카메라 각도, 오른쪽을 향한 붉은 섬광, 그리고 레퍼런스와 일치하는 손의 묘사를 완벽하게 구현했습니다.  ★위반: [gpt-high] 염재보 외에는 등장시키지 말라는 명시적 제한에도 배경에 다른 사람들의 시신을 다수 포함했다."
   }
  ],
  "refs": [
   {
    "label": "PREVIOUS SHOT STILL — a visually related earlier shot of this same place: the location's look, materials, fixed features and lighting mood are LOCKED to this photo; never copy its camera framing. The clothing, hair and overall look of 염재보 — who appear both in that photo and in this shot — are LOCKED to that photo. Anyone else visible in that photo is NOT in this shot: never carry their face, body or clothing onto anyone here. If this photo shows a character who cannot move (dead or unconscious), that character's exact pose, position and orientation are ALSO LOCKED — treat their body as a fixed prop of the set that only the camera moves around.",
    "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/db00517b-165b-4005-a287-f28936b59b95/images/272bd7af-9294-4490-98c6-88a40e9deacc/scene/recipe/S61sh15_sel.png",
    "asset_id": "5447aebd-dbcc-4746-b2f7-5c0ef352ff37",
    "role": "prev_still"
   },
   {
    "label": "HAND OWNER REFERENCE — 염재보: the person whose hand is on the object in this shot. Use this photograph ONLY to get that hand right — its size, build, skin tone and texture, apparent age, grooming, nails, sleeve and anything worn on the wrist or fingers. Do NOT bring their face, body or clothing into the frame; the framing stays on the object and reaches no further than the wrist and forearm.",
    "path": "<bytes:1331451>",
    "asset_id": "bad3b41f-75ca-4d66-8fce-161d96d99096",
    "role": "character_ref"
   }
  ],
  "critique_skipped": true,
  "shot_run_produced": false,
  "shot_run_uid": "06aa1445-4a54-7291-8736-0b3a227055f7",
  "ref_mode": "prev만 (배경 전용·공유 계획)",
  "share_plan": {
   "ref_plan": "prev",
   "prev_anchor_tag": "S61sh15"
  },
  "lane_policy": "share_plan_prev_bgonly"
 },
 "S61sh32::cine": {
  "applied": true,
  "attempted_at": "2026-09-09T08:07:50.612463+00:00",
  "fingerprint": "fe7c9dad8b24d99ecf7556db60b7528f7a007300bee3ea500342a7caa3019f18",
  "fingerprint_version": 2,
  "provider": "mai",
  "endpoint": "openrouter/chat-completions",
  "model": "microsoft/mai-image-2.6",
  "pack": "24.202608252115",
  "source_file": "S61sh32_sel.png",
  "source_sha256": "1b2455b2f0aa3965bb09750e33a021390755e0bbb5a19cba64a78b6bf22c7ee1",
  "file": "S61sh32_cine.png",
  "staged_sha256": "0af424cf71003804059bfd110b8d44697522c5f4ade71496db732e3c7dbc4258",
  "latency_ms": 23170
 },
 "S62sh12::signage": {
  "fp": "7023d369b64647f2",
  "inscriptions": [],
  "cues": [],
  "dropped": []
 },
 "S62sh12": {
  "input_fingerprint": "95ce0c7079ad9ac0",
  "prompt": "Create ONE FINAL photorealistic live-action film still of the moment below — early 20th-century Korea, Jeolla region; all people are Korean unless explicitly stated otherwise. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 대열이 무너진 채 각기 다른 방향을 향해 뒤엉켜 있는 일본인 군인들의 전경.\n\nLOCATION (lock): The wooded slope of the mountain camp's descent trail, where the returning column is exposed to fire from both sides. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: wide shot\n- KEY BACKGROUND ELEMENTS: Forest path (Occupied by a disrupted formation and fallen soldiers) — Seen diagonally uphill from above its downhill edge; used as Makes broken spacing and conflicting body directions readable within one continuous route; Trees on both sides of the path (Flank the soldiers' position) — Border the left and right sides of the uphill corridor; used as Provide opposing threat directions for the soldiers' attention; Soldiers' rifles (Turned toward different sides of the forest) — Barrels form divergent lateral diagonals rather than a common firing line; used as Express the formation's loss of coordinated direction at ordinary human-relative scale.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Continuous daylight and restrained saturation keep the confused formation legible with sober, unembellished contrast.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The people visible in that photo are NOT in this shot: never carry their faces, bodies or clothing onto anyone here. Each person in THIS shot is defined solely by the PEOPLE list, their own CHARACTER REFERENCE images and their wardrobe notes. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The descending forest path is now the route of an uphill return, with scattered gunfire breaking from the trees.\n\nNo people appear unless the shot text itself says so.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.",
  "roll_prompts": {
   "A": "Create ONE FINAL photorealistic live-action film still of the moment below — early 20th-century Korea, Jeolla region; all people are Korean unless explicitly stated otherwise. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 대열이 무너진 채 각기 다른 방향을 향해 뒤엉켜 있는 일본인 군인들의 전경.\n\nLOCATION (lock): The wooded slope of the mountain camp's descent trail, where the returning column is exposed to fire from both sides. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: wide shot\n- KEY BACKGROUND ELEMENTS: Forest path (Occupied by a disrupted formation and fallen soldiers) — Seen diagonally uphill from above its downhill edge; used as Makes broken spacing and conflicting body directions readable within one continuous route; Trees on both sides of the path (Flank the soldiers' position) — Border the left and right sides of the uphill corridor; used as Provide opposing threat directions for the soldiers' attention; Soldiers' rifles (Turned toward different sides of the forest) — Barrels form divergent lateral diagonals rather than a common firing line; used as Express the formation's loss of coordinated direction at ordinary human-relative scale.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Continuous daylight and restrained saturation keep the confused formation legible with sober, unembellished contrast.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The people visible in that photo are NOT in this shot: never carry their faces, bodies or clothing onto anyone here. Each person in THIS shot is defined solely by the PEOPLE list, their own CHARACTER REFERENCE images and their wardrobe notes. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The descending forest path is now the route of an uphill return, with scattered gunfire breaking from the trees.\n\nNo people appear unless the shot text itself says so.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.",
   "B": "Create ONE FINAL photorealistic live-action film still of the moment below — early 20th-century Korea, Jeolla region; all people are Korean unless explicitly stated otherwise. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 대열이 무너진 채 각기 다른 방향을 향해 뒤엉켜 있는 일본인 군인들의 전경.\n\nLOCATION (lock): The wooded slope of the mountain camp's descent trail, where the returning column is exposed to fire from both sides. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: wide shot\n- KEY BACKGROUND ELEMENTS: Forest path (Occupied by a disrupted formation and fallen soldiers) — Seen diagonally uphill from above its downhill edge; used as Makes broken spacing and conflicting body directions readable within one continuous route; Trees on both sides of the path (Flank the soldiers' position) — Border the left and right sides of the uphill corridor; used as Provide opposing threat directions for the soldiers' attention; Soldiers' rifles (Turned toward different sides of the forest) — Barrels form divergent lateral diagonals rather than a common firing line; used as Express the formation's loss of coordinated direction at ordinary human-relative scale.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Continuous daylight and restrained saturation keep the confused formation legible with sober, unembellished contrast.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The people visible in that photo are NOT in this shot: never carry their faces, bodies or clothing onto anyone here. Each person in THIS shot is defined solely by the PEOPLE list, their own CHARACTER REFERENCE images and their wardrobe notes. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The descending forest path is now the route of an uphill return, with scattered gunfire breaking from the trees.\n\nNo people appear unless the shot text itself says so.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.\n\nCOMPOSITION VARIATION (this candidate only): keep the place, the people and the moment exactly as specified, and keep from the CAMERA & FRAME contract its subject scale, screen placement and key background elements exactly. For this candidate only, its camera angle and height are deliberately left open: choose a camera position distinctly different from the obvious one, as a film director picking a second setup on the same blocking. Stay true to any explicit framing the SHOT TEXT itself dictates."
  },
  "shot_run_spend_attempt_count": 1,
  "readings": [
   {
    "label": "A",
    "direction": "병사들의 총구가 좌우 숲 등 엇갈린 여러 방향을 향하고 있으며, 완전히 무너진 대열의 방향성이 정확하게 나타남.",
    "built_space": "숲이 우거진 내리막 산길 경사면으로 레퍼런스의 지형 조건을 잘 일치시킴.",
    "entities": "지시된 일본군 군복을 입은 다수의 병사들과 쓰러진 자들, 그리고 구식 소총들이 정확히 묘사됨.",
    "hard_violations": [
     "[gpt-high] 아래쪽 가장자리에서 오르막을 바라보라는 지형·시점 지시와 반대로, 위쪽에서 내리막과 계곡을 내려다본다."
    ],
    "physics": "모든 인물이 경사면에 물리적으로 안착해 있고, 무기들도 손에 들려 있거나 지면에 놓여 있어 지지 상태가 자연스러움."
   },
   {
    "label": "B",
    "direction": "한 명의 병사만이 좌측 숲을 향해 사격 자세를 취할 뿐, 서로 다른 방향으로 뒤엉킨 대열과 엇갈린 총구의 지시를 충분히 구현하지 못함.",
    "built_space": "레퍼런스와 동일한 산길 경사면 지형 및 숲의 배경을 적절히 구성함.",
    "entities": "일본군 병사들과 소총이 등장하나, 요구된 혼란스러운 대열 묘사를 충족하기에는 역부족임.",
    "hard_violations": [
     "[gemini-pro] 좌측 상단 허공에 아무도 쥐고 있지 않은 소총이 수직으로 떠 있음",
     "[gpt-high] 아래쪽 가장자리에서 오르막을 바라보라는 지형·시점 지시와 반대로, 위쪽에서 내리막과 계곡을 내려다본다.",
     "[gpt-high] 왼쪽 조준 병사의 등 뒤에서 오른쪽 위로 뻗은 별도 소총은 손이나 연결된 멜빵 등 지지점 없이 떠 있다."
    ],
    "physics": "좌측 상단 배경의 소총이 공중에 떠 있으며, 이를 지지하거나 잡고 있는 대상이 전혀 없음."
   }
  ],
  "cross_model_order": {
   "policy": "cross_model_order_v2_slot0_fwd_slot1_rev_parallel2",
   "slots": [
    {
     "model": "gemini-pro",
     "order": "forward",
     "display_to_canonical": {
      "A": "A",
      "B": "B"
     },
     "raw": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 8,
        "verdict_ko": "대열이 붕괴되어 여러 방향을 경계하는 혼란스러운 전경을 훌륭하게 연출했으며, 물리적 오류 없이 프롬프트의 지시를 충실히 따름."
       },
       {
        "label": "B",
        "score": 3,
        "verdict_ko": "사방을 향해 엇갈린 총구와 뒤엉킨 대열의 묘사가 부족하며, 무엇보다 허공에 떠 있는 소총이 렌더링되는 치명적인 물리적 오류가 발생함."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "병사들의 총구가 좌우 숲 등 엇갈린 여러 방향을 향하고 있으며, 완전히 무너진 대열의 방향성이 정확하게 나타남.",
        "built_space": "숲이 우거진 내리막 산길 경사면으로 레퍼런스의 지형 조건을 잘 일치시킴.",
        "entities": "지시된 일본군 군복을 입은 다수의 병사들과 쓰러진 자들, 그리고 구식 소총들이 정확히 묘사됨.",
        "hard_violations": [],
        "physics": "모든 인물이 경사면에 물리적으로 안착해 있고, 무기들도 손에 들려 있거나 지면에 놓여 있어 지지 상태가 자연스러움."
       },
       {
        "label": "B",
        "direction": "한 명의 병사만이 좌측 숲을 향해 사격 자세를 취할 뿐, 서로 다른 방향으로 뒤엉킨 대열과 엇갈린 총구의 지시를 충분히 구현하지 못함.",
        "built_space": "레퍼런스와 동일한 산길 경사면 지형 및 숲의 배경을 적절히 구성함.",
        "entities": "일본군 병사들과 소총이 등장하나, 요구된 혼란스러운 대열 묘사를 충족하기에는 역부족임.",
        "hard_violations": [
         "좌측 상단 허공에 아무도 쥐고 있지 않은 소총이 수직으로 떠 있음"
        ],
        "physics": "좌측 상단 배경의 소총이 공중에 떠 있으며, 이를 지지하거나 잡고 있는 대상이 전혀 없음."
       }
      ],
      "all_candidates_fail": false
     },
     "normalized": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 8,
        "verdict_ko": "대열이 붕괴되어 여러 방향을 경계하는 혼란스러운 전경을 훌륭하게 연출했으며, 물리적 오류 없이 프롬프트의 지시를 충실히 따름."
       },
       {
        "label": "B",
        "score": 3,
        "verdict_ko": "사방을 향해 엇갈린 총구와 뒤엉킨 대열의 묘사가 부족하며, 무엇보다 허공에 떠 있는 소총이 렌더링되는 치명적인 물리적 오류가 발생함."
       }
      ],
      "all_candidates_fail": false,
      "readings": [
       {
        "label": "A",
        "direction": "병사들의 총구가 좌우 숲 등 엇갈린 여러 방향을 향하고 있으며, 완전히 무너진 대열의 방향성이 정확하게 나타남.",
        "built_space": "숲이 우거진 내리막 산길 경사면으로 레퍼런스의 지형 조건을 잘 일치시킴.",
        "entities": "지시된 일본군 군복을 입은 다수의 병사들과 쓰러진 자들, 그리고 구식 소총들이 정확히 묘사됨.",
        "hard_violations": [],
        "physics": "모든 인물이 경사면에 물리적으로 안착해 있고, 무기들도 손에 들려 있거나 지면에 놓여 있어 지지 상태가 자연스러움."
       },
       {
        "label": "B",
        "direction": "한 명의 병사만이 좌측 숲을 향해 사격 자세를 취할 뿐, 서로 다른 방향으로 뒤엉킨 대열과 엇갈린 총구의 지시를 충분히 구현하지 못함.",
        "built_space": "레퍼런스와 동일한 산길 경사면 지형 및 숲의 배경을 적절히 구성함.",
        "entities": "일본군 병사들과 소총이 등장하나, 요구된 혼란스러운 대열 묘사를 충족하기에는 역부족임.",
        "hard_violations": [
         "좌측 상단 허공에 아무도 쥐고 있지 않은 소총이 수직으로 떠 있음"
        ],
        "physics": "좌측 상단 배경의 소총이 공중에 떠 있으며, 이를 지지하거나 잡고 있는 대상이 전혀 없음."
       }
      ]
     },
     "ok": true
    },
    {
     "model": "gpt-high",
     "order": "reverse",
     "display_to_canonical": {
      "A": "B",
      "B": "A"
     },
     "raw": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 2,
        "verdict_ko": "흩어진 병사들은 보이지만 오르막을 바라보라는 시점을 반대로 재현했고, 지지 없는 소총과 오른쪽 숲을 겨누지 않는 총구들이 핵심 연출을 훼손합니다."
       },
       {
        "label": "B",
        "score": 4,
        "verdict_ko": "양쪽 숲을 향한 상반된 조준과 쓰러진 병사들의 뒤엉킴은 더 충실하지만, 참고 사진의 내리막 시점을 반복하여 필수 오르막 구도를 위반합니다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "왼쪽 나무 옆 병사는 시선과 총구를 왼쪽 비탈의 수풀로 향한다. 오른쪽 전경 병사는 오른쪽 숲을 보며 이동하지만 소총 총구는 발치 쪽으로 내려가 있다. 중앙에서 수평으로 총을 든 병사의 총은 오른쪽 끝이 개머리판이어서 오른쪽 숲을 조준하는 모습이 아니다. 뒤쪽 병사들은 서로 다른 방향으로 몸을 돌리고, 주저앉은 병사는 아래를 본다. 몸 방향의 혼란은 있지만 양쪽 숲으로 갈라지는 조준은 약하다.",
        "built_space": "인공 구조물이나 고정 설비는 없다. 하나의 좁은 흙길, 왼쪽의 뿌리 드러난 높은 사면, 오른쪽의 낮은 숲과 먼 계곡이 보인다. 나무 배치와 지면 재질은 참고 장소에 가깝다. 그러나 길이 전경에서 화면 중앙의 낮은 곳으로 내려가므로 카메라는 위쪽에서 내리막을 본다. 요구된 아래쪽 가장자리에서 비스듬히 오르막을 보는 공간 방향과 반대다.",
        "entities": "부분적으로 가려진 인물을 포함해 약 일곱 명의 병사가 보이며, 중앙 한 명은 주저앉아 있다. 보이는 얼굴들은 성인 동아시아계 남성으로 읽히고 군복과 군모는 일본군 설정에 부합하지만 국적이나 개별 신원은 외모만으로 확인할 수 없다. 목재 개머리판과 금속 총열을 가진 장총들이 있으나 참고 소품의 부싯돌식 격발 장치와 황동 띠를 정확히 재현하지는 않는다. 읽을 수 있는 글자나 화면 분할은 없다.",
        "hard_violations": [
         "아래쪽 가장자리에서 오르막을 바라보라는 지형·시점 지시와 반대로, 위쪽에서 내리막과 계곡을 내려다본다.",
         "왼쪽 조준 병사의 등 뒤에서 오른쪽 위로 뻗은 별도 소총은 손이나 연결된 멜빵 등 지지점 없이 떠 있다."
        ],
        "physics": "왼쪽 조준 병사는 한 발을 높은 뿌리 쪽에, 다른 발을 낮은 비탈에 딛고 총을 양손으로 받친다. 중앙 병사는 엉덩이와 다리가 흙길에 닿아 있고, 오른쪽 이동 병사는 앞발로 지면을 지지하면서 뒷발을 든다. 바닥의 소총은 흙과 돌에 놓여 있다. 다만 왼쪽 병사 뒤에는 등에 걸린 총과 별도로 오른쪽 위를 향한 소총이 있으며, 이를 잡은 손이나 매단 끈이 보이지 않아 아무것도 지지하지 않는다."
       },
       {
        "label": "B",
        "direction": "왼쪽 전경의 선 병사와 그 뒤 병사는 총구와 시선을 왼쪽 사면의 숲으로 향한다. 오른쪽 장교는 견착한 소총과 시선을 오른쪽 숲으로 향해 두 위협 방향이 명확히 갈린다. 중앙 병사는 왼쪽으로 팔을 뻗고, 배낭을 멘 병사는 오른쪽을 보면서 총열을 위로 세운다. 왼쪽 아래 병사는 사면을 짚고 숙이며 오른쪽 아래 병사는 길 가장자리로 몸을 낮춘다. 쓰러진 몸들도 길을 따라 서로 다른 각도로 놓여 있어 공통 사격선이 무너진 상태가 읽힌다.",
        "built_space": "인공 구조물이나 고정 설비 없이 하나의 숲길에 병사들이 밀집해 있다. 왼쪽의 높은 흙사면과 노출된 뿌리, 양옆의 나무, 오른쪽 아래로 열린 계곡은 참고 장소의 특징을 유지한다. 그러나 배경으로 갈수록 길이 낮아지고 계곡이 드러나므로 여기서도 위쪽에서 내리막을 보는 구도다. 요구된 오르막 방향의 관찰 시점은 아니다.",
        "entities": "서 있거나 웅크린 병사와 누운 병사를 합쳐 약 아홉 명이 보인다. 보이는 인물들은 성인 동아시아계 남성으로 읽히며 군모, 제복, 가죽 탄띠와 배낭은 일본군 집단 설정에 맞는다. 별도 민간인이나 적군은 보이지 않는다. 장총의 목재와 금속 재질은 실물처럼 보이지만 참고 소품의 특징적인 부싯돌식 장치와 황동 부품까지 일치하지는 않는다. 숲 사이의 옅은 연기는 산발적인 총격 상황과 부합하며 읽을 수 있는 글자나 그래픽은 없다.",
        "hard_violations": [
         "아래쪽 가장자리에서 오르막을 바라보라는 지형·시점 지시와 반대로, 위쪽에서 내리막과 계곡을 내려다본다."
        ],
        "physics": "조준하는 병사들은 비탈과 길에 발을 벌려 딛고 소총을 손과 어깨로 지지한다. 왼쪽 아래 병사는 손과 무릎으로 사면에 기대며, 오른쪽 아래 병사는 무릎과 하체를 지면에 대고 몸을 낮춘다. 중앙과 전경의 쓰러진 병사들은 등·옆구리·다리가 흙길에 닿아 있다. 떨어진 소총은 사면이나 길 위에 놓이고, 세운 소총은 병사의 손에 잡혀 있다. 서로 겹친 몸은 있으나 명백하게 지지 없이 떠 있는 인물이나 물체는 보이지 않는다."
       }
      ],
      "all_candidates_fail": true
     },
     "normalized": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 2,
        "verdict_ko": "흩어진 병사들은 보이지만 오르막을 바라보라는 시점을 반대로 재현했고, 지지 없는 소총과 오른쪽 숲을 겨누지 않는 총구들이 핵심 연출을 훼손합니다."
       },
       {
        "label": "A",
        "score": 4,
        "verdict_ko": "양쪽 숲을 향한 상반된 조준과 쓰러진 병사들의 뒤엉킴은 더 충실하지만, 참고 사진의 내리막 시점을 반복하여 필수 오르막 구도를 위반합니다."
       }
      ],
      "all_candidates_fail": true,
      "readings": [
       {
        "label": "B",
        "direction": "왼쪽 나무 옆 병사는 시선과 총구를 왼쪽 비탈의 수풀로 향한다. 오른쪽 전경 병사는 오른쪽 숲을 보며 이동하지만 소총 총구는 발치 쪽으로 내려가 있다. 중앙에서 수평으로 총을 든 병사의 총은 오른쪽 끝이 개머리판이어서 오른쪽 숲을 조준하는 모습이 아니다. 뒤쪽 병사들은 서로 다른 방향으로 몸을 돌리고, 주저앉은 병사는 아래를 본다. 몸 방향의 혼란은 있지만 양쪽 숲으로 갈라지는 조준은 약하다.",
        "built_space": "인공 구조물이나 고정 설비는 없다. 하나의 좁은 흙길, 왼쪽의 뿌리 드러난 높은 사면, 오른쪽의 낮은 숲과 먼 계곡이 보인다. 나무 배치와 지면 재질은 참고 장소에 가깝다. 그러나 길이 전경에서 화면 중앙의 낮은 곳으로 내려가므로 카메라는 위쪽에서 내리막을 본다. 요구된 아래쪽 가장자리에서 비스듬히 오르막을 보는 공간 방향과 반대다.",
        "entities": "부분적으로 가려진 인물을 포함해 약 일곱 명의 병사가 보이며, 중앙 한 명은 주저앉아 있다. 보이는 얼굴들은 성인 동아시아계 남성으로 읽히고 군복과 군모는 일본군 설정에 부합하지만 국적이나 개별 신원은 외모만으로 확인할 수 없다. 목재 개머리판과 금속 총열을 가진 장총들이 있으나 참고 소품의 부싯돌식 격발 장치와 황동 띠를 정확히 재현하지는 않는다. 읽을 수 있는 글자나 화면 분할은 없다.",
        "hard_violations": [
         "아래쪽 가장자리에서 오르막을 바라보라는 지형·시점 지시와 반대로, 위쪽에서 내리막과 계곡을 내려다본다.",
         "왼쪽 조준 병사의 등 뒤에서 오른쪽 위로 뻗은 별도 소총은 손이나 연결된 멜빵 등 지지점 없이 떠 있다."
        ],
        "physics": "왼쪽 조준 병사는 한 발을 높은 뿌리 쪽에, 다른 발을 낮은 비탈에 딛고 총을 양손으로 받친다. 중앙 병사는 엉덩이와 다리가 흙길에 닿아 있고, 오른쪽 이동 병사는 앞발로 지면을 지지하면서 뒷발을 든다. 바닥의 소총은 흙과 돌에 놓여 있다. 다만 왼쪽 병사 뒤에는 등에 걸린 총과 별도로 오른쪽 위를 향한 소총이 있으며, 이를 잡은 손이나 매단 끈이 보이지 않아 아무것도 지지하지 않는다."
       },
       {
        "label": "A",
        "direction": "왼쪽 전경의 선 병사와 그 뒤 병사는 총구와 시선을 왼쪽 사면의 숲으로 향한다. 오른쪽 장교는 견착한 소총과 시선을 오른쪽 숲으로 향해 두 위협 방향이 명확히 갈린다. 중앙 병사는 왼쪽으로 팔을 뻗고, 배낭을 멘 병사는 오른쪽을 보면서 총열을 위로 세운다. 왼쪽 아래 병사는 사면을 짚고 숙이며 오른쪽 아래 병사는 길 가장자리로 몸을 낮춘다. 쓰러진 몸들도 길을 따라 서로 다른 각도로 놓여 있어 공통 사격선이 무너진 상태가 읽힌다.",
        "built_space": "인공 구조물이나 고정 설비 없이 하나의 숲길에 병사들이 밀집해 있다. 왼쪽의 높은 흙사면과 노출된 뿌리, 양옆의 나무, 오른쪽 아래로 열린 계곡은 참고 장소의 특징을 유지한다. 그러나 배경으로 갈수록 길이 낮아지고 계곡이 드러나므로 여기서도 위쪽에서 내리막을 보는 구도다. 요구된 오르막 방향의 관찰 시점은 아니다.",
        "entities": "서 있거나 웅크린 병사와 누운 병사를 합쳐 약 아홉 명이 보인다. 보이는 인물들은 성인 동아시아계 남성으로 읽히며 군모, 제복, 가죽 탄띠와 배낭은 일본군 집단 설정에 맞는다. 별도 민간인이나 적군은 보이지 않는다. 장총의 목재와 금속 재질은 실물처럼 보이지만 참고 소품의 특징적인 부싯돌식 장치와 황동 부품까지 일치하지는 않는다. 숲 사이의 옅은 연기는 산발적인 총격 상황과 부합하며 읽을 수 있는 글자나 그래픽은 없다.",
        "hard_violations": [
         "아래쪽 가장자리에서 오르막을 바라보라는 지형·시점 지시와 반대로, 위쪽에서 내리막과 계곡을 내려다본다."
        ],
        "physics": "조준하는 병사들은 비탈과 길에 발을 벌려 딛고 소총을 손과 어깨로 지지한다. 왼쪽 아래 병사는 손과 무릎으로 사면에 기대며, 오른쪽 아래 병사는 무릎과 하체를 지면에 대고 몸을 낮춘다. 중앙과 전경의 쓰러진 병사들은 등·옆구리·다리가 흙길에 닿아 있다. 떨어진 소총은 사면이나 길 위에 놓이고, 세운 소총은 병사의 손에 잡혀 있다. 서로 겹친 몸은 있으나 명백하게 지지 없이 떠 있는 인물이나 물체는 보이지 않는다."
       }
      ]
     },
     "ok": true
    }
   ],
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "slot_winner_match": true,
   "slot_winner": {
    "gemini-pro": "A",
    "gpt-high": "A"
   },
   "route": "cross_slot_agree"
  },
  "dual": {
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "normalized": {
    "A": 2.0,
    "B": 0.875
   },
   "adjusted": {
    "A": 1.75,
    "B": 0.625
   },
   "violations": {
    "B": [
     "[gemini-pro] 좌측 상단 허공에 아무도 쥐고 있지 않은 소총이 수직으로 떠 있음",
     "[gpt-high] 아래쪽 가장자리에서 오르막을 바라보라는 지형·시점 지시와 반대로, 위쪽에서 내리막과 계곡을 내려다본다.",
     "[gpt-high] 왼쪽 조준 병사의 등 뒤에서 오른쪽 위로 뻗은 별도 소총은 손이나 연결된 멜빵 등 지지점 없이 떠 있다."
    ],
    "A": [
     "[gpt-high] 아래쪽 가장자리에서 오르막을 바라보라는 지형·시점 지시와 반대로, 위쪽에서 내리막과 계곡을 내려다본다."
    ]
   },
   "per_model_winner": {
    "gemini-pro": "A",
    "gpt-high": "A"
   },
   "agreed": true
  },
  "totals": {
   "A": 1750,
   "B": 625
  },
  "selected": "A",
  "ranking": [
   "A",
   "B"
  ],
  "verdicts": [
   {
    "label": "A",
    "score": 1750,
    "verdict_ko": "대열이 붕괴되어 여러 방향을 경계하는 혼란스러운 전경을 훌륭하게 연출했으며, 물리적 오류 없이 프롬프트의 지시를 충실히 따름.  ★위반: [gpt-high] 아래쪽 가장자리에서 오르막을 바라보라는 지형·시점 지시와 반대로, 위쪽에서 내리막과 계곡을 내려다본다."
   },
   {
    "label": "B",
    "score": 625,
    "verdict_ko": "사방을 향해 엇갈린 총구와 뒤엉킨 대열의 묘사가 부족하며, 무엇보다 허공에 떠 있는 소총이 렌더링되는 치명적인 물리적 오류가 발생함.  ★위반: [gemini-pro] 좌측 상단 허공에 아무도 쥐고 있지 않은 소총이 수직으로 떠 있음 / [gpt-high] 아래쪽 가장자리에서 오르막을 바라보라는 지형·시점 지시와 반대로, 위쪽에서 내리막과 계곡을 내려다본다. / [gpt-high] 왼쪽 조준 병사의 등 뒤에서 오른쪽 위로 뻗은 별도 소총은 손이나 연결된 멜빵 등 지지점 없이 떠 있다."
   }
  ],
  "refs": [
   {
    "label": "PREVIOUS SHOT STILL — a visually related earlier shot of this same place: the location's look, materials, fixed features and lighting mood are LOCKED to this photo; never copy its camera framing. The people visible in that photo are NOT in this shot: never carry their faces, bodies or clothing onto anyone here. Each person in THIS shot is defined solely by the PEOPLE list, their own CHARACTER REFERENCE images and their wardrobe notes. If this photo shows a character who cannot move (dead or unconscious), that character's exact pose, position and orientation are ALSO LOCKED — treat their body as a fixed prop of the set that only the camera moves around.",
    "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/db00517b-165b-4005-a287-f28936b59b95/images/272bd7af-9294-4490-98c6-88a40e9deacc/scene/recipe/S59sh11_sel.png",
    "asset_id": "b9a52fcc-b430-4863-a400-27d31821b15e",
    "role": "prev_still"
   },
   {
    "label": "PROP REFERENCE — 근대식 소총: the exact object appearing in this shot; match its look, material and wear exactly.",
    "path": "<bytes:828568>",
    "asset_id": "7543806c-8087-4caa-8d7b-56d10864fd3c",
    "role": "prop_ref"
   }
  ],
  "critique_skipped": true,
  "shot_run_produced": false,
  "shot_run_uid": "06aa1445-4bfe-70de-aa1c-6159ec501f6c",
  "ref_mode": "prev+엔티티",
  "share_plan": {
   "ref_plan": "prev",
   "prev_anchor_tag": "S59sh11"
  }
 },
 "S62sh12::cine": {
  "applied": false,
  "attempted_at": "2026-09-09T09:37:18.375625+00:00",
  "fingerprint": "92f6a6d0ff5a70f7e882fb680f442ba386d604b4e89e97f38189bd14c394dcaf",
  "fingerprint_version": 2,
  "provider": "mai",
  "endpoint": "openrouter/chat-completions",
  "model": "microsoft/mai-image-2.6",
  "multiroll_tag": "still_S62sh12_cine",
  "slot": "",
  "pack": "24.202608252115",
  "source_file": "S62sh12_sel.png",
  "source_sha256": "042cfb66a41f5cd25492ccacc73e7816554d90b38db8f45d7ff422cc9933ef86",
  "error": "RuntimeError: moderation blocked: {\"message\": \"Response content blocked by label 'MultiSeverity_ViolenceScore'.\", \"code\": 400}",
  "moderation_refusals": 2,
  "declined": true,
  "declined_reason": "moderation"
 },
 "S62sh19::signage": {
  "fp": "c7a0e0ec4eec977f",
  "inscriptions": [],
  "cues": [],
  "dropped": []
 },
 "S62sh19": {
  "input_fingerprint": "b2c3a4d2e556349a",
  "prompt": "Create ONE FINAL photorealistic live-action film still of the moment below — early 20th-century Korea, Jeolla region; all people are Korean unless explicitly stated otherwise. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 나무 틈새로 성큼 걸어 나오는 mid-stride 순간의 피와 흙투성이인 안규홍의 거대한 전신.\n\nLOCATION (lock): A gap between trees opening onto the mountain camp's sloping descent trail, directly before the halted soldiers. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: wide shot\n- KEY BACKGROUND ELEMENTS: Trees flanking the emergence point (안규홍 is stepping through the gap between them); used as Narrow edge framing establishes the opening without obscuring his head or feet; Uphill forest path (The confrontation is forming along this route) — The path runs obliquely from the tree opening toward the right edge; used as Provides ground contact and a readable direction toward his unseen opponent.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Natural daylight with restrained saturation and naturalistic contrast keeps the blood and earth legible without turning the entrance into a heroic tableau.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The people visible in that photo are NOT in this shot: never carry their faces, bodies or clothing onto anyone here. Each person in THIS shot is defined solely by the PEOPLE list, their own CHARACTER REFERENCE images and their wardrobe notes. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nPREVIOUS STILL USAGE (follow exactly — what to take from the attached still and what to exclude; it governs place and objects only, never who is in this shot or how the camera sees them): Take the wooded slope, tree trunks, leaf litter, and daylight. Exclude soldiers who have already fallen from any standing group and avoid duplicating discarded weapons.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): Several fallen soldiers lie along the wooded uphill route. 안규홍: He steps out covered in blood and dirt, with his back wound bleeding and his sickles and captured rifle retained.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 안규홍 (한국인 남성, 청년 나이대, 검은 머리, 단단히 묶은 머리) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.",
  "roll_prompts": {
   "A": "Create ONE FINAL photorealistic live-action film still of the moment below — early 20th-century Korea, Jeolla region; all people are Korean unless explicitly stated otherwise. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 나무 틈새로 성큼 걸어 나오는 mid-stride 순간의 피와 흙투성이인 안규홍의 거대한 전신.\n\nLOCATION (lock): A gap between trees opening onto the mountain camp's sloping descent trail, directly before the halted soldiers. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: wide shot\n- KEY BACKGROUND ELEMENTS: Trees flanking the emergence point (안규홍 is stepping through the gap between them); used as Narrow edge framing establishes the opening without obscuring his head or feet; Uphill forest path (The confrontation is forming along this route) — The path runs obliquely from the tree opening toward the right edge; used as Provides ground contact and a readable direction toward his unseen opponent.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Natural daylight with restrained saturation and naturalistic contrast keeps the blood and earth legible without turning the entrance into a heroic tableau.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The people visible in that photo are NOT in this shot: never carry their faces, bodies or clothing onto anyone here. Each person in THIS shot is defined solely by the PEOPLE list, their own CHARACTER REFERENCE images and their wardrobe notes. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nPREVIOUS STILL USAGE (follow exactly — what to take from the attached still and what to exclude; it governs place and objects only, never who is in this shot or how the camera sees them): Take the wooded slope, tree trunks, leaf litter, and daylight. Exclude soldiers who have already fallen from any standing group and avoid duplicating discarded weapons.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): Several fallen soldiers lie along the wooded uphill route. 안규홍: He steps out covered in blood and dirt, with his back wound bleeding and his sickles and captured rifle retained.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 안규홍 (한국인 남성, 청년 나이대, 검은 머리, 단단히 묶은 머리) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.",
   "B": "Create ONE FINAL photorealistic live-action film still of the moment below — early 20th-century Korea, Jeolla region; all people are Korean unless explicitly stated otherwise. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 나무 틈새로 성큼 걸어 나오는 mid-stride 순간의 피와 흙투성이인 안규홍의 거대한 전신.\n\nLOCATION (lock): A gap between trees opening onto the mountain camp's sloping descent trail, directly before the halted soldiers. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: wide shot\n- KEY BACKGROUND ELEMENTS: Trees flanking the emergence point (안규홍 is stepping through the gap between them); used as Narrow edge framing establishes the opening without obscuring his head or feet; Uphill forest path (The confrontation is forming along this route) — The path runs obliquely from the tree opening toward the right edge; used as Provides ground contact and a readable direction toward his unseen opponent.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Natural daylight with restrained saturation and naturalistic contrast keeps the blood and earth legible without turning the entrance into a heroic tableau.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The people visible in that photo are NOT in this shot: never carry their faces, bodies or clothing onto anyone here. Each person in THIS shot is defined solely by the PEOPLE list, their own CHARACTER REFERENCE images and their wardrobe notes. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nPREVIOUS STILL USAGE (follow exactly — what to take from the attached still and what to exclude; it governs place and objects only, never who is in this shot or how the camera sees them): Take the wooded slope, tree trunks, leaf litter, and daylight. Exclude soldiers who have already fallen from any standing group and avoid duplicating discarded weapons.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): Several fallen soldiers lie along the wooded uphill route. 안규홍: He steps out covered in blood and dirt, with his back wound bleeding and his sickles and captured rifle retained.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 안규홍 (한국인 남성, 청년 나이대, 검은 머리, 단단히 묶은 머리) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.\n\nCOMPOSITION VARIATION (this candidate only): keep the place, the people and the moment exactly as specified, and keep from the CAMERA & FRAME contract its subject scale, screen placement and key background elements exactly. For this candidate only, its camera angle and height are deliberately left open: choose a camera position distinctly different from the obvious one, as a film director picking a second setup on the same blocking. Stay true to any explicit framing the SHOT TEXT itself dictates."
  },
  "shot_run_spend_attempt_count": 1,
  "readings": [
   {
    "label": "A",
    "direction": "주인공의 시선과 몸의 방향이 화면 우측을 향하고 있으며, 산길 역시 우측 위로 경사져 이어집니다.",
    "built_space": "화면 좌우 가장자리에 곧게 뻗은 나무들이 프레임을 형성하여 틈새를 만들어내며, 흙과 낙엽이 깔린 오르막길이 우측으로 비스듬히 나 있습니다.",
    "entities": "안규홍의 얼굴, 머리 모양, 복장이 레퍼런스와 일치하며 옷 전체에 피와 흙이 묻어 있습니다. 오른손에는 소총을, 왼손에는 두 개의 낫을 쥐고 있습니다. 배경의 오르막길에는 이전 샷의 복장(황갈색 군복)을 한 군인들이 쓰러져 있습니다.",
    "hard_violations": [
     "[gpt-high] 이전 사진의 쓰러진 병사들을 고정된 자세와 위치로 보존하지 않고, 바로 누워 비탈에 기대는 등의 다른 자세와 배치로 변경했다."
    ],
    "physics": "주인공은 땅을 확실히 딛고 자연스러운 걷기(mid-stride) 동작을 취하고 있으며, 손에 든 무기들도 물리적으로 올바르게 지탱되고 있습니다."
   },
   {
    "label": "B",
    "direction": "주인공의 시선은 정면을 향하고 있으며, 산길은 화면 우측 위로 경사져 있습니다.",
    "built_space": "양옆으로 빽빽한 나무들이 배치되어 틈새를 형성하고, 우측으로 비스듬히 올라가는 흙길이 잘 묘사되어 있습니다.",
    "entities": "안규홍의 외모와 의상은 레퍼런스와 일치하나 피와 흙의 묘사가 다소 약합니다. 왼손에 소총을, 오른손에 단 한 개의 낫만을 들고 있습니다. 오르막길에는 쓰러진 군인들이 배치되어 있습니다.",
    "hard_violations": [
     "[gpt-high] 이전 사진에서 고정되어야 하는 쓰러진 병사들의 자세와 접지 위치를 유지하지 않고, 길 양쪽에 다른 자세의 병사들을 재배치했다."
    ],
    "physics": "땅을 딛고 걸어 나오는 동작이 안정적이며, 쥐고 있는 무기와 몸의 균형 등 물리적 지탱 요소에 이상이 없습니다."
   }
  ],
  "cross_model_order": {
   "policy": "cross_model_order_v2_slot0_fwd_slot1_rev_parallel2",
   "slots": [
    {
     "model": "gemini-pro",
     "order": "forward",
     "display_to_canonical": {
      "A": "A",
      "B": "B"
     },
     "raw": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 10,
        "verdict_ko": "프롬프트가 요구한 '피와 흙투성이', '낫들(복수)', '노획한 소총'의 디테일을 모두 정확하게 구현하였으며, 전반적인 분위기와 구도도 훌륭합니다."
       },
       {
        "label": "B",
        "score": 7,
        "verdict_ko": "프레임과 배경, 인물의 외형은 잘 맞췄으나 낫을 하나만 들고 있고 피와 흙으로 덮인 정도가 프롬프트의 강렬한 묘사에 비해 부족합니다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "주인공의 시선과 몸의 방향이 화면 우측을 향하고 있으며, 산길 역시 우측 위로 경사져 이어집니다.",
        "built_space": "화면 좌우 가장자리에 곧게 뻗은 나무들이 프레임을 형성하여 틈새를 만들어내며, 흙과 낙엽이 깔린 오르막길이 우측으로 비스듬히 나 있습니다.",
        "entities": "안규홍의 얼굴, 머리 모양, 복장이 레퍼런스와 일치하며 옷 전체에 피와 흙이 묻어 있습니다. 오른손에는 소총을, 왼손에는 두 개의 낫을 쥐고 있습니다. 배경의 오르막길에는 이전 샷의 복장(황갈색 군복)을 한 군인들이 쓰러져 있습니다.",
        "hard_violations": [],
        "physics": "주인공은 땅을 확실히 딛고 자연스러운 걷기(mid-stride) 동작을 취하고 있으며, 손에 든 무기들도 물리적으로 올바르게 지탱되고 있습니다."
       },
       {
        "label": "B",
        "direction": "주인공의 시선은 정면을 향하고 있으며, 산길은 화면 우측 위로 경사져 있습니다.",
        "built_space": "양옆으로 빽빽한 나무들이 배치되어 틈새를 형성하고, 우측으로 비스듬히 올라가는 흙길이 잘 묘사되어 있습니다.",
        "entities": "안규홍의 외모와 의상은 레퍼런스와 일치하나 피와 흙의 묘사가 다소 약합니다. 왼손에 소총을, 오른손에 단 한 개의 낫만을 들고 있습니다. 오르막길에는 쓰러진 군인들이 배치되어 있습니다.",
        "hard_violations": [],
        "physics": "땅을 딛고 걸어 나오는 동작이 안정적이며, 쥐고 있는 무기와 몸의 균형 등 물리적 지탱 요소에 이상이 없습니다."
       }
      ],
      "all_candidates_fail": false
     },
     "normalized": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 10,
        "verdict_ko": "프롬프트가 요구한 '피와 흙투성이', '낫들(복수)', '노획한 소총'의 디테일을 모두 정확하게 구현하였으며, 전반적인 분위기와 구도도 훌륭합니다."
       },
       {
        "label": "B",
        "score": 7,
        "verdict_ko": "프레임과 배경, 인물의 외형은 잘 맞췄으나 낫을 하나만 들고 있고 피와 흙으로 덮인 정도가 프롬프트의 강렬한 묘사에 비해 부족합니다."
       }
      ],
      "all_candidates_fail": false,
      "readings": [
       {
        "label": "A",
        "direction": "주인공의 시선과 몸의 방향이 화면 우측을 향하고 있으며, 산길 역시 우측 위로 경사져 이어집니다.",
        "built_space": "화면 좌우 가장자리에 곧게 뻗은 나무들이 프레임을 형성하여 틈새를 만들어내며, 흙과 낙엽이 깔린 오르막길이 우측으로 비스듬히 나 있습니다.",
        "entities": "안규홍의 얼굴, 머리 모양, 복장이 레퍼런스와 일치하며 옷 전체에 피와 흙이 묻어 있습니다. 오른손에는 소총을, 왼손에는 두 개의 낫을 쥐고 있습니다. 배경의 오르막길에는 이전 샷의 복장(황갈색 군복)을 한 군인들이 쓰러져 있습니다.",
        "hard_violations": [],
        "physics": "주인공은 땅을 확실히 딛고 자연스러운 걷기(mid-stride) 동작을 취하고 있으며, 손에 든 무기들도 물리적으로 올바르게 지탱되고 있습니다."
       },
       {
        "label": "B",
        "direction": "주인공의 시선은 정면을 향하고 있으며, 산길은 화면 우측 위로 경사져 있습니다.",
        "built_space": "양옆으로 빽빽한 나무들이 배치되어 틈새를 형성하고, 우측으로 비스듬히 올라가는 흙길이 잘 묘사되어 있습니다.",
        "entities": "안규홍의 외모와 의상은 레퍼런스와 일치하나 피와 흙의 묘사가 다소 약합니다. 왼손에 소총을, 오른손에 단 한 개의 낫만을 들고 있습니다. 오르막길에는 쓰러진 군인들이 배치되어 있습니다.",
        "hard_violations": [],
        "physics": "땅을 딛고 걸어 나오는 동작이 안정적이며, 쥐고 있는 무기와 몸의 균형 등 물리적 지탱 요소에 이상이 없습니다."
       }
      ]
     },
     "ok": true
    },
    {
     "model": "gpt-high",
     "order": "reverse",
     "display_to_canonical": {
      "A": "B",
      "B": "A"
     },
     "raw": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 3,
        "verdict_ko": "전신과 실제 보행 지지는 확보했지만 오른쪽 상대가 아닌 렌즈 쪽으로 전진하며, 낫은 하나만 보이고 쓰러진 병사들의 고정 배치를 바꿨다."
       },
       {
        "label": "B",
        "score": 4,
        "verdict_ko": "오른쪽을 향한 큰 보폭, 피투성이 전신, 낫 두 자루와 소총은 더 충실하지만 병사들의 고정 자세·배치를 변경해 최종 사용에는 부적합하다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "안규홍의 얼굴과 시선, 전진 방향이 거의 카메라 정면을 향한다. 요구된 화면 오른쪽의 보이지 않는 상대를 향하는 동선은 약하다. 손에 든 소총의 총구는 화면 왼쪽 아래의 땅을 향하며 누구를 조준하지 않는다. 낫날도 아래로 내려가 있다.",
        "built_space": "인공 구조물이나 고정 설비는 없다. 앞쪽 왼편의 굵은 나무 두 그루와 오른편 한 그루가 통로를 둘러싸고, 인물의 머리와 발은 가리지 않는다. 길은 인물 뒤에서 오른쪽 위로 이어지지만 인물의 걸음은 화면 앞으로 빠져나온다. 나무껍질, 낙엽, 흙비탈과 낮의 숲은 참고 장소와 유사하다. 여러 병사가 길 양쪽에 새롭게 배열되어 있어 이전 사진의 고정된 인체 배치와 일치하지 않는다.",
        "entities": "중앙에는 검은 머리를 뒤로 단단히 넘겨 묶은 청년 한국인 남성으로 보이는 인물 한 명이 있다. 얼굴과 마른 체격, 밝은 저고리와 바지, 천 허리띠, 짚신은 인물 참고와 대체로 맞는다. 얼굴과 옷에는 흙과 피가 보이지만 등 상처는 정면 구도라 확인할 수 없다. 손에는 소총 한 자루와 낫 한 자루만 보이며, 요구된 복수의 낫은 확인되지 않는다. 주변에는 적어도 여섯 명의 쓰러진 군복 인물과 여러 소총이 보인다. 읽을 수 있는 글자는 없다.",
        "hard_violations": [
         "이전 사진에서 고정되어야 하는 쓰러진 병사들의 자세와 접지 위치를 유지하지 않고, 길 양쪽에 다른 자세의 병사들을 재배치했다."
        ],
        "physics": "화면 왼쪽의 앞발이 흙길을 딛고 몸무게를 지지하며, 반대쪽 발은 뒤에서 들려 앞으로 이동하는 보행 순간이다. 공중에 뜬 몸은 아니다. 소총과 낫은 각각 손으로 잡고 있고 소총 끈은 중력 방향으로 처진다. 병사들은 비탈과 낙엽 위에 몸을 대고 있으며, 바닥의 총들도 지면에 놓여 있다."
       },
       {
        "label": "B",
        "direction": "안규홍은 몸과 얼굴을 화면 오른쪽으로 돌리고 그쪽 화면 밖을 응시한다. 오른쪽으로 벌어진 큰 보폭이 보이지 않는 상대에게 접근하는 방향을 명확히 만든다. 소총 총구는 오른쪽 아래 지면을 향하고, 두 낫의 날도 오른쪽 아래로 늘어져 있어 공격 조준이 아닌 휴대 상태다.",
        "built_space": "인공 구조물이나 설비는 없다. 좌우의 굵은 나무 두 그루 사이로 인물의 전신이 드러난다. 두 줄기가 가장자리의 상당 부분을 차지해 요구한 좁은 가장자리 프레이밍보다는 두껍다. 낙엽 덮인 길과 비탈은 나무 틈에서 오른쪽 위로 이어져 요구된 사선 경로를 더 잘 보여 준다. 숲의 재질과 자연광은 참고와 유사하지만 바닥과 오른쪽 비탈의 병사들은 이전 사진과 다른 자세로 놓여 있다.",
        "entities": "검은 머리를 뒤로 묶은 청년 한국인 남성 한 명이 중심에 있으며, 얼굴 윤곽과 체격은 참고 인물에 대체로 부합한다. 밝은 전통 상하의, 허리띠와 짚신이 보이고 얼굴·상의·바지에 피와 흙이 묻어 있다. 등 윗부분에는 피에 젖은 천이 돌출되어 보이나 상처 자체는 명확하지 않다. 한 손에는 소총 한 자루, 다른 손에는 구별되는 낫 두 자루가 있다. 바닥과 오른쪽 비탈에는 쓰러진 군복 인물 세 명과 소총들이 보인다. 읽을 수 있는 글자는 없다.",
        "hard_violations": [
         "이전 사진의 쓰러진 병사들을 고정된 자세와 위치로 보존하지 않고, 바로 누워 비탈에 기대는 등의 다른 자세와 배치로 변경했다."
        ],
        "physics": "오른쪽으로 내디딘 앞발과 뒤쪽 발이 흙길에 접촉하고, 벌어진 다리와 앞으로 기울어진 몸통이 체중 이동을 지지한다. 성큼 걷는 순간으로 가능한 자세이며 지지 없이 떠 있지 않다. 소총과 두 낫은 손에 잡혀 있고 총의 멜빵은 아래로 처진다. 병사들의 몸과 바닥의 소총은 비탈의 지지를 받는다."
       }
      ],
      "all_candidates_fail": true
     },
     "normalized": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 3,
        "verdict_ko": "전신과 실제 보행 지지는 확보했지만 오른쪽 상대가 아닌 렌즈 쪽으로 전진하며, 낫은 하나만 보이고 쓰러진 병사들의 고정 배치를 바꿨다."
       },
       {
        "label": "A",
        "score": 4,
        "verdict_ko": "오른쪽을 향한 큰 보폭, 피투성이 전신, 낫 두 자루와 소총은 더 충실하지만 병사들의 고정 자세·배치를 변경해 최종 사용에는 부적합하다."
       }
      ],
      "all_candidates_fail": true,
      "readings": [
       {
        "label": "B",
        "direction": "안규홍의 얼굴과 시선, 전진 방향이 거의 카메라 정면을 향한다. 요구된 화면 오른쪽의 보이지 않는 상대를 향하는 동선은 약하다. 손에 든 소총의 총구는 화면 왼쪽 아래의 땅을 향하며 누구를 조준하지 않는다. 낫날도 아래로 내려가 있다.",
        "built_space": "인공 구조물이나 고정 설비는 없다. 앞쪽 왼편의 굵은 나무 두 그루와 오른편 한 그루가 통로를 둘러싸고, 인물의 머리와 발은 가리지 않는다. 길은 인물 뒤에서 오른쪽 위로 이어지지만 인물의 걸음은 화면 앞으로 빠져나온다. 나무껍질, 낙엽, 흙비탈과 낮의 숲은 참고 장소와 유사하다. 여러 병사가 길 양쪽에 새롭게 배열되어 있어 이전 사진의 고정된 인체 배치와 일치하지 않는다.",
        "entities": "중앙에는 검은 머리를 뒤로 단단히 넘겨 묶은 청년 한국인 남성으로 보이는 인물 한 명이 있다. 얼굴과 마른 체격, 밝은 저고리와 바지, 천 허리띠, 짚신은 인물 참고와 대체로 맞는다. 얼굴과 옷에는 흙과 피가 보이지만 등 상처는 정면 구도라 확인할 수 없다. 손에는 소총 한 자루와 낫 한 자루만 보이며, 요구된 복수의 낫은 확인되지 않는다. 주변에는 적어도 여섯 명의 쓰러진 군복 인물과 여러 소총이 보인다. 읽을 수 있는 글자는 없다.",
        "hard_violations": [
         "이전 사진에서 고정되어야 하는 쓰러진 병사들의 자세와 접지 위치를 유지하지 않고, 길 양쪽에 다른 자세의 병사들을 재배치했다."
        ],
        "physics": "화면 왼쪽의 앞발이 흙길을 딛고 몸무게를 지지하며, 반대쪽 발은 뒤에서 들려 앞으로 이동하는 보행 순간이다. 공중에 뜬 몸은 아니다. 소총과 낫은 각각 손으로 잡고 있고 소총 끈은 중력 방향으로 처진다. 병사들은 비탈과 낙엽 위에 몸을 대고 있으며, 바닥의 총들도 지면에 놓여 있다."
       },
       {
        "label": "A",
        "direction": "안규홍은 몸과 얼굴을 화면 오른쪽으로 돌리고 그쪽 화면 밖을 응시한다. 오른쪽으로 벌어진 큰 보폭이 보이지 않는 상대에게 접근하는 방향을 명확히 만든다. 소총 총구는 오른쪽 아래 지면을 향하고, 두 낫의 날도 오른쪽 아래로 늘어져 있어 공격 조준이 아닌 휴대 상태다.",
        "built_space": "인공 구조물이나 설비는 없다. 좌우의 굵은 나무 두 그루 사이로 인물의 전신이 드러난다. 두 줄기가 가장자리의 상당 부분을 차지해 요구한 좁은 가장자리 프레이밍보다는 두껍다. 낙엽 덮인 길과 비탈은 나무 틈에서 오른쪽 위로 이어져 요구된 사선 경로를 더 잘 보여 준다. 숲의 재질과 자연광은 참고와 유사하지만 바닥과 오른쪽 비탈의 병사들은 이전 사진과 다른 자세로 놓여 있다.",
        "entities": "검은 머리를 뒤로 묶은 청년 한국인 남성 한 명이 중심에 있으며, 얼굴 윤곽과 체격은 참고 인물에 대체로 부합한다. 밝은 전통 상하의, 허리띠와 짚신이 보이고 얼굴·상의·바지에 피와 흙이 묻어 있다. 등 윗부분에는 피에 젖은 천이 돌출되어 보이나 상처 자체는 명확하지 않다. 한 손에는 소총 한 자루, 다른 손에는 구별되는 낫 두 자루가 있다. 바닥과 오른쪽 비탈에는 쓰러진 군복 인물 세 명과 소총들이 보인다. 읽을 수 있는 글자는 없다.",
        "hard_violations": [
         "이전 사진의 쓰러진 병사들을 고정된 자세와 위치로 보존하지 않고, 바로 누워 비탈에 기대는 등의 다른 자세와 배치로 변경했다."
        ],
        "physics": "오른쪽으로 내디딘 앞발과 뒤쪽 발이 흙길에 접촉하고, 벌어진 다리와 앞으로 기울어진 몸통이 체중 이동을 지지한다. 성큼 걷는 순간으로 가능한 자세이며 지지 없이 떠 있지 않다. 소총과 두 낫은 손에 잡혀 있고 총의 멜빵은 아래로 처진다. 병사들의 몸과 바닥의 소총은 비탈의 지지를 받는다."
       }
      ]
     },
     "ok": true
    }
   ],
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "slot_winner_match": true,
   "slot_winner": {
    "gemini-pro": "A",
    "gpt-high": "A"
   },
   "route": "cross_slot_agree"
  },
  "dual": {
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "normalized": {
    "A": 2.0,
    "B": 1.45
   },
   "adjusted": {
    "A": 1.75,
    "B": 1.2
   },
   "violations": {
    "B": [
     "[gpt-high] 이전 사진에서 고정되어야 하는 쓰러진 병사들의 자세와 접지 위치를 유지하지 않고, 길 양쪽에 다른 자세의 병사들을 재배치했다."
    ],
    "A": [
     "[gpt-high] 이전 사진의 쓰러진 병사들을 고정된 자세와 위치로 보존하지 않고, 바로 누워 비탈에 기대는 등의 다른 자세와 배치로 변경했다."
    ]
   },
   "per_model_winner": {
    "gemini-pro": "A",
    "gpt-high": "A"
   },
   "agreed": true
  },
  "totals": {
   "A": 1750,
   "B": 1200
  },
  "selected": "A",
  "ranking": [
   "A",
   "B"
  ],
  "verdicts": [
   {
    "label": "A",
    "score": 1750,
    "verdict_ko": "프롬프트가 요구한 '피와 흙투성이', '낫들(복수)', '노획한 소총'의 디테일을 모두 정확하게 구현하였으며, 전반적인 분위기와 구도도 훌륭합니다.  ★위반: [gpt-high] 이전 사진의 쓰러진 병사들을 고정된 자세와 위치로 보존하지 않고, 바로 누워 비탈에 기대는 등의 다른 자세와 배치로 변경했다."
   },
   {
    "label": "B",
    "score": 1200,
    "verdict_ko": "프레임과 배경, 인물의 외형은 잘 맞췄으나 낫을 하나만 들고 있고 피와 흙으로 덮인 정도가 프롬프트의 강렬한 묘사에 비해 부족합니다.  ★위반: [gpt-high] 이전 사진에서 고정되어야 하는 쓰러진 병사들의 자세와 접지 위치를 유지하지 않고, 길 양쪽에 다른 자세의 병사들을 재배치했다."
   }
  ],
  "refs": [
   {
    "label": "PREVIOUS SHOT STILL — a visually related earlier shot of this same place: the location's look, materials, fixed features and lighting mood are LOCKED to this photo; never copy its camera framing. The people visible in that photo are NOT in this shot: never carry their faces, bodies or clothing onto anyone here. Each person in THIS shot is defined solely by the PEOPLE list, their own CHARACTER REFERENCE images and their wardrobe notes. If this photo shows a character who cannot move (dead or unconscious), that character's exact pose, position and orientation are ALSO LOCKED — treat their body as a fixed prop of the set that only the camera moves around.",
    "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/db00517b-165b-4005-a287-f28936b59b95/images/272bd7af-9294-4490-98c6-88a40e9deacc/scene/recipe/S62sh12_sel.png",
    "asset_id": "6100486a-77b6-4608-9ca6-f18738ad46c7",
    "role": "prev_still"
   },
   {
    "label": "CHARACTER REFERENCE — 안규홍: the exact person appearing in this shot; match face, hair and build exactly.",
    "path": "<bytes:1028384>",
    "asset_id": "3b5894db-ce55-4f34-b966-5e725c3aa952",
    "role": "character_ref"
   }
  ],
  "critique_skipped": true,
  "shot_run_produced": false,
  "shot_run_uid": "06aa1446-5fff-72d3-9139-f2c9437d83ac",
  "ref_mode": "prev+엔티티",
  "share_plan": {
   "ref_plan": "prev",
   "prev_anchor_tag": "S62sh12"
  }
 },
 "S62sh19::cine": {
  "applied": true,
  "attempted_at": "2026-09-09T08:10:48.143753+00:00",
  "fingerprint": "94392c902d64174444419cdd620308ce4f9b93f81ffcc32941a12d438dbc3d9c",
  "fingerprint_version": 2,
  "provider": "mai",
  "endpoint": "openrouter/chat-completions",
  "model": "microsoft/mai-image-2.6",
  "pack": "24.202608252115",
  "source_file": "S62sh19_sel.png",
  "source_sha256": "33b89b269517f7e867ae234d899d408bdb86cf4a728e88ba04cf6512db6a2e7b",
  "file": "S62sh19_cine.png",
  "staged_sha256": "27095c9463b1d9190ede382d2b380bd6f3bc437ad5e9345ba3715eed5f394d6f",
  "latency_ms": 29373
 },
 "S62sh22::signage": {
  "fp": "899e40d7a2a785c3",
  "inscriptions": [],
  "cues": [],
  "dropped": []
 },
 "S62sh22": {
  "input_fingerprint": "43a1a68d8e66e36e",
  "prompt": "Create ONE FINAL photorealistic live-action film still of the moment below — early 20th-century Korea, Jeolla region; all people are Korean unless explicitly stated otherwise. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 두 사람의 팽팽한 시선이 교차하는 산길 숲속의 긴장된 대치 전경.\n\nLOCATION (lock): The forested mountain path below the camp, at the face-to-face confrontation between the trees. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: wide shot\n- KEY BACKGROUND ELEMENTS: Mountain path between the opponents (An open interval separates the two men) — The confrontation runs laterally across the image; used as Makes their physical separation the central compositional subject; Downhill continuation of the path (Extends behind 토미이시) — Recedes beyond the right-hand figure; used as Preserves the established retreat geography without foregrounding the retreat; Surrounding trees (Enclose the forest route); used as Provide an irregular perimeter around the open confrontation space.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Maintain the preceding daylight balance and restrained contrast, allowing the empty separation between the men to carry the tension.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The clothing, hair and overall look of 안규홍 — who appear both in that photo and in this shot — are LOCKED to that photo. Anyone else visible in that photo is NOT in this shot: never carry their face, body or clothing onto anyone here. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nPREVIOUS STILL USAGE (follow exactly — what to take from the attached still and what to exclude; it governs place and objects only, never who is in this shot or how the camera sees them): Take the forest slope, exposed soil, tree trunks, and daytime color. Exclude the fleeing civilian guide and do not restore the earlier intact military formation.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The forest path remains strewn with fallen soldiers from the successive shots. 안규홍: He stands halted, carrying his sickles and captured rifle, his face and clothing covered in blood and dirt and his back still bleeding. 토미이시: He stands halted with his sword retained.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 안규홍 (한국인 남성, 청년 나이대, 검은 머리, 단단히 묶은 머리); 토미이시 (일본인 남성, 성인 나이대, 자연스러운 검은 머리) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.",
  "roll_prompts": {
   "A": "Create ONE FINAL photorealistic live-action film still of the moment below — early 20th-century Korea, Jeolla region; all people are Korean unless explicitly stated otherwise. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 두 사람의 팽팽한 시선이 교차하는 산길 숲속의 긴장된 대치 전경.\n\nLOCATION (lock): The forested mountain path below the camp, at the face-to-face confrontation between the trees. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: wide shot\n- KEY BACKGROUND ELEMENTS: Mountain path between the opponents (An open interval separates the two men) — The confrontation runs laterally across the image; used as Makes their physical separation the central compositional subject; Downhill continuation of the path (Extends behind 토미이시) — Recedes beyond the right-hand figure; used as Preserves the established retreat geography without foregrounding the retreat; Surrounding trees (Enclose the forest route); used as Provide an irregular perimeter around the open confrontation space.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Maintain the preceding daylight balance and restrained contrast, allowing the empty separation between the men to carry the tension.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The clothing, hair and overall look of 안규홍 — who appear both in that photo and in this shot — are LOCKED to that photo. Anyone else visible in that photo is NOT in this shot: never carry their face, body or clothing onto anyone here. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nPREVIOUS STILL USAGE (follow exactly — what to take from the attached still and what to exclude; it governs place and objects only, never who is in this shot or how the camera sees them): Take the forest slope, exposed soil, tree trunks, and daytime color. Exclude the fleeing civilian guide and do not restore the earlier intact military formation.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The forest path remains strewn with fallen soldiers from the successive shots. 안규홍: He stands halted, carrying his sickles and captured rifle, his face and clothing covered in blood and dirt and his back still bleeding. 토미이시: He stands halted with his sword retained.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 안규홍 (한국인 남성, 청년 나이대, 검은 머리, 단단히 묶은 머리); 토미이시 (일본인 남성, 성인 나이대, 자연스러운 검은 머리) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.",
   "B": "Create ONE FINAL photorealistic live-action film still of the moment below — early 20th-century Korea, Jeolla region; all people are Korean unless explicitly stated otherwise. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 두 사람의 팽팽한 시선이 교차하는 산길 숲속의 긴장된 대치 전경.\n\nLOCATION (lock): The forested mountain path below the camp, at the face-to-face confrontation between the trees. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: wide shot\n- KEY BACKGROUND ELEMENTS: Mountain path between the opponents (An open interval separates the two men) — The confrontation runs laterally across the image; used as Makes their physical separation the central compositional subject; Downhill continuation of the path (Extends behind 토미이시) — Recedes beyond the right-hand figure; used as Preserves the established retreat geography without foregrounding the retreat; Surrounding trees (Enclose the forest route); used as Provide an irregular perimeter around the open confrontation space.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Maintain the preceding daylight balance and restrained contrast, allowing the empty separation between the men to carry the tension.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The clothing, hair and overall look of 안규홍 — who appear both in that photo and in this shot — are LOCKED to that photo. Anyone else visible in that photo is NOT in this shot: never carry their face, body or clothing onto anyone here. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nPREVIOUS STILL USAGE (follow exactly — what to take from the attached still and what to exclude; it governs place and objects only, never who is in this shot or how the camera sees them): Take the forest slope, exposed soil, tree trunks, and daytime color. Exclude the fleeing civilian guide and do not restore the earlier intact military formation.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The forest path remains strewn with fallen soldiers from the successive shots. 안규홍: He stands halted, carrying his sickles and captured rifle, his face and clothing covered in blood and dirt and his back still bleeding. 토미이시: He stands halted with his sword retained.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 안규홍 (한국인 남성, 청년 나이대, 검은 머리, 단단히 묶은 머리); 토미이시 (일본인 남성, 성인 나이대, 자연스러운 검은 머리) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.\n\nCOMPOSITION VARIATION (this candidate only): keep the place, the people and the moment exactly as specified, and keep from the CAMERA & FRAME contract its subject scale, screen placement and key background elements exactly. For this candidate only, its camera angle and height are deliberately left open: choose a camera position distinctly different from the obvious one, as a film director picking a second setup on the same blocking. Stay true to any explicit framing the SHOT TEXT itself dictates."
  },
  "shot_run_spend_attempt_count": 1,
  "readings": [
   {
    "label": "A",
    "direction": "안규홍은 우측의 토미이시를 노려보고, 토미이시는 좌측의 안규홍을 응시함. 안규홍의 소총은 우측 아래 바닥을 향함.",
    "built_space": "인공 구조물 없음. 숲속 중앙으로 산길이 이어지며, 두 사람의 대치 공간을 형성함. 지시된 '토미이시 뒤로 이어지는 길' 대신 중앙 뒤편으로 길이 남. 우측 경사면에 시신들이 위치함.",
    "entities": "안규홍(청년, 피 묻은 한복, 묶은 머리) 일치. 토미이시(성인, 일본군복, 수염, 칼 착용) 일치. 쓰러진 군인들 존재함.",
    "hard_violations": [],
    "physics": "두 사람 모두 두 발로 흙바닥을 안정적으로 딛고 있음. 안규홍의 오른손이 소총을, 왼손이 두 개의 낫을 지탱하고 있음."
   },
   {
    "label": "B",
    "direction": "안규홍은 우측의 토미이시를, 토미이시는 좌측의 안규홍을 마주봄. 안규홍의 소총은 우측 아래를 향함.",
    "built_space": "인공 구조물 없음. 두 사람 사이에 산길이 있고, 지시문대로 토미이시 뒤쪽(화면 우측)으로 멀어지는 산길이 묘사됨.",
    "entities": "안규홍(청년, 피 묻은 한복, 묶은 머리) 일치. 토미이시(성인, 일본군복, 수염, 칼 착용) 일치. 쓰러진 군인들이 좌측 하단에 있음.",
    "hard_violations": [
     "[gemini-pro] physically impossible staging: 좌측 하단의 쓰러진 군인들은 3D 공간의 원근감이나 지형의 경사와 맞지 않게 이전 샷의 2D 이미지를 평면적으로 잘라 붙인 듯한 콜라주 형태를 띰.",
     "[gemini-pro] a person placed where the staging does not put them: 이전 샷에서 안규홍의 전방(우측)에 위치했던 시신들이 샷 구도가 바뀌면서 공간적으로 불가능한 안규홍의 뒤쪽(화면 좌측)으로 배치됨.",
     "[gpt-high] 중앙 후방의 병사가 얼굴을 땅 쪽으로 둔 채 엎드려 웅크린 자세로 놓여 있다. 이전 스틸의 누워 있는 병사들의 자세와 접촉점을 고정하라는 조건에 어긋나며, 카메라 이동만으로 설명할 수 없다."
    ],
    "physics": "두 사람 모두 발을 땅에 딛고 서 있으나, 좌측 하단의 시신들은 바닥에 닿아있지 않고 허공에 뜬 듯한 평면적 오버레이로 물리적 지탱이 없음. 안규홍의 왼손에 쥔 낫의 손잡이가 융합되어 있음."
   }
  ],
  "cross_model_order": {
   "policy": "cross_model_order_v2_slot0_fwd_slot1_rev_parallel2",
   "slots": [
    {
     "model": "gemini-pro",
     "order": "forward",
     "display_to_canonical": {
      "A": "A",
      "B": "B"
     },
     "raw": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 7,
        "verdict_ko": "캐릭터의 외형과 긴장감 있는 대치 구도를 잘 살렸으나, 토미이시 뒤로 이어지는 산길 배경 지시와 이전 샷의 시신 자세 유지를 완벽히 따르지는 못했습니다."
       },
       {
        "label": "B",
        "score": 2,
        "verdict_ko": "토미이시 뒤로 이어지는 길의 구도는 지켰으나, 좌측 하단의 시신들이 원근감 없이 평면적으로 복사되어 붙여진 콜라주 오류와 공간적 위치 오류로 인해 실격입니다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "안규홍은 우측의 토미이시를 노려보고, 토미이시는 좌측의 안규홍을 응시함. 안규홍의 소총은 우측 아래 바닥을 향함.",
        "built_space": "인공 구조물 없음. 숲속 중앙으로 산길이 이어지며, 두 사람의 대치 공간을 형성함. 지시된 '토미이시 뒤로 이어지는 길' 대신 중앙 뒤편으로 길이 남. 우측 경사면에 시신들이 위치함.",
        "entities": "안규홍(청년, 피 묻은 한복, 묶은 머리) 일치. 토미이시(성인, 일본군복, 수염, 칼 착용) 일치. 쓰러진 군인들 존재함.",
        "hard_violations": [],
        "physics": "두 사람 모두 두 발로 흙바닥을 안정적으로 딛고 있음. 안규홍의 오른손이 소총을, 왼손이 두 개의 낫을 지탱하고 있음."
       },
       {
        "label": "B",
        "direction": "안규홍은 우측의 토미이시를, 토미이시는 좌측의 안규홍을 마주봄. 안규홍의 소총은 우측 아래를 향함.",
        "built_space": "인공 구조물 없음. 두 사람 사이에 산길이 있고, 지시문대로 토미이시 뒤쪽(화면 우측)으로 멀어지는 산길이 묘사됨.",
        "entities": "안규홍(청년, 피 묻은 한복, 묶은 머리) 일치. 토미이시(성인, 일본군복, 수염, 칼 착용) 일치. 쓰러진 군인들이 좌측 하단에 있음.",
        "hard_violations": [
         "physically impossible staging: 좌측 하단의 쓰러진 군인들은 3D 공간의 원근감이나 지형의 경사와 맞지 않게 이전 샷의 2D 이미지를 평면적으로 잘라 붙인 듯한 콜라주 형태를 띰.",
         "a person placed where the staging does not put them: 이전 샷에서 안규홍의 전방(우측)에 위치했던 시신들이 샷 구도가 바뀌면서 공간적으로 불가능한 안규홍의 뒤쪽(화면 좌측)으로 배치됨."
        ],
        "physics": "두 사람 모두 발을 땅에 딛고 서 있으나, 좌측 하단의 시신들은 바닥에 닿아있지 않고 허공에 뜬 듯한 평면적 오버레이로 물리적 지탱이 없음. 안규홍의 왼손에 쥔 낫의 손잡이가 융합되어 있음."
       }
      ],
      "all_candidates_fail": false
     },
     "normalized": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 7,
        "verdict_ko": "캐릭터의 외형과 긴장감 있는 대치 구도를 잘 살렸으나, 토미이시 뒤로 이어지는 산길 배경 지시와 이전 샷의 시신 자세 유지를 완벽히 따르지는 못했습니다."
       },
       {
        "label": "B",
        "score": 2,
        "verdict_ko": "토미이시 뒤로 이어지는 길의 구도는 지켰으나, 좌측 하단의 시신들이 원근감 없이 평면적으로 복사되어 붙여진 콜라주 오류와 공간적 위치 오류로 인해 실격입니다."
       }
      ],
      "all_candidates_fail": false,
      "readings": [
       {
        "label": "A",
        "direction": "안규홍은 우측의 토미이시를 노려보고, 토미이시는 좌측의 안규홍을 응시함. 안규홍의 소총은 우측 아래 바닥을 향함.",
        "built_space": "인공 구조물 없음. 숲속 중앙으로 산길이 이어지며, 두 사람의 대치 공간을 형성함. 지시된 '토미이시 뒤로 이어지는 길' 대신 중앙 뒤편으로 길이 남. 우측 경사면에 시신들이 위치함.",
        "entities": "안규홍(청년, 피 묻은 한복, 묶은 머리) 일치. 토미이시(성인, 일본군복, 수염, 칼 착용) 일치. 쓰러진 군인들 존재함.",
        "hard_violations": [],
        "physics": "두 사람 모두 두 발로 흙바닥을 안정적으로 딛고 있음. 안규홍의 오른손이 소총을, 왼손이 두 개의 낫을 지탱하고 있음."
       },
       {
        "label": "B",
        "direction": "안규홍은 우측의 토미이시를, 토미이시는 좌측의 안규홍을 마주봄. 안규홍의 소총은 우측 아래를 향함.",
        "built_space": "인공 구조물 없음. 두 사람 사이에 산길이 있고, 지시문대로 토미이시 뒤쪽(화면 우측)으로 멀어지는 산길이 묘사됨.",
        "entities": "안규홍(청년, 피 묻은 한복, 묶은 머리) 일치. 토미이시(성인, 일본군복, 수염, 칼 착용) 일치. 쓰러진 군인들이 좌측 하단에 있음.",
        "hard_violations": [
         "physically impossible staging: 좌측 하단의 쓰러진 군인들은 3D 공간의 원근감이나 지형의 경사와 맞지 않게 이전 샷의 2D 이미지를 평면적으로 잘라 붙인 듯한 콜라주 형태를 띰.",
         "a person placed where the staging does not put them: 이전 샷에서 안규홍의 전방(우측)에 위치했던 시신들이 샷 구도가 바뀌면서 공간적으로 불가능한 안규홍의 뒤쪽(화면 좌측)으로 배치됨."
        ],
        "physics": "두 사람 모두 발을 땅에 딛고 서 있으나, 좌측 하단의 시신들은 바닥에 닿아있지 않고 허공에 뜬 듯한 평면적 오버레이로 물리적 지탱이 없음. 안규홍의 왼손에 쥔 낫의 손잡이가 융합되어 있음."
       }
      ]
     },
     "ok": true
    },
    {
     "model": "gpt-high",
     "order": "reverse",
     "display_to_canonical": {
      "A": "B",
      "B": "A"
     },
     "raw": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 3,
        "verdict_ko": "서로를 보는 시선과 무기 휴대는 맞지만, 중앙의 쓰러진 병사가 엎드린 자세로 바뀌어 이전 장면의 시신 고정 조건을 위반하고 굵은 나무가 대치 간격을 가릅니다."
       },
       {
        "label": "B",
        "score": 8,
        "verdict_ko": "두 사람 사이의 빈 공간을 중심에 둔 와이드 대치와 오른쪽 사면의 쓰러진 병사 배치가 더 충실하지만, 퇴로는 토미이시 뒤보다 화면 중앙으로 이어집니다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "왼쪽 안규홍은 오른쪽 토미이시의 얼굴을 보고, 토미이시는 왼쪽 안규홍을 바라본다. 소총 총구는 오른쪽 아래 지면을 향하며 상대를 겨누지 않는다. 두 낫의 날은 아래로 굽어 있고, 토미이시의 칼집은 허리에서 오른쪽 아래로 내려간다. 돌진하거나 도주하는 사람은 없다.",
        "built_space": "건축물이나 인공 고정 설비는 없다. 두 사람은 흙길 좌우에 서 있지만, 그 사이 왼쪽에 굵은 나무 한 그루가 크게 들어와 열린 간격을 분할한다. 길은 토미이시 바로 뒤가 아니라 두 사람 사이 중앙 뒤로 이어진다. 노출된 흙, 뿌리, 수직 나무줄기와 낮의 녹색 숲은 참조와 유사하나, 병사들이 놓인 주요 사면은 화면 왼쪽에 나타난다.",
        "entities": "서 있는 남성은 두 명이다. 안규홍은 청년 남성의 외형, 묶은 검은 머리, 피와 흙이 묻은 밝은 전통 의복, 등의 피 묻은 천, 소총 한 정과 낫 두 자루를 갖추어 참조의 주요 특징을 따른다. 토미이시는 성인 남성의 외형과 검은 수염, 군모, 갈색 군복, 훈장, 가죽 장화 및 군도를 갖추었다. 작은 측면 얼굴이므로 정확한 얼굴 일치나 국적 자체는 확정하기 어렵다. 쓰러진 병사는 왼쪽 사면, 왼쪽 전경, 중앙 후방에 각각 한 명씩 보인다. 이들은 병사 잔류 조건에 해당하지만 중앙 병사의 엎드린 자세는 참조와 다르다. 읽을 수 있는 글자나 화면 표식은 없다.",
        "hard_violations": [
         "중앙 후방의 병사가 얼굴을 땅 쪽으로 둔 채 엎드려 웅크린 자세로 놓여 있다. 이전 스틸의 누워 있는 병사들의 자세와 접촉점을 고정하라는 조건에 어긋나며, 카메라 이동만으로 설명할 수 없다."
        ],
        "physics": "두 남성의 발은 흙길에 닿아 체중을 지탱한다. 안규홍의 소총과 낫은 손에 잡혀 있고, 소총 멜빵은 중력에 따라 처진다. 토미이시의 군도는 허리 장비에 매달려 지지된다. 쓰러진 병사들은 사면과 지면에 몸을 대고 있어 공중에 뜬 신체나 무기는 없다. 토미이시의 자세는 다소 차렷처럼 경직되어 있으나 물리적으로 불가능하지는 않다."
       },
       {
        "label": "B",
        "direction": "안규홍은 화면 오른쪽 토미이시를, 토미이시는 화면 왼쪽 안규홍을 바라보아 시선이 서로 맞선다. 안규홍의 소총 총구는 오른쪽 아래 흙길을 향하고, 낫 두 자루는 손 아래로 굽어 있다. 토미이시의 군도는 칼집에 들어간 채 뒤쪽 아래로 기울어 있다. 무기를 상대에게 겨누거나 이동하는 동작은 없다.",
        "built_space": "건축물과 인공 고정 설비는 없다. 두 남성이 길 양쪽에 떨어져 서고, 전경에서 두 사람 사이까지 빈 흙길이 이어져 물리적 거리가 중심 소재가 된다. 굵은 나무들이 양옆을 둘러싸며 중앙 간격을 막지 않는다. 오른쪽의 노출된 사면과 그 위 병사들은 참조의 지형 관계에 더 가깝다. 다만 멀어지는 길은 오른쪽 인물 뒤보다 화면 중앙에 놓여, 지정된 퇴로 위치가 정확하지 않다.",
        "entities": "서 있는 인물은 안규홍과 토미이시에 해당하는 남성 두 명이다. 안규홍의 젊은 외형, 묶은 검은 머리, 피 묻은 얼굴과 밝은 전통 의복, 허리띠, 짚신, 등의 혈흔이 참조를 따른다. 손에는 소총 한 정과 낫 두 자루가 보인다. 토미이시는 참조와 유사한 성인 남성의 옆얼굴과 수염, 군모, 훈장 달린 갈색 군복, 가죽 장화와 군도를 갖추었다. 국적은 외형만으로 확정할 수 없다. 병사는 왼쪽 나무 뒤에 부분적으로 한 명, 오른쪽 중경 사면에 두 명, 오른쪽 전경에 부분적으로 한 명이 보인다. 가림 때문에 모든 시신의 정확한 접촉점까지 확인되지는 않지만, 오른쪽 두 병사의 누운 배치는 참조와 유사하다. 민간인 안내자나 온전한 군대 대형, 읽을 수 있는 글자는 없다.",
        "hard_violations": [],
        "physics": "두 남성 모두 두 발로 지면을 딛고 서 있다. 안규홍의 손이 소총과 낫 손잡이를 지지하며, 멜빵은 아래로 늘어진다. 토미이시의 칼집은 허리 장비에 연결되어 있다. 병사들의 몸과 장비는 흙 사면에 놓여 있으며 지지 없이 떠 있는 것은 보이지 않는다. 두 사람의 정지 자세는 가능하지만 토미이시의 곧게 내린 팔과 비교적 가지런한 발은 자연스러운 중단 동작보다 경직되어 보인다."
       }
      ],
      "all_candidates_fail": false
     },
     "normalized": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 3,
        "verdict_ko": "서로를 보는 시선과 무기 휴대는 맞지만, 중앙의 쓰러진 병사가 엎드린 자세로 바뀌어 이전 장면의 시신 고정 조건을 위반하고 굵은 나무가 대치 간격을 가릅니다."
       },
       {
        "label": "A",
        "score": 8,
        "verdict_ko": "두 사람 사이의 빈 공간을 중심에 둔 와이드 대치와 오른쪽 사면의 쓰러진 병사 배치가 더 충실하지만, 퇴로는 토미이시 뒤보다 화면 중앙으로 이어집니다."
       }
      ],
      "all_candidates_fail": false,
      "readings": [
       {
        "label": "B",
        "direction": "왼쪽 안규홍은 오른쪽 토미이시의 얼굴을 보고, 토미이시는 왼쪽 안규홍을 바라본다. 소총 총구는 오른쪽 아래 지면을 향하며 상대를 겨누지 않는다. 두 낫의 날은 아래로 굽어 있고, 토미이시의 칼집은 허리에서 오른쪽 아래로 내려간다. 돌진하거나 도주하는 사람은 없다.",
        "built_space": "건축물이나 인공 고정 설비는 없다. 두 사람은 흙길 좌우에 서 있지만, 그 사이 왼쪽에 굵은 나무 한 그루가 크게 들어와 열린 간격을 분할한다. 길은 토미이시 바로 뒤가 아니라 두 사람 사이 중앙 뒤로 이어진다. 노출된 흙, 뿌리, 수직 나무줄기와 낮의 녹색 숲은 참조와 유사하나, 병사들이 놓인 주요 사면은 화면 왼쪽에 나타난다.",
        "entities": "서 있는 남성은 두 명이다. 안규홍은 청년 남성의 외형, 묶은 검은 머리, 피와 흙이 묻은 밝은 전통 의복, 등의 피 묻은 천, 소총 한 정과 낫 두 자루를 갖추어 참조의 주요 특징을 따른다. 토미이시는 성인 남성의 외형과 검은 수염, 군모, 갈색 군복, 훈장, 가죽 장화 및 군도를 갖추었다. 작은 측면 얼굴이므로 정확한 얼굴 일치나 국적 자체는 확정하기 어렵다. 쓰러진 병사는 왼쪽 사면, 왼쪽 전경, 중앙 후방에 각각 한 명씩 보인다. 이들은 병사 잔류 조건에 해당하지만 중앙 병사의 엎드린 자세는 참조와 다르다. 읽을 수 있는 글자나 화면 표식은 없다.",
        "hard_violations": [
         "중앙 후방의 병사가 얼굴을 땅 쪽으로 둔 채 엎드려 웅크린 자세로 놓여 있다. 이전 스틸의 누워 있는 병사들의 자세와 접촉점을 고정하라는 조건에 어긋나며, 카메라 이동만으로 설명할 수 없다."
        ],
        "physics": "두 남성의 발은 흙길에 닿아 체중을 지탱한다. 안규홍의 소총과 낫은 손에 잡혀 있고, 소총 멜빵은 중력에 따라 처진다. 토미이시의 군도는 허리 장비에 매달려 지지된다. 쓰러진 병사들은 사면과 지면에 몸을 대고 있어 공중에 뜬 신체나 무기는 없다. 토미이시의 자세는 다소 차렷처럼 경직되어 있으나 물리적으로 불가능하지는 않다."
       },
       {
        "label": "A",
        "direction": "안규홍은 화면 오른쪽 토미이시를, 토미이시는 화면 왼쪽 안규홍을 바라보아 시선이 서로 맞선다. 안규홍의 소총 총구는 오른쪽 아래 흙길을 향하고, 낫 두 자루는 손 아래로 굽어 있다. 토미이시의 군도는 칼집에 들어간 채 뒤쪽 아래로 기울어 있다. 무기를 상대에게 겨누거나 이동하는 동작은 없다.",
        "built_space": "건축물과 인공 고정 설비는 없다. 두 남성이 길 양쪽에 떨어져 서고, 전경에서 두 사람 사이까지 빈 흙길이 이어져 물리적 거리가 중심 소재가 된다. 굵은 나무들이 양옆을 둘러싸며 중앙 간격을 막지 않는다. 오른쪽의 노출된 사면과 그 위 병사들은 참조의 지형 관계에 더 가깝다. 다만 멀어지는 길은 오른쪽 인물 뒤보다 화면 중앙에 놓여, 지정된 퇴로 위치가 정확하지 않다.",
        "entities": "서 있는 인물은 안규홍과 토미이시에 해당하는 남성 두 명이다. 안규홍의 젊은 외형, 묶은 검은 머리, 피 묻은 얼굴과 밝은 전통 의복, 허리띠, 짚신, 등의 혈흔이 참조를 따른다. 손에는 소총 한 정과 낫 두 자루가 보인다. 토미이시는 참조와 유사한 성인 남성의 옆얼굴과 수염, 군모, 훈장 달린 갈색 군복, 가죽 장화와 군도를 갖추었다. 국적은 외형만으로 확정할 수 없다. 병사는 왼쪽 나무 뒤에 부분적으로 한 명, 오른쪽 중경 사면에 두 명, 오른쪽 전경에 부분적으로 한 명이 보인다. 가림 때문에 모든 시신의 정확한 접촉점까지 확인되지는 않지만, 오른쪽 두 병사의 누운 배치는 참조와 유사하다. 민간인 안내자나 온전한 군대 대형, 읽을 수 있는 글자는 없다.",
        "hard_violations": [],
        "physics": "두 남성 모두 두 발로 지면을 딛고 서 있다. 안규홍의 손이 소총과 낫 손잡이를 지지하며, 멜빵은 아래로 늘어진다. 토미이시의 칼집은 허리 장비에 연결되어 있다. 병사들의 몸과 장비는 흙 사면에 놓여 있으며 지지 없이 떠 있는 것은 보이지 않는다. 두 사람의 정지 자세는 가능하지만 토미이시의 곧게 내린 팔과 비교적 가지런한 발은 자연스러운 중단 동작보다 경직되어 보인다."
       }
      ]
     },
     "ok": true
    }
   ],
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "slot_winner_match": true,
   "slot_winner": {
    "gemini-pro": "A",
    "gpt-high": "A"
   },
   "route": "cross_slot_agree"
  },
  "dual": {
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "normalized": {
    "A": 2.0,
    "B": 0.661
   },
   "adjusted": {
    "A": 2.0,
    "B": 0.411
   },
   "violations": {
    "B": [
     "[gemini-pro] physically impossible staging: 좌측 하단의 쓰러진 군인들은 3D 공간의 원근감이나 지형의 경사와 맞지 않게 이전 샷의 2D 이미지를 평면적으로 잘라 붙인 듯한 콜라주 형태를 띰.",
     "[gemini-pro] a person placed where the staging does not put them: 이전 샷에서 안규홍의 전방(우측)에 위치했던 시신들이 샷 구도가 바뀌면서 공간적으로 불가능한 안규홍의 뒤쪽(화면 좌측)으로 배치됨.",
     "[gpt-high] 중앙 후방의 병사가 얼굴을 땅 쪽으로 둔 채 엎드려 웅크린 자세로 놓여 있다. 이전 스틸의 누워 있는 병사들의 자세와 접촉점을 고정하라는 조건에 어긋나며, 카메라 이동만으로 설명할 수 없다."
    ]
   },
   "per_model_winner": {
    "gemini-pro": "A",
    "gpt-high": "A"
   },
   "agreed": true
  },
  "totals": {
   "A": 2000,
   "B": 411
  },
  "selected": "A",
  "ranking": [
   "A",
   "B"
  ],
  "verdicts": [
   {
    "label": "A",
    "score": 2000,
    "verdict_ko": "캐릭터의 외형과 긴장감 있는 대치 구도를 잘 살렸으나, 토미이시 뒤로 이어지는 산길 배경 지시와 이전 샷의 시신 자세 유지를 완벽히 따르지는 못했습니다."
   },
   {
    "label": "B",
    "score": 411,
    "verdict_ko": "토미이시 뒤로 이어지는 길의 구도는 지켰으나, 좌측 하단의 시신들이 원근감 없이 평면적으로 복사되어 붙여진 콜라주 오류와 공간적 위치 오류로 인해 실격입니다.  ★위반: [gemini-pro] physically impossible staging: 좌측 하단의 쓰러진 군인들은 3D 공간의 원근감이나 지형의 경사와 맞지 않게 이전 샷의 2D 이미지를 평면적으로 잘라 붙인 듯한 콜라주 형태를 띰. / [gemini-pro] a person placed where the staging does not put them: 이전 샷에서 안규홍의 전방(우측)에 위치했던 시신들이 샷 구도가 바뀌면서 공간적으로 불가능한 안규홍의 뒤쪽(화면 좌측)으로 배치됨. / [gpt-high] 중앙 후방의 병사가 얼굴을 땅 쪽으로 둔 채 엎드려 웅크린 자세로 놓여 있다. 이전 스틸의 누워 있는 병사들의 자세와 접촉점을 고정하라는 조건에 어긋나며, 카메라 이동만으로 설명할 수 없다."
   }
  ],
  "refs": [
   {
    "label": "PREVIOUS SHOT STILL — a visually related earlier shot of this same place: the location's look, materials, fixed features and lighting mood are LOCKED to this photo; never copy its camera framing. The clothing, hair and overall look of 안규홍 — who appear both in that photo and in this shot — are LOCKED to that photo. Anyone else visible in that photo is NOT in this shot: never carry their face, body or clothing onto anyone here. If this photo shows a character who cannot move (dead or unconscious), that character's exact pose, position and orientation are ALSO LOCKED — treat their body as a fixed prop of the set that only the camera moves around.",
    "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/db00517b-165b-4005-a287-f28936b59b95/images/272bd7af-9294-4490-98c6-88a40e9deacc/scene/recipe/S62sh19_sel.png",
    "asset_id": "3a061761-ee1e-449f-86e1-71794772d320",
    "role": "prev_still"
   },
   {
    "label": "CHARACTER REFERENCE — 안규홍: the exact person appearing in this shot; match face, hair and build exactly.",
    "path": "<bytes:1028384>",
    "asset_id": "3b5894db-ce55-4f34-b966-5e725c3aa952",
    "role": "character_ref"
   },
   {
    "label": "CHARACTER REFERENCE — 토미이시: the exact person appearing in this shot; match face, hair and build exactly.",
    "path": "<bytes:797699>",
    "asset_id": "c861ba0f-8750-4262-8324-25cf5c6c9dcd",
    "role": "character_ref"
   }
  ],
  "critique_skipped": true,
  "shot_run_produced": false,
  "shot_run_uid": "06aa1446-61a3-79a3-a014-4021bdb8a326",
  "ref_mode": "prev+엔티티",
  "share_plan": {
   "ref_plan": "prev",
   "prev_anchor_tag": "S62sh19"
  }
 },
 "S62sh22::cine": {
  "applied": true,
  "attempted_at": "2026-09-09T08:13:35.977579+00:00",
  "fingerprint": "4d46e487c740fb953100e290b6dca1909acaf0d3275a100566b99e8b96d5974f",
  "fingerprint_version": 2,
  "provider": "mai",
  "endpoint": "openrouter/chat-completions",
  "model": "microsoft/mai-image-2.6",
  "pack": "24.202608252115",
  "source_file": "S62sh22_sel.png",
  "source_sha256": "d46f02875cd323aeae3d374b16fdd54ad3561db13cc6d8e57d8b6abe0b380ff3",
  "file": "S62sh22_cine.png",
  "staged_sha256": "e34babf5f6cf9267c39cfba90ded20249feb5a3212052e070f3563f3edfab5dd",
  "latency_ms": 31164
 },
 "S63sh18::signage": {
  "fp": "9d4a3495f59e5ea2",
  "inscriptions": [],
  "cues": [],
  "dropped": []
 },
 "groupbg::camp_duel_slope": {
  "input_fingerprint": "7c6ae42096b6060c",
  "meta": {
   "model": "gpt-image-2",
   "size": "1536x864",
   "pack": "11.202607220237",
   "contract": "bgfirst_full_v3",
   "group_sig": {
    "key": "camp_duel_slope",
    "tags": [
     "S63sh18",
     "S63sh53",
     "S63sh61"
    ]
   },
   "context_sig": "19266244f5830479"
  },
  "prompt": "Create ONE empty live-action location background photograph — NO PEOPLE, no figures, no body parts, no silhouettes, no shadows or reflections of people anywhere.\n\"Empty\" means no people only: KEEP the location's inherent occupants and stock that define the place — animals in an animal shelter, pen or farm, goods and produce in a market, moored boats in a harbour — unless the shot text explicitly removes them.\nThe FIRST attached image is a thin-line storyboard sketch of a shot that happens at this location: use it ONLY as spatial evidence — what this place contains, how its ground, structures and landmarks are arranged and proportioned. Ignore the sketched people and arrows entirely, and do NOT copy its line style: render a fully photographic, physically plausible real place that fits THE LOCATION text below.\nHUMAN-SCALE CALIBRATION: derive every structure's true size from human-scale elements — a door ≈ 2m, a window ≈ 1–1.5m wide, one storey ≈ 2.5–3m; never inflate a small structure or shrink a large one.\n\nTHE LOCATION — early 20th-century Korea, Jeolla region; all people are Korean unless explicitly stated otherwise: The dirt slope at the wooded edge of the mountain camp, below the spot where the fighters lost their weapons.\n\nLOCATION DETAIL (descriptive evidence about this place from the production's location records — evidence, not a staging order): use it to understand what kind of place this is and what it permanently contains. Carry over ONLY its enduring physical features — layout, structures, ground and materials, fixed equipment and installations. Do NOT treat any lighting, weather or time-of-day wording, momentary object states, depicted events or subjective impressions in it as fixed facts of the place; the TIME OF DAY lock below is the sole authority for time and lighting.\n파청 산채 공터·훈련장·취사장·막사 주변: 산 중턱에 통나무를 세워 만든 조잡한 목책 안의 넓은 흙바닥 공터. 나무와 거적으로 만든 움막과 막사가 흩어져 있고, 돌과 진흙으로 만든 취사용 아궁이와 무쇠 솥, 모닥불 터가 있다. 전투 후에는 솥이 뒤집히고 막사가 불타거나 무너진다. (특징: 통나무를 박아 세운 목책; 넓게 다져진 흙바닥 (공터/훈련장); 돌과 진흙으로 만든 다수의 아궁이와 무쇠솥; 나무와 거적으로 엮어 만든 움막과 막사; 부서진 토기 밥그릇과 뒤집힌 무쇠솥 (전투 후); 불에 탄 나무 잔해와 흩어진 쌀밥 (전투 후))\n\nSCENE EVIDENCE (verbatim quotes from the screenplay about this place — treat them as evidence of what the location physically contains and looks like; stage the PLACE those moments happen in, but do NOT depict the momentary actions, people or staged props themselves):\n- 그대로 뒤엉켜 비탈을 구른다.\n- 토미이시가 달려들어 규홍 위에 올라탄다.\n두 손으로 목을 조른다.\n- 규홍은 낫을 쥔 채 그대로 서 있다.\n\nTIME OF DAY (lock): day.\n\nRender ONE photorealistic empty location photograph, 16:9, neutral enough that every shot of this place can be staged from it later. No readable writing anywhere: surfaces that would carry writing may be present, but stage any wording out of legibility — an oblique angle, distance, shallow focus. No captions, watermarks or overlay text.",
  "effective_prompt": "Create ONE empty live-action location background photograph — NO PEOPLE, no figures, no body parts, no silhouettes, no shadows or reflections of people anywhere.\n\"Empty\" means no people only: KEEP the location's inherent occupants and stock that define the place — animals in an animal shelter, pen or farm, goods and produce in a market, moored boats in a harbour — unless the shot text explicitly removes them.\nThe FIRST attached image is a thin-line storyboard sketch of a shot that happens at this location: use it ONLY as spatial evidence — what this place contains, how its ground, structures and landmarks are arranged and proportioned. Ignore the sketched people and arrows entirely, and do NOT copy its line style: render a fully photographic, physically plausible real place that fits THE LOCATION text below.\nHUMAN-SCALE CALIBRATION: derive every structure's true size from human-scale elements — a door ≈ 2m, a window ≈ 1–1.5m wide, one storey ≈ 2.5–3m; never inflate a small structure or shrink a large one.\n\nTHE LOCATION — early 20th-century Korea, Jeolla region; all people are Korean unless explicitly stated otherwise: The dirt slope at the wooded edge of the mountain camp, below the spot where the fighters lost their weapons.\n\nLOCATION DETAIL (descriptive evidence about this place from the production's location records — evidence, not a staging order): use it to understand what kind of place this is and what it permanently contains. Carry over ONLY its enduring physical features — layout, structures, ground and materials, fixed equipment and installations. Do NOT treat any lighting, weather or time-of-day wording, momentary object states, depicted events or subjective impressions in it as fixed facts of the place; the TIME OF DAY lock below is the sole authority for time and lighting.\n파청 산채 공터·훈련장·취사장·막사 주변: 산 중턱에 통나무를 세워 만든 조잡한 목책 안의 넓은 흙바닥 공터. 나무와 거적으로 만든 움막과 막사가 흩어져 있고, 돌과 진흙으로 만든 취사용 아궁이와 무쇠 솥, 모닥불 터가 있다. 전투 후에는 솥이 뒤집히고 막사가 불타거나 무너진다. (특징: 통나무를 박아 세운 목책; 넓게 다져진 흙바닥 (공터/훈련장); 돌과 진흙으로 만든 다수의 아궁이와 무쇠솥; 나무와 거적으로 엮어 만든 움막과 막사; 부서진 토기 밥그릇과 뒤집힌 무쇠솥 (전투 후); 불에 탄 나무 잔해와 흩어진 쌀밥 (전투 후))\n\nSCENE EVIDENCE (verbatim quotes from the screenplay about this place — treat them as evidence of what the location physically contains and looks like; stage the PLACE those moments happen in, but do NOT depict the momentary actions, people or staged props themselves):\n- 그대로 뒤엉켜 비탈을 구른다.\n- 토미이시가 달려들어 규홍 위에 올라탄다.\n두 손으로 목을 조른다.\n- 규홍은 낫을 쥔 채 그대로 서 있다.\n\nTIME OF DAY (lock): day.\n\nRender ONE photorealistic empty location photograph, 16:9, neutral enough that every shot of this place can be staged from it later. No readable writing anywhere: surfaces that would carry writing may be present, but stage any wording out of legibility — an oblique angle, distance, shallow focus. No captions, watermarks or overlay text.",
  "bg_path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/db00517b-165b-4005-a287-f28936b59b95/images/272bd7af-9294-4490-98c6-88a40e9deacc/scene/recipe/groupbg_camp_duel_slope_32846b.png",
  "asset_id": "5f21ff4c-2698-4dd1-b40c-e02f271c372a",
  "input_asset_ids": [
   "2e0c93ad-afb4-454a-b9da-50dde103a91c"
  ],
  "origin_tag": "S63sh18",
  "place_text": "The dirt slope at the wooded edge of the mountain camp, below the spot where the fighters lost their weapons.",
  "origin_inputs": {
   "place_text": "The dirt slope at the wooded edge of the mountain camp, below the spot where the fighters lost their weapons.",
   "time_of_day_en": "day",
   "conti_asset_id": "2e0c93ad-afb4-454a-b9da-50dde103a91c"
  }
 },
 "S63sh18::bgfirst_bg": {
  "input_fingerprint": "234b4e8b52fc8b7b",
  "prompt": "Create the EMPTY BACKGROUND PLATE for one film shot — NO PEOPLE, no figures, no body parts, no sketch lines, no arrows anywhere.\n\"Empty\" means no people only: KEEP the location's inherent occupants and stock that define the place — animals in an animal shelter, pen or farm, goods in a market, moored boats in a harbour — unless the shot text explicitly removes them.\nThe FIRST attached image is a thin-line storyboard sketch: use ONLY its camera angle, horizon, perspective and the placement/size of buildings and set masses — ignore the sketched people and arrows entirely. The SECOND attached image (LOCATION PHOTOGRAPH) is the real place: take its architecture, materials, signage and fixed features, and RE-PROJECT them into the sketch's camera. If the photograph's camera differs from the sketch's, the sketch's camera wins.\nHUMAN-SCALE CALIBRATION: derive every structure's true size from human-scale elements — a door ≈ 2m, a window ≈ 1–1.5m wide, one storey ≈ 2.5–3m; never inflate a small structure or shrink a large one.\n\nSHOT TEXT this background must serve (Korean): 엎어진 안규홍의 등 위에 올라탄 자세로 양손으로 목을 꽉 조르고 있는 토미이시의 핏대 선 얼굴.\n\nLOCATION (lock): The dirt slope at the wooded edge of the mountain camp, below the spot where the fighters lost their weapons.\n\nTIME OF DAY (lock): day.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: close-up\n- KEY BACKGROUND ELEMENTS: Earth beneath the struggle (The men have fallen onto the ground after rolling down the slope); used as A narrow lower-edge strip establishes how little space remains beneath 안규홍.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Natural daylight and sober contrast retain facial strain and the blood on 안규홍's clothing without subjective blur or stylized lighting.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nRender ONE photorealistic empty location photograph, 16:9, that this shot can be staged inside later. No readable writing anywhere: surfaces that would carry writing may be present, but stage any wording out of legibility — an oblique angle, distance, shallow focus. No captions, watermarks or overlay text.",
  "effective_prompt": "Create the EMPTY BACKGROUND PLATE for one film shot — NO PEOPLE, no figures, no body parts, no sketch lines, no arrows anywhere.\n\"Empty\" means no people only: KEEP the location's inherent occupants and stock that define the place — animals in an animal shelter, pen or farm, goods in a market, moored boats in a harbour — unless the shot text explicitly removes them.\nThe FIRST attached image is a thin-line storyboard sketch: use ONLY its camera angle, horizon, perspective and the placement/size of buildings and set masses — ignore the sketched people and arrows entirely. The SECOND attached image (LOCATION PHOTOGRAPH) is the real place: take its architecture, materials, signage and fixed features, and RE-PROJECT them into the sketch's camera. If the photograph's camera differs from the sketch's, the sketch's camera wins.\nHUMAN-SCALE CALIBRATION: derive every structure's true size from human-scale elements — a door ≈ 2m, a window ≈ 1–1.5m wide, one storey ≈ 2.5–3m; never inflate a small structure or shrink a large one.\n\nSHOT TEXT this background must serve (Korean): 엎어진 안규홍의 등 위에 올라탄 자세로 양손으로 목을 꽉 조르고 있는 토미이시의 핏대 선 얼굴.\n\nLOCATION (lock): The dirt slope at the wooded edge of the mountain camp, below the spot where the fighters lost their weapons.\n\nTIME OF DAY (lock): day.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: close-up\n- KEY BACKGROUND ELEMENTS: Earth beneath the struggle (The men have fallen onto the ground after rolling down the slope); used as A narrow lower-edge strip establishes how little space remains beneath 안규홍.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Natural daylight and sober contrast retain facial strain and the blood on 안규홍's clothing without subjective blur or stylized lighting.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nRender ONE photorealistic empty location photograph, 16:9, that this shot can be staged inside later. No readable writing anywhere: surfaces that would carry writing may be present, but stage any wording out of legibility — an oblique angle, distance, shallow focus. No captions, watermarks or overlay text.",
  "bg_path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/db00517b-165b-4005-a287-f28936b59b95/images/272bd7af-9294-4490-98c6-88a40e9deacc/scene/recipe/S63sh18__bgfirst_bg.png",
  "asset_id": "873edbd7-5e0f-43e3-9e9a-d6ee007f4134",
  "input_asset_ids": [
   "2e0c93ad-afb4-454a-b9da-50dde103a91c",
   "5f21ff4c-2698-4dd1-b40c-e02f271c372a"
  ]
 },
 "S63sh18": {
  "input_fingerprint": "8e92910f870f22e4",
  "prompt": "Create ONE FINAL photorealistic live-action film still of the moment below — early 20th-century Korea, Jeolla region; all people are Korean unless explicitly stated otherwise. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 엎어진 안규홍의 등 위에 올라탄 자세로 양손으로 목을 꽉 조르고 있는 토미이시의 핏대 선 얼굴.\n\nLOCATION (lock): The dirt slope at the wooded edge of the mountain camp, below the spot where the fighters lost their weapons. The shot takes place here — the attached LOCATION PHOTOGRAPH shows the exact spot.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: close-up\n- KEY BACKGROUND ELEMENTS: Earth beneath the struggle (The men have fallen onto the ground after rolling down the slope); used as A narrow lower-edge strip establishes how little space remains beneath 안규홍.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Natural daylight and sober contrast retain facial strain and the blood on 안규홍's clothing without subjective blur or stylized lighting.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): A dislodged sickle and the dropped sword lie on the dirt near the slope; the camp ruins and battle debris persist. 안규홍: He is pinned in a choking struggle, with blood spreading afresh through his jeogori from the reopened back wound; his face remains blood- and dirt-covered. 토미이시: He straddles in a choking posture, using both hands and no longer holding his sword.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 안규홍 (한국인 남성, 청년 나이대, 검은 머리, 단단히 묶은 머리); 토미이시 (일본인 남성, 성인 나이대, 자연스러운 검은 머리) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.",
  "roll_prompts": {
   "A": "Stage the shot. The FIRST attached image (SHOT BACKGROUND) is the finished empty background of this shot — keep it EXACTLY: its camera, perspective, architecture, lighting and every fixed feature stay untouched. The SECOND attached image (LAYOUT SKETCH) tells you ONLY where the people go: each sketched person's position, screen size, pose and the gaze/motion arrows. Ignore the sketch's background lines. The CHARACTER REFERENCE photographs show the real people.\nPlace the real people into the background at exactly the sketched positions, sizes and poses, following the arrow directions. No sketch lines or arrows may remain.\n\nCreate ONE FINAL photorealistic live-action film still of the moment below — early 20th-century Korea, Jeolla region; all people are Korean unless explicitly stated otherwise. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 엎어진 안규홍의 등 위에 올라탄 자세로 양손으로 목을 꽉 조르고 있는 토미이시의 핏대 선 얼굴.\n\nLOCATION (lock): The dirt slope at the wooded edge of the mountain camp, below the spot where the fighters lost their weapons. The shot takes place here — the FIRST attached image (SHOT BACKGROUND) is this exact place, already built: its ground, structures, horizon, materials and lighting are the finished truth of this location and must not be redesigned or replaced. No location photograph is attached — read the place from that image alone, and add no scenery, structure, vehicle or fixture that it does not already show. This lock governs the place only; the figures in the shot follow the staging and pose instructions.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: close-up\n- KEY BACKGROUND ELEMENTS: Earth beneath the struggle (The men have fallen onto the ground after rolling down the slope); used as A narrow lower-edge strip establishes how little space remains beneath 안규홍.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Natural daylight and sober contrast retain facial strain and the blood on 안규홍's clothing without subjective blur or stylized lighting.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): A dislodged sickle and the dropped sword lie on the dirt near the slope; the camp ruins and battle debris persist. 안규홍: He is pinned in a choking struggle, with blood spreading afresh through his jeogori from the reopened back wound; his face remains blood- and dirt-covered. 토미이시: He straddles in a choking posture, using both hands and no longer holding his sword.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 안규홍 (한국인 남성, 청년 나이대, 검은 머리, 단단히 묶은 머리); 토미이시 (일본인 남성, 성인 나이대, 자연스러운 검은 머리) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.",
   "B": "Create ONE FINAL photorealistic live-action film still of the moment below — early 20th-century Korea, Jeolla region; all people are Korean unless explicitly stated otherwise. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 엎어진 안규홍의 등 위에 올라탄 자세로 양손으로 목을 꽉 조르고 있는 토미이시의 핏대 선 얼굴.\n\nLOCATION (lock): The dirt slope at the wooded edge of the mountain camp, below the spot where the fighters lost their weapons. The shot takes place here — the attached LOCATION PHOTOGRAPH shows the exact spot.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: close-up\n- KEY BACKGROUND ELEMENTS: Earth beneath the struggle (The men have fallen onto the ground after rolling down the slope); used as A narrow lower-edge strip establishes how little space remains beneath 안규홍.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Natural daylight and sober contrast retain facial strain and the blood on 안규홍's clothing without subjective blur or stylized lighting.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): A dislodged sickle and the dropped sword lie on the dirt near the slope; the camp ruins and battle debris persist. 안규홍: He is pinned in a choking struggle, with blood spreading afresh through his jeogori from the reopened back wound; his face remains blood- and dirt-covered. 토미이시: He straddles in a choking posture, using both hands and no longer holding his sword.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 안규홍 (한국인 남성, 청년 나이대, 검은 머리, 단단히 묶은 머리); 토미이시 (일본인 남성, 성인 나이대, 자연스러운 검은 머리) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.\n\nCOMPOSITION VARIATION (this candidate only): keep the place, the people and the moment exactly as specified, and keep from the CAMERA & FRAME contract its subject scale, screen placement and key background elements exactly. For this candidate only, its camera angle and height are deliberately left open: choose a camera position distinctly different from the obvious one, as a film director picking a second setup on the same blocking. Stay true to any explicit framing the SHOT TEXT itself dictates."
  },
  "roll_refs": {
   "A": [
    {
     "label": "SHOT BACKGROUND",
     "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/db00517b-165b-4005-a287-f28936b59b95/images/272bd7af-9294-4490-98c6-88a40e9deacc/scene/recipe/S63sh18__bgfirst_bg.png",
     "asset_id": "873edbd7-5e0f-43e3-9e9a-d6ee007f4134",
     "role": "bgfirst_bg"
    },
    {
     "label": "LAYOUT SKETCH (people placement only)",
     "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/db00517b-165b-4005-a287-f28936b59b95/images/272bd7af-9294-4490-98c6-88a40e9deacc/conti/conti_S63sh18.png",
     "asset_id": "2e0c93ad-afb4-454a-b9da-50dde103a91c",
     "role": "conti_light"
    },
    {
     "label": "CHARACTER REFERENCE — 안규홍: the exact person appearing in this shot; match face, hair and build exactly.",
     "path": "<bytes:1483645>",
     "asset_id": "b52dff85-b663-44f9-aa9c-46b828990f55",
     "role": "character_ref"
    },
    {
     "label": "CHARACTER REFERENCE — 토미이시: the exact person appearing in this shot; match face, hair and build exactly.",
     "path": "<bytes:797699>",
     "asset_id": "c861ba0f-8750-4262-8324-25cf5c6c9dcd",
     "role": "character_ref"
    }
   ],
   "B": [
    {
     "label": "LOCATION PHOTOGRAPH — the exact place of this shot: its architecture, materials, fixed features and lighting mood are spatial truth; stage the moment inside this place. Never copy its camera framing.",
     "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/db00517b-165b-4005-a287-f28936b59b95/images/272bd7af-9294-4490-98c6-88a40e9deacc/scene/recipe/groupbg_camp_duel_slope_32846b.png",
     "asset_id": "5f21ff4c-2698-4dd1-b40c-e02f271c372a",
     "role": "bgfirst_group_bg"
    },
    {
     "label": "CHARACTER REFERENCE — 안규홍: the exact person appearing in this shot; match face, hair and build exactly.",
     "path": "<bytes:1483645>",
     "asset_id": "b52dff85-b663-44f9-aa9c-46b828990f55",
     "role": "character_ref"
    },
    {
     "label": "CHARACTER REFERENCE — 토미이시: the exact person appearing in this shot; match face, hair and build exactly.",
     "path": "<bytes:797699>",
     "asset_id": "c861ba0f-8750-4262-8324-25cf5c6c9dcd",
     "role": "character_ref"
    }
   ]
  },
  "shot_run_spend_attempt_count": 1,
  "judge_flip_skipped": "judge_fn_owns_order",
  "readings": [
   {
    "label": "A",
    "direction": "토미이시의 시선은 아래에 있는 안규홍을 향함.",
    "built_space": "산채의 흙바닥과 배경의 나무 울타리 및 막사 구조물이 보임.",
    "entities": "토미이시, 안규홍, 바닥의 낫과 검. 안규홍이 엎드리지 않고 하늘/정면을 향해 비스듬히 누워있음.",
    "hard_violations": [
     "[gemini-pro] 자세 위반: 프롬프트에 명시된 '엎어진 안규홍의 등 위'라는 지시와 정반대로 안규홍이 앞/위를 향해 누워있음",
     "[gemini-pro] 프레이밍 위반: 요구된 클로즈업 샷이 아닌 너무 넓은 화면 비율"
    ],
    "physics": "두 사람 모두 지면에 지지되어 있으나, 요구된 엎드린 자세의 물리적 압박감이 아님."
   },
   {
    "label": "B",
    "direction": "토미이시의 시선과 힘이 바닥에 엎드린 안규홍을 향해 집중됨.",
    "built_space": "소나무가 있는 숲 경계의 흙바닥.",
    "entities": "토미이시, 안규홍, 바닥의 낫과 꺾인 검. 안규홍이 완전히 엎드려 있고 토미이시가 그 등 위에 올라타 있음.",
    "hard_violations": [
     "[gpt-high] 지정된 낫 한 자루와 떨어진 검 한 자루 외에 전경 하단에 별도의 큰 칼날이 추가되어 있다."
    ],
    "physics": "안규홍은 흙바닥에 완전히 밀착되어 엎드려 있고, 토미이시의 체중이 그 위에 물리적으로 실려 있음."
   }
  ],
  "cross_model_order": {
   "policy": "cross_model_order_v2_slot0_fwd_slot1_rev_parallel2",
   "slots": [
    {
     "model": "gemini-pro",
     "order": "forward",
     "display_to_canonical": {
      "A": "A",
      "B": "B"
     },
     "raw": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 2,
        "verdict_ko": "프롬프트의 가장 핵심인 '엎드린 안규홍의 등 위'라는 자세를 위반하여 안규홍을 정면으로 눕혔고, 요구된 클로즈업 샷 크기도 지키지 않았습니다."
       },
       {
        "label": "B",
        "score": 9,
        "verdict_ko": "'엎어진 안규홍의 등 위에 올라타 목을 조르는' 핵심 자세와 토미이시의 핏대 선 표정을 요구된 클로즈업 구도로 매우 정확하게 구현했습니다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "토미이시의 시선은 아래에 있는 안규홍을 향함.",
        "built_space": "산채의 흙바닥과 배경의 나무 울타리 및 막사 구조물이 보임.",
        "entities": "토미이시, 안규홍, 바닥의 낫과 검. 안규홍이 엎드리지 않고 하늘/정면을 향해 비스듬히 누워있음.",
        "hard_violations": [
         "자세 위반: 프롬프트에 명시된 '엎어진 안규홍의 등 위'라는 지시와 정반대로 안규홍이 앞/위를 향해 누워있음",
         "프레이밍 위반: 요구된 클로즈업 샷이 아닌 너무 넓은 화면 비율"
        ],
        "physics": "두 사람 모두 지면에 지지되어 있으나, 요구된 엎드린 자세의 물리적 압박감이 아님."
       },
       {
        "label": "B",
        "direction": "토미이시의 시선과 힘이 바닥에 엎드린 안규홍을 향해 집중됨.",
        "built_space": "소나무가 있는 숲 경계의 흙바닥.",
        "entities": "토미이시, 안규홍, 바닥의 낫과 꺾인 검. 안규홍이 완전히 엎드려 있고 토미이시가 그 등 위에 올라타 있음.",
        "hard_violations": [],
        "physics": "안규홍은 흙바닥에 완전히 밀착되어 엎드려 있고, 토미이시의 체중이 그 위에 물리적으로 실려 있음."
       }
      ],
      "all_candidates_fail": false
     },
     "normalized": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 2,
        "verdict_ko": "프롬프트의 가장 핵심인 '엎드린 안규홍의 등 위'라는 자세를 위반하여 안규홍을 정면으로 눕혔고, 요구된 클로즈업 샷 크기도 지키지 않았습니다."
       },
       {
        "label": "B",
        "score": 9,
        "verdict_ko": "'엎어진 안규홍의 등 위에 올라타 목을 조르는' 핵심 자세와 토미이시의 핏대 선 표정을 요구된 클로즈업 구도로 매우 정확하게 구현했습니다."
       }
      ],
      "all_candidates_fail": false,
      "readings": [
       {
        "label": "A",
        "direction": "토미이시의 시선은 아래에 있는 안규홍을 향함.",
        "built_space": "산채의 흙바닥과 배경의 나무 울타리 및 막사 구조물이 보임.",
        "entities": "토미이시, 안규홍, 바닥의 낫과 검. 안규홍이 엎드리지 않고 하늘/정면을 향해 비스듬히 누워있음.",
        "hard_violations": [
         "자세 위반: 프롬프트에 명시된 '엎어진 안규홍의 등 위'라는 지시와 정반대로 안규홍이 앞/위를 향해 누워있음",
         "프레이밍 위반: 요구된 클로즈업 샷이 아닌 너무 넓은 화면 비율"
        ],
        "physics": "두 사람 모두 지면에 지지되어 있으나, 요구된 엎드린 자세의 물리적 압박감이 아님."
       },
       {
        "label": "B",
        "direction": "토미이시의 시선과 힘이 바닥에 엎드린 안규홍을 향해 집중됨.",
        "built_space": "소나무가 있는 숲 경계의 흙바닥.",
        "entities": "토미이시, 안규홍, 바닥의 낫과 꺾인 검. 안규홍이 완전히 엎드려 있고 토미이시가 그 등 위에 올라타 있음.",
        "hard_violations": [],
        "physics": "안규홍은 흙바닥에 완전히 밀착되어 엎드려 있고, 토미이시의 체중이 그 위에 물리적으로 실려 있음."
       }
      ]
     },
     "ok": true
    },
    {
     "model": "gpt-high",
     "order": "reverse",
     "display_to_canonical": {
      "A": "B",
      "B": "A"
     },
     "raw": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 3,
        "verdict_ko": "토미이시의 힘주는 얼굴은 더 가깝지만, 추가 칼날이 등장하고 안규홍이 옆으로 돌아누워 있어 엎어진 등 위에서 양손으로 조르는 지정 동작과 어긋납니다."
       },
       {
        "label": "B",
        "score": 6,
        "verdict_ko": "엎어진 안규홍을 올라타 양손으로 목 조르는 관계와 장소는 맞지만, 얼굴 클로즈업 대신 무릎과 넓은 야영지까지 보여 주어 핵심 프레이밍을 놓쳤습니다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "토미이시는 아래쪽 안규홍의 머리와 목을 향해 얼굴을 숙이고 이를 악문다. 안규홍은 땅 가까이에서 화면 오른쪽 아래를 본다. 목 앞에 손이 모여 있지만, 두 손으로 목 양옆을 압박하기보다는 팔을 감아 손을 맞잡은 조르기에 가깝다. 바닥의 무기들은 누구를 겨누고 있지 않다.",
        "built_space": "뒤로 나무가 선 흙비탈이 보이며 인공 시설은 식별되지 않는다. 숲 가장자리라는 환경은 맞지만, 참조 야영지의 정확한 지점인지는 이 크롭으로 확인하기 어렵다. 흙과 돌이 하단의 좁은 띠를 넘어 오른쪽과 전경을 넓게 차지한다. 토미이시의 얼굴은 크게 잡혔지만 두 사람의 몸과 무기까지 상당히 포함된다.",
        "entities": "성인 동아시아계 남성 두 명만 보인다. 토미이시의 짧은 검은 머리, 수염, 갈색 군복과 가죽띠는 참조와 대체로 맞고 모자는 없다. 안규홍의 검은 머리, 젊은 얼굴, 때 묻은 흰 저고리는 부합하지만 묶은 머리의 매듭은 명확하지 않다. 얼굴과 손에는 피가 있으나 보이는 저고리에서 새로 번지는 등 출혈은 뚜렷하지 않다. 오른쪽에 낫 한 자루, 하단에 손잡이와 코등이가 있는 검 한 자루가 있고, 그 앞에 별도의 큰 칼날이 추가로 보인다. 읽을 수 있는 글자는 없다.",
        "hard_violations": [
         "지정된 낫 한 자루와 떨어진 검 한 자루 외에 전경 하단에 별도의 큰 칼날이 추가되어 있다."
        ],
        "physics": "안규홍은 얼굴 옆면과 옆으로 기운 몸통을 흙바닥에 대고 있으며, 토미이시는 그 몸 위에 밀착해 체중을 싣는다. 공중에 지지 없이 떠 있는 인체는 없다. 다만 안규홍의 가슴이 카메라 쪽으로 돌아가고 한쪽 팔이 위로 들려 있어, 엎어진 등을 올라탄 자세보다는 옆으로 누운 상대를 붙든 자세로 읽힌다. 무기들은 흙바닥에 놓여 있다."
       },
       {
        "label": "B",
        "direction": "토미이시의 시선과 숙인 얼굴은 바로 아래 안규홍의 목과 머리로 향한다. 두 손은 안규홍의 목 양옆을 감싸 압박하고, 안규홍은 아래쪽 땅을 보며 양손으로 공격자의 손목 부근을 붙든다. 떨어진 낫과 검은 바닥에 놓여 공격 대상을 향해 사용되지 않는다.",
        "built_space": "왼쪽의 무너진 움막 일부와 뒤쪽 목책, 오른쪽의 나무 지지틀 한 조와 돌무더기, 큰 검은 솥 하나와 작은 검은 용기 하나, 숲비탈과 먼 산이 보인다. 재료와 배치는 참조 장소에 잘 부합하며 명백한 시설 중복은 없다. 그러나 두 사람의 상체뿐 아니라 토미이시의 무릎과 부츠, 넓은 바닥과 배경 시설까지 담아 얼굴 클로즈업이 아닌 중간 거리의 투쟁 장면이 되었다.",
        "entities": "참조 인물에 대응하는 성인 동아시아계 남성 두 명만 보인다. 토미이시는 짧은 검은 머리와 수염, 갈색 군복, 붉은 계급장과 가죽띠를 갖췄으며 모자는 없다. 안규홍은 젊은 얼굴, 뒤로 모은 검은 머리, 피와 흙이 묻은 얼굴 및 흰 저고리로 참조 특징을 따른다. 저고리의 어깨와 등 윗부분에 혈흔이 보인다. 국적은 외형만으로 확정할 수 없다. 오른쪽 흙바닥에는 낫 한 자루와 검 한 자루가 있으며 토미이시는 검을 들고 있지 않다. 읽을 수 있는 글자는 없다.",
        "hard_violations": [],
        "physics": "안규홍의 배와 하체는 바닥에 놓이고 양팔꿈치와 전완이 땅을 짚어 상체를 조금 받친다. 토미이시는 양 무릎을 상대 몸 양옆의 땅에 대고 등 위를 걸터앉아 체중을 싣는다. 양팔에서 목을 잡은 손까지 연결이 자연스럽고, 안규홍의 손도 공격자의 손목과 접촉한다. 인체와 무기 모두 지지면이 확인되며 떠 있는 대상은 없다."
       }
      ],
      "all_candidates_fail": true
     },
     "normalized": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 3,
        "verdict_ko": "토미이시의 힘주는 얼굴은 더 가깝지만, 추가 칼날이 등장하고 안규홍이 옆으로 돌아누워 있어 엎어진 등 위에서 양손으로 조르는 지정 동작과 어긋납니다."
       },
       {
        "label": "A",
        "score": 6,
        "verdict_ko": "엎어진 안규홍을 올라타 양손으로 목 조르는 관계와 장소는 맞지만, 얼굴 클로즈업 대신 무릎과 넓은 야영지까지 보여 주어 핵심 프레이밍을 놓쳤습니다."
       }
      ],
      "all_candidates_fail": true,
      "readings": [
       {
        "label": "B",
        "direction": "토미이시는 아래쪽 안규홍의 머리와 목을 향해 얼굴을 숙이고 이를 악문다. 안규홍은 땅 가까이에서 화면 오른쪽 아래를 본다. 목 앞에 손이 모여 있지만, 두 손으로 목 양옆을 압박하기보다는 팔을 감아 손을 맞잡은 조르기에 가깝다. 바닥의 무기들은 누구를 겨누고 있지 않다.",
        "built_space": "뒤로 나무가 선 흙비탈이 보이며 인공 시설은 식별되지 않는다. 숲 가장자리라는 환경은 맞지만, 참조 야영지의 정확한 지점인지는 이 크롭으로 확인하기 어렵다. 흙과 돌이 하단의 좁은 띠를 넘어 오른쪽과 전경을 넓게 차지한다. 토미이시의 얼굴은 크게 잡혔지만 두 사람의 몸과 무기까지 상당히 포함된다.",
        "entities": "성인 동아시아계 남성 두 명만 보인다. 토미이시의 짧은 검은 머리, 수염, 갈색 군복과 가죽띠는 참조와 대체로 맞고 모자는 없다. 안규홍의 검은 머리, 젊은 얼굴, 때 묻은 흰 저고리는 부합하지만 묶은 머리의 매듭은 명확하지 않다. 얼굴과 손에는 피가 있으나 보이는 저고리에서 새로 번지는 등 출혈은 뚜렷하지 않다. 오른쪽에 낫 한 자루, 하단에 손잡이와 코등이가 있는 검 한 자루가 있고, 그 앞에 별도의 큰 칼날이 추가로 보인다. 읽을 수 있는 글자는 없다.",
        "hard_violations": [
         "지정된 낫 한 자루와 떨어진 검 한 자루 외에 전경 하단에 별도의 큰 칼날이 추가되어 있다."
        ],
        "physics": "안규홍은 얼굴 옆면과 옆으로 기운 몸통을 흙바닥에 대고 있으며, 토미이시는 그 몸 위에 밀착해 체중을 싣는다. 공중에 지지 없이 떠 있는 인체는 없다. 다만 안규홍의 가슴이 카메라 쪽으로 돌아가고 한쪽 팔이 위로 들려 있어, 엎어진 등을 올라탄 자세보다는 옆으로 누운 상대를 붙든 자세로 읽힌다. 무기들은 흙바닥에 놓여 있다."
       },
       {
        "label": "A",
        "direction": "토미이시의 시선과 숙인 얼굴은 바로 아래 안규홍의 목과 머리로 향한다. 두 손은 안규홍의 목 양옆을 감싸 압박하고, 안규홍은 아래쪽 땅을 보며 양손으로 공격자의 손목 부근을 붙든다. 떨어진 낫과 검은 바닥에 놓여 공격 대상을 향해 사용되지 않는다.",
        "built_space": "왼쪽의 무너진 움막 일부와 뒤쪽 목책, 오른쪽의 나무 지지틀 한 조와 돌무더기, 큰 검은 솥 하나와 작은 검은 용기 하나, 숲비탈과 먼 산이 보인다. 재료와 배치는 참조 장소에 잘 부합하며 명백한 시설 중복은 없다. 그러나 두 사람의 상체뿐 아니라 토미이시의 무릎과 부츠, 넓은 바닥과 배경 시설까지 담아 얼굴 클로즈업이 아닌 중간 거리의 투쟁 장면이 되었다.",
        "entities": "참조 인물에 대응하는 성인 동아시아계 남성 두 명만 보인다. 토미이시는 짧은 검은 머리와 수염, 갈색 군복, 붉은 계급장과 가죽띠를 갖췄으며 모자는 없다. 안규홍은 젊은 얼굴, 뒤로 모은 검은 머리, 피와 흙이 묻은 얼굴 및 흰 저고리로 참조 특징을 따른다. 저고리의 어깨와 등 윗부분에 혈흔이 보인다. 국적은 외형만으로 확정할 수 없다. 오른쪽 흙바닥에는 낫 한 자루와 검 한 자루가 있으며 토미이시는 검을 들고 있지 않다. 읽을 수 있는 글자는 없다.",
        "hard_violations": [],
        "physics": "안규홍의 배와 하체는 바닥에 놓이고 양팔꿈치와 전완이 땅을 짚어 상체를 조금 받친다. 토미이시는 양 무릎을 상대 몸 양옆의 땅에 대고 등 위를 걸터앉아 체중을 싣는다. 양팔에서 목을 잡은 손까지 연결이 자연스럽고, 안규홍의 손도 공격자의 손목과 접촉한다. 인체와 무기 모두 지지면이 확인되며 떠 있는 대상은 없다."
       }
      ]
     },
     "ok": true
    }
   ],
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "slot_winner_match": false,
   "slot_winner": {
    "gemini-pro": "B",
    "gpt-high": "A"
   },
   "route": "cross_slot_combined"
  },
  "dual": {
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "normalized": {
    "A": 1.222,
    "B": 1.5
   },
   "adjusted": {
    "A": 0.972,
    "B": 1.25
   },
   "violations": {
    "A": [
     "[gemini-pro] 자세 위반: 프롬프트에 명시된 '엎어진 안규홍의 등 위'라는 지시와 정반대로 안규홍이 앞/위를 향해 누워있음",
     "[gemini-pro] 프레이밍 위반: 요구된 클로즈업 샷이 아닌 너무 넓은 화면 비율"
    ],
    "B": [
     "[gpt-high] 지정된 낫 한 자루와 떨어진 검 한 자루 외에 전경 하단에 별도의 큰 칼날이 추가되어 있다."
    ]
   },
   "per_model_winner": {
    "gemini-pro": "B",
    "gpt-high": "A"
   },
   "agreed": false
  },
  "totals": {
   "A": 972,
   "B": 1250
  },
  "selected": "B",
  "ranking": [
   "B",
   "A"
  ],
  "verdicts": [
   {
    "label": "A",
    "score": 972,
    "verdict_ko": "프롬프트의 가장 핵심인 '엎드린 안규홍의 등 위'라는 자세를 위반하여 안규홍을 정면으로 눕혔고, 요구된 클로즈업 샷 크기도 지키지 않았습니다.  ★위반: [gemini-pro] 자세 위반: 프롬프트에 명시된 '엎어진 안규홍의 등 위'라는 지시와 정반대로 안규홍이 앞/위를 향해 누워있음 / [gemini-pro] 프레이밍 위반: 요구된 클로즈업 샷이 아닌 너무 넓은 화면 비율"
   },
   {
    "label": "B",
    "score": 1250,
    "verdict_ko": "'엎어진 안규홍의 등 위에 올라타 목을 조르는' 핵심 자세와 토미이시의 핏대 선 표정을 요구된 클로즈업 구도로 매우 정확하게 구현했습니다.  ★위반: [gpt-high] 지정된 낫 한 자루와 떨어진 검 한 자루 외에 전경 하단에 별도의 큰 칼날이 추가되어 있다."
   }
  ],
  "refs": [
   {
    "label": "LOCATION PHOTOGRAPH — the exact place of this shot: its architecture, materials, fixed features and lighting mood are spatial truth; stage the moment inside this place. Never copy its camera framing.",
    "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/db00517b-165b-4005-a287-f28936b59b95/images/272bd7af-9294-4490-98c6-88a40e9deacc/scene/recipe/groupbg_camp_duel_slope_32846b.png",
    "asset_id": "5f21ff4c-2698-4dd1-b40c-e02f271c372a",
    "role": "bgfirst_group_bg"
   },
   {
    "label": "CHARACTER REFERENCE — 안규홍: the exact person appearing in this shot; match face, hair and build exactly.",
    "path": "<bytes:1483645>",
    "asset_id": "b52dff85-b663-44f9-aa9c-46b828990f55",
    "role": "character_ref"
   },
   {
    "label": "CHARACTER REFERENCE — 토미이시: the exact person appearing in this shot; match face, hair and build exactly.",
    "path": "<bytes:797699>",
    "asset_id": "c861ba0f-8750-4262-8324-25cf5c6c9dcd",
    "role": "character_ref"
   }
  ],
  "critique_skipped": true,
  "shot_run_produced": false,
  "shot_run_uid": "06aa1446-6353-72f8-9484-0664dc859704",
  "bgfirst": {
   "bg_path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/db00517b-165b-4005-a287-f28936b59b95/images/272bd7af-9294-4490-98c6-88a40e9deacc/scene/recipe/S63sh18__bgfirst_bg.png",
   "bg_asset_id": "873edbd7-5e0f-43e3-9e9a-d6ee007f4134",
   "bg_record_key": "S63sh18::bgfirst_bg",
   "chain_winner": false,
   "authority": "groupbg",
   "group_key": "camp_duel_slope",
   "groupbg_asset_id": "5f21ff4c-2698-4dd1-b40c-e02f271c372a"
  },
  "ref_mode": "그룹 배경+엔티티 (2택1: 무콘티 승)",
  "share_plan": {
   "ref_plan": "background"
  }
 },
 "S63sh18::cine": {
  "applied": true,
  "attempted_at": "2026-09-09T08:19:22.680103+00:00",
  "fingerprint": "8f540e9785c917c9918131c96961e06ac6d7024e48b01fb84100c520ff8a178c",
  "fingerprint_version": 2,
  "provider": "mai",
  "endpoint": "openrouter/chat-completions",
  "model": "microsoft/mai-image-2.6",
  "pack": "24.202608252115",
  "source_file": "S63sh18_sel.png",
  "source_sha256": "3c198145746fb3147c3694426d197411950328b2542cab5b83795e04b5f5202c",
  "file": "S63sh18_cine.png",
  "staged_sha256": "804432475ec2c4a1ba7164f7843c55eca0110158f325bdf7b1d5a8235ae6b6bd",
  "latency_ms": 24744
 },
 "S63sh53::signage": {
  "fp": "905c5ae0be5889bb",
  "inscriptions": [],
  "cues": [],
  "dropped": []
 },
 "S63sh53": {
  "input_fingerprint": "c21c2d3fe79d4608",
  "prompt": "Create ONE FINAL photorealistic live-action film still of the moment below — early 20th-century Korea, Jeolla region; all people are Korean unless explicitly stated otherwise. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 돌진하는 힘이 실린 안규홍의 낫이 토미이시의 몸통에 깊숙이 꽂혀 타격 부위가 움푹 들어간 순간.\n\nLOCATION (lock): The open dirt fighting patch on the mountain camp's wooded slope, among the weapons dropped during the struggle. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: close-up\n- KEY BACKGROUND ELEMENTS: Sickle (Embedded in 토미이시's torso while 안규홍 maintains his grip) — The handle crosses obliquely from the lower-left hands toward the center-right contact; used as Connects the forward force of the hands to the arrested torso, occupying only a small portion of the image; Earth ground (Visible beyond the joined bodies); used as A soft peripheral patch retains the ground-level setting and bodily scale.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Maintain natural daylight, muted saturation and controlled contrast, presenting the impact with blunt physical clarity rather than spectacle.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The clothing, hair and overall look of 토미이시 — who appear both in that photo and in this shot — are LOCKED to that photo. Anyone else visible in that photo is NOT in this shot: never carry their face, body or clothing onto anyone here. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The dirt bears a deep sword puncture; the surrounding camp wreckage remains unchanged. 토미이시: He has dirt across his face, a deep torn flank wound and a second penetrating sickle wound; his sword grip is failing.\n\nTHE HAND THAT IS DOING THIS: the object at the centre of this shot is being held, operated, read, aimed or handed over by 안규홍 right now, so 안규홍's hand — and as much of the wrist and forearm as the framing reaches — is in the frame, gripping or resting on that object exactly the way its real use requires. Match that hand to 안규홍: its size, build, skin, age, grooming, sleeve and anything worn on it belong to that person and to no one else.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 토미이시 (일본인 남성, 성인 나이대, 자연스러운 검은 머리) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.",
  "roll_prompts": {
   "A": "Create ONE FINAL photorealistic live-action film still of the moment below — early 20th-century Korea, Jeolla region; all people are Korean unless explicitly stated otherwise. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 돌진하는 힘이 실린 안규홍의 낫이 토미이시의 몸통에 깊숙이 꽂혀 타격 부위가 움푹 들어간 순간.\n\nLOCATION (lock): The open dirt fighting patch on the mountain camp's wooded slope, among the weapons dropped during the struggle. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: close-up\n- KEY BACKGROUND ELEMENTS: Sickle (Embedded in 토미이시's torso while 안규홍 maintains his grip) — The handle crosses obliquely from the lower-left hands toward the center-right contact; used as Connects the forward force of the hands to the arrested torso, occupying only a small portion of the image; Earth ground (Visible beyond the joined bodies); used as A soft peripheral patch retains the ground-level setting and bodily scale.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Maintain natural daylight, muted saturation and controlled contrast, presenting the impact with blunt physical clarity rather than spectacle.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The clothing, hair and overall look of 토미이시 — who appear both in that photo and in this shot — are LOCKED to that photo. Anyone else visible in that photo is NOT in this shot: never carry their face, body or clothing onto anyone here. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The dirt bears a deep sword puncture; the surrounding camp wreckage remains unchanged. 토미이시: He has dirt across his face, a deep torn flank wound and a second penetrating sickle wound; his sword grip is failing.\n\nTHE HAND THAT IS DOING THIS: the object at the centre of this shot is being held, operated, read, aimed or handed over by 안규홍 right now, so 안규홍's hand — and as much of the wrist and forearm as the framing reaches — is in the frame, gripping or resting on that object exactly the way its real use requires. Match that hand to 안규홍: its size, build, skin, age, grooming, sleeve and anything worn on it belong to that person and to no one else.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 토미이시 (일본인 남성, 성인 나이대, 자연스러운 검은 머리) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.",
   "B": "Create ONE FINAL photorealistic live-action film still of the moment below — early 20th-century Korea, Jeolla region; all people are Korean unless explicitly stated otherwise. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 돌진하는 힘이 실린 안규홍의 낫이 토미이시의 몸통에 깊숙이 꽂혀 타격 부위가 움푹 들어간 순간.\n\nLOCATION (lock): The open dirt fighting patch on the mountain camp's wooded slope, among the weapons dropped during the struggle. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: close-up\n- KEY BACKGROUND ELEMENTS: Sickle (Embedded in 토미이시's torso while 안규홍 maintains his grip) — The handle crosses obliquely from the lower-left hands toward the center-right contact; used as Connects the forward force of the hands to the arrested torso, occupying only a small portion of the image; Earth ground (Visible beyond the joined bodies); used as A soft peripheral patch retains the ground-level setting and bodily scale.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Maintain natural daylight, muted saturation and controlled contrast, presenting the impact with blunt physical clarity rather than spectacle.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The clothing, hair and overall look of 토미이시 — who appear both in that photo and in this shot — are LOCKED to that photo. Anyone else visible in that photo is NOT in this shot: never carry their face, body or clothing onto anyone here. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The dirt bears a deep sword puncture; the surrounding camp wreckage remains unchanged. 토미이시: He has dirt across his face, a deep torn flank wound and a second penetrating sickle wound; his sword grip is failing.\n\nTHE HAND THAT IS DOING THIS: the object at the centre of this shot is being held, operated, read, aimed or handed over by 안규홍 right now, so 안규홍's hand — and as much of the wrist and forearm as the framing reaches — is in the frame, gripping or resting on that object exactly the way its real use requires. Match that hand to 안규홍: its size, build, skin, age, grooming, sleeve and anything worn on it belong to that person and to no one else.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 토미이시 (일본인 남성, 성인 나이대, 자연스러운 검은 머리) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.\n\nCOMPOSITION VARIATION (this candidate only): keep the place, the people and the moment exactly as specified, and keep from the CAMERA & FRAME contract its subject scale, screen placement and key background elements exactly. For this candidate only, its camera angle and height are deliberately left open: choose a camera position distinctly different from the obvious one, as a film director picking a second setup on the same blocking. Stay true to any explicit framing the SHOT TEXT itself dictates."
  },
  "shot_run_spend_attempt_count": 1,
  "initial_roll_all_fail": true,
  "readings": [
   {
    "label": "A",
    "direction": "안규홍의 손이 낫을 쥐고 있으나, 타겟인 토미이시의 몸통(torso)을 향하지 않고 턱과 목 부위에 닿아 있으며 전혀 꽂혀 있지 않습니다.",
    "built_space": "산속 야영지의 흙바닥 비탈길이며, 배경에 버려진 칼과 낫 등의 무기들이 흩어져 있어 장소 설정과 일치합니다.",
    "entities": "토미이시의 얼굴과 피투성이 군복은 레퍼런스와 일치하며(모자 없음), 안규홍의 팔은 흰옷 소매를 입고 있습니다. 낫 소품의 형태도 지시된 레퍼런스와 유사합니다.",
    "hard_violations": [],
    "physics": "낫은 안규홍의 손에 지탱되고 있으나, 타격으로 꽂힌 상태가 아니라 뒷날이 목에 얹혀 있는 잘못된 자세입니다. 토미이시의 몸은 바닥에 누워 지탱되고 있습니다."
   },
   {
    "label": "B",
    "direction": "안규홍의 두 손이 낫을 쥐고 프롬프트가 지시한 타겟인 토미이시의 하복부(몸통)를 향해 조준하여 꽂아 넣고 있습니다.",
    "built_space": "흙바닥 비탈길과 나무가 있는 산속 배경으로, 버려진 칼과 낫 등의 무기들이 바닥에 자연스럽게 흩어져 있습니다.",
    "entities": "낫과 안규홍의 손이 묘사되었으나, 토미이시가 레퍼런스에 없는 일본군 모자를 쓰고 있습니다. 화면 왼쪽 가장자리에 안규홍으로 추정되는 인물의 얼굴 일부가 노출되었습니다.",
    "hard_violations": [
     "[gemini-pro] Invented object: 텍스트와 레퍼런스에 의해 외형이 엄격히 고정(LOCKED)된 토미이시에게 레퍼런스에 없는 군모(모자)를 씌움."
    ],
    "physics": "낫은 두 손에 의해 강한 힘으로 쥐여 몸통에 단단히 꽂힌 채 지탱되고 있으며, 토미이시는 무릎을 꿇고 지면에 자세를 유지하고 있습니다."
   }
  ],
  "cross_model_order": {
   "policy": "cross_model_order_v2_slot0_fwd_slot1_rev_parallel2",
   "slots": [
    {
     "model": "gemini-pro",
     "order": "forward",
     "display_to_canonical": {
      "A": "A",
      "B": "B"
     },
     "raw": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 3,
        "verdict_ko": "클로즈업 프레이밍과 캐릭터 외형(모자 없음)은 지시를 따랐으나, 낫이 몸통에 꽂히는 핵심 액션을 묘사하지 못하고 엉뚱하게 목 부위에 얹혀 있어 방향성 및 연출 지시를 크게 실패했습니다."
       },
       {
        "label": "B",
        "score": 1,
        "verdict_ko": "낫이 몸통에 꽂히는 텍스트의 핵심 액션과 구도는 잘 구현했으나, 외형이 고정(LOCKED)된 캐릭터에게 레퍼런스에 없는 모자를 씌운 치명적 오류(Hard Violation)가 있으며 프레이밍이 너무 넓습니다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "안규홍의 손이 낫을 쥐고 있으나, 타겟인 토미이시의 몸통(torso)을 향하지 않고 턱과 목 부위에 닿아 있으며 전혀 꽂혀 있지 않습니다.",
        "built_space": "산속 야영지의 흙바닥 비탈길이며, 배경에 버려진 칼과 낫 등의 무기들이 흩어져 있어 장소 설정과 일치합니다.",
        "entities": "토미이시의 얼굴과 피투성이 군복은 레퍼런스와 일치하며(모자 없음), 안규홍의 팔은 흰옷 소매를 입고 있습니다. 낫 소품의 형태도 지시된 레퍼런스와 유사합니다.",
        "hard_violations": [],
        "physics": "낫은 안규홍의 손에 지탱되고 있으나, 타격으로 꽂힌 상태가 아니라 뒷날이 목에 얹혀 있는 잘못된 자세입니다. 토미이시의 몸은 바닥에 누워 지탱되고 있습니다."
       },
       {
        "label": "B",
        "direction": "안규홍의 두 손이 낫을 쥐고 프롬프트가 지시한 타겟인 토미이시의 하복부(몸통)를 향해 조준하여 꽂아 넣고 있습니다.",
        "built_space": "흙바닥 비탈길과 나무가 있는 산속 배경으로, 버려진 칼과 낫 등의 무기들이 바닥에 자연스럽게 흩어져 있습니다.",
        "entities": "낫과 안규홍의 손이 묘사되었으나, 토미이시가 레퍼런스에 없는 일본군 모자를 쓰고 있습니다. 화면 왼쪽 가장자리에 안규홍으로 추정되는 인물의 얼굴 일부가 노출되었습니다.",
        "hard_violations": [
         "Invented object: 텍스트와 레퍼런스에 의해 외형이 엄격히 고정(LOCKED)된 토미이시에게 레퍼런스에 없는 군모(모자)를 씌움."
        ],
        "physics": "낫은 두 손에 의해 강한 힘으로 쥐여 몸통에 단단히 꽂힌 채 지탱되고 있으며, 토미이시는 무릎을 꿇고 지면에 자세를 유지하고 있습니다."
       }
      ],
      "all_candidates_fail": true
     },
     "normalized": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 3,
        "verdict_ko": "클로즈업 프레이밍과 캐릭터 외형(모자 없음)은 지시를 따랐으나, 낫이 몸통에 꽂히는 핵심 액션을 묘사하지 못하고 엉뚱하게 목 부위에 얹혀 있어 방향성 및 연출 지시를 크게 실패했습니다."
       },
       {
        "label": "B",
        "score": 1,
        "verdict_ko": "낫이 몸통에 꽂히는 텍스트의 핵심 액션과 구도는 잘 구현했으나, 외형이 고정(LOCKED)된 캐릭터에게 레퍼런스에 없는 모자를 씌운 치명적 오류(Hard Violation)가 있으며 프레이밍이 너무 넓습니다."
       }
      ],
      "all_candidates_fail": true,
      "readings": [
       {
        "label": "A",
        "direction": "안규홍의 손이 낫을 쥐고 있으나, 타겟인 토미이시의 몸통(torso)을 향하지 않고 턱과 목 부위에 닿아 있으며 전혀 꽂혀 있지 않습니다.",
        "built_space": "산속 야영지의 흙바닥 비탈길이며, 배경에 버려진 칼과 낫 등의 무기들이 흩어져 있어 장소 설정과 일치합니다.",
        "entities": "토미이시의 얼굴과 피투성이 군복은 레퍼런스와 일치하며(모자 없음), 안규홍의 팔은 흰옷 소매를 입고 있습니다. 낫 소품의 형태도 지시된 레퍼런스와 유사합니다.",
        "hard_violations": [],
        "physics": "낫은 안규홍의 손에 지탱되고 있으나, 타격으로 꽂힌 상태가 아니라 뒷날이 목에 얹혀 있는 잘못된 자세입니다. 토미이시의 몸은 바닥에 누워 지탱되고 있습니다."
       },
       {
        "label": "B",
        "direction": "안규홍의 두 손이 낫을 쥐고 프롬프트가 지시한 타겟인 토미이시의 하복부(몸통)를 향해 조준하여 꽂아 넣고 있습니다.",
        "built_space": "흙바닥 비탈길과 나무가 있는 산속 배경으로, 버려진 칼과 낫 등의 무기들이 바닥에 자연스럽게 흩어져 있습니다.",
        "entities": "낫과 안규홍의 손이 묘사되었으나, 토미이시가 레퍼런스에 없는 일본군 모자를 쓰고 있습니다. 화면 왼쪽 가장자리에 안규홍으로 추정되는 인물의 얼굴 일부가 노출되었습니다.",
        "hard_violations": [
         "Invented object: 텍스트와 레퍼런스에 의해 외형이 엄격히 고정(LOCKED)된 토미이시에게 레퍼런스에 없는 군모(모자)를 씌움."
        ],
        "physics": "낫은 두 손에 의해 강한 힘으로 쥐여 몸통에 단단히 꽂힌 채 지탱되고 있으며, 토미이시는 무릎을 꿇고 지면에 자세를 유지하고 있습니다."
       }
      ]
     },
     "ok": true
    },
    {
     "model": "gpt-high",
     "order": "reverse",
     "display_to_canonical": {
      "A": "B",
      "B": "A"
     },
     "raw": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 6,
        "verdict_ko": "왼쪽 아래의 두 손에서 몸통으로 이어지는 낫의 방향과 접촉은 더 충실하지만, 구도가 넓고 깊은 함몰이 약하며 참조에 없는 군모를 착용했다."
       },
       {
        "label": "B",
        "score": 3,
        "verdict_ko": "얼굴과 군복은 참조에 가깝지만, 낫 끝이 몸통에 박히지 않고 턱 앞에 드러나 있어 핵심 타격 순간을 구현하지 못했다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "안규홍의 두 손이 왼쪽 아래에서 자루를 잡고 중앙 오른쪽의 토미이시 몸통을 향해 밀고 있다. 굽은 날 끝은 가슴 주머니 아래쪽 옷에 가려져 몸통과 접촉한 것으로 읽힌다. 토미이시는 고개와 눈을 아래로 숙여 타격 부위 쪽을 향한다. 다만 깊숙한 관통과 뚜렷한 함몰은 확인하기 어렵다.",
        "built_space": "인공 구조물이나 고정 설비는 없다. 뒤쪽에 나무가 선 흙 경사면, 돌과 솔잎이 보인다. 왼쪽 바닥에 소총 한 자루, 오른쪽에 긴 자루의 낫 한 자루, 아래쪽에 칼 한 자루, 오른쪽 뒤에 일부 가려진 무기가 보인다. 참조의 숲속 싸움터 재질은 유지하지만, 얼굴부터 허리까지와 넓은 바닥을 포함해 요구한 접촉부 중심 클로즈업보다 넓다.",
        "entities": "토미이시는 검은 머리와 수염이 있는 성인 동아시아계 남성으로, 얼굴의 흙과 상처, 올리브색 군복, 붉은 견장과 갈색 가죽 띠가 참조와 대체로 맞는다. 국적은 외형만으로 확인할 수 없다. 참조에 없던 군모가 추가되었다. 안규홍은 흰 소매와 더러워진 성인 남성의 맨손·팔로 보인다. 낫은 낡은 나무 자루와 녹슨 곡선 금속 날이지만 참조보다 자루가 길어 보인다. 기존 옆구리 열상과 두 번째 관통상의 구분은 불명확하며, 늘어진 손에는 칼이 보이지 않는다. 읽을 수 있는 글자는 없다.",
        "hard_violations": [],
        "physics": "낫 자루는 두 손이 확실히 감싸 쥐고 있어 지지된다. 손과 팔에서 몸통 접촉부로 힘이 전달되는 연결이 보이며 토미이시는 상체를 앞으로 접고 있다. 하체 접지는 화면 밖이므로 판단할 수 없지만 공중에 떠 있는 몸으로 보이지는 않는다. 바닥의 무기는 흙에 놓여 있다. 옷의 압축은 약해 강한 돌진으로 몸통이 깊게 패인 순간의 물리적 표현은 부족하다."
       },
       {
        "label": "B",
        "direction": "안규홍의 손은 왼쪽 아래에서 자루를 중앙으로 밀지만, 낫의 뾰족한 끝은 토미이시 몸통이 아니라 턱 앞의 왼쪽 공간을 향한다. 날 전체가 몸통 밖에 드러나 있어 관통으로 읽히지 않는다. 토미이시의 눈은 화면 왼쪽 위, 안규홍의 상체 방향을 향한다.",
        "built_space": "인공 구조물이나 고정 설비는 없다. 돌과 솔잎이 섞인 흙 경사면과 뒤쪽 나무가 참조의 야외 장소를 대체로 유지한다. 오른쪽 뒤에 소총 한 자루와 긴 나무 자루의 무기 한 자루가 보이고, 오른쪽 아래에는 별도의 곡선 날 일부가 걸쳐 있다. 토미이시의 상체가 중앙과 오른쪽을, 안규홍의 흰옷 상체가 왼쪽을 차지한다. 얼굴부터 허리까지 포함되고 낫이 크게 드러나므로 작은 접촉부와 주변의 부드러운 흙 패치에 집중하는 지정 구도와 다르다.",
        "entities": "토미이시는 검은 짧은 머리와 수염, 얼굴의 오염과 상처를 지닌 성인 동아시아계 남성으로 참조 인상에 가깝고 군모는 없다. 올리브색 군복과 붉은 견장, 갈색 가죽 띠는 맞지만 참조의 가슴 훈장은 보이지 않는다. 안규홍은 흰 소매와 성인 남성의 맨손·팔로 나타난다. 낫의 낡은 나무와 녹슨 금속 재질은 참조에 가깝지만, 몸통에 박힌 낫이라는 요구는 충족하지 않는다. 군복의 혈흔은 보이되 깊은 옆구리 열상과 두 번째 관통 구멍은 확인되지 않는다. 칼을 놓치는 손은 화면 밖이며 읽을 수 있는 글자는 없다.",
        "hard_violations": [],
        "physics": "낫은 안규홍의 한 손이 자루를 단단히 잡아 지지한다. 토미이시는 뒤로 기울고 오른팔을 아래로 뻗으며, 몸 아래쪽은 프레임 경계와 지면으로 이어져 명백한 부유는 보이지 않는다. 배경 무기는 땅에 놓여 있다. 그러나 날 끝이 노출되어 있고 몸통의 함몰도 없어, 현재 자세는 깊게 꽂힌 낫이 돌진을 멈춘 충돌의 결과로 읽히지 않는다."
       }
      ],
      "all_candidates_fail": true
     },
     "normalized": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 6,
        "verdict_ko": "왼쪽 아래의 두 손에서 몸통으로 이어지는 낫의 방향과 접촉은 더 충실하지만, 구도가 넓고 깊은 함몰이 약하며 참조에 없는 군모를 착용했다."
       },
       {
        "label": "A",
        "score": 3,
        "verdict_ko": "얼굴과 군복은 참조에 가깝지만, 낫 끝이 몸통에 박히지 않고 턱 앞에 드러나 있어 핵심 타격 순간을 구현하지 못했다."
       }
      ],
      "all_candidates_fail": true,
      "readings": [
       {
        "label": "B",
        "direction": "안규홍의 두 손이 왼쪽 아래에서 자루를 잡고 중앙 오른쪽의 토미이시 몸통을 향해 밀고 있다. 굽은 날 끝은 가슴 주머니 아래쪽 옷에 가려져 몸통과 접촉한 것으로 읽힌다. 토미이시는 고개와 눈을 아래로 숙여 타격 부위 쪽을 향한다. 다만 깊숙한 관통과 뚜렷한 함몰은 확인하기 어렵다.",
        "built_space": "인공 구조물이나 고정 설비는 없다. 뒤쪽에 나무가 선 흙 경사면, 돌과 솔잎이 보인다. 왼쪽 바닥에 소총 한 자루, 오른쪽에 긴 자루의 낫 한 자루, 아래쪽에 칼 한 자루, 오른쪽 뒤에 일부 가려진 무기가 보인다. 참조의 숲속 싸움터 재질은 유지하지만, 얼굴부터 허리까지와 넓은 바닥을 포함해 요구한 접촉부 중심 클로즈업보다 넓다.",
        "entities": "토미이시는 검은 머리와 수염이 있는 성인 동아시아계 남성으로, 얼굴의 흙과 상처, 올리브색 군복, 붉은 견장과 갈색 가죽 띠가 참조와 대체로 맞는다. 국적은 외형만으로 확인할 수 없다. 참조에 없던 군모가 추가되었다. 안규홍은 흰 소매와 더러워진 성인 남성의 맨손·팔로 보인다. 낫은 낡은 나무 자루와 녹슨 곡선 금속 날이지만 참조보다 자루가 길어 보인다. 기존 옆구리 열상과 두 번째 관통상의 구분은 불명확하며, 늘어진 손에는 칼이 보이지 않는다. 읽을 수 있는 글자는 없다.",
        "hard_violations": [],
        "physics": "낫 자루는 두 손이 확실히 감싸 쥐고 있어 지지된다. 손과 팔에서 몸통 접촉부로 힘이 전달되는 연결이 보이며 토미이시는 상체를 앞으로 접고 있다. 하체 접지는 화면 밖이므로 판단할 수 없지만 공중에 떠 있는 몸으로 보이지는 않는다. 바닥의 무기는 흙에 놓여 있다. 옷의 압축은 약해 강한 돌진으로 몸통이 깊게 패인 순간의 물리적 표현은 부족하다."
       },
       {
        "label": "A",
        "direction": "안규홍의 손은 왼쪽 아래에서 자루를 중앙으로 밀지만, 낫의 뾰족한 끝은 토미이시 몸통이 아니라 턱 앞의 왼쪽 공간을 향한다. 날 전체가 몸통 밖에 드러나 있어 관통으로 읽히지 않는다. 토미이시의 눈은 화면 왼쪽 위, 안규홍의 상체 방향을 향한다.",
        "built_space": "인공 구조물이나 고정 설비는 없다. 돌과 솔잎이 섞인 흙 경사면과 뒤쪽 나무가 참조의 야외 장소를 대체로 유지한다. 오른쪽 뒤에 소총 한 자루와 긴 나무 자루의 무기 한 자루가 보이고, 오른쪽 아래에는 별도의 곡선 날 일부가 걸쳐 있다. 토미이시의 상체가 중앙과 오른쪽을, 안규홍의 흰옷 상체가 왼쪽을 차지한다. 얼굴부터 허리까지 포함되고 낫이 크게 드러나므로 작은 접촉부와 주변의 부드러운 흙 패치에 집중하는 지정 구도와 다르다.",
        "entities": "토미이시는 검은 짧은 머리와 수염, 얼굴의 오염과 상처를 지닌 성인 동아시아계 남성으로 참조 인상에 가깝고 군모는 없다. 올리브색 군복과 붉은 견장, 갈색 가죽 띠는 맞지만 참조의 가슴 훈장은 보이지 않는다. 안규홍은 흰 소매와 성인 남성의 맨손·팔로 나타난다. 낫의 낡은 나무와 녹슨 금속 재질은 참조에 가깝지만, 몸통에 박힌 낫이라는 요구는 충족하지 않는다. 군복의 혈흔은 보이되 깊은 옆구리 열상과 두 번째 관통 구멍은 확인되지 않는다. 칼을 놓치는 손은 화면 밖이며 읽을 수 있는 글자는 없다.",
        "hard_violations": [],
        "physics": "낫은 안규홍의 한 손이 자루를 단단히 잡아 지지한다. 토미이시는 뒤로 기울고 오른팔을 아래로 뻗으며, 몸 아래쪽은 프레임 경계와 지면으로 이어져 명백한 부유는 보이지 않는다. 배경 무기는 땅에 놓여 있다. 그러나 날 끝이 노출되어 있고 몸통의 함몰도 없어, 현재 자세는 깊게 꽂힌 낫이 돌진을 멈춘 충돌의 결과로 읽히지 않는다."
       }
      ]
     },
     "ok": true
    }
   ],
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "slot_winner_match": false,
   "slot_winner": {
    "gemini-pro": "A",
    "gpt-high": "B"
   },
   "route": "cross_slot_combined"
  },
  "dual": {
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "normalized": {
    "A": 1.5,
    "B": 1.333
   },
   "adjusted": {
    "A": 1.5,
    "B": 1.083
   },
   "violations": {
    "B": [
     "[gemini-pro] Invented object: 텍스트와 레퍼런스에 의해 외형이 엄격히 고정(LOCKED)된 토미이시에게 레퍼런스에 없는 군모(모자)를 씌움."
    ]
   },
   "per_model_winner": {
    "gemini-pro": "A",
    "gpt-high": "B"
   },
   "agreed": false
  },
  "totals": {
   "A": 1500,
   "B": 1083
  },
  "selected": "A",
  "ranking": [
   "A",
   "B"
  ],
  "verdicts": [
   {
    "label": "A",
    "score": 1500,
    "verdict_ko": "클로즈업 프레이밍과 캐릭터 외형(모자 없음)은 지시를 따랐으나, 낫이 몸통에 꽂히는 핵심 액션을 묘사하지 못하고 엉뚱하게 목 부위에 얹혀 있어 방향성 및 연출 지시를 크게 실패했습니다."
   },
   {
    "label": "B",
    "score": 1083,
    "verdict_ko": "낫이 몸통에 꽂히는 텍스트의 핵심 액션과 구도는 잘 구현했으나, 외형이 고정(LOCKED)된 캐릭터에게 레퍼런스에 없는 모자를 씌운 치명적 오류(Hard Violation)가 있으며 프레이밍이 너무 넓습니다.  ★위반: [gemini-pro] Invented object: 텍스트와 레퍼런스에 의해 외형이 엄격히 고정(LOCKED)된 토미이시에게 레퍼런스에 없는 군모(모자)를 씌움."
   }
  ],
  "refs": [
   {
    "label": "PREVIOUS SHOT STILL — a visually related earlier shot of this same place: the location's look, materials, fixed features and lighting mood are LOCKED to this photo; never copy its camera framing. The clothing, hair and overall look of 토미이시 — who appear both in that photo and in this shot — are LOCKED to that photo. Anyone else visible in that photo is NOT in this shot: never carry their face, body or clothing onto anyone here. If this photo shows a character who cannot move (dead or unconscious), that character's exact pose, position and orientation are ALSO LOCKED — treat their body as a fixed prop of the set that only the camera moves around.",
    "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/db00517b-165b-4005-a287-f28936b59b95/images/272bd7af-9294-4490-98c6-88a40e9deacc/scene/recipe/S63sh18_sel.png",
    "asset_id": "fdb2118f-1d4c-4f64-8b14-a9a4ce41b13a",
    "role": "prev_still"
   },
   {
    "label": "CHARACTER REFERENCE — 토미이시: the exact person appearing in this shot; match face, hair and build exactly.",
    "path": "<bytes:1407391>",
    "asset_id": "2743b2cf-b5c5-4196-b248-300e6bbbc12e",
    "role": "character_ref"
   },
   {
    "label": "PROP REFERENCE — 낫: the exact object appearing in this shot; match its look, material and wear exactly.",
    "path": "<bytes:977465>",
    "asset_id": "e9884693-b2aa-4769-9ee0-4c7b50db40c2",
    "role": "prop_ref"
   }
  ],
  "critique_skipped": true,
  "needs_reshoot": true,
  "shot_run_produced": false,
  "shot_run_uid": "06aa1446-677e-7729-bb74-e7e02344d4d2",
  "ref_mode": "prev+엔티티",
  "share_plan": {
   "ref_plan": "prev",
   "prev_anchor_tag": "S63sh18"
  }
 },
 "S63sh53::cine": {
  "applied": true,
  "attempted_at": "2026-09-09T08:22:16.622286+00:00",
  "fingerprint": "02bc8f06bf95e7d0322862e5ba3663d6e6c793f6b36e4c506f2c507216a6559e",
  "fingerprint_version": 2,
  "provider": "mai",
  "endpoint": "openrouter/chat-completions",
  "model": "microsoft/mai-image-2.6",
  "pack": "24.202608252115",
  "source_file": "S63sh53_sel.png",
  "source_sha256": "6b261c6dbed094512f3b68cee85ae31e34aae6a2fe53967b606c8b0c8be4ff44",
  "file": "S63sh53_cine.png",
  "staged_sha256": "ee033ed031d5f39af4c2557b57fff5d7b515b73c323879662652318abc21c948",
  "latency_ms": 25086
 },
 "S63sh61::signage": {
  "fp": "745cb34447ed480f",
  "inscriptions": [],
  "cues": [],
  "dropped": []
 },
 "S63sh61": {
  "input_fingerprint": "2ecfa4e2cec7e83e",
  "prompt": "Create ONE FINAL photorealistic live-action film still of the moment below — early 20th-century Korea, Jeolla region; all people are Korean unless explicitly stated otherwise. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 자신의 발밑에 널브러진 토미이시의 시신을 묵묵히 내려다보는 안규홍의 가라앉은 시선.\n\nLOCATION (lock): The dirt ground at the end of the duel on the mountain camp's wooded slope. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: close-up\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Unembellished daylight with restrained saturation preserves the blood-and-earth-covered face and the exhausted gravity of the pause.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The people visible in that photo are NOT in this shot: never carry their faces, bodies or clothing onto anyone here. Each person in THIS shot is defined solely by the PEOPLE list, their own CHARACTER REFERENCE images and their wardrobe notes. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nIMMOBILE CHARACTER POSE — CANONICAL (identical wherever this character appears in ANY panel; on any conflict THIS POSE WINS): His dead body is collapsed on the ground at An Gyuhong's feet, with his sword released from his hand. The screenplay does not specify his final torso orientation, head angle, or the resting positions of his arms and legs.\n\nIMMOBILE BODIES OBEY GRAVITY: a person who is dead or unconscious\nexerts NO muscular effort. Every part of their body — head, torso,\narms, hands, fingers, legs — rests fully on whatever supports it\n(floor, wall, furniture, their own lap) and hangs or slumps with\ngravity. NEVER show any part of an immobile person's body lifted,\nraised, held up in the air, or posed as if presenting something:\nan object in their grip stays clenched in a hand that itself lies\nfallen on a support — the hand does not hold the object up. If the\ncanonical pose leaves a body part unspecified, resolve it as the\nmost gravity-compliant, fully-supported position.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The sword lies dropped on the dirt again; the sword puncture, broken cookware and camp wreckage remain. 안규홍: He stands holding the bloody sickle, exhausted and blood- and dirt-covered, with his back wound still bleeding.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 안규홍 (한국인 남성, 청년 나이대, 검은 머리, 단단히 묶은 머리) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.",
  "roll_prompts": {
   "A": "Create ONE FINAL photorealistic live-action film still of the moment below — early 20th-century Korea, Jeolla region; all people are Korean unless explicitly stated otherwise. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 자신의 발밑에 널브러진 토미이시의 시신을 묵묵히 내려다보는 안규홍의 가라앉은 시선.\n\nLOCATION (lock): The dirt ground at the end of the duel on the mountain camp's wooded slope. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: close-up\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Unembellished daylight with restrained saturation preserves the blood-and-earth-covered face and the exhausted gravity of the pause.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The people visible in that photo are NOT in this shot: never carry their faces, bodies or clothing onto anyone here. Each person in THIS shot is defined solely by the PEOPLE list, their own CHARACTER REFERENCE images and their wardrobe notes. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nIMMOBILE CHARACTER POSE — CANONICAL (identical wherever this character appears in ANY panel; on any conflict THIS POSE WINS): His dead body is collapsed on the ground at An Gyuhong's feet, with his sword released from his hand. The screenplay does not specify his final torso orientation, head angle, or the resting positions of his arms and legs.\n\nIMMOBILE BODIES OBEY GRAVITY: a person who is dead or unconscious\nexerts NO muscular effort. Every part of their body — head, torso,\narms, hands, fingers, legs — rests fully on whatever supports it\n(floor, wall, furniture, their own lap) and hangs or slumps with\ngravity. NEVER show any part of an immobile person's body lifted,\nraised, held up in the air, or posed as if presenting something:\nan object in their grip stays clenched in a hand that itself lies\nfallen on a support — the hand does not hold the object up. If the\ncanonical pose leaves a body part unspecified, resolve it as the\nmost gravity-compliant, fully-supported position.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The sword lies dropped on the dirt again; the sword puncture, broken cookware and camp wreckage remain. 안규홍: He stands holding the bloody sickle, exhausted and blood- and dirt-covered, with his back wound still bleeding.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 안규홍 (한국인 남성, 청년 나이대, 검은 머리, 단단히 묶은 머리) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.",
   "B": "Create ONE FINAL photorealistic live-action film still of the moment below — early 20th-century Korea, Jeolla region; all people are Korean unless explicitly stated otherwise. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 자신의 발밑에 널브러진 토미이시의 시신을 묵묵히 내려다보는 안규홍의 가라앉은 시선.\n\nLOCATION (lock): The dirt ground at the end of the duel on the mountain camp's wooded slope. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: close-up\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Unembellished daylight with restrained saturation preserves the blood-and-earth-covered face and the exhausted gravity of the pause.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The people visible in that photo are NOT in this shot: never carry their faces, bodies or clothing onto anyone here. Each person in THIS shot is defined solely by the PEOPLE list, their own CHARACTER REFERENCE images and their wardrobe notes. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nIMMOBILE CHARACTER POSE — CANONICAL (identical wherever this character appears in ANY panel; on any conflict THIS POSE WINS): His dead body is collapsed on the ground at An Gyuhong's feet, with his sword released from his hand. The screenplay does not specify his final torso orientation, head angle, or the resting positions of his arms and legs.\n\nIMMOBILE BODIES OBEY GRAVITY: a person who is dead or unconscious\nexerts NO muscular effort. Every part of their body — head, torso,\narms, hands, fingers, legs — rests fully on whatever supports it\n(floor, wall, furniture, their own lap) and hangs or slumps with\ngravity. NEVER show any part of an immobile person's body lifted,\nraised, held up in the air, or posed as if presenting something:\nan object in their grip stays clenched in a hand that itself lies\nfallen on a support — the hand does not hold the object up. If the\ncanonical pose leaves a body part unspecified, resolve it as the\nmost gravity-compliant, fully-supported position.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The sword lies dropped on the dirt again; the sword puncture, broken cookware and camp wreckage remain. 안규홍: He stands holding the bloody sickle, exhausted and blood- and dirt-covered, with his back wound still bleeding.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 안규홍 (한국인 남성, 청년 나이대, 검은 머리, 단단히 묶은 머리) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.\n\nCOMPOSITION VARIATION (this candidate only): keep the place, the people and the moment exactly as specified, and keep from the CAMERA & FRAME contract its subject scale, screen placement and key background elements exactly. For this candidate only, its camera angle and height are deliberately left open: choose a camera position distinctly different from the obvious one, as a film director picking a second setup on the same blocking. Stay true to any explicit framing the SHOT TEXT itself dictates."
  },
  "shot_run_spend_attempt_count": 1,
  "initial_roll_all_fail": true,
  "readings": [
   {
    "label": "A",
    "direction": "안규홍의 시선은 바닥에 누운 토미이시를 향해 아래로 향하고 있다.",
    "built_space": "산비탈 흙바닥. 솥 등 부서진 취사 도구와 큰 낫 등의 무기들이 바닥에 넓게 흩어져 있다.",
    "entities": "안규홍(손에 낫을 쥐고 있지 않고 맨손임), 토미이시(바닥에 누워 있으나 오른손에 칼을 쥐고 있음). 화면 좌측 하단에 흰 소매를 입은 정체불명의 손이 나타남.",
    "hard_violations": [
     "[gemini-pro] 화면 좌측 하단에 몸체와 연결되지 않은 정체불명의 팔과 손이 칼을 쥐고 나타남 (존재하지 않는 신체 부위 및 인물 추가)"
    ],
    "physics": "안규홍은 바닥에 안정적으로 서 있다. 토미이시는 죽은 상태임에도 손으로 칼을 꽉 쥐고 지탱하고 있다. 좌측 하단의 칼을 쥔 손은 연결된 몸통 없이 프레임 밖에서 허공을 통해 들어와 있다."
   },
   {
    "label": "B",
    "direction": "안규홍의 시선은 자신의 발밑에 쓰러진 토미이시의 시신을 향해 아래로 고정되어 있다.",
    "built_space": "산비탈 흙바닥. 엎어진 항아리와 잔해가 바닥에 배치되어 카메라 위치 및 각도에 맞는 지형이 묘사됨.",
    "entities": "안규홍(피 묻은 낫을 오른손에 쥠), 토미이시(군복을 입은 채 바닥에 쓰러진 시신, 칼은 그의 목 아래 바닥에 떨어져 있음).",
    "hard_violations": [
     "[gemini-pro] 죽은 토미이시의 시신 왼손이 어떤 지지대도 없이 허공으로 뻗은 채 굳어 있음 (시신의 중력 지침 및 물리적 지탱 법칙 위반)"
    ],
    "physics": "안규홍은 오른손으로 낫의 손잡이를 쥐고 섰다. 토미이시의 시신은 몸통이 바닥에 닿아 있으나, 왼손이 아무런 지지대 없이 허공을 향해 들려 있어 중력에 어긋난다."
   }
  ],
  "cross_model_order": {
   "policy": "cross_model_order_v2_slot0_fwd_slot1_rev_parallel2",
   "slots": [
    {
     "model": "gemini-pro",
     "order": "forward",
     "display_to_canonical": {
      "A": "A",
      "B": "B"
     },
     "raw": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 2,
        "verdict_ko": "지정된 샷 크기와 소품(피 묻은 낫) 묘사에서 프롬프트를 더 잘 따랐으나, 죽은 시신의 손이 허공에 들려 있어 물리 및 시신 지침을 위반하여 실격입니다."
       },
       {
        "label": "A",
        "score": 0,
        "verdict_ko": "프레이밍이 지정된 클로즈업보다 훨씬 넓고 안규홍이 낫을 쥐고 있지 않으며, 화면 좌측 하단에 정체불명의 팔이 추가로 등장해 실격입니다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "안규홍의 시선은 바닥에 누운 토미이시를 향해 아래로 향하고 있다.",
        "built_space": "산비탈 흙바닥. 솥 등 부서진 취사 도구와 큰 낫 등의 무기들이 바닥에 넓게 흩어져 있다.",
        "entities": "안규홍(손에 낫을 쥐고 있지 않고 맨손임), 토미이시(바닥에 누워 있으나 오른손에 칼을 쥐고 있음). 화면 좌측 하단에 흰 소매를 입은 정체불명의 손이 나타남.",
        "hard_violations": [
         "화면 좌측 하단에 몸체와 연결되지 않은 정체불명의 팔과 손이 칼을 쥐고 나타남 (존재하지 않는 신체 부위 및 인물 추가)"
        ],
        "physics": "안규홍은 바닥에 안정적으로 서 있다. 토미이시는 죽은 상태임에도 손으로 칼을 꽉 쥐고 지탱하고 있다. 좌측 하단의 칼을 쥔 손은 연결된 몸통 없이 프레임 밖에서 허공을 통해 들어와 있다."
       },
       {
        "label": "B",
        "direction": "안규홍의 시선은 자신의 발밑에 쓰러진 토미이시의 시신을 향해 아래로 고정되어 있다.",
        "built_space": "산비탈 흙바닥. 엎어진 항아리와 잔해가 바닥에 배치되어 카메라 위치 및 각도에 맞는 지형이 묘사됨.",
        "entities": "안규홍(피 묻은 낫을 오른손에 쥠), 토미이시(군복을 입은 채 바닥에 쓰러진 시신, 칼은 그의 목 아래 바닥에 떨어져 있음).",
        "hard_violations": [
         "죽은 토미이시의 시신 왼손이 어떤 지지대도 없이 허공으로 뻗은 채 굳어 있음 (시신의 중력 지침 및 물리적 지탱 법칙 위반)"
        ],
        "physics": "안규홍은 오른손으로 낫의 손잡이를 쥐고 섰다. 토미이시의 시신은 몸통이 바닥에 닿아 있으나, 왼손이 아무런 지지대 없이 허공을 향해 들려 있어 중력에 어긋난다."
       }
      ],
      "all_candidates_fail": true
     },
     "normalized": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 2,
        "verdict_ko": "지정된 샷 크기와 소품(피 묻은 낫) 묘사에서 프롬프트를 더 잘 따랐으나, 죽은 시신의 손이 허공에 들려 있어 물리 및 시신 지침을 위반하여 실격입니다."
       },
       {
        "label": "A",
        "score": 0,
        "verdict_ko": "프레이밍이 지정된 클로즈업보다 훨씬 넓고 안규홍이 낫을 쥐고 있지 않으며, 화면 좌측 하단에 정체불명의 팔이 추가로 등장해 실격입니다."
       }
      ],
      "all_candidates_fail": true,
      "readings": [
       {
        "label": "A",
        "direction": "안규홍의 시선은 바닥에 누운 토미이시를 향해 아래로 향하고 있다.",
        "built_space": "산비탈 흙바닥. 솥 등 부서진 취사 도구와 큰 낫 등의 무기들이 바닥에 넓게 흩어져 있다.",
        "entities": "안규홍(손에 낫을 쥐고 있지 않고 맨손임), 토미이시(바닥에 누워 있으나 오른손에 칼을 쥐고 있음). 화면 좌측 하단에 흰 소매를 입은 정체불명의 손이 나타남.",
        "hard_violations": [
         "화면 좌측 하단에 몸체와 연결되지 않은 정체불명의 팔과 손이 칼을 쥐고 나타남 (존재하지 않는 신체 부위 및 인물 추가)"
        ],
        "physics": "안규홍은 바닥에 안정적으로 서 있다. 토미이시는 죽은 상태임에도 손으로 칼을 꽉 쥐고 지탱하고 있다. 좌측 하단의 칼을 쥔 손은 연결된 몸통 없이 프레임 밖에서 허공을 통해 들어와 있다."
       },
       {
        "label": "B",
        "direction": "안규홍의 시선은 자신의 발밑에 쓰러진 토미이시의 시신을 향해 아래로 고정되어 있다.",
        "built_space": "산비탈 흙바닥. 엎어진 항아리와 잔해가 바닥에 배치되어 카메라 위치 및 각도에 맞는 지형이 묘사됨.",
        "entities": "안규홍(피 묻은 낫을 오른손에 쥠), 토미이시(군복을 입은 채 바닥에 쓰러진 시신, 칼은 그의 목 아래 바닥에 떨어져 있음).",
        "hard_violations": [
         "죽은 토미이시의 시신 왼손이 어떤 지지대도 없이 허공으로 뻗은 채 굳어 있음 (시신의 중력 지침 및 물리적 지탱 법칙 위반)"
        ],
        "physics": "안규홍은 오른손으로 낫의 손잡이를 쥐고 섰다. 토미이시의 시신은 몸통이 바닥에 닿아 있으나, 왼손이 아무런 지지대 없이 허공을 향해 들려 있어 중력에 어긋난다."
       }
      ]
     },
     "ok": true
    },
    {
     "model": "gpt-high",
     "order": "reverse",
     "display_to_canonical": {
      "A": "B",
      "B": "A"
     },
     "raw": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 6,
        "verdict_ko": "두 후보 모두 지정된 클로즈업보다 넓지만, A는 안규홍의 가라앉은 시선이 아래쪽 시신을 향하고 얼굴도 더 크게 보여 핵심 순간에 더 가깝다."
       },
       {
        "label": "B",
        "score": 3,
        "verdict_ko": "허벅지까지 담은 넓은 구도이며 시선은 뒤쪽 시신이 아닌 자기 앞의 땅을 향하고, 토미이시가 놓아야 할 검도 손에 쥐고 있다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "안규홍은 고개와 눈을 화면 왼쪽 아래로 내려, 그곳에 누운 토미이시의 머리와 상체 쪽을 본다. 들고 있는 낫의 끝도 시신의 얼굴 가까운 아래쪽을 향하지만 휘두르는 동작은 아니다. 떨어진 검날은 바닥을 따라 왼쪽 아래로 뻗는다.",
        "built_space": "나무줄기, 돌, 낙엽과 흙으로 된 비탈이 이전 장면의 장소와 부합한다. 건축물이나 고정 설비는 없다. 왼쪽에 넘어진 솥 하나와 뚜껑 하나가 보인다. 안규홍은 오른쪽 전경, 시신은 왼쪽 아래에 놓인다. 다만 얼굴뿐 아니라 등과 팔, 시신 상체까지 포함해 클로즈업보다 넓다.",
        "entities": "안규홍은 참조와 유사한 젊은 동아시아계 남성 얼굴, 뒤로 묶은 검은 머리, 피와 흙이 묻은 흰 저고리를 지녔다. 얼굴 상처와 등 쪽 혈흔도 보인다. 오른손에는 나무 손잡이의 금속 낫이 있으나 낫날의 피는 뚜렷하지 않다. 군복과 가죽 교차 끈을 착용한 수염 난 토미이시가 누워 있고, 검은 손에서 분리되어 바닥에 있다. 별도의 검 자상과 깨진 조리도구의 파손 정도는 명확하지 않다. 읽을 수 있는 글자는 없다.",
        "hard_violations": [],
        "physics": "낫은 안규홍의 오른손이 손잡이를 감싸 지탱한다. 하체는 화면 밖이므로 발의 접촉은 확인할 수 없다. 시신의 머리와 몸통은 흙과 돌에 놓여 있다. 손바닥을 위로 둔 손은 손목과 소매가 지면 및 검 부근에 겹쳐 있어 손 전체의 접촉은 불명확하지만, 명백하게 공중에 떠 있다고 단정할 근거는 부족하다. 솥과 뚜껑, 떨어진 검은 지면에 놓여 있다."
       },
       {
        "label": "B",
        "direction": "안규홍의 눈은 정면 아래, 자기 앞의 화면 하단을 향한다. 토미이시의 머리와 상체는 안규홍의 왼쪽 뒤에 있어 그 시선이 시신에 닿지 않는다. 시신의 손에 있는 검날은 오른쪽 아래 화면 밖으로 뻗고, 뒤쪽의 큰 곡날 도구는 바닥에 놓여 있다.",
        "built_space": "흙과 낙엽, 바위, 나무가 있는 낮의 숲 비탈은 이전 장소와 대체로 일치한다. 왼쪽에는 검은 조리용기와 뚜껑류 네 점, 천과 야영 잔해가 보인다. 안규홍은 중앙 전경에, 시신은 왼쪽 뒤에 배치된다. 머리부터 허벅지까지 담은 구도여서 지정된 클로즈업과 크게 다르다. 건축 설비나 반사는 없다.",
        "entities": "안규홍의 젊은 남성 얼굴, 묶은 검은 머리, 흰 저고리와 가슴의 혈흔은 인물 참조에 가깝다. 얼굴에는 피와 흙이 있으나 A보다 오염이 덜 두드러진다. 양손 부근은 화면 하단에 걸려 있어 들고 있어야 할 낫의 상태를 확정하기 어렵다. 토미이시는 군복과 가죽 끈을 착용했지만, 화면 왼쪽 아래 손이 검 손잡이를 쥐고 있어 검을 놓았다는 명시 조건과 어긋난다. 등 상처는 구도상 보이지 않는다. 읽을 수 있는 글자는 없다.",
        "hard_violations": [],
        "physics": "안규홍의 상체는 자연스럽게 수직으로 서 있으나 발은 화면 밖이다. 시신의 머리는 돌과 지면에, 몸통과 보이는 팔은 비탈에 기대어 놓여 있다. 검을 쥔 손도 바닥 가까이 떨어져 있어 검을 공중에 들어 올린 자세는 아니지만, 검을 놓은 상태는 아니다. 조리도구와 긴 곡날 도구는 흙바닥이 지탱하며 명백한 부유는 없다."
       }
      ],
      "all_candidates_fail": true
     },
     "normalized": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 6,
        "verdict_ko": "두 후보 모두 지정된 클로즈업보다 넓지만, A는 안규홍의 가라앉은 시선이 아래쪽 시신을 향하고 얼굴도 더 크게 보여 핵심 순간에 더 가깝다."
       },
       {
        "label": "A",
        "score": 3,
        "verdict_ko": "허벅지까지 담은 넓은 구도이며 시선은 뒤쪽 시신이 아닌 자기 앞의 땅을 향하고, 토미이시가 놓아야 할 검도 손에 쥐고 있다."
       }
      ],
      "all_candidates_fail": true,
      "readings": [
       {
        "label": "B",
        "direction": "안규홍은 고개와 눈을 화면 왼쪽 아래로 내려, 그곳에 누운 토미이시의 머리와 상체 쪽을 본다. 들고 있는 낫의 끝도 시신의 얼굴 가까운 아래쪽을 향하지만 휘두르는 동작은 아니다. 떨어진 검날은 바닥을 따라 왼쪽 아래로 뻗는다.",
        "built_space": "나무줄기, 돌, 낙엽과 흙으로 된 비탈이 이전 장면의 장소와 부합한다. 건축물이나 고정 설비는 없다. 왼쪽에 넘어진 솥 하나와 뚜껑 하나가 보인다. 안규홍은 오른쪽 전경, 시신은 왼쪽 아래에 놓인다. 다만 얼굴뿐 아니라 등과 팔, 시신 상체까지 포함해 클로즈업보다 넓다.",
        "entities": "안규홍은 참조와 유사한 젊은 동아시아계 남성 얼굴, 뒤로 묶은 검은 머리, 피와 흙이 묻은 흰 저고리를 지녔다. 얼굴 상처와 등 쪽 혈흔도 보인다. 오른손에는 나무 손잡이의 금속 낫이 있으나 낫날의 피는 뚜렷하지 않다. 군복과 가죽 교차 끈을 착용한 수염 난 토미이시가 누워 있고, 검은 손에서 분리되어 바닥에 있다. 별도의 검 자상과 깨진 조리도구의 파손 정도는 명확하지 않다. 읽을 수 있는 글자는 없다.",
        "hard_violations": [],
        "physics": "낫은 안규홍의 오른손이 손잡이를 감싸 지탱한다. 하체는 화면 밖이므로 발의 접촉은 확인할 수 없다. 시신의 머리와 몸통은 흙과 돌에 놓여 있다. 손바닥을 위로 둔 손은 손목과 소매가 지면 및 검 부근에 겹쳐 있어 손 전체의 접촉은 불명확하지만, 명백하게 공중에 떠 있다고 단정할 근거는 부족하다. 솥과 뚜껑, 떨어진 검은 지면에 놓여 있다."
       },
       {
        "label": "A",
        "direction": "안규홍의 눈은 정면 아래, 자기 앞의 화면 하단을 향한다. 토미이시의 머리와 상체는 안규홍의 왼쪽 뒤에 있어 그 시선이 시신에 닿지 않는다. 시신의 손에 있는 검날은 오른쪽 아래 화면 밖으로 뻗고, 뒤쪽의 큰 곡날 도구는 바닥에 놓여 있다.",
        "built_space": "흙과 낙엽, 바위, 나무가 있는 낮의 숲 비탈은 이전 장소와 대체로 일치한다. 왼쪽에는 검은 조리용기와 뚜껑류 네 점, 천과 야영 잔해가 보인다. 안규홍은 중앙 전경에, 시신은 왼쪽 뒤에 배치된다. 머리부터 허벅지까지 담은 구도여서 지정된 클로즈업과 크게 다르다. 건축 설비나 반사는 없다.",
        "entities": "안규홍의 젊은 남성 얼굴, 묶은 검은 머리, 흰 저고리와 가슴의 혈흔은 인물 참조에 가깝다. 얼굴에는 피와 흙이 있으나 A보다 오염이 덜 두드러진다. 양손 부근은 화면 하단에 걸려 있어 들고 있어야 할 낫의 상태를 확정하기 어렵다. 토미이시는 군복과 가죽 끈을 착용했지만, 화면 왼쪽 아래 손이 검 손잡이를 쥐고 있어 검을 놓았다는 명시 조건과 어긋난다. 등 상처는 구도상 보이지 않는다. 읽을 수 있는 글자는 없다.",
        "hard_violations": [],
        "physics": "안규홍의 상체는 자연스럽게 수직으로 서 있으나 발은 화면 밖이다. 시신의 머리는 돌과 지면에, 몸통과 보이는 팔은 비탈에 기대어 놓여 있다. 검을 쥔 손도 바닥 가까이 떨어져 있어 검을 공중에 들어 올린 자세는 아니지만, 검을 놓은 상태는 아니다. 조리도구와 긴 곡날 도구는 흙바닥이 지탱하며 명백한 부유는 없다."
       }
      ]
     },
     "ok": true
    }
   ],
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "slot_winner_match": true,
   "slot_winner": {
    "gemini-pro": "B",
    "gpt-high": "B"
   },
   "route": "cross_slot_agree"
  },
  "dual": {
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "normalized": {
    "A": 0.5,
    "B": 2.0
   },
   "adjusted": {
    "A": 0.25,
    "B": 1.75
   },
   "violations": {
    "A": [
     "[gemini-pro] 화면 좌측 하단에 몸체와 연결되지 않은 정체불명의 팔과 손이 칼을 쥐고 나타남 (존재하지 않는 신체 부위 및 인물 추가)"
    ],
    "B": [
     "[gemini-pro] 죽은 토미이시의 시신 왼손이 어떤 지지대도 없이 허공으로 뻗은 채 굳어 있음 (시신의 중력 지침 및 물리적 지탱 법칙 위반)"
    ]
   },
   "per_model_winner": {
    "gemini-pro": "B",
    "gpt-high": "B"
   },
   "agreed": true
  },
  "totals": {
   "B": 1750,
   "A": 250
  },
  "selected": "B",
  "ranking": [
   "B",
   "A"
  ],
  "verdicts": [
   {
    "label": "B",
    "score": 1750,
    "verdict_ko": "지정된 샷 크기와 소품(피 묻은 낫) 묘사에서 프롬프트를 더 잘 따랐으나, 죽은 시신의 손이 허공에 들려 있어 물리 및 시신 지침을 위반하여 실격입니다.  ★위반: [gemini-pro] 죽은 토미이시의 시신 왼손이 어떤 지지대도 없이 허공으로 뻗은 채 굳어 있음 (시신의 중력 지침 및 물리적 지탱 법칙 위반)"
   },
   {
    "label": "A",
    "score": 250,
    "verdict_ko": "프레이밍이 지정된 클로즈업보다 훨씬 넓고 안규홍이 낫을 쥐고 있지 않으며, 화면 좌측 하단에 정체불명의 팔이 추가로 등장해 실격입니다.  ★위반: [gemini-pro] 화면 좌측 하단에 몸체와 연결되지 않은 정체불명의 팔과 손이 칼을 쥐고 나타남 (존재하지 않는 신체 부위 및 인물 추가)"
   }
  ],
  "refs": [
   {
    "label": "PREVIOUS SHOT STILL — a visually related earlier shot of this same place: the location's look, materials, fixed features and lighting mood are LOCKED to this photo; never copy its camera framing. The people visible in that photo are NOT in this shot: never carry their faces, bodies or clothing onto anyone here. Each person in THIS shot is defined solely by the PEOPLE list, their own CHARACTER REFERENCE images and their wardrobe notes. If this photo shows a character who cannot move (dead or unconscious), that character's exact pose, position and orientation are ALSO LOCKED — treat their body as a fixed prop of the set that only the camera moves around.",
    "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/db00517b-165b-4005-a287-f28936b59b95/images/272bd7af-9294-4490-98c6-88a40e9deacc/scene/recipe/S63sh53_sel.png",
    "asset_id": "812bbe87-4667-4e62-bd9c-1eb73d0c0e7c",
    "role": "prev_still"
   },
   {
    "label": "CHARACTER REFERENCE — 안규홍: the exact person appearing in this shot; match face, hair and build exactly.",
    "path": "<bytes:1483645>",
    "asset_id": "b52dff85-b663-44f9-aa9c-46b828990f55",
    "role": "character_ref"
   }
  ],
  "critique_skipped": true,
  "needs_reshoot": true,
  "shot_run_produced": false,
  "shot_run_uid": "06aa1446-692e-7503-bab5-2b175929017e",
  "ref_mode": "prev+엔티티",
  "share_plan": {
   "ref_plan": "prev",
   "prev_anchor_tag": "S63sh53"
  }
 },
 "S63sh61::cine": {
  "applied": true,
  "attempted_at": "2026-09-09T09:37:57.757026+00:00",
  "fingerprint": "a08d7d83f2d1b86352a6d7d2731bfad324f849de61887da76d6b94525ef87cd1",
  "fingerprint_version": 2,
  "provider": "mai",
  "endpoint": "openrouter/chat-completions",
  "model": "microsoft/mai-image-2.6",
  "multiroll_tag": "still_S63sh61_cine",
  "slot": "",
  "pack": "24.202608252115",
  "source_file": "S63sh61_sel.png",
  "source_sha256": "0e27bab41ff466c9835ac38f729c53c8aee7a4f29a4651b46a0a27cbf4a820bf",
  "file": "S63sh61_cine.png",
  "staged_sha256": "e50c8f78f0bfdfe659cbdfc84d7709e30765352906a66ce6abd1b2c7419a3e28",
  "latency_ms": 26320
 },
 "S64sh3::signage": {
  "fp": "2fc43cc58512b5fe",
  "inscriptions": [],
  "cues": [],
  "dropped": []
 },
 "S64sh3": {
  "input_fingerprint": "0aa303057ad46c19",
  "prompt": "Create ONE FINAL photorealistic live-action film still of the moment below — early 20th-century Korea, Jeolla region; all people are Korean unless explicitly stated otherwise. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 두 눈을 뜬 채 흙바닥에 굳어 있는 임정현의 창백한 시신 클로즈업.\n\nLOCATION (lock): The dirt ground beside a ruined hut in the mountain camp, where the fallen scribe remains as the smoke clears. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: close-up\n- KEY BACKGROUND ELEMENTS: Earth around 임정현's head (His body lies on the ground); used as Surrounds the face with a restrained contextual border rather than turning the eyes into an isolated insert.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Daylight filtered by the scene's gradually clearing gunpowder smoke remains subdued and naturalistic, preserving the pallor without a supernatural cast.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The clothing, hair and overall look of 임정현 — who appear both in that photo and in this shot — are LOCKED to that photo. Anyone else visible in that photo is NOT in this shot: never carry their face, body or clothing onto anyone here. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): Powder smoke slowly clears over burned or collapsed huts, the overturned pot and bowls and potatoes scattered through blood and dirt; the dropped sword remains on the ground.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 임정현 (한국인 남성, 성인 나이대, 자연스러운 검은 머리) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.",
  "roll_prompts": {
   "A": "Create ONE FINAL photorealistic live-action film still of the moment below — early 20th-century Korea, Jeolla region; all people are Korean unless explicitly stated otherwise. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 두 눈을 뜬 채 흙바닥에 굳어 있는 임정현의 창백한 시신 클로즈업.\n\nLOCATION (lock): The dirt ground beside a ruined hut in the mountain camp, where the fallen scribe remains as the smoke clears. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: close-up\n- KEY BACKGROUND ELEMENTS: Earth around 임정현's head (His body lies on the ground); used as Surrounds the face with a restrained contextual border rather than turning the eyes into an isolated insert.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Daylight filtered by the scene's gradually clearing gunpowder smoke remains subdued and naturalistic, preserving the pallor without a supernatural cast.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The clothing, hair and overall look of 임정현 — who appear both in that photo and in this shot — are LOCKED to that photo. Anyone else visible in that photo is NOT in this shot: never carry their face, body or clothing onto anyone here. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): Powder smoke slowly clears over burned or collapsed huts, the overturned pot and bowls and potatoes scattered through blood and dirt; the dropped sword remains on the ground.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 임정현 (한국인 남성, 성인 나이대, 자연스러운 검은 머리) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.",
   "B": "Create ONE FINAL photorealistic live-action film still of the moment below — early 20th-century Korea, Jeolla region; all people are Korean unless explicitly stated otherwise. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 두 눈을 뜬 채 흙바닥에 굳어 있는 임정현의 창백한 시신 클로즈업.\n\nLOCATION (lock): The dirt ground beside a ruined hut in the mountain camp, where the fallen scribe remains as the smoke clears. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: close-up\n- KEY BACKGROUND ELEMENTS: Earth around 임정현's head (His body lies on the ground); used as Surrounds the face with a restrained contextual border rather than turning the eyes into an isolated insert.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Daylight filtered by the scene's gradually clearing gunpowder smoke remains subdued and naturalistic, preserving the pallor without a supernatural cast.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The clothing, hair and overall look of 임정현 — who appear both in that photo and in this shot — are LOCKED to that photo. Anyone else visible in that photo is NOT in this shot: never carry their face, body or clothing onto anyone here. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): Powder smoke slowly clears over burned or collapsed huts, the overturned pot and bowls and potatoes scattered through blood and dirt; the dropped sword remains on the ground.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 임정현 (한국인 남성, 성인 나이대, 자연스러운 검은 머리) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.\n\nCOMPOSITION VARIATION (this candidate only): keep the place, the people and the moment exactly as specified, and keep from the CAMERA & FRAME contract its subject scale, screen placement and key background elements exactly. For this candidate only, its camera angle and height are deliberately left open: choose a camera position distinctly different from the obvious one, as a film director picking a second setup on the same blocking. Stay true to any explicit framing the SHOT TEXT itself dictates."
  },
  "shot_run_spend_attempt_count": 1,
  "readings": [
   {
    "label": "A",
    "direction": "두 눈을 뜨고 위쪽 허공을 응시하고 있습니다.",
    "built_space": "흙바닥 위에 쌀알과 깨진 그릇이 흩어져 있으며, 뒤쪽으로 오두막의 엮은 벽면이 보입니다.",
    "entities": "임정현의 얼굴 특징, 상처, 옷차림이 참조 이미지와 완벽하게 일치합니다.",
    "hard_violations": [],
    "physics": "머리와 몸통이 흙바닥에 안정적으로 지탱되어 누워 있습니다."
   },
   {
    "label": "B",
    "direction": "두 눈을 뜨고 위쪽을 응시하고 있습니다.",
    "built_space": "바닥에 칼과 파편이 있으며, 배경에 돌무더기와 불탄 나무 구조물이 있습니다.",
    "entities": "임정현의 얼굴에 참조에 없던 눈물 자국이 추가되었으며, 인상도 약간 다릅니다.",
    "hard_violations": [
     "[gemini-pro] 프롬프트에 없는 추가 인물의 신체(배경 양쪽의 발과 다리)가 화면에 포함됨",
     "[gpt-high] 왼쪽 배경에 다른 사람의 다리와 신발이 추가되어, 임정현 외에는 신체 일부도 등장시키지 말라는 인물 제한을 위반합니다."
    ],
    "physics": "바닥에 누워 신체를 지탱하고 있습니다."
   }
  ],
  "cross_model_order": {
   "policy": "cross_model_order_v2_slot0_fwd_slot1_rev_parallel2",
   "slots": [
    {
     "model": "gemini-pro",
     "order": "forward",
     "display_to_canonical": {
      "A": "A",
      "B": "B"
     },
     "raw": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 7,
        "verdict_ko": "참조 이미지의 고정된 인물 상태와 바닥의 디테일(쌀알, 파편)을 정확히 유지하며 요구된 클로즈업을 사실적으로 구현했습니다."
       },
       {
        "label": "B",
        "score": 3,
        "verdict_ko": "지시문에 명시되지 않은 다른 사람의 신체 일부(발, 다리)가 배경에 등장하여 치명적인 오류를 범했습니다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "두 눈을 뜨고 위쪽 허공을 응시하고 있습니다.",
        "built_space": "흙바닥 위에 쌀알과 깨진 그릇이 흩어져 있으며, 뒤쪽으로 오두막의 엮은 벽면이 보입니다.",
        "entities": "임정현의 얼굴 특징, 상처, 옷차림이 참조 이미지와 완벽하게 일치합니다.",
        "hard_violations": [],
        "physics": "머리와 몸통이 흙바닥에 안정적으로 지탱되어 누워 있습니다."
       },
       {
        "label": "B",
        "direction": "두 눈을 뜨고 위쪽을 응시하고 있습니다.",
        "built_space": "바닥에 칼과 파편이 있으며, 배경에 돌무더기와 불탄 나무 구조물이 있습니다.",
        "entities": "임정현의 얼굴에 참조에 없던 눈물 자국이 추가되었으며, 인상도 약간 다릅니다.",
        "hard_violations": [
         "프롬프트에 없는 추가 인물의 신체(배경 양쪽의 발과 다리)가 화면에 포함됨"
        ],
        "physics": "바닥에 누워 신체를 지탱하고 있습니다."
       }
      ],
      "all_candidates_fail": false
     },
     "normalized": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 7,
        "verdict_ko": "참조 이미지의 고정된 인물 상태와 바닥의 디테일(쌀알, 파편)을 정확히 유지하며 요구된 클로즈업을 사실적으로 구현했습니다."
       },
       {
        "label": "B",
        "score": 3,
        "verdict_ko": "지시문에 명시되지 않은 다른 사람의 신체 일부(발, 다리)가 배경에 등장하여 치명적인 오류를 범했습니다."
       }
      ],
      "all_candidates_fail": false,
      "readings": [
       {
        "label": "A",
        "direction": "두 눈을 뜨고 위쪽 허공을 응시하고 있습니다.",
        "built_space": "흙바닥 위에 쌀알과 깨진 그릇이 흩어져 있으며, 뒤쪽으로 오두막의 엮은 벽면이 보입니다.",
        "entities": "임정현의 얼굴 특징, 상처, 옷차림이 참조 이미지와 완벽하게 일치합니다.",
        "hard_violations": [],
        "physics": "머리와 몸통이 흙바닥에 안정적으로 지탱되어 누워 있습니다."
       },
       {
        "label": "B",
        "direction": "두 눈을 뜨고 위쪽을 응시하고 있습니다.",
        "built_space": "바닥에 칼과 파편이 있으며, 배경에 돌무더기와 불탄 나무 구조물이 있습니다.",
        "entities": "임정현의 얼굴에 참조에 없던 눈물 자국이 추가되었으며, 인상도 약간 다릅니다.",
        "hard_violations": [
         "프롬프트에 없는 추가 인물의 신체(배경 양쪽의 발과 다리)가 화면에 포함됨"
        ],
        "physics": "바닥에 누워 신체를 지탱하고 있습니다."
       }
      ]
     },
     "ok": true
    },
    {
     "model": "gpt-high",
     "order": "reverse",
     "display_to_canonical": {
      "A": "B",
      "B": "A"
     },
     "raw": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 3,
        "verdict_ko": "뜬눈과 지면에 닿은 머리는 맞지만, 배경에 다른 사람의 다리와 신발이 추가되어 인물 제한을 위반하며 폐허가 얼굴 주변의 절제된 흙 테두리보다 크게 드러납니다."
       },
       {
        "label": "B",
        "score": 8,
        "verdict_ko": "임정현의 얼굴·복장·뜬눈·지면 접촉 자세를 가장 충실히 유지하지만, 참고 사진의 구도를 거의 반복하고 오두막 배경을 넓게 남겨 요구된 얼굴 중심 클로즈업에는 덜 밀착합니다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "두 눈은 열려 있고 시선은 화면 왼쪽 위의 허공을 향하며, 응시하는 구체적인 대상은 없습니다. 시신의 무초점 시선으로 읽힙니다. 오른쪽 가장자리의 칼은 흙 위에 대각선으로 놓여 있으며 사람을 겨누거나 사용되는 상태가 아닙니다.",
        "built_space": "머리 뒤로 왼쪽과 오른쪽의 무너진 돌무더기, 그 사이를 가로지르는 탄 목재 한 줄과 여러 기둥 잔해가 보입니다. 참고 사진에서 두드러지는 목조·엮은 벽 오두막 대신 석재와 탄 골조가 배경 대부분을 차지합니다. 카메라 각도 변경은 허용되지만 같은 장소임을 확인할 고정 구조의 연속성은 약합니다.",
        "entities": "전경에는 검은 머리의 성인 동아시아계 남성 한 명이 있고 한국인 임정현이라는 설정과 외형상 모순되지는 않습니다. 짙은 누비옷과 흙 묻은 질감은 유사하지만 얼굴 윤곽과 머리 정돈 상태는 참고 인물과 차이가 있습니다. 두 눈의 홍채와 동공은 정상적으로 보이며 얼굴은 창백합니다. 흙, 쌀알, 깨진 그릇, 혈흔과 칼이 보입니다. 왼쪽 뒤에는 별도 인물의 것으로 읽히는 바지 입은 다리와 신발들이 있습니다. 읽을 수 있는 글자는 없습니다.",
        "hard_violations": [
         "왼쪽 배경에 다른 사람의 다리와 신발이 추가되어, 임정현 외에는 신체 일부도 등장시키지 말라는 인물 제한을 위반합니다."
        ],
        "physics": "임정현의 뒤통수와 머리카락은 흙에 닿고 목과 어깨는 누운 몸통에 연결되어 있어 지면이 몸을 지지합니다. 부유하거나 비정상적으로 꺾인 신체는 보이지 않습니다. 칼과 도자기 조각은 흙 위에 놓여 있고, 배경의 다리와 신발도 지면에 놓인 것으로 보입니다."
       },
       {
        "label": "B",
        "direction": "두 눈을 뜬 채 카메라 쪽 허공을 바라보며 특정 대상을 따라가는 시선은 없습니다. 참고 사진의 눈동자 방향과 고개 기울기가 거의 그대로 유지되어 굳어 있는 시신이라는 순간에 부합합니다. 겨누거나 움직이는 물체는 없습니다.",
        "built_space": "뒤에는 어두운 출입구 한 곳과 이를 나누거나 둘러싼 목재 기둥, 좌우의 엮은 벽면이 보입니다. 참고 사진의 낡은 목재와 엮은 벽 재질은 유지되지만 출입구의 화면상 위치와 벽면 비중이 달라지고 오른쪽 돌무더기는 보이지 않습니다. 얼굴 아래와 왼쪽에는 흙바닥이 이어집니다. 구조물 배경이 상당히 넓어 얼굴을 둘러싼 절제된 흙 테두리라는 요구에는 다소 느슨합니다.",
        "entities": "참고 사진과 매우 유사한 검은 머리의 성인 동아시아계 남성 한 명만 보입니다. 얼굴 윤곽, 머리카락, 짙은 누비옷과 여밈, 흙 묻은 상태가 임정현의 외형 잠금에 잘 맞습니다. 두 눈은 정상적인 홍채와 동공을 유지한 채 열려 있고 피부는 자연스럽게 창백합니다. 주변의 쌀알, 깨진 그릇 조각과 혈흔도 참고 사진과 대응합니다. 칼·솥·감자는 이 크롭에서 확인되지 않으며 이를 보여주기 위해 화면을 넓힐 필요는 없습니다. 글자나 추가 인물은 없습니다.",
        "hard_violations": [],
        "physics": "뒤통수와 아래쪽 귀 주변이 흙바닥에 닿고 어깨와 몸통도 바닥에 누워 있어 지지 관계가 명확합니다. 참고 사진의 머리 기울기와 접촉 위치가 유지됩니다. 도자기 조각과 쌀알은 지면에 놓이고 혈흔도 흙 표면에 남아 있으며, 지지 없이 떠 있는 물체는 없습니다."
       }
      ],
      "all_candidates_fail": false
     },
     "normalized": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 3,
        "verdict_ko": "뜬눈과 지면에 닿은 머리는 맞지만, 배경에 다른 사람의 다리와 신발이 추가되어 인물 제한을 위반하며 폐허가 얼굴 주변의 절제된 흙 테두리보다 크게 드러납니다."
       },
       {
        "label": "A",
        "score": 8,
        "verdict_ko": "임정현의 얼굴·복장·뜬눈·지면 접촉 자세를 가장 충실히 유지하지만, 참고 사진의 구도를 거의 반복하고 오두막 배경을 넓게 남겨 요구된 얼굴 중심 클로즈업에는 덜 밀착합니다."
       }
      ],
      "all_candidates_fail": false,
      "readings": [
       {
        "label": "B",
        "direction": "두 눈은 열려 있고 시선은 화면 왼쪽 위의 허공을 향하며, 응시하는 구체적인 대상은 없습니다. 시신의 무초점 시선으로 읽힙니다. 오른쪽 가장자리의 칼은 흙 위에 대각선으로 놓여 있으며 사람을 겨누거나 사용되는 상태가 아닙니다.",
        "built_space": "머리 뒤로 왼쪽과 오른쪽의 무너진 돌무더기, 그 사이를 가로지르는 탄 목재 한 줄과 여러 기둥 잔해가 보입니다. 참고 사진에서 두드러지는 목조·엮은 벽 오두막 대신 석재와 탄 골조가 배경 대부분을 차지합니다. 카메라 각도 변경은 허용되지만 같은 장소임을 확인할 고정 구조의 연속성은 약합니다.",
        "entities": "전경에는 검은 머리의 성인 동아시아계 남성 한 명이 있고 한국인 임정현이라는 설정과 외형상 모순되지는 않습니다. 짙은 누비옷과 흙 묻은 질감은 유사하지만 얼굴 윤곽과 머리 정돈 상태는 참고 인물과 차이가 있습니다. 두 눈의 홍채와 동공은 정상적으로 보이며 얼굴은 창백합니다. 흙, 쌀알, 깨진 그릇, 혈흔과 칼이 보입니다. 왼쪽 뒤에는 별도 인물의 것으로 읽히는 바지 입은 다리와 신발들이 있습니다. 읽을 수 있는 글자는 없습니다.",
        "hard_violations": [
         "왼쪽 배경에 다른 사람의 다리와 신발이 추가되어, 임정현 외에는 신체 일부도 등장시키지 말라는 인물 제한을 위반합니다."
        ],
        "physics": "임정현의 뒤통수와 머리카락은 흙에 닿고 목과 어깨는 누운 몸통에 연결되어 있어 지면이 몸을 지지합니다. 부유하거나 비정상적으로 꺾인 신체는 보이지 않습니다. 칼과 도자기 조각은 흙 위에 놓여 있고, 배경의 다리와 신발도 지면에 놓인 것으로 보입니다."
       },
       {
        "label": "A",
        "direction": "두 눈을 뜬 채 카메라 쪽 허공을 바라보며 특정 대상을 따라가는 시선은 없습니다. 참고 사진의 눈동자 방향과 고개 기울기가 거의 그대로 유지되어 굳어 있는 시신이라는 순간에 부합합니다. 겨누거나 움직이는 물체는 없습니다.",
        "built_space": "뒤에는 어두운 출입구 한 곳과 이를 나누거나 둘러싼 목재 기둥, 좌우의 엮은 벽면이 보입니다. 참고 사진의 낡은 목재와 엮은 벽 재질은 유지되지만 출입구의 화면상 위치와 벽면 비중이 달라지고 오른쪽 돌무더기는 보이지 않습니다. 얼굴 아래와 왼쪽에는 흙바닥이 이어집니다. 구조물 배경이 상당히 넓어 얼굴을 둘러싼 절제된 흙 테두리라는 요구에는 다소 느슨합니다.",
        "entities": "참고 사진과 매우 유사한 검은 머리의 성인 동아시아계 남성 한 명만 보입니다. 얼굴 윤곽, 머리카락, 짙은 누비옷과 여밈, 흙 묻은 상태가 임정현의 외형 잠금에 잘 맞습니다. 두 눈은 정상적인 홍채와 동공을 유지한 채 열려 있고 피부는 자연스럽게 창백합니다. 주변의 쌀알, 깨진 그릇 조각과 혈흔도 참고 사진과 대응합니다. 칼·솥·감자는 이 크롭에서 확인되지 않으며 이를 보여주기 위해 화면을 넓힐 필요는 없습니다. 글자나 추가 인물은 없습니다.",
        "hard_violations": [],
        "physics": "뒤통수와 아래쪽 귀 주변이 흙바닥에 닿고 어깨와 몸통도 바닥에 누워 있어 지지 관계가 명확합니다. 참고 사진의 머리 기울기와 접촉 위치가 유지됩니다. 도자기 조각과 쌀알은 지면에 놓이고 혈흔도 흙 표면에 남아 있으며, 지지 없이 떠 있는 물체는 없습니다."
       }
      ]
     },
     "ok": true
    }
   ],
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "slot_winner_match": true,
   "slot_winner": {
    "gemini-pro": "A",
    "gpt-high": "A"
   },
   "route": "cross_slot_agree"
  },
  "dual": {
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "normalized": {
    "A": 2.0,
    "B": 0.804
   },
   "adjusted": {
    "A": 2.0,
    "B": 0.554
   },
   "violations": {
    "B": [
     "[gemini-pro] 프롬프트에 없는 추가 인물의 신체(배경 양쪽의 발과 다리)가 화면에 포함됨",
     "[gpt-high] 왼쪽 배경에 다른 사람의 다리와 신발이 추가되어, 임정현 외에는 신체 일부도 등장시키지 말라는 인물 제한을 위반합니다."
    ]
   },
   "per_model_winner": {
    "gemini-pro": "A",
    "gpt-high": "A"
   },
   "agreed": true
  },
  "totals": {
   "A": 2000,
   "B": 554
  },
  "selected": "A",
  "ranking": [
   "A",
   "B"
  ],
  "verdicts": [
   {
    "label": "A",
    "score": 2000,
    "verdict_ko": "참조 이미지의 고정된 인물 상태와 바닥의 디테일(쌀알, 파편)을 정확히 유지하며 요구된 클로즈업을 사실적으로 구현했습니다."
   },
   {
    "label": "B",
    "score": 554,
    "verdict_ko": "지시문에 명시되지 않은 다른 사람의 신체 일부(발, 다리)가 배경에 등장하여 치명적인 오류를 범했습니다.  ★위반: [gemini-pro] 프롬프트에 없는 추가 인물의 신체(배경 양쪽의 발과 다리)가 화면에 포함됨 / [gpt-high] 왼쪽 배경에 다른 사람의 다리와 신발이 추가되어, 임정현 외에는 신체 일부도 등장시키지 말라는 인물 제한을 위반합니다."
   }
  ],
  "refs": [
   {
    "label": "PREVIOUS SHOT STILL — a visually related earlier shot of this same place: the location's look, materials, fixed features and lighting mood are LOCKED to this photo; never copy its camera framing. The clothing, hair and overall look of 임정현 — who appear both in that photo and in this shot — are LOCKED to that photo. Anyone else visible in that photo is NOT in this shot: never carry their face, body or clothing onto anyone here. If this photo shows a character who cannot move (dead or unconscious), that character's exact pose, position and orientation are ALSO LOCKED — treat their body as a fixed prop of the set that only the camera moves around.",
    "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/db00517b-165b-4005-a287-f28936b59b95/images/272bd7af-9294-4490-98c6-88a40e9deacc/scene/recipe/S58sh3_sel.png",
    "asset_id": "9387ab21-88fc-4951-8324-6eb7152a36aa",
    "role": "prev_still"
   }
  ],
  "critique_skipped": true,
  "shot_run_produced": false,
  "shot_run_uid": "06aa1446-6adc-7033-b553-0ada5e8b3634",
  "ref_mode": "prev만 (배경 전용·공유 계획)",
  "share_plan": {
   "ref_plan": "prev",
   "prev_anchor_tag": "S58sh3"
  },
  "lane_policy": "share_plan_prev_bgonly",
  "locked_char_refs_excluded": [
   "임정현(C15)"
  ]
 },
 "S64sh3::cine": {
  "applied": true,
  "attempted_at": "2026-09-09T08:25:46.032423+00:00",
  "fingerprint": "340137ab9a9c266a7cfb21b734cab78593668d74fdd2fd93bacb485f2863777b",
  "fingerprint_version": 2,
  "provider": "mai",
  "endpoint": "openrouter/chat-completions",
  "model": "microsoft/mai-image-2.6",
  "pack": "24.202608252115",
  "source_file": "S64sh3_sel.png",
  "source_sha256": "c4f93deaebb9a720016c512b90349caa36db07381790f7caff3bf316d517dc7a",
  "file": "S64sh3_cine.png",
  "staged_sha256": "893aeea21a80714344fbdfa34b7c48d6fb139e419b965909344a376e48734603",
  "latency_ms": 29738
 },
 "S64sh10::signage": {
  "fp": "052bfa4e08b162aa",
  "inscriptions": [],
  "cues": [],
  "dropped": []
 },
 "S64sh10": {
  "input_fingerprint": "967abbab4608e165",
  "prompt": "Create ONE FINAL photorealistic live-action film still of the moment below — early 20th-century Korea, Jeolla region; all people are Korean unless explicitly stated otherwise. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 흙바닥에 넋을 잃고 주저앉은 복만을 묵묵히 내려다보는 안규홍의 가라앉은 시선.\n\nLOCATION (lock): The devastated outdoor clearing between ruined huts and trees in the mountain camp, where survivors regroup. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: medium shot\n- KEY BACKGROUND ELEMENTS: Earth between 복만 and 안규홍 (복만 remains seated on the ground as 안규홍 finishes rising); used as A narrow visible interval connects their different heights within the same space; Collapsed hut (Collapsed in the aftermath of the fighting) — Only a peripheral portion is visible beyond 안규홍; used as Provides subdued context without competing with the downward exchange.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Subdued natural daylight through the thinning gunpowder smoke sustains a quiet, desaturated aftermath without triumphant highlights.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The people visible in that photo are NOT in this shot: never carry their faces, bodies or clothing onto anyone here. Each person in THIS shot is defined solely by the PEOPLE list, their own CHARACTER REFERENCE images and their wardrobe notes. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): Smoke thins over the ruined huts, overturned pot, shattered dishes and food mixed with blood and dirt. 안규홍: He has risen with his bloodied sickle; his face is blood- and dirt-covered and his back wound still bleeds.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 안규홍 (한국인 남성, 청년 나이대, 검은 머리, 단단히 묶은 머리) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.",
  "roll_prompts": {
   "A": "Create ONE FINAL photorealistic live-action film still of the moment below — early 20th-century Korea, Jeolla region; all people are Korean unless explicitly stated otherwise. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 흙바닥에 넋을 잃고 주저앉은 복만을 묵묵히 내려다보는 안규홍의 가라앉은 시선.\n\nLOCATION (lock): The devastated outdoor clearing between ruined huts and trees in the mountain camp, where survivors regroup. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: medium shot\n- KEY BACKGROUND ELEMENTS: Earth between 복만 and 안규홍 (복만 remains seated on the ground as 안규홍 finishes rising); used as A narrow visible interval connects their different heights within the same space; Collapsed hut (Collapsed in the aftermath of the fighting) — Only a peripheral portion is visible beyond 안규홍; used as Provides subdued context without competing with the downward exchange.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Subdued natural daylight through the thinning gunpowder smoke sustains a quiet, desaturated aftermath without triumphant highlights.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The people visible in that photo are NOT in this shot: never carry their faces, bodies or clothing onto anyone here. Each person in THIS shot is defined solely by the PEOPLE list, their own CHARACTER REFERENCE images and their wardrobe notes. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): Smoke thins over the ruined huts, overturned pot, shattered dishes and food mixed with blood and dirt. 안규홍: He has risen with his bloodied sickle; his face is blood- and dirt-covered and his back wound still bleeds.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 안규홍 (한국인 남성, 청년 나이대, 검은 머리, 단단히 묶은 머리) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.",
   "B": "Create ONE FINAL photorealistic live-action film still of the moment below — early 20th-century Korea, Jeolla region; all people are Korean unless explicitly stated otherwise. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 흙바닥에 넋을 잃고 주저앉은 복만을 묵묵히 내려다보는 안규홍의 가라앉은 시선.\n\nLOCATION (lock): The devastated outdoor clearing between ruined huts and trees in the mountain camp, where survivors regroup. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: medium shot\n- KEY BACKGROUND ELEMENTS: Earth between 복만 and 안규홍 (복만 remains seated on the ground as 안규홍 finishes rising); used as A narrow visible interval connects their different heights within the same space; Collapsed hut (Collapsed in the aftermath of the fighting) — Only a peripheral portion is visible beyond 안규홍; used as Provides subdued context without competing with the downward exchange.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Subdued natural daylight through the thinning gunpowder smoke sustains a quiet, desaturated aftermath without triumphant highlights.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The people visible in that photo are NOT in this shot: never carry their faces, bodies or clothing onto anyone here. Each person in THIS shot is defined solely by the PEOPLE list, their own CHARACTER REFERENCE images and their wardrobe notes. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): Smoke thins over the ruined huts, overturned pot, shattered dishes and food mixed with blood and dirt. 안규홍: He has risen with his bloodied sickle; his face is blood- and dirt-covered and his back wound still bleeds.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 안규홍 (한국인 남성, 청년 나이대, 검은 머리, 단단히 묶은 머리) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.\n\nCOMPOSITION VARIATION (this candidate only): keep the place, the people and the moment exactly as specified, and keep from the CAMERA & FRAME contract its subject scale, screen placement and key background elements exactly. For this candidate only, its camera angle and height are deliberately left open: choose a camera position distinctly different from the obvious one, as a film director picking a second setup on the same blocking. Stay true to any explicit framing the SHOT TEXT itself dictates."
  },
  "shot_run_spend_attempt_count": 1,
  "readings": [
   {
    "label": "A",
    "direction": "안규홍은 바닥에 주저앉은 복만을 내려다보고 있으며, 복만은 고개를 숙인 채 시선을 바닥에 두고 있다.",
    "built_space": "부서진 초가집과 나무 울타리가 있는 야외 흙바닥이다. 레퍼런스에 있던 엎어진 가마솥과 쏟아진 밥, 식기류가 바닥에 배치되어 있으나, 지시와 달리 다수의 시체들이 공간 곳곳에 널려 있다.",
    "entities": "안규홍과 복만이 존재하며 안규홍은 피 묻은 낫을 들고 있다. 그러나 샷 텍스트에 없는 다수의 인물(시체)들이 배경에 무더기로 추가되어 있다.",
    "hard_violations": [
     "[gemini-pro] 프롬프트에 등장하지 않는 다수의 인물(시체)들이 배경에 추가됨",
     "[gpt-high] 이번 샷에 등장하지 말아야 할 쓰러진 배경 인물들을 다수 추가했다. 이전 사진의 사람들을 이 장면에 가져오지 말라는 명시적 제한을 위반한다."
    ],
    "physics": "안규홍은 두 발로 서 있고 복만은 바닥에 앉아 있다. 안규홍이 쥔 낫은 손에 의해 정상적으로 지탱된다."
   },
   {
    "label": "B",
    "direction": "안규홍은 넋을 잃고 주저앉은 복만을 향해 시선을 아래로 향하고 있으며, 복만은 멍한 표정으로 바닥 쪽을 바라보고 있다.",
    "built_space": "부서진 초가집과 나무 울타리가 배경에 있는 야외 흙바닥이다. 프롬프트에 명시된 엎어진 가마솥과 식기류는 누락되었으나, 공간의 전반적인 구조는 레퍼런스를 따른다.",
    "entities": "안규홍은 레퍼런스와 일치하는 복장과 머리 모양을 하고 있으며, 피 묻은 낫을 들고 등에 상처가 있다. 복만은 흙바닥에 주저앉아 있다. 지시대로 추가된 인물이나 시체는 전혀 없다.",
    "hard_violations": [],
    "physics": "안규홍은 발로 땅을 딛고 서서 몸을 숙이고 있으며, 복만은 흙바닥에 안정적으로 앉아 있다. 안규홍의 손은 낫의 자루를 제대로 쥐고 있다."
   }
  ],
  "cross_model_order": {
   "policy": "cross_model_order_v2_slot0_fwd_slot1_rev_parallel2",
   "slots": [
    {
     "model": "gemini-pro",
     "order": "forward",
     "display_to_canonical": {
      "A": "A",
      "B": "B"
     },
     "raw": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 2,
        "verdict_ko": "이전 샷의 인물(시체)들을 절대 포함하지 말라는 프롬프트의 엄격한 지시를 어기고 다수의 시체를 배경에 배치한 치명적인 오류가 있습니다."
       },
       {
        "label": "B",
        "score": 8,
        "verdict_ko": "엎어진 가마솥 등 일부 배경 소품이 누락되었으나, 명시된 인물만 화면에 남기는 지시를 정확히 따랐으며 두 인물의 액션과 구도를 매우 충실하게 구현했습니다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "안규홍은 바닥에 주저앉은 복만을 내려다보고 있으며, 복만은 고개를 숙인 채 시선을 바닥에 두고 있다.",
        "built_space": "부서진 초가집과 나무 울타리가 있는 야외 흙바닥이다. 레퍼런스에 있던 엎어진 가마솥과 쏟아진 밥, 식기류가 바닥에 배치되어 있으나, 지시와 달리 다수의 시체들이 공간 곳곳에 널려 있다.",
        "entities": "안규홍과 복만이 존재하며 안규홍은 피 묻은 낫을 들고 있다. 그러나 샷 텍스트에 없는 다수의 인물(시체)들이 배경에 무더기로 추가되어 있다.",
        "hard_violations": [
         "프롬프트에 등장하지 않는 다수의 인물(시체)들이 배경에 추가됨"
        ],
        "physics": "안규홍은 두 발로 서 있고 복만은 바닥에 앉아 있다. 안규홍이 쥔 낫은 손에 의해 정상적으로 지탱된다."
       },
       {
        "label": "B",
        "direction": "안규홍은 넋을 잃고 주저앉은 복만을 향해 시선을 아래로 향하고 있으며, 복만은 멍한 표정으로 바닥 쪽을 바라보고 있다.",
        "built_space": "부서진 초가집과 나무 울타리가 배경에 있는 야외 흙바닥이다. 프롬프트에 명시된 엎어진 가마솥과 식기류는 누락되었으나, 공간의 전반적인 구조는 레퍼런스를 따른다.",
        "entities": "안규홍은 레퍼런스와 일치하는 복장과 머리 모양을 하고 있으며, 피 묻은 낫을 들고 등에 상처가 있다. 복만은 흙바닥에 주저앉아 있다. 지시대로 추가된 인물이나 시체는 전혀 없다.",
        "hard_violations": [],
        "physics": "안규홍은 발로 땅을 딛고 서서 몸을 숙이고 있으며, 복만은 흙바닥에 안정적으로 앉아 있다. 안규홍의 손은 낫의 자루를 제대로 쥐고 있다."
       }
      ],
      "all_candidates_fail": false
     },
     "normalized": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 2,
        "verdict_ko": "이전 샷의 인물(시체)들을 절대 포함하지 말라는 프롬프트의 엄격한 지시를 어기고 다수의 시체를 배경에 배치한 치명적인 오류가 있습니다."
       },
       {
        "label": "B",
        "score": 8,
        "verdict_ko": "엎어진 가마솥 등 일부 배경 소품이 누락되었으나, 명시된 인물만 화면에 남기는 지시를 정확히 따랐으며 두 인물의 액션과 구도를 매우 충실하게 구현했습니다."
       }
      ],
      "all_candidates_fail": false,
      "readings": [
       {
        "label": "A",
        "direction": "안규홍은 바닥에 주저앉은 복만을 내려다보고 있으며, 복만은 고개를 숙인 채 시선을 바닥에 두고 있다.",
        "built_space": "부서진 초가집과 나무 울타리가 있는 야외 흙바닥이다. 레퍼런스에 있던 엎어진 가마솥과 쏟아진 밥, 식기류가 바닥에 배치되어 있으나, 지시와 달리 다수의 시체들이 공간 곳곳에 널려 있다.",
        "entities": "안규홍과 복만이 존재하며 안규홍은 피 묻은 낫을 들고 있다. 그러나 샷 텍스트에 없는 다수의 인물(시체)들이 배경에 무더기로 추가되어 있다.",
        "hard_violations": [
         "프롬프트에 등장하지 않는 다수의 인물(시체)들이 배경에 추가됨"
        ],
        "physics": "안규홍은 두 발로 서 있고 복만은 바닥에 앉아 있다. 안규홍이 쥔 낫은 손에 의해 정상적으로 지탱된다."
       },
       {
        "label": "B",
        "direction": "안규홍은 넋을 잃고 주저앉은 복만을 향해 시선을 아래로 향하고 있으며, 복만은 멍한 표정으로 바닥 쪽을 바라보고 있다.",
        "built_space": "부서진 초가집과 나무 울타리가 배경에 있는 야외 흙바닥이다. 프롬프트에 명시된 엎어진 가마솥과 식기류는 누락되었으나, 공간의 전반적인 구조는 레퍼런스를 따른다.",
        "entities": "안규홍은 레퍼런스와 일치하는 복장과 머리 모양을 하고 있으며, 피 묻은 낫을 들고 등에 상처가 있다. 복만은 흙바닥에 주저앉아 있다. 지시대로 추가된 인물이나 시체는 전혀 없다.",
        "hard_violations": [],
        "physics": "안규홍은 발로 땅을 딛고 서서 몸을 숙이고 있으며, 복만은 흙바닥에 안정적으로 앉아 있다. 안규홍의 손은 낫의 자루를 제대로 쥐고 있다."
       }
      ]
     },
     "ok": true
    },
    {
     "model": "gpt-high",
     "order": "reverse",
     "display_to_canonical": {
      "A": "B",
      "B": "A"
     },
     "raw": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 6,
        "verdict_ko": "복만을 향한 안규홍의 낮은 시선과 두 사람만의 정적은 맞지만, 전신에 가까운 넓은 구도와 크게 드러난 움막은 지정된 미디엄 샷에서 벗어난다."
       },
       {
        "label": "B",
        "score": 2,
        "verdict_ko": "일어선 안규홍과 주저앉은 복만의 관계는 보이나, 금지된 배경 인물들을 다수 추가했고 움막과 주변 참상을 넓게 보여 핵심 구도를 어겼다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "오른쪽 안규홍은 상체와 고개를 왼쪽 아래로 숙이고, 시선을 앉아 있는 복만의 얼굴 쪽에 둔다. 복만은 고개를 떨군 채 무릎 앞 흙바닥을 본다. 손에 든 낫의 날끝은 아래쪽 땅을 향하며 복만을 겨누는 동작은 아니다.",
        "built_space": "왼쪽 배경에 지붕과 벽체가 상당 부분 보이는 움막 한 채, 오른쪽 가장자리에 다른 움막 일부, 뒤쪽에 목책과 나무들이 있다. 복만은 왼쪽 전경 흙바닥에 앉고 안규홍은 오른쪽에 서 있어 그 사이 땅이 보인다. 다만 두 사람의 하체까지 거의 모두 담은 구도이고, 움막도 주변부의 작은 파편이 아니라 배경의 큰 면적을 차지한다. 흙, 목재, 짚과 옅은 연기는 장소 참고의 재질 및 낮 분위기와 대체로 맞는다.",
        "entities": "인물은 두 명이다. 안규홍은 검은 머리를 묶은 청년 남성으로, 참고와 대체로 유사한 얼굴 윤곽 및 체격, 누런 흰색 저고리와 바지, 천 허리띠를 보인다. 얼굴과 옷에는 피와 흙이 묻었고 등에도 혈흔이 있으며 손에는 피 묻은 낫이 있다. 복만은 흐트러진 짧은 머리의 중년 남성으로 표현되어 있고, 흙 묻은 한복 차림과 멍한 표정이 보인다. 두 사람 모두 한국인 시대극 설정과 외형상 부합한다. 왼쪽 아래에는 솥 테두리 일부, 땅에는 그릇 파편과 목재 잔해가 보인다. 읽을 수 있는 글자는 없다.",
        "hard_violations": [],
        "physics": "복만은 엉덩이와 접거나 뻗은 다리를 흙바닥에 대고 앉아 있으며 손은 무릎과 허벅지 쪽에 놓여 있다. 안규홍은 지면에 닿은 발과 다리로 몸을 지탱하면서 허리를 굽힌 상태로, 일어서기를 마무리하는 중간 자세로는 가능하지만 완전히 일어선 모습은 아니다. 낫의 손잡이는 손으로 확실히 쥐고 있다. 지지 없이 떠 있는 몸이나 물체는 보이지 않는다."
       },
       {
        "label": "B",
        "direction": "안규홍은 오른쪽에서 고개를 왼쪽 아래로 내려 복만 쪽을 바라본다. 복만은 자신의 무릎 앞 땅으로 시선을 떨군다. 안규홍의 낫은 허벅지 옆에 내려져 있고 날끝은 지면을 향한다. 공격이나 이동 동작은 보이지 않는다.",
        "built_space": "중앙 배경에는 지붕과 정면이 크게 드러난 움막 한 채가 있고, 양쪽 가장자리에 다른 움막의 일부와 뒤편 목책이 보인다. 중앙 움막 앞에는 선반 또는 탁자 하나가 있다. 복만은 왼쪽 전경에 앉고 안규홍은 오른쪽에 서 있다. 두 사람 사이에는 피가 고인 길쭉한 흙바닥이 넓게 노출된다. 안규홍 뒤에는 넘어진 솥 하나와 여러 그릇, 쏟아진 밥이 있다. 요청한 좁은 간격과 주변부 움막 대신 야영지와 배경 인물들을 폭넓게 담았으며, 이전 사진의 인물 배치를 그대로 유지한 모습도 아니다.",
        "entities": "주요 인물 외에 화면 양쪽과 뒤쪽에 쓰러진 남성들이 다수 보인다. 이는 이번 장면에 이전 사진의 사람들을 가져오지 말라는 지시와 충돌한다. 안규홍은 묶은 검은 머리, 청년 남성의 외형, 참고와 유사한 밝은 한복 및 천 허리띠를 보인다. 얼굴과 등에는 뚜렷한 혈흔이 있고 손에는 피 묻은 낫이 있다. 복만은 짧은 검은 머리의 성인 남성으로 흙바닥에 주저앉아 있다. 솥, 깨진 그릇, 밥과 피 묻은 흙은 보이지만, 이들을 많이 담은 것이 지정 구도 충실성을 높이지는 않는다. 읽을 수 있는 글자는 없다.",
        "hard_violations": [
         "이번 샷에 등장하지 말아야 할 쓰러진 배경 인물들을 다수 추가했다. 이전 사진의 사람들을 이 장면에 가져오지 말라는 명시적 제한을 위반한다."
        ],
        "physics": "복만의 엉덩이와 다리는 지면에 놓이고 양손은 다리 사이에 내려져 있어 앉은 자세의 지지가 자연스럽다. 안규홍은 곧게 일어선 자세이며 발은 화면 아래로 잘리지만 보이는 다리와 몸통에 부유를 시사하는 징후는 없다. 낫은 손잡이를 쥔 손으로 지탱된다. 배경의 쓰러진 인물들과 솥, 그릇도 땅에 닿아 있으며 떠 있는 물체는 보이지 않는다."
       }
      ],
      "all_candidates_fail": true
     },
     "normalized": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 6,
        "verdict_ko": "복만을 향한 안규홍의 낮은 시선과 두 사람만의 정적은 맞지만, 전신에 가까운 넓은 구도와 크게 드러난 움막은 지정된 미디엄 샷에서 벗어난다."
       },
       {
        "label": "A",
        "score": 2,
        "verdict_ko": "일어선 안규홍과 주저앉은 복만의 관계는 보이나, 금지된 배경 인물들을 다수 추가했고 움막과 주변 참상을 넓게 보여 핵심 구도를 어겼다."
       }
      ],
      "all_candidates_fail": true,
      "readings": [
       {
        "label": "B",
        "direction": "오른쪽 안규홍은 상체와 고개를 왼쪽 아래로 숙이고, 시선을 앉아 있는 복만의 얼굴 쪽에 둔다. 복만은 고개를 떨군 채 무릎 앞 흙바닥을 본다. 손에 든 낫의 날끝은 아래쪽 땅을 향하며 복만을 겨누는 동작은 아니다.",
        "built_space": "왼쪽 배경에 지붕과 벽체가 상당 부분 보이는 움막 한 채, 오른쪽 가장자리에 다른 움막 일부, 뒤쪽에 목책과 나무들이 있다. 복만은 왼쪽 전경 흙바닥에 앉고 안규홍은 오른쪽에 서 있어 그 사이 땅이 보인다. 다만 두 사람의 하체까지 거의 모두 담은 구도이고, 움막도 주변부의 작은 파편이 아니라 배경의 큰 면적을 차지한다. 흙, 목재, 짚과 옅은 연기는 장소 참고의 재질 및 낮 분위기와 대체로 맞는다.",
        "entities": "인물은 두 명이다. 안규홍은 검은 머리를 묶은 청년 남성으로, 참고와 대체로 유사한 얼굴 윤곽 및 체격, 누런 흰색 저고리와 바지, 천 허리띠를 보인다. 얼굴과 옷에는 피와 흙이 묻었고 등에도 혈흔이 있으며 손에는 피 묻은 낫이 있다. 복만은 흐트러진 짧은 머리의 중년 남성으로 표현되어 있고, 흙 묻은 한복 차림과 멍한 표정이 보인다. 두 사람 모두 한국인 시대극 설정과 외형상 부합한다. 왼쪽 아래에는 솥 테두리 일부, 땅에는 그릇 파편과 목재 잔해가 보인다. 읽을 수 있는 글자는 없다.",
        "hard_violations": [],
        "physics": "복만은 엉덩이와 접거나 뻗은 다리를 흙바닥에 대고 앉아 있으며 손은 무릎과 허벅지 쪽에 놓여 있다. 안규홍은 지면에 닿은 발과 다리로 몸을 지탱하면서 허리를 굽힌 상태로, 일어서기를 마무리하는 중간 자세로는 가능하지만 완전히 일어선 모습은 아니다. 낫의 손잡이는 손으로 확실히 쥐고 있다. 지지 없이 떠 있는 몸이나 물체는 보이지 않는다."
       },
       {
        "label": "A",
        "direction": "안규홍은 오른쪽에서 고개를 왼쪽 아래로 내려 복만 쪽을 바라본다. 복만은 자신의 무릎 앞 땅으로 시선을 떨군다. 안규홍의 낫은 허벅지 옆에 내려져 있고 날끝은 지면을 향한다. 공격이나 이동 동작은 보이지 않는다.",
        "built_space": "중앙 배경에는 지붕과 정면이 크게 드러난 움막 한 채가 있고, 양쪽 가장자리에 다른 움막의 일부와 뒤편 목책이 보인다. 중앙 움막 앞에는 선반 또는 탁자 하나가 있다. 복만은 왼쪽 전경에 앉고 안규홍은 오른쪽에 서 있다. 두 사람 사이에는 피가 고인 길쭉한 흙바닥이 넓게 노출된다. 안규홍 뒤에는 넘어진 솥 하나와 여러 그릇, 쏟아진 밥이 있다. 요청한 좁은 간격과 주변부 움막 대신 야영지와 배경 인물들을 폭넓게 담았으며, 이전 사진의 인물 배치를 그대로 유지한 모습도 아니다.",
        "entities": "주요 인물 외에 화면 양쪽과 뒤쪽에 쓰러진 남성들이 다수 보인다. 이는 이번 장면에 이전 사진의 사람들을 가져오지 말라는 지시와 충돌한다. 안규홍은 묶은 검은 머리, 청년 남성의 외형, 참고와 유사한 밝은 한복 및 천 허리띠를 보인다. 얼굴과 등에는 뚜렷한 혈흔이 있고 손에는 피 묻은 낫이 있다. 복만은 짧은 검은 머리의 성인 남성으로 흙바닥에 주저앉아 있다. 솥, 깨진 그릇, 밥과 피 묻은 흙은 보이지만, 이들을 많이 담은 것이 지정 구도 충실성을 높이지는 않는다. 읽을 수 있는 글자는 없다.",
        "hard_violations": [
         "이번 샷에 등장하지 말아야 할 쓰러진 배경 인물들을 다수 추가했다. 이전 사진의 사람들을 이 장면에 가져오지 말라는 명시적 제한을 위반한다."
        ],
        "physics": "복만의 엉덩이와 다리는 지면에 놓이고 양손은 다리 사이에 내려져 있어 앉은 자세의 지지가 자연스럽다. 안규홍은 곧게 일어선 자세이며 발은 화면 아래로 잘리지만 보이는 다리와 몸통에 부유를 시사하는 징후는 없다. 낫은 손잡이를 쥔 손으로 지탱된다. 배경의 쓰러진 인물들과 솥, 그릇도 땅에 닿아 있으며 떠 있는 물체는 보이지 않는다."
       }
      ]
     },
     "ok": true
    }
   ],
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "slot_winner_match": true,
   "slot_winner": {
    "gemini-pro": "B",
    "gpt-high": "B"
   },
   "route": "cross_slot_agree"
  },
  "dual": {
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "normalized": {
    "A": 0.583,
    "B": 2.0
   },
   "adjusted": {
    "A": 0.333,
    "B": 2.0
   },
   "violations": {
    "A": [
     "[gemini-pro] 프롬프트에 등장하지 않는 다수의 인물(시체)들이 배경에 추가됨",
     "[gpt-high] 이번 샷에 등장하지 말아야 할 쓰러진 배경 인물들을 다수 추가했다. 이전 사진의 사람들을 이 장면에 가져오지 말라는 명시적 제한을 위반한다."
    ]
   },
   "per_model_winner": {
    "gemini-pro": "B",
    "gpt-high": "B"
   },
   "agreed": true
  },
  "totals": {
   "A": 333,
   "B": 2000
  },
  "selected": "B",
  "ranking": [
   "B",
   "A"
  ],
  "verdicts": [
   {
    "label": "A",
    "score": 333,
    "verdict_ko": "이전 샷의 인물(시체)들을 절대 포함하지 말라는 프롬프트의 엄격한 지시를 어기고 다수의 시체를 배경에 배치한 치명적인 오류가 있습니다.  ★위반: [gemini-pro] 프롬프트에 등장하지 않는 다수의 인물(시체)들이 배경에 추가됨 / [gpt-high] 이번 샷에 등장하지 말아야 할 쓰러진 배경 인물들을 다수 추가했다. 이전 사진의 사람들을 이 장면에 가져오지 말라는 명시적 제한을 위반한다."
   },
   {
    "label": "B",
    "score": 2000,
    "verdict_ko": "엎어진 가마솥 등 일부 배경 소품이 누락되었으나, 명시된 인물만 화면에 남기는 지시를 정확히 따랐으며 두 인물의 액션과 구도를 매우 충실하게 구현했습니다."
   }
  ],
  "refs": [
   {
    "label": "PREVIOUS SHOT STILL — a visually related earlier shot of this same place: the location's look, materials, fixed features and lighting mood are LOCKED to this photo; never copy its camera framing. The people visible in that photo are NOT in this shot: never carry their faces, bodies or clothing onto anyone here. Each person in THIS shot is defined solely by the PEOPLE list, their own CHARACTER REFERENCE images and their wardrobe notes. If this photo shows a character who cannot move (dead or unconscious), that character's exact pose, position and orientation are ALSO LOCKED — treat their body as a fixed prop of the set that only the camera moves around.",
    "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/db00517b-165b-4005-a287-f28936b59b95/images/272bd7af-9294-4490-98c6-88a40e9deacc/scene/recipe/S60sh2_sel.png",
    "asset_id": "9a8a1465-bc86-433f-b821-e073cfb330f4",
    "role": "prev_still"
   },
   {
    "label": "CHARACTER REFERENCE — 안규홍: the exact person appearing in this shot; match face, hair and build exactly.",
    "path": "<bytes:1028384>",
    "asset_id": "3b5894db-ce55-4f34-b966-5e725c3aa952",
    "role": "character_ref"
   }
  ],
  "critique_skipped": true,
  "shot_run_produced": false,
  "shot_run_uid": "06aa1446-6c73-7e57-9ab0-40db58d46739",
  "ref_mode": "prev+엔티티",
  "share_plan": {
   "ref_plan": "prev",
   "prev_anchor_tag": "S60sh2"
  }
 },
 "S64sh10::cine": {
  "applied": true,
  "attempted_at": "2026-09-09T08:27:22.271729+00:00",
  "fingerprint": "10a4797f3ffef6610905626d4548a949a3116f6e36b814590b15dfcc1fc4212e",
  "fingerprint_version": 2,
  "provider": "mai",
  "endpoint": "openrouter/chat-completions",
  "model": "microsoft/mai-image-2.6",
  "pack": "24.202608252115",
  "source_file": "S64sh10_sel.png",
  "source_sha256": "7d6db2741d9ebb69eba1422ec8fd1d33f0adb2c75c79a5abf75be16b39b8b1de",
  "file": "S64sh10_cine.png",
  "staged_sha256": "a09c7e586f5787907b13020feb69a389ae4aff304e3d84054b04bcc7b795500c",
  "latency_ms": 29683
 },
 "S64sh18::signage": {
  "fp": "d22969015af16880",
  "inscriptions": [],
  "cues": [],
  "dropped": []
 },
 "S64sh18": {
  "input_fingerprint": "9bc42de716ada3e5",
  "prompt": "Create ONE FINAL photorealistic live-action film still of the moment below — early 20th-century Korea, Jeolla region; all people are Korean unless explicitly stated otherwise. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 아래쪽을 예의주시하며 미간이 매섭게 좁혀진 안규홍의 날카로운 눈빛 클로즈업.\n\nLOCATION (lock): The survivors' gathering spot in the ruined mountain-camp clearing, facing the lower forest approach. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: close-up\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Keep the established subdued daylight and restrained contrast, letting the tightened expression rather than a new light effect signal danger.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The clothing, hair and overall look of 안규홍 — who appear both in that photo and in this shot — are LOCKED to that photo. Anyone else visible in that photo is NOT in this shot: never carry their face, body or clothing onto anyone here. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The camp retains its burned and collapsed huts, overturned pot, broken bowls and bloodied food as the powder smoke clears. 안규홍: He stands holding the bloody sickle, with his back wound bleeding and his face coated in blood and dirt.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 안규홍 (한국인 남성, 청년 나이대, 검은 머리, 단단히 묶은 머리) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.",
  "roll_prompts": {
   "A": "Create ONE FINAL photorealistic live-action film still of the moment below — early 20th-century Korea, Jeolla region; all people are Korean unless explicitly stated otherwise. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 아래쪽을 예의주시하며 미간이 매섭게 좁혀진 안규홍의 날카로운 눈빛 클로즈업.\n\nLOCATION (lock): The survivors' gathering spot in the ruined mountain-camp clearing, facing the lower forest approach. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: close-up\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Keep the established subdued daylight and restrained contrast, letting the tightened expression rather than a new light effect signal danger.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The clothing, hair and overall look of 안규홍 — who appear both in that photo and in this shot — are LOCKED to that photo. Anyone else visible in that photo is NOT in this shot: never carry their face, body or clothing onto anyone here. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The camp retains its burned and collapsed huts, overturned pot, broken bowls and bloodied food as the powder smoke clears. 안규홍: He stands holding the bloody sickle, with his back wound bleeding and his face coated in blood and dirt.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 안규홍 (한국인 남성, 청년 나이대, 검은 머리, 단단히 묶은 머리) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.",
   "B": "Create ONE FINAL photorealistic live-action film still of the moment below — early 20th-century Korea, Jeolla region; all people are Korean unless explicitly stated otherwise. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 아래쪽을 예의주시하며 미간이 매섭게 좁혀진 안규홍의 날카로운 눈빛 클로즈업.\n\nLOCATION (lock): The survivors' gathering spot in the ruined mountain-camp clearing, facing the lower forest approach. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: close-up\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Keep the established subdued daylight and restrained contrast, letting the tightened expression rather than a new light effect signal danger.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The clothing, hair and overall look of 안규홍 — who appear both in that photo and in this shot — are LOCKED to that photo. Anyone else visible in that photo is NOT in this shot: never carry their face, body or clothing onto anyone here. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The camp retains its burned and collapsed huts, overturned pot, broken bowls and bloodied food as the powder smoke clears. 안규홍: He stands holding the bloody sickle, with his back wound bleeding and his face coated in blood and dirt.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 안규홍 (한국인 남성, 청년 나이대, 검은 머리, 단단히 묶은 머리) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.\n\nCOMPOSITION VARIATION (this candidate only): keep the place, the people and the moment exactly as specified, and keep from the CAMERA & FRAME contract its subject scale, screen placement and key background elements exactly. For this candidate only, its camera angle and height are deliberately left open: choose a camera position distinctly different from the obvious one, as a film director picking a second setup on the same blocking. Stay true to any explicit framing the SHOT TEXT itself dictates."
  },
  "shot_run_spend_attempt_count": 1,
  "readings": [
   {
    "label": "A",
    "direction": "인물은 렌즈 정면을 똑바로 응시하고 있으며 프롬프트가 지시한 아래쪽을 향하지 않습니다.",
    "built_space": "심한 아웃포커싱으로 인해 배경이 흐릿하며, 나무와 불분명한 구조물만 보여 레퍼런스의 캠프 장소인지 명확히 확인하기 어렵습니다.",
    "entities": "얼굴에 흙과 피가 묻은 안규홍의 모습이 보이며, 앞쪽에 낫을 쥐고 있는 손과 낫의 일부가 흐릿하게 나타납니다.",
    "hard_violations": [
     "[gemini-pro] 시선 방향 오류 (프롬프트의 '아래쪽을 예의주시하며'를 위반하고 카메라 정면을 응시함)"
    ],
    "physics": "손이 낫의 자루를 쥐고 있으며, 인물의 자세는 안정적입니다."
   },
   {
    "label": "B",
    "direction": "인물은 프롬프트의 지시대로 화면의 우측 아래쪽을 날카롭고 강렬하게 예의주시하고 있습니다.",
    "built_space": "배경에 이전 샷 레퍼런스에서 설정된 부서진 오두막, 울타리, 흩어진 가마솥과 파편들, 피어오르는 연기가 정확한 위치와 비례로 배치되어 있습니다.",
    "entities": "얼굴에 피와 흙이 묻은 안규홍의 모습이 레퍼런스와 일치하게 나타나며, 미간을 좁힌 날카로운 눈빛이 정확히 표현되었습니다. 클로즈업 프레이밍으로 인해 낫은 보이지 않으나 이는 허용됩니다.",
    "hard_violations": [],
    "physics": "서서 고개를 약간 숙이고 아래를 내려다보는 자연스러운 신체 균형을 유지하고 있습니다."
   }
  ],
  "cross_model_order": {
   "policy": "cross_model_order_v2_slot0_fwd_slot1_rev_parallel2",
   "slots": [
    {
     "model": "gemini-pro",
     "order": "forward",
     "display_to_canonical": {
      "A": "A",
      "B": "B"
     },
     "raw": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 10,
        "verdict_ko": "프롬프트가 요구한 대로 미간을 좁히고 아래쪽을 날카롭게 예주시하는 표정과 시선 방향을 완벽하게 구현했으며, 배경의 캠프 디테일도 레퍼런스와 일치합니다."
       },
       {
        "label": "A",
        "score": 3,
        "verdict_ko": "클로즈업 프레이밍과 인물의 외형은 비슷하게 구현했으나, 텍스트가 명시한 '아래쪽을 예의주시'하는 시선 대신 카메라 정면을 응시하여 핵심 지시를 위반했습니다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "인물은 렌즈 정면을 똑바로 응시하고 있으며 프롬프트가 지시한 아래쪽을 향하지 않습니다.",
        "built_space": "심한 아웃포커싱으로 인해 배경이 흐릿하며, 나무와 불분명한 구조물만 보여 레퍼런스의 캠프 장소인지 명확히 확인하기 어렵습니다.",
        "entities": "얼굴에 흙과 피가 묻은 안규홍의 모습이 보이며, 앞쪽에 낫을 쥐고 있는 손과 낫의 일부가 흐릿하게 나타납니다.",
        "hard_violations": [
         "시선 방향 오류 (프롬프트의 '아래쪽을 예의주시하며'를 위반하고 카메라 정면을 응시함)"
        ],
        "physics": "손이 낫의 자루를 쥐고 있으며, 인물의 자세는 안정적입니다."
       },
       {
        "label": "B",
        "direction": "인물은 프롬프트의 지시대로 화면의 우측 아래쪽을 날카롭고 강렬하게 예의주시하고 있습니다.",
        "built_space": "배경에 이전 샷 레퍼런스에서 설정된 부서진 오두막, 울타리, 흩어진 가마솥과 파편들, 피어오르는 연기가 정확한 위치와 비례로 배치되어 있습니다.",
        "entities": "얼굴에 피와 흙이 묻은 안규홍의 모습이 레퍼런스와 일치하게 나타나며, 미간을 좁힌 날카로운 눈빛이 정확히 표현되었습니다. 클로즈업 프레이밍으로 인해 낫은 보이지 않으나 이는 허용됩니다.",
        "hard_violations": [],
        "physics": "서서 고개를 약간 숙이고 아래를 내려다보는 자연스러운 신체 균형을 유지하고 있습니다."
       }
      ],
      "all_candidates_fail": false
     },
     "normalized": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 10,
        "verdict_ko": "프롬프트가 요구한 대로 미간을 좁히고 아래쪽을 날카롭게 예주시하는 표정과 시선 방향을 완벽하게 구현했으며, 배경의 캠프 디테일도 레퍼런스와 일치합니다."
       },
       {
        "label": "A",
        "score": 3,
        "verdict_ko": "클로즈업 프레이밍과 인물의 외형은 비슷하게 구현했으나, 텍스트가 명시한 '아래쪽을 예의주시'하는 시선 대신 카메라 정면을 응시하여 핵심 지시를 위반했습니다."
       }
      ],
      "all_candidates_fail": false,
      "readings": [
       {
        "label": "A",
        "direction": "인물은 렌즈 정면을 똑바로 응시하고 있으며 프롬프트가 지시한 아래쪽을 향하지 않습니다.",
        "built_space": "심한 아웃포커싱으로 인해 배경이 흐릿하며, 나무와 불분명한 구조물만 보여 레퍼런스의 캠프 장소인지 명확히 확인하기 어렵습니다.",
        "entities": "얼굴에 흙과 피가 묻은 안규홍의 모습이 보이며, 앞쪽에 낫을 쥐고 있는 손과 낫의 일부가 흐릿하게 나타납니다.",
        "hard_violations": [
         "시선 방향 오류 (프롬프트의 '아래쪽을 예의주시하며'를 위반하고 카메라 정면을 응시함)"
        ],
        "physics": "손이 낫의 자루를 쥐고 있으며, 인물의 자세는 안정적입니다."
       },
       {
        "label": "B",
        "direction": "인물은 프롬프트의 지시대로 화면의 우측 아래쪽을 날카롭고 강렬하게 예의주시하고 있습니다.",
        "built_space": "배경에 이전 샷 레퍼런스에서 설정된 부서진 오두막, 울타리, 흩어진 가마솥과 파편들, 피어오르는 연기가 정확한 위치와 비례로 배치되어 있습니다.",
        "entities": "얼굴에 피와 흙이 묻은 안규홍의 모습이 레퍼런스와 일치하게 나타나며, 미간을 좁힌 날카로운 눈빛이 정확히 표현되었습니다. 클로즈업 프레이밍으로 인해 낫은 보이지 않으나 이는 허용됩니다.",
        "hard_violations": [],
        "physics": "서서 고개를 약간 숙이고 아래를 내려다보는 자연스러운 신체 균형을 유지하고 있습니다."
       }
      ]
     },
     "ok": true
    },
    {
     "model": "gpt-high",
     "order": "reverse",
     "display_to_canonical": {
      "A": "B",
      "B": "A"
     },
     "raw": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 8,
        "verdict_ko": "화면 밖 왼쪽 아래를 경계하는 얼굴 클로즈업과 좁혀진 미간이 요청에 더 가깝고, 불필요한 낫 동작 없이 이전 장면의 분위기를 잇는다."
       },
       {
        "label": "B",
        "score": 6,
        "verdict_ko": "미간의 긴장은 선명하지만 시선이 아래쪽 접근로보다 카메라 정면에 가깝고, 전경으로 들어 올린 손과 낫이 눈빛 중심의 순간을 바꾼다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "고개를 숙인 채 양눈이 화면 밖 왼쪽의 다소 낮은 지점을 향한다. 아래쪽을 경계하는 해석은 가능하지만 숲 접근로 자체는 보이지 않아 정확한 주시 대상은 확인할 수 없다. 렌즈를 직접 응시하지 않으며 무기나 지시하는 손은 프레임 밖이다.",
        "built_space": "인물 뒤 왼쪽과 오른쪽에 초가 건물 두 채가 부분적으로 보이고, 왼쪽에는 검게 탄 수직 기둥과 기댄 목재가 있다. 흙바닥에는 기울어진 둥근 용기 하나와 엎어진 솥 모양 용기 하나, 깨진 그릇과 잔해가 보인다. 이전 장면의 초가·목재·연기 낀 산중 공터와 대체로 이어지지만 용기 배치는 정확히 같다고 확인하기 어렵다. 인물은 오른쪽 전경에 있고 배경 구조물과 겹치는 방식은 자연스럽다.",
        "entities": "젊은 한국인 남성으로 설정된 안규홍 한 명만 보인다. 검은 머리를 뒤로 모았으며 얼굴형과 굵은 눈썹은 참조 인물과 대체로 맞지만 앞머리 가닥은 더 많이 풀려 있다. 때 묻은 미색 저고리, 얼굴의 흙과 피가 유지되고 눈은 정상적인 홍채와 동공을 갖는다. 낫, 등 상처, 하의와 신발은 클로즈업 밖이므로 확인 대상이 아니다. 읽을 수 있는 글자나 다른 사람은 없다.",
        "hard_violations": [],
        "physics": "머리가 목과 어깨에 자연스럽게 연결되고 상체를 앞으로 기울인 자세가 보인다. 발은 프레임 밖이지만 공중에 떠 있다는 징후는 없다. 옷은 어깨와 몸에 걸려 자연스럽게 접히며, 뒤쪽 용기와 목재는 바닥에 놓이거나 다른 구조물에 기대어 있다."
       },
       {
        "label": "B",
        "direction": "턱과 이마는 낮췄지만 양눈은 거의 카메라 정면을 노려본다. 아래쪽 숲 접근로에 시선이 분명히 떨어지는 모습은 아니다. 화면 하단에서 잡은 낫은 오른쪽으로 뻗으며 특정 대상을 겨누거나 베는 접촉은 보이지 않는다.",
        "built_space": "배경 왼쪽 중앙에 흐릿한 초가 지붕 하나가 보이고, 왼쪽과 오른쪽에는 굵은 수직 목재 또는 나무줄기가 하나씩 보인다. 흙바닥에는 검은 용기 형태 하나와 잔해가 있으나 얕은 초점 때문에 세부 개수와 기존 배치를 확인하기 어렵다. 산중 폐허의 재질과 낮의 연무는 이어지지만 이전 장면의 건물 배치를 식별하기는 A보다 어렵다. 인물은 중앙 전경에 있다.",
        "entities": "젊은 한국인 남성으로 설정된 안규홍 한 명이며 검은 머리를 단단히 뒤로 모았다. 참조와 비슷한 얼굴형, 미색 저고리, 얼굴의 흙과 혈흔이 보이고 눈의 해부학적 표현도 정상이다. 하단에는 피 묻은 손과 어두운 곡선형 낫날 일부가 보이지만 날의 혈흔은 선명하지 않다. 등 상처와 하체는 프레임 밖이다. 다른 사람이나 읽을 수 있는 글자는 없다.",
        "hard_violations": [],
        "physics": "머리와 상체의 연결은 자연스럽고, 전경의 손이 낫자루를 쥐고 있어 낫의 지지점도 있다. 팔을 앞으로 들어 낫을 가슴 높이 부근에 둔 자세는 물리적으로 가능하다. 다만 이는 단순히 아래쪽을 예의주시하는 눈빛 외에 추가된 동작이다. 배경 잔해는 지면에 놓여 있으며 떠 있는 물체는 보이지 않는다."
       }
      ],
      "all_candidates_fail": false
     },
     "normalized": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 8,
        "verdict_ko": "화면 밖 왼쪽 아래를 경계하는 얼굴 클로즈업과 좁혀진 미간이 요청에 더 가깝고, 불필요한 낫 동작 없이 이전 장면의 분위기를 잇는다."
       },
       {
        "label": "A",
        "score": 6,
        "verdict_ko": "미간의 긴장은 선명하지만 시선이 아래쪽 접근로보다 카메라 정면에 가깝고, 전경으로 들어 올린 손과 낫이 눈빛 중심의 순간을 바꾼다."
       }
      ],
      "all_candidates_fail": false,
      "readings": [
       {
        "label": "B",
        "direction": "고개를 숙인 채 양눈이 화면 밖 왼쪽의 다소 낮은 지점을 향한다. 아래쪽을 경계하는 해석은 가능하지만 숲 접근로 자체는 보이지 않아 정확한 주시 대상은 확인할 수 없다. 렌즈를 직접 응시하지 않으며 무기나 지시하는 손은 프레임 밖이다.",
        "built_space": "인물 뒤 왼쪽과 오른쪽에 초가 건물 두 채가 부분적으로 보이고, 왼쪽에는 검게 탄 수직 기둥과 기댄 목재가 있다. 흙바닥에는 기울어진 둥근 용기 하나와 엎어진 솥 모양 용기 하나, 깨진 그릇과 잔해가 보인다. 이전 장면의 초가·목재·연기 낀 산중 공터와 대체로 이어지지만 용기 배치는 정확히 같다고 확인하기 어렵다. 인물은 오른쪽 전경에 있고 배경 구조물과 겹치는 방식은 자연스럽다.",
        "entities": "젊은 한국인 남성으로 설정된 안규홍 한 명만 보인다. 검은 머리를 뒤로 모았으며 얼굴형과 굵은 눈썹은 참조 인물과 대체로 맞지만 앞머리 가닥은 더 많이 풀려 있다. 때 묻은 미색 저고리, 얼굴의 흙과 피가 유지되고 눈은 정상적인 홍채와 동공을 갖는다. 낫, 등 상처, 하의와 신발은 클로즈업 밖이므로 확인 대상이 아니다. 읽을 수 있는 글자나 다른 사람은 없다.",
        "hard_violations": [],
        "physics": "머리가 목과 어깨에 자연스럽게 연결되고 상체를 앞으로 기울인 자세가 보인다. 발은 프레임 밖이지만 공중에 떠 있다는 징후는 없다. 옷은 어깨와 몸에 걸려 자연스럽게 접히며, 뒤쪽 용기와 목재는 바닥에 놓이거나 다른 구조물에 기대어 있다."
       },
       {
        "label": "A",
        "direction": "턱과 이마는 낮췄지만 양눈은 거의 카메라 정면을 노려본다. 아래쪽 숲 접근로에 시선이 분명히 떨어지는 모습은 아니다. 화면 하단에서 잡은 낫은 오른쪽으로 뻗으며 특정 대상을 겨누거나 베는 접촉은 보이지 않는다.",
        "built_space": "배경 왼쪽 중앙에 흐릿한 초가 지붕 하나가 보이고, 왼쪽과 오른쪽에는 굵은 수직 목재 또는 나무줄기가 하나씩 보인다. 흙바닥에는 검은 용기 형태 하나와 잔해가 있으나 얕은 초점 때문에 세부 개수와 기존 배치를 확인하기 어렵다. 산중 폐허의 재질과 낮의 연무는 이어지지만 이전 장면의 건물 배치를 식별하기는 A보다 어렵다. 인물은 중앙 전경에 있다.",
        "entities": "젊은 한국인 남성으로 설정된 안규홍 한 명이며 검은 머리를 단단히 뒤로 모았다. 참조와 비슷한 얼굴형, 미색 저고리, 얼굴의 흙과 혈흔이 보이고 눈의 해부학적 표현도 정상이다. 하단에는 피 묻은 손과 어두운 곡선형 낫날 일부가 보이지만 날의 혈흔은 선명하지 않다. 등 상처와 하체는 프레임 밖이다. 다른 사람이나 읽을 수 있는 글자는 없다.",
        "hard_violations": [],
        "physics": "머리와 상체의 연결은 자연스럽고, 전경의 손이 낫자루를 쥐고 있어 낫의 지지점도 있다. 팔을 앞으로 들어 낫을 가슴 높이 부근에 둔 자세는 물리적으로 가능하다. 다만 이는 단순히 아래쪽을 예의주시하는 눈빛 외에 추가된 동작이다. 배경 잔해는 지면에 놓여 있으며 떠 있는 물체는 보이지 않는다."
       }
      ]
     },
     "ok": true
    }
   ],
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "slot_winner_match": true,
   "slot_winner": {
    "gemini-pro": "B",
    "gpt-high": "B"
   },
   "route": "cross_slot_agree"
  },
  "dual": {
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "normalized": {
    "A": 1.05,
    "B": 2.0
   },
   "adjusted": {
    "A": 0.8,
    "B": 2.0
   },
   "violations": {
    "A": [
     "[gemini-pro] 시선 방향 오류 (프롬프트의 '아래쪽을 예의주시하며'를 위반하고 카메라 정면을 응시함)"
    ]
   },
   "per_model_winner": {
    "gemini-pro": "B",
    "gpt-high": "B"
   },
   "agreed": true
  },
  "totals": {
   "B": 2000,
   "A": 800
  },
  "selected": "B",
  "ranking": [
   "B",
   "A"
  ],
  "verdicts": [
   {
    "label": "B",
    "score": 2000,
    "verdict_ko": "프롬프트가 요구한 대로 미간을 좁히고 아래쪽을 날카롭게 예주시하는 표정과 시선 방향을 완벽하게 구현했으며, 배경의 캠프 디테일도 레퍼런스와 일치합니다."
   },
   {
    "label": "A",
    "score": 800,
    "verdict_ko": "클로즈업 프레이밍과 인물의 외형은 비슷하게 구현했으나, 텍스트가 명시한 '아래쪽을 예의주시'하는 시선 대신 카메라 정면을 응시하여 핵심 지시를 위반했습니다.  ★위반: [gemini-pro] 시선 방향 오류 (프롬프트의 '아래쪽을 예의주시하며'를 위반하고 카메라 정면을 응시함)"
   }
  ],
  "refs": [
   {
    "label": "PREVIOUS SHOT STILL — a visually related earlier shot of this same place: the location's look, materials, fixed features and lighting mood are LOCKED to this photo; never copy its camera framing. The clothing, hair and overall look of 안규홍 — who appear both in that photo and in this shot — are LOCKED to that photo. Anyone else visible in that photo is NOT in this shot: never carry their face, body or clothing onto anyone here. If this photo shows a character who cannot move (dead or unconscious), that character's exact pose, position and orientation are ALSO LOCKED — treat their body as a fixed prop of the set that only the camera moves around.",
    "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/db00517b-165b-4005-a287-f28936b59b95/images/272bd7af-9294-4490-98c6-88a40e9deacc/scene/recipe/S64sh10_sel.png",
    "asset_id": "ac4fa429-d723-4d7d-84b8-1ddc32cce154",
    "role": "prev_still"
   },
   {
    "label": "CHARACTER REFERENCE — 안규홍: the exact person appearing in this shot; match face, hair and build exactly.",
    "path": "<bytes:1028384>",
    "asset_id": "3b5894db-ce55-4f34-b966-5e725c3aa952",
    "role": "character_ref"
   }
  ],
  "critique_skipped": true,
  "shot_run_produced": false,
  "shot_run_uid": "06aa1446-6e10-7df1-ac97-a12faecd3792",
  "ref_mode": "prev+엔티티",
  "share_plan": {
   "ref_plan": "prev",
   "prev_anchor_tag": "S64sh10"
  }
 },
 "S64sh18::cine": {
  "applied": true,
  "attempted_at": "2026-09-09T08:28:51.721901+00:00",
  "fingerprint": "50ef328475e024503e90ad42f78353937951619266935518ff1a5393359ac7e1",
  "fingerprint_version": 2,
  "provider": "mai",
  "endpoint": "openrouter/chat-completions",
  "model": "microsoft/mai-image-2.6",
  "pack": "24.202608252115",
  "source_file": "S64sh18_sel.png",
  "source_sha256": "6619ca610da8adf32c139ed1d0fc35384b0bdde54cf201308842ed69c40f980f",
  "file": "S64sh18_cine.png",
  "staged_sha256": "9d5d635aea2c6ff38d651f0e9fbb554e548b6e1904123c5377534c9ce90e8224",
  "latency_ms": 25826
 },
 "S65sh1::signage": {
  "fp": "5e102001b4ea4194",
  "inscriptions": [],
  "cues": [],
  "dropped": []
 },
 "S65sh1::bgfirst_bg": {
  "input_fingerprint": "6174985e103bda1b",
  "prompt": "Create the EMPTY BACKGROUND PLATE for one film shot — NO PEOPLE, no figures, no body parts, no sketch lines, no arrows anywhere.\n\"Empty\" means no people only: KEEP the location's inherent occupants and stock that define the place — animals in an animal shelter, pen or farm, goods in a market, moored boats in a harbour — unless the shot text explicitly removes them.\nThe FIRST attached image is a thin-line storyboard sketch: use ONLY its camera angle, horizon, perspective and the placement/size of buildings and set masses — ignore the sketched people and arrows entirely. The SECOND attached image (LOCATION PHOTOGRAPH) is the real place: take its architecture, materials, signage and fixed features, and RE-PROJECT them into the sketch's camera. If the photograph's camera differs from the sketch's, the sketch's camera wins.\nHUMAN-SCALE CALIBRATION: derive every structure's true size from human-scale elements — a door ≈ 2m, a window ≈ 1–1.5m wide, one storey ≈ 2.5–3m; never inflate a small structure or shrink a large one.\n\nSHOT TEXT this background must serve (Korean): 빽빽한 편백나무 사이로 가파른 산길을 오르기 위해 일제히 한쪽 발을 내디딘 거대한 수의 일본인 군인 대열 전경.\n\nLOCATION (lock): The steep cypress-lined approach road below the mountain camp, broad enough for horses and military supply carts.\n\nTIME OF DAY (lock): day.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: wide shot\n- FRAME LAYOUT: Approaching column on the mountain path in the middle-center of the frame, midground, moves toward uphill continuation beyond the lower-left frame edge; Rear of the column and supply carts in the upper-right of the frame, background.\n- KEY BACKGROUND ELEMENTS: Dense cypress trees (The column emerges between them); used as Create irregular foreground and background openings without concealing all the footfalls; Steep mountain path (Filled by the advancing military column) — Ascends from upper-right background toward lower-left foreground; used as Organizes the soldiers, lead horse and carts into a single readable approach; Japanese soldiers behind the lead group (Advancing with bayonet-fitted rifles; planting feet in slightly different phases with varied knee bends, shoulder angles and spacing) — Seen obliquely from above, following the uphill direction rather than facing the lens; used as Convey the column's numerical weight without duplicated poses or a posed group arrangement; Supply carts and cart-mounted machine gun (Loaded carts follow the lead group, with a machine gun on a cart) — Oblique upper and side views follow the path's diagonal; used as Remain small within the rear column, establishing military equipment at realistic scale.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Natural daylight with muted saturation and sober contrast makes the scale of the approaching force legible without ornamental atmosphere.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nRender ONE photorealistic empty location photograph, 16:9, that this shot can be staged inside later. No readable writing anywhere: surfaces that would carry writing may be present, but stage any wording out of legibility — an oblique angle, distance, shallow focus. No captions, watermarks or overlay text.",
  "effective_prompt": "Create the EMPTY BACKGROUND PLATE for one film shot — NO PEOPLE, no figures, no body parts, no sketch lines, no arrows anywhere.\n\"Empty\" means no people only: KEEP the location's inherent occupants and stock that define the place — animals in an animal shelter, pen or farm, goods in a market, moored boats in a harbour — unless the shot text explicitly removes them.\nThe FIRST attached image is a thin-line storyboard sketch: use ONLY its camera angle, horizon, perspective and the placement/size of buildings and set masses — ignore the sketched people and arrows entirely. The SECOND attached image (LOCATION PHOTOGRAPH) is the real place: take its architecture, materials, signage and fixed features, and RE-PROJECT them into the sketch's camera. If the photograph's camera differs from the sketch's, the sketch's camera wins.\nHUMAN-SCALE CALIBRATION: derive every structure's true size from human-scale elements — a door ≈ 2m, a window ≈ 1–1.5m wide, one storey ≈ 2.5–3m; never inflate a small structure or shrink a large one.\n\nSHOT TEXT this background must serve (Korean): 빽빽한 편백나무 사이로 가파른 산길을 오르기 위해 일제히 한쪽 발을 내디딘 거대한 수의 일본인 군인 대열 전경.\n\nLOCATION (lock): The steep cypress-lined approach road below the mountain camp, broad enough for horses and military supply carts.\n\nTIME OF DAY (lock): day.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: wide shot\n- FRAME LAYOUT: Approaching column on the mountain path in the middle-center of the frame, midground, moves toward uphill continuation beyond the lower-left frame edge; Rear of the column and supply carts in the upper-right of the frame, background.\n- KEY BACKGROUND ELEMENTS: Dense cypress trees (The column emerges between them); used as Create irregular foreground and background openings without concealing all the footfalls; Steep mountain path (Filled by the advancing military column) — Ascends from upper-right background toward lower-left foreground; used as Organizes the soldiers, lead horse and carts into a single readable approach; Japanese soldiers behind the lead group (Advancing with bayonet-fitted rifles; planting feet in slightly different phases with varied knee bends, shoulder angles and spacing) — Seen obliquely from above, following the uphill direction rather than facing the lens; used as Convey the column's numerical weight without duplicated poses or a posed group arrangement; Supply carts and cart-mounted machine gun (Loaded carts follow the lead group, with a machine gun on a cart) — Oblique upper and side views follow the path's diagonal; used as Remain small within the rear column, establishing military equipment at realistic scale.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Natural daylight with muted saturation and sober contrast makes the scale of the approaching force legible without ornamental atmosphere.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nRender ONE photorealistic empty location photograph, 16:9, that this shot can be staged inside later. No readable writing anywhere: surfaces that would carry writing may be present, but stage any wording out of legibility — an oblique angle, distance, shallow focus. No captions, watermarks or overlay text.",
  "bg_path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/db00517b-165b-4005-a287-f28936b59b95/images/272bd7af-9294-4490-98c6-88a40e9deacc/scene/recipe/S65sh1__bgfirst_bg.png",
  "asset_id": "51b6927c-7432-4f4c-a975-8e04bcbbbbd8",
  "input_asset_ids": [
   "052a900c-9848-4657-a468-c7a3d9bbcd1c",
   "8c17c7ae-efff-4bb5-b88c-7c5528f00716"
  ]
 },
 "S65sh1": {
  "input_fingerprint": "4507a3aa2f99ad48",
  "prompt": "Create ONE FINAL photorealistic live-action film still of the moment below — early 20th-century Korea, Jeolla region; all people are Korean unless explicitly stated otherwise. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 빽빽한 편백나무 사이로 가파른 산길을 오르기 위해 일제히 한쪽 발을 내디딘 거대한 수의 일본인 군인 대열 전경.\n\nLOCATION (lock): The steep cypress-lined approach road below the mountain camp, broad enough for horses and military supply carts. The shot takes place here — the attached LOCATION STRUCTURE PHOTOGRAPH is the single authority for this exact place — its fixed structure and permanent site details are LOCKED to it. No separate location photograph exists for this place. Build everything else strictly from the location text above and the shot text; the layout sketch (when attached) governs framing and placement only, and the shot text governs time of day, lighting and action.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: wide shot\n- FRAME LAYOUT: Approaching column on the mountain path in the middle-center of the frame, midground, moves toward uphill continuation beyond the lower-left frame edge; Rear of the column and supply carts in the upper-right of the frame, background.\n- KEY BACKGROUND ELEMENTS: Dense cypress trees (The column emerges between them); used as Create irregular foreground and background openings without concealing all the footfalls; Steep mountain path (Filled by the advancing military column) — Ascends from upper-right background toward lower-left foreground; used as Organizes the soldiers, lead horse and carts into a single readable approach; Japanese soldiers behind the lead group (Advancing with bayonet-fitted rifles; planting feet in slightly different phases with varied knee bends, shoulder angles and spacing) — Seen obliquely from above, following the uphill direction rather than facing the lens; used as Convey the column's numerical weight without duplicated poses or a posed group arrangement; Supply carts and cart-mounted machine gun (Loaded carts follow the lead group, with a machine gun on a cart) — Oblique upper and side views follow the path's diagonal; used as Remain small within the rear column, establishing military equipment at realistic scale.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Natural daylight with muted saturation and sober contrast makes the scale of the approaching force legible without ornamental atmosphere.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): A military-supply cart carries a machine gun along the steep forest path; horses accompany the column.\n\nNo people appear unless the shot text itself says so.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.",
  "roll_prompts": {
   "A": "Stage the shot. The FIRST attached image (SHOT BACKGROUND) is the finished empty background of this shot — keep it EXACTLY: its camera, perspective, architecture, lighting and every fixed feature stay untouched. The SECOND attached image (LAYOUT SKETCH) tells you ONLY where the people go: each sketched person's position, screen size, pose and the gaze/motion arrows. Ignore the sketch's background lines. The CHARACTER REFERENCE photographs show the real people.\nPlace the real people into the background at exactly the sketched positions, sizes and poses, following the arrow directions. No sketch lines or arrows may remain.\n\nCreate ONE FINAL photorealistic live-action film still of the moment below — early 20th-century Korea, Jeolla region; all people are Korean unless explicitly stated otherwise. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 빽빽한 편백나무 사이로 가파른 산길을 오르기 위해 일제히 한쪽 발을 내디딘 거대한 수의 일본인 군인 대열 전경.\n\nLOCATION (lock): The steep cypress-lined approach road below the mountain camp, broad enough for horses and military supply carts. The shot takes place here — the FIRST attached image (SHOT BACKGROUND) is this exact place, already built: its ground, structures, horizon, materials and lighting are the finished truth of this location and must not be redesigned or replaced. No location photograph is attached — read the place from that image alone, and add no scenery, structure, vehicle or fixture that it does not already show. This lock governs the place only; the figures in the shot follow the staging and pose instructions.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: wide shot\n- FRAME LAYOUT: Approaching column on the mountain path in the middle-center of the frame, midground, moves toward uphill continuation beyond the lower-left frame edge; Rear of the column and supply carts in the upper-right of the frame, background.\n- KEY BACKGROUND ELEMENTS: Dense cypress trees (The column emerges between them); used as Create irregular foreground and background openings without concealing all the footfalls; Steep mountain path (Filled by the advancing military column) — Ascends from upper-right background toward lower-left foreground; used as Organizes the soldiers, lead horse and carts into a single readable approach; Japanese soldiers behind the lead group (Advancing with bayonet-fitted rifles; planting feet in slightly different phases with varied knee bends, shoulder angles and spacing) — Seen obliquely from above, following the uphill direction rather than facing the lens; used as Convey the column's numerical weight without duplicated poses or a posed group arrangement; Supply carts and cart-mounted machine gun (Loaded carts follow the lead group, with a machine gun on a cart) — Oblique upper and side views follow the path's diagonal; used as Remain small within the rear column, establishing military equipment at realistic scale.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Natural daylight with muted saturation and sober contrast makes the scale of the approaching force legible without ornamental atmosphere.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): A military-supply cart carries a machine gun along the steep forest path; horses accompany the column.\n\nNo people appear unless the shot text itself says so.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.",
   "B": "Create ONE FINAL photorealistic live-action film still of the moment below — early 20th-century Korea, Jeolla region; all people are Korean unless explicitly stated otherwise. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 빽빽한 편백나무 사이로 가파른 산길을 오르기 위해 일제히 한쪽 발을 내디딘 거대한 수의 일본인 군인 대열 전경.\n\nLOCATION (lock): The steep cypress-lined approach road below the mountain camp, broad enough for horses and military supply carts. The shot takes place here — the attached LOCATION STRUCTURE PHOTOGRAPH is the single authority for this exact place — its fixed structure and permanent site details are LOCKED to it. No separate location photograph exists for this place. Build everything else strictly from the location text above and the shot text; the layout sketch (when attached) governs framing and placement only, and the shot text governs time of day, lighting and action.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: wide shot\n- FRAME LAYOUT: Approaching column on the mountain path in the middle-center of the frame, midground, moves toward uphill continuation beyond the lower-left frame edge; Rear of the column and supply carts in the upper-right of the frame, background.\n- KEY BACKGROUND ELEMENTS: Dense cypress trees (The column emerges between them); used as Create irregular foreground and background openings without concealing all the footfalls; Steep mountain path (Filled by the advancing military column) — Ascends from upper-right background toward lower-left foreground; used as Organizes the soldiers, lead horse and carts into a single readable approach; Japanese soldiers behind the lead group (Advancing with bayonet-fitted rifles; planting feet in slightly different phases with varied knee bends, shoulder angles and spacing) — Seen obliquely from above, following the uphill direction rather than facing the lens; used as Convey the column's numerical weight without duplicated poses or a posed group arrangement; Supply carts and cart-mounted machine gun (Loaded carts follow the lead group, with a machine gun on a cart) — Oblique upper and side views follow the path's diagonal; used as Remain small within the rear column, establishing military equipment at realistic scale.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Natural daylight with muted saturation and sober contrast makes the scale of the approaching force legible without ornamental atmosphere.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): A military-supply cart carries a machine gun along the steep forest path; horses accompany the column.\n\nNo people appear unless the shot text itself says so.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.\n\nCOMPOSITION VARIATION (this candidate only): keep the place, the people and the moment exactly as specified, and keep from the CAMERA & FRAME contract its subject scale, screen placement and key background elements exactly. For this candidate only, its camera angle and height are deliberately left open: choose a camera position distinctly different from the obvious one, as a film director picking a second setup on the same blocking. Stay true to any explicit framing the SHOT TEXT itself dictates."
  },
  "roll_refs": {
   "A": [
    {
     "label": "SHOT BACKGROUND",
     "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/db00517b-165b-4005-a287-f28936b59b95/images/272bd7af-9294-4490-98c6-88a40e9deacc/scene/recipe/S65sh1__bgfirst_bg.png",
     "asset_id": "51b6927c-7432-4f4c-a975-8e04bcbbbbd8",
     "role": "bgfirst_bg"
    },
    {
     "label": "LAYOUT SKETCH (people placement only)",
     "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/db00517b-165b-4005-a287-f28936b59b95/images/272bd7af-9294-4490-98c6-88a40e9deacc/conti/conti_S65sh1.png",
     "asset_id": "052a900c-9848-4657-a468-c7a3d9bbcd1c",
     "role": "conti_light"
    },
    {
     "label": "PROP REFERENCE — 근대식 소총: the exact object appearing in this shot; match its look, material and wear exactly.",
     "path": "<bytes:828568>",
     "asset_id": "7543806c-8087-4caa-8d7b-56d10864fd3c",
     "role": "prop_ref"
    }
   ],
   "B": [
    {
     "label": "LOCATION STRUCTURE PHOTOGRAPH — the confirmed photograph of this exact place and its fixed structure: it is the SINGLE authority for the location, the structure's shape, proportions, materials, colors, openings and every permanent site detail. Never copy its camera framing, time of day or lighting — the shot text is the authority for those.",
     "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/db00517b-165b-4005-a287-f28936b59b95/images/background_chain/seed_bg_mountain_camp_sel.png",
     "asset_id": "8c17c7ae-efff-4bb5-b88c-7c5528f00716",
     "role": "location_seed_bg"
    },
    {
     "label": "PROP REFERENCE — 근대식 소총: the exact object appearing in this shot; match its look, material and wear exactly.",
     "path": "<bytes:828568>",
     "asset_id": "7543806c-8087-4caa-8d7b-56d10864fd3c",
     "role": "prop_ref"
    }
   ]
  },
  "shot_run_spend_attempt_count": 1,
  "judge_flip_skipped": "judge_fn_owns_order",
  "readings": [
   {
    "label": "A",
    "direction": "군인 대열과 말 탄 장교가 우측에서 좌측으로 이어지는 평탄한 길을 따라 수평으로 이동하고 있으며, 시선은 모두 좌측 전방을 향해 있습니다.",
    "built_space": "화면 중앙 뒤편으로 나무 울타리와 초가 건물, 바위가 있는 산채가 보이며 주변에 편백나무가 빽빽하지만, 요구된 대각선 오르막길 대신 산채 앞을 지나가는 평탄한 가로 방향의 길이 전경을 가로지르고 있습니다.",
    "entities": "소규모의 일본군 병사들이 어깨에 소총을 메고 있으며 총검은 없습니다(규모 부족 및 디테일 누락). 보급 마차가 있으나, 요구된 기관총 수레 대신 야포(대포)가 뒤따르고 있습니다(객체 오류).",
    "hard_violations": [
     "[gemini-pro] 지정된 카메라 구도 및 동선 위반: 프롬프트가 명시한 '우측 상단에서 좌측 하단으로 이어지는 오르막길'과 대각선 구도를 완전히 무시하고 평탄한 가로 방향의 길로 연출함."
    ],
    "physics": "말과 걷고 있는 병사들 모두 평지를 걷는 자연스러운 보행 자세를 취하고 있으며, 발과 말굽, 마차 바퀴 모두 지면에 잘 닿아 있어 물리적인 오류나 떠 있는 객체는 없습니다."
   },
   {
    "label": "B",
    "direction": "거대한 군인 대열이 우측 상단의 배경에서 좌측 하단의 전경을 향해 대각선으로 이동하고 있으며, 병사들의 시선과 몸의 방향 모두 화면 좌측 하단의 경로를 향해 있습니다.",
    "built_space": "배경의 산마루에 레퍼런스와 일치하는 원형 나무 울타리와 초가 지붕 건물 2채가 있는 산채가 위치하며, 빽빽한 편백나무 사이로 가파른 흙길이 우상단에서 좌하단으로 이어지게 잘 구성되어 있습니다.",
    "entities": "일본군 병사들이 거대한 대열을 이루고 있으며(일치), 손에 근대식 소총을 들고 있으나 총검은 보이지 않습니다(부분 불일치). 대열 뒤쪽으로 보급 마차가 이어지며(일치), 기관총이 탑재된 수레도 존재하지만 후방이 아닌 우측 전경에 크게 배치되었습니다(배치 불일치).",
    "hard_violations": [],
    "physics": "전경에 있는 병사들은 바위와 흙바닥의 경사를 딛고 올라가기 위해 무릎을 굽히고 체중을 싣는 등 물리적으로 올바른 등반 자세를 취하고 있으며, 마차와 말의 바퀴 및 발굽도 지면에 안정적으로 닿아 있습니다."
   }
  ],
  "cross_model_order": {
   "policy": "cross_model_order_v2_slot0_fwd_slot1_rev_parallel2",
   "slots": [
    {
     "model": "gemini-pro",
     "order": "forward",
     "display_to_canonical": {
      "A": "A",
      "B": "B"
     },
     "raw": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 8,
        "verdict_ko": "지정된 우상단에서 좌하단으로 이어지는 대각선 구도와 가파른 산길, 거대한 군인 대열을 훌륭하게 구현했으나, 기관총이 탑재된 수레가 지시된 후미가 아닌 전경에 배치되었고 소총의 총검이 누락되었습니다."
       },
       {
        "label": "A",
        "score": 3,
        "verdict_ko": "지정된 대각선 구도와 가파른 오르막길 설정을 완전히 무시하고 평탄한 가로 방향의 길로 연출한 심각한 구도 위반이 있으며, 요구된 기관총 대신 대포가 등장합니다."
       }
      ],
      "readings": [
       {
        "label": "B",
        "direction": "거대한 군인 대열이 우측 상단의 배경에서 좌측 하단의 전경을 향해 대각선으로 이동하고 있으며, 병사들의 시선과 몸의 방향 모두 화면 좌측 하단의 경로를 향해 있습니다.",
        "built_space": "배경의 산마루에 레퍼런스와 일치하는 원형 나무 울타리와 초가 지붕 건물 2채가 있는 산채가 위치하며, 빽빽한 편백나무 사이로 가파른 흙길이 우상단에서 좌하단으로 이어지게 잘 구성되어 있습니다.",
        "entities": "일본군 병사들이 거대한 대열을 이루고 있으며(일치), 손에 근대식 소총을 들고 있으나 총검은 보이지 않습니다(부분 불일치). 대열 뒤쪽으로 보급 마차가 이어지며(일치), 기관총이 탑재된 수레도 존재하지만 후방이 아닌 우측 전경에 크게 배치되었습니다(배치 불일치).",
        "hard_violations": [],
        "physics": "전경에 있는 병사들은 바위와 흙바닥의 경사를 딛고 올라가기 위해 무릎을 굽히고 체중을 싣는 등 물리적으로 올바른 등반 자세를 취하고 있으며, 마차와 말의 바퀴 및 발굽도 지면에 안정적으로 닿아 있습니다."
       },
       {
        "label": "A",
        "direction": "군인 대열과 말 탄 장교가 우측에서 좌측으로 이어지는 평탄한 길을 따라 수평으로 이동하고 있으며, 시선은 모두 좌측 전방을 향해 있습니다.",
        "built_space": "화면 중앙 뒤편으로 나무 울타리와 초가 건물, 바위가 있는 산채가 보이며 주변에 편백나무가 빽빽하지만, 요구된 대각선 오르막길 대신 산채 앞을 지나가는 평탄한 가로 방향의 길이 전경을 가로지르고 있습니다.",
        "entities": "소규모의 일본군 병사들이 어깨에 소총을 메고 있으며 총검은 없습니다(규모 부족 및 디테일 누락). 보급 마차가 있으나, 요구된 기관총 수레 대신 야포(대포)가 뒤따르고 있습니다(객체 오류).",
        "hard_violations": [
         "지정된 카메라 구도 및 동선 위반: 프롬프트가 명시한 '우측 상단에서 좌측 하단으로 이어지는 오르막길'과 대각선 구도를 완전히 무시하고 평탄한 가로 방향의 길로 연출함."
        ],
        "physics": "말과 걷고 있는 병사들 모두 평지를 걷는 자연스러운 보행 자세를 취하고 있으며, 발과 말굽, 마차 바퀴 모두 지면에 잘 닿아 있어 물리적인 오류나 떠 있는 객체는 없습니다."
       }
      ],
      "all_candidates_fail": false
     },
     "normalized": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 8,
        "verdict_ko": "지정된 우상단에서 좌하단으로 이어지는 대각선 구도와 가파른 산길, 거대한 군인 대열을 훌륭하게 구현했으나, 기관총이 탑재된 수레가 지시된 후미가 아닌 전경에 배치되었고 소총의 총검이 누락되었습니다."
       },
       {
        "label": "A",
        "score": 3,
        "verdict_ko": "지정된 대각선 구도와 가파른 오르막길 설정을 완전히 무시하고 평탄한 가로 방향의 길로 연출한 심각한 구도 위반이 있으며, 요구된 기관총 대신 대포가 등장합니다."
       }
      ],
      "all_candidates_fail": false,
      "readings": [
       {
        "label": "B",
        "direction": "거대한 군인 대열이 우측 상단의 배경에서 좌측 하단의 전경을 향해 대각선으로 이동하고 있으며, 병사들의 시선과 몸의 방향 모두 화면 좌측 하단의 경로를 향해 있습니다.",
        "built_space": "배경의 산마루에 레퍼런스와 일치하는 원형 나무 울타리와 초가 지붕 건물 2채가 있는 산채가 위치하며, 빽빽한 편백나무 사이로 가파른 흙길이 우상단에서 좌하단으로 이어지게 잘 구성되어 있습니다.",
        "entities": "일본군 병사들이 거대한 대열을 이루고 있으며(일치), 손에 근대식 소총을 들고 있으나 총검은 보이지 않습니다(부분 불일치). 대열 뒤쪽으로 보급 마차가 이어지며(일치), 기관총이 탑재된 수레도 존재하지만 후방이 아닌 우측 전경에 크게 배치되었습니다(배치 불일치).",
        "hard_violations": [],
        "physics": "전경에 있는 병사들은 바위와 흙바닥의 경사를 딛고 올라가기 위해 무릎을 굽히고 체중을 싣는 등 물리적으로 올바른 등반 자세를 취하고 있으며, 마차와 말의 바퀴 및 발굽도 지면에 안정적으로 닿아 있습니다."
       },
       {
        "label": "A",
        "direction": "군인 대열과 말 탄 장교가 우측에서 좌측으로 이어지는 평탄한 길을 따라 수평으로 이동하고 있으며, 시선은 모두 좌측 전방을 향해 있습니다.",
        "built_space": "화면 중앙 뒤편으로 나무 울타리와 초가 건물, 바위가 있는 산채가 보이며 주변에 편백나무가 빽빽하지만, 요구된 대각선 오르막길 대신 산채 앞을 지나가는 평탄한 가로 방향의 길이 전경을 가로지르고 있습니다.",
        "entities": "소규모의 일본군 병사들이 어깨에 소총을 메고 있으며 총검은 없습니다(규모 부족 및 디테일 누락). 보급 마차가 있으나, 요구된 기관총 수레 대신 야포(대포)가 뒤따르고 있습니다(객체 오류).",
        "hard_violations": [
         "지정된 카메라 구도 및 동선 위반: 프롬프트가 명시한 '우측 상단에서 좌측 하단으로 이어지는 오르막길'과 대각선 구도를 완전히 무시하고 평탄한 가로 방향의 길로 연출함."
        ],
        "physics": "말과 걷고 있는 병사들 모두 평지를 걷는 자연스러운 보행 자세를 취하고 있으며, 발과 말굽, 마차 바퀴 모두 지면에 잘 닿아 있어 물리적인 오류나 떠 있는 객체는 없습니다."
       }
      ]
     },
     "ok": true
    },
    {
     "model": "gpt-high",
     "order": "reverse",
     "display_to_canonical": {
      "A": "B",
      "B": "A"
     },
     "raw": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 4,
        "verdict_ko": "거대한 병력과 가파른 오르막의 발디딤은 잘 보이지만, 선두와 기관총 수레가 전경을 크게 차지하고 수레의 말들이 병사들과 반대 방향을 향해 지정된 대각선 접근 구도를 벗어납니다."
       },
       {
        "label": "B",
        "score": 6,
        "verdict_ko": "중앙 중경의 병사들과 우상단 후방의 보급 수레 배치는 더 충실하지만, 선두 기마병과 장교가 대열을 마주 보고 있어 일제히 오르는 행군의 방향성이 깨지고 급경사 표현도 약합니다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "보병들은 우상단에서 좌하단으로 전진하며 시선도 대체로 앞쪽 노면을 향합니다. 그러나 오른쪽 전경의 수레에 연결된 말들은 엉덩이를 카메라 쪽에 두고 상단 후방을 향해 보병과 반대로 서 있습니다. 보병의 소총과 총검은 주로 화면 오른쪽 또는 위쪽으로 비스듬히 들려 있고, 전경 기관총 총구는 오른쪽을 향합니다. 어느 총도 특정 인물을 조준하는 모습은 아닙니다.",
        "built_space": "높은 시점의 넓은 화면에 편백나무 사이의 돌과 뿌리가 드러난 넓은 흙길이 보입니다. 상단에는 초가 건물 두 채, 목책으로 둘러싸인 마당, 출입구와 바위군이 있어 참조 장소의 주요 구성요소를 반영합니다. 길의 전경은 가파르게 솟아 보이지만, 선두 보병이 중경보다 전경에 크게 배치되고 기관총 수레도 우하단 전경을 차지합니다. 이는 병력 중심을 중경에, 작은 수레들을 우상단 후방에 두라는 배치와 다릅니다.",
        "entities": "다수의 동아시아계 성인 남성 보병과 기마병, 여러 말, 적재 수레들과 수레 위 기관총 한 정이 보입니다. 얼굴만으로 일본인 여부를 확정할 수는 없지만 군복과 장비로 군인 대열임은 분명합니다. 병사들은 카키색 군복과 모자, 군장, 목재 개머리판의 장총을 갖추고 있으며 일부 총검이 보입니다. 다만 참조 총기의 특징인 외부 격발 장치와 황동 띠까지 같은지는 확인하기 어렵습니다. 낮의 절제된 색감이며 읽을 수 있는 글자는 없습니다.",
        "hard_violations": [],
        "physics": "선두 병사들은 한쪽 군화를 노면에 딛고 다른 발을 들거나 내밀어 오르막에서 체중을 옮깁니다. 무릎과 어깨 각도에는 차이가 있어 동일 자세를 그대로 복제한 행렬은 아닙니다. 소총은 손으로 받치거나 끈으로 지지되고 기마병은 안장에 앉아 있습니다. 수레는 바퀴로 땅을 지지하며 기관총은 적재함 위 받침에 놓여 있습니다. 명백하게 지지 없이 떠 있는 몸이나 물체는 보이지 않습니다."
       },
       {
        "label": "B",
        "direction": "중앙 보병들과 뒤쪽 대열은 우상단에서 좌하단 방향으로 걷고, 소총과 총검은 어깨 옆에서 위로 향합니다. 반면 좌하단의 선두 말과 기마병, 그 옆에서 손을 든 장교는 우상단의 병사들을 마주 보고 있습니다. 따라서 보병의 접근 방향은 맞지만 선두까지 같은 방향으로 오르는 흐름은 아닙니다. 장교의 칼끝은 우하단 노면 쪽을 향하고 수레 기관총 총구는 화면 왼쪽을 향하며, 사람을 겨누는 장면은 아닙니다.",
        "built_space": "중앙 중경에 보병들이 있고, 대열과 적재 수레들은 우상단 배경으로 이어집니다. 기관총 수레도 선두 병사들 뒤에 작게 보여 지정된 화면 배치에 더 가깝습니다. 왼쪽에는 초가 두 채, 마당, 목책, 큰 바위들이 있으며 참조 장소의 재료와 주요 배치가 비교적 잘 유지됩니다. 높은 사선 시점에서 길과 수레 윗면이 보이고, 나무 사이로 발디딤도 확인됩니다. 다만 길이 완만한 사면 통로처럼 보여 좌하단으로 가파르게 상승한다는 요구는 뚜렷하지 않습니다.",
        "entities": "동아시아계 성인 남성 보병들, 기마병 한 명과 앞쪽 장교, 말들, 상자를 실은 후방 수레 두 대와 그 앞의 기관총 수레가 보입니다. 군인들은 짙은 군복과 둥근 군모 또는 철모를 착용하며, 얼굴만으로 국적을 판별할 수는 없습니다. 목재 부분이 있는 장총과 긴 총검은 확인되지만 참조 총기의 세부 격발 구조와 마모까지 일치하는지는 식별되지 않습니다. 병력은 길게 이어지지만 A보다 수적 압박감이 약합니다. 자연광의 낮 장면이며 판독 가능한 문구는 없습니다.",
        "hard_violations": [],
        "physics": "중앙 병사들은 지면을 딛는 발과 들어 올린 발이 구분되고, 소총은 손과 어깨끈으로 지지됩니다. 좌하단 말은 지지하는 다리를 남겨 둔 채 앞다리 하나를 들고 있으며 기마병의 엉덩이는 안장에 놓여 있습니다. 장교 역시 한쪽 발에 체중을 싣고 걸음을 옮깁니다. 수레 바퀴는 길에 닿고 적재 상자와 기관총은 수레 위에 지지되어 있습니다. 명백한 부유나 불가능한 관절 자세는 보이지 않습니다."
       }
      ],
      "all_candidates_fail": true
     },
     "normalized": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 4,
        "verdict_ko": "거대한 병력과 가파른 오르막의 발디딤은 잘 보이지만, 선두와 기관총 수레가 전경을 크게 차지하고 수레의 말들이 병사들과 반대 방향을 향해 지정된 대각선 접근 구도를 벗어납니다."
       },
       {
        "label": "A",
        "score": 6,
        "verdict_ko": "중앙 중경의 병사들과 우상단 후방의 보급 수레 배치는 더 충실하지만, 선두 기마병과 장교가 대열을 마주 보고 있어 일제히 오르는 행군의 방향성이 깨지고 급경사 표현도 약합니다."
       }
      ],
      "all_candidates_fail": true,
      "readings": [
       {
        "label": "B",
        "direction": "보병들은 우상단에서 좌하단으로 전진하며 시선도 대체로 앞쪽 노면을 향합니다. 그러나 오른쪽 전경의 수레에 연결된 말들은 엉덩이를 카메라 쪽에 두고 상단 후방을 향해 보병과 반대로 서 있습니다. 보병의 소총과 총검은 주로 화면 오른쪽 또는 위쪽으로 비스듬히 들려 있고, 전경 기관총 총구는 오른쪽을 향합니다. 어느 총도 특정 인물을 조준하는 모습은 아닙니다.",
        "built_space": "높은 시점의 넓은 화면에 편백나무 사이의 돌과 뿌리가 드러난 넓은 흙길이 보입니다. 상단에는 초가 건물 두 채, 목책으로 둘러싸인 마당, 출입구와 바위군이 있어 참조 장소의 주요 구성요소를 반영합니다. 길의 전경은 가파르게 솟아 보이지만, 선두 보병이 중경보다 전경에 크게 배치되고 기관총 수레도 우하단 전경을 차지합니다. 이는 병력 중심을 중경에, 작은 수레들을 우상단 후방에 두라는 배치와 다릅니다.",
        "entities": "다수의 동아시아계 성인 남성 보병과 기마병, 여러 말, 적재 수레들과 수레 위 기관총 한 정이 보입니다. 얼굴만으로 일본인 여부를 확정할 수는 없지만 군복과 장비로 군인 대열임은 분명합니다. 병사들은 카키색 군복과 모자, 군장, 목재 개머리판의 장총을 갖추고 있으며 일부 총검이 보입니다. 다만 참조 총기의 특징인 외부 격발 장치와 황동 띠까지 같은지는 확인하기 어렵습니다. 낮의 절제된 색감이며 읽을 수 있는 글자는 없습니다.",
        "hard_violations": [],
        "physics": "선두 병사들은 한쪽 군화를 노면에 딛고 다른 발을 들거나 내밀어 오르막에서 체중을 옮깁니다. 무릎과 어깨 각도에는 차이가 있어 동일 자세를 그대로 복제한 행렬은 아닙니다. 소총은 손으로 받치거나 끈으로 지지되고 기마병은 안장에 앉아 있습니다. 수레는 바퀴로 땅을 지지하며 기관총은 적재함 위 받침에 놓여 있습니다. 명백하게 지지 없이 떠 있는 몸이나 물체는 보이지 않습니다."
       },
       {
        "label": "A",
        "direction": "중앙 보병들과 뒤쪽 대열은 우상단에서 좌하단 방향으로 걷고, 소총과 총검은 어깨 옆에서 위로 향합니다. 반면 좌하단의 선두 말과 기마병, 그 옆에서 손을 든 장교는 우상단의 병사들을 마주 보고 있습니다. 따라서 보병의 접근 방향은 맞지만 선두까지 같은 방향으로 오르는 흐름은 아닙니다. 장교의 칼끝은 우하단 노면 쪽을 향하고 수레 기관총 총구는 화면 왼쪽을 향하며, 사람을 겨누는 장면은 아닙니다.",
        "built_space": "중앙 중경에 보병들이 있고, 대열과 적재 수레들은 우상단 배경으로 이어집니다. 기관총 수레도 선두 병사들 뒤에 작게 보여 지정된 화면 배치에 더 가깝습니다. 왼쪽에는 초가 두 채, 마당, 목책, 큰 바위들이 있으며 참조 장소의 재료와 주요 배치가 비교적 잘 유지됩니다. 높은 사선 시점에서 길과 수레 윗면이 보이고, 나무 사이로 발디딤도 확인됩니다. 다만 길이 완만한 사면 통로처럼 보여 좌하단으로 가파르게 상승한다는 요구는 뚜렷하지 않습니다.",
        "entities": "동아시아계 성인 남성 보병들, 기마병 한 명과 앞쪽 장교, 말들, 상자를 실은 후방 수레 두 대와 그 앞의 기관총 수레가 보입니다. 군인들은 짙은 군복과 둥근 군모 또는 철모를 착용하며, 얼굴만으로 국적을 판별할 수는 없습니다. 목재 부분이 있는 장총과 긴 총검은 확인되지만 참조 총기의 세부 격발 구조와 마모까지 일치하는지는 식별되지 않습니다. 병력은 길게 이어지지만 A보다 수적 압박감이 약합니다. 자연광의 낮 장면이며 판독 가능한 문구는 없습니다.",
        "hard_violations": [],
        "physics": "중앙 병사들은 지면을 딛는 발과 들어 올린 발이 구분되고, 소총은 손과 어깨끈으로 지지됩니다. 좌하단 말은 지지하는 다리를 남겨 둔 채 앞다리 하나를 들고 있으며 기마병의 엉덩이는 안장에 놓여 있습니다. 장교 역시 한쪽 발에 체중을 싣고 걸음을 옮깁니다. 수레 바퀴는 길에 닿고 적재 상자와 기관총은 수레 위에 지지되어 있습니다. 명백한 부유나 불가능한 관절 자세는 보이지 않습니다."
       }
      ]
     },
     "ok": true
    }
   ],
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "slot_winner_match": false,
   "slot_winner": {
    "gemini-pro": "B",
    "gpt-high": "A"
   },
   "route": "cross_slot_combined"
  },
  "dual": {
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "normalized": {
    "A": 1.375,
    "B": 1.667
   },
   "adjusted": {
    "A": 1.125,
    "B": 1.667
   },
   "violations": {
    "A": [
     "[gemini-pro] 지정된 카메라 구도 및 동선 위반: 프롬프트가 명시한 '우측 상단에서 좌측 하단으로 이어지는 오르막길'과 대각선 구도를 완전히 무시하고 평탄한 가로 방향의 길로 연출함."
    ]
   },
   "per_model_winner": {
    "gemini-pro": "B",
    "gpt-high": "A"
   },
   "agreed": false
  },
  "totals": {
   "B": 1667,
   "A": 1125
  },
  "selected": "B",
  "ranking": [
   "B",
   "A"
  ],
  "verdicts": [
   {
    "label": "B",
    "score": 1667,
    "verdict_ko": "지정된 우상단에서 좌하단으로 이어지는 대각선 구도와 가파른 산길, 거대한 군인 대열을 훌륭하게 구현했으나, 기관총이 탑재된 수레가 지시된 후미가 아닌 전경에 배치되었고 소총의 총검이 누락되었습니다."
   },
   {
    "label": "A",
    "score": 1125,
    "verdict_ko": "지정된 대각선 구도와 가파른 오르막길 설정을 완전히 무시하고 평탄한 가로 방향의 길로 연출한 심각한 구도 위반이 있으며, 요구된 기관총 대신 대포가 등장합니다.  ★위반: [gemini-pro] 지정된 카메라 구도 및 동선 위반: 프롬프트가 명시한 '우측 상단에서 좌측 하단으로 이어지는 오르막길'과 대각선 구도를 완전히 무시하고 평탄한 가로 방향의 길로 연출함."
   }
  ],
  "refs": [
   {
    "label": "LOCATION STRUCTURE PHOTOGRAPH — the confirmed photograph of this exact place and its fixed structure: it is the SINGLE authority for the location, the structure's shape, proportions, materials, colors, openings and every permanent site detail. Never copy its camera framing, time of day or lighting — the shot text is the authority for those.",
    "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/db00517b-165b-4005-a287-f28936b59b95/images/background_chain/seed_bg_mountain_camp_sel.png",
    "asset_id": "8c17c7ae-efff-4bb5-b88c-7c5528f00716",
    "role": "location_seed_bg"
   },
   {
    "label": "PROP REFERENCE — 근대식 소총: the exact object appearing in this shot; match its look, material and wear exactly.",
    "path": "<bytes:828568>",
    "asset_id": "7543806c-8087-4caa-8d7b-56d10864fd3c",
    "role": "prop_ref"
   }
  ],
  "critique_skipped": true,
  "shot_run_produced": false,
  "shot_run_uid": "06aa1446-6fb0-7e3c-be88-613355e5f4ab",
  "bgfirst": {
   "bg_path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/db00517b-165b-4005-a287-f28936b59b95/images/272bd7af-9294-4490-98c6-88a40e9deacc/scene/recipe/S65sh1__bgfirst_bg.png",
   "bg_asset_id": "51b6927c-7432-4f4c-a975-8e04bcbbbbd8",
   "bg_record_key": "S65sh1::bgfirst_bg",
   "chain_winner": false,
   "authority": "seed_bg"
  },
  "ref_mode": "seed-bg+엔티티 (2택1: 무콘티 승)",
  "share_plan": {
   "ref_plan": "background"
  },
  "lane_policy": "ab_select_ready"
 },
 "S65sh1::cine": {
  "applied": true,
  "attempted_at": "2026-09-09T08:33:44.087147+00:00",
  "fingerprint": "91a1099234e3969085d61e6d14218ae202df270313f9a4ced0d8ca190eb42122",
  "fingerprint_version": 2,
  "provider": "mai",
  "endpoint": "openrouter/chat-completions",
  "model": "microsoft/mai-image-2.6",
  "pack": "24.202608252115",
  "source_file": "S65sh1_sel.png",
  "source_sha256": "a5b88e752f6beea856ae20fe6ba8277f47f0690c45d9c3d1983d21d41b738307",
  "file": "S65sh1_cine.png",
  "staged_sha256": "951e40702a6fcfa44a2511afd64ec6d934e4a698f201855a0a7eb190083ef02e",
  "latency_ms": 28328
 },
 "S65sh14::signage": {
  "fp": "05be488359d28da7",
  "inscriptions": [],
  "cues": [],
  "dropped": []
 },
 "S65sh14": {
  "input_fingerprint": "098cb3162024b8a1",
  "prompt": "Create ONE FINAL photorealistic live-action film still of the moment below — early 20th-century Korea, Jeolla region; all people are Korean unless explicitly stated otherwise. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 허공을 가른 헌병 장교의 서늘한 군도가 배상도의 목에 사선으로 맹렬하게 파고든 타격의 찰나.\n\nLOCATION (lock): The front of the military column on the mountain camp's uphill forest road, beside the mounted commander's horse. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: close-up\n- KEY BACKGROUND ELEMENTS: Saber (Meeting 배상도's neck during the diagonal strike) — Runs from the officer's upper-right arm toward the center-left contact; used as Links attacker and victim through one readable diagonal while occupying a small fraction of the image; Reins (Beginning to slip from 배상도's hand) — A short visible section leads from the lower frame toward the horse outside the crop; used as Provides a secondary action cue at the lower edge without requiring a separate insert.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Maintain the established natural daylight and muted contrast, rendering the sudden violence plainly without a flash, colored accent or glamorized blade highlight.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The people visible in that photo are NOT in this shot: never carry their faces, bodies or clothing onto anyone here. Each person in THIS shot is defined solely by the PEOPLE list, their own CHARACTER REFERENCE images and their wardrobe notes. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nPREVIOUS STILL USAGE (follow exactly — what to take from the attached still and what to exclude; it governs place and objects only, never who is in this shot or how the camera sees them): Take the cypress-lined uphill path, exposed soil, supply cart, and daylight visible in the reference. Exclude duplicated military equipment and any dust cloud unique to the earlier advance.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The machine gun remains mounted on the supply cart in the crowded mountain column; the lead horse is beside the front of the formation. 헌병 장교: He remains mounted with his military sword drawn in a cutting stroke. 배상도: His neck is being cut while his grip on the reins fails.\n\nTHE HAND THAT IS DOING THIS: the object at the centre of this shot is being held, operated, read, aimed or handed over by 헌병 장교 right now, so 헌병 장교's hand — and as much of the wrist and forearm as the framing reaches — is in the frame, gripping or resting on that object exactly the way its real use requires. Match that hand to 헌병 장교: its size, build, skin, age, grooming, sleeve and anything worn on it belong to that person and to no one else.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 배상도 (한국인 남성, 성인 나이대, 자연스러운 검은 머리); 헌병 장교 (일본인 남성, 성인 나이대, 자연스러운 검은 머리) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.",
  "roll_prompts": {
   "A": "Create ONE FINAL photorealistic live-action film still of the moment below — early 20th-century Korea, Jeolla region; all people are Korean unless explicitly stated otherwise. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 허공을 가른 헌병 장교의 서늘한 군도가 배상도의 목에 사선으로 맹렬하게 파고든 타격의 찰나.\n\nLOCATION (lock): The front of the military column on the mountain camp's uphill forest road, beside the mounted commander's horse. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: close-up\n- KEY BACKGROUND ELEMENTS: Saber (Meeting 배상도's neck during the diagonal strike) — Runs from the officer's upper-right arm toward the center-left contact; used as Links attacker and victim through one readable diagonal while occupying a small fraction of the image; Reins (Beginning to slip from 배상도's hand) — A short visible section leads from the lower frame toward the horse outside the crop; used as Provides a secondary action cue at the lower edge without requiring a separate insert.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Maintain the established natural daylight and muted contrast, rendering the sudden violence plainly without a flash, colored accent or glamorized blade highlight.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The people visible in that photo are NOT in this shot: never carry their faces, bodies or clothing onto anyone here. Each person in THIS shot is defined solely by the PEOPLE list, their own CHARACTER REFERENCE images and their wardrobe notes. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nPREVIOUS STILL USAGE (follow exactly — what to take from the attached still and what to exclude; it governs place and objects only, never who is in this shot or how the camera sees them): Take the cypress-lined uphill path, exposed soil, supply cart, and daylight visible in the reference. Exclude duplicated military equipment and any dust cloud unique to the earlier advance.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The machine gun remains mounted on the supply cart in the crowded mountain column; the lead horse is beside the front of the formation. 헌병 장교: He remains mounted with his military sword drawn in a cutting stroke. 배상도: His neck is being cut while his grip on the reins fails.\n\nTHE HAND THAT IS DOING THIS: the object at the centre of this shot is being held, operated, read, aimed or handed over by 헌병 장교 right now, so 헌병 장교's hand — and as much of the wrist and forearm as the framing reaches — is in the frame, gripping or resting on that object exactly the way its real use requires. Match that hand to 헌병 장교: its size, build, skin, age, grooming, sleeve and anything worn on it belong to that person and to no one else.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 배상도 (한국인 남성, 성인 나이대, 자연스러운 검은 머리); 헌병 장교 (일본인 남성, 성인 나이대, 자연스러운 검은 머리) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.",
   "B": "Create ONE FINAL photorealistic live-action film still of the moment below — early 20th-century Korea, Jeolla region; all people are Korean unless explicitly stated otherwise. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 허공을 가른 헌병 장교의 서늘한 군도가 배상도의 목에 사선으로 맹렬하게 파고든 타격의 찰나.\n\nLOCATION (lock): The front of the military column on the mountain camp's uphill forest road, beside the mounted commander's horse. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: close-up\n- KEY BACKGROUND ELEMENTS: Saber (Meeting 배상도's neck during the diagonal strike) — Runs from the officer's upper-right arm toward the center-left contact; used as Links attacker and victim through one readable diagonal while occupying a small fraction of the image; Reins (Beginning to slip from 배상도's hand) — A short visible section leads from the lower frame toward the horse outside the crop; used as Provides a secondary action cue at the lower edge without requiring a separate insert.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Maintain the established natural daylight and muted contrast, rendering the sudden violence plainly without a flash, colored accent or glamorized blade highlight.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The people visible in that photo are NOT in this shot: never carry their faces, bodies or clothing onto anyone here. Each person in THIS shot is defined solely by the PEOPLE list, their own CHARACTER REFERENCE images and their wardrobe notes. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nPREVIOUS STILL USAGE (follow exactly — what to take from the attached still and what to exclude; it governs place and objects only, never who is in this shot or how the camera sees them): Take the cypress-lined uphill path, exposed soil, supply cart, and daylight visible in the reference. Exclude duplicated military equipment and any dust cloud unique to the earlier advance.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The machine gun remains mounted on the supply cart in the crowded mountain column; the lead horse is beside the front of the formation. 헌병 장교: He remains mounted with his military sword drawn in a cutting stroke. 배상도: His neck is being cut while his grip on the reins fails.\n\nTHE HAND THAT IS DOING THIS: the object at the centre of this shot is being held, operated, read, aimed or handed over by 헌병 장교 right now, so 헌병 장교's hand — and as much of the wrist and forearm as the framing reaches — is in the frame, gripping or resting on that object exactly the way its real use requires. Match that hand to 헌병 장교: its size, build, skin, age, grooming, sleeve and anything worn on it belong to that person and to no one else.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 배상도 (한국인 남성, 성인 나이대, 자연스러운 검은 머리); 헌병 장교 (일본인 남성, 성인 나이대, 자연스러운 검은 머리) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.\n\nCOMPOSITION VARIATION (this candidate only): keep the place, the people and the moment exactly as specified, and keep from the CAMERA & FRAME contract its subject scale, screen placement and key background elements exactly. For this candidate only, its camera angle and height are deliberately left open: choose a camera position distinctly different from the obvious one, as a film director picking a second setup on the same blocking. Stay true to any explicit framing the SHOT TEXT itself dictates."
  },
  "shot_run_spend_attempt_count": 1,
  "readings": [
   {
    "label": "A",
    "direction": "우측 상단에서 뻗어 나온 헌병 장교의 군도가 사선으로 내려와 좌측 중앙에 위치한 배상도의 목에 닿아 있습니다. 배상도의 시선은 위쪽을 향하고 있습니다.",
    "built_space": "삼나무가 늘어선 오르막 흙길의 배경이 레퍼런스의 환경과 일치하게 묘사되었으며, 클로즈업 샷에 맞게 배경의 스케일이 적절히 조절되었습니다.",
    "entities": "배상도는 피 묻은 셔츠와 검은 조끼 등 캐릭터 레퍼런스와 일치하는 외형을 갖추고 있습니다. 헌병 장교는 제복을 입고 있으며, 제시된 군도와 고삐 소품이 명확히 등장합니다. (단, 군도의 칼등이 목에 닿는 방향으로 쥐어져 있습니다.)",
    "hard_violations": [],
    "physics": "헌병 장교의 오른손이 군도의 손잡이를 쥐고 공간에 지지되어 있으며, 화면 하단에서는 배상도의 손이 가죽 고삐를 쥐고 있어 허공에 떠 있는 사물 없이 물리적 접촉이 올바릅니다."
   },
   {
    "label": "B",
    "direction": "헌병 장교의 시선이 아래의 배상도를 향하고 있으나, 지시된 무기의 겨냥이나 타격 방향성은 전혀 존재하지 않습니다.",
    "built_space": "오르막길과 멀리 보이는 초가집, 울타리 등 레퍼런스의 배경 요소들이 화면 깊숙이 넓게 배치되어 있습니다.",
    "entities": "헌병 장교의 제복과 군도(칼집에 꽂힌 상태)는 레퍼런스와 일치하나, 배상도는 검은 조끼를 입지 않았고 피와 오염이 없는 깨끗한 옷을 입고 있어 캐릭터 레퍼런스의 상태와 불일치합니다.",
    "hard_violations": [
     "[gemini-pro] 프롬프트의 핵심 액션인 '군도가 목에 파고든 타격의 찰나'를 완전히 누락하고 군도가 칼집에 들어간 상태로 연출됨",
     "[gemini-pro] 프롬프트가 엄격하게 지시한 '클로즈업(Close-up)' 구도를 무시하고 인물들의 전신과 배경이 넓게 보이는 샷으로 연출됨"
    ],
    "physics": "헌병 장교는 말 안장에 앉아 있고 배상도는 두 손으로 고삐를 쥐고 있어 인물과 사물의 지지 상태는 정상적입니다."
   }
  ],
  "cross_model_order": {
   "policy": "cross_model_order_v2_slot0_fwd_slot1_rev_parallel2",
   "slots": [
    {
     "model": "gemini-pro",
     "order": "forward",
     "display_to_canonical": {
      "A": "A",
      "B": "B"
     },
     "raw": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 9,
        "verdict_ko": "프롬프트가 구체적으로 지시한 사선 방향의 군도 타격 액션, 클로즈업 구도, 화면 하단의 고삐 등 핵심 요구사항을 매우 충실하게 시각화했습니다."
       },
       {
        "label": "B",
        "score": 2,
        "verdict_ko": "가장 중요한 액션인 '목을 파고드는 군도의 타격'이 전혀 묘사되지 않고 칼이 칼집에 꽂혀 있으며, 요구된 클로즈업 구도와 배상도의 의상(피 묻은 조끼) 또한 레퍼런스를 무시했습니다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "우측 상단에서 뻗어 나온 헌병 장교의 군도가 사선으로 내려와 좌측 중앙에 위치한 배상도의 목에 닿아 있습니다. 배상도의 시선은 위쪽을 향하고 있습니다.",
        "built_space": "삼나무가 늘어선 오르막 흙길의 배경이 레퍼런스의 환경과 일치하게 묘사되었으며, 클로즈업 샷에 맞게 배경의 스케일이 적절히 조절되었습니다.",
        "entities": "배상도는 피 묻은 셔츠와 검은 조끼 등 캐릭터 레퍼런스와 일치하는 외형을 갖추고 있습니다. 헌병 장교는 제복을 입고 있으며, 제시된 군도와 고삐 소품이 명확히 등장합니다. (단, 군도의 칼등이 목에 닿는 방향으로 쥐어져 있습니다.)",
        "hard_violations": [],
        "physics": "헌병 장교의 오른손이 군도의 손잡이를 쥐고 공간에 지지되어 있으며, 화면 하단에서는 배상도의 손이 가죽 고삐를 쥐고 있어 허공에 떠 있는 사물 없이 물리적 접촉이 올바릅니다."
       },
       {
        "label": "B",
        "direction": "헌병 장교의 시선이 아래의 배상도를 향하고 있으나, 지시된 무기의 겨냥이나 타격 방향성은 전혀 존재하지 않습니다.",
        "built_space": "오르막길과 멀리 보이는 초가집, 울타리 등 레퍼런스의 배경 요소들이 화면 깊숙이 넓게 배치되어 있습니다.",
        "entities": "헌병 장교의 제복과 군도(칼집에 꽂힌 상태)는 레퍼런스와 일치하나, 배상도는 검은 조끼를 입지 않았고 피와 오염이 없는 깨끗한 옷을 입고 있어 캐릭터 레퍼런스의 상태와 불일치합니다.",
        "hard_violations": [
         "프롬프트의 핵심 액션인 '군도가 목에 파고든 타격의 찰나'를 완전히 누락하고 군도가 칼집에 들어간 상태로 연출됨",
         "프롬프트가 엄격하게 지시한 '클로즈업(Close-up)' 구도를 무시하고 인물들의 전신과 배경이 넓게 보이는 샷으로 연출됨"
        ],
        "physics": "헌병 장교는 말 안장에 앉아 있고 배상도는 두 손으로 고삐를 쥐고 있어 인물과 사물의 지지 상태는 정상적입니다."
       }
      ],
      "all_candidates_fail": false
     },
     "normalized": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 9,
        "verdict_ko": "프롬프트가 구체적으로 지시한 사선 방향의 군도 타격 액션, 클로즈업 구도, 화면 하단의 고삐 등 핵심 요구사항을 매우 충실하게 시각화했습니다."
       },
       {
        "label": "B",
        "score": 2,
        "verdict_ko": "가장 중요한 액션인 '목을 파고드는 군도의 타격'이 전혀 묘사되지 않고 칼이 칼집에 꽂혀 있으며, 요구된 클로즈업 구도와 배상도의 의상(피 묻은 조끼) 또한 레퍼런스를 무시했습니다."
       }
      ],
      "all_candidates_fail": false,
      "readings": [
       {
        "label": "A",
        "direction": "우측 상단에서 뻗어 나온 헌병 장교의 군도가 사선으로 내려와 좌측 중앙에 위치한 배상도의 목에 닿아 있습니다. 배상도의 시선은 위쪽을 향하고 있습니다.",
        "built_space": "삼나무가 늘어선 오르막 흙길의 배경이 레퍼런스의 환경과 일치하게 묘사되었으며, 클로즈업 샷에 맞게 배경의 스케일이 적절히 조절되었습니다.",
        "entities": "배상도는 피 묻은 셔츠와 검은 조끼 등 캐릭터 레퍼런스와 일치하는 외형을 갖추고 있습니다. 헌병 장교는 제복을 입고 있으며, 제시된 군도와 고삐 소품이 명확히 등장합니다. (단, 군도의 칼등이 목에 닿는 방향으로 쥐어져 있습니다.)",
        "hard_violations": [],
        "physics": "헌병 장교의 오른손이 군도의 손잡이를 쥐고 공간에 지지되어 있으며, 화면 하단에서는 배상도의 손이 가죽 고삐를 쥐고 있어 허공에 떠 있는 사물 없이 물리적 접촉이 올바릅니다."
       },
       {
        "label": "B",
        "direction": "헌병 장교의 시선이 아래의 배상도를 향하고 있으나, 지시된 무기의 겨냥이나 타격 방향성은 전혀 존재하지 않습니다.",
        "built_space": "오르막길과 멀리 보이는 초가집, 울타리 등 레퍼런스의 배경 요소들이 화면 깊숙이 넓게 배치되어 있습니다.",
        "entities": "헌병 장교의 제복과 군도(칼집에 꽂힌 상태)는 레퍼런스와 일치하나, 배상도는 검은 조끼를 입지 않았고 피와 오염이 없는 깨끗한 옷을 입고 있어 캐릭터 레퍼런스의 상태와 불일치합니다.",
        "hard_violations": [
         "프롬프트의 핵심 액션인 '군도가 목에 파고든 타격의 찰나'를 완전히 누락하고 군도가 칼집에 들어간 상태로 연출됨",
         "프롬프트가 엄격하게 지시한 '클로즈업(Close-up)' 구도를 무시하고 인물들의 전신과 배경이 넓게 보이는 샷으로 연출됨"
        ],
        "physics": "헌병 장교는 말 안장에 앉아 있고 배상도는 두 손으로 고삐를 쥐고 있어 인물과 사물의 지지 상태는 정상적입니다."
       }
      ]
     },
     "ok": true
    },
    {
     "model": "gpt-high",
     "order": "reverse",
     "display_to_canonical": {
      "A": "B",
      "B": "A"
     },
     "raw": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 2,
        "verdict_ko": "장소와 기승 상태는 이어지지만, 군도가 칼집에 있고 배상도가 고삐를 단단히 잡고 있어 목을 베는 타격의 찰나를 구현하지 못했다."
       },
       {
        "label": "B",
        "score": 8,
        "verdict_ko": "밀착 구도에서 장교의 손과 군도가 우상단에서 배상도의 목으로 이어지는 사선을 구현했으나, 맹렬한 타격과 고삐를 놓치는 순간의 표현은 다소 약하다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "배상도는 정면보다 약간 위를 바라보고, 우상단의 장교는 배상도 쪽으로 고개를 숙인다. 장교의 팔은 배상도의 어깨 뒤로 내려가지만 손과 뽑힌 칼날은 보이지 않는다. 군도는 장교의 허리 옆 칼집에 있어 목을 향하는 사선이나 접촉이 없다. 고삐는 배상도의 양손에서 오른쪽 말 쪽으로 이어진다.",
        "built_space": "침엽수 사이의 돌과 흙이 드러난 경사로, 상단의 울타리와 지붕들이 참고 장소를 따른다. 수레는 가까운 오른쪽 배경과 먼 중앙에 하나씩 보이며, 가까운 수레 앞에는 말 두 필이 보인다. 장교가 탄 말과 안장은 오른쪽 전경에 일부 보인다. 배상도는 말 옆에 있지만 얼굴과 목의 클로즈업보다 상반신과 배경을 넓게 담았다.",
        "entities": "등장 인물은 성인 동아시아계 남성 두 명으로, 장교의 일부 얼굴과 배상도의 얼굴이 보인다. 국적은 외모만으로 확인할 수 없다. 배상도는 검은 머리지만 참고의 얼굴 인상과 차이가 있고, 검은 조끼와 두드러진 얼굴 상처도 빠졌다. 장교의 군복, 훈장, 모자와 승마 장비는 대체로 부합한다. 고삐와 칼집에 든 군도는 있으나 핵심 소품인 타격 중의 노출된 칼날은 없다. 수레의 기관총은 식별되지 않으며 읽을 수 있는 글자는 보이지 않는다.",
        "hard_violations": [],
        "physics": "장교의 허벅지와 안장이 맞닿아 기승을 지지한다. 배상도의 하체는 화면 밖이며 공중에 떠 있다는 징후는 없다. 고삐는 양손으로 확실히 잡혀 있고 군도는 허리의 칼집에 지지된다. 물체의 지지는 자연스럽지만, 베기 동작이나 충격에 따른 반응, 손아귀가 풀리는 행동은 보이지 않는다."
       },
       {
        "label": "B",
        "direction": "장교의 오른쪽 위 팔에서 손잡이와 칼날이 중앙 왼쪽 배상도의 목으로 사선 연결되고, 칼날이 목 앞쪽에 닿아 있다. 배상도는 고개를 뒤로 젖혀 화면 오른쪽 위 장교 쪽을 바라본다. 장교의 눈은 화면 밖이다. 하단의 고삐는 배상도의 손에서 오른쪽 화면 밖 말 방향으로 이어진다.",
        "built_space": "배상도의 얼굴과 목, 장교의 손과 팔을 중심으로 가까이 잡았으며, 오른쪽에는 장교의 허벅지와 안장 일부가 보인다. 배경은 참고와 유사한 침엽수 줄기들과 노출된 흙의 경사로이다. 수레, 기관총, 울타리와 건물은 이 밀착 구도에서 확인되지 않지만 이를 보여주기 위해 화면을 넓히지 않았다. 중복 설비나 불가능한 반사는 보이지 않는다.",
        "entities": "검은 머리의 성인 동아시아계 남성인 배상도가 보이며, 얼굴과 체격은 참고에 대체로 가깝다. 피와 흙이 묻은 밝은 저고리와 검은 조끼도 이어진다. 장교는 얼굴 없이 성인 남성의 맨손, 손목, 군복 소매와 몸통 일부로 나타나며 참고 복장과 대체로 맞는다. 군도는 감긴 손잡이, 둥근 가드와 금속 칼날을 갖춰 소품 참고와 유사하지만 칼날의 곡률과 마모 표현은 다르다. 고삐와 안장도 보이고, 추가 인물이나 읽을 수 있는 글자는 없다.",
        "hard_violations": [],
        "physics": "군도는 장교의 손이 손잡이를 감싸 쥐어 지지하고, 손목과 팔이 칼날의 사선으로 연결된다. 장교의 허벅지는 안장에 놓여 기승 상태가 읽힌다. 배상도의 젖혀진 목과 상체는 칼날 접촉에 반응하는 자세로 가능하며, 하체가 잘렸다는 이유로 부유한다고 볼 근거는 없다. 고삐는 배상도의 손가락 사이에 여전히 걸려 있어 지지가 자연스럽지만, 완전히 미끄러지는 순간은 명확하지 않다. 목 접촉은 보이나 격렬하게 파고드는 충격보다는 칼을 댄 순간에 가깝게 읽히는 부분이 있다."
       }
      ],
      "all_candidates_fail": false
     },
     "normalized": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 2,
        "verdict_ko": "장소와 기승 상태는 이어지지만, 군도가 칼집에 있고 배상도가 고삐를 단단히 잡고 있어 목을 베는 타격의 찰나를 구현하지 못했다."
       },
       {
        "label": "A",
        "score": 8,
        "verdict_ko": "밀착 구도에서 장교의 손과 군도가 우상단에서 배상도의 목으로 이어지는 사선을 구현했으나, 맹렬한 타격과 고삐를 놓치는 순간의 표현은 다소 약하다."
       }
      ],
      "all_candidates_fail": false,
      "readings": [
       {
        "label": "B",
        "direction": "배상도는 정면보다 약간 위를 바라보고, 우상단의 장교는 배상도 쪽으로 고개를 숙인다. 장교의 팔은 배상도의 어깨 뒤로 내려가지만 손과 뽑힌 칼날은 보이지 않는다. 군도는 장교의 허리 옆 칼집에 있어 목을 향하는 사선이나 접촉이 없다. 고삐는 배상도의 양손에서 오른쪽 말 쪽으로 이어진다.",
        "built_space": "침엽수 사이의 돌과 흙이 드러난 경사로, 상단의 울타리와 지붕들이 참고 장소를 따른다. 수레는 가까운 오른쪽 배경과 먼 중앙에 하나씩 보이며, 가까운 수레 앞에는 말 두 필이 보인다. 장교가 탄 말과 안장은 오른쪽 전경에 일부 보인다. 배상도는 말 옆에 있지만 얼굴과 목의 클로즈업보다 상반신과 배경을 넓게 담았다.",
        "entities": "등장 인물은 성인 동아시아계 남성 두 명으로, 장교의 일부 얼굴과 배상도의 얼굴이 보인다. 국적은 외모만으로 확인할 수 없다. 배상도는 검은 머리지만 참고의 얼굴 인상과 차이가 있고, 검은 조끼와 두드러진 얼굴 상처도 빠졌다. 장교의 군복, 훈장, 모자와 승마 장비는 대체로 부합한다. 고삐와 칼집에 든 군도는 있으나 핵심 소품인 타격 중의 노출된 칼날은 없다. 수레의 기관총은 식별되지 않으며 읽을 수 있는 글자는 보이지 않는다.",
        "hard_violations": [],
        "physics": "장교의 허벅지와 안장이 맞닿아 기승을 지지한다. 배상도의 하체는 화면 밖이며 공중에 떠 있다는 징후는 없다. 고삐는 양손으로 확실히 잡혀 있고 군도는 허리의 칼집에 지지된다. 물체의 지지는 자연스럽지만, 베기 동작이나 충격에 따른 반응, 손아귀가 풀리는 행동은 보이지 않는다."
       },
       {
        "label": "A",
        "direction": "장교의 오른쪽 위 팔에서 손잡이와 칼날이 중앙 왼쪽 배상도의 목으로 사선 연결되고, 칼날이 목 앞쪽에 닿아 있다. 배상도는 고개를 뒤로 젖혀 화면 오른쪽 위 장교 쪽을 바라본다. 장교의 눈은 화면 밖이다. 하단의 고삐는 배상도의 손에서 오른쪽 화면 밖 말 방향으로 이어진다.",
        "built_space": "배상도의 얼굴과 목, 장교의 손과 팔을 중심으로 가까이 잡았으며, 오른쪽에는 장교의 허벅지와 안장 일부가 보인다. 배경은 참고와 유사한 침엽수 줄기들과 노출된 흙의 경사로이다. 수레, 기관총, 울타리와 건물은 이 밀착 구도에서 확인되지 않지만 이를 보여주기 위해 화면을 넓히지 않았다. 중복 설비나 불가능한 반사는 보이지 않는다.",
        "entities": "검은 머리의 성인 동아시아계 남성인 배상도가 보이며, 얼굴과 체격은 참고에 대체로 가깝다. 피와 흙이 묻은 밝은 저고리와 검은 조끼도 이어진다. 장교는 얼굴 없이 성인 남성의 맨손, 손목, 군복 소매와 몸통 일부로 나타나며 참고 복장과 대체로 맞는다. 군도는 감긴 손잡이, 둥근 가드와 금속 칼날을 갖춰 소품 참고와 유사하지만 칼날의 곡률과 마모 표현은 다르다. 고삐와 안장도 보이고, 추가 인물이나 읽을 수 있는 글자는 없다.",
        "hard_violations": [],
        "physics": "군도는 장교의 손이 손잡이를 감싸 쥐어 지지하고, 손목과 팔이 칼날의 사선으로 연결된다. 장교의 허벅지는 안장에 놓여 기승 상태가 읽힌다. 배상도의 젖혀진 목과 상체는 칼날 접촉에 반응하는 자세로 가능하며, 하체가 잘렸다는 이유로 부유한다고 볼 근거는 없다. 고삐는 배상도의 손가락 사이에 여전히 걸려 있어 지지가 자연스럽지만, 완전히 미끄러지는 순간은 명확하지 않다. 목 접촉은 보이나 격렬하게 파고드는 충격보다는 칼을 댄 순간에 가깝게 읽히는 부분이 있다."
       }
      ]
     },
     "ok": true
    }
   ],
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "slot_winner_match": true,
   "slot_winner": {
    "gemini-pro": "A",
    "gpt-high": "A"
   },
   "route": "cross_slot_agree"
  },
  "dual": {
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "normalized": {
    "A": 2.0,
    "B": 0.472
   },
   "adjusted": {
    "A": 2.0,
    "B": 0.222
   },
   "violations": {
    "B": [
     "[gemini-pro] 프롬프트의 핵심 액션인 '군도가 목에 파고든 타격의 찰나'를 완전히 누락하고 군도가 칼집에 들어간 상태로 연출됨",
     "[gemini-pro] 프롬프트가 엄격하게 지시한 '클로즈업(Close-up)' 구도를 무시하고 인물들의 전신과 배경이 넓게 보이는 샷으로 연출됨"
    ]
   },
   "per_model_winner": {
    "gemini-pro": "A",
    "gpt-high": "A"
   },
   "agreed": true
  },
  "totals": {
   "A": 2000,
   "B": 222
  },
  "selected": "A",
  "ranking": [
   "A",
   "B"
  ],
  "verdicts": [
   {
    "label": "A",
    "score": 2000,
    "verdict_ko": "프롬프트가 구체적으로 지시한 사선 방향의 군도 타격 액션, 클로즈업 구도, 화면 하단의 고삐 등 핵심 요구사항을 매우 충실하게 시각화했습니다."
   },
   {
    "label": "B",
    "score": 222,
    "verdict_ko": "가장 중요한 액션인 '목을 파고드는 군도의 타격'이 전혀 묘사되지 않고 칼이 칼집에 꽂혀 있으며, 요구된 클로즈업 구도와 배상도의 의상(피 묻은 조끼) 또한 레퍼런스를 무시했습니다.  ★위반: [gemini-pro] 프롬프트의 핵심 액션인 '군도가 목에 파고든 타격의 찰나'를 완전히 누락하고 군도가 칼집에 들어간 상태로 연출됨 / [gemini-pro] 프롬프트가 엄격하게 지시한 '클로즈업(Close-up)' 구도를 무시하고 인물들의 전신과 배경이 넓게 보이는 샷으로 연출됨"
   }
  ],
  "refs": [
   {
    "label": "PREVIOUS SHOT STILL — a visually related earlier shot of this same place: the location's look, materials, fixed features and lighting mood are LOCKED to this photo; never copy its camera framing. The people visible in that photo are NOT in this shot: never carry their faces, bodies or clothing onto anyone here. Each person in THIS shot is defined solely by the PEOPLE list, their own CHARACTER REFERENCE images and their wardrobe notes. If this photo shows a character who cannot move (dead or unconscious), that character's exact pose, position and orientation are ALSO LOCKED — treat their body as a fixed prop of the set that only the camera moves around.",
    "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/db00517b-165b-4005-a287-f28936b59b95/images/272bd7af-9294-4490-98c6-88a40e9deacc/scene/recipe/S65sh1_sel.png",
    "asset_id": "32bbf9fa-bf02-4afa-ab0f-9e1880260398",
    "role": "prev_still"
   },
   {
    "label": "CHARACTER REFERENCE — 배상도: the exact person appearing in this shot; match face, hair and build exactly.",
    "path": "<bytes:1240915>",
    "asset_id": "38d95f0e-1467-4f8b-8886-b47f270b7adc",
    "role": "character_ref"
   },
   {
    "label": "CHARACTER REFERENCE — 헌병 장교: the exact person appearing in this shot; match face, hair and build exactly.",
    "path": "<bytes:862147>",
    "asset_id": "83cd18a8-56b7-4e8a-8937-e155fe16c041",
    "role": "character_ref"
   },
   {
    "label": "PROP REFERENCE — 칼: the exact object appearing in this shot; match its look, material and wear exactly.",
    "path": "<bytes:780122>",
    "asset_id": "82614014-6091-423c-89ca-1e15d209de72",
    "role": "prop_ref"
   }
  ],
  "critique_skipped": true,
  "shot_run_produced": false,
  "shot_run_uid": "06aa1446-72c7-7ad0-a30f-6b3485047d9c",
  "ref_mode": "prev+엔티티",
  "share_plan": {
   "ref_plan": "prev",
   "prev_anchor_tag": "S65sh1"
  }
 },
 "S65sh14::cine": {
  "applied": true,
  "attempted_at": "2026-09-09T08:36:50.032942+00:00",
  "fingerprint": "9c1ca796c81441ca2a568882fec9eb8689bc8f89e2469ad7fcf940dcddbaeb3e",
  "fingerprint_version": 2,
  "provider": "mai",
  "endpoint": "openrouter/chat-completions",
  "model": "microsoft/mai-image-2.6",
  "pack": "24.202608252115",
  "source_file": "S65sh14_sel.png",
  "source_sha256": "66ae22d47398182f2a15cbb55976b870af4260166680b6ded18214c304136ae1",
  "file": "S65sh14_cine.png",
  "staged_sha256": "cb57e21ee63c954e9873c10e1c772c806b10bd460bee84bd82a6b13475951a2b",
  "latency_ms": 26863
 },
 "S65sh19::signage": {
  "fp": "8cb69397b2526969",
  "inscriptions": [],
  "cues": [],
  "dropped": []
 },
 "S65sh19": {
  "input_fingerprint": "6abcf3c5adb09a91",
  "prompt": "Create ONE FINAL photorealistic live-action film still of the moment below — early 20th-century Korea, Jeolla region; all people are Korean unless explicitly stated otherwise. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 피가 털린 군도를 곧게 든 채 다시 산 위를 서늘하게 쏘아보는 헌병 장교의 굳은 정면.\n\nLOCATION (lock): The mounted command position at the head of the column on the forest road ascending toward the mountain camp. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: medium shot\n- KEY BACKGROUND ELEMENTS: 군도 (Drawn and held upright after blood has been flicked from it) — The blade rises beside the commander's torso, with the grip visible near the lower frame; used as Connects the preceding killing to the commander's unbroken uphill attention; 산길 주변 나무 (Trees border the mountain path); used as Provide a restrained background behind the mounted figure without revealing a new spatial axis.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Daylight with restrained saturation and naturalistic contrast keeps his expression cold without stylized facial shadows.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The clothing, hair and overall look of 헌병 장교 — who appear both in that photo and in this shot — are LOCKED to that photo. Anyone else visible in that photo is NOT in this shot: never carry their face, body or clothing onto anyone here. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nPREVIOUS STILL USAGE (follow exactly — what to take from the attached still and what to exclude; it governs place and objects only, never who is in this shot or how the camera sees them): Take the uphill forest path, cypress trunks, supply cart, and daytime light. Exclude the civilian guide from the advancing group and do not reproduce airborne blood from the preceding strike.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The machine-gun cart is advancing again; the lead horse's reins are no longer held from the ground. 헌병 장교: He remains mounted holding his drawn sword after flicking blood from its blade.\n\nTHE HAND THAT IS DOING THIS: the object at the centre of this shot is being held, operated, read, aimed or handed over by 헌병 장교 right now, so 헌병 장교's hand — and as much of the wrist and forearm as the framing reaches — is in the frame, gripping or resting on that object exactly the way its real use requires. Match that hand to 헌병 장교: its size, build, skin, age, grooming, sleeve and anything worn on it belong to that person and to no one else.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 헌병 장교 (일본인 남성, 성인 나이대, 자연스러운 검은 머리) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.",
  "roll_prompts": {
   "A": "Create ONE FINAL photorealistic live-action film still of the moment below — early 20th-century Korea, Jeolla region; all people are Korean unless explicitly stated otherwise. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 피가 털린 군도를 곧게 든 채 다시 산 위를 서늘하게 쏘아보는 헌병 장교의 굳은 정면.\n\nLOCATION (lock): The mounted command position at the head of the column on the forest road ascending toward the mountain camp. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: medium shot\n- KEY BACKGROUND ELEMENTS: 군도 (Drawn and held upright after blood has been flicked from it) — The blade rises beside the commander's torso, with the grip visible near the lower frame; used as Connects the preceding killing to the commander's unbroken uphill attention; 산길 주변 나무 (Trees border the mountain path); used as Provide a restrained background behind the mounted figure without revealing a new spatial axis.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Daylight with restrained saturation and naturalistic contrast keeps his expression cold without stylized facial shadows.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The clothing, hair and overall look of 헌병 장교 — who appear both in that photo and in this shot — are LOCKED to that photo. Anyone else visible in that photo is NOT in this shot: never carry their face, body or clothing onto anyone here. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nPREVIOUS STILL USAGE (follow exactly — what to take from the attached still and what to exclude; it governs place and objects only, never who is in this shot or how the camera sees them): Take the uphill forest path, cypress trunks, supply cart, and daytime light. Exclude the civilian guide from the advancing group and do not reproduce airborne blood from the preceding strike.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The machine-gun cart is advancing again; the lead horse's reins are no longer held from the ground. 헌병 장교: He remains mounted holding his drawn sword after flicking blood from its blade.\n\nTHE HAND THAT IS DOING THIS: the object at the centre of this shot is being held, operated, read, aimed or handed over by 헌병 장교 right now, so 헌병 장교's hand — and as much of the wrist and forearm as the framing reaches — is in the frame, gripping or resting on that object exactly the way its real use requires. Match that hand to 헌병 장교: its size, build, skin, age, grooming, sleeve and anything worn on it belong to that person and to no one else.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 헌병 장교 (일본인 남성, 성인 나이대, 자연스러운 검은 머리) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.",
   "B": "Create ONE FINAL photorealistic live-action film still of the moment below — early 20th-century Korea, Jeolla region; all people are Korean unless explicitly stated otherwise. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 피가 털린 군도를 곧게 든 채 다시 산 위를 서늘하게 쏘아보는 헌병 장교의 굳은 정면.\n\nLOCATION (lock): The mounted command position at the head of the column on the forest road ascending toward the mountain camp. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: medium shot\n- KEY BACKGROUND ELEMENTS: 군도 (Drawn and held upright after blood has been flicked from it) — The blade rises beside the commander's torso, with the grip visible near the lower frame; used as Connects the preceding killing to the commander's unbroken uphill attention; 산길 주변 나무 (Trees border the mountain path); used as Provide a restrained background behind the mounted figure without revealing a new spatial axis.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Daylight with restrained saturation and naturalistic contrast keeps his expression cold without stylized facial shadows.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The clothing, hair and overall look of 헌병 장교 — who appear both in that photo and in this shot — are LOCKED to that photo. Anyone else visible in that photo is NOT in this shot: never carry their face, body or clothing onto anyone here. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nPREVIOUS STILL USAGE (follow exactly — what to take from the attached still and what to exclude; it governs place and objects only, never who is in this shot or how the camera sees them): Take the uphill forest path, cypress trunks, supply cart, and daytime light. Exclude the civilian guide from the advancing group and do not reproduce airborne blood from the preceding strike.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The machine-gun cart is advancing again; the lead horse's reins are no longer held from the ground. 헌병 장교: He remains mounted holding his drawn sword after flicking blood from its blade.\n\nTHE HAND THAT IS DOING THIS: the object at the centre of this shot is being held, operated, read, aimed or handed over by 헌병 장교 right now, so 헌병 장교's hand — and as much of the wrist and forearm as the framing reaches — is in the frame, gripping or resting on that object exactly the way its real use requires. Match that hand to 헌병 장교: its size, build, skin, age, grooming, sleeve and anything worn on it belong to that person and to no one else.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 헌병 장교 (일본인 남성, 성인 나이대, 자연스러운 검은 머리) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.\n\nCOMPOSITION VARIATION (this candidate only): keep the place, the people and the moment exactly as specified, and keep from the CAMERA & FRAME contract its subject scale, screen placement and key background elements exactly. For this candidate only, its camera angle and height are deliberately left open: choose a camera position distinctly different from the obvious one, as a film director picking a second setup on the same blocking. Stay true to any explicit framing the SHOT TEXT itself dictates."
  },
  "shot_run_spend_attempt_count": 1,
  "readings": [
   {
    "label": "A",
    "direction": "인물은 정면을 바라보고 있으며, 가슴 앞쪽 중앙에 군도를 수직으로 들고 있습니다.",
    "built_space": "숲속 산길 배경이며, 인물 뒤쪽에 덮개가 있는 수레가 보입니다.",
    "entities": "캐릭터 레퍼런스의 헌병 장교가 아닌 이전 스틸컷의 민간인 얼굴이 잘못 생성되었으며, 장교의 필수 복장인 군모와 훈장이 없습니다.",
    "hard_violations": [
     "[gemini-pro] 이전 샷에 등장한 다른 인물(민간인)의 얼굴을 장교에게 덧씌움",
     "[gemini-pro] 캐릭터 레퍼런스에 있는 필수 복장(군모, 훈장) 누락",
     "[gpt-high] 소매 패치의 한자 표기가 판독 가능한 형태로 노출되어, 이미지 어디에도 읽을 수 있는 문자를 두지 말라는 금지 조건을 위반한다."
    ],
    "physics": "두 손으로 군도의 손잡이를 쥐고 있으며, 하단에 말 머리와 귀가 보여 말 위에 앉아있는 자세를 유지하고 있습니다."
   },
   {
    "label": "B",
    "direction": "인물은 정면(산 위쪽)을 차갑게 응시하고 있으며, 오른손으로 군도를 몸통 옆에 꼿꼿이 세워 들고 있습니다.",
    "built_space": "오르막 산길 배경으로 양옆에 나무들이 늘어서 있으며, 인물 뒤편으로 기관총이 실린 수레가 배치되어 있습니다.",
    "entities": "헌병 장교의 얼굴, 군모, 훈장 및 제복이 캐릭터 레퍼런스와 정확히 일치하며, 들고 있는 군도 또한 소품 레퍼런스의 형태와 부합합니다.",
    "hard_violations": [
     "[gpt-high] 소매 패치의 한자 표기가 판독 가능한 형태로 노출되어, 이미지 어디에도 읽을 수 있는 문자를 두지 말라는 금지 조건을 위반한다."
    ],
    "physics": "오른손이 군도의 손잡이를 자연스럽게 쥐어 지탱하고 있으며, 화면 하단에 말의 일부가 나타나 기마 상태임을 물리적으로 뒷받침합니다."
   }
  ],
  "cross_model_order": {
   "policy": "cross_model_order_v2_slot0_fwd_slot1_rev_parallel2",
   "slots": [
    {
     "model": "gemini-pro",
     "order": "forward",
     "display_to_canonical": {
      "A": "A",
      "B": "B"
     },
     "raw": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 10,
        "verdict_ko": "캐릭터 레퍼런스의 헌병 장교 얼굴과 복장(군모, 훈장 포함)을 완벽하게 재현했으며, 군도를 세워 든 채 정면을 쏘아보는 샷의 요구사항을 정확히 구현했습니다."
       },
       {
        "label": "A",
        "score": 2,
        "verdict_ko": "프롬프트에서 엄격히 금지한 이전 샷의 민간인 얼굴을 헌병 장교에게 잘못 적용하였고, 필수적인 군모와 훈장도 누락하여 지시사항을 크게 위반했습니다."
       }
      ],
      "readings": [
       {
        "label": "B",
        "direction": "인물은 정면(산 위쪽)을 차갑게 응시하고 있으며, 오른손으로 군도를 몸통 옆에 꼿꼿이 세워 들고 있습니다.",
        "built_space": "오르막 산길 배경으로 양옆에 나무들이 늘어서 있으며, 인물 뒤편으로 기관총이 실린 수레가 배치되어 있습니다.",
        "entities": "헌병 장교의 얼굴, 군모, 훈장 및 제복이 캐릭터 레퍼런스와 정확히 일치하며, 들고 있는 군도 또한 소품 레퍼런스의 형태와 부합합니다.",
        "hard_violations": [],
        "physics": "오른손이 군도의 손잡이를 자연스럽게 쥐어 지탱하고 있으며, 화면 하단에 말의 일부가 나타나 기마 상태임을 물리적으로 뒷받침합니다."
       },
       {
        "label": "A",
        "direction": "인물은 정면을 바라보고 있으며, 가슴 앞쪽 중앙에 군도를 수직으로 들고 있습니다.",
        "built_space": "숲속 산길 배경이며, 인물 뒤쪽에 덮개가 있는 수레가 보입니다.",
        "entities": "캐릭터 레퍼런스의 헌병 장교가 아닌 이전 스틸컷의 민간인 얼굴이 잘못 생성되었으며, 장교의 필수 복장인 군모와 훈장이 없습니다.",
        "hard_violations": [
         "이전 샷에 등장한 다른 인물(민간인)의 얼굴을 장교에게 덧씌움",
         "캐릭터 레퍼런스에 있는 필수 복장(군모, 훈장) 누락"
        ],
        "physics": "두 손으로 군도의 손잡이를 쥐고 있으며, 하단에 말 머리와 귀가 보여 말 위에 앉아있는 자세를 유지하고 있습니다."
       }
      ],
      "all_candidates_fail": false
     },
     "normalized": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 10,
        "verdict_ko": "캐릭터 레퍼런스의 헌병 장교 얼굴과 복장(군모, 훈장 포함)을 완벽하게 재현했으며, 군도를 세워 든 채 정면을 쏘아보는 샷의 요구사항을 정확히 구현했습니다."
       },
       {
        "label": "A",
        "score": 2,
        "verdict_ko": "프롬프트에서 엄격히 금지한 이전 샷의 민간인 얼굴을 헌병 장교에게 잘못 적용하였고, 필수적인 군모와 훈장도 누락하여 지시사항을 크게 위반했습니다."
       }
      ],
      "all_candidates_fail": false,
      "readings": [
       {
        "label": "B",
        "direction": "인물은 정면(산 위쪽)을 차갑게 응시하고 있으며, 오른손으로 군도를 몸통 옆에 꼿꼿이 세워 들고 있습니다.",
        "built_space": "오르막 산길 배경으로 양옆에 나무들이 늘어서 있으며, 인물 뒤편으로 기관총이 실린 수레가 배치되어 있습니다.",
        "entities": "헌병 장교의 얼굴, 군모, 훈장 및 제복이 캐릭터 레퍼런스와 정확히 일치하며, 들고 있는 군도 또한 소품 레퍼런스의 형태와 부합합니다.",
        "hard_violations": [],
        "physics": "오른손이 군도의 손잡이를 자연스럽게 쥐어 지탱하고 있으며, 화면 하단에 말의 일부가 나타나 기마 상태임을 물리적으로 뒷받침합니다."
       },
       {
        "label": "A",
        "direction": "인물은 정면을 바라보고 있으며, 가슴 앞쪽 중앙에 군도를 수직으로 들고 있습니다.",
        "built_space": "숲속 산길 배경이며, 인물 뒤쪽에 덮개가 있는 수레가 보입니다.",
        "entities": "캐릭터 레퍼런스의 헌병 장교가 아닌 이전 스틸컷의 민간인 얼굴이 잘못 생성되었으며, 장교의 필수 복장인 군모와 훈장이 없습니다.",
        "hard_violations": [
         "이전 샷에 등장한 다른 인물(민간인)의 얼굴을 장교에게 덧씌움",
         "캐릭터 레퍼런스에 있는 필수 복장(군모, 훈장) 누락"
        ],
        "physics": "두 손으로 군도의 손잡이를 쥐고 있으며, 하단에 말 머리와 귀가 보여 말 위에 앉아있는 자세를 유지하고 있습니다."
       }
      ]
     },
     "ok": true
    },
    {
     "model": "gpt-high",
     "order": "reverse",
     "display_to_canonical": {
      "A": "B",
      "B": "A"
     },
     "raw": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 5,
        "verdict_ko": "정면 미디엄 숏, 몸통 옆에 곧게 세운 군도, 장교의 얼굴·군모·훈장은 더 충실하지만, 뒤로 올라가는 산길과 반대인 시선 및 소매의 읽히는 문자 표기가 문제다."
       },
       {
        "label": "B",
        "score": 3,
        "verdict_ko": "기승과 칼을 쥔 손은 명확하지만, 산 위를 향하지 않는 시선에 더해 장교의 얼굴·군모·훈장 연속성이 깨지고 칼도 몸통 앞을 비스듬히 가로지른다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "장교는 얼굴을 정면으로 두고 카메라 부근을 응시한다. 보이는 산길은 장교 뒤쪽 화면 상단으로 올라가므로, 시선이 그 오르막을 향한다고 읽히지 않는다. 군도는 화면 왼쪽 몸통 옆에서 거의 수직으로 올라가며 칼끝이 위를 향한다. 뒤 수레의 기관총 총구는 대체로 카메라 쪽을 향하지만, 이 숏에는 기관총의 조준 대상이 지정되지 않았다.",
        "built_space": "인공 건축물은 없고, 나무줄기가 양옆을 둘러싼 돌 섞인 흙길 한 줄기가 뒤쪽 높은 곳으로 이어진다. 장교 뒤 화면 오른쪽에 기관총을 실은 수레 한 대가 일부 보인다. 참고의 숲·노면·주간 광선과 대체로 연결된다. 장교는 허리 부근까지 담겼고 말의 갈색 몸 일부가 아래쪽에 보이며, 중복 수레나 추가 인물은 없다. 다만 화면에 드러난 오르막은 장교가 바라보는 쪽이 아니라 등 뒤에 있다.",
        "entities": "성인 동아시아계 남성 장교 한 명이며, 외관만으로 일본 국적을 확인할 수는 없다. 얼굴은 인물 참고에 비교적 가깝고, 군모 아래 검은 머리, 갈색 계열 군복, 붉은 견장, 가슴 훈장, 가죽 대각선 띠와 허리띠가 유지된다. 칼 한 자루는 금속 날, 원형 코등이, 어두운 손잡이를 갖춰 소품 참고와 대체로 맞지만 날의 곡률과 표면 무늬는 덜 일치한다. 손잡이를 쥔 맨손과 군복 소매가 프레임 하단에 보인다. 민간인 안내자와 공중의 피는 없다. 소매 패치에는 흰 한자 표기가 식별된다.",
        "hard_violations": [
         "소매 패치의 한자 표기가 판독 가능한 형태로 노출되어, 이미지 어디에도 읽을 수 있는 문자를 두지 말라는 금지 조건을 위반한다."
        ],
        "physics": "오른손 손가락이 칼자루를 감싸고 있어 세워 든 칼의 지지가 명확하다. 팔꿈치를 굽힌 자세로 칼을 유지하는 것은 가능하다. 하단의 말 몸과 기승 자세가 장교의 지지를 설명하며, 안장 접촉부와 발이 잘렸다는 이유만으로 공중에 떠 있다고 볼 근거는 없다. 수레는 노면에 놓이고 기관총은 수레 위 거치대에 지지된다. 피를 턴 뒤 멈춘 순간은 가능하지만 수레가 다시 움직이는지는 정지 화면으로 확인되지 않는다."
       },
       {
        "label": "B",
        "direction": "장교의 눈은 카메라 부근 정면을 향한다. 산길은 그의 뒤에서 화면 상단 왼쪽으로 올라가므로, 보이는 산 위를 쏘아보는 방향과 맞지 않는다. 군도는 가슴 앞에서 화면 오른쪽 위로 기울어져 올라가며, 몸통 옆에 곧게 세운 배치보다 대각선 배치에 가깝다. 전경 말의 두 귀는 대체로 카메라 쪽을 향한다.",
        "built_space": "나무들이 둘러싼 돌 섞인 흙길이 장교 뒤에서 높은 곳으로 굽어 올라간다. 뒤쪽 오른편에 천으로 덮인 수레 한 대와 바퀴 일부가 보인다. 하단 전경에는 말의 머리 윗부분과 귀 두 개가 있고 그 뒤에 장교가 앉아 있다. 숲의 재질과 낮 조명은 참고에 대체로 부합하며, 건축물이나 불가능한 반사는 없다. 오르막이 시선 반대편에 드러나는 문제는 A와 같다.",
        "entities": "성인 동아시아계 남성 한 명으로, 검은 머리와 갈색 군복은 맞지만 장교 참고의 얼굴보다 이전 숏 안내자의 얼굴과 가르마에 더 가까워 보인다. 인물 참고의 군모와 가슴 훈장이 보이는 영역에서 빠져 있다. 붉은 견장, 가죽 띠와 허리띠는 유지된다. 칼의 곡선 날과 물결무늬, 감긴 손잡이는 소품 참고와 비교적 가깝다. 오른손과 손목이 칼자루를 잡고 있고, 왼손은 말 목 부근에 일부 보인다. 말 한 마리와 덮개 수레가 있으며 기관총은 확인되지 않는다. 공중의 피와 추가 인물은 없지만 소매 패치의 흰 한자 표기가 보인다.",
        "hard_violations": [
         "소매 패치의 한자 표기가 판독 가능한 형태로 노출되어, 이미지 어디에도 읽을 수 있는 문자를 두지 말라는 금지 조건을 위반한다."
        ],
        "physics": "장교는 전경 말의 목 뒤에 자연스럽게 자리하며 기승에 의한 지지가 읽힌다. 안장과 다리는 화면 밖이다. 오른손이 코등이 아래 손잡이를 감싸 칼을 지탱하고, 굽힌 팔로 가슴 앞에 칼을 유지하는 자세는 물리적으로 가능하다. 수레 바퀴는 길 위에 놓여 있다. 떠 있는 신체나 지지 없는 물체는 없으며, 말과 수레의 실제 전진 여부는 이 정지 순간만으로 확정할 수 없다."
       }
      ],
      "all_candidates_fail": true
     },
     "normalized": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 5,
        "verdict_ko": "정면 미디엄 숏, 몸통 옆에 곧게 세운 군도, 장교의 얼굴·군모·훈장은 더 충실하지만, 뒤로 올라가는 산길과 반대인 시선 및 소매의 읽히는 문자 표기가 문제다."
       },
       {
        "label": "A",
        "score": 3,
        "verdict_ko": "기승과 칼을 쥔 손은 명확하지만, 산 위를 향하지 않는 시선에 더해 장교의 얼굴·군모·훈장 연속성이 깨지고 칼도 몸통 앞을 비스듬히 가로지른다."
       }
      ],
      "all_candidates_fail": true,
      "readings": [
       {
        "label": "B",
        "direction": "장교는 얼굴을 정면으로 두고 카메라 부근을 응시한다. 보이는 산길은 장교 뒤쪽 화면 상단으로 올라가므로, 시선이 그 오르막을 향한다고 읽히지 않는다. 군도는 화면 왼쪽 몸통 옆에서 거의 수직으로 올라가며 칼끝이 위를 향한다. 뒤 수레의 기관총 총구는 대체로 카메라 쪽을 향하지만, 이 숏에는 기관총의 조준 대상이 지정되지 않았다.",
        "built_space": "인공 건축물은 없고, 나무줄기가 양옆을 둘러싼 돌 섞인 흙길 한 줄기가 뒤쪽 높은 곳으로 이어진다. 장교 뒤 화면 오른쪽에 기관총을 실은 수레 한 대가 일부 보인다. 참고의 숲·노면·주간 광선과 대체로 연결된다. 장교는 허리 부근까지 담겼고 말의 갈색 몸 일부가 아래쪽에 보이며, 중복 수레나 추가 인물은 없다. 다만 화면에 드러난 오르막은 장교가 바라보는 쪽이 아니라 등 뒤에 있다.",
        "entities": "성인 동아시아계 남성 장교 한 명이며, 외관만으로 일본 국적을 확인할 수는 없다. 얼굴은 인물 참고에 비교적 가깝고, 군모 아래 검은 머리, 갈색 계열 군복, 붉은 견장, 가슴 훈장, 가죽 대각선 띠와 허리띠가 유지된다. 칼 한 자루는 금속 날, 원형 코등이, 어두운 손잡이를 갖춰 소품 참고와 대체로 맞지만 날의 곡률과 표면 무늬는 덜 일치한다. 손잡이를 쥔 맨손과 군복 소매가 프레임 하단에 보인다. 민간인 안내자와 공중의 피는 없다. 소매 패치에는 흰 한자 표기가 식별된다.",
        "hard_violations": [
         "소매 패치의 한자 표기가 판독 가능한 형태로 노출되어, 이미지 어디에도 읽을 수 있는 문자를 두지 말라는 금지 조건을 위반한다."
        ],
        "physics": "오른손 손가락이 칼자루를 감싸고 있어 세워 든 칼의 지지가 명확하다. 팔꿈치를 굽힌 자세로 칼을 유지하는 것은 가능하다. 하단의 말 몸과 기승 자세가 장교의 지지를 설명하며, 안장 접촉부와 발이 잘렸다는 이유만으로 공중에 떠 있다고 볼 근거는 없다. 수레는 노면에 놓이고 기관총은 수레 위 거치대에 지지된다. 피를 턴 뒤 멈춘 순간은 가능하지만 수레가 다시 움직이는지는 정지 화면으로 확인되지 않는다."
       },
       {
        "label": "A",
        "direction": "장교의 눈은 카메라 부근 정면을 향한다. 산길은 그의 뒤에서 화면 상단 왼쪽으로 올라가므로, 보이는 산 위를 쏘아보는 방향과 맞지 않는다. 군도는 가슴 앞에서 화면 오른쪽 위로 기울어져 올라가며, 몸통 옆에 곧게 세운 배치보다 대각선 배치에 가깝다. 전경 말의 두 귀는 대체로 카메라 쪽을 향한다.",
        "built_space": "나무들이 둘러싼 돌 섞인 흙길이 장교 뒤에서 높은 곳으로 굽어 올라간다. 뒤쪽 오른편에 천으로 덮인 수레 한 대와 바퀴 일부가 보인다. 하단 전경에는 말의 머리 윗부분과 귀 두 개가 있고 그 뒤에 장교가 앉아 있다. 숲의 재질과 낮 조명은 참고에 대체로 부합하며, 건축물이나 불가능한 반사는 없다. 오르막이 시선 반대편에 드러나는 문제는 A와 같다.",
        "entities": "성인 동아시아계 남성 한 명으로, 검은 머리와 갈색 군복은 맞지만 장교 참고의 얼굴보다 이전 숏 안내자의 얼굴과 가르마에 더 가까워 보인다. 인물 참고의 군모와 가슴 훈장이 보이는 영역에서 빠져 있다. 붉은 견장, 가죽 띠와 허리띠는 유지된다. 칼의 곡선 날과 물결무늬, 감긴 손잡이는 소품 참고와 비교적 가깝다. 오른손과 손목이 칼자루를 잡고 있고, 왼손은 말 목 부근에 일부 보인다. 말 한 마리와 덮개 수레가 있으며 기관총은 확인되지 않는다. 공중의 피와 추가 인물은 없지만 소매 패치의 흰 한자 표기가 보인다.",
        "hard_violations": [
         "소매 패치의 한자 표기가 판독 가능한 형태로 노출되어, 이미지 어디에도 읽을 수 있는 문자를 두지 말라는 금지 조건을 위반한다."
        ],
        "physics": "장교는 전경 말의 목 뒤에 자연스럽게 자리하며 기승에 의한 지지가 읽힌다. 안장과 다리는 화면 밖이다. 오른손이 코등이 아래 손잡이를 감싸 칼을 지탱하고, 굽힌 팔로 가슴 앞에 칼을 유지하는 자세는 물리적으로 가능하다. 수레 바퀴는 길 위에 놓여 있다. 떠 있는 신체나 지지 없는 물체는 없으며, 말과 수레의 실제 전진 여부는 이 정지 순간만으로 확정할 수 없다."
       }
      ]
     },
     "ok": true
    }
   ],
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "slot_winner_match": true,
   "slot_winner": {
    "gemini-pro": "B",
    "gpt-high": "B"
   },
   "route": "cross_slot_agree"
  },
  "dual": {
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "normalized": {
    "A": 0.8,
    "B": 2.0
   },
   "adjusted": {
    "A": 0.55,
    "B": 1.75
   },
   "violations": {
    "A": [
     "[gemini-pro] 이전 샷에 등장한 다른 인물(민간인)의 얼굴을 장교에게 덧씌움",
     "[gemini-pro] 캐릭터 레퍼런스에 있는 필수 복장(군모, 훈장) 누락",
     "[gpt-high] 소매 패치의 한자 표기가 판독 가능한 형태로 노출되어, 이미지 어디에도 읽을 수 있는 문자를 두지 말라는 금지 조건을 위반한다."
    ],
    "B": [
     "[gpt-high] 소매 패치의 한자 표기가 판독 가능한 형태로 노출되어, 이미지 어디에도 읽을 수 있는 문자를 두지 말라는 금지 조건을 위반한다."
    ]
   },
   "per_model_winner": {
    "gemini-pro": "B",
    "gpt-high": "B"
   },
   "agreed": true
  },
  "totals": {
   "B": 1750,
   "A": 550
  },
  "selected": "B",
  "ranking": [
   "B",
   "A"
  ],
  "verdicts": [
   {
    "label": "B",
    "score": 1750,
    "verdict_ko": "캐릭터 레퍼런스의 헌병 장교 얼굴과 복장(군모, 훈장 포함)을 완벽하게 재현했으며, 군도를 세워 든 채 정면을 쏘아보는 샷의 요구사항을 정확히 구현했습니다.  ★위반: [gpt-high] 소매 패치의 한자 표기가 판독 가능한 형태로 노출되어, 이미지 어디에도 읽을 수 있는 문자를 두지 말라는 금지 조건을 위반한다."
   },
   {
    "label": "A",
    "score": 550,
    "verdict_ko": "프롬프트에서 엄격히 금지한 이전 샷의 민간인 얼굴을 헌병 장교에게 잘못 적용하였고, 필수적인 군모와 훈장도 누락하여 지시사항을 크게 위반했습니다.  ★위반: [gemini-pro] 이전 샷에 등장한 다른 인물(민간인)의 얼굴을 장교에게 덧씌움 / [gemini-pro] 캐릭터 레퍼런스에 있는 필수 복장(군모, 훈장) 누락 / [gpt-high] 소매 패치의 한자 표기가 판독 가능한 형태로 노출되어, 이미지 어디에도 읽을 수 있는 문자를 두지 말라는 금지 조건을 위반한다."
   }
  ],
  "refs": [
   {
    "label": "PREVIOUS SHOT STILL — a visually related earlier shot of this same place: the location's look, materials, fixed features and lighting mood are LOCKED to this photo; never copy its camera framing. The clothing, hair and overall look of 헌병 장교 — who appear both in that photo and in this shot — are LOCKED to that photo. Anyone else visible in that photo is NOT in this shot: never carry their face, body or clothing onto anyone here. If this photo shows a character who cannot move (dead or unconscious), that character's exact pose, position and orientation are ALSO LOCKED — treat their body as a fixed prop of the set that only the camera moves around.",
    "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/db00517b-165b-4005-a287-f28936b59b95/images/272bd7af-9294-4490-98c6-88a40e9deacc/scene/recipe/S65sh14_sel.png",
    "asset_id": "74897c29-1460-4d71-b285-63195dff19de",
    "role": "prev_still"
   },
   {
    "label": "CHARACTER REFERENCE — 헌병 장교: the exact person appearing in this shot; match face, hair and build exactly.",
    "path": "<bytes:862147>",
    "asset_id": "83cd18a8-56b7-4e8a-8937-e155fe16c041",
    "role": "character_ref"
   },
   {
    "label": "PROP REFERENCE — 칼: the exact object appearing in this shot; match its look, material and wear exactly.",
    "path": "<bytes:780122>",
    "asset_id": "82614014-6091-423c-89ca-1e15d209de72",
    "role": "prop_ref"
   }
  ],
  "critique_skipped": true,
  "shot_run_produced": false,
  "shot_run_uid": "06aa1446-7471-7d0c-83cf-729d9dc40994",
  "ref_mode": "prev+엔티티",
  "share_plan": {
   "ref_plan": "prev",
   "prev_anchor_tag": "S65sh14"
  }
 },
 "S65sh19::cine": {
  "applied": true,
  "attempted_at": "2026-09-09T08:38:26.293929+00:00",
  "fingerprint": "8cc6ada818a062802fe312ead658ca47912f3283e8401553857c4bb1db530185",
  "fingerprint_version": 2,
  "provider": "mai",
  "endpoint": "openrouter/chat-completions",
  "model": "microsoft/mai-image-2.6",
  "pack": "24.202608252115",
  "source_file": "S65sh19_sel.png",
  "source_sha256": "35832238ae04cf0215f8c791f18a86fc5cccd6b6332761fcec55e65e69911339",
  "file": "S65sh19_cine.png",
  "staged_sha256": "3e8b3d10c326acb7c74193e5365a2ca9d704af8ab29e451f8b22069dfbbddf77",
  "latency_ms": 29148
 },
 "S66sh1::signage": {
  "fp": "990ad0b625e08efe",
  "inscriptions": [],
  "cues": [],
  "dropped": []
 },
 "S66sh1": {
  "input_fingerprint": "bd3ff567f2f569e4",
  "prompt": "Create ONE FINAL photorealistic live-action film still of the moment below — early 20th-century Korea, Jeolla region; all people are Korean unless explicitly stated otherwise. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 거대한 일본군 병력이 보이는 산 아래쪽을 향해 의연하게 시선을 고정한 안규홍의 피투성이 정면.\n\nLOCATION (lock): The downhill-facing edge of the ruined mountain camp, with a view toward the approaching force below the trees. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: close-up\n- KEY BACKGROUND ELEMENTS: 산채의 나무 (Present around the militia's position); used as Remain softly separated behind the face, retaining location without competing with his eyeline.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Daylight, muted saturation, and naturalistic contrast preserve the bloodied face's texture while keeping his resolve readable.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The clothing, hair and overall look of 안규홍 — who appear both in that photo and in this shot — are LOCKED to that photo. Anyone else visible in that photo is NOT in this shot: never carry their face, body or clothing onto anyone here. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): Daylight remains over the devastated camp, with burned and collapsed huts, broken cookware and food scattered through blood and dirt. 안규홍: He stands blood- and dirt-covered with the bloody sickle, his back wound still bleeding and his neck bare.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 안규홍 (한국인 남성, 청년 나이대, 검은 머리, 단단히 묶은 머리) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.",
  "roll_prompts": {
   "A": "Create ONE FINAL photorealistic live-action film still of the moment below — early 20th-century Korea, Jeolla region; all people are Korean unless explicitly stated otherwise. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 거대한 일본군 병력이 보이는 산 아래쪽을 향해 의연하게 시선을 고정한 안규홍의 피투성이 정면.\n\nLOCATION (lock): The downhill-facing edge of the ruined mountain camp, with a view toward the approaching force below the trees. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: close-up\n- KEY BACKGROUND ELEMENTS: 산채의 나무 (Present around the militia's position); used as Remain softly separated behind the face, retaining location without competing with his eyeline.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Daylight, muted saturation, and naturalistic contrast preserve the bloodied face's texture while keeping his resolve readable.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The clothing, hair and overall look of 안규홍 — who appear both in that photo and in this shot — are LOCKED to that photo. Anyone else visible in that photo is NOT in this shot: never carry their face, body or clothing onto anyone here. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): Daylight remains over the devastated camp, with burned and collapsed huts, broken cookware and food scattered through blood and dirt. 안규홍: He stands blood- and dirt-covered with the bloody sickle, his back wound still bleeding and his neck bare.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 안규홍 (한국인 남성, 청년 나이대, 검은 머리, 단단히 묶은 머리) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.",
   "B": "Create ONE FINAL photorealistic live-action film still of the moment below — early 20th-century Korea, Jeolla region; all people are Korean unless explicitly stated otherwise. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 거대한 일본군 병력이 보이는 산 아래쪽을 향해 의연하게 시선을 고정한 안규홍의 피투성이 정면.\n\nLOCATION (lock): The downhill-facing edge of the ruined mountain camp, with a view toward the approaching force below the trees. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: close-up\n- KEY BACKGROUND ELEMENTS: 산채의 나무 (Present around the militia's position); used as Remain softly separated behind the face, retaining location without competing with his eyeline.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Daylight, muted saturation, and naturalistic contrast preserve the bloodied face's texture while keeping his resolve readable.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The clothing, hair and overall look of 안규홍 — who appear both in that photo and in this shot — are LOCKED to that photo. Anyone else visible in that photo is NOT in this shot: never carry their face, body or clothing onto anyone here. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): Daylight remains over the devastated camp, with burned and collapsed huts, broken cookware and food scattered through blood and dirt. 안규홍: He stands blood- and dirt-covered with the bloody sickle, his back wound still bleeding and his neck bare.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 안규홍 (한국인 남성, 청년 나이대, 검은 머리, 단단히 묶은 머리) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.\n\nCOMPOSITION VARIATION (this candidate only): keep the place, the people and the moment exactly as specified, and keep from the CAMERA & FRAME contract its subject scale, screen placement and key background elements exactly. For this candidate only, its camera angle and height are deliberately left open: choose a camera position distinctly different from the obvious one, as a film director picking a second setup on the same blocking. Stay true to any explicit framing the SHOT TEXT itself dictates."
  },
  "shot_run_spend_attempt_count": 1,
  "readings": [
   {
    "label": "A",
    "direction": "안규홍은 정면(카메라 방향)을 응시하고 있으나, 텍스트 상 그가 바라보아야 할 일본군 병력은 그의 등 뒤 산비탈에 위치해 있어 시선과 타겟의 위치가 모순됨.",
    "built_space": "오른쪽의 부서진 움막은 남아있으나, 레퍼런스에 있던 왼쪽 움막과 울타리가 완전히 사라지고 일본군이 있는 산비탈로 임의 변경됨.",
    "entities": "안규홍(피투성이 얼굴, 피 묻은 낫을 든 모습), 배경에 다수의 일본군.",
    "hard_violations": [
     "[gemini-pro] 프레이밍 위반 (클로즈업 지시를 무시하고 허리와 손에 든 낫이 보이는 미디엄 샷으로 연출됨)",
     "[gemini-pro] 로케이션 훼손 (레퍼런스에 명시된 왼쪽 배경 구조물을 임의로 산비탈로 변경함)",
     "[gpt-high] 안규홍이 바라보는 목표여야 할 접근 일본군을 그의 등 뒤에 배치하여, 인물과 목표의 필수 공간 관계를 뒤집었다."
    ],
    "physics": "두 발로 땅에 서 있으며, 오른손으로 낫의 자루를 정상적으로 쥐고 있음."
   },
   {
    "label": "B",
    "direction": "안규홍은 카메라(산 아래쪽 방향)를 향해 결연하게 시선을 고정하고 있음.",
    "built_space": "레퍼런스와 동일하게 좌우의 부서진 움막과 울타리, 바닥에 흩어진 가마솥 및 파편들이 올바른 위치와 비율로 정확히 유지됨.",
    "entities": "안규홍(정면을 응시하는 피투성이 얼굴, 묶은 검은 머리). 낫은 클로즈업 프레이밍에 맞게 자연스럽게 프레임 밖으로 제외됨.",
    "hard_violations": [],
    "physics": "바닥에 두 발을 딛고 안정적인 자세로 서 있음."
   }
  ],
  "cross_model_order": {
   "policy": "cross_model_order_v2_slot0_fwd_slot1_rev_parallel2",
   "slots": [
    {
     "model": "gemini-pro",
     "order": "forward",
     "display_to_canonical": {
      "A": "A",
      "B": "B"
     },
     "raw": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 10,
        "verdict_ko": "요구된 클로즈업 프레이밍을 정확히 지켰으며, 레퍼런스의 로케이션(배경 구조물)을 온전히 유지한 채 산 아래쪽을 향해 정면을 응시하는 안규홍의 결연한 표정을 완벽하게 구현했습니다."
       },
       {
        "label": "A",
        "score": 3,
        "verdict_ko": "클로즈업 지시를 어기고 프레이밍을 넓혔으며, 시선 방향과 모순되게 등 뒤에 일본군을 배치하기 위해 레퍼런스의 왼쪽 배경 구조물을 완전히 훼손했습니다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "안규홍은 정면(카메라 방향)을 응시하고 있으나, 텍스트 상 그가 바라보아야 할 일본군 병력은 그의 등 뒤 산비탈에 위치해 있어 시선과 타겟의 위치가 모순됨.",
        "built_space": "오른쪽의 부서진 움막은 남아있으나, 레퍼런스에 있던 왼쪽 움막과 울타리가 완전히 사라지고 일본군이 있는 산비탈로 임의 변경됨.",
        "entities": "안규홍(피투성이 얼굴, 피 묻은 낫을 든 모습), 배경에 다수의 일본군.",
        "hard_violations": [
         "프레이밍 위반 (클로즈업 지시를 무시하고 허리와 손에 든 낫이 보이는 미디엄 샷으로 연출됨)",
         "로케이션 훼손 (레퍼런스에 명시된 왼쪽 배경 구조물을 임의로 산비탈로 변경함)"
        ],
        "physics": "두 발로 땅에 서 있으며, 오른손으로 낫의 자루를 정상적으로 쥐고 있음."
       },
       {
        "label": "B",
        "direction": "안규홍은 카메라(산 아래쪽 방향)를 향해 결연하게 시선을 고정하고 있음.",
        "built_space": "레퍼런스와 동일하게 좌우의 부서진 움막과 울타리, 바닥에 흩어진 가마솥 및 파편들이 올바른 위치와 비율로 정확히 유지됨.",
        "entities": "안규홍(정면을 응시하는 피투성이 얼굴, 묶은 검은 머리). 낫은 클로즈업 프레이밍에 맞게 자연스럽게 프레임 밖으로 제외됨.",
        "hard_violations": [],
        "physics": "바닥에 두 발을 딛고 안정적인 자세로 서 있음."
       }
      ],
      "all_candidates_fail": false
     },
     "normalized": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 10,
        "verdict_ko": "요구된 클로즈업 프레이밍을 정확히 지켰으며, 레퍼런스의 로케이션(배경 구조물)을 온전히 유지한 채 산 아래쪽을 향해 정면을 응시하는 안규홍의 결연한 표정을 완벽하게 구현했습니다."
       },
       {
        "label": "A",
        "score": 3,
        "verdict_ko": "클로즈업 지시를 어기고 프레이밍을 넓혔으며, 시선 방향과 모순되게 등 뒤에 일본군을 배치하기 위해 레퍼런스의 왼쪽 배경 구조물을 완전히 훼손했습니다."
       }
      ],
      "all_candidates_fail": false,
      "readings": [
       {
        "label": "A",
        "direction": "안규홍은 정면(카메라 방향)을 응시하고 있으나, 텍스트 상 그가 바라보아야 할 일본군 병력은 그의 등 뒤 산비탈에 위치해 있어 시선과 타겟의 위치가 모순됨.",
        "built_space": "오른쪽의 부서진 움막은 남아있으나, 레퍼런스에 있던 왼쪽 움막과 울타리가 완전히 사라지고 일본군이 있는 산비탈로 임의 변경됨.",
        "entities": "안규홍(피투성이 얼굴, 피 묻은 낫을 든 모습), 배경에 다수의 일본군.",
        "hard_violations": [
         "프레이밍 위반 (클로즈업 지시를 무시하고 허리와 손에 든 낫이 보이는 미디엄 샷으로 연출됨)",
         "로케이션 훼손 (레퍼런스에 명시된 왼쪽 배경 구조물을 임의로 산비탈로 변경함)"
        ],
        "physics": "두 발로 땅에 서 있으며, 오른손으로 낫의 자루를 정상적으로 쥐고 있음."
       },
       {
        "label": "B",
        "direction": "안규홍은 카메라(산 아래쪽 방향)를 향해 결연하게 시선을 고정하고 있음.",
        "built_space": "레퍼런스와 동일하게 좌우의 부서진 움막과 울타리, 바닥에 흩어진 가마솥 및 파편들이 올바른 위치와 비율로 정확히 유지됨.",
        "entities": "안규홍(정면을 응시하는 피투성이 얼굴, 묶은 검은 머리). 낫은 클로즈업 프레이밍에 맞게 자연스럽게 프레임 밖으로 제외됨.",
        "hard_violations": [],
        "physics": "바닥에 두 발을 딛고 안정적인 자세로 서 있음."
       }
      ]
     },
     "ok": true
    },
    {
     "model": "gpt-high",
     "order": "reverse",
     "display_to_canonical": {
      "A": "B",
      "B": "A"
     },
     "raw": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 8,
        "verdict_ko": "피투성이 얼굴의 정면 클로즈업과 기존 산채·복장을 잘 유지하지만, 시선이 렌즈에 가까우며 배경 초막이 나무보다 두드러진다."
       },
       {
        "label": "B",
        "score": 3,
        "verdict_ko": "일본군을 안규홍의 시선 반대편 등 뒤에 배치했고, 낫과 상체까지 보여 주도록 넓혀 정면 클로즈업 지시도 놓쳤다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "얼굴은 거의 정확한 정면이고 두 눈은 카메라 쪽 전방의 약간 낮은 지점을 응시한다. 결연한 표정은 보이지만 시선이 렌즈에 매우 가까워 직접 응시처럼도 읽힌다. 목표인 산 아래 일본군은 화면 밖이므로 실제 목표와의 정렬은 확인할 수 없다. 무기나 이동 중인 인물은 보이지 않는다.",
        "built_space": "얼굴과 어깨가 화면을 크게 차지한다. 뒤에는 왼쪽과 오른쪽 초막 각 한 채, 왼쪽의 탄 수직 기둥 하나와 무너진 목재, 흙바닥의 솥과 파편이 보인다. 왼쪽 바닥에 큰 솥 두 개가 뚜렷하고 그 뒤 작은 용기 하나, 오른쪽에도 어두운 용기 하나가 보인다. 이전 장면의 초가·목재·잔해 배치는 대체로 이어진다. 숲은 상단 배경에 있지만 초막과 바닥 잔해가 요구된 부드러운 나무 배경보다 눈에 띈다. 반사면은 없다.",
        "entities": "안규홍 한 명만 분명하게 보인다. 젊은 동아시아계 남성의 얼굴, 검은 눈썹, 뒤로 단단히 모은 검은 머리와 거친 회백색 옷이 참조와 대체로 일치한다. 이마·양 볼·턱·맨목에 피와 흙이 묻어 있고 눈은 정상적인 홍채와 동공을 갖는다. 머리 묶음 끝은 상단 밖이다. 낫, 등 상처, 하체는 올바른 클로즈업 범위 밖이므로 확인할 수 없다. 읽을 수 있는 글자는 없다.",
        "hard_violations": [],
        "physics": "머리는 목과 어깨에 자연스럽게 연결되고 옷은 어깨 위에서 아래로 드리워진다. 발과 지면 접점은 프레임 밖이지만 떠 있는 자세나 비정상적인 신체 연결은 보이지 않는다. 솥과 목재 잔해는 흙바닥에 놓여 있다. 피는 피부 굴곡을 따라 번지고 목 아래로 흐른 흔적으로 표현되어 있다."
       },
       {
        "label": "B",
        "direction": "안규홍은 몸을 약간 틀고 카메라 쪽 전방을 응시한다. 그러나 대규모 군인 행렬은 그의 왼쪽 뒤편 경사에 있어, 눈길이 실제로 보이는 군대에 닿지 않는다. 군인들은 전경의 산채 쪽으로 올라오는 배치이며 소총과 총검은 대체로 위쪽 대각선을 향한다. 안규홍의 낫은 화면 왼쪽 아래에 들려 있고 날끝은 아래로 향한다.",
        "built_space": "머리 전체부터 허리 부근과 낫을 든 손까지 포함한 넓은 상반신 구도여서 클로즈업이 아니다. 오른쪽 초막 한 채와 중앙 뒤의 무너진 지붕 부분, 왼쪽 탄 기둥 하나, 바닥의 솥·그릇·목재 잔해가 보인다. 왼쪽 뒤로 군인들이 가득한 경사 통로가 열려 있다. 기존 산채의 재료감은 비슷하지만, 안규홍이 바라봐야 할 접근 병력을 그의 뒤쪽 공간에 놓았다. 숲과 군대가 얼굴에서 시선을 분산시킨다. 반사면은 없다.",
        "entities": "안규홍은 젊은 동아시아계 남성으로 보이며 묶은 검은 머리, 얼굴 윤곽, 더러운 회백색 옷이 참조와 대체로 맞는다. 얼굴에는 피와 흙이 있고 목은 장신구 없이 드러나지만 A보다 피투성이의 강도는 약하다. 오른손에는 피 묻은 굽은 금속 낫과 나무 손잡이가 보인다. 뒤에는 군복·군모와 소총을 갖춘 다수의 군인이 있으며 일본군 역할로 읽히지만 개별 국적이나 얼굴은 식별하기 어렵다. 등 상처는 보이지 않으며 읽을 수 있는 글자는 없다.",
        "hard_violations": [
         "안규홍이 바라보는 목표여야 할 접근 일본군을 그의 등 뒤에 배치하여, 인물과 목표의 필수 공간 관계를 뒤집었다."
        ],
        "physics": "낫의 나무 손잡이를 오른손으로 실제로 쥐고 있어 도구의 지지가 명확하다. 상체의 전방 기울임과 팔의 굽힘은 가능한 자세이며 하체는 화면 밖이다. 가까운 군인들은 경사 지면에 발을 딛고 올라오는 모습이고, 뒤쪽 군인들의 접점은 앞사람에게 가려진다. 솥과 부서진 목재는 지면에 놓여 있으며 근거 없이 떠 있는 물체는 보이지 않는다."
       }
      ],
      "all_candidates_fail": false
     },
     "normalized": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 8,
        "verdict_ko": "피투성이 얼굴의 정면 클로즈업과 기존 산채·복장을 잘 유지하지만, 시선이 렌즈에 가까우며 배경 초막이 나무보다 두드러진다."
       },
       {
        "label": "A",
        "score": 3,
        "verdict_ko": "일본군을 안규홍의 시선 반대편 등 뒤에 배치했고, 낫과 상체까지 보여 주도록 넓혀 정면 클로즈업 지시도 놓쳤다."
       }
      ],
      "all_candidates_fail": false,
      "readings": [
       {
        "label": "B",
        "direction": "얼굴은 거의 정확한 정면이고 두 눈은 카메라 쪽 전방의 약간 낮은 지점을 응시한다. 결연한 표정은 보이지만 시선이 렌즈에 매우 가까워 직접 응시처럼도 읽힌다. 목표인 산 아래 일본군은 화면 밖이므로 실제 목표와의 정렬은 확인할 수 없다. 무기나 이동 중인 인물은 보이지 않는다.",
        "built_space": "얼굴과 어깨가 화면을 크게 차지한다. 뒤에는 왼쪽과 오른쪽 초막 각 한 채, 왼쪽의 탄 수직 기둥 하나와 무너진 목재, 흙바닥의 솥과 파편이 보인다. 왼쪽 바닥에 큰 솥 두 개가 뚜렷하고 그 뒤 작은 용기 하나, 오른쪽에도 어두운 용기 하나가 보인다. 이전 장면의 초가·목재·잔해 배치는 대체로 이어진다. 숲은 상단 배경에 있지만 초막과 바닥 잔해가 요구된 부드러운 나무 배경보다 눈에 띈다. 반사면은 없다.",
        "entities": "안규홍 한 명만 분명하게 보인다. 젊은 동아시아계 남성의 얼굴, 검은 눈썹, 뒤로 단단히 모은 검은 머리와 거친 회백색 옷이 참조와 대체로 일치한다. 이마·양 볼·턱·맨목에 피와 흙이 묻어 있고 눈은 정상적인 홍채와 동공을 갖는다. 머리 묶음 끝은 상단 밖이다. 낫, 등 상처, 하체는 올바른 클로즈업 범위 밖이므로 확인할 수 없다. 읽을 수 있는 글자는 없다.",
        "hard_violations": [],
        "physics": "머리는 목과 어깨에 자연스럽게 연결되고 옷은 어깨 위에서 아래로 드리워진다. 발과 지면 접점은 프레임 밖이지만 떠 있는 자세나 비정상적인 신체 연결은 보이지 않는다. 솥과 목재 잔해는 흙바닥에 놓여 있다. 피는 피부 굴곡을 따라 번지고 목 아래로 흐른 흔적으로 표현되어 있다."
       },
       {
        "label": "A",
        "direction": "안규홍은 몸을 약간 틀고 카메라 쪽 전방을 응시한다. 그러나 대규모 군인 행렬은 그의 왼쪽 뒤편 경사에 있어, 눈길이 실제로 보이는 군대에 닿지 않는다. 군인들은 전경의 산채 쪽으로 올라오는 배치이며 소총과 총검은 대체로 위쪽 대각선을 향한다. 안규홍의 낫은 화면 왼쪽 아래에 들려 있고 날끝은 아래로 향한다.",
        "built_space": "머리 전체부터 허리 부근과 낫을 든 손까지 포함한 넓은 상반신 구도여서 클로즈업이 아니다. 오른쪽 초막 한 채와 중앙 뒤의 무너진 지붕 부분, 왼쪽 탄 기둥 하나, 바닥의 솥·그릇·목재 잔해가 보인다. 왼쪽 뒤로 군인들이 가득한 경사 통로가 열려 있다. 기존 산채의 재료감은 비슷하지만, 안규홍이 바라봐야 할 접근 병력을 그의 뒤쪽 공간에 놓았다. 숲과 군대가 얼굴에서 시선을 분산시킨다. 반사면은 없다.",
        "entities": "안규홍은 젊은 동아시아계 남성으로 보이며 묶은 검은 머리, 얼굴 윤곽, 더러운 회백색 옷이 참조와 대체로 맞는다. 얼굴에는 피와 흙이 있고 목은 장신구 없이 드러나지만 A보다 피투성이의 강도는 약하다. 오른손에는 피 묻은 굽은 금속 낫과 나무 손잡이가 보인다. 뒤에는 군복·군모와 소총을 갖춘 다수의 군인이 있으며 일본군 역할로 읽히지만 개별 국적이나 얼굴은 식별하기 어렵다. 등 상처는 보이지 않으며 읽을 수 있는 글자는 없다.",
        "hard_violations": [
         "안규홍이 바라보는 목표여야 할 접근 일본군을 그의 등 뒤에 배치하여, 인물과 목표의 필수 공간 관계를 뒤집었다."
        ],
        "physics": "낫의 나무 손잡이를 오른손으로 실제로 쥐고 있어 도구의 지지가 명확하다. 상체의 전방 기울임과 팔의 굽힘은 가능한 자세이며 하체는 화면 밖이다. 가까운 군인들은 경사 지면에 발을 딛고 올라오는 모습이고, 뒤쪽 군인들의 접점은 앞사람에게 가려진다. 솥과 부서진 목재는 지면에 놓여 있으며 근거 없이 떠 있는 물체는 보이지 않는다."
       }
      ]
     },
     "ok": true
    }
   ],
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "slot_winner_match": true,
   "slot_winner": {
    "gemini-pro": "B",
    "gpt-high": "B"
   },
   "route": "cross_slot_agree"
  },
  "dual": {
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "normalized": {
    "A": 0.675,
    "B": 2.0
   },
   "adjusted": {
    "A": 0.425,
    "B": 2.0
   },
   "violations": {
    "A": [
     "[gemini-pro] 프레이밍 위반 (클로즈업 지시를 무시하고 허리와 손에 든 낫이 보이는 미디엄 샷으로 연출됨)",
     "[gemini-pro] 로케이션 훼손 (레퍼런스에 명시된 왼쪽 배경 구조물을 임의로 산비탈로 변경함)",
     "[gpt-high] 안규홍이 바라보는 목표여야 할 접근 일본군을 그의 등 뒤에 배치하여, 인물과 목표의 필수 공간 관계를 뒤집었다."
    ]
   },
   "per_model_winner": {
    "gemini-pro": "B",
    "gpt-high": "B"
   },
   "agreed": true
  },
  "totals": {
   "B": 2000,
   "A": 425
  },
  "selected": "B",
  "ranking": [
   "B",
   "A"
  ],
  "verdicts": [
   {
    "label": "B",
    "score": 2000,
    "verdict_ko": "요구된 클로즈업 프레이밍을 정확히 지켰으며, 레퍼런스의 로케이션(배경 구조물)을 온전히 유지한 채 산 아래쪽을 향해 정면을 응시하는 안규홍의 결연한 표정을 완벽하게 구현했습니다."
   },
   {
    "label": "A",
    "score": 425,
    "verdict_ko": "클로즈업 지시를 어기고 프레이밍을 넓혔으며, 시선 방향과 모순되게 등 뒤에 일본군을 배치하기 위해 레퍼런스의 왼쪽 배경 구조물을 완전히 훼손했습니다.  ★위반: [gemini-pro] 프레이밍 위반 (클로즈업 지시를 무시하고 허리와 손에 든 낫이 보이는 미디엄 샷으로 연출됨) / [gemini-pro] 로케이션 훼손 (레퍼런스에 명시된 왼쪽 배경 구조물을 임의로 산비탈로 변경함) / [gpt-high] 안규홍이 바라보는 목표여야 할 접근 일본군을 그의 등 뒤에 배치하여, 인물과 목표의 필수 공간 관계를 뒤집었다."
   }
  ],
  "refs": [
   {
    "label": "PREVIOUS SHOT STILL — a visually related earlier shot of this same place: the location's look, materials, fixed features and lighting mood are LOCKED to this photo; never copy its camera framing. The clothing, hair and overall look of 안규홍 — who appear both in that photo and in this shot — are LOCKED to that photo. Anyone else visible in that photo is NOT in this shot: never carry their face, body or clothing onto anyone here. If this photo shows a character who cannot move (dead or unconscious), that character's exact pose, position and orientation are ALSO LOCKED — treat their body as a fixed prop of the set that only the camera moves around.",
    "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/db00517b-165b-4005-a287-f28936b59b95/images/272bd7af-9294-4490-98c6-88a40e9deacc/scene/recipe/S64sh18_sel.png",
    "asset_id": "c36ecdb9-6d4c-44d9-a9a2-d8efd3d5f0ef",
    "role": "prev_still"
   },
   {
    "label": "CHARACTER REFERENCE — 안규홍: the exact person appearing in this shot; match face, hair and build exactly.",
    "path": "<bytes:1028384>",
    "asset_id": "3b5894db-ce55-4f34-b966-5e725c3aa952",
    "role": "character_ref"
   }
  ],
  "critique_skipped": true,
  "shot_run_produced": false,
  "shot_run_uid": "06aa1446-761d-7b01-a5ce-5856f199b2ef",
  "ref_mode": "prev+엔티티",
  "share_plan": {
   "ref_plan": "prev",
   "prev_anchor_tag": "S64sh18"
  }
 },
 "S66sh1::cine": {
  "applied": true,
  "attempted_at": "2026-09-09T08:39:59.472063+00:00",
  "fingerprint": "6981c4bc834a4a1bfa5635d941200a5c8ea5f15fc4e19386599c68887e4e8410",
  "fingerprint_version": 2,
  "provider": "mai",
  "endpoint": "openrouter/chat-completions",
  "model": "microsoft/mai-image-2.6",
  "pack": "24.202608252115",
  "source_file": "S66sh1_sel.png",
  "source_sha256": "598beca0d805d3332d095d8f214a432d0e5d35978e3a476d7348b9d511ec624d",
  "file": "S66sh1_cine.png",
  "staged_sha256": "e7f3a535d9b173a1aa30bf1243eaf6fd5f497166e8997a98ced3e841bf3d9303",
  "latency_ms": 24811
 },
 "S66sh9::signage": {
  "fp": "5caba7a6bd174b8c",
  "inscriptions": [],
  "cues": [],
  "dropped": []
 },
 "S66sh9": {
  "input_fingerprint": "017fe015efd27fcd",
  "prompt": "Create ONE FINAL photorealistic live-action film still of the moment below — early 20th-century Korea, Jeolla region; all people are Korean unless explicitly stated otherwise. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 서로를 마주 본 채 피 묻은 입가에 옅은 미소를 머금은 조선인 의병들의 흙투성이 얼굴들.\n\nLOCATION (lock): The survivors' outdoor gathering place among the trees and wreckage of the mountain camp. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: medium shot\n- KEY BACKGROUND ELEMENTS: 의병들 사이의 나무 (Trees remain around the surviving militia); used as Create separated background intervals behind the staggered faces rather than a uniform group-portrait backdrop.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Softly balanced daylight and muted earth tones allow brief human warmth in the faint smiles without diminishing the scene's grave tension.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established — the place, its fixed features and wear, each person's clothing and state — persists. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The ruined camp remains in daylight, with scattered bowls, food and battle debris.\n\nPEOPLE: the SHOT TEXT alone decides who is visible in this shot. People known to appear somewhere in this scene: 안규홍 (한국인 남성, 청년 나이대, 검은 머리, 단단히 묶은 머리). That list is scene-level, not a cast list for this frame — it may name someone this shot does not show, and it may omit someone this shot does show. If the shot text names a person who is not on the list, draw that person exactly as the shot text describes them; the list does not override the shot text. Never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.",
  "roll_prompts": {
   "A": "Create ONE FINAL photorealistic live-action film still of the moment below — early 20th-century Korea, Jeolla region; all people are Korean unless explicitly stated otherwise. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 서로를 마주 본 채 피 묻은 입가에 옅은 미소를 머금은 조선인 의병들의 흙투성이 얼굴들.\n\nLOCATION (lock): The survivors' outdoor gathering place among the trees and wreckage of the mountain camp. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: medium shot\n- KEY BACKGROUND ELEMENTS: 의병들 사이의 나무 (Trees remain around the surviving militia); used as Create separated background intervals behind the staggered faces rather than a uniform group-portrait backdrop.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Softly balanced daylight and muted earth tones allow brief human warmth in the faint smiles without diminishing the scene's grave tension.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established — the place, its fixed features and wear, each person's clothing and state — persists. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The ruined camp remains in daylight, with scattered bowls, food and battle debris.\n\nPEOPLE: the SHOT TEXT alone decides who is visible in this shot. People known to appear somewhere in this scene: 안규홍 (한국인 남성, 청년 나이대, 검은 머리, 단단히 묶은 머리). That list is scene-level, not a cast list for this frame — it may name someone this shot does not show, and it may omit someone this shot does show. If the shot text names a person who is not on the list, draw that person exactly as the shot text describes them; the list does not override the shot text. Never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.",
   "B": "Create ONE FINAL photorealistic live-action film still of the moment below — early 20th-century Korea, Jeolla region; all people are Korean unless explicitly stated otherwise. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 서로를 마주 본 채 피 묻은 입가에 옅은 미소를 머금은 조선인 의병들의 흙투성이 얼굴들.\n\nLOCATION (lock): The survivors' outdoor gathering place among the trees and wreckage of the mountain camp. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: medium shot\n- KEY BACKGROUND ELEMENTS: 의병들 사이의 나무 (Trees remain around the surviving militia); used as Create separated background intervals behind the staggered faces rather than a uniform group-portrait backdrop.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Softly balanced daylight and muted earth tones allow brief human warmth in the faint smiles without diminishing the scene's grave tension.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established — the place, its fixed features and wear, each person's clothing and state — persists. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The ruined camp remains in daylight, with scattered bowls, food and battle debris.\n\nPEOPLE: the SHOT TEXT alone decides who is visible in this shot. People known to appear somewhere in this scene: 안규홍 (한국인 남성, 청년 나이대, 검은 머리, 단단히 묶은 머리). That list is scene-level, not a cast list for this frame — it may name someone this shot does not show, and it may omit someone this shot does show. If the shot text names a person who is not on the list, draw that person exactly as the shot text describes them; the list does not override the shot text. Never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.\n\nCOMPOSITION VARIATION (this candidate only): keep the place, the people and the moment exactly as specified, and keep from the CAMERA & FRAME contract its subject scale, screen placement and key background elements exactly. For this candidate only, its camera angle and height are deliberately left open: choose a camera position distinctly different from the obvious one, as a film director picking a second setup on the same blocking. Stay true to any explicit framing the SHOT TEXT itself dictates."
  },
  "shot_run_spend_attempt_count": 1,
  "readings": [
   {
    "label": "A",
    "direction": "인물들이 서로를 향해 시선을 두며 마주보고 있음.",
    "built_space": "지시된 산채의 잔해와 나무가 배경 및 전경에 배치되어 공간을 형성함.",
    "entities": "피 묻은 입가와 흙투성이 얼굴을 한 의병들이며 옅은 미소를 띠고 있음. 안규홍 레퍼런스와 일치하는 인물이 포함되어 있으나, 여러 인물의 얼굴이 동일하게 복제된 형태임.",
    "hard_violations": [
     "[gemini-pro] 신체와 제대로 연결되지 않고 소총을 쥐고 있는 중앙 하단의 불가능한 구조의 손",
     "[gemini-pro] 오른쪽 나무 기둥을 관통하여 융합되어 버린 소총의 총신"
    ],
    "physics": "하단 중앙의 소총을 쥐고 있는 손은 어떤 인물의 팔과도 자연스럽게 연결되지 않아 물리적으로 불가능한 위치에 있음. 우측 나무 뒤 인물이 든 소총은 나무 기둥과 물리적으로 겹쳐짐."
   },
   {
    "label": "B",
    "direction": "네 명의 인물이 양옆으로 나뉘어 서로를 마주보며 시선을 교환하고 있음.",
    "built_space": "나무들이 인물들 사이의 간격을 나누며 전경에 위치하고, 배경에는 산채 잔해가 위치해 프레이밍 지시를 정확히 구현함.",
    "entities": "흙투성이 얼굴에 미소를 머금은 의병들. 두 번째 인물이 안규홍 레퍼런스의 얼굴 및 헤어스타일과 정확히 일치함. (맨 오른쪽 인물은 입가에 피가 묻어있지 않은 아쉬움이 있음)",
    "hard_violations": [],
    "physics": "인물들의 자세와 시선이 자연스러우며, 허공에 떠 있거나 물리적 법칙에 어긋나는 객체 및 신체가 없음."
   }
  ],
  "cross_model_order": {
   "policy": "cross_model_order_v2_slot0_fwd_slot1_rev_parallel2",
   "slots": [
    {
     "model": "gemini-pro",
     "order": "forward",
     "display_to_canonical": {
      "A": "A",
      "B": "B"
     },
     "raw": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 8,
        "verdict_ko": "나무를 활용해 인물들의 얼굴을 공간적으로 분리한 미디엄 샷의 구도를 훌륭하게 구현하였으며, 치명적인 물리적 오류 없이 지시사항을 잘 반영함."
       },
       {
        "label": "A",
        "score": 2,
        "verdict_ko": "소총을 쥐고 있는 신체와 연결되지 않은 기형적인 손과 나무를 관통한 무기 등 심각한 물리적 오류(Hard Violation)가 발생하여 실격됨."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "인물들이 서로를 향해 시선을 두며 마주보고 있음.",
        "built_space": "지시된 산채의 잔해와 나무가 배경 및 전경에 배치되어 공간을 형성함.",
        "entities": "피 묻은 입가와 흙투성이 얼굴을 한 의병들이며 옅은 미소를 띠고 있음. 안규홍 레퍼런스와 일치하는 인물이 포함되어 있으나, 여러 인물의 얼굴이 동일하게 복제된 형태임.",
        "hard_violations": [
         "신체와 제대로 연결되지 않고 소총을 쥐고 있는 중앙 하단의 불가능한 구조의 손",
         "오른쪽 나무 기둥을 관통하여 융합되어 버린 소총의 총신"
        ],
        "physics": "하단 중앙의 소총을 쥐고 있는 손은 어떤 인물의 팔과도 자연스럽게 연결되지 않아 물리적으로 불가능한 위치에 있음. 우측 나무 뒤 인물이 든 소총은 나무 기둥과 물리적으로 겹쳐짐."
       },
       {
        "label": "B",
        "direction": "네 명의 인물이 양옆으로 나뉘어 서로를 마주보며 시선을 교환하고 있음.",
        "built_space": "나무들이 인물들 사이의 간격을 나누며 전경에 위치하고, 배경에는 산채 잔해가 위치해 프레이밍 지시를 정확히 구현함.",
        "entities": "흙투성이 얼굴에 미소를 머금은 의병들. 두 번째 인물이 안규홍 레퍼런스의 얼굴 및 헤어스타일과 정확히 일치함. (맨 오른쪽 인물은 입가에 피가 묻어있지 않은 아쉬움이 있음)",
        "hard_violations": [],
        "physics": "인물들의 자세와 시선이 자연스러우며, 허공에 떠 있거나 물리적 법칙에 어긋나는 객체 및 신체가 없음."
       }
      ],
      "all_candidates_fail": false
     },
     "normalized": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 8,
        "verdict_ko": "나무를 활용해 인물들의 얼굴을 공간적으로 분리한 미디엄 샷의 구도를 훌륭하게 구현하였으며, 치명적인 물리적 오류 없이 지시사항을 잘 반영함."
       },
       {
        "label": "A",
        "score": 2,
        "verdict_ko": "소총을 쥐고 있는 신체와 연결되지 않은 기형적인 손과 나무를 관통한 무기 등 심각한 물리적 오류(Hard Violation)가 발생하여 실격됨."
       }
      ],
      "all_candidates_fail": false,
      "readings": [
       {
        "label": "A",
        "direction": "인물들이 서로를 향해 시선을 두며 마주보고 있음.",
        "built_space": "지시된 산채의 잔해와 나무가 배경 및 전경에 배치되어 공간을 형성함.",
        "entities": "피 묻은 입가와 흙투성이 얼굴을 한 의병들이며 옅은 미소를 띠고 있음. 안규홍 레퍼런스와 일치하는 인물이 포함되어 있으나, 여러 인물의 얼굴이 동일하게 복제된 형태임.",
        "hard_violations": [
         "신체와 제대로 연결되지 않고 소총을 쥐고 있는 중앙 하단의 불가능한 구조의 손",
         "오른쪽 나무 기둥을 관통하여 융합되어 버린 소총의 총신"
        ],
        "physics": "하단 중앙의 소총을 쥐고 있는 손은 어떤 인물의 팔과도 자연스럽게 연결되지 않아 물리적으로 불가능한 위치에 있음. 우측 나무 뒤 인물이 든 소총은 나무 기둥과 물리적으로 겹쳐짐."
       },
       {
        "label": "B",
        "direction": "네 명의 인물이 양옆으로 나뉘어 서로를 마주보며 시선을 교환하고 있음.",
        "built_space": "나무들이 인물들 사이의 간격을 나누며 전경에 위치하고, 배경에는 산채 잔해가 위치해 프레이밍 지시를 정확히 구현함.",
        "entities": "흙투성이 얼굴에 미소를 머금은 의병들. 두 번째 인물이 안규홍 레퍼런스의 얼굴 및 헤어스타일과 정확히 일치함. (맨 오른쪽 인물은 입가에 피가 묻어있지 않은 아쉬움이 있음)",
        "hard_violations": [],
        "physics": "인물들의 자세와 시선이 자연스러우며, 허공에 떠 있거나 물리적 법칙에 어긋나는 객체 및 신체가 없음."
       }
      ]
     },
     "ok": true
    },
    {
     "model": "gpt-high",
     "order": "reverse",
     "display_to_canonical": {
      "A": "B",
      "B": "A"
     },
     "raw": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 7,
        "verdict_ko": "서로 향한 시선과 옅은 미소는 잘 맞지만, 얼굴들이 두 줄로 대칭 배치되고 크게 잘려 엇갈린 얼굴 사이를 나무가 나누는 미디엄 숏 구성이 B보다 약하다."
       },
       {
        "label": "B",
        "score": 8,
        "verdict_ko": "허리 부근까지 담은 미디엄 숏에서 엇갈린 의병들의 얼굴과 그 사이 나무, 상호 시선을 더 충실하게 구현했지만 일부 얼굴의 피 묻은 입가와 이전 부상 상태는 부족하다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "왼쪽 두 남자는 오른쪽 동료들을, 오른쪽 두 남자는 왼쪽 동료들을 바라본다. 앞쪽 두 사람과 안쪽 두 사람이 각각 마주 보는 관계로 읽히며 카메라를 응시하지 않는다. 겨누는 무기나 이동 중인 물체는 보이지 않는다.",
        "built_space": "남자 네 명이 좌우 두 명씩 서 있고, 중앙의 굵은 나무줄기 두 개가 그 사이 배경을 나눈다. 뒤에는 초가지붕 막사 한 동이 부분적으로 보이고, 흙바닥에 항아리·그릇·부서진 목재가 놓여 있다. 이전 장면의 초가와 폐허 재료는 이어지지만, 얼굴 배치는 비교적 대칭적이다. 안쪽 왼쪽 인물은 허리 부근까지, 앞쪽 인물들은 주로 가슴 위로 잘린다. 복제된 고정 설비나 부자연스러운 반사는 확인되지 않는다.",
        "entities": "한국인 의병 설정에 부합하는 외양의 청년~중년 남성 네 명이며 검은 머리, 때 묻은 베이지색·갈색 전통 의복을 갖췄다. 왼쪽 인물들의 묶은 머리와 얼굴은 안규홍 참고와 유사하지만 어느 한 사람이 안규홍인지 구별은 선명하지 않다. 모두 미소가 옅고 얼굴에 흙과 상처가 있다. 왼쪽 앞과 오른쪽 안쪽 인물은 입가의 피가 보이지만, 오른쪽 앞 인물은 입 주변 혈흔이 거의 없다. 낮의 부드러운 광선과 흙빛 색조를 유지하며 읽을 수 있는 글자는 없다.",
        "hard_violations": [],
        "physics": "네 사람의 상체는 세워져 있으며 공중에 뜨거나 비정상적으로 꺾인 몸은 없다. 발과 지면의 접촉은 화면 밖이라 직접 확인할 수 없지만, 서 있는 자세와 모순되는 단서는 없다. 항아리와 그릇은 흙바닥에 놓여 있고 나무 상자와 판자는 땅이나 다른 잔해에 기대어 지지된다."
       },
       {
        "label": "B",
        "direction": "왼쪽의 세 남자는 오른쪽 두 동료를 향하고, 오른쪽 두 남자는 왼쪽 무리를 바라본다. 앞쪽 양끝 인물과 안쪽 인물들의 시선이 서로 다른 깊이에서 교차해 마주 보는 순간이 분명하다. 양끝의 장총 총구는 위쪽을 향하며 동료를 겨누지 않는다. 아래쪽의 다른 총들은 낮게 들려 있고, 서로에게 조준하는 동작은 보이지 않는다.",
        "built_space": "남자 다섯 명이 왼쪽 세 명과 오른쪽 두 명으로 깊이를 달리해 서 있다. 굵은 나무줄기 세 개가 인물 사이와 뒤에 자리해 배경을 여러 간격으로 분리한다. 뒤쪽 좌우에는 초가 막사 두 동의 일부가 보이고, 가운데 흙바닥에는 그릇·음식·항아리·목재 잔해가 흩어져 있다. 참고의 낮 산중 폐영지와 재료 구성이 이어진다. 대체로 허리 부근까지 담아 미디엄 숏 요구가 더 명확하며, 나무에 가려진 인물도 나무 뒤에 서 있는 것으로 자연스럽게 읽힌다.",
        "entities": "한국인 의병 설정에 부합하는 외양의 청년~중년 남성 다섯 명이 검은 머리와 낡은 전통 의복을 갖췄다. 왼쪽 앞 두 인물의 얼굴과 묶은 머리는 안규홍 참고와 유사하나 개별 정체성 구분은 다소 모호하다. 여러 얼굴에 흙과 피가 있고 특히 오른쪽 앞 인물은 입가 혈흔이 뚜렷하다. 오른쪽 안쪽 인물은 얼굴과 입가가 상대적으로 깨끗해 부상 상태의 일관성이 약하다. 미소는 대체로 절제되지만 왼쪽 일부는 A보다 조금 더 밝다. 장총들도 보이며, 읽을 수 있는 글자는 없다.",
        "hard_violations": [],
        "physics": "인물들은 몸을 약간 돌리거나 기울인 채 서 있으며 상체의 무게 배분이 자연스럽다. 발은 프레임 밖이지만 부유하거나 지지 불가능한 자세는 보이지 않는다. 양끝의 세운 장총은 손으로 잡고 있고, 낮게 든 총들에도 손의 접촉이 보인다. 가운데 왼쪽 인물의 수평에 가까운 총은 손바닥이 아래에서 받친다. 그릇과 음식, 전투 잔해는 모두 지면에 놓여 있다."
       }
      ],
      "all_candidates_fail": false
     },
     "normalized": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 7,
        "verdict_ko": "서로 향한 시선과 옅은 미소는 잘 맞지만, 얼굴들이 두 줄로 대칭 배치되고 크게 잘려 엇갈린 얼굴 사이를 나무가 나누는 미디엄 숏 구성이 B보다 약하다."
       },
       {
        "label": "A",
        "score": 8,
        "verdict_ko": "허리 부근까지 담은 미디엄 숏에서 엇갈린 의병들의 얼굴과 그 사이 나무, 상호 시선을 더 충실하게 구현했지만 일부 얼굴의 피 묻은 입가와 이전 부상 상태는 부족하다."
       }
      ],
      "all_candidates_fail": false,
      "readings": [
       {
        "label": "B",
        "direction": "왼쪽 두 남자는 오른쪽 동료들을, 오른쪽 두 남자는 왼쪽 동료들을 바라본다. 앞쪽 두 사람과 안쪽 두 사람이 각각 마주 보는 관계로 읽히며 카메라를 응시하지 않는다. 겨누는 무기나 이동 중인 물체는 보이지 않는다.",
        "built_space": "남자 네 명이 좌우 두 명씩 서 있고, 중앙의 굵은 나무줄기 두 개가 그 사이 배경을 나눈다. 뒤에는 초가지붕 막사 한 동이 부분적으로 보이고, 흙바닥에 항아리·그릇·부서진 목재가 놓여 있다. 이전 장면의 초가와 폐허 재료는 이어지지만, 얼굴 배치는 비교적 대칭적이다. 안쪽 왼쪽 인물은 허리 부근까지, 앞쪽 인물들은 주로 가슴 위로 잘린다. 복제된 고정 설비나 부자연스러운 반사는 확인되지 않는다.",
        "entities": "한국인 의병 설정에 부합하는 외양의 청년~중년 남성 네 명이며 검은 머리, 때 묻은 베이지색·갈색 전통 의복을 갖췄다. 왼쪽 인물들의 묶은 머리와 얼굴은 안규홍 참고와 유사하지만 어느 한 사람이 안규홍인지 구별은 선명하지 않다. 모두 미소가 옅고 얼굴에 흙과 상처가 있다. 왼쪽 앞과 오른쪽 안쪽 인물은 입가의 피가 보이지만, 오른쪽 앞 인물은 입 주변 혈흔이 거의 없다. 낮의 부드러운 광선과 흙빛 색조를 유지하며 읽을 수 있는 글자는 없다.",
        "hard_violations": [],
        "physics": "네 사람의 상체는 세워져 있으며 공중에 뜨거나 비정상적으로 꺾인 몸은 없다. 발과 지면의 접촉은 화면 밖이라 직접 확인할 수 없지만, 서 있는 자세와 모순되는 단서는 없다. 항아리와 그릇은 흙바닥에 놓여 있고 나무 상자와 판자는 땅이나 다른 잔해에 기대어 지지된다."
       },
       {
        "label": "A",
        "direction": "왼쪽의 세 남자는 오른쪽 두 동료를 향하고, 오른쪽 두 남자는 왼쪽 무리를 바라본다. 앞쪽 양끝 인물과 안쪽 인물들의 시선이 서로 다른 깊이에서 교차해 마주 보는 순간이 분명하다. 양끝의 장총 총구는 위쪽을 향하며 동료를 겨누지 않는다. 아래쪽의 다른 총들은 낮게 들려 있고, 서로에게 조준하는 동작은 보이지 않는다.",
        "built_space": "남자 다섯 명이 왼쪽 세 명과 오른쪽 두 명으로 깊이를 달리해 서 있다. 굵은 나무줄기 세 개가 인물 사이와 뒤에 자리해 배경을 여러 간격으로 분리한다. 뒤쪽 좌우에는 초가 막사 두 동의 일부가 보이고, 가운데 흙바닥에는 그릇·음식·항아리·목재 잔해가 흩어져 있다. 참고의 낮 산중 폐영지와 재료 구성이 이어진다. 대체로 허리 부근까지 담아 미디엄 숏 요구가 더 명확하며, 나무에 가려진 인물도 나무 뒤에 서 있는 것으로 자연스럽게 읽힌다.",
        "entities": "한국인 의병 설정에 부합하는 외양의 청년~중년 남성 다섯 명이 검은 머리와 낡은 전통 의복을 갖췄다. 왼쪽 앞 두 인물의 얼굴과 묶은 머리는 안규홍 참고와 유사하나 개별 정체성 구분은 다소 모호하다. 여러 얼굴에 흙과 피가 있고 특히 오른쪽 앞 인물은 입가 혈흔이 뚜렷하다. 오른쪽 안쪽 인물은 얼굴과 입가가 상대적으로 깨끗해 부상 상태의 일관성이 약하다. 미소는 대체로 절제되지만 왼쪽 일부는 A보다 조금 더 밝다. 장총들도 보이며, 읽을 수 있는 글자는 없다.",
        "hard_violations": [],
        "physics": "인물들은 몸을 약간 돌리거나 기울인 채 서 있으며 상체의 무게 배분이 자연스럽다. 발은 프레임 밖이지만 부유하거나 지지 불가능한 자세는 보이지 않는다. 양끝의 세운 장총은 손으로 잡고 있고, 낮게 든 총들에도 손의 접촉이 보인다. 가운데 왼쪽 인물의 수평에 가까운 총은 손바닥이 아래에서 받친다. 그릇과 음식, 전투 잔해는 모두 지면에 놓여 있다."
       }
      ]
     },
     "ok": true
    }
   ],
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "slot_winner_match": false,
   "slot_winner": {
    "gemini-pro": "B",
    "gpt-high": "A"
   },
   "route": "cross_slot_combined"
  },
  "dual": {
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "normalized": {
    "A": 1.25,
    "B": 1.875
   },
   "adjusted": {
    "A": 1.0,
    "B": 1.875
   },
   "violations": {
    "A": [
     "[gemini-pro] 신체와 제대로 연결되지 않고 소총을 쥐고 있는 중앙 하단의 불가능한 구조의 손",
     "[gemini-pro] 오른쪽 나무 기둥을 관통하여 융합되어 버린 소총의 총신"
    ]
   },
   "per_model_winner": {
    "gemini-pro": "B",
    "gpt-high": "A"
   },
   "agreed": false
  },
  "totals": {
   "B": 1875,
   "A": 1000
  },
  "selected": "B",
  "ranking": [
   "B",
   "A"
  ],
  "verdicts": [
   {
    "label": "B",
    "score": 1875,
    "verdict_ko": "나무를 활용해 인물들의 얼굴을 공간적으로 분리한 미디엄 샷의 구도를 훌륭하게 구현하였으며, 치명적인 물리적 오류 없이 지시사항을 잘 반영함."
   },
   {
    "label": "A",
    "score": 1000,
    "verdict_ko": "소총을 쥐고 있는 신체와 연결되지 않은 기형적인 손과 나무를 관통한 무기 등 심각한 물리적 오류(Hard Violation)가 발생하여 실격됨.  ★위반: [gemini-pro] 신체와 제대로 연결되지 않고 소총을 쥐고 있는 중앙 하단의 불가능한 구조의 손 / [gemini-pro] 오른쪽 나무 기둥을 관통하여 융합되어 버린 소총의 총신"
   }
  ],
  "refs": [
   {
    "label": "PREVIOUS SHOT STILL — a visually related earlier shot of this same place: the location's look, materials, fixed features, lighting mood and each person's clothing are LOCKED to this photo; never copy its camera framing. If this photo shows a character who cannot move (dead or unconscious), that character's exact pose, position and orientation are ALSO LOCKED — treat their body as a fixed prop of the set that only the camera moves around.",
    "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/db00517b-165b-4005-a287-f28936b59b95/images/272bd7af-9294-4490-98c6-88a40e9deacc/scene/recipe/S66sh1_sel.png",
    "asset_id": "3106acf2-e5a0-4486-8efc-7bcc7dfd026f",
    "role": "prev_still"
   },
   {
    "label": "CHARACTER REFERENCE — 안규홍: the exact person appearing in this shot; match face, hair and build exactly.",
    "path": "<bytes:1190344>",
    "asset_id": "d304dd05-fa51-4e13-8f3a-5e5586f7b520",
    "role": "character_ref"
   }
  ],
  "critique_skipped": true,
  "shot_run_produced": false,
  "shot_run_uid": "06aa1446-77a0-7935-b322-a1ed7885f13d",
  "ref_mode": "prev+엔티티",
  "share_plan": {
   "ref_plan": "prev",
   "prev_anchor_tag": "S66sh1"
  }
 },
 "S66sh9::cine": {
  "applied": false,
  "attempted_at": "2026-09-09T09:38:25.095679+00:00",
  "fingerprint": "e4ee76859611859fad710ba3a85977197efa2909aab7f35f177125f2a3e49d7d",
  "fingerprint_version": 2,
  "provider": "mai",
  "endpoint": "openrouter/chat-completions",
  "model": "microsoft/mai-image-2.6",
  "multiroll_tag": "still_S66sh9_cine",
  "slot": "",
  "pack": "24.202608252115",
  "source_file": "S66sh9_sel.png",
  "source_sha256": "4212be04f405f462f1bd917b0a624d30c109c04810cb9d70cafc6b65b171f51c",
  "error": "RuntimeError: moderation blocked: {\"message\": \"Response content blocked by label 'MultiSeverity_ViolenceScore'.\", \"code\": 400}",
  "moderation_refusals": 2,
  "declined": true,
  "declined_reason": "moderation"
 },
 "S66sh22::signage": {
  "fp": "e5595f04d8ffc2bc",
  "inscriptions": [],
  "cues": [],
  "dropped": []
 },
 "groupbg::final_cypress_passage": {
  "input_fingerprint": "8c2f3fa7deb911ce",
  "meta": {
   "model": "gpt-image-2",
   "size": "1536x864",
   "pack": "11.202607220237",
   "contract": "bgfirst_full_v3",
   "group_sig": {
    "key": "final_cypress_passage",
    "tags": [
     "S66sh22"
    ]
   },
   "context_sig": "74859291cfdebc85"
  },
  "prompt": "Create ONE empty live-action location background photograph — NO PEOPLE, no figures, no body parts, no silhouettes, no shadows or reflections of people anywhere.\n\"Empty\" means no people only: KEEP the location's inherent occupants and stock that define the place — animals in an animal shelter, pen or farm, goods and produce in a market, moored boats in a harbour — unless the shot text explicitly removes them.\nThe FIRST attached image is a thin-line storyboard sketch of a shot that happens at this location: use it ONLY as spatial evidence — what this place contains, how its ground, structures and landmarks are arranged and proportioned. Ignore the sketched people and arrows entirely, and do NOT copy its line style: render a fully photographic, physically plausible real place that fits THE LOCATION text below.\nHUMAN-SCALE CALIBRATION: derive every structure's true size from human-scale elements — a door ≈ 2m, a window ≈ 1–1.5m wide, one storey ≈ 2.5–3m; never inflate a small structure or shrink a large one.\n\nTHE LOCATION — early 20th-century Korea, Jeolla region; all people are Korean unless explicitly stated otherwise: The shaded forest path leading onward between cypress trees from the ruined mountain camp, with its far end darkened by tree cover.\n\nLOCATION DETAIL (descriptive evidence about this place from the production's location records — evidence, not a staging order): use it to understand what kind of place this is and what it permanently contains. Carry over ONLY its enduring physical features — layout, structures, ground and materials, fixed equipment and installations. Do NOT treat any lighting, weather or time-of-day wording, momentary object states, depicted events or subjective impressions in it as fixed facts of the place; the TIME OF DAY lock below is the sole authority for time and lighting.\n파청 산채 공터·훈련장·취사장·막사 주변: 산 중턱에 통나무를 세워 만든 조잡한 목책 안의 넓은 흙바닥 공터. 나무와 거적으로 만든 움막과 막사가 흩어져 있고, 돌과 진흙으로 만든 취사용 아궁이와 무쇠 솥, 모닥불 터가 있다. 전투 후에는 솥이 뒤집히고 막사가 불타거나 무너진다. (특징: 통나무를 박아 세운 목책; 넓게 다져진 흙바닥 (공터/훈련장); 돌과 진흙으로 만든 다수의 아궁이와 무쇠솥; 나무와 거적으로 엮어 만든 움막과 막사; 부서진 토기 밥그릇과 뒤집힌 무쇠솥 (전투 후); 불에 탄 나무 잔해와 흩어진 쌀밥 (전투 후))\n\nSCENE EVIDENCE (verbatim quotes from the screenplay about this place — treat them as evidence of what the location physically contains and looks like; stage the PLACE those moments happen in, but do NOT depict the momentary actions, people or staged props themselves):\n- 편백 사이로 피투성이 사람들이 정면을 향해 걸어온다.\n- 어디로 가는지는 보이지 않는다.\n\nTIME OF DAY (lock): day.\n\nRender ONE photorealistic empty location photograph, 16:9, neutral enough that every shot of this place can be staged from it later. No readable writing anywhere: surfaces that would carry writing may be present, but stage any wording out of legibility — an oblique angle, distance, shallow focus. No captions, watermarks or overlay text.",
  "effective_prompt": "Create ONE empty live-action location background photograph — NO PEOPLE, no figures, no body parts, no silhouettes, no shadows or reflections of people anywhere.\n\"Empty\" means no people only: KEEP the location's inherent occupants and stock that define the place — animals in an animal shelter, pen or farm, goods and produce in a market, moored boats in a harbour — unless the shot text explicitly removes them.\nThe FIRST attached image is a thin-line storyboard sketch of a shot that happens at this location: use it ONLY as spatial evidence — what this place contains, how its ground, structures and landmarks are arranged and proportioned. Ignore the sketched people and arrows entirely, and do NOT copy its line style: render a fully photographic, physically plausible real place that fits THE LOCATION text below.\nHUMAN-SCALE CALIBRATION: derive every structure's true size from human-scale elements — a door ≈ 2m, a window ≈ 1–1.5m wide, one storey ≈ 2.5–3m; never inflate a small structure or shrink a large one.\n\nTHE LOCATION — early 20th-century Korea, Jeolla region; all people are Korean unless explicitly stated otherwise: The shaded forest path leading onward between cypress trees from the ruined mountain camp, with its far end darkened by tree cover.\n\nLOCATION DETAIL (descriptive evidence about this place from the production's location records — evidence, not a staging order): use it to understand what kind of place this is and what it permanently contains. Carry over ONLY its enduring physical features — layout, structures, ground and materials, fixed equipment and installations. Do NOT treat any lighting, weather or time-of-day wording, momentary object states, depicted events or subjective impressions in it as fixed facts of the place; the TIME OF DAY lock below is the sole authority for time and lighting.\n파청 산채 공터·훈련장·취사장·막사 주변: 산 중턱에 통나무를 세워 만든 조잡한 목책 안의 넓은 흙바닥 공터. 나무와 거적으로 만든 움막과 막사가 흩어져 있고, 돌과 진흙으로 만든 취사용 아궁이와 무쇠 솥, 모닥불 터가 있다. 전투 후에는 솥이 뒤집히고 막사가 불타거나 무너진다. (특징: 통나무를 박아 세운 목책; 넓게 다져진 흙바닥 (공터/훈련장); 돌과 진흙으로 만든 다수의 아궁이와 무쇠솥; 나무와 거적으로 엮어 만든 움막과 막사; 부서진 토기 밥그릇과 뒤집힌 무쇠솥 (전투 후); 불에 탄 나무 잔해와 흩어진 쌀밥 (전투 후))\n\nSCENE EVIDENCE (verbatim quotes from the screenplay about this place — treat them as evidence of what the location physically contains and looks like; stage the PLACE those moments happen in, but do NOT depict the momentary actions, people or staged props themselves):\n- 편백 사이로 피투성이 사람들이 정면을 향해 걸어온다.\n- 어디로 가는지는 보이지 않는다.\n\nTIME OF DAY (lock): day.\n\nRender ONE photorealistic empty location photograph, 16:9, neutral enough that every shot of this place can be staged from it later. No readable writing anywhere: surfaces that would carry writing may be present, but stage any wording out of legibility — an oblique angle, distance, shallow focus. No captions, watermarks or overlay text.",
  "bg_path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/db00517b-165b-4005-a287-f28936b59b95/images/272bd7af-9294-4490-98c6-88a40e9deacc/scene/recipe/groupbg_final_cypress_passage_424bfd.png",
  "asset_id": "cc46d830-70ab-4b69-803c-f0db7dc29047",
  "input_asset_ids": [
   "0b420f52-818e-4a0b-9031-18736e4cd7ee"
  ],
  "origin_tag": "S66sh22",
  "place_text": "The shaded forest path leading onward between cypress trees from the ruined mountain camp, with its far end darkened by tree cover.",
  "origin_inputs": {
   "place_text": "The shaded forest path leading onward between cypress trees from the ruined mountain camp, with its far end darkened by tree cover.",
   "time_of_day_en": "day",
   "conti_asset_id": "0b420f52-818e-4a0b-9031-18736e4cd7ee"
  }
 },
 "S66sh22::bgfirst_bg": {
  "input_fingerprint": "347866e3ff298a0d",
  "prompt": "Create the EMPTY BACKGROUND PLATE for one film shot — NO PEOPLE, no figures, no body parts, no sketch lines, no arrows anywhere.\n\"Empty\" means no people only: KEEP the location's inherent occupants and stock that define the place — animals in an animal shelter, pen or farm, goods in a market, moored boats in a harbour — unless the shot text explicitly removes them.\nThe FIRST attached image is a thin-line storyboard sketch: use ONLY its camera angle, horizon, perspective and the placement/size of buildings and set masses — ignore the sketched people and arrows entirely. The SECOND attached image (LOCATION PHOTOGRAPH) is the real place: take its architecture, materials, signage and fixed features, and RE-PROJECT them into the sketch's camera. If the photograph's camera differs from the sketch's, the sketch's camera wins.\nHUMAN-SCALE CALIBRATION: derive every structure's true size from human-scale elements — a door ≈ 2m, a window ≈ 1–1.5m wide, one storey ≈ 2.5–3m; never inflate a small structure or shrink a large one.\n\nSHOT TEXT this background must serve (Korean): 어둠이 깔린 숲길 끝을 향해 한쪽 발을 내디딘 mid-stride 자세인 안규홍의 굳건한 뒷모습.\n\nLOCATION (lock): The shaded forest path leading onward between cypress trees from the ruined mountain camp, with its far end darkened by tree cover.\n\nTIME OF DAY (lock): day.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: wide shot\n- KEY BACKGROUND ELEMENTS: 피 묻은 낫 (Bloodied and carried in 안규홍's lowered hand) — Hangs beside his body, seen obliquely from behind; used as A small, legible continuation of his resolve beside the walking figure; 편백 (The walking route passes between the trees); used as Trunks at the lateral edges contain the rear view without exposing the destination; 발밑의 이동 경로 (Only the nearby ground traversed by 안규홍 is visible; the route's endpoint is outside the frame); used as Provides space to read the advancing foot while withholding where the group is going.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Maintain the scene's daylight and sober naturalistic contrast through the final walking image, reserving darkness for the subsequent cut to black.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nRender ONE photorealistic empty location photograph, 16:9, that this shot can be staged inside later. No readable writing anywhere: surfaces that would carry writing may be present, but stage any wording out of legibility — an oblique angle, distance, shallow focus. No captions, watermarks or overlay text.",
  "effective_prompt": "Create the EMPTY BACKGROUND PLATE for one film shot — NO PEOPLE, no figures, no body parts, no sketch lines, no arrows anywhere.\n\"Empty\" means no people only: KEEP the location's inherent occupants and stock that define the place — animals in an animal shelter, pen or farm, goods in a market, moored boats in a harbour — unless the shot text explicitly removes them.\nThe FIRST attached image is a thin-line storyboard sketch: use ONLY its camera angle, horizon, perspective and the placement/size of buildings and set masses — ignore the sketched people and arrows entirely. The SECOND attached image (LOCATION PHOTOGRAPH) is the real place: take its architecture, materials, signage and fixed features, and RE-PROJECT them into the sketch's camera. If the photograph's camera differs from the sketch's, the sketch's camera wins.\nHUMAN-SCALE CALIBRATION: derive every structure's true size from human-scale elements — a door ≈ 2m, a window ≈ 1–1.5m wide, one storey ≈ 2.5–3m; never inflate a small structure or shrink a large one.\n\nSHOT TEXT this background must serve (Korean): 어둠이 깔린 숲길 끝을 향해 한쪽 발을 내디딘 mid-stride 자세인 안규홍의 굳건한 뒷모습.\n\nLOCATION (lock): The shaded forest path leading onward between cypress trees from the ruined mountain camp, with its far end darkened by tree cover.\n\nTIME OF DAY (lock): day.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: wide shot\n- KEY BACKGROUND ELEMENTS: 피 묻은 낫 (Bloodied and carried in 안규홍's lowered hand) — Hangs beside his body, seen obliquely from behind; used as A small, legible continuation of his resolve beside the walking figure; 편백 (The walking route passes between the trees); used as Trunks at the lateral edges contain the rear view without exposing the destination; 발밑의 이동 경로 (Only the nearby ground traversed by 안규홍 is visible; the route's endpoint is outside the frame); used as Provides space to read the advancing foot while withholding where the group is going.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Maintain the scene's daylight and sober naturalistic contrast through the final walking image, reserving darkness for the subsequent cut to black.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nRender ONE photorealistic empty location photograph, 16:9, that this shot can be staged inside later. No readable writing anywhere: surfaces that would carry writing may be present, but stage any wording out of legibility — an oblique angle, distance, shallow focus. No captions, watermarks or overlay text.",
  "bg_path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/db00517b-165b-4005-a287-f28936b59b95/images/272bd7af-9294-4490-98c6-88a40e9deacc/scene/recipe/S66sh22__bgfirst_bg.png",
  "asset_id": "7a3d2d0c-bd05-4b60-8194-d5bdceca9381",
  "input_asset_ids": [
   "0b420f52-818e-4a0b-9031-18736e4cd7ee",
   "cc46d830-70ab-4b69-803c-f0db7dc29047"
  ]
 },
 "S66sh22": {
  "input_fingerprint": "732dd809699a8f52",
  "prompt": "Create ONE FINAL photorealistic live-action film still of the moment below — early 20th-century Korea, Jeolla region; all people are Korean unless explicitly stated otherwise. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 어둠이 깔린 숲길 끝을 향해 한쪽 발을 내디딘 mid-stride 자세인 안규홍의 굳건한 뒷모습.\n\nLOCATION (lock): The shaded forest path leading onward between cypress trees from the ruined mountain camp, with its far end darkened by tree cover. The shot takes place here — the attached LOCATION PHOTOGRAPH shows the exact spot.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: wide shot\n- KEY BACKGROUND ELEMENTS: 피 묻은 낫 (Bloodied and carried in 안규홍's lowered hand) — Hangs beside his body, seen obliquely from behind; used as A small, legible continuation of his resolve beside the walking figure; 편백 (The walking route passes between the trees); used as Trunks at the lateral edges contain the rear view without exposing the destination; 발밑의 이동 경로 (Only the nearby ground traversed by 안규홍 is visible; the route's endpoint is outside the frame); used as Provides space to read the advancing foot while withholding where the group is going.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Maintain the scene's daylight and sober naturalistic contrast through the final walking image, reserving darkness for the subsequent cut to black.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): It remains daytime among the cypress trees; the devastated camp and its scattered cookware and bloodied food are left behind. 안규홍: He walks at the front with a limp, the bloodied sickle hanging low in his hand; his face is blood- and dirt-covered, his back wound remains open and his neck is bare.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 안규홍 (한국인 남성, 청년 나이대, 검은 머리, 단단히 묶은 머리) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.",
  "roll_prompts": {
   "A": "Stage the shot. The FIRST attached image (SHOT BACKGROUND) is the finished empty background of this shot — keep it EXACTLY: its camera, perspective, architecture, lighting and every fixed feature stay untouched. The SECOND attached image (LAYOUT SKETCH) tells you ONLY where the people go: each sketched person's position, screen size, pose and the gaze/motion arrows. Ignore the sketch's background lines. The CHARACTER REFERENCE photographs show the real people.\nPlace the real people into the background at exactly the sketched positions, sizes and poses, following the arrow directions. No sketch lines or arrows may remain.\n\nCreate ONE FINAL photorealistic live-action film still of the moment below — early 20th-century Korea, Jeolla region; all people are Korean unless explicitly stated otherwise. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 어둠이 깔린 숲길 끝을 향해 한쪽 발을 내디딘 mid-stride 자세인 안규홍의 굳건한 뒷모습.\n\nLOCATION (lock): The shaded forest path leading onward between cypress trees from the ruined mountain camp, with its far end darkened by tree cover. The shot takes place here — the FIRST attached image (SHOT BACKGROUND) is this exact place, already built: its ground, structures, horizon, materials and lighting are the finished truth of this location and must not be redesigned or replaced. No location photograph is attached — read the place from that image alone, and add no scenery, structure, vehicle or fixture that it does not already show. This lock governs the place only; the figures in the shot follow the staging and pose instructions.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: wide shot\n- KEY BACKGROUND ELEMENTS: 피 묻은 낫 (Bloodied and carried in 안규홍's lowered hand) — Hangs beside his body, seen obliquely from behind; used as A small, legible continuation of his resolve beside the walking figure; 편백 (The walking route passes between the trees); used as Trunks at the lateral edges contain the rear view without exposing the destination; 발밑의 이동 경로 (Only the nearby ground traversed by 안규홍 is visible; the route's endpoint is outside the frame); used as Provides space to read the advancing foot while withholding where the group is going.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Maintain the scene's daylight and sober naturalistic contrast through the final walking image, reserving darkness for the subsequent cut to black.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): It remains daytime among the cypress trees; the devastated camp and its scattered cookware and bloodied food are left behind. 안규홍: He walks at the front with a limp, the bloodied sickle hanging low in his hand; his face is blood- and dirt-covered, his back wound remains open and his neck is bare.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 안규홍 (한국인 남성, 청년 나이대, 검은 머리, 단단히 묶은 머리) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.",
   "B": "Create ONE FINAL photorealistic live-action film still of the moment below — early 20th-century Korea, Jeolla region; all people are Korean unless explicitly stated otherwise. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 어둠이 깔린 숲길 끝을 향해 한쪽 발을 내디딘 mid-stride 자세인 안규홍의 굳건한 뒷모습.\n\nLOCATION (lock): The shaded forest path leading onward between cypress trees from the ruined mountain camp, with its far end darkened by tree cover. The shot takes place here — the attached LOCATION PHOTOGRAPH shows the exact spot.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- FRAMING SCALE: wide shot\n- KEY BACKGROUND ELEMENTS: 피 묻은 낫 (Bloodied and carried in 안규홍's lowered hand) — Hangs beside his body, seen obliquely from behind; used as A small, legible continuation of his resolve beside the walking figure; 편백 (The walking route passes between the trees); used as Trunks at the lateral edges contain the rear view without exposing the destination; 발밑의 이동 경로 (Only the nearby ground traversed by 안규홍 is visible; the route's endpoint is outside the frame); used as Provides space to read the advancing foot while withholding where the group is going.\nCompose the frame exactly as specified above — subject scale and screen placement. The viewing angle is deliberately left open here; choose the angle that serves the shot. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Maintain the scene's daylight and sober naturalistic contrast through the final walking image, reserving darkness for the subsequent cut to black.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging within whatever this brief already fixes,\nnot literal fantasy imagery.\n\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person it puts in\nthis shot, imagine that description and still show them as a\nconcrete, fully-formed human being:\nrender the human form as fully as the framing shows — never reduce\na person to a shape, blob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands naturally positioned for the action\nthe SHOT TEXT already gives them, gaze on the target the SHOT TEXT\nimplies; avoid a stiff attention stance (feet together, arms hanging\nstraight down) and a blank stare into the lens. If the SHOT TEXT or\nany POSE/IMMOBILE contract above stages stillness, death, sleep,\nunconsciousness, restraint, drill or an explicit direct-to-camera\nlook, follow THAT exactly — this clause never overrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nBODY & SUPPORT (default only — every explicit direction above wins): a body relates to what holds it. A seated person sits the way the seat is built to be used — hips on the seat, back toward the backrest, legs falling naturally toward the floor; people sharing adjacent seats each occupy their own seat, side by side. Hands, hips and feet keep believable contact with whatever they rest on. When the text stages someone frozen, stunned or holding still, the body stops mid-action exactly where the moment caught it — weight already committed to one side, hands where the interrupted movement left them — rather than resetting into a symmetric at-attention stance. If the SHOT TEXT or any contract above stages a specific arrangement, follow THAT exactly.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): It remains daytime among the cypress trees; the devastated camp and its scattered cookware and bloodied food are left behind. 안규홍: He walks at the front with a limp, the bloodied sickle hanging low in his hand; his face is blood- and dirt-covered, his back wound remains open and his neck is bare.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 안규홍 (한국인 남성, 청년 나이대, 검은 머리, 단단히 묶은 머리) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nCHARACTER REFERENCE ROLE (follow exactly): the attached CHARACTER REFERENCE images establish identity only — face, hair, build and clothing. The pose, gaze direction, camera angle and framing inside those reference images belong to the reference photos, not to this shot; never copy them. Pose and gaze in this shot follow only the CAMERA and action text of this prompt. When this prompt stages a character's face as covered or hidden — by a costume head, mask, helmet, hood, or a body turned away — that staging wins: keep the covering exactly as described and never pull the reference face into view from under it; the reference then guides only what stays visible, such as build and clothing.\n\nNo readable writing anywhere in this image. Surfaces that would carry writing may be present, but nothing a viewer could read: stage every one of them out of legibility — a hand across, an oblique angle, shallow focus, or simply turned away. No caption, subtitle, watermark, logo or overlay.\n\nCOMPOSITION VARIATION (this candidate only): keep the place, the people and the moment exactly as specified, and keep from the CAMERA & FRAME contract its subject scale, screen placement and key background elements exactly. For this candidate only, its camera angle and height are deliberately left open: choose a camera position distinctly different from the obvious one, as a film director picking a second setup on the same blocking. Stay true to any explicit framing the SHOT TEXT itself dictates."
  },
  "roll_refs": {
   "A": [
    {
     "label": "SHOT BACKGROUND",
     "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/db00517b-165b-4005-a287-f28936b59b95/images/272bd7af-9294-4490-98c6-88a40e9deacc/scene/recipe/S66sh22__bgfirst_bg.png",
     "asset_id": "7a3d2d0c-bd05-4b60-8194-d5bdceca9381",
     "role": "bgfirst_bg"
    },
    {
     "label": "LAYOUT SKETCH (people placement only)",
     "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/db00517b-165b-4005-a287-f28936b59b95/images/272bd7af-9294-4490-98c6-88a40e9deacc/conti/conti_S66sh22.png",
     "asset_id": "0b420f52-818e-4a0b-9031-18736e4cd7ee",
     "role": "conti_light"
    },
    {
     "label": "CHARACTER REFERENCE — 안규홍: the exact person appearing in this shot; match face, hair and build exactly.",
     "path": "<bytes:1028384>",
     "asset_id": "3b5894db-ce55-4f34-b966-5e725c3aa952",
     "role": "character_ref"
    }
   ],
   "B": [
    {
     "label": "LOCATION PHOTOGRAPH — the exact place of this shot: its architecture, materials, fixed features and lighting mood are spatial truth; stage the moment inside this place. Never copy its camera framing.",
     "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/db00517b-165b-4005-a287-f28936b59b95/images/272bd7af-9294-4490-98c6-88a40e9deacc/scene/recipe/groupbg_final_cypress_passage_424bfd.png",
     "asset_id": "cc46d830-70ab-4b69-803c-f0db7dc29047",
     "role": "bgfirst_group_bg"
    },
    {
     "label": "CHARACTER REFERENCE — 안규홍: the exact person appearing in this shot; match face, hair and build exactly.",
     "path": "<bytes:1028384>",
     "asset_id": "3b5894db-ce55-4f34-b966-5e725c3aa952",
     "role": "character_ref"
    }
   ]
  },
  "shot_run_spend_attempt_count": 1,
  "judge_flip_skipped": "judge_fn_owns_order",
  "readings": [
   {
    "label": "A",
    "direction": "안규홍의 몸과 시선이 카메라를 등지고 숲길 안쪽(어둠이 깔린 끝)을 향해 똑바로 향해 있음.",
    "built_space": "참조 이미지에 있는 텐트 뼈대, 가마솥, 아궁이 등 야영지의 주요 구조물들이 화면 좌우측에 정확히 배치되어 있음.",
    "entities": "안규홍(상의 탈의 상태, 등에 길게 난 상처, 기준에 맞는 바지와 짚신), 핏물이 떨어지는 낫, 편백나무 숲길.",
    "hard_violations": [],
    "physics": "오른손으로 낫의 자루를 정상적으로 쥐고 있으며, 왼발 뒤꿈치가 들려 앞으로 나아가는 자연스러운 보행(mid-stride)을 지탱하고 있음."
   },
   {
    "label": "B",
    "direction": "몸은 숲길 안쪽을 향하나, 고개를 오른쪽 뒤로 돌려 카메라 측면을 응시하고 있음.",
    "built_space": "숲길과 우측 목책만 보일 뿐, 참조 이미지의 핵심 공간인 무너진 야영지(텐트, 조리 도구 등)가 전혀 존재하지 않음.",
    "entities": "안규홍(등이 찢어진 한복 상의, 피가 묻은 얼굴), 피 묻은 낫, 편백나무 숲길.",
    "hard_violations": [],
    "physics": "오른손으로 낫을 쥐고 있고 왼발을 지면에 딛고 있는 걷는 자세가 물리적으로 성립함."
   }
  ],
  "cross_model_order": {
   "policy": "cross_model_order_v2_slot0_fwd_slot1_rev_parallel2",
   "slots": [
    {
     "model": "gemini-pro",
     "order": "forward",
     "display_to_canonical": {
      "A": "A",
      "B": "B"
     },
     "raw": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 8,
        "verdict_ko": "지정된 로케이션의 디테일(무너진 텐트와 가마솥)을 완벽히 재현하고 굳건히 걷는 뒷모습을 정확히 연출했으나, 상의를 완전히 탈의하여 복장 규정(우선순위 4)에서 감점이 발생했습니다."
       },
       {
        "label": "B",
        "score": 4,
        "verdict_ko": "찢어진 한복과 상처의 묘사는 좋으나, 로케이션 사진의 주요 배경 구조물(야영지 흔적)을 완전히 누락하였고 고개를 뒤로 돌려 프롬프트가 요구한 '앞을 향해 걷는 굳건한 뒷모습'의 구도를 훼손했습니다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "안규홍의 몸과 시선이 카메라를 등지고 숲길 안쪽(어둠이 깔린 끝)을 향해 똑바로 향해 있음.",
        "built_space": "참조 이미지에 있는 텐트 뼈대, 가마솥, 아궁이 등 야영지의 주요 구조물들이 화면 좌우측에 정확히 배치되어 있음.",
        "entities": "안규홍(상의 탈의 상태, 등에 길게 난 상처, 기준에 맞는 바지와 짚신), 핏물이 떨어지는 낫, 편백나무 숲길.",
        "hard_violations": [],
        "physics": "오른손으로 낫의 자루를 정상적으로 쥐고 있으며, 왼발 뒤꿈치가 들려 앞으로 나아가는 자연스러운 보행(mid-stride)을 지탱하고 있음."
       },
       {
        "label": "B",
        "direction": "몸은 숲길 안쪽을 향하나, 고개를 오른쪽 뒤로 돌려 카메라 측면을 응시하고 있음.",
        "built_space": "숲길과 우측 목책만 보일 뿐, 참조 이미지의 핵심 공간인 무너진 야영지(텐트, 조리 도구 등)가 전혀 존재하지 않음.",
        "entities": "안규홍(등이 찢어진 한복 상의, 피가 묻은 얼굴), 피 묻은 낫, 편백나무 숲길.",
        "hard_violations": [],
        "physics": "오른손으로 낫을 쥐고 있고 왼발을 지면에 딛고 있는 걷는 자세가 물리적으로 성립함."
       }
      ],
      "all_candidates_fail": false
     },
     "normalized": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 8,
        "verdict_ko": "지정된 로케이션의 디테일(무너진 텐트와 가마솥)을 완벽히 재현하고 굳건히 걷는 뒷모습을 정확히 연출했으나, 상의를 완전히 탈의하여 복장 규정(우선순위 4)에서 감점이 발생했습니다."
       },
       {
        "label": "B",
        "score": 4,
        "verdict_ko": "찢어진 한복과 상처의 묘사는 좋으나, 로케이션 사진의 주요 배경 구조물(야영지 흔적)을 완전히 누락하였고 고개를 뒤로 돌려 프롬프트가 요구한 '앞을 향해 걷는 굳건한 뒷모습'의 구도를 훼손했습니다."
       }
      ],
      "all_candidates_fail": false,
      "readings": [
       {
        "label": "A",
        "direction": "안규홍의 몸과 시선이 카메라를 등지고 숲길 안쪽(어둠이 깔린 끝)을 향해 똑바로 향해 있음.",
        "built_space": "참조 이미지에 있는 텐트 뼈대, 가마솥, 아궁이 등 야영지의 주요 구조물들이 화면 좌우측에 정확히 배치되어 있음.",
        "entities": "안규홍(상의 탈의 상태, 등에 길게 난 상처, 기준에 맞는 바지와 짚신), 핏물이 떨어지는 낫, 편백나무 숲길.",
        "hard_violations": [],
        "physics": "오른손으로 낫의 자루를 정상적으로 쥐고 있으며, 왼발 뒤꿈치가 들려 앞으로 나아가는 자연스러운 보행(mid-stride)을 지탱하고 있음."
       },
       {
        "label": "B",
        "direction": "몸은 숲길 안쪽을 향하나, 고개를 오른쪽 뒤로 돌려 카메라 측면을 응시하고 있음.",
        "built_space": "숲길과 우측 목책만 보일 뿐, 참조 이미지의 핵심 공간인 무너진 야영지(텐트, 조리 도구 등)가 전혀 존재하지 않음.",
        "entities": "안규홍(등이 찢어진 한복 상의, 피가 묻은 얼굴), 피 묻은 낫, 편백나무 숲길.",
        "hard_violations": [],
        "physics": "오른손으로 낫을 쥐고 있고 왼발을 지면에 딛고 있는 걷는 자세가 물리적으로 성립함."
       }
      ]
     },
     "ok": true
    },
    {
     "model": "gpt-high",
     "order": "reverse",
     "display_to_canonical": {
      "A": "B",
      "B": "A"
     },
     "raw": {
      "winner": "A",
      "ranking": [
       "A",
       "B"
      ],
      "verdicts": [
       {
        "label": "A",
        "score": 7,
        "verdict_ko": "야영지를 뒤로한 편백 숲길의 전신 뒷모습과 낮게 든 피 묻은 낫은 더 충실하지만, 옆으로 돌린 시선과 선명하지 않은 등 상처가 아쉽다."
       },
       {
        "label": "B",
        "score": 6,
        "verdict_ko": "전방을 향한 뒷모습과 열린 등 상처는 맞지만, 야영지와 취사도구를 뒤에 남기지 않은 구도이며 참고 의상의 상의도 사라졌다."
       }
      ],
      "readings": [
       {
        "label": "A",
        "direction": "몸과 걸음은 화면 중앙에서 뒤쪽의 어두운 숲길로 향한다. 다만 고개를 오른쪽으로 돌려 얼굴 옆면이 보이고, 시선은 진행 방향의 숲길 끝보다 오른쪽 아래를 향한다. 오른손의 낫은 몸 옆에서 아래로 늘어져 있으며 날끝은 화면 왼쪽 아래로 굽는다. 무언가를 겨누거나 휘두르는 동작은 아니다.",
        "built_space": "중앙 흙길 양옆을 편백 줄기와 목책이 둘러싼다. 왼쪽에는 가로대를 연결한 목책 구간, 오른쪽에는 촘촘한 세로 말뚝 구간이 보인다. 인물은 그 사이 통로 위에 있다. 야영지의 천막과 취사 설비는 보이지 않아 그것들을 뒤로하고 숲길에 들어선 배치에 가깝다. 가까운 보행 지면과 수목에 가려지는 길만 보이며 목적지는 노출되지 않는다. 참고 장소의 재질과 공간 유형은 맞지만 개별 나무와 목책 배열의 정확한 일치까지 확인되지는 않는다.",
        "entities": "인물은 한 명이며, 보이는 옆얼굴과 체격은 청년 동아시아 남성으로 참고 인물과 대체로 부합한다. 검은 머리는 뒤로 단단히 묶었고, 때 묻은 밝은 저고리와 바지, 허리띠, 짚신은 참고 의상을 따른다. 얼굴 옆면에는 피와 흙으로 보이는 얼룩이 있다. 목에 장식이나 감개는 없지만 저고리 깃이 목 아래를 일부 덮는다. 등에는 세로로 찢어진 옷 사이 피부가 보이나, 열린 상처 자체는 분명하지 않다. 오른손에는 붉게 얼룩진 굽은 금속 낫이 있다. 편백 숲과 낮의 자연광이 보이며 읽을 수 있는 글자나 추가 인물은 없다.",
        "hard_violations": [],
        "physics": "오른발이 지면에서 체중을 지탱하고 왼발은 뒤쪽에서 들려 밑창이 보인다. 지지발이 있는 정상적인 보행 중간 순간으로, 공중에 떠 있는 몸은 아니다. 절뚝거림은 이 한 장면만으로 뚜렷하지 않다. 낫 손잡이는 오른손이 쥐고 있으며 날은 그 아래로 내려온다. 옷의 주름과 낫의 무게 방향에도 명백한 물리적 모순은 없다."
       },
       {
        "label": "B",
        "direction": "인물의 머리와 몸은 카메라 반대편, 화면 중앙 깊숙한 어두운 숲길을 향한다. 얼굴과 눈은 보이지 않지만 고개 방향은 진행 방향과 일치한다. 오른손의 낫은 다리 옆으로 낮게 내려오며 날끝은 화면 왼쪽 아래로 굽는다. 공격하거나 다른 대상을 겨누는 모습은 아니다.",
        "built_space": "왼쪽에 무너진 천막 한 동, 오른쪽에 뚜껑 있는 솥 하나를 얹은 돌 화덕 하나가 보인다. 왼쪽 중경에는 작은 취사대와 그릇들이 있고, 전경에는 넘어진 큰 용기와 흩어진 그릇들이 있다. 양옆 목책과 편백, 중앙 통로의 배치는 장소 참고 사진에 매우 가깝다. 그러나 인물이 오른쪽 화덕 바로 옆의 야영지 내부를 지나고 있어, 취사도구와 폐허를 뒤로한 숲길이라는 장면 배치보다 참고 사진의 야영지 구도를 재현한다. 길의 실제 목적지는 수목에 가려져 있다.",
        "entities": "추가 인물 없이 한 명의 청년 남성이 보인다. 얼굴이 완전히 가려져 한국인 여부와 정확한 얼굴 일치는 확인할 수 없다. 검은 머리는 위쪽으로 묶었고 목은 맨살로 드러난다. 밝은 바지와 허리끈, 짚신은 참고 의상 계열이지만 저고리가 없어 상반신이 완전히 노출되어 있다. 등에 벌어진 붉은 상처와 피부의 때가 명확하다. 얼굴의 피와 흙은 뒷모습 때문에 평가할 수 없다. 오른손에는 피로 보이는 붉은 얼룩이 있는 낫이 있다. 편백과 낮의 자연광은 맞으며 읽을 수 있는 글자는 없다.",
        "hard_violations": [],
        "physics": "오른발이 흙길에 닿아 몸을 지탱하고 왼발은 뒤에서 들려 밑창이 보인다. 걷는 도중 가능한 자세이며 몸을 지지하는 발이 명확하다. 다만 절뚝이는 보행의 특징은 확실하지 않다. 오른손이 낫 손잡이를 잡고 있어 낫이 떠 있지 않다. 천막 잔해는 나무와 지면에 걸쳐 있고 솥은 돌 화덕에 놓여 있어 주변 물체의 지지도 자연스럽다."
       }
      ],
      "all_candidates_fail": false
     },
     "normalized": {
      "winner": "B",
      "ranking": [
       "B",
       "A"
      ],
      "verdicts": [
       {
        "label": "B",
        "score": 7,
        "verdict_ko": "야영지를 뒤로한 편백 숲길의 전신 뒷모습과 낮게 든 피 묻은 낫은 더 충실하지만, 옆으로 돌린 시선과 선명하지 않은 등 상처가 아쉽다."
       },
       {
        "label": "A",
        "score": 6,
        "verdict_ko": "전방을 향한 뒷모습과 열린 등 상처는 맞지만, 야영지와 취사도구를 뒤에 남기지 않은 구도이며 참고 의상의 상의도 사라졌다."
       }
      ],
      "all_candidates_fail": false,
      "readings": [
       {
        "label": "B",
        "direction": "몸과 걸음은 화면 중앙에서 뒤쪽의 어두운 숲길로 향한다. 다만 고개를 오른쪽으로 돌려 얼굴 옆면이 보이고, 시선은 진행 방향의 숲길 끝보다 오른쪽 아래를 향한다. 오른손의 낫은 몸 옆에서 아래로 늘어져 있으며 날끝은 화면 왼쪽 아래로 굽는다. 무언가를 겨누거나 휘두르는 동작은 아니다.",
        "built_space": "중앙 흙길 양옆을 편백 줄기와 목책이 둘러싼다. 왼쪽에는 가로대를 연결한 목책 구간, 오른쪽에는 촘촘한 세로 말뚝 구간이 보인다. 인물은 그 사이 통로 위에 있다. 야영지의 천막과 취사 설비는 보이지 않아 그것들을 뒤로하고 숲길에 들어선 배치에 가깝다. 가까운 보행 지면과 수목에 가려지는 길만 보이며 목적지는 노출되지 않는다. 참고 장소의 재질과 공간 유형은 맞지만 개별 나무와 목책 배열의 정확한 일치까지 확인되지는 않는다.",
        "entities": "인물은 한 명이며, 보이는 옆얼굴과 체격은 청년 동아시아 남성으로 참고 인물과 대체로 부합한다. 검은 머리는 뒤로 단단히 묶었고, 때 묻은 밝은 저고리와 바지, 허리띠, 짚신은 참고 의상을 따른다. 얼굴 옆면에는 피와 흙으로 보이는 얼룩이 있다. 목에 장식이나 감개는 없지만 저고리 깃이 목 아래를 일부 덮는다. 등에는 세로로 찢어진 옷 사이 피부가 보이나, 열린 상처 자체는 분명하지 않다. 오른손에는 붉게 얼룩진 굽은 금속 낫이 있다. 편백 숲과 낮의 자연광이 보이며 읽을 수 있는 글자나 추가 인물은 없다.",
        "hard_violations": [],
        "physics": "오른발이 지면에서 체중을 지탱하고 왼발은 뒤쪽에서 들려 밑창이 보인다. 지지발이 있는 정상적인 보행 중간 순간으로, 공중에 떠 있는 몸은 아니다. 절뚝거림은 이 한 장면만으로 뚜렷하지 않다. 낫 손잡이는 오른손이 쥐고 있으며 날은 그 아래로 내려온다. 옷의 주름과 낫의 무게 방향에도 명백한 물리적 모순은 없다."
       },
       {
        "label": "A",
        "direction": "인물의 머리와 몸은 카메라 반대편, 화면 중앙 깊숙한 어두운 숲길을 향한다. 얼굴과 눈은 보이지 않지만 고개 방향은 진행 방향과 일치한다. 오른손의 낫은 다리 옆으로 낮게 내려오며 날끝은 화면 왼쪽 아래로 굽는다. 공격하거나 다른 대상을 겨누는 모습은 아니다.",
        "built_space": "왼쪽에 무너진 천막 한 동, 오른쪽에 뚜껑 있는 솥 하나를 얹은 돌 화덕 하나가 보인다. 왼쪽 중경에는 작은 취사대와 그릇들이 있고, 전경에는 넘어진 큰 용기와 흩어진 그릇들이 있다. 양옆 목책과 편백, 중앙 통로의 배치는 장소 참고 사진에 매우 가깝다. 그러나 인물이 오른쪽 화덕 바로 옆의 야영지 내부를 지나고 있어, 취사도구와 폐허를 뒤로한 숲길이라는 장면 배치보다 참고 사진의 야영지 구도를 재현한다. 길의 실제 목적지는 수목에 가려져 있다.",
        "entities": "추가 인물 없이 한 명의 청년 남성이 보인다. 얼굴이 완전히 가려져 한국인 여부와 정확한 얼굴 일치는 확인할 수 없다. 검은 머리는 위쪽으로 묶었고 목은 맨살로 드러난다. 밝은 바지와 허리끈, 짚신은 참고 의상 계열이지만 저고리가 없어 상반신이 완전히 노출되어 있다. 등에 벌어진 붉은 상처와 피부의 때가 명확하다. 얼굴의 피와 흙은 뒷모습 때문에 평가할 수 없다. 오른손에는 피로 보이는 붉은 얼룩이 있는 낫이 있다. 편백과 낮의 자연광은 맞으며 읽을 수 있는 글자는 없다.",
        "hard_violations": [],
        "physics": "오른발이 흙길에 닿아 몸을 지탱하고 왼발은 뒤에서 들려 밑창이 보인다. 걷는 도중 가능한 자세이며 몸을 지지하는 발이 명확하다. 다만 절뚝이는 보행의 특징은 확실하지 않다. 오른손이 낫 손잡이를 잡고 있어 낫이 떠 있지 않다. 천막 잔해는 나무와 지면에 걸쳐 있고 솥은 돌 화덕에 놓여 있어 주변 물체의 지지도 자연스럽다."
       }
      ]
     },
     "ok": true
    }
   ],
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "slot_winner_match": false,
   "slot_winner": {
    "gemini-pro": "A",
    "gpt-high": "B"
   },
   "route": "cross_slot_combined"
  },
  "dual": {
   "models": [
    "gemini-pro",
    "gpt-high"
   ],
   "normalized": {
    "A": 1.857,
    "B": 1.5
   },
   "adjusted": {
    "A": 1.857,
    "B": 1.5
   },
   "violations": {},
   "per_model_winner": {
    "gemini-pro": "A",
    "gpt-high": "B"
   },
   "agreed": false
  },
  "totals": {
   "A": 1857,
   "B": 1500
  },
  "selected": "A",
  "ranking": [
   "A",
   "B"
  ],
  "verdicts": [
   {
    "label": "A",
    "score": 1857,
    "verdict_ko": "지정된 로케이션의 디테일(무너진 텐트와 가마솥)을 완벽히 재현하고 굳건히 걷는 뒷모습을 정확히 연출했으나, 상의를 완전히 탈의하여 복장 규정(우선순위 4)에서 감점이 발생했습니다."
   },
   {
    "label": "B",
    "score": 1500,
    "verdict_ko": "찢어진 한복과 상처의 묘사는 좋으나, 로케이션 사진의 주요 배경 구조물(야영지 흔적)을 완전히 누락하였고 고개를 뒤로 돌려 프롬프트가 요구한 '앞을 향해 걷는 굳건한 뒷모습'의 구도를 훼손했습니다."
   }
  ],
  "refs": [
   {
    "label": "LOCATION PHOTOGRAPH — the exact place of this shot: its architecture, materials, fixed features and lighting mood are spatial truth; stage the moment inside this place. Never copy its camera framing.",
    "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/db00517b-165b-4005-a287-f28936b59b95/images/272bd7af-9294-4490-98c6-88a40e9deacc/scene/recipe/groupbg_final_cypress_passage_424bfd.png",
    "asset_id": "cc46d830-70ab-4b69-803c-f0db7dc29047",
    "role": "bgfirst_group_bg"
   },
   {
    "label": "CHARACTER REFERENCE — 안규홍: the exact person appearing in this shot; match face, hair and build exactly.",
    "path": "<bytes:1028384>",
    "asset_id": "3b5894db-ce55-4f34-b966-5e725c3aa952",
    "role": "character_ref"
   }
  ],
  "critique_skipped": true,
  "shot_run_produced": false,
  "shot_run_uid": "06aa1447-9514-7c66-b829-5a1dd1d5820e",
  "bgfirst": {
   "bg_path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/db00517b-165b-4005-a287-f28936b59b95/images/272bd7af-9294-4490-98c6-88a40e9deacc/scene/recipe/S66sh22__bgfirst_bg.png",
   "bg_asset_id": "7a3d2d0c-bd05-4b60-8194-d5bdceca9381",
   "bg_record_key": "S66sh22::bgfirst_bg",
   "chain_winner": true,
   "authority": "groupbg",
   "group_key": "final_cypress_passage",
   "groupbg_asset_id": "cc46d830-70ab-4b69-803c-f0db7dc29047"
  },
  "ref_mode": "재투영 배경+콘티+엔티티 (2택1: 체인 승)",
  "share_plan": {
   "ref_plan": "background"
  }
 },
 "S66sh22::cine": {
  "applied": false,
  "attempted_at": "2026-09-09T09:39:01.786270+00:00",
  "fingerprint": "3a8a3a488b99923e3375d3eec5fbfec95b43973d9d176730ba0dec50b3d0c6cc",
  "fingerprint_version": 2,
  "provider": "mai",
  "endpoint": "openrouter/chat-completions",
  "model": "microsoft/mai-image-2.6",
  "multiroll_tag": "still_S66sh22_cine",
  "slot": "",
  "pack": "24.202608252115",
  "source_file": "S66sh22_sel.png",
  "source_sha256": "eb9381dd2ab695452a32a7bd903423b4eb73cbc174c60d5dc0e30adcffcbe023",
  "error": "RuntimeError: moderation blocked: {\"message\": \"Input content violated mainline safety policies.\", \"code\": 400}",
  "moderation_refusals": 2,
  "declined": true,
  "declined_reason": "moderation"
 },
 "S18sh1::cine::fb": {
  "applied": true,
  "attempted_at": "2026-09-09T09:22:03.025684+00:00",
  "fingerprint": "33b6197e0f576b0264d4f232bd00437c4a7510e4ee923ef0f533fa18abfc215f",
  "fingerprint_version": 2,
  "provider": "grok",
  "endpoint": "openrouter/chat-completions",
  "model": "x-ai/grok-imagine-image-2.0",
  "multiroll_tag": "still_S18sh1_cine_fb",
  "slot": "fb",
  "pack": "24.202608252115",
  "source_file": "S18sh1_sel.png",
  "source_sha256": "f1ec102f72664b861b034717d815e6ab87e0b11fb44fca2f21f19711d7d286be",
  "file": "S18sh1_cine_fb.png",
  "staged_sha256": "f850e88d2dd9fe16f8096e93b78aaf0082ff8453ca1f6e81636d0d73c4d7de89",
  "latency_ms": 16738
 },
 "S30sh38::cine::fb": {
  "applied": true,
  "attempted_at": "2026-09-09T09:26:39.681490+00:00",
  "fingerprint": "7e7f4bfffdcf813f18c0ad0a04489b45f54771fde82e692ea80728040437b0dd",
  "fingerprint_version": 2,
  "provider": "grok",
  "endpoint": "openrouter/chat-completions",
  "model": "x-ai/grok-imagine-image-2.0",
  "multiroll_tag": "still_S30sh38_cine_fb",
  "slot": "fb",
  "pack": "24.202608252115",
  "source_file": "S30sh38_sel.png",
  "source_sha256": "7b33733d54403bffe679b3f559cb159fac0cc1e8f07b47136fd11ab8ba3573ae",
  "file": "S30sh38_cine_fb.png",
  "staged_sha256": "adc569ae46ecf7833e51491fbff70d8b5c13435b7424875e5754c088d3edab1c",
  "latency_ms": 18779
 },
 "S30sh50::cine::fb": {
  "applied": true,
  "attempted_at": "2026-09-09T09:26:58.821263+00:00",
  "fingerprint": "ff96502188f8adf177c8eadb1cb975de4998a90c6bd9335847761b9c0274fc2b",
  "fingerprint_version": 2,
  "provider": "grok",
  "endpoint": "openrouter/chat-completions",
  "model": "x-ai/grok-imagine-image-2.0",
  "multiroll_tag": "still_S30sh50_cine_fb",
  "slot": "fb",
  "pack": "24.202608252115",
  "source_file": "S30sh50_sel.png",
  "source_sha256": "de615d897afd6a3ae427dd8dcaec423b5465329391caa118eb004af7aff9997f",
  "file": "S30sh50_cine_fb.png",
  "staged_sha256": "3dd7e48cf416ecd78fbea1e3d835bb25d98df74c0994a63c9fdb140c21bf8917",
  "latency_ms": 13152
 },
 "S31sh1::cine::fb": {
  "applied": true,
  "attempted_at": "2026-09-09T09:27:12.518544+00:00",
  "fingerprint": "4ec8f7ad3b3e1285e820637bf42c83582419efe6ef1dccb226bbc07465a93c5e",
  "fingerprint_version": 2,
  "provider": "grok",
  "endpoint": "openrouter/chat-completions",
  "model": "x-ai/grok-imagine-image-2.0",
  "multiroll_tag": "still_S31sh1_cine_fb",
  "slot": "fb",
  "pack": "24.202608252115",
  "source_file": "S31sh1_sel.png",
  "source_sha256": "274b282fdae72f3fb934348c88c15d5ac0444519d30bac93df714691fbea0588",
  "file": "S31sh1_cine_fb.png",
  "staged_sha256": "a7c7fec97ab4db5301078dfc21db738764de6f185a2b5ffcc7f7acad5c719663",
  "latency_ms": 14974
 },
 "S31sh12::cine::fb": {
  "applied": true,
  "attempted_at": "2026-09-09T09:27:27.851530+00:00",
  "fingerprint": "e138c731f4de6b6575f55d940d7825ddb7992de58d84bbca1565d4442e46eaf6",
  "fingerprint_version": 2,
  "provider": "grok",
  "endpoint": "openrouter/chat-completions",
  "model": "x-ai/grok-imagine-image-2.0",
  "multiroll_tag": "still_S31sh12_cine_fb",
  "slot": "fb",
  "pack": "24.202608252115",
  "source_file": "S31sh12_sel.png",
  "source_sha256": "149a3e69065ff8f9e3a412adbbaa26d3bb0a66f5ae1c8030f4724721b8d0f325",
  "file": "S31sh12_cine_fb.png",
  "staged_sha256": "8af8b16114f2c3d7f3f6826f8fc20ce298a9611ba26cc4e33fb5454c21da9816",
  "latency_ms": 15000
 },
 "S32sh1::cine::fb": {
  "applied": true,
  "attempted_at": "2026-09-09T09:27:43.333111+00:00",
  "fingerprint": "30b75d92e267954f9981b800b76dbe72e827d9c2b29bacfbc10daef04a8efa13",
  "fingerprint_version": 2,
  "provider": "grok",
  "endpoint": "openrouter/chat-completions",
  "model": "x-ai/grok-imagine-image-2.0",
  "multiroll_tag": "still_S32sh1_cine_fb",
  "slot": "fb",
  "pack": "24.202608252115",
  "source_file": "S32sh1_sel.png",
  "source_sha256": "691084ad969e12835d84eebaf5fe5c22b3cf4f3d6082da6b86817f90dfa3e6a5",
  "file": "S32sh1_cine_fb.png",
  "staged_sha256": "e3b61203e8f23cad1a9574e11e626351b9fa952f416a6f7ce38784e2d65e65e1",
  "latency_ms": 22751
 },
 "S39sh18::cine::fb": {
  "applied": true,
  "attempted_at": "2026-09-09T09:28:09.181713+00:00",
  "fingerprint": "3c80056faeb81d8ac0375aab07ee0a90db0636858dcac342d37e6c22f04a444b",
  "fingerprint_version": 2,
  "provider": "grok",
  "endpoint": "openrouter/chat-completions",
  "model": "x-ai/grok-imagine-image-2.0",
  "multiroll_tag": "still_S39sh18_cine_fb",
  "slot": "fb",
  "pack": "24.202608252115",
  "source_file": "S39sh18_sel.png",
  "source_sha256": "95c32f1981917fd5e86e59622b646f3f493ceba2df74bfea3e1748a7853fc884",
  "file": "S39sh18_cine_fb.png",
  "staged_sha256": "2bb4589e58a83ab905af881a8bb70a8f17a4873becabf213993a8b68d34e48ac",
  "latency_ms": 15462
 },
 "S40sh4::cine::fb": {
  "applied": true,
  "attempted_at": "2026-09-09T09:28:25.172266+00:00",
  "fingerprint": "00d78b655f7dbc85d233420053ea24c3fb80df3ae85afab1ba87acca4b85f7d9",
  "fingerprint_version": 2,
  "provider": "grok",
  "endpoint": "openrouter/chat-completions",
  "model": "x-ai/grok-imagine-image-2.0",
  "multiroll_tag": "still_S40sh4_cine_fb",
  "slot": "fb",
  "pack": "24.202608252115",
  "source_file": "S40sh4_sel.png",
  "source_sha256": "4a1b454883eeb00dee59e0aac98e014d7b2ed63d7ca9f4710fac759499d5f679",
  "file": "S40sh4_cine_fb.png",
  "staged_sha256": "eed6e8aa02de99c163e095a54b5662bf5f6550edf39eb53d7457a22c86f91ecc",
  "latency_ms": 14137
 },
 "S42sh4::cine::fb": {
  "applied": true,
  "attempted_at": "2026-09-09T09:28:40.380105+00:00",
  "fingerprint": "4e8ed308073c3f3baa3aa0dbedff15118f195bc2a19ac9b9a51eeab009a00a6c",
  "fingerprint_version": 2,
  "provider": "grok",
  "endpoint": "openrouter/chat-completions",
  "model": "x-ai/grok-imagine-image-2.0",
  "multiroll_tag": "still_S42sh4_cine_fb",
  "slot": "fb",
  "pack": "24.202608252115",
  "source_file": "S42sh4_sel.png",
  "source_sha256": "9256f30bc06f6ed59c964bdf542a19e8beb2aad59b07f0994d558ce3a9459869",
  "file": "S42sh4_cine_fb.png",
  "staged_sha256": "a019860c4130d5413762d2ae27c33b0fea1784e8655e40c05026cf6fc2dd0ea6",
  "latency_ms": 14157
 },
 "S44sh10::cine::fb": {
  "applied": true,
  "attempted_at": "2026-09-09T09:30:32.642955+00:00",
  "fingerprint": "4d986fac325ad733a5136d6583b8ce26ad5cb34a7dda4b8e2102cbda0125600a",
  "fingerprint_version": 2,
  "provider": "grok",
  "endpoint": "openrouter/chat-completions",
  "model": "x-ai/grok-imagine-image-2.0",
  "multiroll_tag": "still_S44sh10_cine_fb",
  "slot": "fb",
  "pack": "24.202608252115",
  "source_file": "S44sh10_sel.png",
  "source_sha256": "b07313fbf4cb4ee61da376bd6b6eeb1722197a05708258eaadd338cdb0db44f8",
  "file": "S44sh10_cine_fb.png",
  "staged_sha256": "47cd8d69167399bfe410367288c6bcf2292999288ed29a8da8fbb07a68d5221b",
  "latency_ms": 13079
 },
 "S45sh5::cine::fb": {
  "applied": true,
  "attempted_at": "2026-09-09T09:31:11.411973+00:00",
  "fingerprint": "4739b8b5183acf8516268b5822cbd10aae3e0c103fa28a04019b516b900ae9a7",
  "fingerprint_version": 2,
  "provider": "grok",
  "endpoint": "openrouter/chat-completions",
  "model": "x-ai/grok-imagine-image-2.0",
  "multiroll_tag": "still_S45sh5_cine_fb",
  "slot": "fb",
  "pack": "24.202608252115",
  "source_file": "S45sh5_sel.png",
  "source_sha256": "a0d58518c872e0ee560399a834309fc74069cc76215891e900f97c4ec8168ca0",
  "file": "S45sh5_cine_fb.png",
  "staged_sha256": "dfe9ac5e8defe6946ed3b7ae9d4a58b2024b1cee05f903320e837817127986a9",
  "latency_ms": 16953
 },
 "S51sh2::cine::fb": {
  "applied": true,
  "attempted_at": "2026-09-09T09:31:50.324697+00:00",
  "fingerprint": "639d3ebe4048acb14d3128173ca45bc9c7baaba494d2e2448823dc42a4bc7362",
  "fingerprint_version": 2,
  "provider": "grok",
  "endpoint": "openrouter/chat-completions",
  "model": "x-ai/grok-imagine-image-2.0",
  "multiroll_tag": "still_S51sh2_cine_fb",
  "slot": "fb",
  "pack": "24.202608252115",
  "source_file": "S51sh2_sel.png",
  "source_sha256": "56c4af61362044bd408de3a9acef5769e52629ae40dc03cb672348e05dd385e0",
  "file": "S51sh2_cine_fb.png",
  "staged_sha256": "2bf16ec4b171141c3b2303a9648b8e38aa739ac11d496fab46e2075e0c134000",
  "latency_ms": 16905
 },
 "S53sh2::cine::fb": {
  "applied": true,
  "attempted_at": "2026-09-09T09:32:26.009088+00:00",
  "fingerprint": "fe9faf4bf8ad9dc4c01e343e569f8d60f4774717282828892b1f0ab90bc748f1",
  "fingerprint_version": 2,
  "provider": "grok",
  "endpoint": "openrouter/chat-completions",
  "model": "x-ai/grok-imagine-image-2.0",
  "multiroll_tag": "still_S53sh2_cine_fb",
  "slot": "fb",
  "pack": "24.202608252115",
  "source_file": "S53sh2_sel.png",
  "source_sha256": "5a7b02e31818e0acda3c9f98561c1eb6b8706540b45b3804d7ba450ac9dab374",
  "file": "S53sh2_cine_fb.png",
  "staged_sha256": "1e43bdc0fecdf2a5c6aa88f5b0cbf4a35a429ac48e712893f657215abbdf4f09",
  "latency_ms": 16659
 },
 "S55sh34::cine::fb": {
  "applied": true,
  "attempted_at": "2026-09-09T09:33:10.082697+00:00",
  "fingerprint": "bc50b1f06a8024c06a2d38d1b6f3e36b87c377d60937786046c064f6ba541934",
  "fingerprint_version": 2,
  "provider": "grok",
  "endpoint": "openrouter/chat-completions",
  "model": "x-ai/grok-imagine-image-2.0",
  "multiroll_tag": "still_S55sh34_cine_fb",
  "slot": "fb",
  "pack": "24.202608252115",
  "source_file": "S55sh34_sel.png",
  "source_sha256": "a940cb9d728311a5720aff64b8e0ee656771b5a7c80b26e0b0e491695d007c80",
  "file": "S55sh34_cine_fb.png",
  "staged_sha256": "94cd4cf4821ae07563045d5b2aa9a22433fffb20ca2cc75c21766e9c619c6301",
  "latency_ms": 22656
 },
 "S60sh11::cine::fb": {
  "applied": true,
  "attempted_at": "2026-09-09T09:35:48.955739+00:00",
  "fingerprint": "5d16ac46e6c58f662faff33fc15fec95178fa6ee47f7bc019148365e456c6d4f",
  "fingerprint_version": 2,
  "provider": "grok",
  "endpoint": "openrouter/chat-completions",
  "model": "x-ai/grok-imagine-image-2.0",
  "multiroll_tag": "still_S60sh11_cine_fb",
  "slot": "fb",
  "pack": "24.202608252115",
  "source_file": "S60sh11_sel.png",
  "source_sha256": "eca46a50e26d507273049bf846984e3340d970caef5f6ad7f1a60ede82304a40",
  "file": "S60sh11_cine_fb.png",
  "staged_sha256": "6bb845386adae7087838ec7a6b28c929201ca4a4aae1ce314d59b5068af4da69",
  "latency_ms": 13350
 },
 "S60sh12::cine::fb": {
  "applied": true,
  "attempted_at": "2026-09-09T09:36:25.005823+00:00",
  "fingerprint": "8c775ecbada79460526624ceb32e8fa5ad22745e79382268d66bb1c5828a7a3a",
  "fingerprint_version": 2,
  "provider": "grok",
  "endpoint": "openrouter/chat-completions",
  "model": "x-ai/grok-imagine-image-2.0",
  "multiroll_tag": "still_S60sh12_cine_fb",
  "slot": "fb",
  "pack": "24.202608252115",
  "source_file": "S60sh12_sel.png",
  "source_sha256": "97a4a2fdd1c34a68c50e496eff39c8e5c428ab56df863333e0626cf21c83418f",
  "file": "S60sh12_cine_fb.png",
  "staged_sha256": "63a04801a5c4fa984bb898559721e04b70615a7e45ae8f8855b057473a76c236",
  "latency_ms": 15028
 },
 "S61sh15::cine::fb": {
  "applied": true,
  "attempted_at": "2026-09-09T09:37:01.318285+00:00",
  "fingerprint": "c8efd4040ce32561603999b09fd745a7ed99317f96a136bc177fd0ca0b33d816",
  "fingerprint_version": 2,
  "provider": "grok",
  "endpoint": "openrouter/chat-completions",
  "model": "x-ai/grok-imagine-image-2.0",
  "multiroll_tag": "still_S61sh15_cine_fb",
  "slot": "fb",
  "pack": "24.202608252115",
  "source_file": "S61sh15_sel.png",
  "source_sha256": "513ba4cbe189a2eaf0fe2ecb6481c107194717a5bd63aedd59ecd69e6fc0e195",
  "file": "S61sh15_cine_fb.png",
  "staged_sha256": "4f6ee3c0bae744dcaaf7142fa37807b1045f8971ee18a0ee61585fc467d3e086",
  "latency_ms": 16614
 },
 "S62sh12::cine::fb": {
  "applied": true,
  "attempted_at": "2026-09-09T09:37:38.273124+00:00",
  "fingerprint": "2300cea6d91c88c56ad85bd5deedb848678689779c4fc7097a0fa10c059eb739",
  "fingerprint_version": 2,
  "provider": "grok",
  "endpoint": "openrouter/chat-completions",
  "model": "x-ai/grok-imagine-image-2.0",
  "multiroll_tag": "still_S62sh12_cine_fb",
  "slot": "fb",
  "pack": "24.202608252115",
  "source_file": "S62sh12_sel.png",
  "source_sha256": "042cfb66a41f5cd25492ccacc73e7816554d90b38db8f45d7ff422cc9933ef86",
  "file": "S62sh12_cine_fb.png",
  "staged_sha256": "40f908ac910daa38274510b0754edc4bd25b601a639600b113ae8c2977ab52c5",
  "latency_ms": 18600
 },
 "S66sh9::cine::fb": {
  "applied": true,
  "attempted_at": "2026-09-09T09:38:45.167870+00:00",
  "fingerprint": "7974653d8e8b09c9e0a09e3d0e6b03f940494b43aeb5cc844198c6c18269b1ab",
  "fingerprint_version": 2,
  "provider": "grok",
  "endpoint": "openrouter/chat-completions",
  "model": "x-ai/grok-imagine-image-2.0",
  "multiroll_tag": "still_S66sh9_cine_fb",
  "slot": "fb",
  "pack": "24.202608252115",
  "source_file": "S66sh9_sel.png",
  "source_sha256": "4212be04f405f462f1bd917b0a624d30c109c04810cb9d70cafc6b65b171f51c",
  "file": "S66sh9_cine_fb.png",
  "staged_sha256": "e8f67cd756c67e53c5d8e97f19af5686c43279dc78f7a59a20ee46261cce184e",
  "latency_ms": 16078
 },
 "S66sh22::cine::fb": {
  "applied": true,
  "attempted_at": "2026-09-09T09:39:28.531237+00:00",
  "fingerprint": "715fa185c7ee18ebcbadedac515d3e7d33d0dc4f94db8e80c858b80e35dd1293",
  "fingerprint_version": 2,
  "provider": "grok",
  "endpoint": "openrouter/chat-completions",
  "model": "x-ai/grok-imagine-image-2.0",
  "multiroll_tag": "still_S66sh22_cine_fb",
  "slot": "fb",
  "pack": "24.202608252115",
  "source_file": "S66sh22_sel.png",
  "source_sha256": "eb9381dd2ab695452a32a7bd903423b4eb73cbc174c60d5dc0e30adcffcbe023",
  "file": "S66sh22_cine_fb.png",
  "staged_sha256": "28810365a4328894f41345b2e3feffbc9f476401e94c4c6dc81521ac9d590bba",
  "latency_ms": 18338
 },
 "S18sh1::cine::fb_grok": {
  "applied": true,
  "attempted_at": "2026-09-09T11:20:53.581051+00:00",
  "fingerprint": "33b6197e0f576b0264d4f232bd00437c4a7510e4ee923ef0f533fa18abfc215f",
  "fingerprint_version": 2,
  "provider": "grok",
  "endpoint": "openrouter/chat-completions",
  "model": "x-ai/grok-imagine-image-2.0",
  "multiroll_tag": "still_S18sh1_cine_fb_grok",
  "slot": "fb_grok",
  "pack": "24.202608252115",
  "source_file": "S18sh1_sel.png",
  "source_sha256": "f1ec102f72664b861b034717d815e6ab87e0b11fb44fca2f21f19711d7d286be",
  "file": "S18sh1_cine_fb_grok.png",
  "staged_sha256": "5a15ee60a54a7befcaf799950e42939d96a30337d5767eb064f677ae4f3b2e49",
  "latency_ms": 15221
 },
 "S30sh38::cine::fb_grok": {
  "applied": true,
  "attempted_at": "2026-09-09T11:25:37.421011+00:00",
  "fingerprint": "7e7f4bfffdcf813f18c0ad0a04489b45f54771fde82e692ea80728040437b0dd",
  "fingerprint_version": 2,
  "provider": "grok",
  "endpoint": "openrouter/chat-completions",
  "model": "x-ai/grok-imagine-image-2.0",
  "multiroll_tag": "still_S30sh38_cine_fb_grok",
  "slot": "fb_grok",
  "pack": "24.202608252115",
  "source_file": "S30sh38_sel.png",
  "source_sha256": "7b33733d54403bffe679b3f559cb159fac0cc1e8f07b47136fd11ab8ba3573ae",
  "file": "S30sh38_cine_fb_grok.png",
  "staged_sha256": "959cb39a739f882e849988d2a17fa128eb5fa4140cfbbc264ace5e1ff1b6dfd0",
  "latency_ms": 13933
 },
 "S30sh50::cine::fb_grok": {
  "applied": true,
  "attempted_at": "2026-09-09T11:25:51.717603+00:00",
  "fingerprint": "ff96502188f8adf177c8eadb1cb975de4998a90c6bd9335847761b9c0274fc2b",
  "fingerprint_version": 2,
  "provider": "grok",
  "endpoint": "openrouter/chat-completions",
  "model": "x-ai/grok-imagine-image-2.0",
  "multiroll_tag": "still_S30sh50_cine_fb_grok",
  "slot": "fb_grok",
  "pack": "24.202608252115",
  "source_file": "S30sh50_sel.png",
  "source_sha256": "de615d897afd6a3ae427dd8dcaec423b5465329391caa118eb004af7aff9997f",
  "file": "S30sh50_cine_fb_grok.png",
  "staged_sha256": "193fc88bcd5a99ceed10e41b86a6b0910e031252aa35c757df870b138dc90bf5",
  "latency_ms": 17380
 },
 "S31sh1::cine::fb_grok": {
  "applied": true,
  "attempted_at": "2026-09-09T11:26:09.619651+00:00",
  "fingerprint": "4ec8f7ad3b3e1285e820637bf42c83582419efe6ef1dccb226bbc07465a93c5e",
  "fingerprint_version": 2,
  "provider": "grok",
  "endpoint": "openrouter/chat-completions",
  "model": "x-ai/grok-imagine-image-2.0",
  "multiroll_tag": "still_S31sh1_cine_fb_grok",
  "slot": "fb_grok",
  "pack": "24.202608252115",
  "source_file": "S31sh1_sel.png",
  "source_sha256": "274b282fdae72f3fb934348c88c15d5ac0444519d30bac93df714691fbea0588",
  "file": "S31sh1_cine_fb_grok.png",
  "staged_sha256": "72ab43dd03fd031901b1d1c22bb99ac4515719ae0780d2eb91c1a6a7f70d7cb1",
  "latency_ms": 14670
 },
 "S31sh12::cine::fb_grok": {
  "applied": true,
  "attempted_at": "2026-09-09T11:26:24.644059+00:00",
  "fingerprint": "e138c731f4de6b6575f55d940d7825ddb7992de58d84bbca1565d4442e46eaf6",
  "fingerprint_version": 2,
  "provider": "grok",
  "endpoint": "openrouter/chat-completions",
  "model": "x-ai/grok-imagine-image-2.0",
  "multiroll_tag": "still_S31sh12_cine_fb_grok",
  "slot": "fb_grok",
  "pack": "24.202608252115",
  "source_file": "S31sh12_sel.png",
  "source_sha256": "149a3e69065ff8f9e3a412adbbaa26d3bb0a66f5ae1c8030f4724721b8d0f325",
  "file": "S31sh12_cine_fb_grok.png",
  "staged_sha256": "2d3547d6f2faf198f82b297df485448d14007b6854d4cf5cf34bb95a35e4ff4b",
  "latency_ms": 17091
 },
 "S32sh1::cine::fb_grok": {
  "applied": true,
  "attempted_at": "2026-09-09T11:26:42.193151+00:00",
  "fingerprint": "30b75d92e267954f9981b800b76dbe72e827d9c2b29bacfbc10daef04a8efa13",
  "fingerprint_version": 2,
  "provider": "grok",
  "endpoint": "openrouter/chat-completions",
  "model": "x-ai/grok-imagine-image-2.0",
  "multiroll_tag": "still_S32sh1_cine_fb_grok",
  "slot": "fb_grok",
  "pack": "24.202608252115",
  "source_file": "S32sh1_sel.png",
  "source_sha256": "691084ad969e12835d84eebaf5fe5c22b3cf4f3d6082da6b86817f90dfa3e6a5",
  "file": "S32sh1_cine_fb_grok.png",
  "staged_sha256": "dae2def47e28d352877c1940d54185ad6e2cbd51634a6b224ed31d697121d8a2",
  "latency_ms": 17317
 },
 "S39sh18::cine::fb_grok": {
  "applied": true,
  "attempted_at": "2026-09-09T11:27:02.867024+00:00",
  "fingerprint": "3c80056faeb81d8ac0375aab07ee0a90db0636858dcac342d37e6c22f04a444b",
  "fingerprint_version": 2,
  "provider": "grok",
  "endpoint": "openrouter/chat-completions",
  "model": "x-ai/grok-imagine-image-2.0",
  "multiroll_tag": "still_S39sh18_cine_fb_grok",
  "slot": "fb_grok",
  "pack": "24.202608252115",
  "source_file": "S39sh18_sel.png",
  "source_sha256": "95c32f1981917fd5e86e59622b646f3f493ceba2df74bfea3e1748a7853fc884",
  "file": "S39sh18_cine_fb_grok.png",
  "staged_sha256": "6fbbe9c71a29760f6ffa616625fa0d8024eb047de37312fe53e965d7941ac473",
  "latency_ms": 17428
 },
 "S40sh4::cine::fb_grok": {
  "applied": true,
  "attempted_at": "2026-09-09T11:27:20.857461+00:00",
  "fingerprint": "00d78b655f7dbc85d233420053ea24c3fb80df3ae85afab1ba87acca4b85f7d9",
  "fingerprint_version": 2,
  "provider": "grok",
  "endpoint": "openrouter/chat-completions",
  "model": "x-ai/grok-imagine-image-2.0",
  "multiroll_tag": "still_S40sh4_cine_fb_grok",
  "slot": "fb_grok",
  "pack": "24.202608252115",
  "source_file": "S40sh4_sel.png",
  "source_sha256": "4a1b454883eeb00dee59e0aac98e014d7b2ed63d7ca9f4710fac759499d5f679",
  "file": "S40sh4_cine_fb_grok.png",
  "staged_sha256": "8cb0d4c22a933ca366f49df83b9a0e8176be261845bfe9d2c86074c457f0c95f",
  "latency_ms": 13383
 },
 "S42sh4::cine::fb_grok": {
  "applied": true,
  "attempted_at": "2026-09-09T11:27:35.413583+00:00",
  "fingerprint": "4e8ed308073c3f3baa3aa0dbedff15118f195bc2a19ac9b9a51eeab009a00a6c",
  "fingerprint_version": 2,
  "provider": "grok",
  "endpoint": "openrouter/chat-completions",
  "model": "x-ai/grok-imagine-image-2.0",
  "multiroll_tag": "still_S42sh4_cine_fb_grok",
  "slot": "fb_grok",
  "pack": "24.202608252115",
  "source_file": "S42sh4_sel.png",
  "source_sha256": "9256f30bc06f6ed59c964bdf542a19e8beb2aad59b07f0994d558ce3a9459869",
  "file": "S42sh4_cine_fb_grok.png",
  "staged_sha256": "4adac0017f8caa914b5d91274588029f1975faa35121872390604d99eedae758",
  "latency_ms": 18664
 },
 "S44sh10::cine::fb_grok": {
  "applied": true,
  "attempted_at": "2026-09-09T11:29:53.978807+00:00",
  "fingerprint": "4d986fac325ad733a5136d6583b8ce26ad5cb34a7dda4b8e2102cbda0125600a",
  "fingerprint_version": 2,
  "provider": "grok",
  "endpoint": "openrouter/chat-completions",
  "model": "x-ai/grok-imagine-image-2.0",
  "multiroll_tag": "still_S44sh10_cine_fb_grok",
  "slot": "fb_grok",
  "pack": "24.202608252115",
  "source_file": "S44sh10_sel.png",
  "source_sha256": "b07313fbf4cb4ee61da376bd6b6eeb1722197a05708258eaadd338cdb0db44f8",
  "file": "S44sh10_cine_fb_grok.png",
  "staged_sha256": "152845adacfab6b3f633d1ae94ad7a5646736bf4b160d2fad5ab53675d228e57",
  "latency_ms": 13047
 },
 "S45sh5::cine::fb_grok": {
  "applied": true,
  "attempted_at": "2026-09-09T11:30:07.730887+00:00",
  "fingerprint": "4739b8b5183acf8516268b5822cbd10aae3e0c103fa28a04019b516b900ae9a7",
  "fingerprint_version": 2,
  "provider": "grok",
  "endpoint": "openrouter/chat-completions",
  "model": "x-ai/grok-imagine-image-2.0",
  "multiroll_tag": "still_S45sh5_cine_fb_grok",
  "slot": "fb_grok",
  "pack": "24.202608252115",
  "source_file": "S45sh5_sel.png",
  "source_sha256": "a0d58518c872e0ee560399a834309fc74069cc76215891e900f97c4ec8168ca0",
  "file": "S45sh5_cine_fb_grok.png",
  "staged_sha256": "5fb799396305e37c367e387af848605efd82816218011301f30d8295c54e8981",
  "latency_ms": 19515
 },
 "S51sh2::cine::fb_grok": {
  "applied": true,
  "attempted_at": "2026-09-09T11:30:30.261919+00:00",
  "fingerprint": "639d3ebe4048acb14d3128173ca45bc9c7baaba494d2e2448823dc42a4bc7362",
  "fingerprint_version": 2,
  "provider": "grok",
  "endpoint": "openrouter/chat-completions",
  "model": "x-ai/grok-imagine-image-2.0",
  "multiroll_tag": "still_S51sh2_cine_fb_grok",
  "slot": "fb_grok",
  "pack": "24.202608252115",
  "source_file": "S51sh2_sel.png",
  "source_sha256": "56c4af61362044bd408de3a9acef5769e52629ae40dc03cb672348e05dd385e0",
  "file": "S51sh2_cine_fb_grok.png",
  "staged_sha256": "fa6b7e15e46e0d7541bb0aa3abdff83c05c40235a8d035bc57d0fa7a38a22834",
  "latency_ms": 14792
 },
 "S53sh2::cine::fb_grok": {
  "applied": true,
  "attempted_at": "2026-09-09T11:30:45.916716+00:00",
  "fingerprint": "fe9faf4bf8ad9dc4c01e343e569f8d60f4774717282828892b1f0ab90bc748f1",
  "fingerprint_version": 2,
  "provider": "grok",
  "endpoint": "openrouter/chat-completions",
  "model": "x-ai/grok-imagine-image-2.0",
  "multiroll_tag": "still_S53sh2_cine_fb_grok",
  "slot": "fb_grok",
  "pack": "24.202608252115",
  "source_file": "S53sh2_sel.png",
  "source_sha256": "5a7b02e31818e0acda3c9f98561c1eb6b8706540b45b3804d7ba450ac9dab374",
  "file": "S53sh2_cine_fb_grok.png",
  "staged_sha256": "11d5b2ec75526083aa79bb24f5c20d5e9d34f778f7ab9e78ef2a9603caf59995",
  "latency_ms": 18707
 },
 "S55sh34::cine::fb_grok": {
  "applied": true,
  "attempted_at": "2026-09-09T11:31:05.486038+00:00",
  "fingerprint": "bc50b1f06a8024c06a2d38d1b6f3e36b87c377d60937786046c064f6ba541934",
  "fingerprint_version": 2,
  "provider": "grok",
  "endpoint": "openrouter/chat-completions",
  "model": "x-ai/grok-imagine-image-2.0",
  "multiroll_tag": "still_S55sh34_cine_fb_grok",
  "slot": "fb_grok",
  "pack": "24.202608252115",
  "source_file": "S55sh34_sel.png",
  "source_sha256": "a940cb9d728311a5720aff64b8e0ee656771b5a7c80b26e0b0e491695d007c80",
  "file": "S55sh34_cine_fb_grok.png",
  "staged_sha256": "384d3df501dc8a8ddfed50a3614bfb23e848b6619761e7ba9d29485b9d572db1",
  "latency_ms": 20056
 },
 "S59sh2::cine::fb_grok": {
  "applied": false,
  "attempted_at": "2026-09-09T11:31:27.248629+00:00",
  "fingerprint": "2edc5a8acffe9a30d3a407e78390a21e6b7ec4254c692e71aece1adfda0a0d6a",
  "fingerprint_version": 2,
  "provider": "grok",
  "endpoint": "openrouter/chat-completions",
  "model": "x-ai/grok-imagine-image-2.0",
  "multiroll_tag": "still_S59sh2_cine_fb_grok",
  "slot": "fb_grok",
  "pack": "24.202608252115",
  "source_file": "S59sh2_sel.png",
  "source_sha256": "c05391a9e15768a293b2998aa2df14dd38aefa6d9a894ef717eefef8ed324047",
  "error": "RuntimeError: moderation blocked: {\"message\": \"Generated image rejected by content moderation.\", \"code\": 400}",
  "moderation_refusals": 1
 },
 "S59sh2::cine::fb_seedream": {
  "applied": true,
  "attempted_at": "2026-09-09T11:31:42.032157+00:00",
  "fingerprint": "5677a02a0db4a49716cd12fdfcd926e6f4adece21e8da02ab7bca6790bd3f200",
  "fingerprint_version": 2,
  "provider": "seedream",
  "endpoint": "openrouter/images",
  "model": "bytedance-seed/seedream-5-0-pro",
  "multiroll_tag": "still_S59sh2_cine_fb_seedream",
  "slot": "fb_seedream",
  "pack": "24.202608252115",
  "source_file": "S59sh2_sel.png",
  "source_sha256": "c05391a9e15768a293b2998aa2df14dd38aefa6d9a894ef717eefef8ed324047",
  "file": "S59sh2_cine_fb_seedream.png",
  "staged_sha256": "3fe315977c9d0250a715d8a7e95e3950d41077198f97d7db17740aa902f8b3a0",
  "latency_ms": 115847
 },
 "S60sh2::cine::fb_grok": {
  "applied": true,
  "attempted_at": "2026-09-09T11:33:38.501624+00:00",
  "fingerprint": "627cdaa6f0633512bf7a7e8f5080302f20d8d46233158c3f65f6f23d5b2589ea",
  "fingerprint_version": 2,
  "provider": "grok",
  "endpoint": "openrouter/chat-completions",
  "model": "x-ai/grok-imagine-image-2.0",
  "multiroll_tag": "still_S60sh2_cine_fb_grok",
  "slot": "fb_grok",
  "pack": "24.202608252115",
  "source_file": "S60sh2_sel.png",
  "source_sha256": "978fbbcbd84fd11dd411f6d34fe361a8e06949b63164f53e1b733b8ac130f669",
  "file": "S60sh2_cine_fb_grok.png",
  "staged_sha256": "10a1806e8fccd4dea4aa6c0cab2c48b48911147fab1bc8348cc5a5721c8eced4",
  "latency_ms": 16532
 },
 "S60sh11::cine::fb_grok": {
  "applied": true,
  "attempted_at": "2026-09-09T11:33:55.362534+00:00",
  "fingerprint": "5d16ac46e6c58f662faff33fc15fec95178fa6ee47f7bc019148365e456c6d4f",
  "fingerprint_version": 2,
  "provider": "grok",
  "endpoint": "openrouter/chat-completions",
  "model": "x-ai/grok-imagine-image-2.0",
  "multiroll_tag": "still_S60sh11_cine_fb_grok",
  "slot": "fb_grok",
  "pack": "24.202608252115",
  "source_file": "S60sh11_sel.png",
  "source_sha256": "eca46a50e26d507273049bf846984e3340d970caef5f6ad7f1a60ede82304a40",
  "file": "S60sh11_cine_fb_grok.png",
  "staged_sha256": "4439a8f9eee3162b3875a75428c42e388ea4ee11daf5433c775cdebbc39b36a2",
  "latency_ms": 15203
 },
 "S60sh12::cine::fb_grok": {
  "applied": true,
  "attempted_at": "2026-09-09T11:34:11.092819+00:00",
  "fingerprint": "8c775ecbada79460526624ceb32e8fa5ad22745e79382268d66bb1c5828a7a3a",
  "fingerprint_version": 2,
  "provider": "grok",
  "endpoint": "openrouter/chat-completions",
  "model": "x-ai/grok-imagine-image-2.0",
  "multiroll_tag": "still_S60sh12_cine_fb_grok",
  "slot": "fb_grok",
  "pack": "24.202608252115",
  "source_file": "S60sh12_sel.png",
  "source_sha256": "97a4a2fdd1c34a68c50e496eff39c8e5c428ab56df863333e0626cf21c83418f",
  "file": "S60sh12_cine_fb_grok.png",
  "staged_sha256": "c0e60df677cc2baab42080ce05d055d8cc510d7721f3c6940dd352f95c3228df",
  "latency_ms": 13191
 },
 "S61sh15::cine::fb_grok": {
  "applied": true,
  "attempted_at": "2026-09-09T11:34:24.737609+00:00",
  "fingerprint": "c8efd4040ce32561603999b09fd745a7ed99317f96a136bc177fd0ca0b33d816",
  "fingerprint_version": 2,
  "provider": "grok",
  "endpoint": "openrouter/chat-completions",
  "model": "x-ai/grok-imagine-image-2.0",
  "multiroll_tag": "still_S61sh15_cine_fb_grok",
  "slot": "fb_grok",
  "pack": "24.202608252115",
  "source_file": "S61sh15_sel.png",
  "source_sha256": "513ba4cbe189a2eaf0fe2ecb6481c107194717a5bd63aedd59ecd69e6fc0e195",
  "file": "S61sh15_cine_fb_grok.png",
  "staged_sha256": "ec94f00eb8d10c87fadb372969572841d08e262be437eeb249691e444d071ff3",
  "latency_ms": 19579
 },
 "S62sh12::cine::fb_grok": {
  "applied": true,
  "attempted_at": "2026-09-09T11:34:44.762761+00:00",
  "fingerprint": "2300cea6d91c88c56ad85bd5deedb848678689779c4fc7097a0fa10c059eb739",
  "fingerprint_version": 2,
  "provider": "grok",
  "endpoint": "openrouter/chat-completions",
  "model": "x-ai/grok-imagine-image-2.0",
  "multiroll_tag": "still_S62sh12_cine_fb_grok",
  "slot": "fb_grok",
  "pack": "24.202608252115",
  "source_file": "S62sh12_sel.png",
  "source_sha256": "042cfb66a41f5cd25492ccacc73e7816554d90b38db8f45d7ff422cc9933ef86",
  "file": "S62sh12_cine_fb_grok.png",
  "staged_sha256": "d868279ebc0afced8c7f508a09da25720d1e01acd671fa6e66ac83b0dc2c35bd",
  "latency_ms": 16908
 },
 "S66sh9::cine::fb_grok": {
  "applied": true,
  "attempted_at": "2026-09-09T11:35:03.489659+00:00",
  "fingerprint": "7974653d8e8b09c9e0a09e3d0e6b03f940494b43aeb5cc844198c6c18269b1ab",
  "fingerprint_version": 2,
  "provider": "grok",
  "endpoint": "openrouter/chat-completions",
  "model": "x-ai/grok-imagine-image-2.0",
  "multiroll_tag": "still_S66sh9_cine_fb_grok",
  "slot": "fb_grok",
  "pack": "24.202608252115",
  "source_file": "S66sh9_sel.png",
  "source_sha256": "4212be04f405f462f1bd917b0a624d30c109c04810cb9d70cafc6b65b171f51c",
  "file": "S66sh9_cine_fb_grok.png",
  "staged_sha256": "ca5136b2698902522631466eee5245ed8b327bc40780a21f37581f75e5be235e",
  "latency_ms": 17495
 },
 "S66sh22::cine::fb_grok": {
  "applied": true,
  "attempted_at": "2026-09-09T11:35:21.497067+00:00",
  "fingerprint": "715fa185c7ee18ebcbadedac515d3e7d33d0dc4f94db8e80c858b80e35dd1293",
  "fingerprint_version": 2,
  "provider": "grok",
  "endpoint": "openrouter/chat-completions",
  "model": "x-ai/grok-imagine-image-2.0",
  "multiroll_tag": "still_S66sh22_cine_fb_grok",
  "slot": "fb_grok",
  "pack": "24.202608252115",
  "source_file": "S66sh22_sel.png",
  "source_sha256": "eb9381dd2ab695452a32a7bd903423b4eb73cbc174c60d5dc0e30adcffcbe023",
  "file": "S66sh22_cine_fb_grok.png",
  "staged_sha256": "dd003f0b4e6a75fb4dced8de387609bc5207f643738916ce1e4fd09fc1a0b7c2",
  "latency_ms": 13585
 }
}